Index

\n Welcome on my awesome homepage.\n{% endblock %}\n```\n\n `{% extends %}` 태그가 여기서 핵심이다. 이 태그는 템플릿 엔진에게 이 템플릿이\n다른 템플릿을 “확장(extends)” 한다라고 알려준다. 템플릿 시스템이 이 템플릿을\n검증할 때, 가장 먼저 부모 템플릿을 찾는다. 그 확장 태그가 템플릿에서 가장 먼저\n있어야 한다. 부모 템플릿에 정의된 블럭의 내용을 보여주려면 `{{ super() }}` 를\n사용하면 된다.\n\n좋은 어플리케이션과 사용자 인터페이스의 모든것은 피드백이다. 사용자가 충분한 피드백을 얻지 못한다면 그들은 결국 그 어플리케이션을 싫어할 것이다. 플라스크는 플래싱 시스템을 가지고 사용자에게 피드백을 주는 정말 간단한 방법을 제공한다. 플래싱 시스템은 기본적으로 요청의 끝에 메시지를 기록하고 그 다음 요청에서만 그 메시지에 접근할 수 있게 한다. 보통은 플래싱을 처리하는 레이아웃 템플릿과 결함되어 사용된다.\n\n## 간단한 플래싱¶\n\n그래서 아래 전체 예제를 준비했다:\n\n```\nfrom flask import Flask, flash, redirect, render_template, \\\n request, url_for\n\napp = Flask(__name__)\napp.secret_key = 'some_secret'\n\n@app.route('/login', methods=['GET', 'POST'])\ndef login():\n error = None\n if request.method == 'POST':\n if request.form['username'] != 'admin' or \\\n request.form['password'] != 'secret':\n error = 'Invalid credentials'\n else:\n flash('You were successfully logged in')\n return redirect(url_for('index'))\n return render_template('login.html', error=error)\n\nif __name__ == \"__main__\":\n app.run()\n```\n\n 그리고 여기에 그 마법을 다룰 `layout.html` 템플릿이 있다: \n\n```\n\n\n{% with messages = get_flashed_messages() %}\n {% if messages %}\n

\n {% for message in messages %}\n
• {{ message }}
\n {% endfor %}\n

Overview

Do you want to log in?\n{% endblock %}\n```\n\n물론 로그인 템플릿도 있다:\n\n```\n{% extends \"layout.html\" %}\n{% block body %}\n

Login

Error: {{ error }}\n {% endif %}\n

\n {% for category, message in messages %}\n
• {{ message }}
\n {% endfor %}\n

jQuery Example

+\n =\n ?\n

calculate server side\n```\n\n여기서는 jQuery가 어떠헥 동작하는지 자세하게 들어가지는 않을 것이고, 단지 위에 있는 일부 코드에 대한 간략한 설명만 있을 것이다:\n\n```\n$(function() { ... })\n```\n\n는 브라우저가 해당 페이지의 기본 구성들을 로딩할 때 실행될 코드를 지정한다. * \n `$('selector')` 는 요소를 선택하고 그 요소에서 실행하게 한다. * \n\n```\nelement.bind('event', func)\n```\n\n는 사용자가 해당 요소에서 클릭했을 때 실행될 함수를 지정한다. 그 함수가 false 를 반환하면, 기본 동작은 시작되지 않을 것이다 (이 경우, # URL로 이동). * \n\n```\n$.getJSON(url, data, func)\n```\n\n은 url 로 GET 요청을 보내고 쿼리 인자로 data 객체의 내요을 보낼 것이다. 일단 데이터가 도착하면, 인자로 반환 값을 가지고 주어진 함수를 호출할 것이다. 여기서는 앞에서 설정한 $SCRIPT_ROOT 변수를 사용할 수 있다는 것에 주목해라. \n여러분이 전체적으로 이해가 안된다면, 깃허브(github)에서 이 예제에 대한 소스 코드.\n\n 플라스크에는 앞에서 나온 HTTP 오류 코드를 가지고 요청을 중단하는 `abort()` 함수가 있다. 그것은 또한 정말 꾸미지 않은 기본적인\n설명을 가진 단순한 흑백의 오류 페이지를 제공할 것이다. \n오류 코드에 따라서 사용자가 실제로 그런 오류를 볼 가능성이 있다.\n\n## 공통 오류 코드¶\n\n다음의 오류 코드는 어플리케이션이 정상적으로 동작했음에도 불구하고 사용자에게 종종 보여지는 것이다:\n\n* 404 Not Found\n * 즐겨쓰는 메시지에 “이봐 친구, 그 URL 입력에 실수가 있어” 가 있다. 인터넷 초보자 조차 그 404를 알고 있는 그렇게 일반적인 것은 다음을 의미한다 : 젠장, 내가 찾고 있는 것이 거기에 없네. 404 페이지에 적어도 index 페이지로 돌아갈 수 있는 링크와 같은 유용한 것이 있도록 하는게 매우 좋은 방식이다.\n * 403 Forbidden\n * 여러분의 웹사이트에 어떤 접근 제어가 있다면, 허용되지 않는 자원에 대해 403 코드를 보내야할 것이다. 그렇기 때문에 사용자가 금지된 자원에 대해 접근하려할 때 사용자가 링크를 잃어버리지 않도록 해야한다.\n * 410 Gone\n * 여러분은 “404 Not Found” 에게 “410 Gone” 이라는 형제가 있었다는 것을 알았는가? 일부 사람들만 실제로 그것을 구현했지만, 그 방식은 전에 존재했지만 현재 삭제된 자원에 대해 404 대신에 401 로 응답하는 것이다. 여러분이 데이터베이스에서 영구적으로 문서를 지우지 않고 삭제됐다고 표시만 한다면, 사용자에게 편의를 제공하며 410 코드를 대신 사용하고 그들이 찾고 있는 것은 영구적으로 삭제됐다는 메시지를 보여줘라.\n * 500 Internal Server Error\n * 보통 프로그래밍 오류나 서버에 한계 부하를 넘었을 때 이 오류가 발생한다. 그 경우에 멋진 페이지를 보여주는 것이 굉장히 좋은 방식인데, 왜냐하면 여러분의 어플리케이션은 머지않아 다시 동작하지 않을 것이기 때문이다 (여기를 또한 살펴봐라: 어플리케이션 에러 로깅하기).\n\n## 오류 핸들러¶\n\n 오류 핸들러는 뷰 함수와 같은 일종의 함수이지만, 그것은 오류가 발생하고 그 오류를 전달했을 때 호출된다. 오류는 대부분 `HTTPException` 이지만, 어떤 경우에는 다른\n오류일 수도 있다: 내부 서버 오류에 대한 핸들러는 그 요류들이 잡히지 않을지\n모르지만 다른 예외 인스턴스에 넘겨질 것이다. 오류 핸들러는 `errorhandler()` 데코레이터와 예외에 대한\n오류 코드를 가지고 등록된다. 플라스크는 오류 코드를 설정하지 않을 것\n이라는 것을 명심하고, 그렇기 때문에 응답을 반환할 때 HTTP 상태 코드를\n제공하도록 해야한다. \n여기에 “404 Page Not Found” 예외에 대한 구현 예제가 있다:\n\n@app.errorhandler(404)\ndef page_not_found(e):\n return render_template('404.html'), 404\n```\n\n예제 템플릿은 다음과 같을 것이다:\n\n```\n{% extends \"layout.html\" %}\n{% block title %}Page Not Found{% endblock %}\n{% block body %}\n

Page Not Found

What you were looking for is just not there.\n

go somewhere nice\n{% endblock %}\n```\n\n플라스크는 보통 데코레이터를 가지고 사용된다. 데코레이터는 간단하고 여러분은 특정 URL에 대해 호출되는 함수 바로 옆에 URL을 놓게 된다. 그러나 이런 방식에 단점이 있다: 데코레이터를 사용하는 여러분의 모든 코드가 가장 앞단에서 임포트되어야 하고 그렇지 않다면 플라스크는 실제로 여러분의 함수를 결코 찾지 못할 것이다.\n\n여러분의 어플리케이션이 빠르게 임포트되어야 한다면 이것은 문제가 될 수 있다. 구글 앱 엔진이나 다른 시스템처럼 그 자체가 그 임포트를 해야할 지도 모른다. 그렇기 때문에 여러분의 어플리케이션이 규모가 커져서 이 방식이 더 이상 맞지 않다는 것을 갑자기 알게된다면 여러분은 중앙집중식 URL 매핑으로 되돌아 갈 수 도 있다.\n 중앙식 URL 맴을 갖도록 활성화하는 시스템은 `add_url_rule()` 함수이다. 데코레이터를 사용하는 대신, 여러분은 어플리케이션의 모든 URL을\n파일로 갖을 수 있다. \n\n## 중앙집중식 URL 맴으로 변환하기¶\n\n현재 어플리케이션이 아래와 같은 모습이라고 상상해보자:\n\n@app.route('/')\ndef index():\n pass\n\n그렇다면 중앙집중 방식에서는 여러분은 데코레이터가 없는 뷰를 가진 하나의 파일 (views.py) 을 가질 것이다:\n\n```\ndef index():\n pass\n\n그렇다면 URL과 함수를 매핑하는 어플리케이션을 설정하는 파일을 다음과 같다:\n\n```\nfrom flask import Flask\nfrom yourapplication import views\napp = Flask(__name__)\napp.add_url_rule('/', view_func=views.index)\napp.add_url_rule('/user/', view_func=views.user)\n```\n\n## 늦은 로딩¶\n\n지금까지 우리는 단지 뷰와 라우팅만 분리했지만, 모듈은 여전히 처음에 로딩된다. 필요할 때 뷰 함수가 실제 로딩되는 기법이 있다. 이것은 함수와 같이 동작하는 도움 클래스와 같이 수행될 수 있지만, 내부적으로 첫 사용에는 실제 함수를 임포는 한다:\n\n```\nfrom werkzeug import import_string, cached_property\n\nclass LazyView(object):\n\n def __init__(self, import_name):\n self.__module__, self.__name__ = import_name.rsplit('.', 1)\n self.import_name = import_name\n\n @cached_property\n def view(self):\n return import_string(self.import_name)\n\n def __call__(self, *args, **kwargs):\n return self.view(*args, **kwargs)\n```\n\n여기서 중요한 것은 __module__ 과 __name__ 알맞게 설정되야 한다. 이것은 여러분이 URL 규칙에 대한 이름을 제공하는 않는 경우에 플라스크가 내부적으로 URL 규칙을 이름짓는 방식을 알기 위해 사용된다.\n\n그리고 나서 여러분은 아래처럼 뷰와 결합할 중앙위치를 정의할 수 있다:\n from flask import Flask from yourapplication.helpers import LazyView app = Flask(__name__) app.add_url_rule(‘/’,\nview_func=LazyView(‘yourapplication.views.index’)) * app.add_url_rule(‘/user/’,\n * view_func=LazyView(‘yourapplication.views.user’))\n 여러분은 키보드로 입력하는 코드의 양이라는 측면에서 프로젝트 명과 점(dot)을 접두어화하고 필요에 따라 LazyView 에 있는 view_func 을 래핑한 것으로 `add_url_rule()` 을 호출하는 함수를 작성함으로 이것을\n더 최적화할 수 있다: \n\n```\ndef url(url_rule, import_name, **options):\n view = LazyView('yourapplication.' + import_name)\n app.add_url_rule(url_rule, view_func=view, **options)\n\nurl('/', 'views.index')\nurl('/user/', 'views.user')\n```\n\n한가지 명심해야할 것은 전후 요청 핸들러가 첫 요청에 올바르게 동작하기 위해 앞단에서 임포트되는 파일에 있어야 한다는 것이다. 나머지 데코레이터도 같은 방식이 적용된다.\n\n기능을 갖춘 DBMS 보다 문서 기반 데이터베이스를 사용하는 것이 요즘은 더 일반적이다. 이번 패턴은 MongoDB와 통합하기 위한 문서 매핑 라이브러리인 MongoKit의 사용법을 보여준다.\n\n이 패턴은 동작하는 MongoDB 서버와 MongoKit 라이브러리가 설치되있는 것을 전제로 한다.\n\nMongoKit을 사용하는 두가지 일반적인 방식이 있다. 여기서 각 방법을 요약할 것이다:\n\n## 선언 부분¶\n\nMongoKit의 기본 동작은 Django 나 SQLAlchemy의 선언적 확장의 공통 방식에 기반을 둔 선언적 방식이다.\n\n아래는 app.py 모듈의 예제이다:\n\n```\nfrom flask import Flask\nfrom mongokit import Connection, Document\n\n# configuration\nMONGODB_HOST = 'localhost'\nMONGODB_PORT = 27017\n\n# create the little application object\napp = Flask(__name__)\napp.config.from_object(__name__)\n\n# connect to the database\nconnection = Connection(app.config['MONGODB_HOST'],\n app.config['MONGODB_PORT'])\n```\n\n여러분의 모델을 정의하기 위해, MongoKit에서 임포트한 Document 클래스는 상속해라. SQLAlchemy 패턴을 봤다면 여러분은 왜 우리가 세션을 갖고 있지 않고 심지어 init_db 함수를 여기서 정의하지 않았는지 궁금해할 지도 모른다. 한편으로, MongoKit은 세션같은 것을 갖지 않는다. 이것은 때때로 더 많이 타이핑을 하지만 엄청나게 빠르다. 다른 면으로, MongoDB는 스키마가 없다. 이것은 여러분이 하나의 입력 질의로부터 어떤 문제도 없이 다음 질의에서 데이터 구조를 변경할 수 있다. MongoKit 또한 스키마가 없지만, 데이터의 무결성을 보장하기 위해 어떤 검증을 구현한다.\n\n여기서 예제 문서가 있다 (예를 들면 이것 또한 app.py 에 넣는다):\n\n```\ndef max_length(length):\n def validate(value):\n if len(value) <= length:\n return True\n raise Exception('%s must be at most %s characters long' % length)\n return validate\n\nclass User(Document):\n structure = {\n 'name': unicode,\n 'email': unicode,\n }\n validators = {\n 'name': max_length(50),\n 'email': max_length(120)\n }\n use_dot_notation = True\n def __repr__(self):\n return '' % (self.name)\n\n# register the User document with our current connection\nconnection.register([User])\n```\n\n이 예제는 여러분에게 스키마 (구조라 불리는) 를 정의하는 법, 최대 문자 길이에 대한 검증자를 보여주고 use_dot_notation 이라 불리는 특별한 MongoKit 기능을 사용한다. 기본 MongoKit 마다 파이썬 딕셔너리 같은 동작을 하지만 use_dot_notation 에 True 로 설정을 하면 여러분은 속성들 사이를 분리하기 위해 점(dot)을 사용해서 어떤 다른 ORM과 근접한 모델을 사용하는 것 처럼 문서를 사용할 수 있다.\n\n여러분은 아래 처럼 데이터베이스에 항목을 넣을 수 있다:\n\n```\n>>> from yourapplication.database import connection\n>>> from yourapplication.models import User\n>>> collection = connection['test'].users\n>>> user = collection.User()\n>>> user['name'] = u'admin'\n>>> user['email'] = u''\n>>> user.save()\n```\n\nMongoKit은 사용된 컬럼 타입에 다소 엄격하고, 여러분은 유니코드인 name 또는 email 에 대한 공통의 str 타입을 사용하지 않아야 한다.\n\n질의하는것 또한 간단하다:\n\n```\n>>> list(collection.User.find())\n[]\n>>> collection.User.find_one({'name': u'admin'})\n\n```\n\n## PyMongo 호환성 계층¶\n\n여러분이 PyMongo를 단지 사용하고 싶다면, MongoKit을 가지고 그것을 할 수 있다. 여러분이 데이터를 얻는데 가장 좋은 성능이 필요하다면 이 프로세스를 사용하면 된다. 이 예제는 Flask와 그것을 연결하는 방법을 보여주는지 않고, 예를 들면 위의 위의 MongoKit 코드를 봐라:\n\n```\nfrom MongoKit import Connection\n\nconnection = Connection()\n```\n\n데이터를 입력하기 위해 여러분은 insert 메소드를 사용할 수있다. 우리는 첫번째로 콜렉션을 얻어야하고, 이것은 SQL 세상에서 테이블과 약간 유사하다.\n\n```\n>>> collection = connection['test'].users\n>>> user = {'name': u'admin', 'email': u''}\n>>> collection.insert(user)\n```\n\nMongoKit는 자동으로 커밋할 것이다.\n\n데이터베이스에 질의하기 위해서, 여러분은 직접 컬렉션을 사용한다.\n\n```\n>>> list(collection.find())\n[{u'_id': ObjectId('4c271729e13823182f000000'), u'name': u'admin', u'email': u''}]\n>>> collection.find_one({'name': u'admin'})\n{u'_id': ObjectId('4c271729e13823182f000000'), u'name': u'admin', u'email': u''}\n```\n\n이 결과 또한 딕셔너리 형태의 객체이다:\n\n```\n>>> r = collection.find_one({'name': u'admin'})\n>>> r['email']\nu'admin'\n```\n\nMongoKit에 대한 더 많은 정보에 대해서, website 으로 가서 살펴봐라.\n\n“파비콘(favicon)”은 탭이나 북마크를 위해 브라우저에서 사용되는 아이콘이다. 이 아이콘은 여러분의 웹사이트를 구분하는데 도움을 주고 그 사이트에 유일한 표식을 준다.\n 일반적인 질문은 플라스크 어플리케이션에 어떻게 파이콘을 추가하는가 이다. 물론 제일 먼저 파이콘이 필요하다. 16 × 16 픽셀과 ICO 파일 형태이어야 한다. 이것은 전제사항은 아니지만 그 형태가 모든 브라우저에서 지원하는 업계 표준이다. 여러분의 static 디렉토리에 그 아이콘을 `favicon.ico` 파일명으로 넣는다. \n자, 그 아이콘을 브라우저에서 찾으려면, 여러분의 HTML에 link 태그를 추가하는게 알맞은 방법이다. 그래서 예를 들면:\n\n```\n\n```\n\n대부분의 브라우저에서는 위의 한줄이 여러분이 해줄 전부이지만, 몇몇 예전 브라우저는 이 표준을 지원하지 않는다. 예전 업계 표준은 웹사이트의 루트에 그 파일을 위치시켜 제공하는 것이다. 여러분의 어플리케이션이 해당 도메인의 루트 경로에 마운트되지 않았다면 루트에서 그 아이콘이 제공되도록 웹서버 설정이 필요하므로 그렇게 할 수 없다면 여러분은 운이 없는 것이다. 그러나 어프리케이션이 루트에 있다면 여러분은 리디렉션으로 간단히 경로를 찾게할 수 있다:\n\n```\napp.add_url_rule('/favicon.ico',\n redirect_to=url_for('static', filename='favicon.ico'))\n```\n\n 여러분이 추가적인 리디렉션 요청을 저장하고 싶다면\n\n```\nsend_from_directory()\n```\n\n를 사용한 뷰 함수 또한 작성할 수 있다: \n\n@app.route('/favicon.ico')\ndef favicon():\n return send_from_directory(os.path.join(app.root_path, 'static'),\n 'favicon.ico', mimetype='image/vnd.microsoft.icon')\n```\n\n명시적인 마임타입(mimetype)을 생략할 수 있고 그 타입은 추측될 것이지만, 그것은 항상 같게 추측될 것이므로 추가적인 추측을 피하기 위해 타입을 지정하는 것이 좋다.\n\n위의 예는 어플리케이션을 통해 아이콘을 제공할 것이고 가능하다면 웹서버 문서를 참고하여 그렇게 할 전담 웹서버를 구성하는 것이 더 좋다.\n\n종종 여러분은 클라이언트로 메모리에 유지하고 싶은 양보다 훨씬 더 큰, 엄청난 양의 데이터를 전송하기를 원한다. 여러분이 그 데이터를 바로 생성하고 있을 때, 파일시스템으로 갔다오지 않고 어떻게 클라이언트로 그것을 전송하는가?\n\n답은 바로 생성기(generators)와 직접 응답이다.\n\n## 기본 사용법¶\n\n이것은 많은 CSV 데이터를 즉각 생성하는 기초적인 뷰 함수이다. 데이터를 생성하는 생성기를 사용하는 내부(inner) 함수를 갖고 그 함수를 호출하면서 응답 객체에 그 함수를 넘기는것이 그 기법이다:\n\n@app.route('/large.csv')\ndef generate_large_csv():\n def generate():\n for row in iter_all_rows():\n yield ','.join(row) + '\\n'\n return Response(generate(), mimetype='text/csv')\n```\n\n 각 `yield` 표현식은 브라우져에 직접 전송된다. 어떤 WSGI 미들웨어는\n스트리밍을 깰수도 있지만, 프로파일러를 가진 디버그 환경과 여러분이 활성화\n시킨 다른 것들을 조심하도록 유의하라. \n\n## 템플릿에서 스트리밍¶\n\n진자2 템플릿 엔진은 또한 부분 단위로 템플릿을 뿌려주는 것을 지원한다. 이 기능은 꽤 일반적이지 않기 때문에 플라스크에 직접적으로 노출되지 않지만, 아래처럼 여러분이 쉽게 직접 구현할 수 있다:\n\ndef stream_template(template_name, **context):\n app.update_template_context(context)\n t = app.jinja_env.get_template(template_name)\n rv = t.stream(context)\n rv.enable_buffering(5)\n return rv\n\n@app.route('/my-large-page.html')\ndef render_large_template():\n rows = iter_all_rows()\n return Response(stream_template('the_template.html', rows=rows))\n```\n\n 여기서의 기법은 어플리케이션의 진자2 환경에서 템플릿 객체를 얻고 문자열 대신 스트림 객체를 반환하는 `render()` 대신 `stream()` 를 호출하는 것이다. 플라스크 템플릿 렌더 함수를\n지나치고 템플릿 객체 자체를 사용하고 있기 때문에 \n\n```\nupdate_template_context()\n```\n\n를 호출하여 렌더 컨텍스트를 갱신하도록\n보장해야한다. 그리고 나서 스트림을 순환하면서 템플릿을 해석한다. 매번 여러분은 yield\n를 호출하기 때문에 서버는 클라이언트로 내용을 밀어내어 보낼 것이고 여러분은 \n\n```\nrv.enable_buffering(size)\n```\n\n을 가지고 템플릿 안에 몇가지 항목을 버퍼링하기를 원할지도\n모른다. `5` 가 보통 기본값이다. \n\n## 컨텍스트를 가진 스트리밍¶\n\n버전 0.9에 추가.\n\n여러분이 데이터를 스트리밍할 때, 요청 컨텍스트는 그 함수가 호출된 순간 이미 종료된다는 것에 주목해라. 플라스크 0.9 는 생성기가 수행하는 동안 요청 컨텍스트를 유지하게 해주는 조력자를 제공한다:\n\n```\nfrom flask import stream_with_context, request, Response\n\n@app.route('/stream')\ndef streamed_response():\n def generate():\n yield 'Hello '\n yield request.args['name']\n yield '!'\n return Response(stream_with_context(generate()))\n```\n\n```\nstream_with_context()\n```\n\n함수 없다면 여러분은 그 시점에 `RuntimeError` 를 얻을 것이다.\n\n플라스크의 설계 원칙 중 한가지는 응답 객체가 생성되고 그 객체를 수정하거나 대체할 수 있는 잠재적인 콜백의 호출 사슬로 그 객체를 전달하는 것이다. 요청 처리가 시작될 때, 아직은 응답 객체를 존재하지 않는다. 뷰 함수나 시스템에 있는 어떤 다른 컴포넌트에 의해 필요할 때 생성된다.\n\n하지만 응답이 아직 존재하지 않는 시점에서 응답을 수정하려 하면 어떻게 되는가? 그에 대한 일반적인 예제는 응답 객체에 있는 쿠키를 설정하기를 원하는 before-request 함수에서 발생할 것이다.\n\n한가지 방법은 그런 상황을 피하는 것이다. 꽤 자주 이렇게 하는게 가능하다. 예를 들면 여러분은 그런 로직을 after-request 콜백으로 대신 옮기도록 시도할 수 있다. 하지만 때때로 거기에 있는 코드를 옮기는 것은 유쾌한 일이 아니거나 코드가 대단히 부자연스럽게 보이게 된다.\n 다른 가능성으로서 여러분은 `g` 객체에 여러 콜백 함수를\n추가하고 요청의 끝부분에서 그것을 호출할 수 있다. 이 방식으로 여러분은\n어플리케이션의 어느 위치로도 코드 실행을 지연시킬 수 있다. \n\n## 데코레이터¶\n\n 다음에 나올 데코레이터가 핵심이다. 그 데코레이터는 `g` 객체에\n있는 리스트에 함수를 등록한다: \n\ndef after_this_request(f):\n if not hasattr(g, 'after_request_callbacks'):\n g.after_request_callbacks = []\n g.after_request_callbacks.append(f)\n return f\n```\n\n## 지연된 함수의 호출¶\n\n 이제 여러분은 요청의 마지막에서 호출될 함수를 표시하기 위해 after_this_request 데코레이터를 사용할 수 있다. 하지만 우리는 여전히 그 함수를 호출할 수도 있다. 이렇게 하기 위해 다음 함수가 `after_request()` 콜백으로 등록될\n필요가 있다: \n\n```\n@app.after_request\ndef call_after_request_callbacks(response):\n for callback in getattr(g, 'after_request_callbacks', ()):\n response = callback(response)\n return response\n```\n\n## 실제적인 예제¶\n\n이제 우리는 이 특정 요청의 마지막에서 호출될 함수를 적절한 어느 시점에라도 쉽게 등록할 수 있다. 예를 들면 여러분은 before-request 함수에서 쿠키에 사용자의 현재 언어를 기억하게 할 수 있다:\n\n@app.before_request\ndef detect_user_language():\n language = request.cookies.get('user_lang')\n if language is None:\n language = guess_language_from_request()\n @after_this_request\n def remember_language(response):\n response.set_cookie('user_lang', language)\n g.language = language\n```\n\n어떤 HTTP 프록시는 임시적인 HTTP 메소드나 새로운 HTTP 메소드 (PATCH 같은) 를 지원하지 않는다. 그런 경우에 프로토콜 전체를 위반하는 방식으로 HTTP 메소드를 다른 HTTP 메소드로 “프록시” 하는 것이 가능하다.\n 이렇게 동작하는 방식은 클라이언트가 HTTP POST로 요청하고\n\n```\nX-HTTP-Method-Override\n```\n\n헤더를 설정하고 그 값으로 의도하는 HTTP 메소드\n( `PATCH` 와 같은)를 설정하면 된다. \n이것은 HTTP 미들웨어로 쉽게 수행할 수 있다:\n\n```\nclass HTTPMethodOverrideMiddleware(object):\n allowed_methods = frozenset([\n 'GET',\n 'HEAD',\n 'POST',\n 'DELETE',\n 'PUT',\n 'PATCH',\n 'OPTIONS'\n ])\n bodyless_methods = frozenset(['GET', 'HEAD', 'OPTIONS', 'DELETE'])\n\n def __call__(self, environ, start_response):\n method = environ.get('HTTP_X_HTTP_METHOD_OVERRIDE', '').upper()\n if method in self.allowed_methods:\n method = method.encode('ascii', 'replace')\n environ['REQUEST_METHOD'] = method\n if method in self.bodyless_methods:\n environ['CONTENT_LENGTH'] = '0'\n return self.app(environ, start_response)\n```\n\n플라스크로 이것을 하려면 아래와 같이 하면 된다:\n\n```\nfrom flask import Flask\n\napp = Flask(__name__)\napp.wsgi_app = HTTPMethodOverrideMiddleware(app.wsgi_app)\n```\n\n코드의 여러 부분이 요청 데이터로 Various pieces of code can consume the request data and preprocess it. For instance JSON data ends up on the request object already read and processed, form data ends up there as well but goes through a different code path. This seems inconvenient when you want to calculate the checksum of the incoming request data. This is necessary sometimes for some APIs.\n\nFortunately this is however very simple to change by wrapping the input stream.\n\nThe following example calculates the SHA1 checksum of the incoming data as it gets read and stores it in the WSGI environment:\n\n```\nimport hashlib\n\nclass ChecksumCalcStream(object):\n\n def __init__(self, stream):\n self._stream = stream\n self._hash = hashlib.sha1()\n\n def read(self, bytes):\n rv = self._stream.read(bytes)\n self._hash.update(rv)\n return rv\n\n def readline(self, size_hint):\n rv = self._stream.readline(size_hint)\n self._hash.update(rv)\n return rv\n\ndef generate_checksum(request):\n env = request.environ\n stream = ChecksumCalcStream(env['wsgi.input'])\n env['wsgi.input'] = stream\n return stream._hash\n```\n\n To use this, all you need to do is to hook the calculating stream in before the request starts consuming data. (Eg: be careful accessing `request.form` or anything of that nature. \n\n```\nbefore_request_handlers\n```\n\nfor instance should be careful not to access it). \nExample usage:\n\n```\n@app.route('/special-api', methods=['POST'])\ndef special_api():\n hash = generate_checksum(request)\n # Accessing this parses the input stream\n files = request.files\n # At this point the hash is fully constructed.\n checksum = hash.hexdigest()\n return 'Hash was: %s' % checksum\n```\n\n여러분이 어떤 것을 이용할 수 있는지에 따라서, 플라스크 어플리케이션을 실행할 수 있는 다양한 방법이 있다. 개발 중에는 내장 서버를 사용할 수 있지만, 운영울 위해서는 모든 배포 옵션을 사용해야 한다. (운영 중에는 내장 개발 서버를 사용하지 마라.) 사용할 수 있는 몇가지 옵션이 있다.\n 만약 다른 WSGI 서버가 있다면, WSGI 어플리케이션을 사용하는 방법에 대한 서버 문서를 참조하라. 단지 `Flask` 어플리케이션 객체가 실제 WSGI 어플리케이션임을 기억하라. \n빠르게 배포하고 실행하기 위한 옵션을 보려면, 퀵스타트의 웹서버에 배포하기 를 참조하라.\n\n만약 Apache 웹서브를 사용하고 있다면, mod_wsgi 를 사용할 것을 고려하라.\n\n## mod_wsgi 설치하기¶\n\nmod_wsgi 가 아직 설치되지 않았다면, 패키지 관리자를 사용하여 설치하거나, 직접 컴파일해야 한다. mod_wsgi installation instructions 가 유닉스 시스템에서의 소스 설치를 다룬다.\n\n만약 우분투/데비안을 사용하고 있다면 apt-get를 사용하여 다음과 같이 활성화할 수 있다:\n\n```\n# apt-get install libapache2-mod-wsgi\n```\n\nFreeBSD에에서는 www/mod_wsgi 포트를 컴파일하거나 pkg-add를 사용하여 mod_wsgi 를 설치하라:\n\n```\n# pkg_add -r mod_wsgi\n```\n\n만약 pkgsrc를 사용하고 있다면, www/ap2-wsgi 패키지를 컴파일하여 mod_wsgi 를 설치할 수 있다.\n\n만약 처음 아파치를 리로드한 후에 세그멘테이션폴트가 나는 자식 프로세스를 발견한다면, 그것들을 무시할 수 있다. 단지 서버를 재시작하면 된다.\n\n## .wsgi 파일 생성하기¶\n\n어플리케이션을 실행하기 위해서는 yourapplication.wsgi 파일이 필요하다. 이 파일은 application 객체를 얻기 위해 시작 시점에 mod_wsgi`가 실행할 코드를 포함하고 있다. 그 객체는 그 파일에서 `application 라고 불리며, application으로 사용된다.\n\n대부분의 어플리케이션에서 다음과 같은 파일이면 충분하다:\n\n```\nfrom yourapplication import app as application\n```\n\n만약 여러분이 application 객체 생성하는 팩토리 함수가 없다면, ‘application’으로 하나를 직접 임포트하여 싱글톤 객체로 사용할 수 있다.\n\n그 파일을 나중에 다시 찾을 수 있는 어딘가에 저장하고(예:/var/www/yourapplication) `yourapplication`과 사용되는 모든 라이브러리가 파이썬 로그 경로에 있는지 확인하라. 만약 파이썬을 설치하는 것을 원하지 않으면, 시스템 전반적으로 `virtual python`_ 객체를 사용하는 것을 고려하라. 실제로 어플리케이션을 virtualenv에 설치해야 하는 것을 기억하라. 다른 방법으로는 임포트 전에 .wsgi 파일 내 경로를 패치하기 위한 옵션이 있다:\n\n```\nimport sys\nsys.path.insert(0, '/path/to/the/application')\n```\n\n여러분이 해야할 마지막 일은 어플리케이션을 위한 아파치 설정 파일을 생성하는 것이다. 이 예제에서 보안적인 이유로 다른 사용자 하에서 어플리케이션을 실행하라고 ‘mod_wsgi’에게 말할 것이다:\n\n```\n\n ServerName example.com\n\n WSGIDaemonProcess yourapplication user=user1 group=group1 threads=5\n WSGIScriptAlias / /var/www/yourapplication/yourapplication.wsgi\n\n \n WSGIProcessGroup yourapplication\n WSGIApplicationGroup %{GLOBAL}\n Order deny,allow\n Allow from all\n \n\n```\n\nNote: WSGIDaemonProcess는 윈도우에서 구현되어 있지 않으며 아파치는 위와 같은 설정을 허용하지 않을 것이다. 윈도우 시스템에서는 해당 라인들을 제거하라:\n\n```\n\n ServerName example.com\n WSGIScriptAlias / C:\\yourdir\\yourapp.wsgi\n \n Order deny,allow\n Allow from all\n \n\n```\n\n더 많은 정보를 위해 `mod_wsgi wiki`_를 참조하라.\n\n## 문제해결¶\n\n만약 어플리케이션이 실행되지 않는다면 아래 문제해결 가이드를 참조하라:\n\n* 문제: 어플리케이션이 실행되지 않으며, 에러로그는 SystemExit ignored를 보여준다\n * \n\n조건에 의해 보호되지 않는 어플리케이션 파일에서 `app.run()` 를 호출한다. 파일에서 `run()` 호출을 제거하고 run.py 파일로 옮기거나, if 블럭 안에 넣어라. * 문제: 어플리케이션이 퍼미션 에러를 준다.\n * 아마 잘못된 사용자에 의해 어플리케이션이 실행되었을 것이다. 어플리케이션이 접근이 필요한 폴더가 적절한 권한이 설정되어 있는지 어플리케이션이 올바른 사용자로 실행되는지 확인하라 (WSGIDaemonProcess 지시어에\n `user` 와 `group` 파라미터) * 문제: 어플리케이션이 에러를 출력하며 죽는다\n * \n\nmod_wsgi는\n `sys.stdout` 와 \n\n```\nsys.stderr`로 어떤 것을 하는 것을 허용하지 않는다는 것을 기억라라. `WSGIRestrictStdout\n```\n\n를 ``off``로 설정하여 이 보호를 해지할 수 있다: > WSGIRestrictStdout Off\n\n다른 대안으로 .wsgi 파일에서의 표준 출력을 다른 스트림으로 변환할 수 있다:\n > import sys sys.stdout = sys.stderr\n * 문제: IO 에러가 나는 자원에 접근하기\n * \n\n아마 어플리케이션이 site-packages 폴더안에 심볼링링크되어 있는 싱글 .py file일 것이다. 이것은 작동하지 않으며 대신 파일이 저장되어 있는 폴더를 pythonpath를 넣거나 어플리케이션을 패키지로 바꿔라.\n\n이러한 이유는 패키지로 설치되지 않은 경우, 모듈 파일명이 자원을 찾기 위해 사용되어 지며, 심볼링링크를 위해 잘못된 파일명이 선택되어 지기 때문이다.\n\n## 자동 리로딩 지원¶\n\n배포 도구를 돕기 위애 여러분은 자동 리로딩 지원을 활성화할 수 있다. .wsgi 파일이 변경되면, `mod_wsgi`는 모든 데몬 프로세스를 리로드할 것이다.\n\n이를 위해 Directory 섹션에 다음과 같은 지시어만 추가하면 된다:\n\n```\nWSGIScriptReloading On\n```\n\n## 가상 환경에서 실행하기¶\n\n가상환경은 필요한 의존적인 것들을 시스템 전반적으로 설치할 필요가 없어서 사용되는 것을 더 잘 컨트롤 할 수 있는 잇점이 있다. 만약 가상환경에서 mod_wsgi를 사용하고 싶다면, .wsgi 파일을 조금 변경할 필요가 있다.\n\n.wsgi 파일 위에 아래와 같은 내용을 추가하라:\n\n```\nactivate_this = '/path/to/env/bin/activate_this.py'\nexecfile(activate_this, dict(__file__=activate_this))\n```\n\n이것은 가상환경 설정에 따라서 로드할 경로를 설정한다. 경로가 절대 경로임을 명심하라.\n\nWSGI 어플리케이션을 포함하고 HTTP를 서비스하는 파이썬으로 개발된 인기있는 서버가 있다. 이 서버들은 실행될 때 독립적이다; 여러분은 그것들을 웹서버를 통해 프록시할 수 있다. 이와 같이 하려면 프록시 설정 단락을 참조하라.\n\n## Gunicorn¶\n\nGunicorn ‘Green Unicorn’은 UNIX를 위한 WSGI HTTP 서버이다. 이것은 루비의 Unicorn 프로젝트에서 포팅된 pre-fork worker 모델이다. 이것은 `eventlet`_와 `greenlet`_을 모두 지원한다. 이 서버에서 플라스크 어플리케이션을 실행하는 것은 매우 간단하다:\n\n```\ngunicorn myproject:app\n```\n\n Gunicorn`_은 많은 커맨드라인 옵션을 제공한다 – ``gunicorn -h` 를 참조하라. 예를 들면 로컬호스트 4000포트로 바인딩하면서( `-b 127.0.0.1:4000` )\n4개의 worker 프로세스로 실행하기 위해서는 ( `-w 4` ) 아래와 같이 하면 된다: \n\n```\ngunicorn -w 4 -b 127.0.0.1:4000 myproject:app\n```\n\n## Tornado¶\n\n`Tornado`_는 `FriendFeed`_를 강화한 확장성있는 넌블러킹 웹서버의 오픈소스 버전이다. 넌블러킹이며 epoll을 사용하기 때문에 수천개의 동시 연결을 처리할 수 있다. 이것은 이상적인 실시간 웹서비스를 의미한다. 플라스크를 이 서비스로 통합하는 것은 복잡하지 않다:\n\n```\nfrom tornado.wsgi import WSGIContainer\nfrom tornado.httpserver import HTTPServer\nfrom tornado.ioloop import IOLoop\nfrom yourapplication import app\n\nhttp_server = HTTPServer(WSGIContainer(app))\nhttp_server.listen(5000)\nIOLoop.instance().start()\n```\n\n## Gevent¶\n\n`Gevent`_는 `libevent`_ 이벤트 루프 위에서 고수준 동기화 API를 제공하기 위해 `greenlet`_를 사용하는 coroutine기반 파이썬 네트워킹 라이브러리이다:\n\n```\nfrom gevent.wsgi import WSGIServer\nfrom yourapplication import app\n\nhttp_server = WSGIServer(('', 5000), app)\nhttp_server.serve_forever()\n```\n\n## Twisted Web¶\n\nTwisted Web`_은 `Twisted`_에 포함되어 있는 웹서버이며, 성숙된 넌블러킹 이벤트 드리븐 네트워킹 라이브러리이다. Twisted Web은 ``twistd` 유틸리티를 사용하여 커맨드라인을 통해 컨트롤할 수 있는 표준 WSGI 컨테이너이다:\n\n```\ntwistd web --wsgi myproject.app\n```\n\n 이 예제는 `myproject` 모듈로부터 ``app``을 호출하는 Flask application를 실행할 것이다. Twisted Web은 많은 플래그와 옵션을 지원하며, `twistd` 유틸리티 또한 많은 것을 제공한다;\n더 많은 정보를 위해 `twistd -h` 와 \n\n```\ntwistd web -h``를 참조하라.\n예를 들어, ``myproject\n```\n\n어플리케이션을 8080포트로 Twisted 웹서버로 서비스하려면 아래와 같이 하면 된다: \n\n```\ntwistd -n web --port 8080 --wsgi myproject.app\n```\n\n## 프록시 설정¶\n\n어플리케이션을 HTTP 프록시 뒤에 있는 서버들 중 하나를 사용하여 배포한다면 어플리케이션을 실행하기 위해 몇가지 헤더를 rewrite할 필요가 있을 것이다. WSGI 환경에서 문제가 있는 두가지 값을 REMOTE_ADDR 와 HTTP_HOST 이다. 이 헤더들은 httpd에 전달하기 위한 설정을 할 수 있다. 또는 미들웨어에서 그 헤더들을 수정할 수 있다. Werkzeug는 몇가지 설정으로 해결할 수 있는 픽서(fixer)을 포함한다. 그러나 특정한 설정을 위해 WSGI 미들웨어를 사용하기를 원할지도 모른다.\n\n아래는 적절한 헤더를 설정하여 로컬포스트 8000번 포트로 서비스되는 어플리케이션으로 프록시하는 간단한 nginx 설정이다:\n\n```\n.. sourcecode:: nginx\n```\n\n * server {\n\nlisten 80;\n\nserver_name _;\n\naccess_log /var/log/nginx/access.log; error_log /var/log/nginx/error.log;\n * location / {\n\nproxy_pass http://127.0.0.1:8000/; proxy_redirect off;\n\nproxy_set_header Host $host; proxy_set_header X-Real-IP $remote_addr; proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;\n\n만약 httpd가 이 헤더들을 제공하지 않는다면 가장 일반적인 설정은 `X-Forwarded-Host`로부터 호스트를 `X-Forwarded-For`로부터 원격 어드레스를 가져오는 것이다:\n\n```\nfrom werkzeug.contrib.fixers import ProxyFix\napp.wsgi_app = ProxyFix(app.wsgi_app)\n```\n\n헤더 신뢰\n\n악의저인 클라이언트에 의해 위조될 수 있는 헤더를 무조건 신뢰할 것이기 때문에 non-proxy 설정에서 이런 미들웨어를 사용하는 것은 보안적인 문제가 있다는 것을 기억하라.\n\n만약 다른 헤더로부터 헤더들은 rewrite하려면, 아래와 같이 픽서를 사용할 수 있다:\n\n```\nclass CustomProxyFix(object):\n\n def __call__(self, environ, start_response):\n host = environ.get('HTTP_X_FHOST', '')\n if host:\n environ['HTTP_HOST'] = host\n return self.app(environ, start_response)\n\napp.wsgi_app = CustomProxyFix(app.wsgi_app)\n```\n\nuWSGI는 nginx, lighttpd, cherokee 와 같은 서버에서 사용할 수 있는 배포 옵션이다; 다른 옵션을 보려면 FastCGI 와 독립적인 WSGI 컨테이너 를 확인하라. uWSGI 프로토토콜로 WSGI 어플리케이션을 사용하기 위해서는 먼저 uWSGI 서버가 필요하다. uWSGI는프로토콜이면서 어플리케이션 서버이다; 어플리케이션서버는 uWSGI, FastCGI, HTTP 프로토콜을 서비스할 수 있다.\n\n가장 인기있는 uWSGI 서버는 `uwsgi`_이며, 설명을 위해 사용할 것이다. 아래와 같이 설치되어 있는지 확인하라.\n\n## uwsgi로 app 시작하기¶\n\n`uwsgi`는 파이썬 모듈에 있는 WSGI callables에서 운영하기 위해 설계되어 있다.\n\nmyapp.py에 flask application이 있다면 아래와 같이 사용하라:\n\n```\n$ uwsgi -s /tmp/uwsgi.sock --module myapp --callable app\n```\n\n또는 아래와 같은 방법도 사용할 수 있다:\n\n```\n$ uwsgi -s /tmp/uwsgi.sock -w myapp:app\n```\n\nnginx를 위한 기본적인 flask uWSGI 설정은 아래와 같다:\n\n```\nlocation = /yourapplication { rewrite ^ /yourapplication/; }\nlocation /yourapplication { try_files $uri @yourapplication; }\nlocation @yourapplication {\n include uwsgi_params;\n uwsgi_param SCRIPT_NAME /yourapplication;\n uwsgi_modifier1 30;\n uwsgi_pass unix:/tmp/uwsgi.sock;\n}\n```\n\n이 설정은 어플리케이션을 `/yourapplication`에 바인드한다. URL 루트에 어플리케이션을 두길 원한다면, WSGI `SCRIPT_NAME`를 알려줄 필요가 없거나 uwsgi modifier를 설정할 필요가 없기 때문에 조금 더 간단하다:\n\nThis configuration binds the application to /yourapplication. If you want to have it in the URL root it’s a bit simpler because you don’t have to tell it the WSGI SCRIPT_NAME or set the uwsgi modifier to make use of it:\n\n```\nlocation / { try_files $uri @yourapplication; }\nlocation @yourapplication {\n include uwsgi_params;\n uwsgi_pass unix:/tmp/uwsgi.sock;\n}\n```\n\nFastCGI는 nginx, lighttpd, cherokee 와 같은 서버에서 사용할 수 있는 배포 옵션이다; 다른 옵션을 보려면 uWSGI 와 독립적인 WSGI 컨테이너 를 확인하라. WSGI 어플리케이션을 사용하기 위해서는 먼저 FastCGI 서버가 필요하다. 가장 인기있는 것은 `flup`_이며, 설명을 위해 사용할 것이다. 아래와 같이 설치되어 있는지 확인하라.\n\n## .fcgi 파일 생성하기¶\n\n먼저 FastCGI 서버 파일을 만드는 것이 필요하다. 그것을 yourapplication.fcgi:: 이라고 부르자.\n #!/usr/bin/python from flup.server.fcgi import WSGIServer from yourapplication import app\n * if __name__ == ‘__main__’:\n * WSGIServer(app).run()\n 이것은 아파치에 적용하기에는 충분하지만 nginx나 lighttpd의 오래된 버전의 경우 FastCGI 서버와ㅏ 통신하기 위해 명시적으로 소켓을 전달해야 한다. `WSGIServer` ::에 소켓 경로를 전달해야 한다. > WSGIServer(application, bindAddress=’/path/to/fcgi.sock’).run()\n\n그 경로는 서버 설정에 정의된 것과 정확히 같아야 한다.\n\nyourapplication.fcgi`을 다시 찾을 수 있는 어딘가에 저장하라. `/var/www/yourapplication 나 비슷한 곳에 두면 된다.\n\n서버가 이 파일을 실행할 수 있도록 실행 가능 비트를 설정하라:\n\n```\n# chmod +x /var/www/yourapplication/yourapplication.fcgi\n```\n\n기본적인 아파치 배포는 위 예제로 충분하나, .fcgi 파일이 example.com/yourapplication.fcgi/news/ 와 같이 어플리케이션 URL에 보일 것이다. URL에 yourapplication.fcgi을 보이지 않게 하는 며가지 설정 방법이 있다. 가장 많이 사용되는 방법은 ScriptAlias 설정 지시어를 사용하는 것이다:\n * \n * ServerName example.com ScriptAlias / /path/to/yourapplication.fcgi/\n\n\nSetHandler fcgid-script Options +FollowSymLinks +ExecCGI \n\n * \n * Options +FollowSymlinks RewriteEngine On RewriteBase / RewriteCond %{REQUEST_FILENAME} !-f RewriteRule ^(.*)$ yourapplication.fcgi/$1 [QSA,L]\n\n/lib/python2.6/site-packages’)\n\nfrom flup.server.fcgi import WSGIServer from yourapplication import app\n * class ScriptNameStripper(object):\n * def __init__(self, app):\n * self.app = app\n * def __call__(self, environ, start_response):\n * environ[‘SCRIPT_NAME’] = ‘’ return self.app(environ, start_response)\n\napp = ScriptNameStripper(app)\n * if __name__ == ‘__main__’:\n * WSGIServer(app).run()\n\n## lighttpd 설정하기¶\n\nlighttpd를 위한 기본적인 FastCGI 설정은 아래와 같다:\n\n```\nfastcgi.server = (\"/yourapplication.fcgi\" =>\n ((\n \"socket\" => \"/tmp/yourapplication-fcgi.sock\",\n \"bin-path\" => \"/var/www/yourapplication/yourapplication.fcgi\",\n \"check-local\" => \"disable\",\n \"max-procs\" => 1\n ))\n)\n\nalias.url = (\n \"/static/\" => \"/path/to/your/static\"\n)\n\nurl.rewrite-once = (\n \"^(/static($|/.*))$\" => \"$1\",\n \"^(/.*)$\" => \"/yourapplication.fcgi$1\"\n```\n\nFastCGI, alias, rewrite modules 모듈을 활성화하는 것을 기억하라. 이 설정은 어플리케이션을 /yourapplication`에 바인드한다. 만약 어플리케이션을 URL 루트에서 실행하기를 원한다면, :class:`~werkzeug.contrib.fixers.LighttpdCGIRootFix 미들웨어 관련 lighttpd 버그의 회피 수단을 적용해야 한다.\n\n어플리케이션을 URL 루트에 마운트하는 경우에만 이것을 적용해야 한다. 또한 FastCGI and Python 에 대한 더 많은 정보를 위해 lighttpd 문서를 확인하라.(명시적으로 소켓을 전달하는 것이 더 이상 필요하지 않음을 주목하라.)\n\nnginx에 FastCGI 어플리케이션 설치는 기본적으로 어떠한 FastCGI 파라미터가 전달되지 않기 때문에 조금 다르다.\n\nnginx를 위한 기본적인 플라스크 FastCGI 설정은 아래와 같다:\n\n```\nlocation = /yourapplication { rewrite ^ /yourapplication/ last; }\nlocation /yourapplication { try_files $uri @yourapplication; }\nlocation @yourapplication {\n include fastcgi_params;\n fastcgi_split_path_info ^(/yourapplication)(.*)$;\n fastcgi_param PATH_INFO $fastcgi_path_info;\n fastcgi_param SCRIPT_NAME $fastcgi_script_name;\n fastcgi_pass unix:/tmp/yourapplication-fcgi.sock;\n}\n```\n\n이 설정은 어플리케이션을 /yourapplication`에 바인드한다. URL 루트에 어플리케이션을 두길 원한다면, `PATH_INFO 와 `SCRIPT_NAME`를 계산하는 방법을 이해할 필요가 없기 때문에 조금 더 간단하다:\n\n```\nlocation / { try_files $uri @yourapplication; }\nlocation @yourapplication {\n include fastcgi_params;\n fastcgi_param PATH_INFO $fastcgi_script_name;\n fastcgi_param SCRIPT_NAME \"\";\n fastcgi_pass unix:/tmp/yourapplication-fcgi.sock;\n}\n```\n\n## FastCGI 프로세스 실행하기¶\n\n Nginx와 다른 서버들은 FastCGI 어플리케이션을 로드하지 않기 때문에, 스스로 로드를 해야 한다. 관리자가 FastCGI 프로세스를 관리할 수 있다. 다른 FastCGI 프로세스 관리자를 찾아 보거나, 부팅할 때 .fcgi 파일을 실행하도록 스크립트를 작성할 수 있다. (예. SysV `init.d` 스크립트를 사용)\n임시 방편으로는 GNU 화면 안에서 항상 `.fcgi` 스크립트를 실행할 수 있다.\n상세를 위해 `man screen` 를 확인하고, 이것이 시스템을 재시작하면 지속되지 않는 수동적인 해결책임을 주의하라: \n\n```\n$ screen\n$ /var/www/yourapplication/yourapplication.fcgi\n```\n\n## 디버깅¶\n\nFastCGI 배포는 대부분 웹서버에서 디버깅하기 어려운 경향이 있다. 아주 종종 서버 로그가 알려는 유일한 정보는 “premature end of headers” 라인과 함께 나타난다. 어플리케이션을 디버깅하기 위해서, 여러분에게 정말로 줄 수 있는 유일한 방법은 정확한 사용자로 바꾸고 손으로 직접 어플리케이션을 실행하는 것이다.\n\n아래 예제는 어플리케이션을 application.fcgi 라고 부르고, 웹서버가 사용자가 `www-data`라고 가정한다:\n\n```\n$ su www-data\n$ cd /var/www/yourapplication\n$ python application.fcgi\nTraceback (most recent call last):\n File \"yourapplication.fcgi\", line 4, in \nImportError: No module named yourapplication\n```\n\n이 경우 에러는 “yourapplication”가 파이썬 경로에 없는 것 같이 보인다. 일반적인 문제는 아래와 같다:\n\n```\n- 상대 경로가 사용된 경우. 현재 작업 디렉토리에 의존하지 마라.\n- 웹서버에 의해 설정되지 않는 환경 변수에 의존적인 코드.\n- 다른 파이썬 해석기가 사용된 경우.\n```\n\n다른 모든 배포 방법을 적용할 수 없다면, CGI는 확실히 가능할 것이다. CGI는 거의 모든 주요 서버에 의해 제공되지만 보통 차선의 성능을 가진다.\n\n또한 CGI와 같은 환경에서 실행할 수 있는 구글의 App Engine 에서 플라스크 어플리케이션을 사용할 수 있는 방법이기도 하다.\n\n## .cgi 파일 만들기¶\n\n먼저 CGI 어플리케이션 파일을 만드는 것이 필요하다. 그것을 yourapplication.cgi:: 이라고 부르자.\n #!/usr/bin/python from wsgiref.handlers import CGIHandler from yourapplication import app\n\nCGIHandler().run(app)\n\n## 서버 설정¶\n\n보통 서버를 설정하는 두가지 방법이 있다. .cgi 를 cgi-bin 으로 복사하거나(mod_rewrite 나 URL을 재작성하는 유사한 것을 사용) 서버 지점을 파일에 직접 작성한다.\n\n아파치의 경우, 아래와 같이 설정 파일 안에 입력할 수 있다:\n\n```\nScriptAlias /app /path/to/the/application.cgi\n```\n\n그러나 공유된 웹호스팅에서는 아파치 설정에 접근할 수 없을 지도 모른다. 이런 경우, 여러분의 어플리케이션이 실행되기를 원하는 공개 디렉토리에 있는 .htaccess 파일에 입력할 수 있으나, 이 경우엔는 ScriptAlias 지시어는 적용되지 않을 것이다:\n\n```\nRewriteEngine On\nRewriteCond %{REQUEST_FILENAME} !-f # Don't interfere with static files\nRewriteRule ^(.*)$ /path/to/the/application.cgi/$1 [L]\n```\n\n더 많은 정보를 위해 사용하는 웹서버의 문서를 참조하라.\n\n코드가 증가하고 어플리케이션의 규모가 커질 때 여러분에게 몇가지 선택사항이 있다.\n\n## 소스코드를 읽어라.¶\n\n플라스크는 이미 존재하고 잘 사용되는 도구인 벡자이크(WSGI)와 진자(템플릿 엔진) 의 상위에서 여러분 자신의 프레임워크를 만들기 위한 방법을 설명하기 위해 부분적으로 시작되었고, 개발되어오면서, 넓은 사용자층에게 유용해졌다. 여러분의 코드가 점점 커진다면, 플라스크를 사용하지 않아야 한다 – 그 이유를 이해해라. 소스를 읽어라. 플라스크의 소스는 읽혀지기 위해 쓰여진것이다; 그것은 공표된 문서이기에 여러분은 그것의 내부 API를 사용할 수 있다. 플라스크는 여러분의 프로젝트에 필요한 훅 포인트 (hook point)를 찾을 수 있도록 상위 라이브러리에 있는 문서화된 API를 고수하고 그것의 내부 유틸리티를 문서화한다.\n\n## 훅(Hook)하고. 확장하라.¶\n\nAPI 문서는 사용가능한 오버라이드, 훅 포인트, 시그널(Signals) 로 가득 차 있다. 여러분은 요청과 응답 객체와 같은 것에 대한 커스텀 클래스를 제공할 수 있다. 여러분이 사용하는 API에 대해 좀 더 깊게 파보고, 플라스크 릴리즈에 특별하게 사용가능한 커스터마이징된 것을 찾아봐라. 여러분의 프로젝트가 여러 유틸리티나 플라스크 확장으로 리팩토링될 수 있는 방식을 찾아봐라. 커뮤니티에 있는 여러 확장 을 살펴보고 , 여러분이 필요한 도구를 찾을 수 없다면 여러분 자신의 확장을 만들기 위한 패턴을 찾아봐라.\n\n## 서브클래싱하라.¶\n\n `Flask` 클래스는 서브클래싱에 대해 설계된 여러 메소드가 있다.\n여러분은 `Flask` (연결된 메소드 문서를 살펴봐라) 를\n서브클래싱하고 어플리케이션 클래스를 인스턴스화한 곳 어디서나 그 서브클래스를\n사용하여 동작을 빠르게 추가하거나 커스터마이징할 수 있다. 이것은\n어플리케이션 팩토리 과 함께 잘 동작한다. \n\n## 미들웨어로 감싸라.¶\n\n어플리케이션 디스패칭 장에서 미들웨어를 적용하는 방법에 대해 자세히 보여줬다. 여러분의 플라스크 인스턴스를 감싸기 위해 WSGI 미들웨어와 여러분의 어플리케이션 과 HTTP 서버 사이에 있는 계층에 수정과 변경을 소개할 수 있다. 벡자이크는 다수의 미들웨어 를 포함하고 있다.\n\n## 분기하라.¶\n\n위에서 언급한 선택사항 중 어느 것에도 해당되지 않는다면, 플라스크를 분기해라. 플라스크 코드의 다수는 벡자이크와 진자2 안에 있는 것이다. 이런 라이브러리가 그 동작의 대다수를 수행한다. 플라스크는 단지 그런 라이브러리를 잘 붙여논 풀같은 것이다. 모든 프로젝트에 대해 하부 프레임워크가 방해가 되는 점이 있다 (왜냐하면 원 개발자들의 가진 가정들 때문에). 이것은 자연스러운 것인데 왜냐하면 그런 경우가 아니라면, 그 프레임워크는 시작부터 굉장히 가파른 학습 곡선과 많은 개발자의 좌절감을 유발하는 매우 복잡한 시스템이 될것이기 때문이다.\n\n이것은 플라스크에만 유일하게 적용되지 않는다. 많은 사람들이 그들이 사용하는 프레임워크의 단점에 대응하기 위해 패치되고 변경된 버전의 프레임워크를 사용한다. 이 방식 또한 플라스크의 라이선스에 반영돼있다. 여러분이 플라스크에 변경하기로 결정했더라도 그 어떤 변경에 대해서도 다시 기여하지 않아도 된다.\n\n물론 분기하는 것에 단점은 플라스크 확장에 대부분 깨질것이라는 점인데 왜냐하면 새로운 프레임워크는 다른 임포트 명을 가질 것이기 때문이다. 더 나아가서 변경의 개수의 따라 상위 변경을 통합하는 것이 복합한 과정일 수 있다. 그런것 때문에 분기하는 것은 최후의 보루가 되어야 할 것이다.\n\n## 프로페셔널처럼 규모를 조절하라.¶\n\n많은 웹 어플리케이션에 대한 코드의 복잡도는 예상되는 사용자 수와 데이터 항목에 대한 규모를 조절하는것에 비하면 큰 이슈는 아니다. 플라스크 그 자체로는 여러분의 어플리케이션 코드, 사용하고자 하는 데이타 저장소 그리고 실행되는 파이썬 구현과 웹서버에 의한 규모 조절이라는 측면에서만 제한된다.\n\n예를 자면 규모를 잘 조절한다는 것은 만약 여러분이 서버의 양을 두배로 했을 때 두배의 성능을 얻는다는 것을 의미한다. 규모 조절이 잘못됐다는 것은 여러분이 새로운 서버를 추가했을 때 어플리케이션이 더 좋은 성능을 내지 못하거나 심지어 대체되는 서버로 지원하지도 못하는 것을 의미한다.\n\n문맥 로컬 프록시가 플라스크에서 규모 조절과 관련된 유일한 제한 요소이다. 그것들은 플라스크에 있는 쓰레드, 프로세스 혹은 그린렛으로 정의되는 문맥에 의존한다. 여러분의 서버가 쓰레드나 그린렛에 기반하지 않은 어떤 동시성의 기능을 사용한다면, 플라스크는 이와 같은 전역적인 프록시를 더 이상 제공할 수 없을 것이다. 하지만 서버의 다수는 기반에서 벡자이크 라이브러리에 의해 잘 지원되는 모든 메소드인 동시성을 수행하기 위해 쓰레드, 그린렛 또는 분리된 프로세스를 사용할 것이다.\n\n## 커뮤니티와 논의해라.¶\n\n플라스크 개발자들은 사용자들에게 크고 작은 코드를 가지고 그 프레임워크에 접근할 수 있도록 유지한다. 여러분의 방식에서 플라스크로 인한 어떤 걸림돌을 만난다면 메일링 리스트나 IRC 채널을 통해 플라스크 개발자들에게 연락하는 것에 망설이지 말도록 해라. 더 큰 어플리케이션에 대한 도구를 개선하기 위해 플라스크와 그 확장 을 개발하는 개발자들에게 가장 좋은 방법은 사용자들에게 피드백을 얻는 것이다.\n\nIf you are curious why Flask does certain things the way it does and not differently, this section is for you. This should give you an idea about some of the design decisions that may appear arbitrary and surprising at first, especially in direct comparison with other frameworks.\n\n## The Explicit Application Object¶\n\n A Python web application based on WSGI has to have one central callable object that implements the actual application. In Flask this is an instance of the `Flask` class. Each Flask application has\nto create an instance of this class itself and pass it the name of the\nmodule, but why can’t Flask do that itself? \nWithout such an explicit application object the following code:\n\nWould look like this instead:\n\n```\nfrom hypothetical_flask import route\n\n@route('/')\ndef index():\n return 'Hello World!'\n```\n\nThere are three major reasons for this. The most important one is that implicit application objects require that there may only be one instance at the time. There are ways to fake multiple applications with a single application object, like maintaining a stack of applications, but this causes some problems I won’t outline here in detail. Now the question is: when does a microframework need more than one application at the same time? A good example for this is unittesting. When you want to test something it can be very helpful to create a minimal application to test specific behavior. When the application object is deleted everything it allocated will be freed again.\n Another thing that becomes possible when you have an explicit object lying around in your code is that you can subclass the base class ( `Flask` ) to alter specific behavior. This would not be\npossible without hacks if the object were created ahead of time for you\nbased on a class that is not exposed to you. But there is another very important reason why Flask depends on an explicit instantiation of that class: the package name. Whenever you create a Flask instance you usually pass it __name__ as package name. Flask depends on that information to properly load resources relative to your module. With Python’s outstanding support for reflection it can then access the package to figure out where the templates and static files are stored (see `open_resource()` ). Now obviously there\nare frameworks around that do not need any configuration and will still be\nable to load templates relative to your application module. But they have\nto use the current working directory for that, which is a very unreliable\nway to determine where the application is. The current working directory\nis process-wide and if you are running multiple applications in one\nprocess (which could happen in a webserver without you knowing) the paths\nwill be off. Worse: many webservers do not set the working directory to\nthe directory of your application but to the document root which does not\nhave to be the same folder. The third reason is “explicit is better than implicit”. That object is your WSGI application, you don’t have to remember anything else. If you want to apply a WSGI middleware, just wrap it and you’re done (though there are better ways to do that so that you do not lose the reference to the application object `wsgi_app()` ). \nFurthermore this design makes it possible to use a factory function to create the application which is very helpful for unittesting and similar things (어플리케이션 팩토리).\n\n## The Routing System¶\n\nFlask uses the Werkzeug routing system which has was designed to automatically order routes by complexity. This means that you can declare routes in arbitrary order and they will still work as expected. This is a requirement if you want to properly implement decorator based routing since decorators could be fired in undefined order when the application is split into multiple modules.\n\nAnother design decision with the Werkzeug routing system is that routes in Werkzeug try to ensure that URLs are unique. Werkzeug will go quite far with that in that it will automatically redirect to a canonical URL if a route is ambiguous.\n\n## One Template Engine¶\n\nFlask decides on one template engine: Jinja2. Why doesn’t Flask have a pluggable template engine interface? You can obviously use a different template engine, but Flask will still configure Jinja2 for you. While that limitation that Jinja2 is always configured will probably go away, the decision to bundle one template engine and use that will not.\n\nTemplate engines are like programming languages and each of those engines has a certain understanding about how things work. On the surface they all work the same: you tell the engine to evaluate a template with a set of variables and take the return value as string.\n\nBut that’s about where similarities end. Jinja2 for example has an extensive filter system, a certain way to do template inheritance, support for reusable blocks (macros) that can be used from inside templates and also from Python code, uses Unicode for all operations, supports iterative template rendering, configurable syntax and more. On the other hand an engine like Genshi is based on XML stream evaluation, template inheritance by taking the availability of XPath into account and more. Mako on the other hand treats templates similar to Python modules.\n\nWhen it comes to connecting a template engine with an application or framework there is more than just rendering templates. For instance, Flask uses Jinja2’s extensive autoescaping support. Also it provides ways to access macros from Jinja2 templates.\n\nA template abstraction layer that would not take the unique features of the template engines away is a science on its own and a too large undertaking for a microframework like Flask.\n\nFurthermore extensions can then easily depend on one template language being present. You can easily use your own templating language, but an extension could still depend on Jinja itself.\n\n## Micro with Dependencies¶\n\nWhy does Flask call itself a microframework and yet it depends on two libraries (namely Werkzeug and Jinja2). Why shouldn’t it? If we look over to the Ruby side of web development there we have a protocol very similar to WSGI. Just that it’s called Rack there, but besides that it looks very much like a WSGI rendition for Ruby. But nearly all applications in Ruby land do not work with Rack directly, but on top of a library with the same name. This Rack library has two equivalents in Python: WebOb (formerly Paste) and Werkzeug. Paste is still around but from my understanding it’s sort of deprecated in favour of WebOb. The development of WebOb and Werkzeug started side by side with similar ideas in mind: be a good implementation of WSGI for other applications to take advantage.\n\nFlask is a framework that takes advantage of the work already done by Werkzeug to properly interface WSGI (which can be a complex task at times). Thanks to recent developments in the Python package infrastructure, packages with dependencies are no longer an issue and there are very few reasons against having libraries that depend on others.\n\n## Thread Locals¶\n\n Flask uses thread local objects (context local objects in fact, they support greenlet contexts as well) for request, session and an extra object you can put your own things on ( `g` ). Why is that and\nisn’t that a bad idea? \nYes it is usually not such a bright idea to use thread locals. They cause troubles for servers that are not based on the concept of threads and make large applications harder to maintain. However Flask is just not designed for large applications or asynchronous servers. Flask wants to make it quick and easy to write a traditional web application.\n\nAlso see the 크게 만들기 section of the documentation for some inspiration for larger applications based on Flask.\n\n## What Flask is, What Flask is Not¶\n\nFlask will never have a database layer. It will not have a form library or anything else in that direction. Flask itself just bridges to Werkzeug to implement a proper WSGI application and to Jinja2 to handle templating. It also binds to a few common standard library packages such as logging. Everything else is up for extensions.\n\nWhy is this the case? Because people have different preferences and requirements and Flask could not meet those if it would force any of this into the core. The majority of web applications will need a template engine in some sort. However not every application needs a SQL database.\n\nThe idea of Flask is to build a good foundation for all applications. Everything else is up to you or extensions.\n\nDate: 2004-01-01\nCategories: \nTags: \n\nThe Flask documentation and example applications are using HTML5. You may notice that in many situations, when end tags are optional they are not used, so that the HTML is cleaner and faster to load. Because there is much confusion about HTML and XHTML among developers, this document tries to answer some of the major questions.\n\n## History of XHTML¶\n\nFor a while, it appeared that HTML was about to be replaced by XHTML. However, barely any websites on the Internet are actual XHTML (which is HTML processed using XML rules). There are a couple of major reasons why this is the case. One of them is Internet Explorer’s lack of proper XHTML support. The XHTML spec states that XHTML must be served with the MIME type application/xhtml+xml, but Internet Explorer refuses to read files with that MIME type. While it is relatively easy to configure Web servers to serve XHTML properly, few people do. This is likely because properly using XHTML can be quite painful.\n\nOne of the most important causes of pain is XML’s draconian (strict and ruthless) error handling. When an XML parsing error is encountered, the browser is supposed to show the user an ugly error message, instead of attempting to recover from the error and display what it can. Most of the (X)HTML generation on the web is based on non-XML template engines (such as Jinja, the one used in Flask) which do not protect you from accidentally creating invalid XHTML. There are XML based template engines, such as Kid and the popular Genshi, but they often come with a larger runtime overhead and, are not as straightforward to use because they have to obey XML rules.\n The majority of users, however, assumed they were properly using XHTML. They wrote an XHTML doctype at the top of the document and self-closed all the necessary tags ( `
` becomes `
` or `

` in XHTML).\nHowever, even if the document properly validates as XHTML, what really\ndetermines XHTML/HTML processing in browsers is the MIME type, which as\nsaid before is often not set properly. So the valid XHTML was being treated\nas invalid HTML. \nXHTML also changed the way JavaScript is used. To properly work with XHTML, programmers have to use the namespaced DOM interface with the XHTML namespace to query for HTML elements.\n\n## History of HTML5¶\n\nDevelopment of the HTML5 specification was started in 2004 under the name “Web Applications 1.0” by the Web Hypertext Application Technology Working Group, or WHATWG (which was formed by the major browser vendors Apple, Mozilla, and Opera) with the goal of writing a new and improved HTML specification, based on existing browser behavior instead of unrealistic and backwards-incompatible specifications.\n For example, in HTML4 `` . However, since people were using\nXHTML-like tags along the lines of `` , browser vendors implemented\nthe XHTML syntax over the syntax defined by the specification. \nIn 2007, the specification was adopted as the basis of a new HTML specification under the umbrella of the W3C, known as HTML5. Currently, it appears that XHTML is losing traction, as the XHTML 2 working group has been disbanded and HTML5 is being implemented by all major browser vendors.\n\n## HTML versus XHTML¶\n\nThe following table gives you a quick overview of features available in HTML 4.01, XHTML 1.1 and HTML5. (XHTML 1.0 is not included, as it was superseded by XHTML 1.1 and the barely-used XHTML5.)\n\nHTML4.01 | XHTML1.1 | HTML5 |\n | --- | --- | --- |\n\nHTML4.01\n XHTML1.1\n HTML5\n\n value\n [1]\n\n
supported\n [2]\n\n \n\n\n```\n\n If you know a bit of JavaScript internals you might know that it’s possible to patch constructors and register callbacks for setters. An attacker can use this (like above) to get all the data you exported in your JSON file. The browser will totally ignore the `application/json` mimetype if `text/javascript` is defined as content type in the script\ntag and evaluate that as JavaScript. Because top-level array elements are\nallowed (albeit useless) and we hooked in our own constructor, after that\npage loaded the data from the JSON response is in the captured array. Because it is a syntax error in JavaScript to have an object literal ( `{...}` ) toplevel an attacker could not just do a request to an\nexternal URL with the script tag to load up the data. So what Flask does\nis to only allow objects as toplevel elements when using `jsonify()` . Make sure to do the same when using an ordinary\nJSON generate function.\n\nFlask like Jinja2 and Werkzeug is totally Unicode based when it comes to text. Not only these libraries, also the majority of web related Python libraries that deal with text. If you don’t know Unicode so far, you should probably read The Absolute Minimum Every Software Developer Absolutely, Positively Must Know About Unicode and Character Sets. This part of the documentation just tries to cover the very basics so that you have a pleasant experience with Unicode related things.\n\n## Automatic Conversion¶\n\nFlask has a few assumptions about your application (which you can change of course) that give you basic and painless Unicode support:\n\n* the encoding for text on your website is UTF-8\n * internally you will always use Unicode exclusively for text except for literal strings with only ASCII character points.\n * encoding and decoding happens whenever you are talking over a protocol that requires bytes to be transmitted.\n\nSo what does this mean to you?\n\nHTTP is based on bytes. Not only the protocol, also the system used to address documents on servers (so called URIs or URLs). However HTML which is usually transmitted on top of HTTP supports a large variety of character sets and which ones are used, are transmitted in an HTTP header. To not make this too complex Flask just assumes that if you are sending Unicode out you want it to be UTF-8 encoded. Flask will do the encoding and setting of the appropriate headers for you.\n\nThe same is true if you are talking to databases with the help of SQLAlchemy or a similar ORM system. Some databases have a protocol that already transmits Unicode and if they do not, SQLAlchemy or your other ORM should take care of that.\n\n## The Golden Rule¶\n\nSo the rule of thumb: if you are not dealing with binary data, work with Unicode. What does working with Unicode in Python 2.x mean?\n\n* as long as you are using ASCII charpoints only (basically numbers, some special characters of latin letters without umlauts or anything fancy) you can use regular string literals (\n `'Hello World'` ). * if you need anything else than ASCII in a string you have to mark this string as Unicode string by prefixing it with a lowercase u. (like\n `u'Hänsel und Gretel'` ) * if you are using non-Unicode characters in your Python files you have to tell Python which encoding your file uses. Again, I recommend UTF-8 for this purpose. To tell the interpreter your encoding you can put the\n\n```\n# -*- coding: utf-8 -*-\n```\n\ninto the first or second line of your Python source file. * Jinja is configured to decode the template files from UTF-8. So make sure to tell your editor to save the file as UTF-8 there as well.\n\n## Encoding and Decoding Yourself¶\n\nIf you are talking with a filesystem or something that is not really based on Unicode you will have to ensure that you decode properly when working with Unicode interface. So for example if you want to load a file on the filesystem and embed it into a Jinja2 template you will have to decode it from the encoding of that file. Here the old problem that text files do not specify their encoding comes into play. So do yourself a favour and limit yourself to UTF-8 for text files as well.\n Anyways. To load such a file with Unicode you can use the built-in `str.decode()` method: \n\n```\ndef read_file(filename, charset='utf-8'):\n with open(filename, 'r') as f:\n return f.read().decode(charset)\n```\n\n To go from Unicode into a specific charset such as UTF-8 you can use the `unicode.encode()` method: \n\n```\ndef write_file(filename, contents, charset='utf-8'):\n with open(filename, 'w') as f:\n f.write(contents.encode(charset))\n```\n\n## Configuring Editors¶\n\nMost editors save as UTF-8 by default nowadays but in case your editor is not configured to do this you have to change it. Here some common ways to set your editor to store as UTF-8:\n\nVim: put\n `set enc=utf-8` to your `.vimrc` file. * \n\nEmacs: either use an encoding cookie or put this into your\n `.emacs` file: > (prefer-coding-system 'utf-8) (setq default-buffer-file-coding-system 'utf-8)\n * \n\nNotepad++:\n\n* Go to Settings -> Preferences ...\n * Select the “New Document/Default Directory” tab\n * Select “UTF-8 without BOM” as encoding\n\nIt is also recommended to use the Unix newline format, you can select it in the same panel but this is not a requirement.\n\nFlask, being a microframework, often requires some repetitive steps to get a third party library working. Because very often these steps could be abstracted to support multiple projects the Flask Extension Registry was created.\n\nIf you want to create your own Flask extension for something that does not exist yet, this guide to extension development will help you get your extension running in no time and to feel like users would expect your extension to behave.\n\n## Anatomy of an Extension¶\n\n Extensions are all located in a package called `flask_something` where “something” is the name of the library you want to bridge. So for\nexample if you plan to add support for a library named simplexml to\nFlask, you would name your extension’s package `flask_simplexml` . \nThe name of the actual extension (the human readable name) however would be something like “Flask-SimpleXML”. Make sure to include the name “Flask” somewhere in that name and that you check the capitalization. This is how users can then register dependencies to your extension in their setup.py files.\n Flask sets up a redirect package called `flask.ext` where users\nshould import the extensions from. If you for instance have a package\ncalled `flask_something` users would import it as `flask.ext.something` . This is done to transition from the old\nnamespace packages. See Extension Import Transition for more details. \nBut how do extensions look like themselves? An extension has to ensure that it works with multiple Flask application instances at once. This is a requirement because many people will use patterns like the 어플리케이션 팩토리 pattern to create their application as needed to aid unittests and to support multiple configurations. Because of that it is crucial that your application supports that kind of behavior.\n\nMost importantly the extension must be shipped with a setup.py file and registered on PyPI. Also the development checkout link should work so that people can easily install the development version into their virtualenv without having to download the library by hand.\n\nFlask extensions must be licensed under a BSD, MIT or more liberal license to be able to be enlisted in the Flask Extension Registry. Keep in mind that the Flask Extension Registry is a moderated place and libraries will be reviewed upfront if they behave as required.\n\n## “Hello Flaskext!”¶\n\nSo let’s get started with creating such a Flask extension. The extension we want to create here will provide very basic support for SQLite3.\n\nFirst we create the following folder structure:\n\n```\nflask-sqlite3/\n flask_sqlite3.py\n LICENSE\n README\n```\n\nHere’s the contents of the most important files:\n\n### setup.py¶\n\nThe next file that is absolutely required is the setup.py file which is used to install your Flask extension. The following contents are something you can work with:\n\n```\n\"\"\"\nFlask-SQLite3\n-------------\n\nThis is the description for that library\n\"\"\"\nfrom setuptools import setup\n\nsetup(\n name='Flask-SQLite3',\n version='1.0',\n url='http://example.com/flask-sqlite3/',\n license='BSD',\n author='',\n author_email='',\n description='Very short description',\n long_description=__doc__,\n py_modules=['flask_sqlite3'],\n # if you would be using a package instead use packages instead\n # of py_modules:\n # packages=['flask_sqlite3'],\n zip_safe=False,\n include_package_data=True,\n platforms='any',\n install_requires=[\n 'Flask'\n ],\n classifiers=[\n 'Environment :: Web Environment',\n 'Intended Audience :: Developers',\n 'License :: OSI Approved :: BSD License',\n 'Operating System :: OS Independent',\n 'Programming Language :: Python',\n 'Topic :: Internet :: WWW/HTTP :: Dynamic Content',\n 'Topic :: Software Development :: Libraries :: Python Modules'\n ]\n)\n```\n\nThat’s a lot of code but you can really just copy/paste that from existing extensions and adapt.\n\n### flask_sqlite3.py¶\n\nNow this is where your extension code goes. But how exactly should such an extension look like? What are the best practices? Continue reading for some insight.\n\n## Initializing Extensions¶\n\nMany extensions will need some kind of initialization step. For example, consider an application that’s currently connecting to SQLite like the documentation suggests (Flask에서 SQLite 3 사용하기). So how does the extension know the name of the application object?\n\nQuite simple: you pass it to it.\n\nThere are two recommended ways for an extension to initialize:\n\ninitialization functions:\n > If your extension is called helloworld you might have a function called\n\n```\ninit_helloworld(app[, extra_args])\n```\n\nthat initializes the extension for that application. It could attach before / after handlers etc. \nclasses:\n > Classes work mostly like initialization functions but can later be used to further change the behavior. For an example look at how the OAuth extension works: there is an OAuth object that provides some helper functions like OAuth.remote_app to create a reference to a remote application that uses OAuth.\n\nWhat to use depends on what you have in mind. For the SQLite 3 extension we will use the class-based approach because it will provide users with an object that handles opening and closing database connections.\n\nWhat’s important about classes is that they encourage to be shared around on module level. In that case, the object itself must not under any circumstances store any application specific state and must be shareable between different application.\n\n## The Extension Code¶\n\nHere’s the contents of the flask_sqlite3.py for copy/paste:\n\n```\nimport sqlite3\n\n# Find the stack on which we want to store the database connection.\n# Starting with Flask 0.9, the _app_ctx_stack is the correct one,\n# before that we need to use the _request_ctx_stack.\ntry:\n from flask import _app_ctx_stack as stack\nexcept ImportError:\n from flask import _request_ctx_stack as stack\n\nclass SQLite3(object):\n\n def __init__(self, app=None):\n if app is not None:\n self.app = app\n self.init_app(self.app)\n else:\n self.app = None\n\n def init_app(self, app):\n app.config.setdefault('SQLITE3_DATABASE', ':memory:')\n # Use the newstyle teardown_appcontext if it's available,\n # otherwise fall back to the request context\n if hasattr(app, 'teardown_appcontext'):\n app.teardown_appcontext(self.teardown)\n else:\n app.teardown_request(self.teardown)\n\n def connect(self):\n return sqlite3.connect(self.app.config['SQLITE3_DATABASE'])\n\n def teardown(self, exception):\n ctx = stack.top\n if hasattr(ctx, 'sqlite3_db'):\n ctx.sqlite3_db.close()\n\n @property\n def connection(self):\n ctx = stack.top\n if ctx is not None:\n if not hasattr(ctx, 'sqlite3_db'):\n ctx.sqlite3_db = self.connect()\n return ctx.sqlite3_db\n```\n\nSo here’s what these lines of code do:\n\nThe\n `__init__` method takes an optional app object and, if supplied, will call `init_app` . * \n\nThe\n `init_app` method exists so that the `SQLite3` object can be instantiated without requiring an app object. This method supports the factory pattern for creating applications. The `init_app` will set the configuration for the database, defaulting to an in memory database if no configuration is supplied. In addition, the `init_app` method attaches the `teardown` handler. It will try to use the newstyle app context handler and if it does not exist, falls back to the request context one. * \n\nNext, we define a\n `connect` method that opens a database connection. * \n\nFinally, we add a\n `connection` property that on first access opens the database connection and stores it on the context. This is also the recommended way to handling resources: fetch resources lazily the first time they are used. \nNote here that we’re attaching our database connection to the top application context via\n `_app_ctx_stack.top` . Extensions should use the top context for storing their own information with a sufficiently complex name. Note that we’re falling back to the \n\n```\n_request_ctx_stack.top\n```\n\nif the application is using an older version of Flask that does not support it. \nSo why did we decide on a class-based approach here? Because using our extension looks something like this:\n\n```\nfrom flask import Flask\nfrom flask_sqlite3 import SQLite3\n\napp = Flask(__name__)\napp.config.from_pyfile('the-config.cfg')\ndb = SQLite3(app)\n```\n\nYou can then use the database from views like this:\n\n```\n@app.route('/')\ndef show_all():\n cur = db.connection.cursor()\n cur.execute(...)\n```\n\nLikewise if you are outside of a request but you are using Flask 0.9 or later with the app context support, you can use the database in the same way:\n\n```\nwith app.app_context():\n cur = db.connection.cursor()\n cur.execute(...)\n```\n\nAt the end of the with block the teardown handles will be executed automatically.\n Additionally, the `init_app` method is used to support the factory pattern\nfor creating apps: \n\n```\ndb = Sqlite3()\n# Then later on.\napp = create_app('the-config.cfg')\ndb.init_app(app)\n```\n\nKeep in mind that supporting this factory pattern for creating apps is required for approved flask extensions (described below).\n Note on `init_app` As you noticed, `init_app` does not assign `app` to `self` . This\nis intentional! Class based Flask extensions must only store the\napplication on the object when the application was passed to the\nconstructor. This tells the extension: I am not interested in using\nmultiple applications. When the extension needs to find the current application and it does not have a reference to it, it must either use the `current_app` context local or change the API in a way\nthat you can pass the application explicitly. \n\n## Using _app_ctx_stack¶\n\n In the example above, before every request, a `sqlite3_db` variable is\nassigned to `_app_ctx_stack.top` . In a view function, this variable is\naccessible using the `connection` property of `SQLite3` . During the\nteardown of a request, the `sqlite3_db` connection is closed. By using\nthis pattern, the same connection to the sqlite3 database is accessible\nto anything that needs it for the duration of the request. If the `_app_ctx_stack` does not exist because the user uses\nan old version of Flask, it is recommended to fall back to `_request_ctx_stack` which is bound to a request. \n\n## Teardown Behavior¶\n\nThis is only relevant if you want to support Flask 0.6 and older\n\nDue to the change in Flask 0.7 regarding functions that are run at the end of the request your extension will have to be extra careful there if it wants to continue to support older versions of Flask. The following pattern is a good way to support both:\n\n```\ndef close_connection(response):\n ctx = _request_ctx_stack.top\n ctx.sqlite3_db.close()\n return response\n\nif hasattr(app, 'teardown_request'):\n app.teardown_request(close_connection)\nelse:\n app.after_request(close_connection)\n```\n\nStrictly speaking the above code is wrong, because teardown functions are passed the exception and typically don’t return anything. However because the return value is discarded this will just work assuming that the code in between does not touch the passed parameter.\n\n## Learn from Others¶\n\nThis documentation only touches the bare minimum for extension development. If you want to learn more, it’s a very good idea to check out existing extensions on the Flask Extension Registry. If you feel lost there is still the mailinglist and the IRC channel to get some ideas for nice looking APIs. Especially if you do something nobody before you did, it might be a very good idea to get some more input. This not only to get an idea about what people might want to have from an extension, but also to avoid having multiple developers working on pretty much the same side by side.\n\nRemember: good API design is hard, so introduce your project on the mailinglist, and let other developers give you a helping hand with designing the API.\n\nThe best Flask extensions are extensions that share common idioms for the API. And this can only work if collaboration happens early.\n\n## Approved Extensions¶\n\nFlask also has the concept of approved extensions. Approved extensions are tested as part of Flask itself to ensure extensions do not break on new releases. These approved extensions are listed on the Flask Extension Registry and marked appropriately. If you want your own extension to be approved you have to follow these guidelines:\n\n* An approved Flask extension requires a maintainer. In the event an extension author would like to move beyond the project, the project should find a new maintainer including full source hosting transition and PyPI access. If no maintainer is available, give access to the Flask core team.\n * An approved Flask extension must provide exactly one package or module named\n `flask_extensionname` . They might also reside inside a `flaskext` namespace packages though this is discouraged now. * It must ship a testing suite that can either be invoked with\n `make test` or `python setup.py test` . For test suites invoked with `make test` the extension has to ensure that all dependencies for the test are installed automatically. If tests are invoked with `python setup.py test` , test dependencies can be specified in the setup.py file. The test suite also has to be part of the distribution. * APIs of approved extensions will be checked for the following characteristics:\n\n* an approved extension has to support multiple applications running in the same Python process.\n * it must be possible to use the factory pattern for creating applications.\n * The license must be BSD/MIT/WTFPL licensed.\n * The naming scheme for official extensions is Flask-ExtensionName or ExtensionName-Flask.\n * Approved extensions must define all their dependencies in the setup.py file unless a dependency cannot be met because it is not available on PyPI.\n * The extension must have documentation that uses one of the two Flask themes for Sphinx documentation.\n * The setup.py description (and thus the PyPI description) has to link to the documentation, website (if there is one) and there must be a link to automatically install the development version (\n `PackageName==dev` ). * The\n `zip_safe` flag in the setup script must be set to `False` , even if the extension would be safe for zipping. * An extension currently has to support Python 2.5, 2.6 as well as Python 2.7\n\n## Extension Import Transition¶\n\nFor a while we recommended using namespace packages for Flask extensions. This turned out to be problematic in practice because many different competing namespace package systems exist and pip would automatically switch between different systems and this caused a lot of problems for users.\n Instead we now recommend naming packages `flask_foo` instead of the now\ndeprecated `flaskext.foo` . Flask 0.8 introduces a redirect import\nsystem that lets uses import from `flask.ext.foo` and it will try `flask_foo` first and if that fails `flaskext.foo` . Flask extensions should urge users to import from `flask.ext.foo` instead of `flask_foo` or `flaskext_foo` so that extensions can\ntransition to the new package name without affecting users.\n\nThe Pocoo styleguide is the styleguide for all Pocoo Projects, including Flask. This styleguide is a requirement for Patches to Flask and a recommendation for Flask extensions.\n\nIn general the Pocoo Styleguide closely follows PEP 8 with some small differences and extensions.\n\n## General Layout¶\n\n* Indentation:\n * 4 real spaces. No tabs, no exceptions.\n * Maximum line length:\n * 79 characters with a soft limit for 84 if absolutely necessary. Try to avoid too nested code by cleverly placing break, continue and return statements.\n * Continuing long statements:\n * \n\nTo continue a statement you can use backslashes in which case you should align the next line with the last dot or equal sign, or indent four spaces:\n > this_is_a_very_long(function_call, 'with many parameters') \\ .that_returns_an_object_with_an_attribute MyModel.query.filter(MyModel.scalar > 120) \\ .order_by(MyModel.name.desc()) \\ .limit(10)\n\nIf you break in a statement with parentheses or braces, align to the braces:\n > this_is_a_very_long(function_call, 'with many parameters', 23, 42, 'and even more')\n\nFor lists or tuples with many items, break immediately after the opening brace:\n > items = [ 'this is the first', 'set of items', 'with more items', 'to come in this line', 'like this' ]\n * Blank lines:\n * \n\nTop level functions and classes are separated by two lines, everything else by one. Do not use too many blank lines to separate logical segments in code. Example:\n > def hello(name): print 'Hello %s!' % name def goodbye(name): print 'See you %s.' % name class MyClass(object): \"\"\"This is a simple docstring\"\"\" def __init__(self, name): self.name = name def get_annoying_name(self): return self.name.upper() + '!!!!111'\n\n## Expressions and Statements¶\n\n* General whitespace rules:\n * \n\n* No whitespace for unary operators that are not words (e.g.:\n `-` , `~` etc.) as well on the inner side of parentheses. * Whitespace is placed between binary operators.\n\nGood:\n > exp = -1.05 value = (item_value / item_count) * offset / exp value = my_list[index] value = my_dict['key']\n\nBad:\n > exp = - 1.05 value = ( item_value / item_count ) * offset / exp value = (item_value/item_count)*offset/exp value=( item_value/item_count ) * offset/exp value = my_list[ index ] value = my_dict ['key']\n * No whitespace for unary operators that are not words (e.g.:\n * Yoda statements are a no-go:\n * \n\nNever compare constant with variable, always variable with constant:\n\nGood:\n > if method == 'md5': pass\n\nBad:\n > if 'md5' == method: pass\n * Comparisons:\n * \n\n* against arbitrary types:\n `==` and `!=` * against singletons with\n `is` and `is not` (eg: `foo is not None` ) * never compare something with True or False (for example never do\n `foo == False` , do `not foo` instead) * against arbitrary types:\n * Negated containment checks:\n * use\n `foo not in bar` instead of `not foo in bar` * Instance checks:\n * \n `isinstance(a, C)` instead of `type(A) is C` , but try to avoid instance checks in general. Check for features. \n\n## Naming Conventions¶\n\n* Class names:\n `CamelCase` , with acronyms kept uppercase ( `HTTPWriter` and not `HttpWriter` ) * Variable names:\n\n * Method and function names:\n\n * Constants:\n\n```\nUPPERCASE_WITH_UNDERSCORES\n```\n\n * precompiled regular expressions:\n `name_re` \nProtected members are prefixed with a single underscore. Double underscores are reserved for mixin classes.\n\nOn classes with keywords, trailing underscores are appended. Clashes with builtins are allowed and must not be resolved by appending an underline to the variable name. If the function needs to access a shadowed builtin, rebind the builtin to a different name instead.\n\n* Function and method arguments:\n * \n\n* class methods:\n `cls` as first parameter * instance methods:\n `self` as first parameter * lambdas for properties might have the first parameter replaced with\n `x` like in \n\n```\ndisplay_name = property(lambda x: x.real_name or x.username)\n```\n\n * class methods:\n\n## Docstrings¶\n\n* Docstring conventions:\n * \n\nAll docstrings are formatted with reStructuredText as understood by Sphinx. Depending on the number of lines in the docstring, they are laid out differently. If it’s just one line, the closing triple quote is on the same line as the opening, otherwise the text is on the same line as the opening quote and the triple quote that closes the string on its own line:\n > def foo(): \"\"\"This is a simple docstring\"\"\" def bar(): \"\"\"This is a longer docstring with so much information in there that it spans three lines. In this case the closing triple quote is on its own line. \"\"\"\n * Module header:\n * \n\nThe module header consists of an utf-8 encoding declaration (if non ASCII letters are used, but it is recommended all the time) and a standard docstring:\n > # -*- coding: utf-8 -*- \"\"\" package.module ~~~~~~~~~~~~~~ A brief description goes here. :copyright: (c) YEAR by AUTHOR. :license: LICENSE_NAME, see LICENSE_FILE for more details. \"\"\"\n\nPlease keep in mind that proper copyrights and license files are a requirement for approved Flask extensions.\n\n## Comments¶\n\n Rules for comments are similar to docstrings. Both are formatted with reStructuredText. If a comment is used to document an attribute, put a colon after the opening pound sign ( `#` ): \n\n```\nclass User(object):\n #: the name of the user as unicode string\n name = Column(String)\n #: the sha1 hash of the password + inline salt\n pw_hash = Column(String)\n```\n\nFlask itself is changing like any software is changing over time. Most of the changes are the nice kind, the kind where you don’t have to change anything in your code to profit from a new release.\n\nHowever every once in a while there are changes that do require some changes in your code or there are changes that make it possible for you to improve your own code quality by taking advantage of new features in Flask.\n\nThis section of the documentation enumerates all the changes in Flask from release to release and how you can change your code to have a painless updating experience.\n If you want to use the easy_install command to upgrade your Flask installation, make sure to pass it the `-U` parameter: \n\n```\n$ easy_install -U Flask\n```\n\n## Version 0.10¶\n\nThe biggest change going from 0.9 to 0.10 is that the cookie serialization format changed from pickle to a specialized JSON format. This change has been done in order to avoid the damage an attacker can do if the secret key is leaked. When you upgrade you will notice two major changes: all sessions that were issued before the upgrade are invalidated and you can only store a limited amount of types in the session. The new sessions are by design much more restricted to only allow JSON with a few small extensions for tuples and strings with HTML markup.\n\nIn order to not break people’s sessions it is possible to continue using the old session system by using the Flask-OldSessions_ extension.\n\n## Version 0.9¶\n\n The behavior of returning tuples from a function was simplified. If you return a tuple it no longer defines the arguments for the response object you’re creating, it’s now always a tuple in the form\n\nwhere at least one item has to be provided. If you depend on\nthe old behavior, you can add it easily by subclassing Flask: \n\n```\nclass TraditionalFlask(Flask):\n def make_response(self, rv):\n if isinstance(rv, tuple):\n return self.response_class(*rv)\n return Flask.make_response(self, rv)\n```\n\n If you maintain an extension that was using `_request_ctx_stack` before, please consider changing to `_app_ctx_stack` if it makes\nsense for your extension. For instance, the app context stack makes sense for\nextensions which connect to databases. Using the app context stack instead of\nthe request stack will make extensions more readily handle use cases outside of\nrequests. \n\n## Version 0.8¶\n\n Flask introduced a new session interface system. We also noticed that there was a naming collision between flask.session the module that implements sessions and `flask.session` which is the global session\nobject. With that introduction we moved the implementation details for\nthe session system into a new module called `flask.sessions` . If you\nused the previously undocumented session support we urge you to upgrade. If invalid JSON data was submitted Flask will now raise a `BadRequest` exception instead of letting the\ndefault `ValueError` bubble up. This has the advantage that you no\nlonger have to handle that error to avoid an internal server error showing\nup for the user. If you were catching this down explicitly in the past\nas ValueError you will need to change this. \nDue to a bug in the test client Flask 0.7 did not trigger teardown handlers when the test client was used in a with statement. This was since fixed but might require some changes in your testsuites if you relied on this behavior.\n\n## Version 0.7¶\n\nIn Flask 0.7 we cleaned up the code base internally a lot and did some backwards incompatible changes that make it easier to implement larger applications with Flask. Because we want to make upgrading as easy as possible we tried to counter the problems arising from these changes by providing a script that can ease the transition.\n\nThe script scans your whole application and generates an unified diff with changes it assumes are safe to apply. However as this is an automated tool it won’t be able to find all use cases and it might miss some. We internally spread a lot of deprecation warnings all over the place to make it easy to find pieces of code that it was unable to upgrade.\n\nWe strongly recommend that you hand review the generated patchfile and only apply the chunks that look good.\n If you are using git as version control system for your project we recommend applying the patch with\n\n```\npath -p1 < patchfile.diff\n```\n\nand then\nusing the interactive commit feature to only apply the chunks that look\ngood. \nTo apply the upgrade script do the following:\n\nDownload the script: flask-07-upgrade.py\n * \n\nRun it in the directory of your application:\n > python flask-07-upgrade.py > patchfile.diff\n * \n\nReview the generated patchfile.\n * \n\nApply the patch:\n > patch -p1 < patchfile.diff\n * \n\nIf you were using per-module template folders you need to move some templates around. Previously if you had a folder named\n `templates` next to a blueprint named `admin` the implicit template path automatically was `admin/index.html` for a template file called `templates/index.html` . This no longer is the case. Now you need to name the template \n\n```\ntemplates/admin/index.html\n```\n\n. The tool will not detect this so you will have to do that on your own. Please note that deprecation warnings are disabled by default starting with Python 2.7. In order to see the deprecation warnings that might be emitted you have to enabled them with the `warnings` module. \nIf you are working with windows and you lack the patch command line utility you can get it as part of various Unix runtime environments for windows including cygwin, msysgit or ming32. Also source control systems like svn, hg or git have builtin support for applying unified diffs as generated by the tool. Check the manual of your version control system for more information.\n\n### Bug in Request Locals¶\n\n Due to a bug in earlier implementations the request local proxies now raise a `RuntimeError` instead of an `AttributeError` when they\nare unbound. If you caught these exceptions with `AttributeError` before, you should catch them with `RuntimeError` now. Additionally the `send_file()` function is now issuing\ndeprecation warnings if you depend on functionality that will be removed\nin Flask 1.0. Previously it was possible to use etags and mimetypes\nwhen file objects were passed. This was unreliable and caused issues\nfor a few setups. If you get a deprecation warning, make sure to\nupdate your application to work with either filenames there or disable\netag attaching and attach them yourself. \nOld code:\n\n```\nreturn send_file(my_file_object)\nreturn send_file(my_file_object)\n```\n\nNew code:\n\n```\nreturn send_file(my_file_object, add_etags=False)\n```\n\n### Upgrading to new Teardown Handling¶\n\n We streamlined the behavior of the callbacks for request handling. For things that modify the response the `after_request()` decorators continue to work as expected, but for things that absolutely\nmust happen at the end of request we introduced the new `teardown_request()` decorator. Unfortunately that\nchange also made after-request work differently under error conditions.\nIt’s not consistently skipped if exceptions happen whereas previously it\nmight have been called twice to ensure it is executed at the end of the\nrequest. \nIf you have database connection code that looks like this:\n\n```\n@app.after_request\ndef after_request(response):\n g.db.close()\n return response\n```\n\nYou are now encouraged to use this instead:\n\n```\n@app.teardown_request\ndef after_request(exception):\n if hasattr(g, 'db'):\n g.db.close()\n```\n\nOn the upside this change greatly improves the internal code flow and makes it easier to customize the dispatching and error handling. This makes it now a lot easier to write unit tests as you can prevent closing down of database connections for a while. You can take advantage of the fact that the teardown callbacks are called when the response context is removed from the stack so a test can query the database after request handling:\n\n```\nwith app.test_client() as client:\n resp = client.get('/')\n # g.db is still bound if there is such a thing\n\n# and here it's gone\n```\n\n### Manual Error Handler Attaching¶\n\n While it is still possible to attach error handlers to `Flask.error_handlers` it’s discouraged to do so and in fact\ndeprecated. In general we no longer recommend custom error handler\nattaching via assignments to the underlying dictionary due to the more\ncomplex internal handling to support arbitrary exception classes and\nblueprints. See `Flask.errorhandler()` for more information. \nThe proper upgrade is to change this:\n\n```\napp.error_handlers[403] = handle_error\n```\n\nInto this:\n\n```\napp.register_error_handler(403, handle_error)\n```\n\nAlternatively you should just attach the function with a decorator:\n\n```\n@app.errorhandler(403)\ndef handle_error(e):\n ...\n```\n\n (Note that\n\n```\nregister_error_handler()\n```\n\nis new in Flask 0.7) \n\n### Blueprint Support¶\n\nBlueprints replace the previous concept of “Modules” in Flask. They provide better semantics for various features and work better with large applications. The update script provided should be able to upgrade your applications automatically, but there might be some cases where it fails to upgrade. What changed?\n\n* Blueprints need explicit names. Modules had an automatic name guesssing scheme where the shortname for the module was taken from the last part of the import module. The upgrade script tries to guess that name but it might fail as this information could change at runtime.\n * Blueprints have an inverse behavior for\n `url_for()` . Previously `.foo` told `url_for()` that it should look for the endpoint foo on the application. Now it means “relative to current module”. The script will inverse all calls to `url_for()` automatically for you. It will do this in a very eager way so you might end up with some unnecessary leading dots in your code if you’re not using modules. * Blueprints do not automatically provide static folders. They will also no longer automatically export templates from a folder called templates next to their location however but it can be enabled from the constructor. Same with static files: if you want to continue serving static files you need to tell the constructor explicitly the path to the static folder (which can be relative to the blueprint’s module path).\n * Rendering templates was simplified. Now the blueprints can provide template folders which are added to a general template searchpath. This means that you need to add another subfolder with the blueprint’s name into that folder if you want\n\n```\nblueprintname/template.html\n```\n\nas the template name. \nIf you continue to use the Module object which is deprecated, Flask will restore the previous behavior as good as possible. However we strongly recommend upgrading to the new blueprints as they provide a lot of useful improvement such as the ability to attach a blueprint multiple times, blueprint specific error handlers and a lot more.\n\n## Version 0.6¶\n\nFlask 0.6 comes with a backwards incompatible change which affects the order of after-request handlers. Previously they were called in the order of the registration, now they are called in reverse order. This change was made so that Flask behaves more like people expected it to work and how other systems handle request pre- and postprocessing. If you depend on the order of execution of post-request functions, be sure to change the order.\n\nAnother change that breaks backwards compatibility is that context processors will no longer override values passed directly to the template rendering function. If for example request is as variable passed directly to the template, the default context processor will not override it with the current request object. This makes it easier to extend context processors later to inject additional variables without breaking existing template not expecting them.\n\n## Version 0.5¶\n\nFlask 0.5 is the first release that comes as a Python package instead of a single module. There were a couple of internal refactoring so if you depend on undocumented internal details you probably have to adapt the imports.\n\nThe following changes may be relevant to your application:\n\n* autoescaping no longer happens for all templates. Instead it is configured to only happen on files ending with\n `.html` , `.htm` , `.xml` and `.xhtml` . If you have templates with different extensions you should override the \n\n```\nselect_jinja_autoescape()\n```\n\nmethod. * Flask no longer supports zipped applications in this release. This functionality might come back in future releases if there is demand for this feature. Removing support for this makes the Flask internal code easier to understand and fixes a couple of small issues that make debugging harder than necessary.\n * The create_jinja_loader function is gone. If you want to customize the Jinja loader now, use the\n\n```\ncreate_jinja_environment()\n```\n\nmethod instead. \n\n## Version 0.4¶\n\n For application developers there are no changes that require changes in your code. In case you are developing on a Flask extension however, and that extension has a unittest-mode you might want to link the activation of that mode to the new `TESTING` flag. \n\n## Version 0.3¶\n\nFlask 0.3 introduces configuration support and logging as well as categories for flashing messages. All these are features that are 100% backwards compatible but you might want to take advantage of them.\n\n### Configuration Support¶\n\nThe configuration support makes it easier to write any kind of application that requires some sort of configuration. (Which most likely is the case for any application out there).\n\nIf you previously had code like this:\n\n```\napp.debug = DEBUG\napp.secret_key = SECRET_KEY\n```\n\nYou no longer have to do that, instead you can just load a configuration into the config object. How this works is outlined in 설정 다루기.\n\n### Logging Integration¶\n\nFlask now configures a logger for you with some basic and useful defaults. If you run your application in production and want to profit from automatic error logging, you might be interested in attaching a proper log handler. Also you can start logging warnings and errors into the logger when appropriately. For more information on that, read 어플리케이션 에러 로깅하기.\n\n### Categories for Flash Messages¶\n\nFlash messages can now have categories attached. This makes it possible to render errors, warnings or regular messages differently for example. This is an opt-in feature because it requires some rethinking in the code.\n\nRead all about that in the 메시지 플래싱(Message Flashing) pattern.\n\nFlask is licensed under a three clause BSD License. It basically means: do whatever you want with it as long as the copyright in Flask sticks around, the conditions are not modified and the disclaimer is present. Furthermore you must not use the names of the authors to promote derivatives of the software without written consent.\n\nThe full license text can be found below (Flask License). For the documentation and artwork different licenses apply.\n\n## General License Definitions¶\n\nThe following section contains the full license texts for Flask and the documentation.\n\n* “AUTHORS” hereby refers to all the authors listed in the Authors section.\n * The “Flask License” applies to all the sourcecode shipped as part of Flask (Flask itself as well as the examples and the unittests) as well as documentation.\n * The “Flask Artwork License” applies to the project’s Horn-Logo.\n"}}},{"rowIdx":95,"cells":{"project":{"kind":"string","value":"predictNMB"},"source":{"kind":"string","value":"cran"},"language":{"kind":"string","value":"R"},"content":{"kind":"string","value":"Package ‘predictNMB’\n June 3, 2023\nType Package\nTitle Evaluate Clinical Prediction Models by Net Monetary Benefit\nVersion 0.2.1\nDescription Estimates when and where a model-guided treatment strategy may\n outperform a treat-all or treat-none approach by Monte Carlo simulation and\n evaluation of the Net Monetary Benefit. Details can be viewed in\n Parsons et al. (2023) .\nLicense GPL (>= 3)\nEncoding UTF-8\nRoxygenNote 7.2.3\nImports assertthat, cutpointr, dplyr, ggplot2, magrittr, pmsampsize,\n rlang, scales, stats, tibble, tidyr\nSuggests spelling, covr, flextable, knitr, parallel, pbapply,\n rmarkdown, testthat (>= 3.0.0), vdiffr, withr\nConfig/testthat/edition 3\nVignetteBuilder knitr\nURL https://docs.ropensci.org/predictNMB/\nBugReports https://github.com/ropensci/predictNMB/issues\nDepends R (>= 3.5.0)\nLanguage en-US\nNeedsCompilation no\nAuthor [aut, cre] (),\n [aut] (),\n [aut] (),\n [rev] (Emi Tanaka reviewed predictNMB for rOpenSci, see\n .),\n [rev] (Tariyawasam reviewed predictNMB for\n rOpenSci, see\n .),\n [ctb] (),\n [ctb] ()\nMaintainer <>\nRepository CRAN\nDate/Publication 2023-06-03 07:40:02 UTC\nR topics documented:\nautoplot.predictNMBscree... 2\nautoplot.predictNMBsi... 4\nce_plo... 6\nce_plot.predictNMBsi... 8\ndo_nmb_si... 9\nevaluate_cutpoint_cos... 11\nevaluate_cutpoint_nm... 12\nevaluate_cutpoint_qaly... 13\nget_inbuilt_cutpoin... 14\nget_inbuilt_cutpoint_method... 14\nget_nmb_sample... 15\nget_sampl... 16\nget_threshold... 17\nprint.predictNMBscree... 19\nprint.predictNMBsi... 19\nscreen_simulation_input... 20\nsummary.predictNMBscree... 22\nsummary.predictNMBsi... 23\ntheme_si... 25\n autoplot.predictNMBscreen\n Create plots of from screened predictNMB simulations.\nDescription\n Create plots of from screened predictNMB simulations.\nUsage\n ## S3 method for class 'predictNMBscreen'\n autoplot(\n object,\n x_axis_var = NULL,\n constants = list(),\n what = c(\"nmb\", \"inb\", \"cutpoints\", \"qalys\", \"costs\"),\n inb_ref_col = NA,\n plot_range = TRUE,\n plot_conf_level = TRUE,\n plot_line = TRUE,\n plot_alpha = 0.5,\n dodge_width = 0,\n conf.level = 0.95,\n methods_order = NULL,\n rename_vector,\n ...\n )\nArguments\n object A predictNMBscreen object.\n x_axis_var The desired screened factor to be displayed along the x axis. For example, if\n the simulation screen was used with many values for event rate, this could be\n \"event_rate\". Defaults to the first detected, varied input.\n constants Named vector. If multiple inputs were screened in this object, this argument can\n be used to modify the selected values for all those except the input that’s varying\n along the x-axis. See the summarising methods vignette.\n what What to summarise: one of \"nmb\", \"inb\", \"cutpoints\", \"qalys\" or \"costs\". De-\n faults to \"nmb\".\n inb_ref_col Which cutpoint method to use as the reference strategy when calculating the\n incremental net monetary benefit. See do_nmb_sim for more information.\n plot_range logical. Whether or not to plot the range of the distribution as a thin line.\n Defaults to TRUE.\n plot_conf_level\n logical. Whether or not to plot the confidence region of the distribution as a\n thicker line. Defaults to TRUE.\n plot_line logical. Whether or not to connect the medians of the distributions for each\n method along the x-axis. Defaults to TRUE.\n plot_alpha Alpha value (transparency) of all plot elements. Defaults to 0.5.\n dodge_width The dodge width of plot elements. Can be used to avoid excessive overlap be-\n tween methods. Defaults to 0.\n conf.level The confidence level of the interval. Defaults to 0.95 (coloured area of distribu-\n tion represents 95% CIs).\n methods_order The order (left to right) to display the cutpoint methods.\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n ... Additional (unused) arguments.\nDetails\n This plot method works with predictNMBscreen objects that are created using screen_simulation_inputs().\n Can be used to visualise distributions from many different simulations and assign a varying input to\n the x-axis of the plot.\nValue\n Returns a ggplot object.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_screen_obj <- screen_simulation_inputs(\n n_sims = 50, n_valid = 10000, sim_auc = seq(0.7, 0.9, 0.1),\n event_rate = c(0.1, 0.2, 0.3),\n fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb,\n cutpoint_methods = c(\"all\", \"none\", \"youden\", \"value_optimising\")\n )\n autoplot(sim_screen_obj)\n autoplot(\n sim_screen_obj,\n x_axis_var = \"event_rate\",\n constants = c(sim_auc = 0.8),\n dodge_width = 0.02,\n rename_vector = c(\n \"Value-Optimising\" = \"value_optimising\",\n \"Treat-None\" = \"none\",\n \"Treat-All\" = \"all\",\n \"Youden Index\" = \"youden\"\n )\n )\n autoplot.predictNMBsim\n Create plots of from predictNMB simulations.\nDescription\n Create plots of from predictNMB simulations.\nUsage\n ## S3 method for class 'predictNMBsim'\n autoplot(\n object,\n what = c(\"nmb\", \"inb\", \"cutpoints\", \"qalys\", \"costs\"),\n inb_ref_col = NA,\n conf.level = 0.95,\n methods_order = NULL,\n n_bins = 40,\n label_wrap_width = 12,\n fill_cols = c(\"grey50\", \"#ADD8E6\"),\n median_line_size = 2,\n median_line_alpha = 0.5,\n median_line_col = \"black\",\n rename_vector,\n ...\n )\nArguments\n object A predictNMBsim object.\n what What to summarise: one of \"nmb\", \"inb\", \"cutpoints\", \"qalys\" or \"costs\". De-\n faults to \"nmb\".\n inb_ref_col Which cutpoint method to use as the reference strategy when calculating the\n incremental net monetary benefit. See do_nmb_sim for more information.\n conf.level The confidence level of the interval. Defaults to 0.95 (coloured area of distribu-\n tion represents 95% CIs).\n methods_order The order (left to right) to display the cutpoint methods.\n n_bins The number of bins used when constructing histograms. Defaults to 40.\n label_wrap_width\n The number of characters in facet labels at which the label is wrapped. Default\n is 12.\n fill_cols Vector containing the colours used for fill aesthetic of histograms. The first\n colour represents the area outside of the confidence region, second colour shows\n the confidence region. Defaults to c(\"grey50\", \"#ADD8E6\").\n median_line_size\n Size of line used to represent the median of distribution. Defaults to 2.\n median_line_alpha\n Alpha (transparency) for line used to represent the median of distribution. De-\n faults to 0.5.\n median_line_col\n Colour of line used to represent the median of distribution. Defaults to \"black\".\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n ... Additional (unused) arguments.\nDetails\n This plot method works with predictNMBsim objects that are created using do_nmb_sim(). Can be\n used to visualise distributions from simulations for different cutpoint methods.\nValue\n Returns a ggplot object.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb,\n cutpoint_methods = c(\"all\", \"none\", \"youden\", \"value_optimising\")\n )\n autoplot(\n sim_obj,\n rename_vector = c(\n \"Value- Optimising\" = \"value_optimising\",\n \"Treat- None\" = \"none\",\n \"Treat- All\" = \"all\",\n \"Youden Index\" = \"youden\"\n )\n ) + theme_sim()\n ce_plot Create a cost-effectiveness plot.\nDescription\n Create a cost-effectiveness plot.\nUsage\n ce_plot(\n object,\n ref_col,\n wtp,\n show_wtp = TRUE,\n methods_order = NULL,\n rename_vector,\n shape = 21,\n wtp_linetype = \"dashed\",\n add_prop_ce = FALSE,\n ...\n )\nArguments\n object A predictNMBsim object.\n ref_col Which cutpoint method to use as the reference strategy when calculating the in-\n cremental net monetary benefit. Often sensible to use a \"all\" or \"none\" approach\n for this.\n wtp A numeric. The willingness to pay (WTP) value used to create a WTP thresh-\n old line on the plot (if show_wtp = TRUE). Defaults to the WTP stored in the\n predictNMBsim object.\n show_wtp A logical. Whether or not to show the willingness to pay threshold.\n methods_order The order (within the legend) to display the cutpoint methods.\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n shape The shape used for ggplot2::geom_point(). Defaults to 21 (hollow cir-\n cles). If shape = \"method\" or shape = \"cost-effective\" (only applicable\n when show_wtp = TRUE) , then the shape will be mapped to that aesthetic.\n wtp_linetype The linetype used for ggplot2::geom_abline() when making the WTP. De-\n faults to \"dashed\".\n add_prop_ce Whether to append the proportion of simulations for that method which were\n cost-effective (beneath the WTP threshold) to their labels in the legend. Only\n applicable when show_wtp = TRUE.\n ... Additional (unused) arguments.\nDetails\n This plot method works with predictNMBsim objects that are created using do_nmb_sim(). Can be\n used to visualise the simulations on a cost-effectiveness plot (costs vs effectiveness)\nValue\n Returns a ggplot object.\nExamples\n get_nmb_evaluation <- get_nmb_sampler(\n qalys_lost = function() rnorm(1, 0.33, 0.03),\n wtp = 28000,\n high_risk_group_treatment_effect = function() exp(rnorm(n = 1, mean = log(0.58), sd = 0.43)),\n high_risk_group_treatment_cost = function() rnorm(n = 1, mean = 161, sd = 49)\n )\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb_evaluation, fx_nmb_evaluation = get_nmb_evaluation\n )\n ce_plot(sim_obj, ref_col = \"all\")\n ce_plot.predictNMBsim Create a cost-effectiveness plot.\nDescription\n Create a cost-effectiveness plot.\nUsage\n ## S3 method for class 'predictNMBsim'\n ce_plot(\n object,\n ref_col,\n wtp,\n show_wtp = TRUE,\n methods_order = NULL,\n rename_vector,\n shape = 21,\n wtp_linetype = \"dashed\",\n add_prop_ce = FALSE,\n ...\n )\nArguments\n object A predictNMBsim object.\n ref_col Which cutpoint method to use as the reference strategy when calculating the in-\n cremental net monetary benefit. Often sensible to use a \"all\" or \"none\" approach\n for this.\n wtp A numeric. The willingness to pay (WTP) value used to create a WTP thresh-\n old line on the plot (if show_wtp = TRUE). Defaults to the WTP stored in the\n predictNMBsim object.\n show_wtp A logical. Whether or not to show the WTP threshold.\n methods_order The order (within the legend) to display the cutpoint methods.\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n shape The shape used for ggplot2::geom_point(). Defaults to 21 (hollow cir-\n cles). If shape = \"method\" or shape = \"cost-effective\" (only applicable\n when show_wtp = TRUE) , then the shape will be mapped to that aesthetic.\n wtp_linetype The linetype used for ggplot2::geom_abline() when making the WTP. De-\n faults to \"dashed\".\n add_prop_ce Whether to append the proportion of simulations for that method which were\n cost-effective (beneath the WTP threshold) to their labels in the legend. Only\n applicable when show_wtp = TRUE.\n ... Additional (unused) arguments.\nDetails\n This plot method works with predictNMBsim objects that are created using do_nmb_sim(). Can be\n used to visualise the simulations on a cost-effectiveness plot (costs vs effectiveness)\nValue\n Returns a ggplot object.\nExamples\n get_nmb_evaluation <- get_nmb_sampler(\n qalys_lost = function() rnorm(1, 0.33, 0.03),\n wtp = 28000,\n high_risk_group_treatment_effect = function() exp(rnorm(n = 1, mean = log(0.58), sd = 0.43)),\n high_risk_group_treatment_cost = function() rnorm(n = 1, mean = 161, sd = 49)\n )\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb_evaluation, fx_nmb_evaluation = get_nmb_evaluation\n )\n ce_plot(sim_obj, ref_col = \"all\")\n do_nmb_sim Do the predictNMB simulation, evaluating the net monetary benefit\n (NMB) of the simulated model.\nDescription\n Do the predictNMB simulation, evaluating the net monetary benefit (NMB) of the simulated model.\nUsage\n do_nmb_sim(\n sample_size,\n n_sims,\n n_valid,\n sim_auc,\n event_rate,\n cutpoint_methods = get_inbuilt_cutpoint_methods(),\n fx_nmb_training,\n fx_nmb_evaluation,\n meet_min_events = TRUE,\n min_events = NA,\n show_progress = FALSE,\n cl = NULL\n )\nArguments\n sample_size Sample size of training set. If missing, a sample size calculation will be per-\n formed and the calculated size will be used.\n n_sims Number of simulations to run.\n n_valid Sample size for evaluation set.\n sim_auc Simulated model discrimination (AUC).\n event_rate Simulated event rate of the binary outcome being predicted. Also known as\n prevalence.\n cutpoint_methods\n A value or vector of cutpoint methods to include. Defaults to use the inbuilt\n methods:\n • \"all\" = treat all patients (cutpoint = 0)\n • \"none\" = treat no patients (cutpoint = 1)\n • \"value_optimising\" = select the cutpoint that maximises NMB\n • \"youden\" = select cutpoint based on the Youden index, also known as the\n J-index (sensitivity + specificity - 1)\n • \"cost_minimising\" = select the cutpoint that minimises expected value of\n costs\n • \"prod_sens_spec\" = product of sensitivity and specificity (sensitivity * speci-\n ficity)\n • \"roc01\" = selects the closest threshold to the (0,1) point on the ROC curve\n User-defined cutpoint methods can be used by passing the name of a function\n that takes the following arguments:\n • predicted (predicted probabilities)\n • actual (the actual, binary outcome)\n • nmb (a named vector containing NMB values assigned to each predicted\n class (i.e. c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)))\n See ?get_thresholds for an example of a user-defined cutpoint function.\n fx_nmb_training\n Function or NMBsampler that returns a named vector of NMB assigned to clas-\n sifications used for obtaining cutpoint on training set.\n fx_nmb_evaluation\n Function or NMBsampler that returns a named vector of NMB assigned to clas-\n sifications used for obtaining cutpoint on evaluation set.\n meet_min_events\n Whether or not to incrementally add samples until the expected number of events\n (sample_size * event_rate) is met. (Applies to sampling of training data\n only.)\n min_events The minimum number of events to include in the training sample. If less than\n this number are included in sample of size sample_size, additional samples are\n added until the min_events is met. The default (NA) will use the expected value\n given the event_rate and the sample_size.\n show_progress Logical. Whether to display a progress bar. Requires the pbapply package.\n cl A cluster made using parallel::makeCluster(). If a cluster is provided, the\n simulation will be done in parallel.\nDetails\n This function runs a simulation for a given set of inputs that represent a healthcare setting using\n model-guided interventions.\n The arguments fx_nmb_training and fx_nmb_evaluation should be functions that capture the\n treatment being used, its costs and effectiveness, and the costs of the outcome being treated/prevented.\n Both of these are functions that return a named vector of NMB values when called and are used\n for obtaining and evaluating cutpoints, respectively. For example, the following function returns\n the appropriately named vector.\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n There is a helper function, get_nmb_sampler(), to help you create these.\nValue\n Returns a predictNMBsim object.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb\n )\n evaluate_cutpoint_cost\n Evaluates a cutpoint by returning the mean treatment cost per sample.\nDescription\n Evaluates a cutpoint by returning the mean treatment cost per sample.\nUsage\n evaluate_cutpoint_cost(predicted, actual, pt, nmb)\nArguments\n predicted A vector of predicted probabilities.\n actual A vector of actual outcomes.\n pt The probability threshold to be evaluated.\n nmb A named vector containing NMB assigned to each classification and the treat-\n ment costs.\nValue\n Returns a numeric value representing the mean cost for that cutpoint and data.\nExamples\n evaluate_cutpoint_cost(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n pt = 0.1,\n nmb = c(\n \"qalys_lost\" = 5,\n \"low_risk_group_treatment_cost\" = 0,\n \"high_risk_group_treatment_cost\" = 1,\n \"low_risk_group_treatment_effect\" = 0,\n \"high_risk_group_treatment_effect\" = 0.3,\n \"outcome_cost\" = 10\n )\n )\n evaluate_cutpoint_nmb Evaluates a cutpoint by returning the mean NMB per sample.\nDescription\n Evaluates a cutpoint by returning the mean NMB per sample.\nUsage\n evaluate_cutpoint_nmb(predicted, actual, pt, nmb)\nArguments\n predicted A vector of predicted probabilities.\n actual A vector of actual outcomes.\n pt The probability threshold to be evaluated.\n nmb A named vector containing NMB assigned to each classification.\nValue\n Returns a numeric value representing the NMB for that cutpoint and data.\nExamples\n evaluate_cutpoint_nmb(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n pt = 0.1,\n nmb = c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n )\n evaluate_cutpoint_qalys\n Evaluates a cutpoint by returning the mean QALYs lost per sample.\nDescription\n Evaluates a cutpoint by returning the mean QALYs lost per sample.\nUsage\n evaluate_cutpoint_qalys(predicted, actual, pt, nmb)\nArguments\n predicted A vector of predicted probabilities.\n actual A vector of actual outcomes.\n pt The probability threshold to be evaluated.\n nmb A named vector containing NMB assigned to each classification and the treat-\n ment effects and QALYS lost due to the event of interest.\nValue\n Returns a numeric value representing the mean QALYs for that cutpoint and data.\nExamples\n evaluate_cutpoint_qalys(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n pt = 0.1,\n nmb = c(\n \"qalys_lost\" = 5,\n \"low_risk_group_treatment_effect\" = 0,\n \"high_risk_group_treatment_effect\" = 0.5\n )\n )\n get_inbuilt_cutpoint Get a cutpoint using the methods inbuilt to predictNMB\nDescription\n Get a cutpoint using the methods inbuilt to predictNMB\nUsage\n get_inbuilt_cutpoint(predicted, actual, nmb, method)\nArguments\n predicted A vector of predicted probabilities\n actual A vector of actual outcomes\n nmb A named vector containing NMB assigned to each classification\n method A cutpoint selection method to be used methods that can be used as the method\n argument\nValue\n Returns a selected cutpoint (numeric).\nExamples\n ## get the list of available methods:\n get_inbuilt_cutpoint_methods()\n ## get the cutpoint that maximises the Youden index for a given set of\n ## probabilities and outcomes\n get_inbuilt_cutpoint(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n method = \"youden\"\n )\n get_inbuilt_cutpoint_methods\n Get a vector of all the inbuilt cutpoint methods\nDescription\n Get a vector of all the inbuilt cutpoint methods\nUsage\n get_inbuilt_cutpoint_methods()\nValue\n Returns a vector cutpoint methods that can be used in do_nmb_sim().\nExamples\n get_inbuilt_cutpoint_methods()\n get_nmb_sampler Make a NMB sampler for use in do_nmb_sim() or\n screen_simulation_inputs()\nDescription\n Make a NMB sampler for use in do_nmb_sim() or screen_simulation_inputs()\nUsage\n get_nmb_sampler(\n outcome_cost,\n wtp,\n qalys_lost,\n high_risk_group_treatment_effect,\n high_risk_group_treatment_cost,\n low_risk_group_treatment_effect = 0,\n low_risk_group_treatment_cost = 0,\n use_expected_values = FALSE,\n nboot = 10000\n )\nArguments\n outcome_cost The cost of the outcome. Must be provided if wtp and qalys_lost are not. Or\n can be used in addition to these arguments to represent additional cost to the\n health burden.\n wtp Willingness-to-pay.\n qalys_lost Quality-adjusted life years (QALYs) lost due to healthcare event being pre-\n dicted.\n high_risk_group_treatment_effect\n The effect of the treatment provided to patients given high risk prediction. Can\n be a number of a function. Provide a function to incorporate uncertainty.\n high_risk_group_treatment_cost\n The cost of the treatment provided to patients given high risk prediction. Can be\n a number of a function. Provide a function to incorporate uncertainty.\n low_risk_group_treatment_effect\n The effect of the treatment provided to patients given low risk prediction. Can be\n a number of a function. Provide a function to incorporate uncertainty. Defaults\n to 0 (no treatment).\n low_risk_group_treatment_cost\n The cost of the treatment provided to patients given low risk prediction. Can be\n a number of a function. Provide a function to incorporate uncertainty. Defaults\n to 0 (no treatment).\n use_expected_values\n Logical. If TRUE, gets the mean of many samples from the produced function\n and returns these every time. This is a sensible choice when using the resulting\n function for selecting the cutpoint. See fx_nmb_training. Defaults to FALSE.\n nboot The number of samples to use when creating a function that returns the expected\n values. Defaults to 10000.\nValue\n Returns a NMBsampler object.\nExamples\n get_nmb_training <- get_nmb_sampler(\n outcome_cost = 100,\n high_risk_group_treatment_effect = function() rbeta(1, 1, 2),\n high_risk_group_treatment_cost = 10,\n use_expected_values = TRUE\n )\n get_nmb_evaluation <- get_nmb_sampler(\n outcome_cost = 100,\n high_risk_group_treatment_effect = function() rbeta(1, 1, 2),\n high_risk_group_treatment_cost = 10\n )\n get_nmb_training()\n get_nmb_training()\n get_nmb_training()\n get_nmb_evaluation()\n get_nmb_evaluation()\n get_nmb_evaluation()\n get_sample Samples data for a prediction model with a specified AUC and preva-\n lence.\nDescription\n Samples data for a prediction model with a specified AUC and prevalence.\nUsage\n get_sample(auc, n_samples, prevalence, min_events = 0)\nArguments\n auc The Area Under the (receiver operating characteristic) Curve.\n n_samples Number of samples to draw.\n prevalence Prevalence or event rate of the binary outcome as a proportion (0.1 = 10%).\n min_events Minimum number of events required in the sample.\nValue\n Returns a data.frame.\nExamples\n get_sample(0.7, 1000, 0.1)\n get_thresholds Gets probability thresholds given predicted probabilities, outcomes\n and NMB.\nDescription\n Gets probability thresholds given predicted probabilities, outcomes and NMB.\nUsage\n get_thresholds(predicted, actual, nmb, cutpoint_methods = NULL)\nArguments\n predicted A vector of predicted probabilities.\n actual A vector of actual outcomes.\n nmb A named vector containing NMB assigned to true positives, true negatives, false\n positives and false negatives\n cutpoint_methods\n Which cutpoint method(s) to return. The default (NULL) uses all the inbuilt\n methods.\nValue\n Returns a list.\nExamples\n # get thresholds using default (all inbuilt) cutpoint methods\n get_thresholds(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n nmb = c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n )\n # get cutpoints using user-defined functions\n # These functions must take the \\code{predicted} and \\code{actual}\n # as arguments. They can also take \\code{nmb} (named vector containing NMB\n # with values for TP, FP, TN, FN).\n fx_roc01 <- function(predicted, actual, ...) {\n cutpointr::cutpointr(\n x = predicted, class = actual, method = cutpointr::minimize_metric,\n metric = cutpointr::roc01,\n silent = TRUE\n )[[\"optimal_cutpoint\"]]\n }\n fx_sum_sens_spec <- function(predicted, actual, ...) {\n cutpointr::cutpointr(\n x = predicted, class = actual, method = cutpointr::maximize_metric,\n metric = cutpointr::sum_sens_spec,\n silent = TRUE\n )[[\"optimal_cutpoint\"]]\n }\n get_thresholds(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n cutpoint_methods = c(\"fx_roc01\", \"fx_sum_sens_spec\"),\n nmb = c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n )\n # get a combination of cutpoints from both user-defined functions and\n # inbuilt methods\n get_thresholds(\n predicted = runif(1000),\n actual = sample(c(0, 1), size = 1000, replace = TRUE),\n cutpoint_methods = c(\n \"fx_roc01\",\n \"fx_sum_sens_spec\",\n \"youden\",\n \"all\",\n \"none\"\n ),\n nmb = c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n )\n print.predictNMBscreen\n Print a summary of a predictNMBscreen object\nDescription\n Print a summary of a predictNMBscreen object\nUsage\n ## S3 method for class 'predictNMBscreen'\n print(x, ...)\nArguments\n x A predictNMBscreen object.\n ... Optional, ignored arguments.\nValue\n print(x) returns x invisibly.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_screen_obj <- screen_simulation_inputs(\n n_sims = 50, n_valid = 10000, sim_auc = seq(0.7, 0.9, 0.1),\n event_rate = 0.1,\n fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb\n )\n print(sim_screen_obj)\n print.predictNMBsim Print a summary of a predictNMBsim object\nDescription\n Print a summary of a predictNMBsim object\nUsage\n ## S3 method for class 'predictNMBsim'\n print(x, ...)\nArguments\n x A predictNMBsim object.\n ... Optional, ignored arguments.\nValue\n print(x) returns x invisibly.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb\n )\n print(sim_obj)\n screen_simulation_inputs\n Screen many simulation inputs: a parent function to do_nmb_sim()\nDescription\n Runs do_nmb_sim() with a range of inputs.\nUsage\n screen_simulation_inputs(\n sample_size,\n n_sims,\n n_valid,\n sim_auc,\n event_rate,\n cutpoint_methods = get_inbuilt_cutpoint_methods(),\n fx_nmb_training,\n fx_nmb_evaluation,\n pair_nmb_train_and_evaluation_functions = FALSE,\n meet_min_events = TRUE,\n min_events = NA,\n show_progress = FALSE,\n cl = NULL\n )\nArguments\n sample_size A value (or vector of values): Sample size of training set. If missing, a sample\n size calculation will be performed and the calculated size will be used.\n n_sims A value (or vector of values): Number of simulations to run.\n n_valid A value (or vector of values): Sample size for evaluation set.\n sim_auc A value (or vector of values): Simulated model discrimination (AUC).\n event_rate A value (or vector of values): simulated event rate of the binary outcome being\n predicted.\n cutpoint_methods\n cutpoint methods to include. Defaults to use the inbuilt methods. This doesn’t\n change across calls to do_nmb_sim().\n fx_nmb_training\n A function or NMBsampler (or list of) that returns named vector of NMB as-\n signed to classifications use for obtaining cutpoint on training set.\n fx_nmb_evaluation\n A function or NMBsampler (or list of) that returns named vector of NMB as-\n signed to classifications use for obtaining cutpoint on evaluation set.\n pair_nmb_train_and_evaluation_functions\n logical. Whether or not to pair the lists of functions passed for fx_nmb_training\n and fx_nmb_evaluation. If two treatment strategies are being used, it may\n make more sense to pair these because selecting a value-optimising or cost-\n minimising threshold using one strategy but evaluating another is likely un-\n wanted.\n meet_min_events\n Whether or not to incrementally add samples until the expected number of events\n (sample_size * event_rate) is met. (Applies to sampling of training data\n only.)\n min_events A value: the minimum number of events to include in the training sample. If\n less than this number are included in sample of size sample_size, additional\n samples are added until the min_events is met. The default (NA) will use the\n expected value given the event_rate and the sample_size.\n show_progress Logical. Whether to display a progress bar.\n cl A cluster made using parallel::makeCluster(). If a cluster is provided, the\n simulation will be done in parallel.\nValue\n Returns a predictNMBscreen object.\nExamples\n # Screen for optimal cutpoints given increasing values of\n # model discrimination (sim_auc)\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_screen_obj <- screen_simulation_inputs(\n n_sims = 50, n_valid = 10000, sim_auc = seq(0.7, 0.9, 0.1),\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb\n )\n summary.predictNMBscreen\n Create table summaries of predictNMBscreen objects.\nDescription\n Create table summaries of predictNMBscreen objects.\nUsage\n ## S3 method for class 'predictNMBscreen'\n summary(\n object,\n what = c(\"nmb\", \"inb\", \"cutpoints\"),\n inb_ref_col = NULL,\n agg_functions = list(median = function(x) {\n round(stats::median(x), digits = 2)\n }, `95% CI` = function(x) {\n paste0(round(stats::quantile(x, probs = c(0.025,\n 0.975)), digits = 1), collapse = \" to \")\n }),\n rename_vector,\n show_full_inputs = FALSE,\n ...\n )\nArguments\n object A predictNMBscreen object.\n what What to summarise: one of \"nmb\", \"inb\" or \"cutpoints\". Defaults to \"nmb\".\n inb_ref_col Which cutpoint method to use as the reference strategy when calculating the\n incremental net monetary benefit. See do_nmb_sim for more information.\n agg_functions A named list of functions to use to aggregate the selected values. Defaults to the\n median and 95% interval.\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n show_full_inputs\n A logical. Whether or not to include the inputs used for simulation alongside\n aggregations.\n ... Additional, ignored arguments.\nDetails\n Table summaries will be based on the what argument. Using \"nmb\" returns the simulated values\n for NMB, with no reference group; \"inb\" returns the difference between simulated values for NMB\n and a set strategy defined by inb_ref_col; \"cutpoints\" returns the cutpoints selected (0, 1).\nValue\n Returns a tibble.\nExamples\n # perform screen with increasing values of model discimination (sim_auc)\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_screen_obj <- screen_simulation_inputs(\n n_sims = 50, n_valid = 10000, sim_auc = seq(0.7, 0.9, 0.1),\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb,\n cutpoint_methods = c(\"all\", \"none\", \"youden\", \"value_optimising\")\n )\n summary(\n sim_screen_obj,\n rename_vector = c(\n \"Value_Optimising\" = \"value_optimising\",\n \"Treat_None\" = \"none\",\n \"Treat_All\" = \"all\",\n \"Youden_Index\" = \"youden\"\n )\n )\n summary.predictNMBsim Create table summaries of predictNMBsim objects.\nDescription\n Create table summaries of predictNMBsim objects.\nUsage\n ## S3 method for class 'predictNMBsim'\n summary(\n object,\n what = c(\"nmb\", \"inb\", \"cutpoints\"),\n inb_ref_col = NULL,\n agg_functions = list(median = function(x) {\n round(stats::median(x), digits = 2)\n }, `95% CI` = function(x) {\n paste0(round(stats::quantile(x, probs = c(0.025,\n 0.975)), digits = 1), collapse = \" to \")\n }),\n rename_vector,\n ...\n )\nArguments\n object A predictNMBsim object.\n what What to summarise: one of \"nmb\", \"inb\" or \"cutpoints\". Defaults to \"nmb\".\n inb_ref_col Which cutpoint method to use as the reference strategy when calculating the\n incremental net monetary benefit. See do_nmb_sim for more information.\n agg_functions A named list of functions to use to aggregate the selected values. Defaults to the\n median and 95% interval.\n rename_vector A named vector for renaming the methods in the summary. The values of the\n vector are the default names and the names given are the desired names in the\n output.\n ... Additional, ignored arguments.\nDetails\n Table summaries will be based on the what argument. Using \"nmb\" returns the simulated values\n for NMB, with no reference group; \"inb\" returns the difference between simulated values for NMB\n and a set strategy defined by inb_ref_col; \"cutpoints\" returns the cutpoints selected (0, 1).\nValue\n Returns a tibble.\nExamples\n # perform simulation with do_nmb_sim()\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb,\n cutpoint_methods = c(\"all\", \"none\", \"youden\", \"value_optimising\")\n )\n summary(\n sim_obj,\n rename_vector = c(\n \"Value_Optimising\" = \"value_optimising\",\n \"Treat_None\" = \"none\",\n \"Treat_All\" = \"all\",\n \"Youden_Index\" = \"youden\"\n )\n )\n theme_sim Returns a ggplot2 theme that reduces clutter in an autoplot() of a\n predictNMBsim object.\nDescription\n Returns a ggplot2 theme that reduces clutter in an autoplot() of a predictNMBsim object.\nUsage\n theme_sim()\nValue\n Returns a ggplot2 theme.\nExamples\n get_nmb <- function() c(\"TP\" = -3, \"TN\" = 0, \"FP\" = -1, \"FN\" = -4)\n sim_obj <- do_nmb_sim(\n sample_size = 200, n_sims = 50, n_valid = 10000, sim_auc = 0.7,\n event_rate = 0.1, fx_nmb_training = get_nmb, fx_nmb_evaluation = get_nmb\n )\n autoplot(sim_obj) + theme_sim()"}}},{"rowIdx":96,"cells":{"project":{"kind":"string","value":"iconr"},"source":{"kind":"string","value":"cran"},"language":{"kind":"string","value":"R"},"content":{"kind":"string","value":"Package ‘iconr’\n October 13, 2022\nTitle Graphical and Spatial Analysis for Prehistoric Iconography\nVersion 0.1.0\nDescription Set of formal methods for studying archaeological iconographic datasets (rock-art, pot-\n tery decoration, stelae, etc.) using network and spatial analysis (Alexan-\n der 2008 ; Huet 2018 ).\nLicense GPL-2\nEncoding UTF-8\nLazyData true\nImports igraph, magick, rgdal, grDevices, graphics, utils\nSuggests ggplot2, knitr, rmarkdown, dplyr, kableExtra, data.tree,\n dendextend\nVignetteBuilder knitr\nRoxygenNote 7.1.1\nURL https://zoometh.github.io/iconr/\nBugReports https://github.com/zoometh/iconr/issues\nNeedsCompilation no\nAuthor [aut, cre] (),\n [aut] (),\n [ctb] ()\nMaintainer <>\nRepository CRAN\nDate/Publication 2021-02-16 09:10:05 UTC\nR topics documented:\ncontemp_nd... 2\nlabels_shado... 3\nlist_compa... 4\nlist_de... 6\nnamed_element... 7\nplot_compa... 8\nplot_dec_grp... 10\nread_ed... 13\nread_nd... 14\nsame_element... 15\nside_plo... 16\n contemp_nds Select Contemporaneous Nodes\nDescription\n Find the connected component, or subgraph, of contemporaneous nodes (connected by normal and\n attribute edges) given a selected node and remove the other components\nUsage\n contemp_nds(nds.df, eds.df, selected.nd)\nArguments\n nds.df Dataframe of the nodes as the one obtained by the function read_nds.\n eds.df Dataframe of the edges as the one obtained by the function read_eds.\n selected.nd The node of the decoration graph for which to extract the connected component.\n It can be either the node order (numeric) or the node name/id (character).\nValue\n A named list of two dataframes: list(nodes, edges), collecting the contemporaneous nodes and\n edges, respectivelly.\nExamples\n # Set data folder\n dataDir <- system.file(\"extdata\", package = \"iconr\")\n # Read a decoration\n nds.df <- read_nds(site = \"Ibahernando\",\n decor = \"Ibahernando\",\n dir = dataDir)\n eds.df <- read_eds(site = \"Ibahernando\",\n decor = \"Ibahernando\",\n dir = dataDir)\n # Extract the subgraph contemporaneous to the node 2\n l_dec_df <- contemp_nds(nds.df, eds.df, selected.nd = 2)\n ## It returns a list of two dataframes, one for nodes and one for edges:\n l_dec_df\n labels_shadow Plot Labels with Contrasting Shadow\nDescription\n Plot labels (text) with a contrasting buffer to make them more visible when located on a similar\n color background. This function is the shadowtext() function developed by . Called\n by plot functions: plot_dec_grph, plot_compar\nUsage\n labels_shadow(x, y = NULL, labels,\n col = \"black\", bg = \"white\",\n theta = seq(0, 2 * pi, length.out = 50),\n r = 0.1,\n cex = 1, ...)\nArguments\n x, y Numeric vector of coordinates where the labels should be plotted. Alternatively,\n a single argument x can be provided with the same syntax as in xy.coords.\n labels Set of labels provided as a character vector.\n col, bg Graphical parameters for the label color and background (buffer) color.\n theta Angles for generating the buffer with possible anisotropy along one direction\n (default is isotropic) and controlling buffer smoothness (angular resolution).\n r Thickness of the buffer relative to the size of the used font, by default 0.1.\n cex Size of the label, by default 1.\n ... Further graphical parameter accepted by text, such as pos, offset, or family.\nValue\n No return value. It creates a contrasting buffer to make labels more visible.\nReferences\n https://rdrr.io/cran/TeachingDemos/man/shadowtext.html\n list_compar Graph Pairwise Comparison on Common Elements\nDescription\n nds_compar identifies common nodes in a pair of graphs.\n eds_compar identifies common edges in a pair of graphs.\n Given a list of graphs, list_compar extract all combinations of graph pairs and compare them on\n common elements (nodes and edges).\nUsage\n nds_compar(grphs, nd.var = \"type\")\n eds_compar(grphs, nd.var = \"type\")\n list_compar(lgrph, nd.var = \"type\",\n verbose = FALSE)\nArguments\n grphs A list of two graphs (pair of graphs) to be compared.\n lgrph A list of any number of graphs to be pairwise compared. The list can be typically\n obtained with the function list_dec\n nd.var An attribute of the graph nodes containing the node variable (ie, field) on which\n the comparison will be done. By default nd.var = \"type\".\n verbose Logical. If TRUE, the names of each graph pair combination are listed on the\n screen. By default verbose = FALSE.\nDetails\n list_compar() calls the functions: nds_compar() and eds_compar() which return respectively\n the common nodes and the common edges of a graph pairwise.\n Nodes are common when they have the same value for a given variable, for example horse, sword,\n etc., for the variable type (nd.var = \"type\").\n Edges are common when they have the same value for starting and ending nodes (horse, sword,\n etc.) and the same type of edge ('=', '+', etc.). For example, a -=- b in graph 1 is equal to a -=- b\n in graph 2, but not equal to a -+- b. Edges of type = (normal edges) are undirected, so that a -=- b\n is equal to b -=- a. But edges of types + (attribute edges) or > (diachronic edges) are directed, so:\n a ->- b is not equal to b ->- a.\n If any of the graphs has multiple nodes/edges with the same value, it is considered to count for as\n many coincidences as the smaller multiplicity. For instance, if there are 2 nodes with value epee in\n graph 1, and 3 nodes with value epee in graph 2, their number of common nodes is min(2, 3) = 2.\nValue\n nds_compar() returns the input pair of graphs, each complemented with a new node attribute\n named comm with value 1 for common nodes and 0 for non-common nodes.\n eds_compar() returns the input pair of graphs, each complemented with a new edge attribute named\n comm with value 1 for common edges and 0 for non-common edges.\n list_compar() returns a list of all combinations of graph pairs. For each pair, both graphs are\n complemented with the node attribute (comm) identifying common nodes and the edge attribute\n (comm) identifying common edges. Each pair is also complemented with an attribute named nd.var\n recording the compared node variable.\nSee Also\n list_dec, plot_compar, same_elements\nExamples\n # Read data\n imgs <- read.table(system.file(\"extdata\", \"imgs.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n nodes <- read.table(system.file(\"extdata\", \"nodes.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n edges <- read.table(system.file(\"extdata\", \"edges.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n # Generate list of graphs from the three data.frames\n lgrph <- list_dec(imgs, nodes, edges)\n # Generate list of all graph comparisons depending on the node \"type\" variable\n g.compar <- list_compar(lgrph, nd.var = \"type\")\n length(g.compar)\n ## Ten pairwise comparisons\n # Inspect the second pairwise comparison of the list\n g.compar[[2]]\n ## The two compared graphs with the name of the comparison variable\n # Inspecting nodes:\n igraph::as_data_frame(g.compar[[2]][[1]], \"vertices\")\n ## Vertices from the first decoration graph\n igraph::as_data_frame(g.compar[[2]][[2]], \"vertices\")\n ## Vertices from the second decoration graph\n # Inspecting edges:\n igraph::as_data_frame(g.compar[[2]][[1]])\n ## Edges of the first decoration graph\n igraph::as_data_frame(g.compar[[2]][[2]])\n ## Edges of the second decoration graph\n list_dec Create Decoration’s Graphs and Store them in a List\nDescription\n Create undirected graphs for each decoration from nodes, edges and imgs dataframes and store the\n graphs in a list. The join between these dataframes is done on the two fields site and decor. Graph\n names refer to imgs$idf.\nUsage\n list_dec(imgs,\n nodes,\n edges)\nArguments\n imgs Dataframe of decorations\n nodes Dataframe of nodes\n edges Dataframe of edges\nValue\n A list of igraph graphs.\nSee Also\n graph_from_data_frame\nExamples\n # Read imgs, nodes and edges dataframes\n imgs <- read.table(system.file(\"extdata\", \"imgs.csv\", package = \"iconr\"),\n sep=\";\", stringsAsFactors = FALSE)\n nodes <- read.table(system.file(\"extdata\", \"nodes.csv\", package = \"iconr\"),\n sep=\";\", stringsAsFactors = FALSE)\n edges <- read.table(system.file(\"extdata\", \"edges.csv\", package = \"iconr\"),\n sep=\";\", stringsAsFactors = FALSE)\n # Create the list of graphs\n lgrph <- list_dec(imgs, nodes, edges)\n # Get the first graph\n g <- lgrph[[1]]\n g\n # Graph name\n g$name\n # Graph label\n g$lbl\n # Graph number of nodes\n igraph::gorder(g)\n # Graph number of edges\n igraph::gsize(g)\n named_elements Textual Notation of Graph Elements\nDescription\n Create a textual notation for nodes or edges.\nUsage\n named_elements(grph,\n focus = \"edges\",\n nd.var = \"type\",\n disamb.marker = \"#\")\nArguments\n grph A decoration graph (object of class igraph).\n focus Textual notation of edges (focus = \"edges\") or nodes (focus = \"nodes\"). By\n default focus = \"edges\".\n nd.var The attribute of the graph nodes containing the node variable (ie, field) for the\n textual annotation. By default nd.var = \"type\".\n disamb.marker Marker used to disambiguated repeated elements. By default disamb.marker =\n \"#\".\nDetails\n Edges of type '=' (normal edges) are undirected, so that the order of their nodes is irrelevant and\n they are presented in alphabetical order. Conversely, edges of types '+' (attribute edges) and '>'\n (diachronic edges) are directed, so that the given order of nodes is preserved.\n Repeated node or edge names are disambiguated by appending the symbol disamb.marker ('#'\n by default) at the end of the second appearance (suffix). Subsequent appearances are marked by\n additional disamb.markers.\nValue\n A character vector of named nodes or edges.\nSee Also\n list_compar, same_elements\nExamples\n # Read data\n imgs <- read.table(system.file(\"extdata\", \"imgs.tsv\", package = \"iconr\"),\n sep=\"\\t\", stringsAsFactors = FALSE)\n nodes <- read.table(system.file(\"extdata\", \"nodes.tsv\", package = \"iconr\"),\n sep=\"\\t\", stringsAsFactors = FALSE)\n edges <- read.table(system.file(\"extdata\", \"edges.tsv\", package = \"iconr\"),\n sep=\"\\t\", stringsAsFactors = FALSE)\n # Generate list of graphs from the three data.frames\n lgrph <- list_dec(imgs, nodes, edges)\n # Textual notation of disambiguated edges\n named_elements(lgrph[[2]], focus = \"edges\", nd.var=\"type\")\n # Textual notation of disambiguated nodes\n named_elements(lgrph[[2]], focus = \"nodes\", nd.var=\"type\")\n plot_compar Plot and Save Comparison Figures Between Pairs of Graphs\nDescription\n Given a list of pairwise graph comparisons, the function plots any given subset selected by graph\n name, displaying side-by-side pairs of graphs and highlighting common nodes or common edges\n with a choice of several graphical parameters.\nUsage\n plot_compar(listg, dec2comp = NULL, focus = \"nodes\",\n dir = getwd(),\n nd.color = c(\"orange\", \"red\"), nd.size = c(0.5, 1),\n ed.color = c(\"orange\", \"red\"), ed.width = c(1, 2),\n lbl.size = 0.5,\n dir.out = dir, out.file.name = NULL,\n img.format = NULL, res = 300)\nArguments\n listg A list of graph pairwise comparisons as returned by list_compar.\n dec2comp A vector with the names of the graphs for which comparisons are to be plotted.\n The user can select to plot all pairwise combinations (by default), all combina-\n tions of a subset, or a single pair.\n focus Either \"nodes\" (default) or \"edges\". It selects the type of comparison to be\n plotted, highlighting common nodes or common edges, respectively.\n dir Data folder including the decoration images. By default the working directory.\n nd.color, nd.size, ed.color, ed.width\n Graphical parameters for color and size/widths of nodes and edges. Each of\n them is a vector with two values for different and common nodes/edges, re-\n spectively. If only one value is provided, this unique value is taken for both\n different and common elements. Labels are displayed with the same color as\n common nodes. For focus = \"nodes\" all edges are plotted with the first value\n of ed.color and ed.width.\n lbl.size Graphical parameter for the size of the labels with the node names. The default\n is 0.5.\n dir.out Folder for the output image. By default, it coincides with the input dir.\n out.file.name Name of the output image, including path from current directory and extension.\n By default the name is automatically generated including site, decor, nd.var,\n and the extension from img.format.\n If set, out.file.name overrides dir.out and img.format.\n img.format, res\n Format and resolution of the saved images. The handled formats are \"png\",\n \"bmp\", \"tiff\"/\"tif\", \"jpeg\"/\"jpg\", and \"pdf\". The default resolution is 300\n (ppi). The resolution does not applies to format pdf.\n if img.format=NULL (default), the plot is sent to the active device.\nDetails\n To highlight common elements between a list of graphs, the user can focus on nodes (focus =\n \"nodes\") or edges (focus = \"edges\"). As stated in the function list_compar, for a given com-\n parison variable (eg. nd.var=\"type\") if there is multiple nodes/edges with the same value, it is\n considered to count for as many coincidences as the smaller multiplicity.\n img.format=NULL (plot to the active device) does not make sense for more than one comparison.\nValue\n Generates graph decoration images, for pairwise comparisons between two or more decorations,\n comparing graphs elements (nodes or edges).\n If img.format=NULL, the plot is sent to the active device and no value is returned.\n If img.format= \"png\" or \"bmp\" or \"tiff\"/\"tif\" or \"jpeg\"/\"jpg\" or \"pdf\", the return value is a\n character vector with the dir/name of every saved image in the indicated format.\nSee Also\n list_compar plot_dec_grph\nExamples\n # Read data\n imgs <- read.table(system.file(\"extdata\", \"imgs.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n nodes <- read.table(system.file(\"extdata\", \"nodes.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n edges <- read.table(system.file(\"extdata\", \"edges.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n # Generate list of graphs from the three dataframes\n lgrph <- list_dec(imgs, nodes, edges)\n # Generate all pairwise comparisons of the graphs with respect to nodes \"type\"\n g.compar <- list_compar(lgrph, nd.var=\"type\")\n # Generate the image showing the comparison on common nodes of graphs\n # '1' and '4', save it in png format, and return its path.\n dataDir <- system.file(\"extdata\", package = \"iconr\")\n outDir <- tempdir()\n plot_compar(g.compar, c(1,4), focus = \"nodes\",\n dir = dataDir,\n dir.out = outDir,\n img.format = \"png\")\n # Generate the image showing the comparison on common edges of all pairwise\n # combinations of graphs '1','3', and '4', save them in pdf format, and return\n # their path.\n # Plot nodes involved in non-common edges in orange and\n # nodes involved in common edges and the corresponding labels in brown.\n plot_compar(g.compar, c(1, 3, 4), focus = \"edges\",\n dir = dataDir,\n nd.color = c(\"orange\", \"brown\"),\n dir.out = outDir,\n img.format = \"pdf\")\n # Save the png image showing the comparison on common nodes of graphs\n # '1' and '4'.\n # Then read and plot the image.\n img.filename <- plot_compar(g.compar, c(1, 4), focus = \"nodes\",\n dir = dataDir,\n dir.out = outDir,\n img.format = \"png\")\n plot(magick::image_read(img.filename))\n # Plot directly on the active device (default) the comparison on common nodes\n # of graphs '1' and '4'.\n plot_compar(g.compar, c(1, 4), focus = \"nodes\",\n dir = dataDir)\n plot_dec_grph Plot a Graph on a Decoration\nDescription\n Plot with nodes only, edges only, or both (geometric graph) over a decoration image.\nUsage\n plot_dec_grph(nodes = NULL,\n edges = NULL,\n imgs,\n site,\n decor,\n dir = getwd(),\n nd.var = 'id',\n nd.color = 'orange',\n nd.size = 0.5,\n lbl.color = 'black',\n lbl.size = 0.5,\n ed.color = c(\"orange\", \"blue\"),\n ed.lwd = 1,\n dir.out = dir,\n out.file.name = NULL,\n img.format = NULL,\n res = 300)\nArguments\n nodes Dataframe of nodes\n edges Dataframe of edges\n imgs Dataframe of decorations\n site Name of the site\n decor Name of the decoration\n dir Data folder including the decoration images. By default the working directory.\n nd.var Field name in the nodes data frame to be displayed as node labels. By default\n the identifier nodes$id.\n nd.color, nd.size, lbl.color, lbl.size, ed.color, ed.lwd\n Graphical parameters for color and size/widths of nodes, edges, and labels.\n ed.color is a vector with two values (the second value is used for diachronic\n edges).\n dir.out Folder for the output image. By default, it coincides with the input dir.\n out.file.name Name of the output image, including path from current directory and extension.\n By default the name is automatically generated including site, decor, nd.var,\n and the extension from img.format.\n If set, out.file.name overrides dir.out and img.format.\n img.format, res\n Format and resolution of the saved images. The handled formats are \"png\",\n \"bmp\", \"tiff\"/\"tif\", \"jpeg\"/\"jpg\", and \"pdf\". The default resolution is 300\n (ppi). The resolution does not applies to format pdf.\n if img.format=NULL (default), the plot is sent to the active device.\nDetails\n Plot nodes only (if edges = NULL), edges only (if nodes = NULL), or both (graph) over a decoration\n image.\nValue\n Generates graph decoration images with nodes, edges, or both, overlapping the decoration image.\n If img.format=NULL, the plot is sent to the active device and no value is returned.\n If img.format= \"png\" or \"bmp\" or \"tiff\"/\"tif\" or \"jpeg\"/\"jpg\" or \"pdf\", the return value is a\n character vector with the dir/name of the saved image in the indicated format.\nExamples\n ## Set data folder\n dataDir <- system.file(\"extdata\", package = \"iconr\")\n ## Decoration to be plotted\n site <- \"Brozas\"\n decor <- \"Brozas\"\n ## Read nodes, edges, and decorations\n nds.df <- read_nds(site, decor, dataDir)\n eds.df <- read_eds(site, decor, dataDir)\n imgs <- read.table(paste0(dataDir, \"/imgs.tsv\"),\n sep=\"\\t\", stringsAsFactors = FALSE)\n ## Plot 'Brozas' nodes and edges on the active device\n ## with node variable \"type\" as labels\n plot_dec_grph(nds.df, eds.df, imgs,\n site, decor,\n dir = dataDir,\n lbl.size = 0.4,\n nd.var = \"type\")\n ## Save only edges of 'Brozas' with bigger widths and in image format jpg.\n outDir <- tempdir()\n img.filename <- plot_dec_grph(nodes = NULL, eds.df, imgs,\n site, decor,\n dir = dataDir,\n ed.lwd = 2,\n dir.out = outDir,\n img.format = \"jpg\")\n ## Then read and plot the image.\n a.dec <- magick::image_read(img.filename)\n ## Inspect the output image\n magick::image_info(a.dec)\n ## Plot the output image\n plot(a.dec)\n read_eds Read Edges of a Decoration\nDescription\n Read edges’ information from a file including all edges and extract edges of one decoration. Ac-\n cepted formats are tab separated values (’tsv’), semicolon separated values (’csv’), or shapefile\n (’shp’).\nUsage\n read_eds(site,\n decor,\n dir = getwd(),\n edges = \"edges\",\n nodes = \"nodes\",\n format = \"tsv\")\nArguments\n site Name of the site.\n decor Name of the decoration.\n dir Path to the working folder, by default it is the working directory.\n edges Name of the edges file (a dataframe or a shapefile).\n nodes Name of the nodes file (a dataframe or a shapefile).\n format File extension indicating a file format from ’tsv’ (tab separated values), ’csv’\n (semicolon separated values) or ’shp’ (shapefile). For ’tsv’ and ’csv’ the coordi-\n nates of the edges will be calculated from the same decoration’s node dataframe.\nDetails\n Subset the dataframe of edges depending on ’site’ and ’decor’.\nValue\n Dataframe of graph edges, including at least the columns \"site\", \"decor\", \"a\", \"b\", \"xa\", \"ya\", \"xb\",\n \"yb\", with values for each edge (row).\nExamples\n # Set data folder\n dataDir <- system.file(\"extdata\", package = \"iconr\")\n # Read .tsv file\n eds.df <- read_eds(site = \"Cerro Muriano\", decor = \"Cerro Muriano 1\",\n dir = dataDir, edges = \"edges\", format = \"tsv\")\n eds.df\n ## Dataframe of edges\n # Read shapefile\n eds.df <- read_eds(site = \"Cerro Muriano\", decor = \"Cerro Muriano 1\",\n dir = dataDir, edges = \"edges\", format = \"shp\")\n eds.df\n ## Dataframe of edges\n read_nds Read Nodes of a Decoration\nDescription\n Read nodes’ information from a file including all nodes and extract nodes of one decoration. Ac-\n cepted formats are tab separated values (’tsv’), semicolon separated values (’csv’), or shapefile\n (’shp’).\nUsage\n read_nds(site,\n decor,\n dir = getwd(),\n nodes = \"nodes\",\n format = \"tsv\")\nArguments\n site Name of the site\n decor Name of the decoration\n dir Path to the working folder, by default it is the working directory\n nodes Name of the nodes file (a dataframe or a shapefile)\n format File extension indicating a file format from ’tsv’ (tab separated values), ’csv’\n (semicolon separated values) or ’shp’ (shapefile). For ’tsv’ and ’csv’ the files\n must include node coordinates (nodes$x, nodes$y).\nValue\n Dataframe of graph nodes, including at least the columns \"site\", \"decor\", \"id\", \"x\", \"y\", with values\n for each node (row).\nExamples\n # Set data folder\n dataDir <- system.file(\"extdata\", package = \"iconr\")\n # Read dataframe of nodes\n nds.df <- read_nds(site = \"Cerro Muriano\", decor = \"Cerro Muriano 1\",\n dir = dataDir, format = \"tsv\")\n nds.df\n ## Dataframe of nodes\n # Read shapefile of nodes\n nds.df <- read_nds(site = \"Cerro Muriano\", decor = \"Cerro Muriano 1\",\n dir = dataDir, format = \"shp\")\n nds.df\n ## Dataframe of nodes\n same_elements Number of Equal Elements Between Each Decoration Pair\nDescription\n Create the (symmetric) dataframe with the count of common nodes or common edges (see list_compar\n for comparison criteria) for each pair of decorations (graphs) from a list. The diagonal of the sym-\n metric dataframe is filled with counts of nodes/edges for each decoration.\nUsage\n same_elements(lgrph, nd.var = \"type\",\n focus = \"nodes\")\nArguments\n lgrph A list of any number of graphs to be pairwise compared. The list can be typically\n obtained with the function list_dec\n nd.var An attribute of the graph vertices containing the node variable (ie, field) on\n which the comparison will be done. By default nd.var = \"type\".\n focus Either \"nodes\" (default) or \"edges\" to select the type of elements to be com-\n pared for the count.\nValue\n A symmetric matrix with the counts of the pairwise coincidences of nodes or edges. The matrix has\n as row and column names the names of the corresponding graphs in the input list.\nSee Also\n list_dec, list_compar, plot_compar\nExamples\n # read imgs, nodes and edges dataframes\n imgs <- read.table(system.file(\"extdata\", \"imgs.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n nodes <- read.table(system.file(\"extdata\", \"nodes.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n edges <- read.table(system.file(\"extdata\", \"edges.tsv\", package = \"iconr\"),\n sep=\"\\t\",stringsAsFactors = FALSE)\n lgrph <- list_dec(imgs,nodes,edges)\n # Counting same nodes\n df.same_nodes <- same_elements(lgrph, nd.var = \"type\",\n focus = \"nodes\")\n df.same_nodes\n ## a symmetric matrix of nodes comparisons\n # same edges\n df.same_edges <- same_elements(lgrph, nd.var = \"type\",\n focus = \"edges\")\n df.same_edges\n ## a symmetric matrix of edges comparisons\n side_plot Plot Two Figures Side-by-Side Identifying Common Elements\nDescription\n Plot two decoration graphs side-by-side identifying common nodes and common edges. This func-\n tion is called by the function plot_compar.\nUsage\n side_plot(grph, dir, nd.var, focus = \"nodes\",\n nd.color = c(\"orange\", \"red\"),\n nd.size = c(0.5, 1),\n ed.color = c(\"orange\", \"red\"),\n ed.width = c(1, 2),\n lbl.size = 0.5)\nArguments\n grph List of two or more ’igraph’ graphs created with the list_compar function.\n dir Working directory which contains the imgs, nodes, edges dataframes and the\n decoration images.\n nd.var Field of nodes on which the comparison will be done.\n focus Focus on nodes or on edges, by default focus = \"nodes\".\n nd.color, nd.size, ed.color, ed.width\n Graphical parameters for the nodes and edges. The different nodes/edges will\n be displayed with the first values of the vectors (eg, \"orange\") while the common\n nodes/edges will be displayed with the second values of the vectors (eg, \"red\").\n lbl.size Size of the labels\nValue\n No return value, group images side-by-side\nSee Also\n plot_compar"}}},{"rowIdx":97,"cells":{"project":{"kind":"string","value":"smartsdk-recipes"},"source":{"kind":"string","value":"readthedoc"},"language":{"kind":"string","value":"JSON"},"content":{"kind":"string","value":"This page explains some mechanisms that are common for the management of services deployed with the recipes.\n\n## Multisite Deployment\n\nIf you happen to have your cluster with nodes distributed in specific areas (A1, A2, ..., AN) and you would like for the deployment of the replicas of your service S to happen only in specific areas, you can achieve this by using placement constraints.\n\nFor example, imagine you have a cluster composed of VMS in Málaga, Madrid and Zurich, but you want the deployment of the replicas of your database to stay only within Spain boundaries due to legal regulations. (DISCLAIMER: This is just a simplification, you should always inform yourself on how to properly comply with data protection regulations.)\n\nFirst, you need define a labeling for the nodes of your cluster where you want your deployment to happen. Your label in this case can be done by region. You connect to any of the swarm manager nodes and execute the following commands.\n\n```\ndocker node update --label-add region=ES malaga-0\ndocker node update --label-add region=ES madrid-0\ndocker node update --label-add region=ES madrid-1\n\ndocker node update --label-add region=CH zurich-0\ndocker node update --label-add region=CH zurich-1\n```\n\n When you are about to deploy your database, you will have to add a constraint to the definition of the service. This means you will have to edit the recipe before deploying. For instance, in the case of MongoDB, it should have the `deploy` part looking something like this: \n\n```\nmongo:\n ...\n deploy:\n placement:\n constraints:\n - region == ES\n```\n\nThis was just a simple example. You can have multiple tags, combine them so the placement only happens in nodes with all the tags, and many other combinations. For more details of this functionality, please refer to the official docker docs on service placement.\n\nNote: For these features to work, you need access to a manager node of the swarm cluster and Docker 17.04+.\n\n## Scalability\n\nAs you probably already know, each recipe deploys a bunch of services, and each service can be of any of two types: Stateless or Stateful. The way to differentiate which type a service is, is by inspecting the implementation of the service to tell if it needs data persistence within itself or not.\n\nThe point is, the way to scale services depends on the type of the service.\n\nScaling stateless services with Docker is pretty straightforward, you can simply increase the number of replicas. Assuming no constraints violations (see previous section), you will be able to dynamically set more or less replicas for each stateless service.\n\n```\ndocker service scale orion=5\n```\n\nMore info on the scaling process is documented here.\n\nAs regards scaling stateful services, there is no silver bullet and it will always depend on the service being discussed.\n\nFor example, Docker handles two types of service deployments: replicated and global, as explained here. A replicated service can be scaled as shown in the previous example, but the only way to scale a global service (which means there will be a maximum of one instance per node), is by adding nodes. To scale down a global service you can either remove nodes or apply constraints (see Multisite Deployment section above).\n\nIn any of the two cases, for a stateful service, someone will have to be responsible for coordinating the data layer among all the instances and deal with replication, partitioning and all sort of issues typically seen in distributed systems.\n\nFinally, the recipes should properly document which of its services are stateless and which are not, so as to have these considerations for Scalability. They should include also, notes on how their stateful services could be scaled.\n\nHere you can find recipes aimed at different usages of the Orion Context Broker. We assume you are already familiar with Orion. If not, refer to the official documentation.\n\nThe easiest and simplest way to try Orion is as explained in Orion's official docker image docs using this docker-compose file. But here, we will explore a distributed configuration for this Generic Enabler.\n\nInstructions on how to prepare your environment to test these recipes are given in https://github.com/smartsdk/smartsdk-recipes.\n\nDate: 2016-01-01\nCategories: \nTags: \n\nThis recipe shows how to deploy an scalable Orion Context Broker service backed with an scalable replica set of MongoDB instances.\n\n Orion needs a mongo database for its backend. If you have already deployed Mongo within your cluster and would like to reuse that database, you can skip the next step (deploying backend). You will just need to pay attention to the variables you define for Orion to link to Mongo, namely, `MONGO_SERVICE_URI` . Make sure\nyou have the correct values in `settings.env` (or `settings.bat` in Windows).\nThe value of `MONGO_SERVICE_URI` should be a routable address for mongo.\nIf deployed within the swarm, the service name (with stack prefix)\nwould suffice. You can read more in the\nofficial docker docs.\nThe default values should be fine for you if you used the\nMongo ReplicaSet Recipe. \nNow you can activate your settings and deploy Orion...\n\n```\n$ source settings.env # In Windows, simply execute settings.bat instead.\n$ docker stack deploy -c docker-compose.yml orion\n```\n\nAt some point, your deployment should look like this...\n\n```\n$ docker service ls\nID NAME MODE REPLICAS IMAGE\nnrxbm6k0a2yn mongo-rs_mongo global 3/3 mongo:3.2\nrgws8vumqye2 mongo-rs_mongo-controller replicated 1/1 smartsdk/mongo-rs-controller-swarm:latest\nzk7nu592vsde orion_orion replicated 3/3 fiware/orion:1.3.0\n```\n\n As shown above, if you see `3/3` in the replicas column it means the 3 replicas\nare up and running. \n\nYou can check the distribution of the containers of a service (a.k.a tasks) through the swarm running the following...\n\n```\n$ docker service ps orion_orion\nID NAME IMAGE NODE DESIRED STATE CURRENT STATE ERROR PORTS\nwwgt3q6nqqg3 orion_orion.1 fiware/orion:1.3.0 ms-worker0 Running Running 9 minutes ago \nl1wavgqra8ry orion_orion.2 fiware/orion:1.3.0 ms-worker1 Running Running 9 minutes ago \nz20v0pnym8ky orion_orion.3 fiware/orion:1.3.0 ms-manager0 Running Running 25 minutes ago\n```\n\n The good news is that, as you can see from the above output, by default docker already took care of deploying all the replicas of the service `context-broker_orion` to different hosts. Of course, with the use of labels, constraints or deploying mode you have the power to customize the distribution of tasks among swarm nodes. You can see the mongo replica recipe to understand the deployment of the `mongo-replica_mongo` service. \nNow, let's query Orion to check it's truly up and running. The question now is... where is Orion actually running? We'll cover the network internals later, but for now let's query the manager node...\n\n```\n$ sh ../query.sh $(docker-machine ip ms-manager0)\n```\n\nYou will get something like...\n\n```\n{\n \"orion\" : {\n \"version\" : \"1.3.0\",\n \"uptime\" : \"0 d, 0 h, 18 m, 13 s\",\n \"git_hash\" : \"cb6813f044607bc01895296223a27e4466ab0913\",\n \"compile_time\" : \"Fri Sep 2 08:19:12 UTC 2016\",\n \"compiled_by\" : \"root\",\n \"compiled_in\" : \"ba19f7d3be65\"\n}\n}\n[]\n```\n\n Thanks to the docker swarm internal routing mesh, you can actually perform the previous query to any node of the swarm, it will be redirected to a node where the request on port `1026` can be attended (i.e, any node running Orion). \nLet's insert some data...\n\n```\n$ sh ../insert.sh $(docker-machine ip ms-worker1)\n```\n\nAnd check it's there...\n\n```\n$ sh ../query.sh $(docker-machine ip ms-worker0)\n...\n[\n {\n \"id\": \"Room1\",\n \"pressure\": {\n \"metadata\": {},\n \"type\": \"Integer\",\n \"value\": 720\n },\n \"temperature\": {\n \"metadata\": {},\n \"type\": \"Float\",\n \"value\": 23\n },\n \"type\": \"Room\"\n }\n]\n```\n\nYes, you can query any of the three nodes.\n\nSwarm's internal load balancer will be load-balancing in a round-robin approach all the requests for an orion service among the orion tasks running in the swarm.\n\n## Rescaling Orion\n\nScaling up and down orion is a simple as runnnig something like...\n\n```\n$ docker service scale orion_orion=2\n```\n\n (this maps to the `replicas` argument in the docker-compose) \nConsequently, one of the nodes (ms-worker1 in my case) is no longer running Orion...\n\n```\n$ docker service ps orion_orion\nID NAME IMAGE NODE DESIRED STATE CURRENT STATE ERROR PORTS\n2tibpye24o5q orion_orion.2 fiware/orion:1.3.0 ms-manager0 Running Running 11 minutes ago \nw9zmn8pp61ql orion_orion.3 fiware/orion:1.3.0 ms-worker0 Running Running 11 minutes ago\n```\n\nBut still responds to the querying as mentioned above...\n\n```\n$ sh ../query.sh $(docker-machine ip ms-worker1)\n{\n \"orion\" : {\n \"version\" : \"1.3.0\",\n \"uptime\" : \"0 d, 0 h, 14 m, 30 s\",\n \"git_hash\" : \"cb6813f044607bc01895296223a27e4466ab0913\",\n \"compile_time\" : \"Fri Sep 2 08:19:12 UTC 2016\",\n \"compiled_by\" : \"root\",\n \"compiled_in\" : \"ba19f7d3be65\"\n}\n}\n[]\n```\n\nYou can see the mongo replica recipe to see how to scale the mongodb backend. But basically, due to the fact that it's a \"global\" service, you can scale it down like shown before. However, scaling it up might require adding a new node to the swarm because there can be only one instance per node.\n\n## Dealing with failures\n\nDocker is taking care of the reconciliation of the services in case a container goes down. Let's show this by running the following (always on the manager node):\n\nSuppose orion container goes down...\n\n```\n$ docker rm -f abc5e37037f0\n```\n\nYou will see it gone, but after a while it will automatically come back.\n\nEven if a whole node goes down, the service will remain working because you had both redundant orion instances and redundant db replicas.\n\n```\n$ docker-machine rm ms-worker0\n```\n\nYou will still get replies to...\n\n```\n$ sh ../query.sh $(docker-machine ip ms-manager0)\n$ sh ../query.sh $(docker-machine ip ms-worker1)\n```\n\nThis recipe shows how to deploy an scalable API Umbrella instance service backed with an scalable replica set of MongoDB instances.\n\nAt the time being, other services, such as Elastic Search for logging api interactions and QoS are not deployed. This is mostly due to the fact that API Umbrella supports only obsolete versions of Elastic Search (i.e. version 2, while current version is 6).\n\n In case you haven't done it yet for other recipes, deploy `backend` and `frontend` networks as described in the\ninstallation guide. API Umbrella needs a mongo database for its backend. If you have already deployed Mongo within your cluster and would like to reuse that database, you can skip the next step (deploying backend). You will just need to pay attention to the variables you define for API Umbrella to link to Mongo, namely, `MONGO_SERVICE_URI` and `REPLICASET_NAME` . \nOtherwise, if you prefer to make a new deployment of MongoDB just for API Umbrella, you can take a shortcut and run...\n\n```\n$ sh deploy_back.sh\n\nCreating config mongo-rs_mongo-healthcheck\nCreating service mongo-rs_mongo\nCreating service mongo-rs_controller\n```\n\nBeside that, given that Ruby driver for MongoDB is not supporting service discovery, you will need to expose the ports of the MongoDB server on the cluster to allow the connection to the Replica Set from API Umbrella.\n\nBe aware that this works only when you deploy (as in the script), your MongoDB in global mode.\n\n```\n$ docker service update --publish-add published=27017,target=27017,protocol=tcp,mode=host mongo-rs_mongo\n\nmongo-rs_mongo\noverall progress: 1 out of 1 tasks\nw697ke0djs3c: running [==================================================>]\nverify: Service converged\n```\n\nWait some time until the backend is ready, you can check the backed deployment running:\n\n```\n$ docker stack ps mongo-rs\nID NAME IMAGE NODE DESIRED STATE CURRENT STATE ERROR PORTS\nmxxrlexvj0r9 mongo-rs_mongo.z69rvapjce827l69b6zehceal mongo:3.2 ms-worker1 Running Starting 9 seconds ago \nd74orl0f0q7a mongo-rs_mongo.fw2ajm8zw4f12ut3sgffgdwsl mongo:3.2 ms-worker0 Running Starting 15 seconds ago \na2wddzw2g2fg mongo-rs_mongo.w697ke0djs3cfdf3bgbrcblam mongo:3.2 ms-manager0 Running Starting 6 seconds ago \nnero0vahaa8h mongo-rs_controller.1 smartsdk/mongo-rs-controller-swarm:latest ms-manager0 Running Running 5 seconds ago\n```\n\n Set the connection url for mongo based on the IPs of your Swarm Cluster (alternatively edit the `frontend.env` file): \n\n```\n$ MONGO_REPLICATE_SET_IPS=192.168.99.100:27017,192.168.99.101:27017,192.168.99.102:27017\n$ export MONGO_REPLICATE_SET_IPS\n```\n\n If you used `miniswarm` to create your cluster, you can get the different IPs\nusing the `docker-machine ip` command, e.g.: \n\n$ docker-machine ip ms-worker0\n\n$ docker-machine ip ms-worker1\n```\n\nwhen all services will be in status ready, your backend is ready to be used:\n\n```\n$ sh deploy_front.sh\n\ngenerating config file\nreplacing target file api-umbrella.yml\nreplace mongodb with mongo-rs_mongo\nreplacing target file api-umbrella.yml\nreplace rs_name with rs\nCreating config api_api-umbrella\nCreating service api_api-umbrella\n```\n\nWhen also the frontend services will be running, your deployment will look like this:\n\n```\n$ docker service ls\n\nID NAME MODE REPLICAS IMAGE PORTS\nca11lmx40tu5 api_api-umbrella replicated 2/2 smartsdk/api-umbrella:0.14.4-1-fiware *:80->80/tcp,*:443->443/tcp\nte1i0vhwtmnw mongo-rs_controller replicated 1/1 smartsdk/mongo-rs-controller-swarm:latest \nrbo2oe2y0d72 mongo-rs_mongo global 3/3 mongo:3.2\n```\n\n If you see `3/3` in the replicas column it means the 3 out of 3 planned replicas\nare up and running. \n\nIn the following walkthrough we will explain how to do the initial configuration of API Umbrella and register your first API. For more details read API Umbrella's documentation.\n\n* Let's create the admin user in API Umbrella. As first thing, get the IP of your master node:\n\n Open the browser at the following endpoint:\n\n```\nhttp:///admin\n```\n\n. \nUnless you also created certificates for your server, API Umbrella will ask you to accept the connection to an insecure instance.\n\nIn the page displayed you can enter the admin user name and the password.\n\nNow you are logged in and you can configure the backend APIs.\n\nN.B.: The usage of the cluster master IP is just a convention, you can reach the services also at the IPs of the worker nodes.\n\nRetrieve\n `X-Admin-Auth-Token` Access and `X-Api-Key` . In the menu select \n\n```\nUsers->Admin Accounts\n```\n\nand click on the username you just created. Copy the `Admin API Access` for your account. \nIn the menu select\n `Users->Api Users` click on the username \n\n```\n@internal.apiumbrella\n```\n\nand copy the API Key (of course you can create new ones instead of reusing API Umbrella defaults). * \n\nRegister an new API. Create a simple API to test that everything works:\n ```bash $ curl -k -X POST \"https://} } EOF \nResponse: { \"api\": { \"backend_host\": \"maps.googleapis.com\", \"backend_protocol\": \"http\", \"balance_algorithm\": \"least_conn\", \"created_at\": \"2018-02-26T13:47:02Z\", \"created_by\": \"c9d7c2cf-737c-46ae-974b-22ebc12cce0c\", \"deleted_at\": null, \"frontend_host\": \"\n\n* Publish the newly registered API.\n\n```bash $ curl -k -X POST \"https://\n\nResponse:\n\n{ \"config_version\": { \"config\": { \"apis\": [ { \"_id\": \"cbe24047-7f74-4eb5-bd7e-211c3f8ede22\", \"version\": 2, \"deleted_at\": null, \"name\": \"distance FIWARE REST\", \"sort_order\": 100000, \"backend_protocol\": \"http\", \"frontend_host\": \"192.168.99.100\", \"backend_host\": \"maps.googleapis.com\", \"balance_algorithm\": \"least_conn\", \"updated_by\": \"c9d7c2cf-737c-46ae-974b-22ebc12cce0c\", \"updated_at\": \"2018-02-26T14:02:08Z\", \"created_at\": \"2018-02-26T13:47:02Z\", \"created_by\": \"c9d7c2cf-737c-46ae-974b-22ebc12cce0c\", \"settings\": { \"require_https\": \"required_return_error\", \"disable_api_key\": false, \"api_key_verification_level\": \"none\", \"require_idp\": \"fiware-oauth2\", \"rate_limit_mode\": \"unlimited\", \"error_templates\": {}, \"_id\": \"4dfe22af-c12a-4733-807d-0a668c413a96\", \"anonymous_rate_limit_behavior\": \"ip_fallback\", \"authenticated_rate_limit_behavior\": \"all\", \"error_data\": {} }, \"servers\": [ { \"host\": \"maps.googleapis.com\", \"port\": 80, \"_id\": \"f0f7a039-d88c-4ef8-8798-a00ad3c8fcdb\" } ], \"url_matches\": [ { \"frontend_prefix\": \"/distance2/\", \"backend_prefix\": \"/\", \"_id\": \"ec719b9f-2020-4eb9-8744-5cb2bae4b625\" } ] } ], \"website_backends\": [] }, \"created_at\": \"2018-02-26T14:03:53Z\", \"updated_at\": \"2018-02-26T14:03:53Z\", \"version\": \"2018-02-26T14:03:53Z\", \"id\": { \"$oid\": \"5a9413c99f9d04008c5a0b6c\" } } } ```\n\nTest your new API, by issuing a query:\n\n* Get a token from FIWARE:\n\n```bash $ wget --no-check-certificate https://raw.githubusercontent.com/fgalan/oauth2-example-orion-client/master/token_script.sh $ bash token_script.sh\n\nUsername: Password: Token:\n b'```'\n\n* Use it to make a query to your API:\n\n```bash $ curl -k \"https://\n b'/distance2/maps/api/distancematrix/json?units=imperial&origins=Washington,DC&destinations=New+York+City,NY&token='\n b'\"'\n\nResponse: { \"destination_addresses\" : [ \"New York, NY, USA\" ], \"origin_addresses\" : [ \"Washington, DC, USA\" ], \"rows\" : [ { \"elements\" : [ { \"distance\" : { \"text\" : \"225 mi\", \"value\" : 361940 }, \"duration\" : { \"text\" : \"3 hours 50 mins\", \"value\" : 13816 }, \"status\" : \"OK\" } ] } ], \"status\" : \"OK\" } ```\n\nThis section contains useful (and sometimes temporary) scripts as well as references to tools, projects and pieces of documentation used for the development of the recipes.\n\nThe basic environment setup is explained in the Installation part of the docs.\n\n## Playing with Recipes?\n\n### miniswarm\n\nHelpful tool to help you quickly setup a local virtualbox-based swarm cluster for testing purposes.\n\n### wait-for-it\n\nUseful shell script used when you need to wait for a service to be started.\n\nNote: This might no longer be needed since docker introduced the healthchecks feature.\n\n### docker-swarm-visualizer\n\n If you'd like to have a basic view of the distribution of containers in your swarm cluster, you can use the `visualzer.yml` file provided in this folder. \n\n```\ndocker stack deploy -c visualizer.yml vis\n```\n\n### portainer\n\nIf you'd like a more sophisticated UI with info about your swarm, you can deploy portainer as follows.\n\n```\ndocker service create \\\n --name portainer \\\n --publish 9000:9000 \\\n --constraint 'node.role == manager' \\\n --mount type=bind,src=/var/run/docker.sock,dst=/var/run/docker.sock \\\n portainer/portainer \\\n -H unix:///var/run/docker.sock\n```\n\nAlternatively, you can make use of the docker-compose file available in this folder.\n\n```\ndocker stack deploy -c portainer.yml portainer\n```\n\n### postman\n\n A well-known tool for experimenting with APIs. Do you want to try the curl-based examples of the recipes from Postman? Import the\n\n```\npostman_collection.json\n```\n\navailable in this folder and make your tests easier. Note: this collection is\nwork in progress, feel free to contribute! \n\n## Writing Docs?\n\n We typically write documentation in markdown format. Then, mkdocs is used to generate the html format. You can see in the root of this project the `mkdocs.yml` config file. \nFor architecture diagrams, we use PlantUML. For the diagrams we follow the conventions and leverage on the features you can find in this project.\n Instead of uploading pictures in this document, we use gravizo's power to convert the .dot or PlantUML files and have them served as pictures online. There is an intermediate conversion done with gravizo's converter. Inspect the source of any recipe's `readme.md` to see an example. \nOther tools for documentation that you may find useful are...\n\n### draw.io\n\nUse this tool when the diagrams start getting too complex of when you foresee the diagram will be complex from the scratch.\n Complex in the sense that making a simple change takes more time understanding the `.dot` than making a manual gui-based change. When using draw.io, keep the source file in the repository under a `/doc` subfolder of the corresponding recipe. \n\n### color names\n\n The reference for color names used in `.dot` files. \n\n### diagramr (deprecated)\n\nTo give more docker-related details we could use this tool to create diagrams from docker-compose files. The tools gives also the .dot file, which would be eventually customized and then turned into a png file using graphviz.\n\n```\n$ dot compose.dot -Tpng -o compose.png\n```\n\nContributions are more than welcome in the form of Pull Requests.\n\nFeel free to open issues if something looks wrong.\n\nBe aware that during the CI process, a number of linters are run:\n\nTo ensure correctness of yml files, we recommend you to check yaml linting rules.\n * \n\nTo ensure consistency of the documentation style, we recommend you to adhere to the MD linting rules.\n\nOnce you make a pull request to the repository, you will be able to observe the results of the compliancy verification in your PR. Merge will be only possible after a successful CI run.\n\n## Important observations for new recipes\n\nPlease add a\n `README.md` with an introduction on how to use the recipe and which things MUST be set in order for it to work. * \n\nState which services are stateful and which are stateless.\n * \n\nProvide the recipe with sensible defaults that work out of the box in the FIWARE Lab. Use as much defaults as possible, so users need little to none configuration before the first testing deployment.\n * \n\nPlease keep recipes in order (respect the folder structure) following categories.\n\n## Contributing to Portainer Recipes\n\nThe SmartDSK Recipes aim to be deployable both using command line tools following this Guide or by using a portainer template.\n\nThe documentation of the portainer template file is documented in the Stack template definition format\n The json file is however not human friendly to edit, so in this project the templates should be written in the\n\nfiles, and before committing changes, call\nthe `make` from the root directory of the project in order to update\nthe file \n\n. Then Add and commit the files \n\nand \n\ntogether. \n\n## Documentation\n\nFor now we are using Mkdocs deploying on Github Pages.\n You will also notice that instead of having a separate `docs` folder,\nthe documentation is composed of the README's content of all subfolders so as\nto keep docs as close to the respective recipes as possible. If you change the structure of the index or add new pages, remember to update `mkdocs.yml` accordingly. \nNote you can preview your changes locally by running\n\nAfter all your changes, remember to run"}}},{"rowIdx":98,"cells":{"project":{"kind":"string","value":"learningstatisticswithr_com_book"},"source":{"kind":"string","value":"free_programming_book"},"language":{"kind":"string","value":"Unknown"},"content":{"kind":"string","value":"# Licensing\n\nThis book is published under a Creative Commons BY-SA license (CC BY-SA) version 4.0. This means that this book can be reused, remixed, retained, revised and redistributed (including commercially) as long as appropriate credit is given to the authors. If you remix, or modify the original version of this open textbook, you must redistribute all versions of this open textbook under the same license - CC BY-SA. https://creativecommons.org/licenses/by-sa/4.0/\n\nLearning statistics with R: A tutorial for psychology students and other beginners. (Version 0.6.1) Dedication This book was brought to you today by the letter ‘R’.\n\nDate: 2013-01-13\nCategories: \nTags: \n\n# Preface\n\n## 0.1 Preface to Version 0.6.1\n\nHi! I’m not Danielle.\n\nThis version is a work in progress to transport the book from LaTeX to bookdown, this allows for the code chunks shown in the text to be interactively rendered and is the first step in updating sections to make use of the tidyverse.\n\nMoving from LaTeX is a bit fiddly so there are broken pieces in this version that I am updating progressively. If you notice errors or have suggestions feel free to raise an issue (or pull request) at http://github.com/ekothe/rbook\n\nCheers \n\n## 0.2 Preface to Version 0.6\n\nThe book hasn’t changed much since 2015 when I released Version 0.5 – it’s probably fair to say that I’ve changed more than it has. I moved from Adelaide to Sydney in 2016 and my teaching profile at UNSW is different to what it was at Adelaide, and I haven’t really had a chance to work on it since arriving here! It’s a little strange looking back at this actually. A few quick comments…\n\n* Weirdly, the book consistently misgenders me, but I suppose I have only myself to blame for that one :-) There’s now a brief footnote on page 12 that mentions this issue; in real life I’ve been working through a gender affirmation process for the last two years and mostly go by she/her pronouns. I am, however, just as lazy as I ever was so I haven’t bothered updating the text in the book.\n * For Version 0.6 I haven’t changed much I’ve made a few minor changes when people have pointed out typos or other errors. In particular it’s worth noting the issue associated with the etaSquared function in the lsr package (which isn’t really being maintained any more) in Section 14.4. The function works fine for the simple examples in the book, but there are definitely bugs in there that I haven’t found time to check! So please take care with that one.\n * The biggest change really is the licensing! I’ve released it under a Creative Commons licence (CC BY-SA 4.0, specifically), and placed all the source files to the associated GitHub repository, if anyone wants to adapt it.\n\nMaybe someone would like to write a version that makes use of the tidyverse… I hear that’s become rather important to R these days :-)\n\nBest, \n\nAnother year, another update. This time around, the update has focused almost entirely on the theory sections of the book. Chapters 9, 10 and 11 have been rewritten, hopefully for the better. Along the same lines, Chapter 17 is entirely new, and focuses on Bayesian statistics. I think the changes have improved the book a great deal. I’ve always felt uncomfortable about the fact that all the inferential statistics in the book are presented from an orthodox perspective, even though I almost always present Bayesian data analyses in my own work. Now that I’ve managed to squeeze Bayesian methods into the book somewhere, I’m starting to feel better about the book as a whole. I wanted to get a few other things done in this update, but as usual I’m running into teaching deadlines, so the update has to go out the way it is!\n\n February 16, 2015\n\n## 0.4 Preface to Version 0.4\n\nA year has gone by since I wrote the last preface. The book has changed in a few important ways: Chapters 3 and 4 do a better job of documenting some of the time saving features of Rstudio, Chapters 12 and 13 now make use of new functions in the lsr package for running chi-square tests and t tests, and the discussion of correlations has been adapted to refer to the new functions in the lsr package. The soft copy of 0.4 now has better internal referencing (i.e., actual hyperlinks between sections), though that was introduced in 0.3.1. There’s a few tweaks here and there, and many typo corrections (thank you to everyone who pointed out typos!), but overall 0.4 isn’t massively different from 0.3.\n\nI wish I’d had more time over the last 12 months to add more content. The absence of any discussion of repeated measures ANOVA and mixed models more generally really does annoy me. My excuse for this lack of progress is that my second child was born at the start of 2013, and so I spent most of last year just trying to keep my head above water. As a consequence, unpaid side projects like this book got sidelined in favour of things that actually pay my salary! Things are a little calmer now, so with any luck version 0.5 will be a bigger step forward.\n\nOne thing that has surprised me is the number of downloads the book gets. I finally got some basic tracking information from the website a couple of months ago, and (after excluding obvious robots) the book has been averaging about 90 downloads per day. That’s encouraging: there’s at least a few people who find the book useful!\n\n February 4, 2014\n\nThere’s a part of me that really doesn’t want to publish this book. It’s not finished.\n\nAnd when I say that, I mean it. The referencing is spotty at best, the chapter summaries are just lists of section titles, there’s no index, there are no exercises for the reader, the organisation is suboptimal, and the coverage of topics is just not comprehensive enough for my liking. Additionally, there are sections with content that I’m not happy with, figures that really need to be redrawn, and I’ve had almost no time to hunt down inconsistencies, typos, or errors. In other words, this book is not finished. If I didn’t have a looming teaching deadline and a baby due in a few weeks, I really wouldn’t be making this available at all.\n\nWhat this means is that if you are an academic looking for teaching materials, a Ph.D. student looking to learn R, or just a member of the general public interested in statistics, I would advise you to be cautious. What you’re looking at is a first draft, and it may not serve your purposes. If we were living in the days when publishing was expensive and the internet wasn’t around, I would never consider releasing a book in this form. The thought of someong shelling out $80 for this (which is what a commercial publisher told me it would retail for when they offered to distribute it) makes me feel more than a little uncomfortable. However, it’s the 21st century, so I can post the pdf on my website for free, and I can distribute hard copies via a print-on-demand service for less than half what a textbook publisher would charge. And so my guilt is assuaged, and I’m willing to share! With that in mind, you can obtain free soft copies and cheap hard copies online, from the following webpages:\n\n* http://www.compcogscisydney.com/learning-statistics-with-r.html\n * http://www.lulu.com/content/13570633\n\nEven so, the warning still stands: what you are looking at is Version 0.3 of a work in progress. If and when it hits Version 1.0, I would be willing to stand behind the work and say, yes, this is a textbook that I would encourage other people to use. At that point, I’ll probably start shamelessly flogging the thing on the internet and generally acting like a tool. But until that day comes, I’d like it to be made clear that I’m really ambivalent about the work as it stands.\n\nAll of the above being said, there is one group of people that I can enthusiastically endorse this book to: the psychology students taking our undergraduate research methods classes (DRIP and DRIP:A) in 2013. For you, this book is ideal, because it was written to accompany your stats lectures. If a problem arises due to a shortcoming of these notes, I can and will adapt content on the fly to fix that problem. Effectively, you’ve got a textbook written specifically for your classes, distributed for free (electronic copy) or at near-cost prices (hard copy). Better yet, the notes have been tested: Version 0.1 of these notes was used in the 2011 class, Version 0.2 was used in the 2012 class, and now you’re looking at the new and improved Version 0.3. I’m not saying these notes are titanium plated awesomeness on a stick – though if you wanted to say so on the student evaluation forms, then you’re totally welcome to – because they’re not. But I am saying that they’ve been tried out in previous years and they seem to work okay. Besides, there’s a group of us around to troubleshoot if any problems come up, and you can guarantee that at least one of your lecturers has read the whole thing cover to cover!\n\nOkay, with all that out of the way, I should say something about what the book aims to be. At its core, it is an introductory statistics textbook pitched primarily at psychology students. As such, it covers the standard topics that you’d expect of such a book: study design, descriptive statistics, the theory of hypothesis testing, \\(t\\)-tests, \\(\\chi^2\\) tests, ANOVA and regression. However, there are also several chapters devoted to the R statistical package, including a chapter on data manipulation and another one on scripts and programming. Moreover, when you look at the content presented in the book, you’ll notice a lot of topics that are traditionally swept under the carpet when teaching statistics to psychology students. The Bayesian/frequentist divide is openly disussed in the probability chapter, and the disagreement between Neyman and Fisher about hypothesis testing makes an appearance. The difference between probability and density is discussed. A detailed treatment of Type I, II and III sums of squares for unbalanced factorial ANOVA is provided. And if you have a look in the Epilogue, it should be clear that my intention is to add a lot more advanced content.\n\nMy reasons for pursuing this approach are pretty simple: the students can handle it, and they even seem to enjoy it. Over the last few years I’ve been pleasantly surprised at just how little difficulty I’ve had in getting undergraduate psych students to learn R. It’s certainly not easy for them, and I’ve found I need to be a little charitable in setting marking standards, but they do eventually get there. Similarly, they don’t seem to have a lot of problems tolerating ambiguity and complexity in presentation of statistical ideas, as long as they are assured that the assessment standards will be set in a fashion that is appropriate for them. So if the students can handle it, why not teach it? The potential gains are pretty enticing. If they learn R, the students get access to CRAN, which is perhaps the largest and most comprehensive library of statistical tools in existence. And if they learn about probability theory in detail, it’s easier for them to switch from orthodox null hypothesis testing to Bayesian methods if they want to. Better yet, they learn data analysis skills that they can take to an employer without being dependent on expensive and proprietary software.\n\nSadly, this book isn’t the silver bullet that makes all this possible. It’s a work in progress, and maybe when it is finished it will be a useful tool. One among many, I would think. There are a number of other books that try to provide a basic introduction to statistics using R, and I’m not arrogant enough to believe that mine is better. Still, I rather like the book, and maybe other people will find it useful, incomplete though it is.\n\n January 13, 2013\n\nDate: 2016-05-08\nCategories: \nTags: \n\n# Chapter 1 Why do we learn statistics?\n\n“Thou shalt not answer questionnaires\n Or quizzes upon World Affairs, Nor with compliance\n Take any test. Thou shalt not sit With statisticians nor commit\" – W.H. Auden1\n\n## 1.1 On the psychology of statistics\n\nTo the surprise of many students, statistics is a fairly significant part of a psychological education. To the surprise of no-one, statistics is very rarely the favourite part of one’s psychological education. After all, if you really loved the idea of doing statistics, you’d probably be enrolled in a statistics class right now, not a psychology class. So, not surprisingly, there’s a pretty large proportion of the student base that isn’t happy about the fact that psychology has so much statistics in it. In view of this, I thought that the right place to start might be to answer some of the more common questions that people have about stats…\n\nA big part of this issue at hand relates to the very idea of statistics. What is it? What’s it there for? And why are scientists so bloody obsessed with it? These are all good questions, when you think about it. So let’s start with the last one. As a group, scientists seem to be bizarrely fixated on running statistical tests on everything. In fact, we use statistics so often that we sometimes forget to explain to people why we do. It’s a kind of article of faith among scientists – and especially social scientists – that your findings can’t be trusted until you’ve done some stats. Undergraduate students might be forgiven for thinking that we’re all completely mad, because no-one takes the time to answer one very simple question:\n Why do you do statistics? Why don’t scientists just use common sense?\n\nIt’s a naive question in some ways, but most good questions are. There’s a lot of good answers to it,2 but for my money, the best answer is a really simple one: we don’t trust ourselves enough. We worry that we’re human, and susceptible to all of the biases, temptations and frailties that humans suffer from. Much of statistics is basically a safeguard. Using “common sense” to evaluate evidence means trusting gut instincts, relying on verbal arguments and on using the raw power of human reason to come up with the right answer. Most scientists don’t think this approach is likely to work.\n\nIn fact, come to think of it, this sounds a lot like a psychological question to me, and since I do work in a psychology department, it seems like a good idea to dig a little deeper here. Is it really plausible to think that this “common sense” approach is very trustworthy? Verbal arguments have to be constructed in language, and all languages have biases – some things are harder to say than others, and not necessarily because they’re false (e.g., quantum electrodynamics is a good theory, but hard to explain in words). The instincts of our “gut” aren’t designed to solve scientific problems, they’re designed to handle day to day inferences – and given that biological evolution is slower than cultural change, we should say that they’re designed to solve the day to day problems for a different world than the one we live in. Most fundamentally, reasoning sensibly requires people to engage in “induction”, making wise guesses and going beyond the immediate evidence of the senses to make generalisations about the world. If you think that you can do that without being influenced by various distractors, well, I have a bridge in Brooklyn I’d like to sell you. Heck, as the next section shows, we can’t even solve “deductive” problems (ones where no guessing is required) without being influenced by our pre-existing biases.\n\n### 1.1.1 The curse of belief bias\n\nPeople are mostly pretty smart. We’re certainly smarter than the other species that we share the planet with (though many people might disagree). Our minds are quite amazing things, and we seem to be capable of the most incredible feats of thought and reason. That doesn’t make us perfect though. And among the many things that psychologists have shown over the years is that we really do find it hard to be neutral, to evaluate evidence impartially and without being swayed by pre-existing biases. A good example of this is the belief bias effect in logical reasoning: if you ask people to decide whether a particular argument is logically valid (i.e., conclusion would be true if the premises were true), we tend to be influenced by the believability of the conclusion, even when we shouldn’t. For instance, here’s a valid argument where the conclusion is believable:\n No cigarettes are inexpensive (Premise 1)\n Some addictive things are inexpensive (Premise 2) Therefore, some addictive things are not cigarettes (Conclusion) \nAnd here’s a valid argument where the conclusion is not believable:\n No addictive things are inexpensive (Premise 1)\n Some cigarettes are inexpensive (Premise 2) Therefore, some cigarettes are not addictive (Conclusion) \nThe logical structure of argument #2 is identical to the structure of argument #1, and they’re both valid. However, in the second argument, there are good reasons to think that premise 1 is incorrect, and as a result it’s probably the case that the conclusion is also incorrect. But that’s entirely irrelevant to the topic at hand: an argument is deductively valid if the conclusion is a logical consequence of the premises. That is, a valid argument doesn’t have to involve true statements.\n\nOn the other hand, here’s an invalid argument that has a believable conclusion:\n No addictive things are inexpensive (Premise 1)\n Some cigarettes are inexpensive (Premise 2) Therefore, some addictive things are not cigarettes (Conclusion) \nAnd finally, an invalid argument with an unbelievable conclusion:\n No cigarettes are inexpensive (Premise 1)\n Some addictive things are inexpensive (Premise 2) Therefore, some cigarettes are not addictive (Conclusion) \nNow, suppose that people really are perfectly able to set aside their pre-existing biases about what is true and what isn’t, and purely evaluate an argument on its logical merits. We’d expect 100% of people to say that the valid arguments are valid, and 0% of people to say that the invalid arguments are valid. So if you ran an experiment looking at this, you’d expect to see data like this:\n\nconclusion feels true | conclusion feels false |\n | --- | --- |\nargument is valid | 100% say “valid” | 100% say “valid” |\nargument is invalid | 0% say “valid” | 0% say “valid” |\n\nIf the psychological data looked like this (or even a good approximation to this), we might feel safe in just trusting our gut instincts. That is, it’d be perfectly okay just to let scientists evaluate data based on their common sense, and not bother with all this murky statistics stuff. However, you guys have taken psych classes, and by now you probably know where this is going…\n\nIn a classic study, Evans, Barston, and Pollard (1983) ran an experiment looking at exactly this. What they found is that when pre-existing biases (i.e., beliefs) were in agreement with the structure of the data, everything went the way you’d hope:\n\nNot perfect, but that’s pretty good. But look what happens when our intuitive feelings about the truth of the conclusion run against the logical structure of the argument:\n\nOh dear, that’s not as good. Apparently, when people are presented with a strong argument that contradicts our pre-existing beliefs, we find it pretty hard to even perceive it to be a strong argument (people only did so 46% of the time). Even worse, when people are presented with a weak argument that agrees with our pre-existing biases, almost no-one can see that the argument is weak (people got that one wrong 92% of the time!)3\n\nIf you think about it, it’s not as if these data are horribly damning. Overall, people did do better than chance at compensating for their prior biases, since about 60% of people’s judgements were correct (you’d expect 50% by chance). Even so, if you were a professional “evaluator of evidence”, and someone came along and offered you a magic tool that improves your chances of making the right decision from 60% to (say) 95%, you’d probably jump at it, right? Of course you would. Thankfully, we actually do have a tool that can do this. But it’s not magic, it’s statistics. So that’s reason #1 why scientists love statistics. It’s just too easy for us to “believe what we want to believe”; so if we want to “believe in the data” instead, we’re going to need a bit of help to keep our personal biases under control. That’s what statistics does: it helps keep us honest.\n\n## 1.2 The cautionary tale of Simpson’s paradox\n\nThe following is a true story (I think…). In 1973, the University of California, Berkeley had some worries about the admissions of students into their postgraduate courses. Specifically, the thing that caused the problem was that the gender breakdown of their admissions, which looked like this…\n\nNumber of applicants | Percent admitted |\n | --- | --- |\nMales | 8442 | 46% |\nFemales | 4321 | 35% |\n\n…and the were worried about being sued.4 Given that there were nearly 13,000 applicants, a difference of 9% in admission rates between males and females is just way too big to be a coincidence. Pretty compelling data, right? And if I were to say to you that these data actually reflect a weak bias in favour of women (sort of!), you’d probably think that I was either crazy or sexist.\n\nOddly, it’s actually sort of true …when people started looking more carefully at the admissions data (Bickel, Hammel, and O’Connell 1975) they told a rather different story. Specifically, when they looked at it on a department by department basis, it turned out that most of the departments actually had a slightly higher success rate for female applicants than for male applicants. Table 1.1 shows the admission figures for the six largest departments (with the names of the departments removed for privacy reasons):\n\nDepartment | Male Applicants | Male Percent Admitted | Female Applicants | Female Percent admitted |\n | --- | --- | --- | --- | --- |\nA | 825 | 62% | 108 | 82% |\nB | 560 | 63% | 25 | 68% |\nC | 325 | 37% | 593 | 34% |\nD | 417 | 33% | 375 | 35% |\nE | 191 | 28% | 393 | 24% |\nF | 272 | 6% | 341 | 7% |\n\nRemarkably, most departments had a higher rate of admissions for females than for males! Yet the overall rate of admission across the university for females was lower than for males. How can this be? How can both of these statements be true at the same time?\n\nHere’s what’s going on. Firstly, notice that the departments are not equal to one another in terms of their admission percentages: some departments (e.g., engineering, chemistry) tended to admit a high percentage of the qualified applicants, whereas others (e.g., English) tended to reject most of the candidates, even if they were high quality. So, among the six departments shown above, notice that department A is the most generous, followed by B, C, D, E and F in that order. Next, notice that males and females tended to apply to different departments. If we rank the departments in terms of the total number of male applicants, we get A>B>D>C>F>E (the “easy” departments are in bold). On the whole, males tended to apply to the departments that had high admission rates. Now compare this to how the female applicants distributed themselves. Ranking the departments in terms of the total number of female applicants produces a quite different ordering C>E>D>F>A>B. In other words, what these data seem to be suggesting is that the female applicants tended to apply to “harder” departments. And in fact, if we look at all Figure 1.1 we see that this trend is systematic, and quite striking. This effect is known as Simpson’s paradox. It’s not common, but it does happen in real life, and most people are very surprised by it when they first encounter it, and many people refuse to even believe that it’s real. It is very real. And while there are lots of very subtle statistical lessons buried in there, I want to use it to make a much more important point …doing research is hard, and there are lots of subtle, counterintuitive traps lying in wait for the unwary. That’s reason #2 why scientists love statistics, and why we teach research methods. Because science is hard, and the truth is sometimes cunningly hidden in the nooks and crannies of complicated data.\n\nBefore leaving this topic entirely, I want to point out something else really critical that is often overlooked in a research methods class. Statistics only solves part of the problem. Remember that we started all this with the concern that Berkeley’s admissions processes might be unfairly biased against female applicants. When we looked at the “aggregated” data, it did seem like the university was discriminating against women, but when we “disaggregate” and looked at the individual behaviour of all the departments, it turned out that the actual departments were, if anything, slightly biased in favour of women. The gender bias in total admissions was caused by the fact that women tended to self-select for harder departments. From a legal perspective, that would probably put the university in the clear. Postgraduate admissions are determined at the level of the individual department (and there are good reasons to do that), and at the level of individual departments, the decisions are more or less unbiased (the weak bias in favour of females at that level is small, and not consistent across departments). Since the university can’t dictate which departments people choose to apply to, and the decision making takes place at the level of the department it can hardly be held accountable for any biases that those choices produce.\n\nThat was the basis for my somewhat glib remarks earlier, but that’s not exactly the whole story, is it? After all, if we’re interested in this from a more sociological and psychological perspective, we might want to ask why there are such strong gender differences in applications. Why do males tend to apply to engineering more often than females, and why is this reversed for the English department? And why is it it the case that the departments that tend to have a female-application bias tend to have lower overall admission rates than those departments that have a male-application bias? Might this not still reflect a gender bias, even though every single department is itself unbiased? It might. Suppose, hypothetically, that males preferred to apply to “hard sciences” and females prefer “humanities”. And suppose further that the reason for why the humanities departments have low admission rates is because the government doesn’t want to fund the humanities (Ph.D. places, for instance, are often tied to government funded research projects). Does that constitute a gender bias? Or just an unenlightened view of the value of the humanities? What if someone at a high level in the government cut the humanities funds because they felt that the humanities are “useless chick stuff”. That seems pretty blatantly gender biased. None of this falls within the purview of statistics, but it matters to the research project. If you’re interested in the overall structural effects of subtle gender biases, then you probably want to look at both the aggregated and disaggregated data. If you’re interested in the decision making process at Berkeley itself then you’re probably only interested in the disaggregated data.\n\nIn short there are a lot of critical questions that you can’t answer with statistics, but the answers to those questions will have a huge impact on how you analyse and interpret data. And this is the reason why you should always think of statistics as a tool to help you learn about your data, no more and no less. It’s a powerful tool to that end, but there’s no substitute for careful thought.\n\n## 1.3 Statistics in psychology\n\nI hope that the discussion above helped explain why science in general is so focused on statistics. But I’m guessing that you have a lot more questions about what role statistics plays in psychology, and specifically why psychology classes always devote so many lectures to stats. So here’s my attempt to answer a few of them…\n\nWhy does psychology have so much statistics?\n\nTo be perfectly honest, there’s a few different reasons, some of which are better than others. The most important reason is that psychology is a statistical science. What I mean by that is that the “things” that we study are people. Real, complicated, gloriously messy, infuriatingly perverse people. The “things” of physics include object like electrons, and while there are all sorts of complexities that arise in physics, electrons don’t have minds of their own. They don’t have opinions, they don’t differ from each other in weird and arbitrary ways, they don’t get bored in the middle of an experiment, and they don’t get angry at the experimenter and then deliberately try to sabotage the data set (not that I’ve ever done that…). At a fundamental level psychology is harder than physics.5\n\nBasically, we teach statistics to you as psychologists because you need to be better at stats than physicists. There’s actually a saying used sometimes in physics, to the effect that “if your experiment needs statistics, you should have done a better experiment”. They have the luxury of being able to say that because their objects of study are pathetically simple in comparison to the vast mess that confronts social scientists. It’s not just psychology, really: most social sciences are desperately reliant on statistics. Not because we’re bad experimenters, but because we’ve picked a harder problem to solve. We teach you stats because you really, really need it.\n\nCan’t someone else do the statistics?\n\nTo some extent, but not completely. It’s true that you don’t need to become a fully trained statistician just to do psychology, but you do need to reach a certain level of statistical competence. In my view, there’s three reasons that every psychological researcher ought to be able to do basic statistics:\n\n* Firstly, there’s the fundamental reason: statistics is deeply intertwined with research design. If you want to be good at designing psychological studies, you need to at least understand the basics of stats.\n * Secondly, if you want to be good at the psychological side of the research, then you need to be able to understand the psychological literature, right? But almost every paper in the psychological literature reports the results of statistical analyses. So if you really want to understand the psychology, you need to be able to understand what other people did with their data. And that means understanding a certain amount of statistics.\n * Thirdly, there’s a big practical problem with being dependent on other people to do all your statistics: statistical analysis is expensive. If you ever get bored and want to look up how much the Australian government charges for university fees, you’ll notice something interesting: statistics is designated as a “national priority” category, and so the fees are much, much lower than for any other area of study. This is because there’s a massive shortage of statisticians out there. So, from your perspective as a psychological researcher, the laws of supply and demand aren’t exactly on your side here! As a result, in almost any real life situation where you want to do psychological research, the cruel facts will be that you don’t have enough money to afford a statistician. So the economics of the situation mean that you have to be pretty self-sufficient.\n\nNote that a lot of these reasons generalise beyond researchers. If you want to be a practicing psychologist and stay on top of the field, it helps to be able to read the scientific literature, which relies pretty heavily on statistics.\n\nI don’t care about jobs, research, or clinical work. Do I need statistics?\n\nOkay, now you’re just messing with me. Still, I think it should matter to you too. Statistics should matter to you in the same way that statistics should matter to everyone: we live in the 21st century, and data are everywhere. Frankly, given the world in which we live these days, a basic knowledge of statistics is pretty damn close to a survival tool! Which is the topic of the next section…\n\n## 1.4 Statistics in everyday life\n\n“We are drowning in information, but we are starved for knowledge”\n\n-Various authors, original probably \n\nWhen I started writing up my lecture notes I took the 20 most recent news articles posted to the ABC news website. Of those 20 articles, it turned out that 8 of them involved a discussion of something that I would call a statistical topic; 6 of those made a mistake. The most common error, if you’re curious, was failing to report baseline data (e.g., the article mentions that 5% of people in situation X have some characteristic Y, but doesn’t say how common the characteristic is for everyone else!) The point I’m trying to make here isn’t that journalists are bad at statistics (though they almost always are), it’s that a basic knowledge of statistics is very helpful for trying to figure out when someone else is either making a mistake or even lying to you. In fact, one of the biggest things that a knowledge of statistics does to you is cause you to get angry at the newspaper or the internet on a far more frequent basis: you can find a good example of this in Section 5.1.5. In later versions of this book I’ll try to include more anecdotes along those lines.\n\n## 1.5 There’s more to research methods than statistics\n\nSo far, most of what I’ve talked about is statistics, and so you’d be forgiven for thinking that statistics is all I care about in life. To be fair, you wouldn’t be far wrong, but research methodology is a broader concept than statistics. So most research methods courses will cover a lot of topics that relate much more to the pragmatics of research design, and in particular the issues that you encounter when trying to do research with humans. However, about 99% of student fears relate to the statistics part of the course, so I’ve focused on the stats in this discussion, and hopefully I’ve convinced you that statistics matters, and more importantly, that it’s not to be feared. That being said, it’s pretty typical for introductory research methods classes to be very stats-heavy. This is not (usually) because the lecturers are evil people. Quite the contrary, in fact. Introductory classes focus a lot on the statistics because you almost always find yourself needing statistics before you need the other research methods training. Why? Because almost all of your assignments in other classes will rely on statistical training, to a much greater extent than they rely on other methodological tools. It’s not common for undergraduate assignments to require you to design your own study from the ground up (in which case you would need to know a lot about research design), but it is common for assignments to ask you to analyse and interpret data that were collected in a study that someone else designed (in which case you need statistics). In that sense, from the perspective of allowing you to do well in all your other classes, the statistics is more urgent.\n\nBut note that “urgent” is different from “important” – they both matter. I really do want to stress that research design is just as important as data analysis, and this book does spend a fair amount of time on it. However, while statistics has a kind of universality, and provides a set of core tools that are useful for most types of psychological research, the research methods side isn’t quite so universal. There are some general principles that everyone should think about, but a lot of research design is very idiosyncratic, and is specific to the area of research that you want to engage in. To the extent that it’s the details that matter, those details don’t usually show up in an introductory stats and research methods class.\n\nThe quote comes from Auden’s 1946 poem Under Which Lyre: A Reactionary Tract for the Times, delivered as part of a commencement address at Harvard University. The history of the poem is kind of interesting: http://harvardmagazine.com/2007/11/a-poets-warning.html↩\n * \n\nIncluding the suggestion that common sense is in short supply among scientists.↩\n * \n\nIn my more cynical moments I feel like this fact alone explains 95% of what I read on the internet.↩\n * \n\nEarlier versions of these notes incorrectly suggested that they actually were sued – apparently that’s not true. There’s a nice commentary on this here: https://www.refsmmat.com/posts/2016-05-08-simpsons-paradox-berkeley.html. A big thank you to for pointing this out to me!↩\n * \n\nWhich might explain why physics is just a teensy bit further advanced as a science than we are.↩\n\n# Chapter 2 A brief introduction to research design\n\nTo consult the statistician after an experiment is finished is often merely to ask him to conduct a post mortem examination. He can perhaps say what the experiment died of.\n\n– \n\nIn this chapter, we’re going to start thinking about the basic ideas that go into designing a study, collecting data, checking whether your data collection works, and so on. It won’t give you enough information to allow you to design studies of your own, but it will give you a lot of the basic tools that you need to assess the studies done by other people. However, since the focus of this book is much more on data analysis than on data collection, I’m only giving a very brief overview. Note that this chapter is “special” in two ways. Firstly, it’s much more psychology-specific than the later chapters. Secondly, it focuses much more heavily on the scientific problem of research methodology, and much less on the statistical problem of data analysis. Nevertheless, the two problems are related to one another, so it’s traditional for stats textbooks to discuss the problem in a little detail. This chapter relies heavily on Campbell and Stanley (1963) for the discussion of study design, and Stevens (1946) for the discussion of scales of measurement. Later versions will attempt to be more precise in the citations.\n\n## 2.1 Introduction to psychological measurement\n\nThe first thing to understand is data collection can be thought of as a kind of measurement. That is, what we’re trying to do here is measure something about human behaviour or the human mind. What do I mean by “measurement”?\n\n### 2.1.1 Some thoughts about psychological measurement\n\nMeasurement itself is a subtle concept, but basically it comes down to finding some way of assigning numbers, or labels, or some other kind of well-defined descriptions to “stuff”. So, any of the following would count as a psychological measurement:\n\n* My age is 33 years.\n * I do not like anchovies.\n * My chromosomal gender is male.\n * My self-identified gender is male.7\n\nIn the short list above, the bolded part is “the thing to be measured”, and the italicised part is “the measurement itself”. In fact, we can expand on this a little bit, by thinking about the set of possible measurements that could have arisen in each case:\n\n* My age (in years) could have been 0, 1, 2, 3 …, etc. The upper bound on what my age could possibly be is a bit fuzzy, but in practice you’d be safe in saying that the largest possible age is 150, since no human has ever lived that long.\n * When asked if I like anchovies, I might have said that I do, or I do not, or I have no opinion, or I sometimes do.\n * My chromosomal gender is almost certainly going to be male (XY) or female (XX), but there are a few other possibilities. I could also have Klinfelter’s syndrome (XXY), which is more similar to male than to female. And I imagine there are other possibilities too.\n * My self-identified gender is also very likely to be male or female, but it doesn’t have to agree with my chromosomal gender. I may also choose to identify with neither, or to explicitly call myself transgender.\n\nAs you can see, for some things (like age) it seems fairly obvious what the set of possible measurements should be, whereas for other things it gets a bit tricky. But I want to point out that even in the case of someone’s age, it’s much more subtle than this. For instance, in the example above, I assumed that it was okay to measure age in years. But if you’re a developmental psychologist, that’s way too crude, and so you often measure age in years and months (if a child is 2 years and 11 months, this is usually written as “2;11”). If you’re interested in newborns, you might want to measure age in days since birth, maybe even hours since birth. In other words, the way in which you specify the allowable measurement values is important.\n\nLooking at this a bit more closely, you might also realise that the concept of “age” isn’t actually all that precise. In general, when we say “age” we implicitly mean “the length of time since birth”. But that’s not always the right way to do it. Suppose you’re interested in how newborn babies control their eye movements. If you’re interested in kids that young, you might also start to worry that “birth” is not the only meaningful point in time to care about. If is born 3 weeks premature and is born 1 week late, would it really make sense to say that they are the “same age” if we encountered them “2 hours after birth”? In one sense, yes: by social convention, we use birth as our reference point for talking about age in everyday life, since it defines the amount of time the person has been operating as an independent entity in the world, but from a scientific perspective that’s not the only thing we care about. When we think about the biology of human beings, it’s often useful to think of ourselves as organisms that have been growing and maturing since conception, and from that perspective Alice and Bianca aren’t the same age at all. So you might want to define the concept of “age” in two different ways: the length of time since conception, and the length of time since birth. When dealing with adults, it won’t make much difference, but when dealing with newborns it might.\n\nMoving beyond these issues, there’s the question of methodology. What specific “measurement method” are you going to use to find out someone’s age? As before, there are lots of different possibilities:\n\n* You could just ask people “how old are you?” The method of self-report is fast, cheap and easy, but it only works with people old enough to understand the question, and some people lie about their age.\n * You could ask an authority (e.g., a parent) “how old is your child?” This method is fast, and when dealing with kids it’s not all that hard since the parent is almost always around. It doesn’t work as well if you want to know “age since conception”, since a lot of parents can’t say for sure when conception took place. For that, you might need a different authority (e.g., an obstetrician).\n * You could look up official records, like birth certificates. This is time consuming and annoying, but it has its uses (e.g., if the person is now dead).\n\n### 2.1.2 Operationalisation: defining your measurement\n\nAll of the ideas discussed in the previous section all relate to the concept of operationalisation. To be a bit more precise about the idea, operationalisation is the process by which we take a meaningful but somewhat vague concept, and turn it into a precise measurement. The process of operationalisation can involve several different things:\n\n* Being precise about what you are trying to measure. For instance, does “age” mean “time since birth” or “time since conception” in the context of your research?\n * Determining what method you will use to measure it. Will you use self-report to measure age, ask a parent, or look up an official record? If you’re using self-report, how will you phrase the question?\n * Defining the set of the allowable values that the measurement can take. Note that these values don’t always have to be numerical, though they often are. When measuring age, the values are numerical, but we still need to think carefully about what numbers are allowed. Do we want age in years, years and months, days, hours? Etc. For other types of measurements (e.g., gender), the values aren’t numerical. But, just as before, we need to think about what values are allowed. If we’re asking people to self-report their gender, what options to we allow them to choose between? Is it enough to allow only “male” or “female”? Do you need an “other” option? Or should we not give people any specific options, and let them answer in their own words? And if you open up the set of possible values to include all verbal response, how will you interpret their answers?\n\nOperationalisation is a tricky business, and there’s no “one, true way” to do it. The way in which you choose to operationalise the informal concept of “age” or “gender” into a formal measurement depends on what you need to use the measurement for. Often you’ll find that the community of scientists who work in your area have some fairly well-established ideas for how to go about it. In other words, operationalisation needs to be thought through on a case by case basis. Nevertheless, while there a lot of issues that are specific to each individual research project, there are some aspects to it that are pretty general.\n\nBefore moving on, I want to take a moment to clear up our terminology, and in the process introduce one more term. Here are four different things that are closely related to each other:\n\n* A theoretical construct. This is the thing that you’re trying to take a measurement of, like “age”, “gender” or an “opinion”. A theoretical construct can’t be directly observed, and often they’re actually a bit vague.\n * A measure. The measure refers to the method or the tool that you use to make your observations. A question in a survey, a behavioural observation or a brain scan could all count as a measure.\n * An operationalisation. The term “operationalisation” refers to the logical connection between the measure and the theoretical construct, or to the process by which we try to derive a measure from a theoretical construct.\n * A variable. Finally, a new term. A variable is what we end up with when we apply our measure to something in the world. That is, variables are the actual “data” that we end up with in our data sets.\n\nIn practice, even scientists tend to blur the distinction between these things, but it’s very helpful to try to understand the differences.\n\n## 2.2 Scales of measurement\n\nAs the previous section indicates, the outcome of a psychological measurement is called a variable. But not all variables are of the same qualitative type, and it’s very useful to understand what types there are. A very useful concept for distinguishing between different types of variables is what’s known as scales of measurement.\n\n### 2.2.1 Nominal scale\n\nA nominal scale variable (also referred to as a categorical variable) is one in which there is no particular relationship between the different possibilities: for these kinds of variables it doesn’t make any sense to say that one of them is “bigger’ or”better\" than any other one, and it absolutely doesn’t make any sense to average them. The classic example for this is “eye colour”. Eyes can be blue, green and brown, among other possibilities, but none of them is any “better” than any other one. As a result, it would feel really weird to talk about an “average eye colour”. Similarly, gender is nominal too: male isn’t better or worse than female, neither does it make sense to try to talk about an “average gender”. In short, nominal scale variables are those for which the only thing you can say about the different possibilities is that they are different. That’s it.\n\nLet’s take a slightly closer look at this. Suppose I was doing research on how people commute to and from work. One variable I would have to measure would be what kind of transportation people use to get to work. This “transport type” variable could have quite a few possible values, including: “train”, “bus”, “car”, “bicycle”, etc. For now, let’s suppose that these four are the only possibilities, and suppose that when I ask 100 people how they got to work today, and I get this:\n\nSo, what’s the average transportation type? Obviously, the answer here is that there isn’t one. It’s a silly question to ask. You can say that travel by car is the most popular method, and travel by train is the least popular method, but that’s about all. Similarly, notice that the order in which I list the options isn’t very interesting. I could have chosen to display the data like this\n\nand nothing really changes.\n\n### 2.2.2 Ordinal scale\n\nOrdinal scale variables have a bit more structure than nominal scale variables, but not by a lot. An ordinal scale variable is one in which there is a natural, meaningful way to order the different possibilities, but you can’t do anything else. The usual example given of an ordinal variable is “finishing position in a race”. You can say that the person who finished first was faster than the person who finished second, but you don’t know how much faster. As a consequence we know that 1st > 2nd, and we know that 2nd > 3rd, but the difference between 1st and 2nd might be much larger than the difference between 2nd and 3rd.\n\nHere’s an more psychologically interesting example. Suppose I’m interested in people’s attitudes to climate change, and I ask them to pick one of these four statements that most closely matches their beliefs:\n * Temperatures are rising, because of human activity\n * Temperatures are rising, but we don’t know why\n * Temperatures are rising, but not because of humans\n * Temperatures are not rising\n\nNotice that these four statements actually do have a natural ordering, in terms of “the extent to which they agree with the current science”. Statement 1 is a close match, statement 2 is a reasonable match, statement 3 isn’t a very good match, and statement 4 is in strong opposition to the science. So, in terms of the thing I’m interested in (the extent to which people endorse the science), I can order the items as 1 > 2 > 3 > 4. Since this ordering exists, it would be very weird to list the options like this…\n * Temperatures are rising, but not because of humans\n * Temperatures are rising, because of human activity\n * Temperatures are not rising\n * Temperatures are rising, but we don’t know why\n\n… because it seems to violate the natural “structure” to the question.\n\nSo, let’s suppose I asked 100 people these questions, and got the following answers:\n\nResponse | Number |\n | --- | --- |\n(1) Temperatures are rising, because of human activity | 51 |\n(2) Temperatures are rising, but we don’t know why | 20 |\n(3) Temperatures are rising, but not because of humans | 10 |\n(4) Temperatures are not rising | 19 |\n\nWhen analysing these data, it seems quite reasonable to try to group (1), (2) and (3) together, and say that 81 of 100 people were willing to at least partially endorse the science. And it’s also quite reasonable to group (2), (3) and (4) together and say that 49 of 100 people registered at least some disagreement with the dominant scientific view. However, it would be entirely bizarre to try to group (1), (2) and (4) together and say that 90 of 100 people said… what? There’s nothing sensible that allows you to group those responses together at all.\n\nThat said, notice that while we can use the natural ordering of these items to construct sensible groupings, what we can’t do is average them. For instance, in my simple example here, the “average” response to the question is 1.97. If you can tell me what that means, I’d love to know. Because that sounds like gibberish to me!\n\n### 2.2.3 Interval scale\n\nIn contrast to nominal and ordinal scale variables, interval scale and ratio scale variables are variables for which the numerical value is genuinely meaningful. In the case of interval scale variables, the differences between the numbers are interpretable, but the variable doesn’t have a “natural” zero value. A good example of an interval scale variable is measuring temperature in degrees celsius. For instance, if it was 15\\(^\\circ\\) yesterday and 18\\(^\\circ\\) today, then the 3\\(^\\circ\\) difference between the two is genuinely meaningful. Moreover, that 3\\(^\\circ\\) difference is exactly the same as the 3\\(^\\circ\\) difference between 7\\(^\\circ\\) and 10\\(^\\circ\\). In short, addition and subtraction are meaningful for interval scale variables.8\n\nHowever, notice that the 0\\(^\\circ\\) does not mean “no temperature at all”: it actually means “the temperature at which water freezes”, which is pretty arbitrary. As a consequence, it becomes pointless to try to multiply and divide temperatures. It is wrong to say that \\(20^\\circ\\) is twice as hot as 10\\(^\\circ\\), just as it is weird and meaningless to try to claim that 20\\(^\\circ\\) is negative two times as hot as -10\\(^\\circ\\).\n\nAgain, lets look at a more psychological example. Suppose I’m interested in looking at how the attitudes of first-year university students have changed over time. Obviously, I’m going to want to record the year in which each student started. This is an interval scale variable. A student who started in 2003 did arrive 5 years before a student who started in 2008. However, it would be completely insane for me to divide 2008 by 2003 and say that the second student started “1.0024 times later” than the first one. That doesn’t make any sense at all.\n\n### 2.2.4 Ratio scale\n\nThe fourth and final type of variable to consider is a ratio scale variable, in which zero really means zero, and it’s okay to multiply and divide. A good psychological example of a ratio scale variable is response time (RT). In a lot of tasks it’s very common to record the amount of time somebody takes to solve a problem or answer a question, because it’s an indicator of how difficult the task is. Suppose that Alan takes 2.3 seconds to respond to a question, whereas Ben takes 3.1 seconds. As with an interval scale variable, addition and subtraction are both meaningful here. Ben really did take 3.1 - 2.3 = 0.8 seconds longer than Alan did. However, notice that multiplication and division also make sense here too: Ben took 3.1 / 2.3 = 1.35 times as long as Alan did to answer the question. And the reason why you can do this is that, for a ratio scale variable such as RT, “zero seconds” really does mean “no time at all”.\n\n### 2.2.5 Continuous versus discrete variables\n\nThere’s a second kind of distinction that you need to be aware of, regarding what types of variables you can run into. This is the distinction between continuous variables and discrete variables. The difference between these is as follows:\n\n* A continuous variable is one in which, for any two values that you can think of, it’s always logically possible to have another value in between.\n * A discrete variable is, in effect, a variable that isn’t continuous. For a discrete variable, it’s sometimes the case that there’s nothing in the middle.\n\nThese definitions probably seem a bit abstract, but they’re pretty simple once you see some examples. For instance, response time is continuous. If Alan takes 3.1 seconds and Ben takes 2.3 seconds to respond to a question, then it’s possible for Cameron’s response time to lie in between, by taking 3.0 seconds. And of course it would also be possible for David to take 3.031 seconds to respond, meaning that his RT would lie in between Cameron’s and Alan’s. And while in practice it might be impossible to measure RT that precisely, it’s certainly possible in principle. Because we can always find a new value for RT in between any two other ones, we say that RT is continuous.\n\nDiscrete variables occur when this rule is violated. For example, nominal scale variables are always discrete: there isn’t a type of transportation that falls “in between” trains and bicycles, not in the strict mathematical way that 2.3 falls in between 2 and 3. So transportation type is discrete. Similarly, ordinal scale variables are always discrete: although “2nd place” does fall between “1st place” and “3rd place”, there’s nothing that can logically fall in between “1st place” and “2nd place”. Interval scale and ratio scale variables can go either way. As we saw above, response time (a ratio scale variable) is continuous. Temperature in degrees celsius (an interval scale variable) is also continuous. However, the year you went to school (an interval scale variable) is discrete. There’s no year in between 2002 and 2003. The number of questions you get right on a true-or-false test (a ratio scale variable) is also discrete: since a true-or-false question doesn’t allow you to be “partially correct”, there’s nothing in between 5/10 and 6/10. Table 2.1 summarises the relationship between the scales of measurement and the discrete/continuity distinction. Cells with a tick mark correspond to things that are possible. I’m trying to hammer this point home, because (a) some textbooks get this wrong, and (b) people very often say things like “discrete variable” when they mean “nominal scale variable”. It’s very unfortunate.\n\ncontinuous | discrete |\n | --- | --- |\nnominal | \\(\\checkmark\\) |\nordinal | \\(\\checkmark\\) |\ninterval | \\(\\checkmark\\) | \\(\\checkmark\\) |\nratio | \\(\\checkmark\\) | \\(\\checkmark\\) |\n\n### 2.2.6 Some complexities\n\nOkay, I know you’re going to be shocked to hear this, but … the real world is much messier than this little classification scheme suggests. Very few variables in real life actually fall into these nice neat categories, so you need to be kind of careful not to treat the scales of measurement as if they were hard and fast rules. It doesn’t work like that: they’re guidelines, intended to help you think about the situations in which you should treat different variables differently. Nothing more.\n\nSo let’s take a classic example, maybe the classic example, of a psychological measurement tool: the Likert scale. The humble Likert scale is the bread and butter tool of all survey design. You yourself have filled out hundreds, maybe thousands of them, and odds are you’ve even used one yourself. Suppose we have a survey question that looks like this:\n Which of the following best describes your opinion of the statement that “all pirates are freaking awesome” …\n\nand then the options presented to the participant are these:\n * Strongly disagree\n * Disagree\n * Neither agree nor disagree\n * Agree\n * Strongly agree\n\nThis set of items is an example of a 5-point Likert scale: people are asked to choose among one of several (in this case 5) clearly ordered possibilities, generally with a verbal descriptor given in each case. However, it’s not necessary that all items be explicitly described. This is a perfectly good example of a 5-point Likert scale too:\n * Strongly disagree\n * Strongly agree\n\nLikert scales are very handy, if somewhat limited, tools. The question is, what kind of variable are they? They’re obviously discrete, since you can’t give a response of 2.5. They’re obviously not nominal scale, since the items are ordered; and they’re not ratio scale either, since there’s no natural zero.\n\nBut are they ordinal scale or interval scale? One argument says that we can’t really prove that the difference between “strongly agree” and “agree” is of the same size as the difference between “agree” and “neither agree nor disagree”. In fact, in everyday life it’s pretty obvious that they’re not the same at all. So this suggests that we ought to treat Likert scales as ordinal variables. On the other hand, in practice most participants do seem to take the whole “on a scale from 1 to 5” part fairly seriously, and they tend to act as if the differences between the five response options were fairly similar to one another. As a consequence, a lot of researchers treat Likert scale data as if it were interval scale. It’s not interval scale, but in practice it’s close enough that we usually think of it as being quasi-interval scale.\n\n## 2.3 Assessing the reliability of a measurement\n\nAt this point we’ve thought a little bit about how to operationalise a theoretical construct and thereby create a psychological measure; and we’ve seen that by applying psychological measures we end up with variables, which can come in many different types. At this point, we should start discussing the obvious question: is the measurement any good? We’ll do this in terms of two related ideas: reliability and validity. Put simply, the reliability of a measure tells you how precisely you are measuring something, whereas the validity of a measure tells you how accurate the measure is. In this section I’ll talk about reliability; we’ll talk about validity in the next chapter.\n\nReliability is actually a very simple concept: it refers to the repeatability or consistency of your measurement. The measurement of my weight by means of a “bathroom scale” is very reliable: if I step on and off the scales over and over again, it’ll keep giving me the same answer. Measuring my intelligence by means of “asking my mum” is very unreliable: some days she tells me I’m a bit thick, and other days she tells me I’m a complete moron. Notice that this concept of reliability is different to the question of whether the measurements are correct (the correctness of a measurement relates to it’s validity). If I’m holding a sack of potatos when I step on and off of the bathroom scales, the measurement will still be reliable: it will always give me the same answer. However, this highly reliable answer doesn’t match up to my true weight at all, therefore it’s wrong. In technical terms, this is a reliable but invalid measurement. Similarly, while my mum’s estimate of my intelligence is a bit unreliable, she might be right. Maybe I’m just not too bright, and so while her estimate of my intelligence fluctuates pretty wildly from day to day, it’s basically right. So that would be an unreliable but valid measure. Of course, to some extent, notice that if my mum’s estimates are too unreliable, it’s going to be very hard to figure out which one of her many claims about my intelligence is actually the right one. To some extent, then, a very unreliable measure tends to end up being invalid for practical purposes; so much so that many people would say that reliability is necessary (but not sufficient) to ensure validity.\n\nOkay, now that we’re clear on the distinction between reliability and validity, let’s have a think about the different ways in which we might measure reliability:\n\n* Test-retest reliability. This relates to consistency over time: if we repeat the measurement at a later date, do we get a the same answer?\n * Inter-rater reliability. This relates to consistency across people: if someone else repeats the measurement (e.g., someone else rates my intelligence) will they produce the same answer?\n * Parallel forms reliability. This relates to consistency across theoretically-equivalent measurements: if I use a different set of bathroom scales to measure my weight, does it give the same answer?\n * Internal consistency reliability. If a measurement is constructed from lots of different parts that perform similar functions (e.g., a personality questionnaire result is added up across several questions) do the individual parts tend to give similar answers.\n\nNot all measurements need to possess all forms of reliability. For instance, educational assessment can be thought of as a form of measurement. One of the subjects that I teach, Computational Cognitive Science, has an assessment structure that has a research component and an exam component (plus other things). The exam component is intended to measure something different from the research component, so the assessment as a whole has low internal consistency. However, within the exam there are several questions that are intended to (approximately) measure the same things, and those tend to produce similar outcomes; so the exam on its own has a fairly high internal consistency. Which is as it should be. You should only demand reliability in those situations where you want to be measure the same thing!\n\n## 2.4 The “role” of variables: predictors and outcomes\n\nOkay, I’ve got one last piece of terminology that I need to explain to you before moving away from variables. Normally, when we do some research we end up with lots of different variables. Then, when we analyse our data we usually try to explain some of the variables in terms of some of the other variables. It’s important to keep the two roles “thing doing the explaining” and “thing being explained” distinct. So let’s be clear about this now. Firstly, we might as well get used to the idea of using mathematical symbols to describe variables, since it’s going to happen over and over again. Let’s denote the “to be explained” variable \\(Y\\), and denote the variables “doing the explaining” as \\(X_1\\), \\(X_2\\), etc.\n\nNow, when we doing an analysis, we have different names for \\(X\\) and \\(Y\\), since they play different roles in the analysis. The classical names for these roles are independent variable (IV) and dependent variable (DV). The IV is the variable that you use to do the explaining (i.e., \\(X\\)) and the DV is the variable being explained (i.e., \\(Y\\)). The logic behind these names goes like this: if there really is a relationship between \\(X\\) and \\(Y\\) then we can say that \\(Y\\) depends on \\(X\\), and if we have designed our study “properly” then \\(X\\) isn’t dependent on anything else. However, I personally find those names horrible: they’re hard to remember and they’re highly misleading, because (a) the IV is never actually “independent of everything else” and (b) if there’s no relationship, then the DV doesn’t actually depend on the IV. And in fact, because I’m not the only person who thinks that IV and DV are just awful names, there are a number of alternatives that I find more appealing. The terms that I’ll use in these notes are predictors and outcomes. The idea here is that what you’re trying to do is use \\(X\\) (the predictors) to make guesses about \\(Y\\) (the outcomes).9 This is summarised in Table 2.2.\n\nrole of the variable | classical name | modern name |\n | --- | --- | --- |\nto be explained | dependent variable (DV) | outcome |\nto do the explaining | independent variable (IV) | predictor |\n\n## 2.5 Experimental and non-experimental research\n\nOne of the big distinctions that you should be aware of is the distinction between “experimental research” and “non-experimental research”. When we make this distinction, what we’re really talking about is the degree of control that the researcher exercises over the people and events in the study.\n\n### 2.5.1 Experimental research\n\nThe key features of experimental research is that the researcher controls all aspects of the study, especially what participants experience during the study. In particular, the researcher manipulates or varies the predictor variables (IVs), and then allows the outcome variable (DV) to vary naturally. The idea here is to deliberately vary the predictors (IVs) to see if they have any causal effects on the outcomes. Moreover, in order to ensure that there’s no chance that something other than the predictor variables is causing the outcomes, everything else is kept constant or is in some other way “balanced” to ensure that they have no effect on the results. In practice, it’s almost impossible to think of everything else that might have an influence on the outcome of an experiment, much less keep it constant. The standard solution to this is randomisation: that is, we randomly assign people to different groups, and then give each group a different treatment (i.e., assign them different values of the predictor variables). We’ll talk more about randomisation later in this course, but for now, it’s enough to say that what randomisation does is minimise (but not eliminate) the chances that there are any systematic difference between groups.\n\nLet’s consider a very simple, completely unrealistic and grossly unethical example. Suppose you wanted to find out if smoking causes lung cancer. One way to do this would be to find people who smoke and people who don’t smoke, and look to see if smokers have a higher rate of lung cancer. This is not a proper experiment, since the researcher doesn’t have a lot of control over who is and isn’t a smoker. And this really matters: for instance, it might be that people who choose to smoke cigarettes also tend to have poor diets, or maybe they tend to work in asbestos mines, or whatever. The point here is that the groups (smokers and non-smokers) actually differ on lots of things, not just smoking. So it might be that the higher incidence of lung cancer among smokers is caused by something else, not by smoking per se. In technical terms, these other things (e.g. diet) are called “confounds”, and we’ll talk about those in just a moment.\n\nIn the meantime, let’s now consider what a proper experiment might look like. Recall that our concern was that smokers and non-smokers might differ in lots of ways. The solution, as long as you have no ethics, is to control who smokes and who doesn’t. Specifically, if we randomly divide participants into two groups, and force half of them to become smokers, then it’s very unlikely that the groups will differ in any respect other than the fact that half of them smoke. That way, if our smoking group gets cancer at a higher rate than the non-smoking group, then we can feel pretty confident that (a) smoking does cause cancer and (b) we’re murderers.\n\n### 2.5.2 Non-experimental research\n\nNon-experimental research is a broad term that covers “any study in which the researcher doesn’t have quite as much control as they do in an experiment”. Obviously, control is something that scientists like to have, but as the previous example illustrates, there are lots of situations in which you can’t or shouldn’t try to obtain that control. Since it’s grossly unethical (and almost certainly criminal) to force people to smoke in order to find out if they get cancer, this is a good example of a situation in which you really shouldn’t try to obtain experimental control. But there are other reasons too. Even leaving aside the ethical issues, our “smoking experiment” does have a few other issues. For instance, when I suggested that we “force” half of the people to become smokers, I must have been talking about starting with a sample of non-smokers, and then forcing them to become smokers. While this sounds like the kind of solid, evil experimental design that a mad scientist would love, it might not be a very sound way of investigating the effect in the real world. For instance, suppose that smoking only causes lung cancer when people have poor diets, and suppose also that people who normally smoke do tend to have poor diets. However, since the “smokers” in our experiment aren’t “natural” smokers (i.e., we forced non-smokers to become smokers; they didn’t take on all of the other normal, real life characteristics that smokers might tend to possess) they probably have better diets. As such, in this silly example they wouldn’t get lung cancer, and our experiment will fail, because it violates the structure of the “natural” world (the technical name for this is an “artifactual” result; see later).\n\nOne distinction worth making between two types of non-experimental research is the difference between quasi-experimental research and case studies. The example I discussed earlier – in which we wanted to examine incidence of lung cancer among smokers and non-smokers, without trying to control who smokes and who doesn’t – is a quasi-experimental design. That is, it’s the same as an experiment, but we don’t control the predictors (IVs). We can still use statistics to analyse the results, it’s just that we have to be a lot more careful.\n\nThe alternative approach, case studies, aims to provide a very detailed description of one or a few instances. In general, you can’t use statistics to analyse the results of case studies, and it’s usually very hard to draw any general conclusions about “people in general” from a few isolated examples. However, case studies are very useful in some situations. Firstly, there are situations where you don’t have any alternative: neuropsychology has this issue a lot. Sometimes, you just can’t find a lot of people with brain damage in a specific area, so the only thing you can do is describe those cases that you do have in as much detail and with as much care as you can. However, there’s also some genuine advantages to case studies: because you don’t have as many people to study, you have the ability to invest lots of time and effort trying to understand the specific factors at play in each case. This is a very valuable thing to do. As a consequence, case studies can complement the more statistically-oriented approaches that you see in experimental and quasi-experimental designs. We won’t talk much about case studies in these lectures, but they are nevertheless very valuable tools!\n\n## 2.6 Assessing the validity of a study\n\nMore than any other thing, a scientist wants their research to be “valid”. The conceptual idea behind validity is very simple: can you trust the results of your study? If not, the study is invalid. However, while it’s easy to state, in practice it’s much harder to check validity than it is to check reliability. And in all honesty, there’s no precise, clearly agreed upon notion of what validity actually is. In fact, there’s lots of different kinds of validity, each of which raises it’s own issues, and not all forms of validity are relevant to all studies. I’m going to talk about five different types:\n\n* Internal validity\n * External validity\n * Construct validity\n * Face validity\n * Ecological validity\n\nTo give you a quick guide as to what matters here… (1) Internal and external validity are the most important, since they tie directly to the fundamental question of whether your study really works. (2) Construct validity asks whether you’re measuring what you think you are. (3) Face validity isn’t terribly important except insofar as you care about “appearances”. (4) Ecological validity is a special case of face validity that corresponds to a kind of appearance that you might care about a lot.\n\n### 2.6.1 Internal validity\n\nInternal validity refers to the extent to which you are able draw the correct conclusions about the causal relationships between variables. It’s called “internal” because it refers to the relationships between things “inside” the study. Let’s illustrate the concept with a simple example. Suppose you’re interested in finding out whether a university education makes you write better. To do so, you get a group of first year students, ask them to write a 1000 word essay, and count the number of spelling and grammatical errors they make. Then you find some third-year students, who obviously have had more of a university education than the first-years, and repeat the exercise. And let’s suppose it turns out that the third-year students produce fewer errors. And so you conclude that a university education improves writing skills. Right? Except… the big problem that you have with this experiment is that the third-year students are older, and they’ve had more experience with writing things. So it’s hard to know for sure what the causal relationship is: Do older people write better? Or people who have had more writing experience? Or people who have had more education? Which of the above is the true cause of the superior performance of the third-years? Age? Experience? Education? You can’t tell. This is an example of a failure of internal validity, because your study doesn’t properly tease apart the causal relationships between the different variables.\n\n### 2.6.2 External validity\n\nExternal validity relates to the generalisability of your findings. That is, to what extent do you expect to see the same pattern of results in “real life” as you saw in your study. To put it a bit more precisely, any study that you do in psychology will involve a fairly specific set of questions or tasks, will occur in a specific environment, and will involve participants that are drawn from a particular subgroup. So, if it turns out that the results don’t actually generalise to people and situations beyond the ones that you studied, then what you’ve got is a lack of external validity.\n\nThe classic example of this issue is the fact that a very large proportion of studies in psychology will use undergraduate psychology students as the participants. Obviously, however, the researchers don’t care only about psychology students; they care about people in general. Given that, a study that uses only psych students as participants always carries a risk of lacking external validity. That is, if there’s something “special” about psychology students that makes them different to the general populace in some relevant respect, then we may start worrying about a lack of external validity.\n\nThat said, it is absolutely critical to realise that a study that uses only psychology students does not necessarily have a problem with external validity. I’ll talk about this again later, but it’s such a common mistake that I’m going to mention it here. The external validity is threatened by the choice of population if (a) the population from which you sample your participants is very narrow (e.g., psych students), and (b) the narrow population that you sampled from is systematically different from the general population, in some respect that is relevant to the psychological phenomenon that you intend to study. The italicised part is the bit that lots of people forget: it is true that psychology undergraduates differ from the general population in lots of ways, and so a study that uses only psych students may have problems with external validity. However, if those differences aren’t very relevant to the phenomenon that you’re studying, then there’s nothing to worry about. To make this a bit more concrete, here’s two extreme examples:\n\n* You want to measure “attitudes of the general public towards psychotherapy”, but all of your participants are psychology students. This study would almost certainly have a problem with external validity.\n * You want to measure the effectiveness of a visual illusion, and your participants are all psychology students. This study is very unlikely to have a problem with external validity\n\nHaving just spent the last couple of paragraphs focusing on the choice of participants (since that’s the big issue that everyone tends to worry most about), it’s worth remembering that external validity is a broader concept. The following are also examples of things that might pose a threat to external validity, depending on what kind of study you’re doing:\n\n* People might answer a “psychology questionnaire” in a manner that doesn’t reflect what they would do in real life.\n * Your lab experiment on (say) “human learning” has a different structure to the learning problems people face in real life.\n\n### 2.6.3 Construct validity\n\nConstruct validity is basically a question of whether you’re measuring what you want to be measuring. A measurement has good construct validity if it is actually measuring the correct theoretical construct, and bad construct validity if it doesn’t. To give very simple (if ridiculous) example, suppose I’m trying to investigate the rates with which university students cheat on their exams. And the way I attempt to measure it is by asking the cheating students to stand up in the lecture theatre so that I can count them. When I do this with a class of 300 students, 0 people claim to be cheaters. So I therefore conclude that the proportion of cheaters in my class is 0%. Clearly this is a bit ridiculous. But the point here is not that this is a very deep methodological example, but rather to explain what construct validity is. The problem with my measure is that while I’m trying to measure “the proportion of people who cheat” what I’m actually measuring is “the proportion of people stupid enough to own up to cheating, or bloody minded enough to pretend that they do”. Obviously, these aren’t the same thing! So my study has gone wrong, because my measurement has very poor construct validity.\n\n### 2.6.4 Face validity\n\nFace validity simply refers to whether or not a measure “looks like” it’s doing what it’s supposed to, nothing more. If I design a test of intelligence, and people look at it and they say “no, that test doesn’t measure intelligence”, then the measure lacks face validity. It’s as simple as that. Obviously, face validity isn’t very important from a pure scientific perspective. After all, what we care about is whether or not the measure actually does what it’s supposed to do, not whether it looks like it does what it’s supposed to do. As a consequence, we generally don’t care very much about face validity. That said, the concept of face validity serves three useful pragmatic purposes:\n\n* Sometimes, an experienced scientist will have a “hunch” that a particular measure won’t work. While these sorts of hunches have no strict evidentiary value, it’s often worth paying attention to them. Because often times people have knowledge that they can’t quite verbalise, so there might be something to worry about even if you can’t quite say why. In other words, when someone you trust criticises the face validity of your study, it’s worth taking the time to think more carefully about your design to see if you can think of reasons why it might go awry. Mind you, if you don’t find any reason for concern, then you should probably not worry: after all, face validity really doesn’t matter much.\n * Often (very often), completely uninformed people will also have a “hunch” that your research is crap. And they’ll criticise it on the internet or something. On close inspection, you’ll often notice that these criticisms are actually focused entirely on how the study “looks”, but not on anything deeper. The concept of face validity is useful for gently explaining to people that they need to substantiate their arguments further.\n * Expanding on the last point, if the beliefs of untrained people are critical (e.g., this is often the case for applied research where you actually want to convince policy makers of something or other) then you have to care about face validity. Simply because – whether you like it or not – a lot of people will use face validity as a proxy for real validity. If you want the government to change a law on scientific, psychological grounds, then it won’t matter how good your studies “really” are. If they lack face validity, you’ll find that politicians ignore you. Of course, it’s somewhat unfair that policy often depends more on appearance than fact, but that’s how things go.\n\n### 2.6.5 Ecological validity\n\nEcological validity is a different notion of validity, which is similar to external validity, but less important. The idea is that, in order to be ecologically valid, the entire set up of the study should closely approximate the real world scenario that is being investigated. In a sense, ecological validity is a kind of face validity – it relates mostly to whether the study “looks” right, but with a bit more rigour to it. To be ecologically valid, the study has to look right in a fairly specific way. The idea behind it is the intuition that a study that is ecologically valid is more likely to be externally valid. It’s no guarantee, of course. But the nice thing about ecological validity is that it’s much easier to check whether a study is ecologically valid than it is to check whether a study is externally valid. An simple example would be eyewitness identification studies. Most of these studies tend to be done in a university setting, often with fairly simple array of faces to look at rather than a line up. The length of time between seeing the “criminal” and being asked to identify the suspect in the “line up” is usually shorter. The “crime” isn’t real, so there’s no chance that the witness being scared, and there’s no police officers present, so there’s not as much chance of feeling pressured. These things all mean that the study definitely lacks ecological validity. They might (but might not) mean that it also lacks external validity.\n\n## 2.7 Confounds, artifacts and other threats to validity\n\nIf we look at the issue of validity in the most general fashion, the two biggest worries that we have are confounds and artifact. These two terms are defined in the following way:\n\n* Confound: A confound is an additional, often unmeasured variable10 that turns out to be related to both the predictors and the outcomes. The existence of confounds threatens the internal validity of the study because you can’t tell whether the predictor causes the outcome, or if the confounding variable causes it, etc.\n * Artifact: A result is said to be “artifactual” if it only holds in the special situation that you happened to test in your study. The possibility that your result is an artifact describes a threat to your external validity, because it raises the possibility that you can’t generalise your results to the actual population that you care about.\n\nAs a general rule confounds are a bigger concern for non-experimental studies, precisely because they’re not proper experiments: by definition, you’re leaving lots of things uncontrolled, so there’s a lot of scope for confounds working their way into your study. Experimental research tends to be much less vulnerable to confounds: the more control you have over what happens during the study, the more you can prevent confounds from appearing.\n\nHowever, there’s always swings and roundabouts, and when we start thinking about artifacts rather than confounds, the shoe is very firmly on the other foot. For the most part, artifactual results tend to be a concern for experimental studies than for non-experimental studies. To see this, it helps to realise that the reason that a lot of studies are non-experimental is precisely because what the researcher is trying to do is examine human behaviour in a more naturalistic context. By working in a more real-world context, you lose experimental control (making yourself vulnerable to confounds) but because you tend to be studying human psychology “in the wild” you reduce the chances of getting an artifactual result. Or, to put it another way, when you take psychology out of the wild and bring it into the lab (which we usually have to do to gain our experimental control), you always run the risk of accidentally studying something different than you wanted to study: which is more or less the definition of an artifact.\n\nBe warned though: the above is a rough guide only. It’s absolutely possible to have confounds in an experiment, and to get artifactual results with non-experimental studies. This can happen for all sorts of reasons, not least of which is researcher error. In practice, it’s really hard to think everything through ahead of time, and even very good researchers make mistakes. But other times it’s unavoidable, simply because the researcher has ethics (e.g., see 2.7.5).\n\nOkay. There’s a sense in which almost any threat to validity can be characterised as a confound or an artifact: they’re pretty vague concepts. So let’s have a look at some of the most common examples…\n\n### 2.7.1 History effects\n\nHistory effects refer to the possibility that specific events may occur during the study itself that might influence the outcomes. For instance, something might happen in between a pre-test and a post-test. Or, in between testing participant 23 and participant 24. Alternatively, it might be that you’re looking at an older study, which was perfectly valid for its time, but the world has changed enough since then that the conclusions are no longer trustworthy. Examples of things that would count as history effects:\n\n* You’re interested in how people think about risk and uncertainty. You started your data collection in December 2010. But finding participants and collecting data takes time, so you’re still finding new people in February 2011. Unfortunately for you (and even more unfortunately for others), the Queensland floods occurred in January 2011, causing billions of dollars of damage and killing many people. Not surprisingly, the people tested in February 2011 express quite different beliefs about handling risk than the people tested in December 2010. Which (if any) of these reflects the “true” beliefs of participants? I think the answer is probably both: the Queensland floods genuinely changed the beliefs of the Australian public, though possibly only temporarily. The key thing here is that the “history” of the people tested in February is quite different to people tested in December.\n * You’re testing the psychological effects of a new anti-anxiety drug. So what you do is measure anxiety before administering the drug (e.g., by self-report, and taking physiological measures, let’s say), then you administer the drug, and then you take the same measures afterwards. In the middle, however, because your labs are in Los Angeles, there’s an earthquake, which increases the anxiety of the participants.\n\n### 2.7.2 Maturation effects\n\nAs with history effects, maturational effects are fundamentally about change over time. However, maturation effects aren’t in response to specific events. Rather, they relate to how people change on their own over time: we get older, we get tired, we get bored, etc. Some examples of maturation effects:\n\n* When doing developmental psychology research, you need to be aware that children grow up quite rapidly. So, suppose that you want to find out whether some educational trick helps with vocabulary size among 3 year olds. One thing that you need to be aware of is that the vocabulary size of children that age is growing at an incredible rate (multiple words per day), all on its own. If you design your study without taking this maturational effect into account, then you won’t be able to tell if your educational trick works.\n * When running a very long experiment in the lab (say, something that goes for 3 hours), it’s very likely that people will begin to get bored and tired, and that this maturational effect will cause performance to decline, regardless of anything else going on in the experiment\n\n### 2.7.3 Repeated testing effects\n\nAn important type of history effect is the effect of repeated testing. Suppose I want to take two measurements of some psychological construct (e.g., anxiety). One thing I might be worried about is if the first measurement has an effect on the second measurement. In other words, this is a history effect in which the “event” that influences the second measurement is the first measurement itself! This is not at all uncommon. Examples of this include:\n\n* Learning and practice: e.g., “intelligence” at time 2 might appear to go up relative to time 1 because participants learned the general rules of how to solve “intelligence-test-style” questions during the first testing session.\n * Familiarity with the testing situation: e.g., if people are nervous at time 1, this might make performance go down; after sitting through the first testing situation, they might calm down a lot precisely because they’ve seen what the testing looks like.\n * Auxiliary changes caused by testing: e.g., if a questionnaire assessing mood is boring, then mood at measurement at time 2 is more likely to become “bored”, precisely because of the boring measurement made at time 1.\n\n### 2.7.4 Selection bias\n\nSelection bias is a pretty broad term. Suppose that you’re running an experiment with two groups of participants, where each group gets a different “treatment”, and you want to see if the different treatments lead to different outcomes. However, suppose that, despite your best efforts, you’ve ended up with a gender imbalance across groups (say, group A has 80% females and group B has 50% females). It might sound like this could never happen, but trust me, it can. This is an example of a selection bias, in which the people “selected into” the two groups have different characteristics. If any of those characteristics turns out to be relevant (say, your treatment works better on females than males) then you’re in a lot of trouble.\n\n### 2.7.5 Differential attrition\n\nOne quite subtle danger to be aware of is called differential attrition, which is a kind of selection bias that is caused by the study itself. Suppose that, for the first time ever in the history of psychology, I manage to find the perfectly balanced and representative sample of people. I start running “Dan’s incredibly long and tedious experiment” on my perfect sample, but then, because my study is incredibly long and tedious, lots of people start dropping out. I can’t stop this: as we’ll discuss later in the chapter on research ethics, participants absolutely have the right to stop doing any experiment, any time, for whatever reason they feel like, and as researchers we are morally (and professionally) obliged to remind people that they do have this right. So, suppose that “Dan’s incredibly long and tedious experiment” has a very high drop out rate. What do you suppose the odds are that this drop out is random? Answer: zero. Almost certainly, the people who remain are more conscientious, more tolerant of boredom etc than those that leave. To the extent that (say) conscientiousness is relevant to the psychological phenomenon that I care about, this attrition can decrease the validity of my results.\n\nWhen thinking about the effects of differential attrition, it is sometimes helpful to distinguish between two different types. The first is homogeneous attrition, in which the attrition effect is the same for all groups, treatments or conditions. In the example I gave above, the differential attrition would be homogeneous if (and only if) the easily bored participants are dropping out of all of the conditions in my experiment at about the same rate. In general, the main effect of homogeneous attrition is likely to be that it makes your sample unrepresentative. As such, the biggest worry that you’ll have is that the generalisability of the results decreases: in other words, you lose external validity.\n\nThe second type of differential attrition is heterogeneous attrition, in which the attrition effect is different for different groups. This is a much bigger problem: not only do you have to worry about your external validity, you also have to worry about your internal validity too. To see why this is the case, let’s consider a very dumb study in which I want to see if insulting people makes them act in a more obedient way. Why anyone would actually want to study that I don’t know, but let’s suppose I really, deeply cared about this. So, I design my experiment with two conditions. In the “treatment” condition, the experimenter insults the participant and then gives them a questionnaire designed to measure obedience. In the “control” condition, the experimenter engages in a bit of pointless chitchat and then gives them the questionnaire. Leaving aside the questionable scientific merits and dubious ethics of such a study, let’s have a think about what might go wrong here. As a general rule, when someone insults me to my face, I tend to get much less co-operative. So, there’s a pretty good chance that a lot more people are going to drop out of the treatment condition than the control condition. And this drop out isn’t going to be random. The people most likely to drop out would probably be the people who don’t care all that much about the importance of obediently sitting through the experiment. Since the most bloody minded and disobedient people all left the treatment group but not the control group, we’ve introduced a confound: the people who actually took the questionnaire in the treatment group were already more likely to be dutiful and obedient than the people in the control group. In short, in this study insulting people doesn’t make them more obedient: it makes the more disobedient people leave the experiment! The internal validity of this experiment is completely shot.\n\n### 2.7.6 Non-response bias\n\nNon-response bias is closely related to selection bias, and to differential attrition. The simplest version of the problem goes like this. You mail out a survey to 1000 people, and only 300 of them reply. The 300 people who replied are almost certainly not a random subsample. People who respond to surveys are systematically different to people who don’t. This introduces a problem when trying to generalise from those 300 people who replied, to the population at large; since you now have a very non-random sample. The issue of non-response bias is more general than this, though. Among the (say) 300 people that did respond to the survey, you might find that not everyone answers every question. If (say) 80 people chose not to answer one of your questions, does this introduce problems? As always, the answer is maybe. If the question that wasn’t answered was on the last page of the questionnaire, and those 80 surveys were returned with the last page missing, there’s a good chance that the missing data isn’t a big deal: probably the pages just fell off. However, if the question that 80 people didn’t answer was the most confrontational or invasive personal question in the questionnaire, then almost certainly you’ve got a problem. In essence, what you’re dealing with here is what’s called the problem of missing data. If the data that is missing was “lost” randomly, then it’s not a big problem. If it’s missing systematically, then it can be a big problem.\n\n### 2.7.7 Regression to the mean\n\nRegression to the mean is a curious variation on selection bias. It refers to any situation where you select data based on an extreme value on some measure. Because the measure has natural variation, it almost certainly means that when you take a subsequent measurement, that later measurement will be less extreme than the first one, purely by chance.\n\nHere’s an example. Suppose I’m interested in whether a psychology education has an adverse effect on very smart kids. To do this, I find the 20 psych I students with the best high school grades and look at how well they’re doing at university. It turns out that they’re doing a lot better than average, but they’re not topping the class at university, even though they did top their classes at high school. What’s going on? The natural first thought is that this must mean that the psychology classes must be having an adverse effect on those students. However, while that might very well be the explanation, it’s more likely that what you’re seeing is an example of “regression to the mean”. To see how it works, let’s take a moment to think about what is required to get the best mark in a class, regardless of whether that class be at high school or at university. When you’ve got a big class, there are going to be lots of very smart people enrolled. To get the best mark you have to be very smart, work very hard, and be a bit lucky. The exam has to ask just the right questions for your idiosyncratic skills, and you have to not make any dumb mistakes (we all do that sometimes) when answering them. And that’s the thing: intelligence and hard work are transferrable from one class to the next. Luck isn’t. The people who got lucky in high school won’t be the same as the people who get lucky at university. That’s the very definition of “luck”. The consequence of this is that, when you select people at the very extreme values of one measurement (the top 20 students), you’re selecting for hard work, skill and luck. But because the luck doesn’t transfer to the second measurement (only the skill and work), these people will all be expected to drop a little bit when you measure them a second time (at university). So their scores fall back a little bit, back towards everyone else. This is regression to the mean.\n\nRegression to the mean is surprisingly common. For instance, if two very tall people have kids, their children will tend to be taller than average, but not as tall as the parents. The reverse happens with very short parents: two very short parents will tend to have short children, but nevertheless those kids will tend to be taller than the parents. It can also be extremely subtle. For instance, there have been studies done that suggested that people learn better from negative feedback than from positive feedback. However, the way that people tried to show this was to give people positive reinforcement whenever they did good, and negative reinforcement when they did bad. And what you see is that after the positive reinforcement, people tended to do worse; but after the negative reinforcement they tended to do better. But! Notice that there’s a selection bias here: when people do very well, you’re selecting for “high” values, and so you should expect (because of regression to the mean) that performance on the next trial should be worse, regardless of whether reinforcement is given. Similarly, after a bad trial, people will tend to improve all on their own. The apparent superiority of negative feedback is an artifact caused by regression to the mean (see Kahneman and Tversky 1973 for discussion).\n\n### 2.7.8 Experimenter bias\n\nExperimenter bias can come in multiple forms. The basic idea is that the experimenter, despite the best of intentions, can accidentally end up influencing the results of the experiment by subtly communicating the “right answer” or the “desired behaviour” to the participants. Typically, this occurs because the experimenter has special knowledge that the participant does not – either the right answer to the questions being asked, or knowledge of the expected pattern of performance for the condition that the participant is in, and so on. The classic example of this happening is the case study of “”, which dates back to 1907 (Pfungst 1911; Hothersall 2004). was a horse that apparently was able to read and count, and perform other human like feats of intelligence. After became famous, psychologists started examining his behaviour more closely. It turned out that – not surprisingly – Hans didn’t know how to do maths. Rather, Hans was responding to the human observers around him. Because they did know how to count, and the horse had learned to change its behaviour when people changed theirs.\n\nThe general solution to the problem of experimenter bias is to engage in double blind studies, where neither the experimenter nor the participant knows which condition the participant is in, or knows what the desired behaviour is. This provides a very good solution to the problem, but it’s important to recognise that it’s not quite ideal, and hard to pull off perfectly. For instance, the obvious way that I could try to construct a double blind study is to have one of my Ph.D. students (one who doesn’t know anything about the experiment) run the study. That feels like it should be enough. The only person (me) who knows all the details (e.g., correct answers to the questions, assignments of participants to conditions) has no interaction with the participants, and the person who does all the talking to people (the Ph.D. student) doesn’t know anything. Except, that last part is very unlikely to be true. In order for the Ph.D. student to run the study effectively, they need to have been briefed by me, the researcher. And, as it happens, the Ph.D. student also knows me, and knows a bit about my general beliefs about people and psychology (e.g., I tend to think humans are much smarter than psychologists give them credit for). As a result of all this, it’s almost impossible for the experimenter to avoid knowing a little bit about what expectations I have. And even a little bit of knowledge can have an effect: suppose the experimenter accidentally conveys the fact that the participants are expected to do well in this task. Well, there’s a thing called the “Pygmalion effect”: if you expect great things of people, they’ll rise to the occasion; but if you expect them to fail, they’ll do that too. In other words, the expectations become a self-fulfilling prophesy.\n\n### 2.7.9 Demand effects and reactivity\n\nWhen talking about experimenter bias, the worry is that the experimenter’s knowledge or desires for the experiment are communicated to the participants, and that these effect people’s behaviour (Rosenthal 1966). However, even if you manage to stop this from happening, it’s almost impossible to stop people from knowing that they’re part of a psychological study. And the mere fact of knowing that someone is watching/studying you can have a pretty big effect on behaviour. This is generally referred to as reactivity or demand effects. The basic idea is captured by the Hawthorne effect: people alter their performance because of the attention that the study focuses on them. The effect takes its name from a the “Hawthorne Works” factory outside of Chicago (see Adair 1984). A study done in the 1920s looking at the effects of lighting on worker productivity at the factory turned out to be an effect of the fact that the workers knew they were being studied, rather than the lighting.\n\nTo get a bit more specific about some of the ways in which the mere fact of being in a study can change how people behave, it helps to think like a social psychologist and look at some of the roles that people might adopt during an experiment, but might not adopt if the corresponding events were occurring in the real world:\n\n* The good participant tries to be too helpful to the researcher: he or she seeks to figure out the experimenter’s hypotheses and confirm them.\n * The negative participant does the exact opposite of the good participant: he or she seeks to break or destroy the study or the hypothesis in some way.\n * The faithful participant is unnaturally obedient: he or she seeks to follow instructions perfectly, regardless of what might have happened in a more realistic setting.\n * The apprehensive participant gets nervous about being tested or studied, so much so that his or her behaviour becomes highly unnatural, or overly socially desirable.\n\n### 2.7.10 Placebo effects\n\nThe placebo effect is a specific type of demand effect that we worry a lot about. It refers to the situation where the mere fact of being treated causes an improvement in outcomes. The classic example comes from clinical trials: if you give people a completely chemically inert drug and tell them that it’s a cure for a disease, they will tend to get better faster than people who aren’t treated at all. In other words, it is people’s belief that they are being treated that causes the improved outcomes, not the drug.\n\n### 2.7.11 Situation, measurement and subpopulation effects\n\nIn some respects, these terms are a catch-all term for “all other threats to external validity”. They refer to the fact that the choice of subpopulation from which you draw your participants, the location, timing and manner in which you run your study (including who collects the data) and the tools that you use to make your measurements might all be influencing the results. Specifically, the worry is that these things might be influencing the results in such a way that the results won’t generalise to a wider array of people, places and measures.\n\n### 2.7.12 Fraud, deception and self-deception\n\nIt is difficult to get a man to understand something, when his salary depends on his not understanding it.\n\n– \n\nOne final thing that I feel like I should mention. While reading what the textbooks often have to say about assessing the validity of the study, I couldn’t help but notice that they seem to make the assumption that the researcher is honest. I find this hilarious. While the vast majority of scientists are honest, in my experience at least, some are not.11 Not only that, as I mentioned earlier, scientists are not immune to belief bias – it’s easy for a researcher to end up deceiving themselves into believing the wrong thing, and this can lead them to conduct subtly flawed research, and then hide those flaws when they write it up. So you need to consider not only the (probably unlikely) possibility of outright fraud, but also the (probably quite common) possibility that the research is unintentionally “slanted”. I opened a few standard textbooks and didn’t find much of a discussion of this problem, so here’s my own attempt to list a few ways in which these issues can arise are:\n\n* Data fabrication. Sometimes, people just make up the data. This is occasionally done with “good” intentions. For instance, the researcher believes that the fabricated data do reflect the truth, and may actually reflect “slightly cleaned up” versions of actual data. On other occasions, the fraud is deliberate and malicious. Some high-profile examples where data fabrication has been alleged or shown include (a psychologist who is thought to have fabricated some of his data), (who has been accused of fabricating his data connecting the MMR vaccine to autism) and (who falsified a lot of his data on stem cell research).\n * Hoaxes. Hoaxes share a lot of similarities with data fabrication, but they differ in the intended purpose. A hoax is often a joke, and many of them are intended to be (eventually) discovered. Often, the point of a hoax is to discredit someone or some field. There’s quite a few well known scientific hoaxes that have occurred over the years (e.g., Piltdown man) some of were deliberate attempts to discredit particular fields of research (e.g., the Sokal affair).\n * Data misrepresentation. While fraud gets most of the headlines, it’s much more common in my experience to see data being misrepresented. When I say this, I’m not referring to newspapers getting it wrong (which they do, almost always). I’m referring to the fact that often, the data don’t actually say what the researchers think they say. My guess is that, almost always, this isn’t the result of deliberate dishonesty, it’s due to a lack of sophistication in the data analyses. For instance, think back to the example of Simpson’s paradox that I discussed in the beginning of these notes. It’s very common to see people present “aggregated” data of some kind; and sometimes, when you dig deeper and find the raw data yourself, you find that the aggregated data tell a different story to the disaggregated data. Alternatively, you might find that some aspect of the data is being hidden, because it tells an inconvenient story (e.g., the researcher might choose not to refer to a particular variable). There’s a lot of variants on this; many of which are very hard to detect.\n * Study “misdesign”. Okay, this one is subtle. Basically, the issue here is that a researcher designs a study that has built-in flaws, and those flaws are never reported in the paper. The data that are reported are completely real, and are correctly analysed, but they are produced by a study that is actually quite wrongly put together. The researcher really wants to find a particular effect, and so the study is set up in such a way as to make it “easy” to (artifactually) observe that effect. One sneaky way to do this – in case you’re feeling like dabbling in a bit of fraud yourself – is to design an experiment in which it’s obvious to the participants what they’re “supposed” to be doing, and then let reactivity work its magic for you. If you want, you can add all the trappings of double blind experimentation etc. It won’t make a difference, since the study materials themselves are subtly telling people what you want them to do. When you write up the results, the fraud won’t be obvious to the reader: what’s obvious to the participant when they’re in the experimental context isn’t always obvious to the person reading the paper. Of course, the way I’ve described this makes it sound like it’s always fraud: probably there are cases where this is done deliberately, but in my experience the bigger concern has been with unintentional misdesign. The researcher believes … and so the study just happens to end up with a built in flaw, and that flaw then magically erases itself when the study is written up for publication.\n * Data mining & post hoc hypothesising. Another way in which the authors of a study can more or less lie about what they found is by engaging in what’s referred to as “data mining”. As we’ll discuss later in the class, if you keep trying to analyse your data in lots of different ways, you’ll eventually find something that “looks” like a real effect but isn’t. This is referred to as “data mining”. It used to be quite rare because data analysis used to take weeks, but now that everyone has very powerful statistical software on their computers, it’s becoming very common. Data mining per se isn’t “wrong”, but the more that you do it, the bigger the risk you’re taking. The thing that is wrong, and I suspect is very common, is unacknowledged data mining. That is, the researcher run every possible analysis known to humanity, finds the one that works, and then pretends that this was the only analysis that they ever conducted. Worse yet, they often “invent” a hypothesis after looking at the data, to cover up the data mining. To be clear: it’s not wrong to change your beliefs after looking at the data, and to reanalyse your data using your new “post hoc” hypotheses. What is wrong (and, I suspect, common) is failing to acknowledge that you’ve done so. If you acknowledge that you did it, then other researchers are able to take your behaviour into account. If you don’t, then they can’t. And that makes your behaviour deceptive. Bad!\n * Publication bias & self-censoring. Finally, a pervasive bias is “non-reporting” of negative results. This is almost impossible to prevent. Journals don’t publish every article that is submitted to them: they prefer to publish articles that find “something”. So, if 20 people run an experiment looking at whether reading Finnegans Wake causes insanity in humans, and 19 of them find that it doesn’t, which one do you think is going to get published? Obviously, it’s the one study that did find that Finnegans Wake causes insanity.12 This is an example of a publication bias: since no-one ever published the 19 studies that didn’t find an effect, a naive reader would never know that they existed. Worse yet, most researchers “internalise” this bias, and end up self-censoring their research. Knowing that negative results aren’t going to be accepted for publication, they never even try to report them. As a friend of mine says “for every experiment that you get published, you also have 10 failures”. And she’s right. The catch is, while some (maybe most) of those studies are failures for boring reasons (e.g. you stuffed something up) others might be genuine “null” results that you ought to acknowledge when you write up the “good” experiment. And telling which is which is often hard to do. A good place to start is a paper by Ioannidis (2005) with the depressing title “Why most published research findings are false”. I’d also suggest taking a look at work by Kühberger, Fritz, and Scherndl (2014) presenting statistical evidence that this actually happens in psychology.\n\nThere’s probably a lot more issues like this to think about, but that’ll do to start with. What I really want to point out is the blindingly obvious truth that real world science is conducted by actual humans, and only the most gullible of people automatically assumes that everyone else is honest and impartial. Actual scientists aren’t usually that naive, but for some reason the world likes to pretend that we are, and the textbooks we usually write seem to reinforce that stereotype.\n\n## 2.8 Summary\n\nThis chapter isn’t really meant to provide a comprehensive discussion of psychological research methods: it would require another volume just as long as this one to do justice to the topic. However, in real life statistics and study design are tightly intertwined, so it’s very handy to discuss some of the key topics. In this chapter, I’ve briefly discussed the following topics:\n\n* Introduction to psychological measurement. What does it mean to operationalise a theoretical construct? What does it mean to have variables and take measurements?\n * Scales of measurement and types of variables. Remember that there are two different distinctions here: there’s the difference between discrete and continuous data, and there’s the difference between the four different scale types (nominal, ordinal, interval and ratio).\n * Reliability of a measurement. If I measure the “same” thing twice, should I expect to see the same result? Only if my measure is reliable. But what does it mean to talk about doing the “same” thing? Well, that’s why we have different types of reliability. Make sure you remember what they are.\n * Terminology: predictors and outcomes. What roles do variables play in an analysis? Can you remember the difference between predictors and outcomes? Dependent and independent variables? Etc.\n * Experimental and non-experimental research designs. What makes an experiment an experiment? Is it a nice white lab coat, or does it have something to do with researcher control over variables?\n * Validity and its threats. Does your study measure what you want it to? How might things go wrong? And is it my imagination, or was that a very long list of possible ways in which things can go wrong?\n\nAll this should make clear to you that study design is a critical part of research methodology. I built this chapter from the classic little book by Stanley (1963), but there are of course a large number of textbooks out there on research design. Spend a few minutes with your favourite search engine and you’ll find dozens.\n\nPresidential Address to the First Indian Statistical Congress, 1938. Source: http://en.wikiquote.org/wiki/Ronald_Fisher↩\n * \n\nWell… now this is awkward, isn’t it? This section is one of the oldest parts of the book, and it’s outdated in a rather embarrassing way. I wrote this in 2010, at which point all of those facts were true. Revisiting this in 2018… well I’m not 33 any more, but that’s not surprising I suppose. I can’t imagine my chromosomes have changed, so I’m going to guess my karyotype was then and is now XY. The self-identified gender, on the other hand… ah. I suppose the fact that the title page now refers to me as Danielle rather than Daniel might possibly be a giveaway, but I don’t typically identify as “male” on a gender questionnaire these days, and I prefer “she/her” pronouns as a default (it’s a long story)! I did think a little about how I was going to handle this in the book, actually. The book has a somewhat distinct authorial voice to it, and I feel like it would be a rather different work if I went back and wrote everything as Danielle and updated all the pronouns in the work. Besides, it would be a lot of work, so I’ve left my name as “Dan” throughout the book, and in ant case “Dan” is a perfectly good nickname for “Danielle”, don’t you think? In any case, it’s not a big deal. I only wanted to mention it to make life a little easier for readers who aren’t sure how to refer to me. I still don’t like anchovies though :-)↩\n * \n\nActually, I’ve been informed by readers with greater physics knowledge than I that temperature isn’t strictly an interval scale, in the sense that the amount of energy required to heat something up by 3\\(^\\circ\\) depends on it’s current temperature. So in the sense that physicists care about, temperature isn’t actually interval scale. But it still makes a cute example, so I’m going to ignore this little inconvenient truth.↩\n * \n\nAnnoyingly, though, there’s a lot of different names used out there. I won’t list all of them – there would be no point in doing that – other than to note that R often uses “response variable” where I’ve used “outcome”, and a traditionalist would use “dependent variable”. Sigh. This sort of terminological confusion is very common, I’m afraid.↩\n * \n\nThe reason why I say that it’s unmeasured is that if you have measured it, then you can use some fancy statistical tricks to deal with the confound. Because of the existence of these statistical solutions to the problem of confounds, we often refer to a confound that we have measured and dealt with as a covariate. Dealing with covariates is a topic for a more advanced course, but I thought I’d mention it in passing, since it’s kind of comforting to at least know that this stuff exists.↩\n * \n\nSome people might argue that if you’re not honest then you’re not a real scientist. Which does have some truth to it I guess, but that’s disingenuous (google the “No true Scotsman” fallacy). The fact is that there are lots of people who are employed ostensibly as scientists, and whose work has all of the trappings of science, but who are outright fraudulent. Pretending that they don’t exist by saying that they’re not scientists is just childish.↩\n * \n\nClearly, the real effect is that only insane people would even try to read Finnegans Wake.↩\n\n# Chapter 3 Getting started with R\n\nRobots are nice to work with.\n\n–\n\nIn this chapter I’ll discuss how to get started in R. I’ll briefly talk about how to download and install R, but most of the chapter will be focused on getting you started typing R commands. Our goal in this chapter is not to learn any statistical concepts: we’re just trying to learn the basics of how R works and get comfortable interacting with the system. To do this, we’ll spend a bit of time using R as a simple calculator, since that’s the easiest thing to do with R. In doing so, you’ll get a bit of a feel for what it’s like to work in R. From there I’ll introduce some very basic programming ideas: in particular, I’ll talk about the idea of defining variables to store information, and a few things that you can do with these variables.\n\nHowever, before going into any of the specifics, it’s worth talking a little about why you might want to use R at all. Given that you’re reading this, you’ve probably got your own reasons. However, if those reasons are “because that’s what my stats class uses”, it might be worth explaining a little why your lecturer has chosen to use R for the class. Of course, I don’t really know why other people choose R, so I’m really talking about why I use it.\n\n* It’s sort of obvious, but worth saying anyway: doing your statistics on a computer is faster, easier and more powerful than doing statistics by hand. Computers excel at mindless repetitive tasks, and a lot of statistical calculations are both mindless and repetitive. For most people, the only reason to ever do statistical calculations with pencil and paper is for learning purposes. In my class I do occasionally suggest doing some calculations that way, but the only real value to it is pedagogical. It does help you to get a “feel” for statistics to do some calculations yourself, so it’s worth doing it once. But only once!\n * Doing statistics in a spreadsheet (e.g., Microsoft Excel) is generally a bad idea in the long run. Although many people are likely feel more familiar with them, spreadsheets are very limited in terms of what analyses they allow you do. If you get into the habit of trying to do your real life data analysis using spreadsheets, then you’ve dug yourself into a very deep hole.\n * Avoiding proprietary software is a very good idea. There are a lot of commercial packages out there that you can buy, some of which I like and some of which I don’t. They’re usually very glossy in their appearance, and generally very powerful (much more powerful than spreadsheets). However, they’re also very expensive: usually, the company sells “student versions” (crippled versions of the real thing) very cheaply; they sell full powered “educational versions” at a price that makes me wince; and they sell commercial licences with a staggeringly high price tag. The business model here is to suck you in during your student days, and then leave you dependent on their tools when you go out into the real world. It’s hard to blame them for trying, but personally I’m not in favour of shelling out thousands of dollars if I can avoid it. And you can avoid it: if you make use of packages like R that are open source and free, you never get trapped having to pay exorbitant licensing fees.\n * Something that you might not appreciate now, but will love later on if you do anything involving data analysis, is the fact that R is highly extensible. When you download and install R, you get all the basic “packages”, and those are very powerful on their own. However, because R is so open and so widely used, it’s become something of a standard tool in statistics, and so lots of people write their own packages that extend the system. And these are freely available too. One of the consequences of this, I’ve noticed, is that if you open up an advanced textbook (a recent one, that is) rather than introductory textbooks, is that a lot of them use R. In other words, if you learn how to do your basic statistics in R, then you’re a lot closer to being able to use the state of the art methods than you would be if you’d started out with a “simpler” system: so if you want to become a genuine expert in psychological data analysis, learning R is a very good use of your time.\n * Related to the previous point: R is a real programming language. As you get better at using R for data analysis, you’re also learning to program. To some people this might seem like a bad thing, but in truth, programming is a core research skill across a lot of the social and behavioural sciences. Think about how many surveys and experiments are done online, or presented on computers. Think about all those online social environments which you might be interested in studying; and maybe collecting data from in an automated fashion. Think about artificial intelligence systems, computer vision and speech recognition. If any of these are things that you think you might want to be involved in – as someone “doing research in psychology”, that is – you’ll need to know a bit of programming. And if you don’t already know how to program, then learning how to do statistics using R is a nice way to start.\n\nThose are the main reasons I use R. It’s not without its flaws: it’s not easy to learn, and it has a few very annoying quirks to it that we’re all pretty much stuck with, but on the whole I think the strengths outweigh the weakness; more so than any other option I’ve encountered so far.\n\n## 3.1 Installing R\n\nOkay, enough with the sales pitch. Let’s get started. Just as with any piece of software, R needs to be installed on a “computer”, which is a magical box that does cool things and delivers free ponies. Or something along those lines: I may be confusing computers with the iPad marketing campaigns. Anyway, R is freely distributed online, and you can download it from the R homepage, which is:\n\nAt the top of the page – under the heading “Download and Install R” – you’ll see separate links for Windows users, Mac users, and Linux users. If you follow the relevant link, you’ll see that the online instructions are pretty self-explanatory, but I’ll walk you through the installation anyway. As of this writing, the current version of R is 3.0.2 (``Frisbee Sailing“), but they usually issue updates every six months, so you’ll probably have a newer version.14\n\n### 3.1.1 Installing R on a Windows computer\n\n The CRAN homepage changes from time to time, and it’s not particularly pretty, or all that well-designed quite frankly. But it’s not difficult to find what you’re after. In general you’ll find a link at the top of the page with the text “Download R for Windows”. If you click on that, it will take you to a page that offers you a few options. Again, at the very top of the page you’ll be told to click on a link that says to click here if you’re installing R for the first time. That’s probably what you want. This will take you to a page that has a prominent link at the top called “Download R 3.0.2 for Windows”. That’s the one you want. Click on that and your browser should start downloading a file called `R-3.0.2-win.exe` , or whatever the equivalent version number is by the time you read this. The file for version 3.0.2 is about 54MB in size, so it may take some time depending on how fast your internet connection is. Once you’ve downloaded the file, double click to install it. As with any software you download online, Windows will ask you some questions about whether you trust the file and so on. After you click through those, it’ll ask you where you want to install it, and what components you want to install. The default values should be fine for most people, so again, just click through. Once all that is done, you should have R installed on your system. You can access it from the Start menu, or from the desktop if you asked it to add a shortcut there. You can now open up R in the usual way if you want to, but what I’m going to suggest is that instead of doing that you should now install RStudio (see Section 3.1.4 for instructions). \n\n### 3.1.2 Installing R on a Mac\n\n When you click on the Mac OS X link, you should find yourself on a page with the title “R for Mac OS X”. The vast majority of Mac users will have a fairly recent version of the operating system: as long as you’re running Mac OS X 10.6 (Snow Leopard) or higher, then you’ll be fine.15 There’s a fairly prominent link on the page called “R-3.0.2.pkg”, which is the one you want. Click on that link and you’ll start downloading the installer file, which is (not surprisingly) called `R-3.0.2.pkg` . It’s about 61MB in size, so the download can take a while on slower internet connections. Once you’ve downloaded `R-3.0.2.pkg` , all you need to do is open it by double clicking on the package file. The installation should go smoothly from there: just follow all the instructions just like you usually do when you install something. Once it’s finished, you’ll find a file called `R.app` in the Applications folder. You can now open up R in the usual way16 if you want to, but what I’m going to suggest is that instead of doing that you should now install RStudio (see Section 3.1.4 for instructions). \n\n### 3.1.3 Installing R on a Linux computer\n\n If you’re successfully managing to run a Linux box, regardless of what distribution, then you should find the instructions on the website easy enough. You can compile R from source yourself if you want, or install it through your package management system, which will probably have R in it. Alternatively, the CRAN site has precompiled binaries for Debian, Red Hat, Suse and Ubuntu and has separate instructions for each. Once you’ve got R installed, you can run it from the command line just by typing `R` . However, if you’re feeling envious of Windows and Mac users for their fancy GUIs, you can download RStudio too (see Section 3.1.4 for instructions). \n\n### 3.1.4 Downloading and installing RStudio\n\nOkay, so regardless of what operating system you’re using, the last thing that I told you to do is to download RStudio. To understand why I’ve suggested this, you need to understand a little bit more about R itself. The term R doesn’t really refer to a specific application on your computer. Rather, it refers to the underlying statistical language. You can use this language through lots of different applications. When you install R initially, it comes with one application that lets you do this: it’s the R.exe application on a Windows machine, and the R.app application on a Mac. But that’s not the only way to do it. There are lots of different applications that you can use that will let you interact with R. One of those is called RStudio, and it’s the one I’m going to suggest that you use. RStudio provides a clean, professional interface to R that I find much nicer to work with than either the Windows or Mac defaults. Like R itself, RStudio is free software: you can find all the details on their webpage. In the meantime, you can download it here:\n\nWhen you visit the RStudio website, you’ll probably be struck by how much cleaner and simpler it is than the CRAN website,17 and how obvious it is what you need to do: click the big green button that says “Download”.\n\nWhen you click on the download button on the homepage it will ask you to choose whether you want the desktop version or the server version. You want the desktop version. After choosing the desktop version it will take you to a page http://www.RStudio.org/download/desktop) that shows several possible downloads: there’s a different one for each operating system. However, the nice people at RStudio have designed the webpage so that it automatically recommends the download that is most appropriate for your computer. Click on the appropriate link, and the RStudio installer file will start downloading.\n\nOnce it’s finished downloading, open the installer file in the usual way to install RStudio. After it’s finished installing, you can start R by opening RStudio. You don’t need to open R.app or R.exe in order to access R. RStudio will take care of that for you. To illustrate what RStudio looks like, Figure 3.1 shows a screenshot of an R session in progress. In this screenshot, you can see that it’s running on a Mac, but it looks almost identical no matter what operating system you have. The Windows version looks more like a Windows application (e.g., the menus are attached to the application window and the colour scheme is slightly different), but it’s more or less identical. There are a few minor differences in where things are located in the menus (I’ll point them out as we go along) and in the shortcut keys, because RStudio is trying to “feel” like a proper Mac application or a proper Windows application, and this means that it has to change its behaviour a little bit depending on what computer it’s running on. Even so, these differences are very small: I started out using the Mac version of RStudio and then started using the Windows version as well in order to write these notes.\n\nThe only “shortcoming” I’ve found with RStudio is that – as of this writing – it’s still a work in progress. The “problem” is that they keep improving it. New features keep turning up the more recent releases, so there’s a good chance that by the time you read this book there will be a version out that has some really neat things that weren’t in the version that I’m using now.\n\n### 3.1.5 Starting up R\n\nOne way or another, regardless of what operating system you’re using and regardless of whether you’re using RStudio, or the default GUI, or even the command line, it’s time to open R and get started. When you do that, the first thing you’ll see (assuming that you’re looking at the R console, that is) is a whole lot of text that doesn’t make much sense. It should look something like this:\n\n```\nR version 3.0.2 (2013-09-25) -- \"Frisbee Sailing\"\nCopyright (C) 2013 The R Foundation for Statistical Computing\nPlatform: x86_64-apple-darwin10.8.0 (64-bit)\n\nR is free software and comes with ABSOLUTELY NO WARRANTY.\nYou are welcome to redistribute it under certain conditions.\nType 'license()' or 'licence()' for distribution details.\n\n Natural language support but running in an English locale\n\nR is a collaborative project with many contributors.\nType 'contributors()' for more information and\n'citation()' on how to cite R or R packages in publications.\n\nType 'demo()' for some demos, 'help()' for on-line help, or\n'help.start()' for an HTML browser interface to help.\nType 'q()' to quit R.\n\n>\n```\n\nMost of this text is pretty uninteresting, and when doing real data analysis you’ll never really pay much attention to it. The important part of it is this…\n `>` \n… which has a flashing cursor next to it. That’s the command prompt. When you see this, it means that R is waiting patiently for you to do something!\n\n## 3.2 Typing commands at the R console\n\n One of the easiest things you can do with R is use it as a simple calculator, so it’s a good place to start. For instance, try typing `10 + 20` , and hitting enter.18 When you do this, you’ve entered a command, and R will “execute” that command. What you see on screen now will be this: \n\n Not a lot of surprises in this extract. But there’s a few things worth talking about, even with such a simple example. Firstly, it’s important that you understand how to read the extract. In this example, what I typed was the `10 + 20` part. I didn’t type the `>` symbol: that’s just the R command prompt and isn’t part of the actual command. And neither did I type the `[1] 30` part. That’s what R printed out in response to my command. Secondly, it’s important to understand how the output is formatted. Obviously, the correct answer to the sum `10 + 20` is `30` , and not surprisingly R has printed that out as part of its response. But it’s also printed out this `[1]` part, which probably doesn’t make a lot of sense to you right now. You’re going to see that a lot. I’ll talk about what this means in a bit more detail later on, but for now you can think of `[1] 30` as if R were saying “the answer to the 1st question you asked is 30”. That’s not quite the truth, but it’s close enough for now. And in any case it’s not really very interesting at the moment: we only asked R to calculate one thing, so obviously there’s only one answer printed on the screen. Later on this will change, and the `[1]` part will start to make a bit more sense. For now, I just don’t want you to get confused or concerned by it. \n\n### 3.2.1 An important digression about formatting\n\nNow that I’ve taught you these rules I’m going to change them pretty much immediately. That is because I want you to be able to copy code from the book directly into R if if you want to test things or conduct your own analyses. However, if you copy this kind of code (that shows the command prompt and the results) directly into R you will get an error\n\n```\n## Error: :1:1: unexpected '>'\n## 1: >\n## ^\n```\n\nSo instead, I’m going to provide code in a slightly different format so that it looks like this…\n `10 + 20` `## [1] 30` \nThere are two main differences.\n\n* In your console, you type after the >, but from now I I won’t show the command prompt in the book.\n * In the book, output is commented out with ##, in your console it appears directly after your code.\n\nThese two differences mean that if you’re working with an electronic version of the book, you can easily copy code out of the book and into the console.\n\nSo for example if you copied the two lines of code from the book you’d get this\n `10 + 20` `## [1] 30` `## [1] 30` \n\n### 3.2.2 Be very careful to avoid typos\n\n Before we go on to talk about other types of calculations that we can do with R, there’s a few other things I want to point out. The first thing is that, while R is good software, it’s still software. It’s pretty stupid, and because it’s stupid it can’t handle typos. It takes it on faith that you meant to type exactly what you did type. For example, suppose that you forgot to hit the shift key when trying to type `+` , and as a result your command ended up being `10 = 20` rather than `10 + 20` . Here’s what happens: `10 = 20` \n\n```\n## Error in 10 = 20: invalid (do_set) left-hand side to assignment\n```\n\n What’s happened here is that R has attempted to interpret `10 = 20` as a command, and spits out an error message because the command doesn’t make any sense to it. When a human looks at this, and then looks down at his or her keyboard and sees that `+` and `=` are on the same key, it’s pretty obvious that the command was a typo. But R doesn’t know this, so it gets upset. And, if you look at it from its perspective, this makes sense. All that R “knows” is that `10` is a legitimate number, `20` is a legitimate number, and `=` is a legitimate part of the language too. In other words, from its perspective this really does look like the user meant to type `10 = 20` , since all the individual parts of that statement are legitimate and it’s too stupid to realise that this is probably a typo. Therefore, R takes it on faith that this is exactly what you meant… it only “discovers” that the command is nonsense when it tries to follow your instructions, typo and all. And then it whinges, and spits out an error. Even more subtle is the fact that some typos won’t produce errors at all, because they happen to correspond to “well-formed” R commands. For instance, suppose that not only did I forget to hit the shift key when trying to type `10 + 20` , I also managed to press the key next to one I meant do. The resulting typo would produce the command `10 - 20` . Clearly, R has no way of knowing that you meant to add 20 to 10, not subtract 20 from 10, so what happens this time is this: `10 - 20` `## [1] -10` \nIn this case, R produces the right answer, but to the the wrong question.\n\nTo some extent, I’m stating the obvious here, but it’s important. The people who wrote R are smart. You, the user, are smart. But R itself is dumb. And because it’s dumb, it has to be mindlessly obedient. It does exactly what you ask it to do. There is no equivalent to “autocorrect” in R, and for good reason. When doing advanced stuff – and even the simplest of statistics is pretty advanced in a lot of ways – it’s dangerous to let a mindless automaton like R try to overrule the human user. But because of this, it’s your responsibility to be careful. Always make sure you type exactly what you mean. When dealing with computers, it’s not enough to type “approximately” the right thing. In general, you absolutely must be precise in what you say to R … like all machines it is too stupid to be anything other than absurdly literal in its interpretation.\n\n### 3.2.3 R is (a bit) flexible with spacing\n\n Of course, now that I’ve been so uptight about the importance of always being precise, I should point out that there are some exceptions. Or, more accurately, there are some situations in which R does show a bit more flexibility than my previous description suggests. The first thing R is smart enough to do is ignore redundant spacing. What I mean by this is that, when I typed `10 + 20` before, I could equally have done this `10 + 20` `## [1] 30` \nor this\n `10+20` `## [1] 30` and I would get exactly the same answer. However, that doesn’t mean that you can insert spaces in any old place. When we looked at the startup documentation in Section 3.1.5 it suggested that you could type `citation()` to get some information about how to cite R. If I do so… `citation()` \n\n```\n## \n## To cite R in publications use:\n## \n## R Core Team (2018). R: A language and environment for\n## statistical computing. R Foundation for Statistical Computing,\n## Vienna, Austria. URL https://www.R-project.org/.\n## \n## A BibTeX entry for LaTeX users is\n## \n## @Manual{,\n## title = {R: A Language and Environment for Statistical Computing},\n## author = {{}},\n## organization = {R Foundation for Statistical Computing},\n## address = {Vienna, Austria},\n## year = {2018},\n## url = {https://www.R-project.org/},\n## }\n## \n## We have invested a lot of time and effort in creating R, please\n## cite it when using it for data analysis. See also\n## 'citation(\"pkgname\")' for citing R packages.\n```\n\n… it tells me to cite the R manual (R Core Team 2013). Let’s see what happens when I try changing the spacing. If I insert spaces in between the word and the parentheses, or inside the parentheses themselves, then all is well. That is, either of these two commands\n `citation ()` `citation( )` \nwill produce exactly the same response. However, what I can’t do is insert spaces in the middle of the word. If I try to do this, R gets upset:\n `citat ion()` \n\n```\n## Error: :1:7: unexpected symbol\n## 1: citat ion\n## ^\n```\n\nThroughout this book I’ll vary the way I use spacing a little bit, just to give you a feel for the different ways in which spacing can be used. I’ll try not to do it too much though, since it’s generally considered to be good practice to be consistent in how you format your commands.\n\n### 3.2.4 R can sometimes tell that you’re not finished yet (but not often)\n\n One more thing I should point out. If you hit enter in a situation where it’s “obvious” to R that you haven’t actually finished typing the command, R is just smart enough to keep waiting. For example, if you type `10 +` and then press enter, even R is smart enough to realise that you probably wanted to type in another number. So here’s what happens (for illustrative purposes I’m breaking my own code formatting rules in this section): \n\n```\n> 10+\n+\n```\n\n and there’s a blinking cursor next to the plus sign. What this means is that R is still waiting for you to finish. It “thinks” you’re still typing your command, so it hasn’t tried to execute it yet. In other words, this plus sign is actually another command prompt. It’s different from the usual one (i.e., the `>` symbol) to remind you that R is going to “add” whatever you type now to what you typed last time. For example, if I then go on to type `3` and hit enter, what I get is this: \n\n And as far as R is concerned, this is exactly the same as if you had typed `10 + 20` . Similarly, consider the `citation()` command that we talked about in the previous section. Suppose you hit enter after typing `citation(` . Once again, R is smart enough to realise that there must be more coming – since you need to add the `)` character – so it waits. I can even hit enter several times and it will keep waiting: \n\n```\n> citation(\n+ \n+ \n+ )\n```\n\nI’ll make use of this a lot in this book. A lot of the commands that we’ll have to type are pretty long, and they’re visually a bit easier to read if I break it up over several lines. If you start doing this yourself, you’ll eventually get yourself in trouble (it happens to us all). Maybe you start typing a command, and then you realise you’ve screwed up. For example,\n\n```\n> citblation( \n+ \n+\n```\n\n You’d probably prefer R not to try running this command, right? If you want to get out of this situation, just hit the ‘escape’ key.19 R will return you to the normal command prompt (i.e. `>` ) without attempting to execute the botched command. That being said, it’s not often the case that R is smart enough to tell that there’s more coming. For instance, in the same way that I can’t add a space in the middle of a word, I can’t hit enter in the middle of a word either. If I hit enter after typing `citat` I get an error, because R thinks I’m interested in an “object” called `citat` and can’t find it: \n\n```\n> citat\nError: object 'citat' not found\n```\n\n What about if I typed `citation` and hit enter? In this case we get something very odd, something that we definitely don’t want, at least at this stage. Here’s what happens: \n\n```\ncitation\n## function (package = \"base\", lib.loc = NULL, auto = NULL) \n## {\n## dir <- system.file(package = package, lib.loc = lib.loc)\n## if (dir == \"\") \n## stop(gettextf(\"package '%s' not found\", package), domain = NA)\n\nBLAH BLAH BLAH\n```\n\n where the `BLAH BLAH BLAH` goes on for rather a long time, and you don’t know enough R yet to understand what all this gibberish actually means (of course, it doesn’t actually say BLAH BLAH BLAH - it says some other things we don’t understand or need to know that I’ve edited for length) This incomprehensible output can be quite intimidating to novice users, and unfortunately it’s very easy to forget to type the parentheses; so almost certainly you’ll do this by accident. Do not panic when this happens. Simply ignore the gibberish. As you become more experienced this gibberish will start to make sense, and you’ll find it quite handy to print this stuff out.20 But for now just try to remember to add the parentheses when typing your commands. \n\n## 3.3 Doing simple calculations with R\n\n Okay, now that we’ve discussed some of the tedious details associated with typing R commands, let’s get back to learning how to use the most powerful piece of statistical software in the world as a $2 calculator. So far, all we know how to do is addition. Clearly, a calculator that only did addition would be a bit stupid, so I should tell you about how to perform other simple calculations using R. But first, some more terminology. Addition is an example of an “operation” that you can perform (specifically, an arithmetic operation), and the operator that performs it is `+` . To people with a programming or mathematics background, this terminology probably feels pretty natural, but to other people it might feel like I’m trying to make something very simple (addition) sound more complicated than it is (by calling it an arithmetic operation). To some extent, that’s true: if addition was the only operation that we were interested in, it’d be a bit silly to introduce all this extra terminology. However, as we go along, we’ll start using more and more different kinds of operations, so it’s probably a good idea to get the language straight now, while we’re still talking about very familiar concepts like addition! \n\n### 3.3.1 Adding, subtracting, multiplying and dividing\n\nSo, now that we have the terminology, let’s learn how to perform some arithmetic operations in R. To that end, Table 3.1 lists the operators that correspond to the basic arithmetic we learned in primary school: addition, subtraction, multiplication and division.\n\noperation | operator | example input | example output |\n | --- | --- | --- | --- |\naddition | | 10 + 2 | 12 |\nsubtraction | | 9 - 3 | 6 |\nmultiplication | | 5 * 5 | 25 |\ndivision | | 10 / 3 | 3 |\npower | | 5 ^ 2 | 25 |\n\n As you can see, R uses fairly standard symbols to denote each of the different operations you might want to perform: addition is done using the `+` operator, subtraction is performed by the `-` operator, and so on. So if I wanted to find out what 57 times 61 is (and who wouldn’t?), I can use R instead of a calculator, like so: `57 * 61` `## [1] 3477` \nSo that’s handy.\n\n### 3.3.2 Taking powers\n\nThe first four operations listed in Table 3.1 are things we all learned in primary school, but they aren’t the only arithmetic operations built into R. There are three other arithmetic operations that I should probably mention: taking powers, doing integer division, and calculating a modulus. Of the three, the only one that is of any real importance for the purposes of this book is taking powers, so I’ll discuss that one here: the other two are discussed in Chapter 7.\n\nFor those of you who can still remember your high school maths, this should be familiar. But for some people high school maths was a long time ago, and others of us didn’t listen very hard in high school. It’s not complicated. As I’m sure everyone will probably remember the moment they read this, the act of multiplying a number \\(x\\) by itself \\(n\\) times is called “raising \\(x\\) to the \\(n\\)-th power”. Mathematically, this is written as \\(x^n\\). Some values of \\(n\\) have special names: in particular \\(x^2\\) is called \\(x\\)-squared, and \\(x^3\\) is called \\(x\\)-cubed. So, the 4th power of 5 is calculated like this: \\[ 5^4 = 5 \\times 5 \\times 5 \\times 5 \\]\n\nOne way that we could calculate \\(5^4\\) in R would be to type in the complete multiplication as it is shown in the equation above. That is, we could do this\n `5 * 5 * 5 * 5` `## [1] 625` \nbut it does seem a bit tedious. It would be very annoying indeed if you wanted to calculate \\(5^{15}\\), since the command would end up being quite long. Therefore, to make our lives easier, we use the power operator instead. When we do that, our command to calculate \\(5^4\\) goes like this:\n `5 ^ 4` `## [1] 625` \nMuch easier.\n\n### 3.3.3 Doing calculations in the right order\n\nOkay. At this point, you know how to take one of the most powerful pieces of statistical software in the world, and use it as a $2 calculator. And as a bonus, you’ve learned a few very basic programming concepts. That’s not nothing (you could argue that you’ve just saved yourself $2) but on the other hand, it’s not very much either. In order to use R more effectively, we need to introduce more programming concepts.\n In most situations where you would want to use a calculator, you might want to do multiple calculations. R lets you do this, just by typing in longer commands.21 In fact, we’ve already seen an example of this earlier, when I typed in `5 * 5 * 5 * 5` . However, let’s try a slightly different example: `1 + 2 * 4` `## [1] 9` Clearly, this isn’t a problem for R either. However, it’s worth stopping for a second, and thinking about what R just did. Clearly, since it gave us an answer of `9` it must have multiplied `2 * 4` (to get an interim answer of 8) and then added 1 to that. But, suppose it had decided to just go from left to right: if R had decided instead to add `1+2` (to get an interim answer of 3) and then multiplied by 4, it would have come up with an answer of `12` . To answer this, you need to know the order of operations that R uses. If you remember back to your high school maths classes, it’s actually the same order that you got taught when you were at school: the “BEDMAS” order.22 That is, first calculate things inside Brackets `()` , then calculate Exponents `^` , then Division `/` and Multiplication `*` , then Addition `+` and Subtraction `-` . So, to continue the example above, if we want to force R to calculate the `1+2` part before the multiplication, all we would have to do is enclose it in brackets: `(1 + 2) * 4` `## [1] 12` This is a fairly useful thing to be able to do. The only other thing I should point out about order of operations is what to expect when you have two operations that have the same priority: that is, how does R resolve ties? For instance, multiplication and division are actually the same priority, but what should we expect when we give R a problem like `4 / 2 * 3` to solve? If it evaluates the multiplication first and then the division, it would calculate a value of two-thirds. But if it evaluates the division first it calculates a value of 6. The answer, in this case, is that R goes from left to right, so in this case the division step would come first: `4 / 2 * 3` `## [1] 6` All of the above being said, it’s helpful to remember that brackets always come first. So, if you’re ever unsure about what order R will do things in, an easy solution is to enclose the thing you want it to do first in brackets. There’s nothing stopping you from typing `(4 / 2) * 3` . By enclosing the division in brackets we make it clear which thing is supposed to happen first. In this instance you wouldn’t have needed to, since R would have done the division first anyway, but when you’re first starting out it’s better to make sure R does what you want! \n\n## 3.4 Storing a number as a variable\n\nOne of the most important things to be able to do in R (or any programming language, for that matter) is to store information in variables. Variables in R aren’t exactly the same thing as the variables we talked about in the last chapter on research methods, but they are similar. At a conceptual level you can think of a variable as label for a certain piece of information, or even several different pieces of information. When doing statistical analysis in R all of your data (the variables you measured in your study) will be stored as variables in R, but as well see later in the book you’ll find that you end up creating variables for other things too. However, before we delve into all the messy details of data sets and statistical analysis, let’s look at the very basics for how we create variables and work with them.\n\n### 3.4.1 Variable assignment using\n\n `<-` and `->` Since we’ve been working with numbers so far, let’s start by creating variables to store our numbers. And since most people like concrete examples, let’s invent one. Suppose I’m trying to calculate how much money I’m going to make from this book. There’s several different numbers I might want to store. Firstly, I need to figure out how many copies I’ll sell. This isn’t exactly Harry Potter, so let’s assume I’m only going to sell one copy per student in my class. That’s 350 sales, so let’s create a variable called `sales` . What I want to do is assign a value to my variable `sales` , and that value should be `350` . We do this by using the assignment operator, which is `<-` . Here’s how we do it: `sales <- 350` When you hit enter, R doesn’t print out any output.23 It just gives you another command prompt. However, behind the scenes R has created a variable called `sales` and given it a value of `350` . You can check that this has happened by asking R to print the variable on screen. And the simplest way to do that is to type the name of the variable and hit enter24. `sales` `## [1] 350` \nSo that’s nice to know. Anytime you can’t remember what R has got stored in a particular variable, you can just type the name of the variable and hit enter.\n Okay, so now we know how to assign variables. Actually, there’s a bit more you should know. Firstly, one of the curious features of R is that there are several different ways of making assignments. In addition to the `<-` operator, we can also use `->` and `=` , and it’s pretty important to understand the differences between them.25 Let’s start by considering `->` , since that’s the easy one (we’ll discuss the use of `=` in Section 3.5.1. As you might expect from just looking at the symbol, it’s almost identical to `<-` . It’s just that the arrow (i.e., the assignment) goes from left to right. So if I wanted to define my `sales` variable using `->` , I would write it like this: `350 -> sales` This has the same effect: and it still means that I’m only going to sell `350` copies. Sigh. Apart from this superficial difference, `<-` and `->` are identical. In fact, as far as R is concerned, they’re actually the same operator, just in a “left form” and a “right form.”26 \n\n### 3.4.2 Doing calculations using variables\n\n Okay, let’s get back to my original story. In my quest to become rich, I’ve written this textbook. To figure out how good a strategy is, I’ve started creating some variables in R. In addition to defining a `sales` variable that counts the number of copies I’m going to sell, I can also create a variable called `royalty` , indicating how much money I get per copy. Let’s say that my royalties are about $7 per book: \n\n```\nsales <- 350\nroyalty <- 7\n```\n\n The nice thing about variables (in fact, the whole point of having variables) is that we can do anything with a variable that we ought to be able to do with the information that it stores. That is, since R allows me to multiply `350` by `7` `350 * 7` `## [1] 2450` it also allows me to multiply `sales` by `royalty` `sales * royalty` `## [1] 2450` As far as R is concerned, the `sales * royalty` command is the same as the `350 * 7` command. Not surprisingly, I can assign the output of this calculation to a new variable, which I’ll call `revenue` . And when we do this, the new variable `revenue` gets the value `2450` . So let’s do that, and then get R to print out the value of `revenue` so that we can verify that it’s done what we asked: \n\n```\nrevenue <- sales * royalty\nrevenue\n```\n\n `## [1] 2450` \nThat’s fairly straightforward. A slightly more subtle thing we can do is reassign the value of my variable, based on its current value. For instance, suppose that one of my students (no doubt under the influence of psychotropic drugs) loves the book so much that he or she donates me an extra $550. The simplest way to capture this is by a command like this:\n\n```\nrevenue <- revenue + 550\nrevenue\n```\n\n `## [1] 3000` In this calculation, R has taken the old value of `revenue` (i.e., 2450) and added 550 to that value, producing a value of 3000. This new value is assigned to the `revenue` variable, overwriting its previous value. In any case, we now know that I’m expecting to make $3000 off this. Pretty sweet, I thinks to myself. Or at least, that’s what I thinks until I do a few more calculation and work out what the implied hourly wage I’m making off this looks like. \n\n### 3.4.3 Rules and conventions for naming variables\n\n In the examples that we’ve seen so far, my variable names ( `sales` and `revenue` ) have just been English-language words written using lowercase letters. However, R allows a lot more flexibility when it comes to naming your variables, as the following list of rules27 illustrates: \n\n* Variable names can only use the upper case alphabetic characters\n `A` - `Z` as well as the lower case characters `a` - `z` . You can also include numeric characters `0` - `9` in the variable name, as well as the period `.` or underscore `_` character. In other words, you can use `SaL.e_s` as a variable name (though I can’t think why you would want to), but you can’t use `Sales?` . * Variable names cannot include spaces: therefore\n `my sales` is not a valid name, but `my.sales` is. * Variable names are case sensitive: that is,\n `Sales` and `sales` are different variable names. * Variable names must start with a letter or a period. You can’t use something like\n `_sales` or `1sales` as a variable name. You can use `.sales` as a variable name if you want, but it’s not usually a good idea. By convention, variables starting with a `.` are used for special purposes, so you should avoid doing so. * Variable names cannot be one of the reserved keywords. These are special names that R needs to keep “safe” from us mere users, so you can’t use them as the names of variables. The keywords are:\n `if` , `else` , `repeat` , `while` , `function` , `for` , `in` , `next` , `break` , `TRUE` , `FALSE` , `NULL` , `Inf` , `NaN` , `NA` , `NA_integer_` , `NA_real_` , `NA_complex_` , and finally, `NA_character_` . Don’t feel especially obliged to memorise these: if you make a mistake and try to use one of the keywords as a variable name, R will complain about it like the whiny little automaton it is. \nIn addition to those rules that R enforces, there are some informal conventions that people tend to follow when naming variables. One of them you’ve already seen: i.e., don’t use variables that start with a period. But there are several others. You aren’t obliged to follow these conventions, and there are many situations in which it’s advisable to ignore them, but it’s generally a good idea to follow them when you can:\n\n* Use informative variable names. As a general rule, using meaningful names like\n `sales` and `revenue` is preferred over arbitrary ones like `variable1` and `variable2` . Otherwise it’s very hard to remember what the contents of different variables are, and it becomes hard to understand what your commands actually do. * Use short variable names. Typing is a pain and no-one likes doing it. So we much prefer to use a name like\n `sales` over a name like \n\n```\nsales.for.this.book.that.you.are.reading\n```\n\n. Obviously there’s a bit of a tension between using informative names (which tend to be long) and using short names (which tend to be meaningless), so use a bit of common sense when trading off these two conventions. * Use one of the conventional naming styles for multi-word variable names. Suppose I want to name a variable that stores “my new salary”. Obviously I can’t include spaces in the variable name, so how should I do this? There are three different conventions that you sometimes see R users employing. Firstly, you can separate the words using periods, which would give you\n `my.new.salary` as the variable name. Alternatively, you could separate words using underscores, as in `my_new_salary` . Finally, you could use capital letters at the beginning of each word (except the first one), which gives you `myNewSalary` as the variable name. I don’t think there’s any strong reason to prefer one over the other,28 but it’s important to be consistent. \n\n## 3.5 Using functions to do calculations\n\n The symbols `+` , `-` , `*` and so on are examples of operators. As we’ve seen, you can do quite a lot of calculations just by using these operators. However, in order to do more advanced calculations (and later on, to do actual statistics), you’re going to need to start using functions.29 I’ll talk in more detail about functions and how they work in Section 8.4, but for now let’s just dive in and use a few. To get started, suppose I wanted to take the square root of 225. The square root, in case your high school maths is a bit rusty, is just the opposite of squaring a number. So, for instance, since “5 squared is 25” I can say that “5 is the square root of 25”. The usual notation for this is \n\\[ \\sqrt{25} = 5 \\]\n\nthough sometimes you’ll also see it written like this \\(25^{0.5} = 5.\\) This second way of writing it is kind of useful to “remind” you of the mathematical fact that “square root of \\(x\\)” is actually the same as “raising \\(x\\) to the power of 0.5”. Personally, I’ve never found this to be terribly meaningful psychologically, though I have to admit it’s quite convenient mathematically. Anyway, it’s not important. What is important is that you remember what a square root is, since we’re going to need it later on.\n To calculate the square root of 25, I can do it in my head pretty easily, since I memorised my multiplication tables when I was a kid. It gets harder when the numbers get bigger, and pretty much impossible if they’re not whole numbers. This is where something like R comes in very handy. Let’s say I wanted to calculate \\(\\sqrt{225}\\), the square root of 225. There’s two ways I could do this using R. Firstly, since the square root of 255 is the same thing as raising 225 to the power of 0.5, I could use the power operator `^` , just like we did earlier: `225 ^ 0.5` `## [1] 15` However, there’s a second way that we can do this, since R also provides a square root function, `sqrt()` . To calculate the square root of 255 using this function, what I do is insert the number `225` in the parentheses. That is, the command I type is this: `sqrt( 225 )` `## [1] 15` and as you might expect from our previous discussion, the spaces in between the parentheses are purely cosmetic. I could have typed `sqrt(225)` or `sqrt( 225 )` and gotten the same result. When we use a function to do something, we generally refer to this as calling the function, and the values that we type into the function (there can be more than one) are referred to as the arguments of that function. Obviously, the `sqrt()` function doesn’t really give us any new functionality, since we already knew how to do square root calculations by using the power operator `^` , though I do think it looks nicer when we use `sqrt()` . However, there are lots of other functions in R: in fact, almost everything of interest that I’ll talk about in this book is an R function of some kind. For example, one function that we will need to use in this book is the absolute value function. Compared to the square root function, it’s extremely simple: it just converts negative numbers to positive numbers, and leaves positive numbers alone. Mathematically, the absolute value of \\(x\\) is written \\(|x|\\) or sometimes \\(\\mbox{abs}(x)\\). Calculating absolute values in R is pretty easy, since R provides the `abs()` function that you can use for this purpose. When you feed it a positive number… `abs( 21 )` `## [1] 21` \nthe absolute value function does nothing to it at all. But when you feed it a negative number, it spits out the positive version of the same number, like this:\n `abs( -13 )` `## [1] 13` \nIn all honesty, there’s nothing that the absolute value function does that you couldn’t do just by looking at the number and erasing the minus sign if there is one. However, there’s a few places later in the book where we have to use absolute values, so I thought it might be a good idea to explain the meaning of the term early on.\n Before moving on, it’s worth noting that – in the same way that R allows us to put multiple operations together into a longer command, like `1 + 2*4` for instance – it also lets us put functions together and even combine functions with operators if we so desire. For example, the following is a perfectly legitimate command: `sqrt( 1 + abs(-8) )` `## [1] 3` When R executes this command, starts out by calculating the value of `abs(-8)` , which produces an intermediate value of `8` . Having done so, the command simplifies to `sqrt( 1 + 8 )` . To solve the square root30 it first needs to add `1 + 8` to get `9` , at which point it evaluates `sqrt(9)` , and so it finally outputs a value of `3` . \n\n### 3.5.1 Function arguments, their names and their defaults\n\n There’s two more fairly important things that you need to understand about how functions work in R, and that’s the use of “named” arguments, and default values\" for arguments. Not surprisingly, that’s not to say that this is the last we’ll hear about how functions work, but they are the last things we desperately need to discuss in order to get you started. To understand what these two concepts are all about, I’ll introduce another function. The `round()` function can be used to round some value to the nearest whole number. For example, I could type this: `round( 3.1415 )` `## [1] 3` Pretty straightforward, really. However, suppose I only wanted to round it to two decimal places: that is, I want to get `3.14` as the output. The `round()` function supports this, by allowing you to input a second argument to the function that specifies the number of decimal places that you want to round the number to. In other words, I could do this: `round( 3.14165, 2 )` `## [1] 3.14` What’s happening here is that I’ve specified two arguments: the first argument is the number that needs to be rounded (i.e., `3.1415` ), the second argument is the number of decimal places that it should be rounded to (i.e., `2` ), and the two arguments are separated by a comma. In this simple example, it’s quite easy to remember which one argument comes first and which one comes second, but for more complicated functions this is not easy. Fortunately, most R functions make use of argument names. For the `round()` function, for example the number that needs to be rounded is specified using the `x` argument, and the number of decimal points that you want it rounded to is specified using the `digits` argument. Because we have these names available to us, we can specify the arguments to the function by name. We do so like this: \n\n `## [1] 3.14` Notice that this is kind of similar in spirit to variable assignment (Section 3.4), except that I used `=` here, rather than `<-` . In both cases we’re specifying specific values to be associated with a label. However, there are some differences between what I was doing earlier on when creating variables, and what I’m doing here when specifying arguments, and so as a consequence it’s important that you use `=` in this context. \nAs you can see, specifying the arguments by name involves a lot more typing, but it’s also a lot easier to read. Because of this, the commands in this book will usually specify arguments by name,31 since that makes it clearer to you what I’m doing. However, one important thing to note is that when specifying the arguments using their names, it doesn’t matter what order you type them in. But if you don’t use the argument names, then you have to input the arguments in the correct order. In other words, these three commands all produce the same output…\n `round( 3.14165, 2 )` `## [1] 3.14` \n\n `## [1] 3.14` \n\n```\nround( digits = 2, x = 3.1415 )\n```\n\n `## [1] 3.14` \nbut this one does not…\n `round( 2, 3.14165 )` `## [1] 2` \nHow do you find out what the correct order is? There’s a few different ways, but the easiest one is to look at the help documentation for the function (see Section 4.12. However, if you’re ever unsure, it’s probably best to actually type in the argument name.\n Okay, so that’s the first thing I said you’d need to know: argument names. The second thing you need to know about is default values. Notice that the first time I called the `round()` function I didn’t actually specify the `digits` argument at all, and yet R somehow knew that this meant it should round to the nearest whole number. How did that happen? The answer is that the `digits` argument has a default value of `0` , meaning that if you decide not to specify a value for `digits` then R will act as if you had typed `digits = 0` . This is quite handy: the vast majority of the time when you want to round a number you want to round it to the nearest whole number, and it would be pretty annoying to have to specify the `digits` argument every single time. On the other hand, sometimes you actually do want to round to something other than the nearest whole number, and it would be even more annoying if R didn’t allow this! Thus, by having `digits = 0` as the default value, we get the best of both worlds. \n\n## 3.6 Letting RStudio help you with your commands\n\nTime for a bit of a digression. At this stage you know how to type in basic commands, including how to use R functions. And it’s probably beginning to dawn on you that there are a lot of R functions, all of which have their own arguments. You’re probably also worried that you’re going to have to remember all of them! Thankfully, it’s not that bad. In fact, very few data analysts bother to try to remember all the commands. What they really do is use tricks to make their lives easier. The first (and arguably most important one) is to use the internet. If you don’t know how a particular R function works, Google it. Second, you can look up the R help documentation. I’ll talk more about these two tricks in Section 4.12. But right now I want to call your attention to a couple of simple tricks that RStudio makes available to you.\n\n### 3.6.1 Autocomplete using “tab”\n\nThe first thing I want to call your attention to is the autocomplete ability in RStudio.32\n Let’s stick to our example above and assume that what you want to do is to round a number. This time around, start typing the name of the function that you want, and then hit the “tab” key. RStudio will then display a little window like the one shown in Figure 3.2. In this figure, I’ve typed the letters `ro` at the command line, and then hit tab. The window has two panels. On the left, there’s a list of variables and functions that start with the letters that I’ve typed shown in black text, and some grey text that tells you where that variable/function is stored. Ignore the grey text for now: it won’t make much sense to you until we’ve talked about packages in Section 4.2. In Figure 3.2 you can see that there’s quite a few things that start with the letters `ro` : there’s something called `rock` , something called `round` , something called `round.Date` and so on. The one we want is `round` , but if you’re typing this yourself you’ll notice that when you hit the tab key the window pops up with the top entry (i.e., `rock` ) highlighted. You can use the up and down arrow keys to select the one that you want. Or, if none of the options look right to you, you can hit the escape key (“esc”) or the left arrow key to make the window go away. In our case, the thing we want is the `round` option, so we’ll select that. When you do this, you’ll see that the panel on the right changes. Previously, it had been telling us something about the `rock` data set (i.e., “Measurements on 48 rock samples…”) that is distributed as part of R. But when we select `round` , it displays information about the `round()` function, exactly as it is shown in Figure 3.2. This display is really handy. The very first thing it says is `round(x, digits = 0)` : what this is telling you is that the `round()` function has two arguments. The first argument is called `x` , and it doesn’t have a default value. The second argument is `digits` , and it has a default value of 0. In a lot of situations, that’s all the information you need. But RStudio goes a bit further, and provides some additional information about the function underneath. Sometimes that additional information is very helpful, sometimes it’s not: RStudio pulls that text from the R help documentation, and my experience is that the helpfulness of that documentation varies wildly. Anyway, if you’ve decided that `round()` is the function that you want to use, you can hit the right arrow or the enter key, and RStudio will finish typing the rest of the function name for you. The RStudio autocomplete tool works slightly differently if you’ve already got the name of the function typed and you’re now trying to type the arguments. For instance, suppose I’ve typed `round(` into the console, and then I hit tab. RStudio is smart enough to recognise that I already know the name of the function that I want, because I’ve already typed it! Instead, it figures that what I’m interested in is the arguments to that function. So that’s what pops up in the little window. You can see this in Figure 3.3. Again, the window has two panels, and you can interact with this window in exactly the same way that you did with the window shown in Figure 3.2. On the left hand panel, you can see a list of the argument names. On the right hand side, it displays some information about what the selected argument does. \n\n### 3.6.2 Browsing your command history\n\nOne thing that R does automatically is keep track of your “command history”. That is, it remembers all the commands that you’ve previously typed. You can access this history in a few different ways. The simplest way is to use the up and down arrow keys. If you hit the up key, the R console will show you the most recent command that you’ve typed. Hit it again, and it will show you the command before that. If you want the text on the screen to go away, hit escape33 Using the up and down keys can be really handy if you’ve typed a long command that had one typo in it. Rather than having to type it all again from scratch, you can use the up key to bring up the command and fix it.\n\nThe second way to get access to your command history is to look at the history panel in RStudio. On the upper right hand side of the RStudio window you’ll see a tab labelled “History”. Click on that, and you’ll see a list of all your recent commands displayed in that panel: it should look something like Figure 3.4. If you double click on one of the commands, it will be copied to the R console. (You can achieve the same result by selecting the command you want with the mouse and then clicking the “To Console” button).34\n\n## 3.7 Storing many numbers as a vector\n\nAt this point we’ve covered functions in enough detail to get us safely through the next couple of chapters (with one small exception: see Section 4.11, so let’s return to our discussion of variables. When I introduced variables in Section 3.4 I showed you how we can use variables to store a single number. In this section, we’ll extend this idea and look at how to store multiple numbers within the one variable. In R, the name for a variable that can store multiple values is a vector. So let’s create one.\n\n### 3.7.1 Creating a vector\n\n Let’s stick to my silly “get rich quick by textbook writing” example. Suppose the textbook company (if I actually had one, that is) sends me sales data on a monthly basis. Since my class start in late February, we might expect most of the sales to occur towards the start of the year. Let’s suppose that I have 100 sales in February, 200 sales in March and 50 sales in April, and no other sales for the rest of the year. What I would like to do is have a variable – let’s call it `sales.by.month` – that stores all this sales data. The first number stored should be `0` since I had no sales in January, the second should be `100` , and so on. The simplest way to do this in R is to use the combine function, `c()` . To do so, all we have to do is type all the numbers you want to store in a comma separated list, like this:35 \n\n```\nsales.by.month <- c(0, 100, 200, 50, 0, 0, 0, 0, 0, 0, 0, 0)\nsales.by.month\n```\n\n To use the correct terminology here, we have a single variable here called `sales.by.month` : this variable is a vector that consists of 12 elements. \n\n### 3.7.2 A handy digression\n\n Now that we’ve learned how to put information into a vector, the next thing to understand is how to pull that information back out again. However, before I do so it’s worth taking a slight detour. If you’ve been following along, typing all the commands into R yourself, it’s possible that the output that you saw when we printed out the `sales.by.month` vector was slightly different to what I showed above. This would have happened if the window (or the RStudio panel) that contains the R console is really, really narrow. If that were the case, you might have seen output that looks something like this: `sales.by.month` \n\n```\n## [1] 0 100 200 50\n## [5] 0 0 0 0\n## [9] 0 0 0 0\n```\n\n Because there wasn’t much room on the screen, R has printed out the results over three lines. But that’s not the important thing to notice. The important point is that the first line has a `[1]` in front of it, whereas the second line starts with `[5]` and the third with `[9]` . It’s pretty clear what’s happening here. For the first row, R has printed out the 1st element through to the 4th element, so it starts that row with a `[1]` . For the second row, R has printed out the 5th element of the vector through to the 8th one, and so it begins that row with a `[5]` so that you can tell where it’s up to at a glance. It might seem a bit odd to you that R does this, but in some ways it’s a kindness, especially when dealing with larger data sets! \n\n### 3.7.3 Getting information out of vectors\n\n To get back to the main story, let’s consider the problem of how to get information out of a vector. At this point, you might have a sneaking suspicion that the answer has something to do with the `[1]` and `[9]` things that R has been printing out. And of course you are correct. Suppose I want to pull out the February sales data only. February is the second month of the year, so let’s try this: `sales.by.month[2]` `## [1] 100` Yep, that’s the February sales all right. But there’s a subtle detail to be aware of here: notice that R outputs `[1] 100` , not `[2] 100` . This is because R is being extremely literal. When we typed in `sales.by.month[2]` , we asked R to find exactly one thing, and that one thing happens to be the second element of our `sales.by.month` vector. So, when it outputs `[1] 100` what R is saying is that the first number that we just asked for is `100` . This behaviour makes more sense when you realise that we can use this trick to create new variables. For example, I could create a `february.sales` variable like this: \n\n```\nfebruary.sales <- sales.by.month[2]\nfebruary.sales\n```\n\n `## [1] 100` Obviously, the new variable `february.sales` should only have one element and so when I print it out this new variable, the R output begins with a `[1]` because `100` is the value of the first (and only) element of `february.sales` . The fact that this also happens to be the value of the second element of `sales.by.month` is irrelevant. We’ll pick this topic up again shortly (Section 3.10). \n\n### 3.7.4 Altering the elements of a vector\n\n Sometimes you’ll want to change the values stored in a vector. Imagine my surprise when the publisher rings me up to tell me that the sales data for May are wrong. There were actually an additional 25 books sold in May, but there was an error or something so they hadn’t told me about it. How can I fix my `sales.by.month` variable? One possibility would be to assign the whole vector again from the beginning, using `c()` . But that’s a lot of typing. Also, it’s a little wasteful: why should R have to redefine the sales figures for all 12 months, when only the 5th one is wrong? Fortunately, we can tell R to change only the 5th element, using this trick: \n\n```\nsales.by.month[5] <- 25\nsales.by.month\n```\n\n Another way to edit variables is to use the `edit()` or `fix()` functions. I won’t discuss them in detail right now, but you can check them out on your own. \n\n### 3.7.5 Useful things to know about vectors\n\n Before moving on, I want to mention a couple of other things about vectors. Firstly, you often find yourself wanting to know how many elements there are in a vector (usually because you’ve forgotten). You can use the `length()` function to do this. It’s quite straightforward: \n\n```\nlength( x = sales.by.month )\n```\n\n `## [1] 12` Secondly, you often want to alter all of the elements of a vector at once. For instance, suppose I wanted to figure out how much money I made in each month. Since I’m earning an exciting $7 per book (no seriously, that’s actually pretty close to what authors get on the very expensive textbooks that you’re expected to purchase), what I want to do is multiply each element in the `sales.by.month` vector by `7` . R makes this pretty easy, as the following example shows: `sales.by.month * 7` \n\n```\n## [1] 0 700 1400 350 175 0 0 0 0 0 0 0\n```\n\nIn other words, when you multiply a vector by a single number, all elements in the vector get multiplied. The same is true for addition, subtraction, division and taking powers. So that’s neat. On the other hand, suppose I wanted to know how much money I was making per day, rather than per month. Since not every month has the same number of days, I need to do something slightly different. Firstly, I’ll create two new vectors:\n\n```\ndays.per.month <- c(31, 28, 31, 30, 31, 30, 31, 31, 30, 31, 30, 31)\nprofit <- sales.by.month * 7\n```\n\n Obviously, the `profit` variable is the same one we created earlier, and the `days.per.month` variable is pretty straightforward. What I want to do is divide every element of `profit` by the corresponding element of `days.per.month` . Again, R makes this pretty easy: \n\n```\nprofit / days.per.month\n```\n\n```\n## [1] 0.000000 25.000000 45.161290 11.666667 5.645161 0.000000 0.000000\n## [8] 0.000000 0.000000 0.000000 0.000000 0.000000\n```\n\n I still don’t like all those zeros, but that’s not what matters here. Notice that the second element of the output is 25, because R has divided the second element of `profit` (i.e. 700) by the second element of `days.per.month` (i.e. 28). Similarly, the third element of the output is equal to 1400 divided by 31, and so on. We’ll talk more about calculations involving vectors later on (and in particular a thing called the “recycling rule”; Section 7.12.2), but that’s enough detail for now. \n\n## 3.8 Storing text data\n\nA lot of the time your data will be numeric in nature, but not always. Sometimes your data really needs to be described using text, not using numbers. To address this, we need to consider the situation where our variables store text. To create a variable that stores the word “hello”, we can type this:\n\n```\ngreeting <- \"hello\"\ngreeting\n```\n\n `## [1] \"hello\"` When interpreting this, it’s important to recognise that the quote marks here aren’t part of the string itself. They’re just something that we use to make sure that R knows to treat the characters that they enclose as a piece of text data, known as a character string. In other words, R treats `\"hello\"` as a string containing the word “hello”; but if I had typed `hello` instead, R would go looking for a variable by that name! You can also use `'hello'` to specify a character string. Okay, so that’s how we store the text. Next, it’s important to recognise that when we do this, R stores the entire word `\"hello\"` as a single element: our `greeting` variable is not a vector of five different letters. Rather, it has only the one element, and that element corresponds to the entire character string `\"hello\"` . To illustrate this, if I actually ask R to find the first element of `greeting` , it prints the whole string: `greeting[1]` `## [1] \"hello\"` Of course, there’s no reason why I can’t create a vector of character strings. For instance, if we were to continue with the example of my attempts to look at the monthly sales data for my book, one variable I might want would include the names of all 12 `months` .36 To do so, I could type in a command like this \n\n```\nmonths <- c(\"January\", \"February\", \"March\", \"April\", \"May\", \"June\",\n \"July\", \"August\", \"September\", \"October\", \"November\", \n \"December\")\n```\n\nThis is a character vector containing 12 elements, each of which is the name of a month. So if I wanted R to tell me the name of the fourth month, all I would do is this:\n `months[4]` `## [1] \"April\"` \n\n### 3.8.1 Working with text\n\n Working with text data is somewhat more complicated than working with numeric data, and I discuss some of the basic ideas in Section 7.8, but for purposes of the current chapter we only need this bare bones sketch. The only other thing I want to do before moving on is show you an example of a function that can be applied to text data. So far, most of the functions that we have seen (i.e., `sqrt()` , `abs()` and `round()` ) only make sense when applied to numeric data (e.g., you can’t calculate the square root of “hello”), and we’ve seen one function that can be applied to pretty much any variable or vector (i.e., `length()` ). So it might be nice to see an example of a function that can be applied to text. The function I’m going to introduce you to is called `nchar()` , and what it does is count the number of individual characters that make up a string. Recall earlier that when we tried to calculate the `length()` of our `greeting` variable it returned a value of `1` : the `greeting` variable contains only the one string, which happens to be `\"hello\"` . But what if I want to know how many letters there are in the word? Sure, I could count them, but that’s boring, and more to the point it’s a terrible strategy if what I wanted to know was the number of letters in War and Peace. That’s where the `nchar()` function is helpful: \n\n```\nnchar( x = greeting )\n```\n\n `## [1] 5` That makes sense, since there are in fact 5 letters in the string `\"hello\"` . Better yet, you can apply `nchar()` to whole vectors. So, for instance, if I want R to tell me how many letters there are in the names of each of the 12 months, I can do this: `nchar( x = months )` \n\n```\n## [1] 7 8 5 5 3 4 4 6 9 7 8 8\n```\n\n So that’s nice to know. The `nchar()` function can do a bit more than this, and there’s a lot of other functions that you can do to extract more information from text or do all sorts of fancy things. However, the goal here is not to teach any of that! The goal right now is just to see an example of a function that actually does work when applied to text. \n\n## 3.9 Storing “true or false” data\n\n Time to move onto a third kind of data. A key concept in that a lot of R relies on is the idea of a logical value. A logical value is an assertion about whether something is true or false. This is implemented in R in a pretty straightforward way. There are two logical values, namely `TRUE` and `FALSE` . Despite the simplicity, a logical values are very useful things. Let’s see how they work. \n\n### 3.9.1 Assessing mathematical truths\n\n In ’s classic book 1984, one of the slogans used by the totalitarian Party was “two plus two equals five”, the idea being that the political domination of human freedom becomes complete when it is possible to subvert even the most basic of truths. It’s a terrifying thought, especially when the protagonist finally breaks down under torture and agrees to the proposition. “Man is infinitely malleable”, the book says. I’m pretty sure that this isn’t true of humans37 but it’s definitely not true of R. R is not infinitely malleable. It has rather firm opinions on the topic of what is and isn’t true, at least as regards basic mathematics. If I ask it to calculate `2 + 2` , it always gives the same answer, and it’s not bloody 5: `2 + 2` `## [1] 4` \nOf course, so far R is just doing the calculations. I haven’t asked it to explicitly assert that \\(2+2 = 4\\) is a true statement. If I want R to make an explicit judgement, I can use a command like this:\n `2 + 2 == 4` `## [1] TRUE` What I’ve done here is use the equality operator, `==` , to force R to make a “true or false” judgement.38 Okay, let’s see what R thinks of the Party slogan: `2+2 == 5` `## [1] FALSE` Booyah! Freedom and ponies for all! Or something like that. Anyway, it’s worth having a look at what happens if I try to force R to believe that two plus two is five by making an assignment statement like `2 + 2 = 5` or `2 + 2 <- 5` . When I do this, here’s what happens: `2 + 2 = 5` \n\n```\n## Error in 2 + 2 = 5: target of assignment expands to non-language object\n```\n\n R doesn’t like this very much. It recognises that `2 + 2` is not a variable (that’s what the “non-language object” part is saying), and it won’t let you try to “reassign” it. While R is pretty flexible, and actually does let you do some quite remarkable things to redefine parts of R itself, there are just some basic, primitive truths that it refuses to give up. It won’t change the laws of addition, and it won’t change the definition of the number `2` . \nThat’s probably for the best.\n\n### 3.9.2 Logical operations\n\nSo now we’ve seen logical operations at work, but so far we’ve only seen the simplest possible example. You probably won’t be surprised to discover that we can combine logical operations with other operations and functions in a more complicated way, like this:\n `3*3 + 4*4 == 5*5` `## [1] TRUE` \nor this\n `sqrt( 25 ) == 5` `## [1] TRUE` \nNot only that, but as Table 3.2 illustrates, there are several other logical operators that you can use, corresponding to some basic mathematical concepts.\n\noperation | operator | example input | answer |\n | --- | --- | --- | --- |\nless than | < | 2 < 3 | |\nless than or equal to | <= | 2 <= 2 | |\ngreater than | > | 2 > 3 | |\ngreater than or equal to | >= | 2 >= 2 | |\nequal to | == | 2 == 3 | |\nnot equal to | != | 2 != 3 | |\n\n Hopefully these are all pretty self-explanatory: for example, the less than operator `<` checks to see if the number on the left is less than the number on the right. If it’s less, then R returns an answer of `TRUE` : `99 < 100` `## [1] TRUE` but if the two numbers are equal, or if the one on the right is larger, then R returns an answer of `FALSE` , as the following two examples illustrate: `100 < 100` `## [1] FALSE` `100 < 99` `## [1] FALSE` In contrast, the less than or equal to operator `<=` will do exactly what it says. It returns a value of `TRUE` if the number of the left hand side is less than or equal to the number on the right hand side. So if we repeat the previous two examples using `<=` , here’s what we get: `100 <= 100` `## [1] TRUE` `100 <= 99` `## [1] FALSE` And at this point I hope it’s pretty obvious what the greater than operator `>` and the greater than or equal to operator `>=` do! Next on the list of logical operators is the not equal to operator `!=` which – as with all the others – does what it says it does. It returns a value of `TRUE` when things on either side are not identical to each other. Therefore, since \\(2+2\\) isn’t equal to \\(5\\), we get: `2 + 2 != 5` `## [1] TRUE` \nWe’re not quite done yet. There are three more logical operations that are worth knowing about, listed in Table 3.3.\n\noperation | operator | example input | answer |\n | --- | --- | --- | --- |\nnot | ! | !(1==1) | |\nor | | | (1==1) | (2==3) | |\nand | & | (1==1) & (2==3) | |\n\n These are the not operator `!` , the and operator `&` , and the or operator `|` . Like the other logical operators, their behaviour is more or less exactly what you’d expect given their names. For instance, if I ask you to assess the claim that “either \\(2+2 = 4\\) or \\(2+2 = 5\\)” you’d say that it’s true. Since it’s an “either-or” statement, all we need is for one of the two parts to be true. That’s what the `|` operator does: \n\n```\n(2+2 == 4) | (2+2 == 5)\n```\n\n `## [1] TRUE` On the other hand, if I ask you to assess the claim that “both \\(2+2 = 4\\) and \\(2+2 = 5\\)” you’d say that it’s false. Since this is an and statement we need both parts to be true. And that’s what the `&` operator does: \n\n```\n(2+2 == 4) & (2+2 == 5)\n```\n\n `## [1] FALSE` \nFinally, there’s the not operator, which is simple but annoying to describe in English. If I ask you to assess my claim that “it is not true that \\(2+2 = 5\\)” then you would say that my claim is true; because my claim is that “\\(2+2 = 5\\) is false”. And I’m right. If we write this as an R command we get this:\n `! (2+2 == 5)` `## [1] TRUE` In other words, since `2+2 == 5` is a `FALSE` statement, it must be the case that `!(2+2 == 5)` is a `TRUE` one. Essentially, what we’ve really done is claim that “not false” is the same thing as “true”. Obviously, this isn’t really quite right in real life. But R lives in a much more black or white world: for R everything is either true or false. No shades of gray are allowed. We can actually see this much more explicitly, like this: `! FALSE` `## [1] TRUE` Of course, in our \\(2+2 = 5\\) example, we didn’t really need to use “not” `!` and “equals to” `==` as two separate operators. We could have just used the “not equals to” operator `!=` like this: `2+2 != 5` `## [1] TRUE` But there are many situations where you really do need to use the `!` operator. We’ll see some later on.39 \n\n### 3.9.3 Storing and using logical data\n\n Up to this point, I’ve introduced numeric data (in Sections 3.4 and 3.7) and character data (in Section 3.8). So you might not be surprised to discover that these `TRUE` and `FALSE` values that R has been producing are actually a third kind of data, called logical data. That is, when I asked R if `2 + 2 == 5` and it said `[1] FALSE` in reply, it was actually producing information that we can store in variables. For instance, I could create a variable called `is.the.Party.correct` , which would store R’s opinion: \n\n```\nis.the.Party.correct <- 2 + 2 == 5\nis.the.Party.correct\n```\n\n `## [1] FALSE` Alternatively, you can assign the value directly, by typing `TRUE` or `FALSE` in your command. Like this: \n\n```\nis.the.Party.correct <- FALSE\nis.the.Party.correct\n```\n\n `## [1] FALSE` Better yet, because it’s kind of tedious to type `TRUE` or `FALSE` over and over again, R provides you with a shortcut: you can use `T` and `F` instead (but it’s case sensitive: `t` and `f` won’t work).40 So this works: \n\n```\nis.the.Party.correct <- F\nis.the.Party.correct\n```\n\n `## [1] FALSE` \nbut this doesn’t:\n\n```\nis.the.Party.correct <- f\n```\n\n### 3.9.4 Vectors of logicals\n\n The next thing to mention is that you can store vectors of logical values in exactly the same way that you can store vectors of numbers (Section 3.7) and vectors of text data (Section 3.8). Again, we can define them directly via the `c()` function, like this: \n\n```\nx <- c(TRUE, TRUE, FALSE)\nx\n```\n\n or you can produce a vector of logicals by applying a logical operator to a vector. This might not make a lot of sense to you, so let’s unpack it slowly. First, let’s suppose we have a vector of numbers (i.e., a “non-logical vector”). For instance, we could use the `sales.by.month` vector that we were using in Section 3.7. Suppose I wanted R to tell me, for each month of the year, whether I actually sold a book in that month. I can do that by typing this: `sales.by.month > 0` \n\nand again, I can store this in a vector if I want, as the example below illustrates:\n\n```\nany.sales.this.month <- sales.by.month > 0\nany.sales.this.month\n```\n\n In other words, `any.sales.this.month` is a logical vector whose elements are `TRUE` only if the corresponding element of `sales.by.month` is greater than zero. For instance, since I sold zero books in January, the first element is `FALSE` . \n\n### 3.9.5 Applying logical operation to text\n\n In a moment (Section 3.10) I’ll show you why these logical operations and logical vectors are so handy, but before I do so I want to very briefly point out that you can apply them to text as well as to logical data. It’s just that we need to be a bit more careful in understanding how R interprets the different operations. In this section I’ll talk about how the equal to operator `==` applies to text, since this is the most important one. Obviously, the not equal to operator `!=` gives the exact opposite answers to `==` so I’m implicitly talking about that one too, but I won’t give specific commands showing the use of `!=` . As for the other operators, I’ll defer a more detailed discussion of this topic to Section 7.8.5. Okay, let’s see how it works. In one sense, it’s very simple. For instance, I can ask R if the word `\"cat\"` is the same as the word `\"dog\"` , like this: `\"cat\" == \"dog\"` `## [1] FALSE` That’s pretty obvious, and it’s good to know that even R can figure that out. Similarly, R does recognise that a `\"cat\"` is a `\"cat\"` : `\"cat\" == \"cat\"` `## [1] TRUE` \nAgain, that’s exactly what we’d expect. However, what you need to keep in mind is that R is not at all tolerant when it comes to grammar and spacing. If two strings differ in any way whatsoever, R will say that they’re not equal to each other, as the following examples indicate:\n `\" cat\" == \"cat\"` `## [1] FALSE` `\"cat\" == \"CAT\"` `## [1] FALSE` `\"cat\" == \"c a t\"` `## [1] FALSE` \n\n## 3.10 Indexing vectors\n\n One last thing to add before finishing up this chapter. So far, whenever I’ve had to get information out of a vector, all I’ve done is typed something like `months[4]` ; and when I do this R prints out the fourth element of the `months` vector. In this section, I’ll show you two additional tricks for getting information out of the vector. \n\n### 3.10.1 Extracting multiple elements\n\n One very useful thing we can do is pull out more than one element at a time. In the previous example, we only used a single number (i.e., `2` ) to indicate which element we wanted. Alternatively, we can use a vector. So, suppose I wanted the data for February, March and April. What I could do is use the vector `c(2,3,4)` to indicate which elements I want R to pull out. That is, I’d type this: \n\n```\nsales.by.month[ c(2,3,4) ]\n```\n\n `## [1] 100 200 50` Notice that the order matters here. If I asked for the data in the reverse order (i.e., April first, then March, then February) by using the vector `c(4,3,2)` , then R outputs the data in the reverse order: \n\n```\nsales.by.month[ c(4,3,2) ]\n```\n\n `## [1] 50 200 100` A second thing to be aware of is that R provides you with handy shortcuts for very common situations. For instance, suppose that I wanted to extract everything from the 2nd month through to the 8th month. One way to do this is to do the same thing I did above, and use the vector `c(2,3,4,5,6,7,8)` to indicate the elements that I want. That works just fine \n\n```\nsales.by.month[ c(2,3,4,5,6,7,8) ]\n```\n\n but it’s kind of a lot of typing. To help make this easier, R lets you use `2:8` as shorthand for `c(2,3,4,5,6,7,8)` , which makes things a lot simpler. First, let’s just check that this is true: `2:8` `## [1] 2 3 4 5 6 7 8` Next, let’s check that we can use the `2:8` shorthand as a way to pull out the 2nd through 8th elements of `sales.by.months` : `sales.by.month[2:8]` \n\nSo that’s kind of neat.\n\n### 3.10.2 Logical indexing\n\n At this point, I can introduce an extremely useful tool called logical indexing. In the last section, I created a logical vector `any.sales.this.month` , whose elements are `TRUE` for any month in which I sold at least one book, and `FALSE` for all the others. However, that big long list of `TRUE` s and `FALSE` s is a little bit hard to read, so what I’d like to do is to have R select the names of the `months` for which I sold any books. Earlier on, I created a vector `months` that contains the names of each of the months. This is where logical indexing is handy. What I need to do is this: \n\n```\nmonths[ sales.by.month > 0 ]\n```\n\n To understand what’s happening here, it’s helpful to notice that `sales.by.month > 0` is the same logical expression that we used to create the `any.sales.this.month` vector in the last section. In fact, I could have just done this: \n\n```\nmonths[ any.sales.this.month ]\n```\n\n and gotten exactly the same result. In order to figure out which elements of `months` to include in the output, what R does is look to see if the corresponding element in `any.sales.this.month` is `TRUE` . Thus, since element 1 of `any.sales.this.month` is `FALSE` , R does not include `\"January\"` as part of the output; but since element 2 of `any.sales.this.month` is `TRUE` , R does include `\"February\"` in the output. Note that there’s no reason why I can’t use the same trick to find the actual sales numbers for those months. The command to do that would just be this: \n\n```\nsales.by.month [ sales.by.month > 0 ]\n```\n\n```\n## [1] 100 200 50 25\n```\n\n In fact, we can do the same thing with text. Here’s an example. Suppose that – to continue the saga of the textbook sales – I later find out that the bookshop only had sufficient stocks for a few months of the year. They tell me that early in the year they had `\"high\"` stocks, which then dropped to `\"low\"` levels, and in fact for one month they were `\"out\"` of copies of the book for a while before they were able to replenish them. Thus I might have a variable called `stock.levels` which looks like this: \n\n```\nstock.levels<-c(\"high\", \"high\", \"low\", \"out\", \"out\", \"high\",\n \"high\", \"high\", \"high\", \"high\", \"high\", \"high\")\n\nstock.levels\n```\n\n```\n## [1] \"high\" \"high\" \"low\" \"out\" \"out\" \"high\" \"high\" \"high\" \"high\" \"high\"\n## [11] \"high\" \"high\"\n```\n\n Thus, if I want to know the months for which the bookshop was out of my book, I could apply the logical indexing trick, but with the character vector `stock.levels` , like this: \n\n```\nmonths[stock.levels == \"out\"]\n```\n\n `## [1] \"April\" \"May\"` \nAlternatively, if I want to know when the bookshop was either low on copies or out of copies, I could do this:\n\n```\nmonths[stock.levels == \"out\" | stock.levels == \"low\"]\n```\n\nor this\n\n```\nmonths[stock.levels != \"high\" ]\n```\n\nEither way, I get the answer I want.\n\nAt this point, I hope you can see why logical indexing is such a useful thing. It’s a very basic, yet very powerful way to manipulate data. We’ll talk a lot more about how to manipulate data in Chapter 7, since it’s a critical skill for real world research that is often overlooked in introductory research methods classes (or at least, that’s been my experience). It does take a bit of practice to become completely comfortable using logical indexing, so it’s a good idea to play around with these sorts of commands. Try creating a few different variables of your own, and then ask yourself questions like “how do I get R to spit out all the elements that are [blah]”. Practice makes perfect, and it’s only by practicing logical indexing that you’ll perfect the art of yelling frustrated insults at your computer.41\n\n## 3.11 Quitting R\n\n There’s one last thing I should cover in this chapter: how to quit R. When I say this, I’m not trying to imply that R is some kind of pathological addition and that you need to call the R QuitLine or wear patches to control the cravings (although you certainly might argue that there’s something seriously pathological about being addicted to R). I just mean how to exit the program. Assuming you’re running R in the usual way (i.e., through RStudio or the default GUI on a Windows or Mac computer), then you can just shut down the application in the normal way. However, R also has a function, called `q()` that you can use to quit, which is pretty handy if you’re running R in a terminal window. \nRegardless of what method you use to quit R, when you do so for the first time R will probably ask you if you want to save the “workspace image”. We’ll talk a lot more about loading and saving data in Section 4.5, but I figured we’d better quickly cover this now otherwise you’re going to get annoyed when you close R at the end of the chapter. If you’re using RStudio, you’ll see a dialog box that looks like the one shown in Figure 3.5. If you’re using a text based interface you’ll see this:\n\n```\nq()\n\n## Save workspace image? [y/n/c]:\n```\n\n The `y/n/c` part here is short for “yes / no / cancel”. Type `y` if you want to save, `n` if you don’t, and `c` if you’ve changed your mind and you don’t want to quit after all. What does this actually mean? What’s going on is that R wants to know if you want to save all those variables that you’ve been creating, so that you can use them later. This sounds like a great idea, so it’s really tempting to type `y` or click the “Save” button. To be honest though, I very rarely do this, and it kind of annoys me a little bit… what R is really asking is if you want it to store these variables in a “default” data file, which it will automatically reload for you next time you open R. And quite frankly, if I’d wanted to save the variables, then I’d have already saved them before trying to quit. Not only that, I’d have saved them to a location of my choice, so that I can find it again later. So I personally never bother with this. \nIn fact, every time I install R on a new machine one of the first things I do is change the settings so that it never asks me again. You can do this in RStudio really easily: use the menu system to find the RStudio option; the dialog box that comes up will give you an option to tell R never to whine about this again (see Figure 3.6. On a Mac, you can open this window by going to the “RStudio” menu and selecting “Preferences”. On a Windows machine you go to the “Tools” menu and select “Global Options”. Under the “General” tab you’ll see an option that reads “Save workspace to .Rdata on exit”. By default this is set to “ask”. If you want R to stop asking, change it to “never”.\n\n## 3.12 Summary\n\nEvery book that tries to introduce basic programming ideas to novices has to cover roughly the same topics, and in roughly the same order. Mine is no exception, and so in the grand tradition of doing it just the same way everyone else did it, this chapter covered the following topics:\n\n* Getting started. We downloaded and installed R and RStudio\n * Basic commands. We talked a bit about the logic of how R works and in particular how to type commands into the R console (Section @ref(#firstcommand), and in doing so learned how to perform basic calculations using the arithmetic operators\n `+` , `-` , `*` , `/` and `^` . * Introduction to functions. We saw several different functions, three that are used to perform numeric calculations (\n `sqrt()` , `abs()` , `round()` , one that applies to text ( `nchar()` ; Section 3.8.1), and one that works on any variable ( `length()` ; Section 3.7.5). In doing so, we talked a bit about how argument names work, and learned about default values for arguments. (Section 3.5.1) * Introduction to variables. We learned the basic idea behind variables, and how to assign values to variables using the assignment operator\n `<-` (Section 3.4). We also learned how to create vectors using the combine function `c()` (Section 3.7). * Data types. Learned the distinction between numeric, character and logical data; including the basics of how to enter and use each of them. (Sections 3.4 to 3.9)\n * Logical operations. Learned how to use the logical operators\n `==` , `!=` , `<` , `>` , `<=` , `=>` , `!` , `&` and `|` . And learned how to use logical indexing. (Section 3.10) \nWe still haven’t arrived at anything that resembles a “data set”, of course. Maybe the next Chapter will get us a bit closer…\n\nSource: Dismal Light (1968).↩\n * \n\nAlthough R is updated frequently, it doesn’t usually make much of a difference for the sort of work we’ll do in this book. In fact, during the writing of the book I upgraded several times, and didn’t have to change much except these sections describing the downloading.↩\n * \n\nIf you’re running an older version of the Mac OS, then you need to follow the link to the “old” page (http://cran.r-project.org/bin/macosx/old/). You should be able to find the installer file that you need at the bottom of the page.↩\n * \n\nTip for advanced Mac users. You can run R from the terminal if you want to. The command is just “R”. It behaves like the normal desktop version, except that help documentation behaves like a “man” page instead of opening in a new window.↩\n * \n\nThis is probably no coincidence: the people who design and distribute the core R language itself are focused on technical stuff. And sometimes they almost seem to forget that there’s an actual human user at the end. The people who design and distribute RStudio are focused on user interface. They want to make R as usable as possible. The two websites reflect that difference.↩\n * \n\nSeriously. If you’re in a position to do so, open up R and start typing. The simple act of typing it rather than “just reading” makes a big difference. It makes the concepts more concrete, and it ties the abstract ideas (programming and statistics) to the actual context in which you need to use them. Statistics is something you do, not just something you read about in a textbook.↩\n * \n\nIf you’re running R from the terminal rather than from RStudio, escape doesn’t work: use CTRL-C instead.↩\n * \n\nFor advanced users: yes, as you’ve probably guessed, R is printing out the source code for the function.↩\n * \n\nIf you’re reading this with R open, a good learning trick is to try typing in a few different variations on what I’ve done here. If you experiment with your commands, you’ll quickly learn what works and what doesn’t↩\n * \n\nFor advanced users: if you want a table showing the complete order of operator precedence in R, type\n `?Syntax` . I haven’t included it in this book since there are quite a few different operators, and we don’t need that much detail. Besides, in practice most people seem to figure it out from seeing examples: until writing this book I never looked at the formal statement of operator precedence for any language I ever coded in, and never ran into any difficulties.↩ * \n\nIf you are using RStudio, and the “environment” panel (formerly known as the “workspace” panel) is visible when you typed the command, then you probably saw something happening there. That’s to be expected, and is quite helpful. However, there’s two things to note here (1) I haven’t yet explained what that panel does, so for now just ignore it, and (2) this is one of the helpful things RStudio does, not a part of R itself.↩\n * \n\nAs we’ll discuss later, by doing this we are implicitly using the\n `print()` function↩ * \n\nActually, in keeping with the R tradition of providing you with a billion different screwdrivers (even when you’re actually looking for a hammer) these aren’t the only options. There’s also the\n `assign()` function, and the `<<-` and `->>` operators. However, we won’t be using these at all in this book.↩ * \n\nA quick reminder: when using operators like\n `<-` and `->` that span multiple characters, you can’t insert spaces in the middle. That is, if you type `- >` or `< -` , R will interpret your command the wrong way. And I will cry.↩ * \n\nActually, you can override any of these rules if you want to, and quite easily. All you have to do is add quote marks or backticks around your non-standard variable name. For instance\n ``my sales ` <- 350` would work just fine, but it’s almost never a good idea to do this.↩ * \n\nFor very advanced users: there is one exception to this. If you’re naming a function, don’t use\n `.` in the name unless you are intending to make use of the S3 object oriented programming system in R. If you don’t know what S3 is, then you definitely don’t want to be using it! For function naming, there’s been a trend among R users to prefer `myFunctionName` .↩ * \n\nA side note for students with a programming background. Technically speaking, operators are functions in R: the addition operator\n `+` is actually a convenient way of calling the addition function `+()` . Thus `10+20` is equivalent to the function call `+(20, 30)` . Not surprisingly, no-one ever uses this version. Because that would be stupid.↩ * \n\nA note for the mathematically inclined: R does support complex numbers, but unless you explicitly specify that you want them it assumes all calculations must be real valued. By default, the square root of a negative number is treated as undefined:\n `sqrt(-9)` will produce `NaN` (not a number) as its output. To get complex numbers, you would type `sqrt(-9+0i)` and R would now return `0+3i` . However, since we won’t have any need for complex numbers in this book, I won’t refer to them again.↩ * \n\nThe two functions discussed previously,\n `sqrt()` and `abs()` , both only have a single argument, `x` . So I could have typed something like `sqrt(x = 225)` or `abs(x = -13)` earlier. The fact that all these functions use `x` as the name of the argument that corresponds the “main” variable that you’re working with is no coincidence. That’s a fairly widely used convention. Quite often, the writers of R functions will try to use conventional names like this to make your life easier. Or at least that’s the theory. In practice it doesn’t always work as well as you’d hope.↩ * \n\nFor advanced users: obviously, this isn’t just an RStudio thing. If you’re running R in a terminal window, tab autocomplete still works, and does so in exactly the way you’d expect. It’s not as visually pretty as the RStudio version, of course, and lacks some of the cooler features that RStudio provides. I don’t bother to document that here: my assumption is that if you are running R in the terminal then you’re already familiar with using tab autocomplete.↩\n * \n\nIncidentally, that always works: if you’ve started typing a command and you want to clear it and start again, hit escape.↩\n * \n\nAnother method is to start typing some text and then hit the Control key and the up arrow together (on Windows or Linux) or the Command key and the up arrow together (on a Mac). This will bring up a window showing all your recent commands that started with the same text as what you’ve currently typed. That can come in quite handy sometimes.↩\n * \n\nNotice that I didn’t specify any argument names here. The\n `c()` function is one of those cases where we don’t use names. We just type all the numbers, and R just dumps them all in a single variable.↩ * \n\nThough actually there’s no real need to do this, since R has an inbuilt variable called\n `month.name` that you can use for this purpose.↩ * \n\nI offer up my teenage attempts to be “cool” as evidence that some things just can’t be done.↩\n * \n\nNote that this is a very different operator to the assignment operator\n `=` that I talked about in Section 3.4. A common typo that people make when trying to write logical commands in R (or other languages, since the “ `=` versus `==` ” distinction is important in most programming languages) is to accidentally type `=` when you really mean `==` . Be especially cautious with this – I’ve been programming in various languages since I was a teenager, and I still screw this up a lot. Hm. I think I see why I wasn’t cool as a teenager. And why I’m still not cool.↩ * \n\nA note for those of you who have taken a computer science class: yes, R does have a function for exclusive-or, namely\n `xor()` . Also worth noting is the fact that R makes the distinction between element-wise operators `&` and `|` and operators that look only at the first element of the vector, namely `&&` and `||` . To see the distinction, compare the behaviour of a command like \n\n```\nc(FALSE,TRUE) & c(TRUE,TRUE)\n```\n\nto the behaviour of something like \n\n```\nc(FALSE,TRUE) && c(TRUE,TRUE)\n```\n\n. If this doesn’t mean anything to you, ignore this footnote entirely. It’s not important for the content of this book.↩ * \n\nWarning!\n `TRUE` and `FALSE` are reserved keywords in R, so you can trust that they always mean what they say they do. Unfortunately, the shortcut versions `T` and `F` do not have this property. It’s even possible to create variables that set up the reverse meanings, by typing commands like `T <- FALSE` and `F <- TRUE` . This is kind of insane, and something that is generally thought to be a design flaw in R. Anyway, the long and short of it is that it’s safer to use `TRUE` and `FALSE` .↩ * \n\nWell, I say that… but in my personal experience it wasn’t until I started learning “regular expressions” that my loathing of computers reached its peak.↩\n\nDate: 2011-09-12\nCategories: \nTags: \n\n# Chapter 4 Additional R concepts\n\nForm follows function\n\n– \n\nIn Chapter 3 our main goal was to get started in R. As we go through the book we’ll run into a lot of new R concepts, which I’ll explain alongside the relevant data analysis concepts. However, there’s still quite a few things that I need to talk about now, otherwise we’ll run into problems when we start trying to work with data and do statistics. So that’s the goal in this chapter: to build on the introductory content from the last chapter, to get you to the point that we can start using R for statistics. Broadly speaking, the chapter comes in two parts. The first half of the chapter is devoted to the “mechanics” of R: installing and loading packages, managing the workspace, navigating the file system, and loading and saving data. In the second half, I’ll talk more about what kinds of variables exist in R, and introduce three new kinds of variables: factors, data frames and formulas. I’ll finish up by talking a little bit about the help documentation in R as well as some other avenues for finding assistance. In general, I’m not trying to be comprehensive in this chapter, I’m trying to make sure that you’ve got the basic foundations needed to tackle the content that comes later in the book. However, a lot of the topics are revisited in more detail later, especially in Chapters 7 and 8.\n\n## 4.1 Using comments\n\n Before discussing any of the more complicated stuff, I want to introduce the comment character, `#` . It has a simple meaning: it tells R to ignore everything else you’ve written on this line. You won’t have much need of the `#` character immediately, but it’s very useful later on when writing scripts (see Chapter 8). However, while you don’t need to use it, I want to be able to include comments in my R extracts. For instance, if you read this:42 \n\n```\nseeker <- 3.1415 # create the first variable\nlover <- 2.7183 # create the second variable\nkeeper <- seeker * lover # now multiply them to create a third one\nprint( keeper ) # print out the value of 'keeper'\n```\n\n `## [1] 8.539539` \nit’s a lot easier to understand what I’m doing than if I just write this:\n\n```\nseeker <- 3.1415\nlover <- 2.7183\nkeeper <- seeker * lover\nprint( keeper )\n```\n\n `## [1] 8.539539` You might have already noticed that the code extracts in Chapter 3 included the `#` character, but from now on, you’ll start seeing `#` characters appearing in the extracts, with some human-readable explanatory remarks next to them. These are still perfectly legitimate commands, since R knows that it should ignore the `#` character and everything after it. But hopefully they’ll help make things a little easier to understand. \n\n## 4.2 Installing and loading packages\n\nIn this section I discuss R packages, since almost all of the functions you might want to use in R come in packages. A package is basically just a big collection of functions, data sets and other R objects that are all grouped together under a common name. Some packages are already installed when you put R on your computer, but the vast majority of them of R packages are out there on the internet, waiting for you to download, install and use them.\n\nWhen I first started writing this book, RStudio didn’t really exist as a viable option for using R, and as a consequence I wrote a very lengthy section that explained how to do package management using raw R commands. It’s not actually terribly hard to work with packages that way, but it’s clunky and unpleasant. Fortunately, we don’t have to do things that way anymore. In this section, I’ll describe how to work with packages using the RStudio tools, because they’re so much simpler. Along the way, you’ll see that whenever you get RStudio to do something (e.g., install a package), you’ll actually see the R commands that get created. I’ll explain them as we go, because I think that helps you understand what’s going on.\n\nHowever, before we get started, there’s a critical distinction that you need to understand, which is the difference between having a package installed on your computer, and having a package loaded in R. As of this writing, there are just over 5000 R packages freely available “out there” on the internet.43 When you install R on your computer, you don’t get all of them: only about 30 or so come bundled with the basic R installation. So right now there are about 30 packages “installed” on your computer, and another 5000 or so that are not installed. So that’s what installed means: it means “it’s on your computer somewhere”. The critical thing to remember is that just because something is on your computer doesn’t mean R can use it. In order for R to be able to use one of your 30 or so installed packages, that package must also be “loaded”. Generally, when you open up R, only a few of these packages (about 7 or 8) are actually loaded. Basically what it boils down to is this:\n A package must be installed before it can be loaded.\n A package must be loaded before it can be used.\n\nThis two step process might seem a little odd at first, but the designers of R had very good reasons to do it this way,44 and you get the hang of it pretty quickly.\n\n### 4.2.1 The package panel in RStudio\n\nRight, lets get started. The first thing you need to do is look in the lower right hand panel in RStudio. You’ll see a tab labelled “Packages”. Click on the tab, and you’ll see a list of packages that looks something like Figure 4.1. Every row in the panel corresponds to a different package, and every column is a useful piece of information about that package.45 Going from left to right, here’s what each column is telling you:\n\n* The check box on the far left column indicates whether or not the package is loaded.\n * The one word of text immediately to the right of the check box is the name of the package.\n * The short passage of text next to the name is a brief description of the package.\n * The number next to the description tells you what version of the package you have installed.\n * The little x-mark next to the version number is a button that you can push to uninstall the package from your computer (you almost never need this).\n\n### 4.2.2 Loading a package\n\n That seems straightforward enough, so let’s try loading and unloading packades. For this example, I’ll use the `foreign` package. The `foreign` package is a collection of tools that are very handy when R needs to interact with files that are produced by other software packages (e.g., SPSS). It comes bundled with R, so it’s one of the ones that you have installed already, but it won’t be one of the ones loaded. Inside the `foreign` package is a function called `read.spss()` . It’s a handy little function that you can use to import an SPSS data file into R, so let’s pretend we want to use it. Currently, the `foreign` package isn’t loaded, so if I ask R to tell me if it knows about a function called `read.spss()` it tells me that there’s no such thing… \n\n `## [1] FALSE` Now let’s load the package. In RStudio, the process is dead simple: go to the package tab, find the entry for the `foreign` package, and check the box on the left hand side. The moment that you do this, you’ll see a command like this appear in the R console: \n\n```\nlibrary(\"foreign\", lib.loc=\"/Library/Frameworks/R.framework/Versions/3.0/Resources/library\")\n```\n\n The `lib.loc` bit will look slightly different on Macs versus on Windows, because that part of the command is just RStudio telling R where to look to find the installed packages. What I’ve shown you above is the Mac version. On a Windows machine, you’ll probably see something that looks like this: \n\n```\nlibrary(\"foreign\", lib.loc=\"C:/Program Files/R/R-3.0.2/library\")\n```\n\n But actually it doesn’t matter much. The `lib.loc` bit is almost always unnecessary. Unless you’ve taken to installing packages in idiosyncratic places (which is something that you can do if you really want) R already knows where to look. So in the vast majority of cases, the command to load the `foreign` package is just this: `library(\"foreign\")` Throughout this book, you’ll often see me typing in `library()` commands. You don’t actually have to type them in yourself: you can use the RStudio package panel to do all your package loading for you. The only reason I include the `library()` commands sometimes is as a reminder to you to make sure that you have the relevant package loaded. Oh, and I suppose we should check to see if our attempt to load the package actually worked. Let’s see if R now knows about the existence of the `read.spss()` function… \n\n `## [1] TRUE` \nYep. All good.\n\n### 4.2.3 Unloading a package\n\n Sometimes, especially after a long session of working with R, you find yourself wanting to get rid of some of those packages that you’ve loaded. The RStudio package panel makes this exactly as easy as loading the package in the first place. Find the entry corresponding to the package you want to unload, and uncheck the box. When you do that for the `foreign` package, you’ll see this command appear on screen: \n\n```\ndetach(\"package:foreign\", unload=TRUE)\n```\n\n```\n## Warning: 'foreign' namespace cannot be unloaded:\n## namespace 'foreign' is imported by 'rio', 'psych' so cannot be unloaded\n```\n\n And the package is unloaded. We can verify this by seeing if the `read.spss()` function still `exists()` : \n\n `## [1] FALSE` \nNope. Definitely gone.\n\n### 4.2.4 A few extra comments\n\n Sections 4.2.2 and 4.2.3 cover the main things you need to know about loading and unloading packages. However, there’s a couple of other details that I want to draw your attention to. A concrete example is the best way to illustrate. One of the other packages that you already have installed on your computer is the `Matrix` package, so let’s load that one and see what happens: \n\n```\nlibrary( Matrix )\n\n## Loading required package: lattice\n```\n\n This is slightly more complex than the output that we got last time, but it’s not too complicated. The `Matrix` package makes use of some of the tools in the `lattice` package, and R has kept track of this dependency. So when you try to load the `Matrix` package, R recognises that you’re also going to need to have the `lattice` package loaded too. As a consequence, both packages get loaded, and R prints out a helpful little note on screen to tell you that it’s done so. R is pretty aggressive about enforcing these dependencies. Suppose, for example, I try to unload the `lattice` package while the `Matrix` package is still loaded. This is easy enough to try: all I have to do is uncheck the box next to “lattice” in the packages panel. But if I try this, here’s what happens: \n\n```\ndetach(\"package:lattice\", unload=TRUE)\n\n## Error: package `lattice' is required by `Matrix' so will not be detached\n```\n\n R refuses to do it. This can be quite useful, since it stops you from accidentally removing something that you still need. So, if I want to remove both `Matrix` and `lattice` , I need to do it in the correct order Something else you should be aware of. Sometimes you’ll attempt to load a package, and R will print out a message on screen telling you that something or other has been “masked”. This will be confusing to you if I don’t explain it now, and it actually ties very closely to the whole reason why R forces you to load packages separately from installing them. Here’s an example. Two of the package that I’ll refer to a lot in this book are called `car` and `psych` . The `car` package is short for “Companion to Applied Regression” (which is a really great book, I’ll add), and it has a lot of tools that I’m quite fond of. The `car` package was written by a guy called , who has written a lot of great statistical tools for social science applications. The `psych` package was written by , and it has a lot of functions that are very useful for psychologists in particular, especially in regards to psychometric techniques. For the most part, `car` and `psych` are quite unrelated to each other. They do different things, so not surprisingly almost all of the function names are different. But… there’s one exception to that. The `car` package and the `psych` package both contain a function called `logit()` .46 This creates a naming conflict. If I load both packages into R, an ambiguity is created. If the user types in `logit(100)` , should R use the `logit()` function in the `car` package, or the one in the `psych` package? The answer is: R uses whichever package you loaded most recently, and it tells you this very explicitly. Here’s what happens when I load the `car` package, and then afterwards load the `psych` package: \n\n```\nlibrary(car)\nlibrary(psych)\n```\n\n The output here is telling you that the `logit` object (i.e., function) in the `car` package is no longer accessible to you. It’s been hidden (or “masked”) from you by the one in the `psych` package.47 \n\n### 4.2.5 Downloading new packages\n\n One of the main selling points for R is that there are thousands of packages that have been written for it, and these are all available online. So whereabouts online are these packages to be found, and how do we download and install them? There is a big repository of packages called the “Comprehensive R Archive Network” (CRAN), and the easiest way of getting and installing a new package is from one of the many CRAN mirror sites. Conveniently for us, R provides a function called `install.packages()` that you can use to do this. Even more conveniently, the RStudio team runs its own CRAN mirror and RStudio has a clean interface that lets you install packages without having to learn how to use the `install.packages()` command48 \nUsing the RStudio tools is, again, dead simple. In the top left hand corner of the packages panel (Figure 4.1) you’ll see a button called “Install Packages”. If you click on that, it will bring up a window like the one shown in Figure 4.2.\n\nThere are a few different buttons and boxes you can play with. Ignore most of them. Just go to the line that says “Packages” and start typing the name of the package that you want. As you type, you’ll see a dropdown menu appear (Figure 4.3), listing names of packages that start with the letters that you’ve typed so far.\n\nYou can select from this list, or just keep typing. Either way, once you’ve got the package name that you want, click on the install button at the bottom of the window. When you do, you’ll see the following command appear in the R console:\n\n```\ninstall.packages(\"psych\")\n```\n\n This is the R command that does all the work. R then goes off to the internet, has a conversation with CRAN, downloads some stuff, and installs it on your computer. You probably don’t care about all the details of R’s little adventure on the web, but the `install.packages()` function is rather chatty, so it reports a bunch of gibberish that you really aren’t all that interested in: \n\n```\ntrying URL 'http://cran.rstudio.com/bin/macosx/contrib/3.0/psych_1.4.1.tgz'\nContent type 'application/x-gzip' length 2737873 bytes (2.6 Mb)\nopened URL\n==================================================\ndownloaded 2.6 Mb\n\nThe downloaded binary packages are in\n //downloaded_packages\n```\n\nDespite the long and tedious response, all thar really means is “I’ve installed the psych package”. I find it best to humour the talkative little automaton. I don’t actually read any of this garbage, I just politely say “thanks” and go back to whatever I was doing.\n\n### 4.2.6 Updating R and R packages\n\n Every now and then the authors of packages release updated versions. The updated versions often add new functionality, fix bugs, and so on. It’s generally a good idea to update your packages periodically. There’s an `update.packages()` function that you can use to do this, but it’s probably easier to stick with the RStudio tool. In the packages panel, click on the “Update Packages” button. This will bring up a window that looks like the one shown in Figure 4.4. In this window, each row refers to a package that needs to be updated. You can to tell R which updates you want to install by checking the boxes on the left. If you’re feeling lazy and just want to update everything, click the “Select All” button, and then click the “Install Updates” button. R then prints out a lot of garbage on the screen, individually downloading and installing all the new packages. This might take a while to complete depending on how good your internet connection is. Go make a cup of coffee. Come back, and all will be well. \nAbout every six months or so, a new version of R is released. You can’t update R from within RStudio (not to my knowledge, at least): to get the new version you can go to the CRAN website and download the most recent version of R, and install it in the same way you did when you originally installed R on your computer. This used to be a slightly frustrating event, because whenever you downloaded the new version of R, you would lose all the packages that you’d downloaded and installed, and would have to repeat the process of re-installing them. This was pretty annoying, and there were some neat tricks you could use to get around this. However, newer versions of R don’t have this problem so I no longer bother explaining the workarounds for that issue.\n\n### 4.2.7 What packages does this book use?\n\nThere are several packages that I make use of in this book. The most prominent ones are:\n\n* \n `lsr` . This is the Learning Statistics with R package that accompanies this book. It doesn’t have a lot of interesting high-powered tools: it’s just a small collection of handy little things that I think can be useful to novice users. As you get more comfortable with R this package should start to feel pretty useless to you. * \n `psych` . This package, written by , includes a lot of tools that are of particular use to psychologists. In particular, there’s several functions that are particularly convenient for producing analyses or summaries that are very common in psych, but less common in other disciplines. * \n `car` . This is the Companion to Applied Regression package, which accompanies the excellent book of the same name by (Fox and Weisberg 2011). It provides a lot of very powerful tools, only some of which we’ll touch in this book. Besides these three, there are a number of packages that I use in a more limited fashion: `gplots` , `sciplot` , `foreign` , `effects` , `R.matlab` , `gdata` , `lmtest` , and probably one or two others that I’ve missed. There are also a number of packages that I refer to but don’t actually use in this book, such as `reshape` , `compute.es` , `HistData` and `multcomp` among others. Finally, there are a number of packages that provide more advanced tools that I hope to talk about in future versions of the book, such as `sem` , `ez` , `nlme` and `lme4` . In any case, whenever I’m using a function that isn’t in the core packages, I’ll make sure to note this in the text. \n\n## 4.3 Managing the workspace\n\n Let’s suppose that you’re reading through this book, and what you’re doing is sitting down with it once a week and working through a whole chapter in each sitting. Not only that, you’ve been following my advice and typing in all these commands into R. So far during this chapter, you’d have typed quite a few commands, although the only ones that actually involved creating variables were the ones you typed during Section 4.1. As a result, you currently have three variables; `seeker` , `lover` , and `keeper` . These three variables are the contents of your workspace, also referred to as the global environment. The workspace is a key concept in R, so in this section we’ll talk a lot about what it is and how to manage its contents. \n\n### 4.3.1 Listing the contents of the workspace\n\n The first thing that you need to know how to do is examine the contents of the workspace. If you’re using RStudio, you will probably find that the easiest way to do this is to use the “Environment” panel in the top right hand corner. Click on that, and you’ll see a list that looks very much like the one shown in Figures 4.5 and 4.6. If you’re using the commmand line, then the `objects()` function may come in handy: `objects()` \n\n```\n## [1] \"a\" \"addArrow\" \"addDistPlot\" \n## [4] \"afl\" \"afl.finalists\" \"afl.margins\" \n## [7] \"afl2\" \"age\" \"age.breaks\" \n## [10] \"age.group\" \"age.group2\" \"age.group3\" \n## [13] \"age.labels\" \"agpp\" \"animals\" \n## [16] \"anova.model\" \"anovaImg\" \"any.sales.this.month\"\n## [19] \"awesome\" \"b\" \"bad.coef\" \n## [22] \"balance\" \"beers\" \"berkeley\" \n## [25] \"berkeley.small\" \"binomPlot\" \"bw\" \n## [28] \"cake.1\" \"cake.2\" \"cake.df\" \n## [31] \"cake.mat1\" \"cake.mat2\" \"cakes\" \n## [34] \"cakes.flipped\" \"cardChoices\" \"cards\" \n## [37] \"chapek9\" \"chapekFrequencies\" \"chi.sq.20\" \n## [40] \"chi.sq.3\" \"chico\" \"chico2\" \n## [43] \"chico3\" \"chiSqImg\" \"choice\" \n## [46] \"choice.2\" \"clin.trial\" \"clin.trial.2\" \n## [49] \"coef\" \"coffee\" \"colour\" \n## [52] \"crit\" \"crit.hi\" \"crit.lo\" \n## [55] \"crit.val\" \"crosstab\" \"d.cor\" \n## [58] \"dan.awake\" \"data\" \"day\" \n## [61] \"days.per.month\" \"def.par\" \"describeImg\" \n## [64] \"dev.from.gp.means\" \"dev.from.grand.mean\" \"df\" \n## [67] \"doubleMax\" \"drawBasicScatterplot\" \"drug.anova\" \n## [70] \"drug.lm\" \"drug.means\" \"drug.regression\" \n## [73] \"druganxifree\" \"drugs\" \"drugs.2\" \n## [76] \"eff\" \"effort\" \"emphCol\" \n## [79] \"emphColLight\" \"emphGrey\" \"eps\" \n## [82] \"es\" \"estImg\" \"eta.squared\" \n## [85] \"eventNames\" \"expected\" \"f\" \n## [88] \"F.3.20\" \"F.stat\" \"fac\" \n## [91] \"february.sales\" \"fibonacci\" \"Fibonacci\" \n## [94] \"fileName\" \"freq\" \"full.model\" \n## [97] \"G\" \"garden\" \"gender\" \n## [100] \"generateRLineTypes\" \"generateRPointShapes\" \"good.coef\" \n## [103] \"gp.mean\" \"gp.means\" \"gp.sizes\" \n## [106] \"grades\" \"grand.mean\" \"greeting\" \n## [109] \"group\" \"h\" \"happiness\" \n## [112] \"harpo\" \"heavy.tailed.data\" \"height\" \n## [115] \"hw\" \"i\" \"interest\" \n## [118] \"IQ\" \"is.MP.speaking\" \"is.the.Party.correct\"\n## [121] \"itng\" \"itng.table\" \"keeper\" \n## [124] \"likert.centred\" \"likert.ordinal\" \"likert.raw\" \n## [127] \"lover\" \"lower.area\" \"m\" \n## [130] \"M\" \"M0\" \"M1\" \n## [133] \"makka.pakka\" \"max.val\" \"mod\" \n## [136] \"mod.1\" \"mod.2\" \"mod.3\" \n## [139] \"mod.4\" \"mod.H\" \"mod.R\" \n## [142] \"model\" \"model.1\" \"model.2\" \n## [145] \"model.3\" \"models\" \"monkey\" \n## [148] \"monkey.1\" \"month\" \"monthly.multiplier\" \n## [151] \"months\" \"ms.diff\" \"ms.res\" \n## [154] \"msg\" \"mu\" \"mu.null\" \n## [157] \"my.anova\" \"my.anova.residuals\" \"my.contrasts\" \n## [160] \"my.var\" \"n\" \"N\" \n## [163] \"ng\" \"nhstImg\" \"nodrug.regression\" \n## [166] \"normal.a\" \"normal.b\" \"normal.c\" \n## [169] \"normal.d\" \"normal.data\" \"null.model\" \n## [172] \"nullProbs\" \"numbers\" \"observed\" \n## [175] \"old\" \"old.text\" \"old_par\" \n## [178] \"oneCorPlot\" \"opinion.dir\" \"opinion.strength\" \n## [181] \"out.0\" \"out.1\" \"out.2\" \n## [184] \"outcome\" \"p.value\" \"parenthood\" \n## [187] \"payments\" \"PJ\" \"plotHist\" \n## [190] \"plotOne\" \"plotSamples\" \"plotTwo\" \n## [193] \"pow\" \"probabilities\" \"profit\" \n## [196] \"projecthome\" \"quadruple\" \"r\" \n## [199] \"R.squared\" \"random.contrasts\" \"regression.1\" \n## [202] \"regression.2\" \"regression.3\" \"regression.model\" \n## [205] \"regression.model.2\" \"regression.model.3\" \"regressionImg\" \n## [208] \"resid\" \"revenue\" \"right.table\" \n## [211] \"row.1\" \"row.2\" \"royalty\" \n## [214] \"rtfm.1\" \"rtfm.2\" \"rtfm.3\" \n## [217] \"s\" \"salem.tabs\" \"sales\" \n## [220] \"sales.by.month\" \"sample.mean\" \"scaled.chi.sq.20\" \n## [223] \"scaled.chi.sq.3\" \"score.A\" \"score.B\" \n## [226] \"sd.true\" \"sd1\" \"seeker\" \n## [229] \"sem.true\" \"setUpPlot\" \"sig\" \n## [232] \"sigEx\" \"simpson\" \"skewed.data\" \n## [235] \"some.data\" \"speaker\" \"speech.by.char\" \n## [238] \"squared.devs\" \"ss.diff\" \"SS.drug\" \n## [241] \"SS.res\" \"ss.res.full\" \"ss.res.null\" \n## [244] \"SS.resid\" \"SS.therapy\" \"ss.tot\" \n## [247] \"SS.tot\" \"SSb\" \"SStot\" \n## [250] \"SSw\" \"stock.levels\" \"suspicious.cases\" \n## [253] \"t.3\" \"teams\" \"text\" \n## [256] \"therapy.means\" \"theta\" \"today\" \n## [259] \"tombliboo\" \"total.paid\" \"trial\" \n## [262] \"ttestImg\" \"type.I.sum\" \"type.II.sum\" \n## [265] \"upper.area\" \"upsy.daisy\" \"utterance\" \n## [268] \"w\" \"W\" \"w.length\" \n## [271] \"width\" \"words\" \"wt.squared.devs\" \n## [274] \"x\" \"X\" \"x1\" \n## [277] \"X1\" \"x2\" \"X2\" \n## [280] \"x3\" \"X3\" \"X4\" \n## [283] \"xlu\" \"xtab.3d\" \"xval\" \n## [286] \"y\" \"Y\" \"Y.pred\" \n## [289] \"y1\" \"Y1\" \"y2\" \n## [292] \"Y2\" \"y3\" \"Y3\" \n## [295] \"Y4\" \"Ybar\" \"yhat.2\" \n## [298] \"yval\" \"yval.1\" \"yval.2\" \n## [301] \"z\" \"Z\" \"z.score\"\n```\n\n Of course, in the true R tradition, the `objects()` function has a lot of fancy capabilities that I’m glossing over in this example. Moreover there are also several other functions that you can use, including `ls()` which is pretty much identical to `objects()` , and `ls.str()` which you can use to get a fairly detailed description of all the variables in the workspace. In fact, the `lsr` package actually includes its own function that you can use for this purpose, called `who()` . The reason for using the `who()` function is pretty straightforward: in my everyday work I find that the output produced by the `objects()` command isn’t quite informative enough, because the only thing it prints out is the name of each variable; but the `ls.str()` function is too informative, because it prints out a lot of additional information that I really don’t like to look at. The `who()` function is a compromise between the two. First, now that we’ve got the `lsr` package installed, we need to load it: `library(lsr)` and now we can use the `who()` function: `who()` \n\n As you can see, the `who()` function lists all the variables and provides some basic information about what kind of variable each one is and how many elements it contains. Personally, I find this output much easier more useful than the very compact output of the `objects()` function, but less overwhelming than the extremely verbose `ls.str()` function. Throughout this book you’ll see me using the `who()` function a lot. You don’t have to use it yourself: in fact, I suspect you’ll find it easier to look at the RStudio environment panel. But for the purposes of writing a textbook I found it handy to have a nice text based description: otherwise there would be about another 100 or so screenshots added to the book.49 \n\n### 4.3.2 Removing variables from the workspace\n\nLooking over that list of variables, it occurs to me that I really don’t need them any more. I created them originally just to make a point, but they don’t serve any useful purpose anymore, and now I want to get rid of them. I’ll show you how to do this, but first I want to warn you – there’s no “undo” option for variable removal. Once a variable is removed, it’s gone forever unless you save it to disk. I’ll show you how to do that in Section 4.5, but quite clearly we have no need for these variables at all, so we can safely get rid of them.\n In RStudio, the easiest way to remove variables is to use the environment panel. Assuming that you’re in grid view (i.e., Figure 4.6), check the boxes next to the variables that you want to delete, then click on the “Clear” button at the top of the panel. When you do this, RStudio will show a dialog box asking you to confirm that you really do want to delete the variables. It’s always worth checking that you really do, because as RStudio is at pains to point out, you can’t undo this. Once a variable is deleted, it’s gone.50 In any case, if you click “yes”, that variable will disappear from the workspace: it will no longer appear in the environment panel, and it won’t show up when you use the `who()` command. Suppose you don’t access to RStudio, and you still want to remove variables. This is where the remove function `rm()` comes in handy. The simplest way to use `rm()` is just to type in a (comma separated) list of all the variables you want to remove. Let’s say I want to get rid of `seeker` and `lover` , but I would like to keep `keeper` . To do this, all I have to do is type: `rm( seeker, lover )` \nThere’s no visible output, but if I now inspect the workspace\n `who()` \n\n I see that there’s only the `keeper` variable left. As you can see, `rm()` can be very handy for keeping the workspace tidy. \n\n## 4.5 Loading and saving data\n\nThere are several different types of files that are likely to be relevant to us when doing data analysis. There are three in particular that are especially important from the perspective of this book:\n\n* Workspace files are those with a .Rdata file extension. This is the standard kind of file that R uses to store data and variables. They’re called “workspace files” because you can use them to save your whole workspace.\n * Comma separated value (CSV) files are those with a .csv file extension. These are just regular old text files, and they can be opened with almost any software. It’s quite typical for people to store data in CSV files, precisely because they’re so simple.\n * Script files are those with a .R file extension. These aren’t data files at all; rather, they’re used to save a collection of commands that you want R to execute later. They’re just text files, but we won’t make use of them until Chapter 8.\n\nThere are also several other types of file that R makes use of,53 but they’re not really all that central to our interests. There are also several other kinds of data file that you might want to import into R. For instance, you might want to open Microsoft Excel spreadsheets (.xlsx files), or data files that have been saved in the native file formats for other statistics software, such as SPSS, SAS, Minitab, Stata or Systat. Finally, you might have to handle databases. R tries hard to play nicely with other software, so it has tools that let you open and work with any of these and many others. I’ll discuss some of these other possibilities elsewhere in this book (Section 7.9), but for now I want to focus primarily on the two kinds of data file that you’re most likely to need: .Rdata files and .csv files. In this section I’ll talk about how to load a workspace file, how to import data from a CSV file, and how to save your workspace to a workspace file. Throughout this section I’ll first describe the (sometimes awkward) R commands that do all the work, and then I’ll show you the (much easier) way to do it using RStudio.\n\n### 4.5.1 Loading workspace files using R\n\n When I used the `list.files()` command to list the contents of the \n\ndirectory (in Section 4.4.2), the output referred to a file called booksales.Rdata. Let’s say I want to load the data from this file into my workspace. The way I do this is with the `load()` function. There are two arguments to this function, but the only one we’re interested in is \n\n* \n `file` . This should be a character string that specifies a path to the file that needs to be loaded. You can use an absolute path or a relative path to do so. \nUsing the absolute file path, the command would look like this:\n\n```\nload( file = \"/Users/dan/Rbook/data/booksales.Rdata\" )\n```\n\n but this is pretty lengthy. Given that the working directory (remember, we changed the directory at the end of Section 4.4.4) is\n\n, I could use a relative file path, like so: \n\n```\nload( file = \"../data/booksales.Rdata\" )\n```\n\nHowever, my preference is usually to change the working directory first, and then load the file. What that would look like is this:\n\n```\nsetwd( \"../data\" ) # move to the data directory\nload( \"booksales.Rdata\" ) # load the data\n```\n\n If I were then to type `who()` I’d see that there are several new variables in my workspace now. Throughout this book, whenever you see me loading a file, I will assume that the file is actually stored in the working directory, or that you’ve changed the working directory so that R is pointing at the directory that contains the file. Obviously, you don’t need type that command yourself: you can use the RStudio file panel to do the work. \n\n### 4.5.2 Loading workspace files using RStudio\n\n Okay, so how do we open an .Rdata file using the RStudio file panel? It’s terribly simple. First, use the file panel to find the folder that contains the file you want to load. If you look at Figure 4.7, you can see that there are several .Rdata files listed. Let’s say I want to load the `booksales.Rdata` file. All I have to do is click on the file name. RStudio brings up a little dialog box asking me to confirm that I do want to load this file. I click yes. The following command then turns up in the console, \n\n```\nload(\"~/Rbook/data/booksales.Rdata\")\n```\n\n and the new variables will appear in the workspace (you’ll see them in the Environment panel in RStudio, or if you type `who()` ). So easy it barely warrants having its own section. \n\n### 4.5.3 Importing data from CSV files using\n\n `loadingcsv` \nOne quite commonly used data format is the humble “comma separated value” file, also called a CSV file, and usually bearing the file extension .csv. CSV files are just plain old-fashioned text files, and what they store is basically just a table of data. This is illustrated in Figure 4.8, which shows a file called booksales.csv that I’ve created. As you can see, each row corresponds to a variable, and each row represents the book sales data for one month. The first row doesn’t contain actual data though: it has the names of the variables.\n If RStudio were not available to you, the easiest way to open this file would be to use the `read.csv()` function.54 This function is pretty flexible, and I’ll talk a lot more about it’s capabilities in Section 7.9 for more details, but for now there’s only two arguments to the function that I’ll mention: \n\n* \n `file` . This should be a character string that specifies a path to the file that needs to be loaded. You can use an absolute path or a relative path to do so. * \n `header` . This is a logical value indicating whether or not the first row of the file contains variable names. The default value is `TRUE` . \nTherefore, to import the CSV file, the command I need is:\n\n```\nbooks <- read.csv( file = \"booksales.csv\" )\n```\n\n There are two very important points to notice here. Firstly, notice that I didn’t try to use the `load()` function, because that function is only meant to be used for .Rdata files. If you try to use `load()` on other types of data, you get an error. Secondly, notice that when I imported the CSV file I assigned the result to a variable, which I imaginatively called `books` .55 file. There’s a reason for this. The idea behind an `.Rdata` file is that it stores a whole workspace. So, if you had the ability to look inside the file yourself you’d see that the data file keeps track of all the variables and their names. So when you `load()` the file, R restores all those original names. CSV files are treated differently: as far as R is concerned, the CSV only stores one variable, but that variable is big table. So when you import that table into the workspace, R expects you to give it a name.] Let’s have a look at what we’ve got: `print( books )` \n\n```\n## Month Days Sales Stock.Levels\n## 1 January 31 0 high\n## 2 February 28 100 high\n## 3 March 31 200 low\n## 4 April 30 50 out\n## 5 May 31 0 out\n## 6 June 30 0 high\n## 7 July 31 0 high\n## 8 August 31 0 high\n## 9 September 30 0 high\n## 10 October 31 0 high\n## 11 November 30 0 high\n## 12 December 31 0 high\n```\n\nClearly, it’s worked, but the format of this output is a bit unfamiliar. We haven’t seen anything like this before. What you’re looking at is a data frame, which is a very important kind of variable in R, and one I’ll discuss in Section 4.8. For now, let’s just be happy that we imported the data and that it looks about right.\n\n### 4.5.4 Importing data from CSV files using RStudio\n\nYet again, it’s easier in RStudio. In the environment panel in RStudio you should see a button called “Import Dataset”. Click on that, and it will give you a couple of options: select the “From Text File…” option, and it will open up a very familiar dialog box asking you to select a file: if you’re on a Mac, it’ll look like the usual Finder window that you use to choose a file; on Windows it looks like an Explorer window. An example of what it looks like on a Mac is shown in Figure 4.9. I’m assuming that you’re familiar with your own computer, so you should have no problem finding the CSV file that you want to import! Find the one you want, then click on the “Open” button. When you do this, you’ll see a window that looks like the one in Figure 4.10.\n\nThe import data set window is relatively straightforward to understand.\n In the top left corner, you need to type the name of the variable you R to create. By default, that will be the same as the file name: our file is called `booksales.csv` , so RStudio suggests the name `booksales` . If you’re happy with that, leave it alone. If not, type something else. Immediately below this are a few things that you can tweak to make sure that the data gets imported correctly: \n\n* Heading. Does the first row of the file contain raw data, or does it contain headings for each variable? The\n `booksales.csv` file has a header at the top, so I selected “yes”. * Separator. What character is used to separate different entries? In most CSV files this will be a comma (it is “comma separated” after all). But you can change this if your file is different.\n * Decimal. What character is used to specify the decimal point? In English speaking countries, this is almost always a period (i.e.,\n `.` ). That’s not universally true: many European countries use a comma. So you can change that if you need to. * Quote. What character is used to denote a block of text? That’s usually going to be a double quote mark. It is for the\n `booksales.csv` file, so that’s what I selected. \nThe nice thing about the RStudio window is that it shows you the raw data file at the top of the window, and it shows you a preview of the data at the bottom. If the data at the bottom doesn’t look right, try changing some of the settings on the left hand side. Once you’re happy, click “Import”. When you do, two commands appear in the R console:\n\n```\nbooksales <- read.csv(\"~/Rbook/data/booksales.csv\")\nView(booksales)\n```\n\nThe first of these commands is the one that loads the data. The second one will display a pretty table showing the data in RStudio.\n\n### 4.5.5 Saving a workspace file using\n\n `save` Not surprisingly, saving data is very similar to loading data. Although RStudio provides a simple way to save files (see below), it’s worth understanding the actual commands involved. There are two commands you can use to do this, `save()` and `save.image()` . If you’re happy to save all of the variables in your workspace into the data file, then you should use `save.image()` . And if you’re happy for R to save the file into the current working directory, all you have to do is this: \n\n```\nsave.image( file = \"myfile.Rdata\" )\n```\n\n Since `file` is the first argument, you can shorten this to \n\n```\nsave.image(\"myfile.Rdata\")\n```\n\n; and if you want to save to a different directory, then (as always) you need to be more explicit about specifying the path to the file, just as we discussed in Section 4.4. Suppose, however, I have several variables in my workspace, and I only want to save some of them. For instance, I might have this as my workspace: \n\n```\nwho()\n## -- Name -- -- Class -- -- Size --\n## data data.frame 3 x 2 \n## handy character 1 \n## junk numeric 1\n```\n\n I want to save `data` and `handy` , but not `junk` . But I don’t want to delete `junk` right now, because I want to use it for something else later on. This is where the `save()` function is useful, since it lets me indicate exactly which variables I want to save. Here is one way I can use the `save` function to solve my problem: \n\n```\nsave(data, handy, file = \"myfile.Rdata\")\n```\n\n Importantly, you must specify the name of the `file` argument. The reason is that if you don’t do so, R will think that `\"myfile.Rdata\"` is actually a variable that you want to save, and you’ll get an error message. Finally, I should mention a second way to specify which variables the `save()` function should save, which is to use the `list` argument. You do so like this: \n\n```\nsave.me <- c(\"data\", \"handy\") # the variables to be saved\nsave( file = \"booksales2.Rdata\", list = save.me ) # the command to save them\n```\n\n### 4.5.6 Saving a workspace file using RStudio\n\nRStudio allows you to save the workspace pretty easily. In the environment panel (Figures 4.5 and 4.6) you can see the “save” button. There’s no text, but it’s the same icon that gets used on every computer everywhere: it’s the one that looks like a floppy disk. You know, those things that haven’t been used in about 20 years. Alternatively, go to the “Session” menu and click on the “Save Workspace As…” option.56 This will bring up the standard “save” dialog box for your operating system (e.g., on a Mac it’ll look a little bit like the loading dialog box in Figure 4.9). Type in the name of the file that you want to save it to, and all the variables in your workspace will be saved to disk. You’ll see an R command like this one\n\n```\nsave.image(\"~/Desktop/Untitled.RData\")\n```\n\nPretty straightforward, really.\n\n### 4.5.7 Other things you might want to save\n\nUntil now, we’ve talked mostly about loading and saving data. Other things you might want to save include:\n\nThe output. Sometimes you might also want to keep a copy of all your interactions with R, including everything that you typed in and everything that R did in response. There are some functions that you can use to get R to write its output to a file rather than to print onscreen (e.g.,\n `sink()` ), but to be honest, if you do want to save the R output, the easiest thing to do is to use the mouse to select the relevant text in the R console, go to the “Edit” menu in RStudio and select “Copy”. The output has now been copied to the clipboard. Now open up your favourite text editor or word processing software, and paste it. And you’re done. However, this will only save the contents of the console, not the plots you’ve drawn (assuming you’ve drawn some). We’ll talk about saving images later on. * \n\nA script. While it is possible – and sometimes handy – to save the R output as a method for keeping a copy of your statistical analyses, another option that people use a lot (especially when you move beyond simple “toy” analyses) is to write scripts. A script is a text file in which you write out all the commands that you want R to run. You can write your script using whatever software you like. In real world data analysis writing scripts is a key skill – and as you become familiar with R you’ll probably find that most of what you do involves scripting rather than typing commands at the R prompt. However, you won’t need to do much scripting initially, so we’ll leave that until Chapter 8.\n\n## 4.6 Useful things to know about variables\n\nIn Chapter 3 I talked a lot about variables, how they’re assigned and some of the things you can do with them, but there’s a lot of additional complexities. That’s not a surprise of course. However, some of those issues are worth drawing your attention to now. So that’s the goal of this section; to cover a few extra topics. As a consequence, this section is basically a bunch of things that I want to briefly mention, but don’t really fit in anywhere else. In short, I’ll talk about several different issues in this section, which are only loosely connected to one another.\n\n### 4.6.1 Special values\n\n The first thing I want to mention are some of the “special” values that you might see R produce. Most likely you’ll see them in situations where you were expecting a number, but there are quite a few other ways you can encounter them. These values are `Inf` , `NaN` , `NA` and `NULL` . These values can crop up in various different places, and so it’s important to understand what they mean. \n\n* Infinity (\n `Inf` ). The easiest of the special values to explain is `Inf` , since it corresponds to a value that is infinitely large. You can also have `-Inf` . The easiest way to get `Inf` is to divide a positive number by 0: `1 / 0` `## [1] Inf` \nIn most real world data analysis situations, if you’re ending up with infinite numbers in your data, then something has gone awry. Hopefully you’ll never have to see them.\n\n* Not a Number (\n `NaN` ). The special value of `NaN` is short for “not a number”, and it’s basically a reserved keyword that means “there isn’t a mathematically defined number for this”. If you can remember your high school maths, remember that it is conventional to say that \\(0/0\\) doesn’t have a proper answer: mathematicians would say that \\(0/0\\) is undefined. R says that it’s not a number: `0 / 0` `## [1] NaN` Nevertheless, it’s still treated as a “numeric” value. To oversimplify, `NaN` corresponds to cases where you asked a proper numerical question that genuinely has no meaningful answer. \n\nNot available (\n `NA` ). `NA` indicates that the value that is “supposed” to be stored here is missing. To understand what this means, it helps to recognise that the `NA` value is something that you’re most likely to see when analysing data from real world experiments. Sometimes you get equipment failures, or you lose some of the data, or whatever. The point is that some of the information that you were “expecting” to get from your study is just plain missing. Note the difference between `NA` and `NaN` . For `NaN` , we really do know what’s supposed to be stored; it’s just that it happens to correspond to something like \\(0/0\\) that doesn’t make any sense at all. In contrast, `NA` indicates that we actually don’t know what was supposed to be there. The information is missing. * \n\nNo value (\n `NULL` ). The `NULL` value takes this “absence” concept even further. It basically asserts that the variable genuinely has no value whatsoever. This is quite different to both `NaN` and `NA` . For `NaN` we actually know what the value is, because it’s something insane like \\(0/0\\). For `NA` , we believe that there is supposed to be a value “out there”, but a dog ate our homework and so we don’t quite know what it is. But for `NULL` we strongly believe that there is no value at all. \n\n### 4.6.2 Assigning names to vector elements\n\nOne thing that is sometimes a little unsatisfying about the way that R prints out a vector is that the elements come out unlabelled. Here’s what I mean. Suppose I’ve got data reporting the quarterly profits for some company. If I just create a no-frills vector, I have to rely on memory to know which element corresponds to which event. That is:\n\n```\nprofit <- c( 3.1, 0.1, -1.4, 1.1 )\nprofit\n```\n\n```\n## [1] 3.1 0.1 -1.4 1.1\n```\n\n You can probably guess that the first element corresponds to the first quarter, the second element to the second quarter, and so on, but that’s only because I’ve told you the back story and because this happens to be a very simple example. In general, it can be quite difficult. This is where it can be helpful to assign `names` to each of the elements. Here’s how you do it: \n\n```\nnames(profit) <- c(\"Q1\",\"Q2\",\"Q3\",\"Q4\")\nprofit\n```\n\n This is a slightly odd looking command, admittedly, but it’s not too difficult to follow. All we’re doing is assigning a vector of labels (character strings) to `names(profit)` . You can always delete the names again by using the command \n\n```\nnames(profit) <- NULL\n```\n\n. It’s also worth noting that you don’t have to do this as a two stage process. You can get the same result with this command: \n\n```\nprofit <- c( \"Q1\" = 3.1, \"Q2\" = 0.1, \"Q3\" = -1.4, \"Q4\" = 1.1 )\nprofit\n```\n\n The important things to notice are that (a) this does make things much easier to read, but (b) the names at the top aren’t the “real” data. The value of `profit[1]` is still `3.1` ; all I’ve done is added a name to `profit[1]` as well. Nevertheless, names aren’t purely cosmetic, since R allows you to pull out particular elements of the vector by referring to their names: `profit[\"Q1\"]` \n\n```\n## Q1 \n## 3.1\n```\n\n And if I ever need to pull out the names themselves, then I just type `names(profit)` . \n\n### 4.6.3 Variable classes\n\nAs we’ve seen, R allows you to store different kinds of data. In particular, the variables we’ve defined so far have either been character data (text), numeric data, or logical data.57 It’s important that we remember what kind of information each variable stores (and even more important that R remembers) since different kinds of variables allow you to do different things to them. For instance, if your variables have numerical information in them, then it’s okay to multiply them together:\n\n```\nx <- 5 # x is numeric\ny <- 4 # y is numeric\nx * y\n```\n\n `## [1] 20` \nBut if they contain character data, multiplication makes no sense whatsoever, and R will complain if you try to do it:\n\n```\nx <- \"apples\" # x is character\ny <- \"oranges\" # y is character\nx * y\n```\n\n Even R is smart enough to know you can’t multiply `\"apples\"` by `\"oranges\"` . It knows this because the quote marks are indicators that the variable is supposed to be treated as text, not as a number. This is quite useful, but notice that it means that R makes a big distinction between `5` and `\"5\"` . Without quote marks, R treats `5` as the number five, and will allow you to do calculations with it. With the quote marks, R treats `\"5\"` as the textual character five, and doesn’t recognise it as a number any more than it recognises `\"p\"` or `\"five\"` as numbers. As a consequence, there’s a big difference between typing `x <- 5` and typing `x <- \"5\"` . In the former, we’re storing the number `5` ; in the latter, we’re storing the character `\"5\"` . Thus, if we try to do multiplication with the character versions, R gets stroppy: \n\n```\nx <- \"5\" # x is character\ny <- \"4\" # y is character\nx * y\n```\n\n Okay, let’s suppose that I’ve forgotten what kind of data I stored in the variable `x` (which happens depressingly often). R provides a function that will let us find out. Or, more precisely, it provides three functions: `class()` , `mode()` and `typeof()` . Why the heck does it provide three functions, you might be wondering? Basically, because R actually keeps track of three different kinds of information about a variable: \n\n* The class of a variable is a “high level” classification, and it captures psychologically (or statistically) meaningful distinctions. For instance\n `\"2011-09-12\"` and `\"my birthday\"` are both text strings, but there’s an important difference between the two: one of them is a date. So it would be nice if we could get R to recognise that `\"2011-09-12\"` is a date, and allow us to do things like add or subtract from it. The class of a variable is what R uses to keep track of things like that. Because the class of a variable is critical for determining what R can or can’t do with it, the `class()` function is very handy. * The mode of a variable refers to the format of the information that the variable stores. It tells you whether R has stored text data or numeric data, for instance, which is kind of useful, but it only makes these “simple” distinctions. It can be useful to know about, but it’s not the main thing we care about. So I’m not going to use the\n `mode()` function very much.58 * The type of a variable is a very low level classification. We won’t use it in this book, but (for those of you that care about these details) this is where you can see the distinction between integer data, double precision numeric, etc. Almost none of you actually will care about this, so I’m not even going to bother demonstrating the\n `typeof()` function. For purposes, it’s the `class()` of the variable that we care most about. Later on, I’ll talk a bit about how you can convince R to “coerce” a variable to change from one class to another (Section 7.10). That’s a useful skill for real world data analysis, but it’s not something that we need right now. In the meantime, the following examples illustrate the use of the `class()` function: \n\n```\nx <- \"hello world\" # x is text\nclass(x)\n```\n\n `## [1] \"character\"` \n\n```\nx <- TRUE # x is logical \nclass(x)\n```\n\n `## [1] \"logical\"` \n\n```\nx <- 100 # x is a number\nclass(x)\n```\n\n `## [1] \"numeric\"` \nExciting, no?\n\n## 4.7 Factors\n\nOkay, it’s time to start introducing some of the data types that are somewhat more specific to statistics. If you remember back to Chapter 2, when we assign numbers to possible outcomes, these numbers can mean quite different things depending on what kind of variable we are attempting to measure. In particular, we commonly make the distinction between nominal, ordinal, interval and ratio scale data. How do we capture this distinction in R? Currently, we only seem to have a single numeric data type. That’s probably not going to be enough, is it?\n\nA little thought suggests that the numeric variable class in R is perfectly suited for capturing ratio scale data. For instance, if I were to measure response time (RT) for five different events, I could store the data in R like this:\n\n```\nRT <- c(342, 401, 590, 391, 554)\n```\n\nwhere the data here are measured in milliseconds, as is conventional in the psychological literature. It’s perfectly sensible to talk about “twice the response time”, \\(2 \\times \\mbox{RT}\\), or the “response time plus 1 second”, \\(\\mbox{RT} + 1000\\), and so both of the following are perfectly reasonable things for R to do:\n `2 * RT` \n\n```\n## [1] 684 802 1180 782 1108\n```\n\n `RT + 1000` \n\n```\n## [1] 1342 1401 1590 1391 1554\n```\n\nAnd to a lesser extent, the “numeric” class is okay for interval scale data, as long as we remember that multiplication and division aren’t terribly interesting for these sorts of variables. That is, if my IQ score is 110 and yours is 120, it’s perfectly okay to say that you’re 10 IQ points smarter than me59, but it’s not okay to say that I’m only 92% as smart as you are, because intelligence doesn’t have a natural zero.60 We might even be willing to tolerate the use of numeric variables to represent ordinal scale variables, such as those that you typically get when you ask people to rank order items (e.g., like we do in Australian elections), though as we will see R actually has a built in tool for representing ordinal data (see Section 7.11.2) However, when it comes to nominal scale data, it becomes completely unacceptable, because almost all of the “usual” rules for what you’re allowed to do with numbers don’t apply to nominal scale data. It is for this reason that R has factors.\n\n### 4.7.1 Introducing factors\n\nSuppose, I was doing a study in which people could belong to one of three different treatment conditions. Each group of people were asked to complete the same task, but each group received different instructions. Not surprisingly, I might want to have a variable that keeps track of what group people were in. So I could type in something like this\n\n```\ngroup <- c(1,1,1,2,2,2,3,3,3)\n```\n\n so that `group[i]` contains the group membership of the `i` -th person in my study. Clearly, this is numeric data, but equally obviously this is a nominal scale variable. There’s no sense in which “group 1” plus “group 2” equals “group 3”, but nevertheless if I try to do that, R won’t stop me because it doesn’t know any better: `group + 2` \n\n```\n## [1] 3 3 3 4 4 4 5 5 5\n```\n\n Apparently R seems to think that it’s allowed to invent “group 4” and “group 5”, even though they didn’t actually exist. Unfortunately, R is too stupid to know any better: it thinks that `3` is an ordinary number in this context, so it sees no problem in calculating `3 + 2` . But since we’re not that stupid, we’d like to stop R from doing this. We can do so by instructing R to treat `group` as a factor. This is easy to do using the `as.factor()` function.61 \n\n```\ngroup <- as.factor(group)\ngroup\n```\n\n It looks more or less the same as before (though it’s not immediately obvious what all that `Levels` rubbish is about), but if we ask R to tell us what the class of the `group` variable is now, it’s clear that it has done what we asked: `class(group)` `## [1] \"factor\"` Neat. Better yet, now that I’ve converted `group` to a factor, look what happens when I try to add 2 to it: `group + 2` \n\n```\n## Warning in Ops.factor(group, 2): '+' not meaningful for factors\n```\n\n```\n## [1] NA NA NA NA NA NA NA NA NA\n```\n\n This time even R is smart enough to know that I’m being an idiot, so it tells me off and then produces a vector of missing values. (i.e., `NA` : see Section 4.6.1). \n\n### 4.7.2 Labelling the factor levels\n\n I have a confession to make. My memory is not infinite in capacity; and it seems to be getting worse as I get older. So it kind of annoys me when I get data sets where there’s a nominal scale variable called `gender` , with two levels corresponding to males and females. But when I go to print out the variable I get something like this: `gender` \n\n```\n## [1] 1 1 1 1 1 2 2 2 2\n## Levels: 1 2\n```\n\nOkaaaay. That’s not helpful at all, and it makes me very sad. Which number corresponds to the males and which one corresponds to the females? Wouldn’t it be nice if R could actually keep track of this? It’s way too hard to remember which number corresponds to which gender. To fix this problem what we need to do is assign meaningful labels to the different levels of each factor. We can do that like this:\n\n```\nlevels(group) <- c(\"group 1\", \"group 2\", \"group 3\")\nprint(group)\n```\n\n```\nlevels(gender) <- c(\"male\", \"female\")\nprint(gender)\n```\n\n```\n## [1] male male male male male female female female female\n## Levels: male female\n```\n\nThat’s much easier on the eye.\n\n### 4.7.3 Moving on…\n\nFactors are very useful things, and we’ll use them a lot in this book: they’re the main way to represent a nominal scale variable. And there are lots of nominal scale variables out there. I’ll talk more about factors in Section 7.11.2, but for now you know enough to be able to get started.\n\n## 4.8 Data frames\n\nIt’s now time to go back and deal with the somewhat confusing thing that happened in Section ?? when we tried to open up a CSV file. Apparently we succeeded in loading the data, but it came to us in a very odd looking format. At the time, I told you that this was a data frame. Now I’d better explain what that means.\n\n### 4.8.1 Introducing data frames\n\n In order to understand why R has created this funny thing called a data frame, it helps to try to see what problem it solves. So let’s go back to the little scenario that I used when introducing factors in Section 4.7. In that section I recorded the `group` and `gender` for all 9 participants in my study. Let’s also suppose I recorded their ages and their `score` on “Dan’s Terribly Exciting Psychological Test”: \n\n```\nage <- c(17, 19, 21, 37, 18, 19, 47, 18, 19)\nscore <- c(12, 10, 11, 15, 16, 14, 25, 21, 29)\n```\n\n Assuming no other variables are in the workspace, if I type `who()` I get this: `who()` \n\n So there are four variables in the workspace, `age` , `gender` , `group` and `score` . And it just so happens that all four of them are the same size (i.e., they’re all vectors with 9 elements). Aaaand it just so happens that `age[1]` corresponds to the age of the first person, and `gender[1]` is the gender of that very same person, etc. In other words, you and I both know that all four of these variables correspond to the same data set, and all four of them are organised in exactly the same way. However, R doesn’t know this! As far as it’s concerned, there’s no reason why the `age` variable has to be the same length as the `gender` variable; and there’s no particular reason to think that `age[1]` has any special relationship to `gender[1]` any more than it has a special relationship to `gender[4]` . In other words, when we store everything in separate variables like this, R doesn’t know anything about the relationships between things. It doesn’t even really know that these variables actually refer to a proper data set. The data frame fixes this: if we store our variables inside a data frame, we’re telling R to treat these variables as a single, fairly coherent data set. To see how they do this, let’s create one. So how do we create a data frame? One way we’ve already seen: if we import our data from a CSV file, R will store it as a data frame. A second way is to create it directly from some existing variables using the `data.frame()` function. All you have to do is type a list of variables that you want to include in the data frame. The output of a `data.frame()` command is, well, a data frame. So, if I want to store all four variables from my experiment in a data frame called `expt` I can do so like this: \n\n```\nexpt <- data.frame ( age, gender, group, score ) \nexpt\n```\n\n```\n## age gender group score\n## 1 17 male group 1 12\n## 2 19 male group 1 10\n## 3 21 male group 1 11\n## 4 37 male group 2 15\n## 5 18 male group 2 16\n## 6 19 female group 2 14\n## 7 47 female group 3 25\n## 8 18 female group 3 21\n## 9 19 female group 3 29\n```\n\n Note that `expt` is a completely self-contained variable. Once you’ve created it, it no longer depends on the original variables from which it was constructed. That is, if we make changes to the original `age` variable, it will not lead to any changes to the age data stored in `expt` . \n\n### 4.8.2 Pulling out the contents of the data frame using\n\n `$` At this point, our workspace contains only the one variable, a data frame called `expt` . But as we can see when we told R to print the variable out, this data frame contains 4 variables, each of which has 9 observations. So how do we get this information out again? After all, there’s no point in storing information if you don’t use it, and there’s no way to use information if you can’t access it. So let’s talk a bit about how to pull information out of a data frame. The first thing we might want to do is pull out one of our stored variables, let’s say `score` . One thing you might try to do is ignore the fact that `score` is locked up inside the `expt` data frame. For instance, you might try to print it out like this: `score` \n\n```\n## Error in eval(expr, envir, enclos): object 'score' not found\n```\n\n This doesn’t work, because R doesn’t go “peeking” inside the data frame unless you explicitly tell it to do so. There’s actually a very good reason for this, which I’ll explain in a moment, but for now let’s just assume R knows what it’s doing. How do we tell R to look inside the data frame? As is always the case with R there are several ways. The simplest way is to use the `$` operator to extract the variable you’re interested in, like this: `expt$score` \n\n```\n## [1] 12 10 11 15 16 14 25 21 29\n```\n\n### 4.8.3 Getting information about a data frame\n\n One problem that sometimes comes up in practice is that you forget what you called all your variables. Normally you might try to type `objects()` or `who()` , but neither of those commands will tell you what the names are for those variables inside a data frame! One way is to ask R to tell you what the names of all the variables stored in the data frame are, which you can do using the `names()` function: `names(expt)` \n\n```\n## [1] \"age\" \"gender\" \"group\" \"score\"\n```\n\n An alternative method is to use the `who()` function, as long as you tell it to look at the variables inside data frames. If you set `expand = TRUE` then it will not only list the variables in the workspace, but it will “expand” any data frames that you’ve got in the workspace, so that you can see what they look like. That is: `who(expand = TRUE)` \n\n```\n## -- Name -- -- Class -- -- Size --\n## a numeric 1 \n## addArrow function \n## addDistPlot function \n## afl data.frame 4296 x 12 \n## $home.team factor 4296 \n## $away.team factor 4296 \n## $home.score numeric 4296 \n## $away.score numeric 4296 \n## $year numeric 4296 \n## $round numeric 4296 \n## $weekday factor 4296 \n## $day numeric 4296 \n## $month numeric 4296 \n## $is.final logical 4296 \n## $venue factor 4296 \n## $attendance numeric 4296 \n## afl.finalists factor 400 \n## afl.margins numeric 176 \n## afl2 data.frame 4296 x 2 \n## $margin numeric 4296 \n## $year numeric 4296 \n## age.breaks numeric 4 \n## age.group factor 11 \n## age.group2 factor 11 \n## age.group3 factor 11 \n## age.labels character 3 \n## agpp data.frame 100 x 3 \n## $id factor 100 \n## $response_before factor 100 \n## $response_after factor 100 \n## animals character 4 \n## anova.model aov 13 \n## anovaImg list 0 \n## any.sales.this.month logical 12 \n## awesome data.frame 10 x 2 \n## $scores numeric 10 \n## $group factor 10 \n## b numeric 1 \n## bad.coef numeric 2 \n## balance numeric 1 \n## beers character 3 \n## berkeley data.frame 39 x 3 \n## $women.apply numeric 39 \n## $total.admit numeric 39 \n## $number.apply numeric 39 \n## berkeley.small data.frame 46 x 2 \n## $women.apply numeric 46 \n## $total.admit numeric 46 \n## binomPlot function \n## bw numeric 1 \n## cake.1 numeric 5 \n## cake.2 numeric 5 \n## cake.df data.frame 5 x 2 \n## $cake.1 numeric 5 \n## $cake.2 numeric 5 \n## cake.mat1 matrix 5 x 2 \n## cake.mat2 matrix 2 x 5 \n## cakes matrix 4 x 5 \n## cakes.flipped matrix 5 x 4 \n## cardChoices xtabs 4 x 4 \n## cards data.frame 200 x 3 \n## $id factor 200 \n## $choice_1 factor 200 \n## $choice_2 factor 200 \n## chapek9 data.frame 180 x 2 \n## $species factor 180 \n## $choice factor 180 \n## chapekFrequencies xtabs 3 x 2 \n## chi.sq.20 numeric 1000 \n## chi.sq.3 numeric 1000 \n## chico data.frame 20 x 3 \n## $id factor 20 \n## $grade_test1 numeric 20 \n## $grade_test2 numeric 20 \n## chico2 data.frame 40 x 4 \n## $id factor 40 \n## $improvement numeric 40 \n## $time factor 40 \n## $grade numeric 40 \n## chico3 data.frame 20 x 4 \n## $id factor 20 \n## $grade_test1 numeric 20 \n## $grade_test2 numeric 20 \n## $improvement numeric 20 \n## chiSqImg list 0 \n## choice data.frame 4 x 10 \n## $id integer 4 \n## $gender factor 4 \n## $MRT/block1/day1 numeric 4 \n## $MRT/block1/day2 numeric 4 \n## $MRT/block2/day1 numeric 4 \n## $MRT/block2/day2 numeric 4 \n## $PC/block1/day1 numeric 4 \n## $PC/block1/day2 numeric 4 \n## $PC/block2/day1 numeric 4 \n## $PC/block2/day2 numeric 4 \n## choice.2 data.frame 16 x 6 \n## $id integer 16 \n## $gender factor 16 \n## $MRT numeric 16 \n## $PC numeric 16 \n## $block factor 16 \n## $day factor 16 \n## clin.trial data.frame 18 x 3 \n## $drug factor 18 \n## $therapy factor 18 \n## $mood.gain numeric 18 \n## clin.trial.2 data.frame 18 x 5 \n## $druganxifree numeric 18 \n## $drugjoyzepam numeric 18 \n## $therapyCBT numeric 18 \n## $mood.gain numeric 18 \n## $druganxifree numeric 18 \n## coef numeric 2 \n## coffee data.frame 18 x 3 \n## $milk factor 18 \n## $sugar factor 18 \n## $babble numeric 18 \n## colour logical 1 \n## crit numeric 1 \n## crit.hi numeric 1 \n## crit.lo numeric 1 \n## crit.val numeric 1 \n## crosstab xtabs 2 x 3 \n## d.cor numeric 1 \n## dan.awake logical 10 \n## data data.frame 12 x 4 \n## $V1 factor 12 \n## $V2 integer 12 \n## $V3 integer 12 \n## $V4 factor 12 \n## day character 1 \n## days.per.month numeric 12 \n## def.par list 66 \n## describeImg list 0 \n## dev.from.gp.means array 18 \n## dev.from.grand.mean array 3 \n## df numeric 1 \n## doubleMax function \n## drawBasicScatterplot function \n## drug.anova aov 13 \n## drug.lm lm 13 \n## drug.means numeric 3 \n## drug.regression lm 12 \n## druganxifree numeric 18 \n## drugs data.frame 10 x 8 \n## $id factor 10 \n## $gender factor 10 \n## $WMC_alcohol numeric 10 \n## $WMC_caffeine numeric 10 \n## $WMC_no.drug numeric 10 \n## $RT_alcohol numeric 10 \n## $RT_caffeine numeric 10 \n## $RT_no.drug numeric 10 \n## drugs.2 data.frame 30 x 5 \n## $id factor 30 \n## $gender factor 30 \n## $drug factor 30 \n## $WMC numeric 30 \n## $RT numeric 30 \n## eff eff 22 \n## effort data.frame 10 x 2 \n## $hours numeric 10 \n## $grade numeric 10 \n## emphCol character 1 \n## emphColLight character 1 \n## emphGrey character 1 \n## eps logical 1 \n## es matrix 4 x 7 \n## estImg list 0 \n## eta.squared numeric 1 \n## eventNames character 5 \n## expected numeric 4 \n## expt data.frame 9 x 4 \n## $age numeric 9 \n## $gender factor 9 \n## $group factor 9 \n## $score numeric 9 \n## f table 14 \n## F.3.20 numeric 1000 \n## F.stat numeric 1 \n## fac factor 3 \n## february.sales numeric 1 \n## fibonacci numeric 6 \n## Fibonacci numeric 7 \n## fileName character 1 \n## freq integer 17 \n## full.model lm 12 \n## G factor 18 \n## garden data.frame 5 x 3 \n## $speaker factor 5 \n## $utterance factor 5 \n## $line numeric 5 \n## generateRLineTypes function \n## generateRPointShapes function \n## good.coef numeric 2 \n## gp.mean array 3 \n## gp.means array 3 \n## gp.sizes array 3 \n## grades numeric 20 \n## grand.mean numeric 1 \n## greeting character 1 \n## h numeric 1 \n## happiness data.frame 10 x 3 \n## $before numeric 10 \n## $after numeric 10 \n## $change numeric 10 \n## harpo data.frame 33 x 2 \n## $grade numeric 33 \n## $tutor factor 33 \n## heavy.tailed.data numeric 100 \n## height numeric 1 \n## hw character 2 \n## i integer 1 \n## interest numeric 1 \n## IQ numeric 10000 \n## is.MP.speaking logical 5 \n## is.the.Party.correct table 14 \n## itng data.frame 10 x 2 \n## $speaker factor 10 \n## $utterance factor 10 \n## itng.table table 3 x 4 \n## likert.centred numeric 10 \n## likert.ordinal ordered 10 \n## likert.raw numeric 10 \n## lower.area numeric 1 \n## m numeric 1 \n## M matrix 2 x 3 \n## M0 lm 12 \n## M1 lm 12 \n## makka.pakka character 4 \n## max.val numeric 1 \n## mod lm 13 \n## mod.1 lm 11 \n## mod.2 lm 13 \n## mod.3 lm 13 \n## mod.4 lm 13 \n## mod.H lm 13 \n## mod.R lm 13 \n## model aov 13 \n## model.1 aov 13 \n## model.2 aov 13 \n## model.3 aov 13 \n## models BFBayesFactor \n## monkey character 1 \n## monkey.1 list 1 \n## month numeric 1 \n## monthly.multiplier numeric 1 \n## months character 12 \n## ms.diff numeric 1 \n## ms.res numeric 1 \n## msg character 1 \n## mu numeric 3 \n## mu.null numeric 1 \n## my.anova aov 13 \n## my.anova.residuals numeric 18 \n## my.contrasts list 2 \n## my.var numeric 1 \n## n numeric 1 \n## N integer 1 \n## ng character 2 \n## nhstImg list 0 \n## nodrug.regression lm 12 \n## normal.a numeric 1000 \n## normal.b numeric 1000 \n## normal.c numeric 1000 \n## normal.d numeric 1000 \n## normal.data numeric 100 \n## null.model lm 11 \n## nullProbs numeric 4 \n## numbers numeric 3 \n## observed table 4 \n## old list 66 \n## old.text character 1 \n## old_par list 72 \n## oneCorPlot function \n## opinion.dir numeric 10 \n## opinion.strength numeric 10 \n## out.0 data.frame 100 x 2 \n## $V1 numeric 100 \n## $V2 numeric 100 \n## out.1 data.frame 100 x 2 \n## $V1 numeric 100 \n## $V2 numeric 100 \n## out.2 data.frame 100 x 2 \n## $V1 numeric 100 \n## $V2 numeric 100 \n## outcome numeric 18 \n## p.value numeric 1 \n## parenthood data.frame 100 x 4 \n## $dan.sleep numeric 100 \n## $baby.sleep numeric 100 \n## $dan.grump numeric 100 \n## $day integer 100 \n## payments numeric 1 \n## PJ character 1 \n## plotHist function \n## plotOne function \n## plotSamples function \n## plotTwo function \n## pow numeric 100 \n## probabilities numeric 5 \n## projecthome character 1 \n## quadruple function \n## r numeric 1 \n## R.squared numeric 1 \n## random.contrasts matrix 3 x 2 \n## regression.1 lm 12 \n## regression.2 lm 12 \n## regression.3 lm 12 \n## regression.model lm 12 \n## regression.model.2 lm 13 \n## regression.model.3 lm 13 \n## regressionImg list 0 \n## resid numeric 18 \n## revenue numeric 1 \n## right.table xtabs 2 x 2 \n## row.1 numeric 3 \n## row.2 numeric 3 \n## royalty numeric 1 \n## rtfm.1 data.frame 8 x 3 \n## $grade numeric 8 \n## $attend numeric 8 \n## $reading numeric 8 \n## rtfm.2 data.frame 8 x 3 \n## $grade numeric 8 \n## $attend factor 8 \n## $reading factor 8 \n## rtfm.3 data.frame 8 x 3 \n## $grade numeric 8 \n## $attend factor 8 \n## $reading factor 8 \n## s numeric 1 \n## [ reached getOption(\"max.print\") -- omitted 85 rows ]\n```\n\n or, since `expand` is the first argument in the `who()` function you can just type `who(TRUE)` . I’ll do that a lot in this book. \n\n### 4.8.4 Looking for more on data frames?\n\nThere’s a lot more that can be said about data frames: they’re fairly complicated beasts, and the longer you use R the more important it is to make sure you really understand them. We’ll talk a lot more about them in Chapter 7.\n\n## 4.9 Lists\n\n The next kind of data I want to mention are lists. Lists are an extremely fundamental data structure in R, and as you start making the transition from a novice to a savvy R user you will use lists all the time. I don’t use lists very often in this book – not directly – but most of the advanced data structures in R are built from lists (e.g., data frames are actually a specific type of list). Because lists are so important to how R stores things, it’s useful to have a basic understanding of them. Okay, so what is a list, exactly? Like data frames, lists are just “collections of variables.” However, unlike data frames – which are basically supposed to look like a nice “rectangular” table of data – there are no constraints on what kinds of variables we include, and no requirement that the variables have any particular relationship to one another. In order to understand what this actually means, the best thing to do is create a list, which we can do using the `list()` function. If I type this as my command: \n\n```\nDan <- list( age = 34,\n nerd = TRUE,\n parents = c(\"Joe\",\"Liz\") \n)\n```\n\n R creates a new list variable called `Dan` , which is a bundle of three different variables: `age` , `nerd` and `parents` . Notice, that the `parents` variable is longer than the others. This is perfectly acceptable for a list, but it wouldn’t be for a data frame. If we now print out the variable, you can see the way that R stores the list: `print( Dan )` \n\n```\n## $age\n## [1] 34\n## \n## $nerd\n## [1] TRUE\n## \n## $parents\n## [1] \"Joe\" \"Liz\"\n```\n\n As you might have guessed from those `$` symbols everywhere, the variables are stored in exactly the same way that they are for a data frame (again, this is not surprising: data frames are a type of list). So you will (I hope) be entirely unsurprised and probably quite bored when I tell you that you can extract the variables from the list using the `$` operator, like so: `Dan$nerd` `## [1] TRUE` If you need to add new entries to the list, the easiest way to do so is to again use `$` , as the following example illustrates. If I type a command like this \n\n```\nDan$children <- \"Alex\"\n```\n\n then R creates a new entry to the end of the list called `children` , and assigns it a value of `\"Alex\"` . If I were now to `print()` this list out, you’d see a new entry at the bottom of the printout. Finally, it’s actually possible for lists to contain other lists, so it’s quite possible that I would end up using a command like `Dan$children$age` to find out how old my son is. Or I could try to remember it myself I suppose. \n\n## 4.10 Formulas\n\nThe last kind of variable that I want to introduce before finally being able to start talking about statistics is the formula. Formulas were originally introduced into R as a convenient way to specify a particular type of statistical model (see Chapter 15) but they’re such handy things that they’ve spread. Formulas are now used in a lot of different contexts, so it makes sense to introduce them early.\n Stated simply, a formula object is a variable, but it’s a special type of variable that specifies a relationship between other variables. A formula is specified using the “tilde operator” `~` . A very simple example of a formula is shown below:62 \n\n```\nformula1 <- out ~ pred\nformula1\n```\n\n `## out ~ pred` The precise meaning of this formula depends on exactly what you want to do with it, but in broad terms it means “the `out` (outcome) variable, analysed in terms of the `pred` (predictor) variable”. That said, although the simplest and most common form of a formula uses the “one variable on the left, one variable on the right” format, there are others. For instance, the following examples are all reasonably common \n\n```\nformula2 <- out ~ pred1 + pred2 # more than one variable on the right\nformula3 <- out ~ pred1 * pred2 # different relationship between predictors \nformula4 <- ~ var1 + var2 # a 'one-sided' formula\n```\n\nand there are many more variants besides. Formulas are pretty flexible things, and so different functions will make use of different formats, depending on what the function is intended to do.\n\n## 4.11 Generic functions\n\n There’s one really important thing that I omitted when I discussed functions earlier on in Section 3.5, and that’s the concept of a generic function. The two most notable examples that you’ll see in the next few chapters are `summary()` and `plot()` , although you’ve already seen an example of one working behind the scenes, and that’s the `print()` function. The thing that makes generics different from the other functions is that their behaviour changes, often quite dramatically, depending on the `class()` of the input you give it. The easiest way to explain the concept is with an example. With that in mind, lets take a closer look at what the `print()` function actually does. I’ll do this by creating a formula, and printing it out in a few different ways. First, let’s stick with what we know: \n\n```\nmy.formula <- blah ~ blah.blah # create a variable of class \"formula\"\nprint( my.formula ) # print it out using the generic print() function\n```\n\n `## blah ~ blah.blah` So far, there’s nothing very surprising here. But there’s actually a lot going on behind the scenes here. When I type `print( my.formula )` , what actually happens is the `print()` function checks the class of the `my.formula` variable. When the function discovers that the variable it’s been given is a formula, it goes looking for a function called `print.formula()` , and then delegates the whole business of printing out the variable to the `print.formula()` function.63 For what it’s worth, the name for a “dedicated” function like `print.formula()` that exists only to be a special case of a generic function like `print()` is a method, and the name for the process in which the generic function passes off all the hard work onto a method is called method dispatch. You won’t need to understand the details at all for this book, but you do need to know the gist of it; if only because a lot of the functions we’ll use are actually generics. Anyway, to help expose a little more of the workings to you, let’s bypass the `print()` function entirely and call the formula method directly: \n\n```\nprint.formula( my.formula ) # print it out using the print.formula() method\n\n## Appears to be deprecated\n```\n\n There’s no difference in the output at all. But this shouldn’t surprise you because it was actually the `print.formula()` method that was doing all the hard work in the first place. The `print()` function itself is a lazy bastard that doesn’t do anything other than select which of the methods is going to do the actual printing. Okay, fair enough, but you might be wondering what would have happened if `print.formula()` didn’t exist? That is, what happens if there isn’t a specific method defined for the class of variable that you’re using? In that case, the generic function passes off the hard work to a “default” method, whose name in this case would be `print.default()` . Let’s see what happens if we bypass the `print()` formula, and try to print out `my.formula` using the `print.default()` function: \n\n```\nprint.default( my.formula ) # print it out using the print.default() method\n```\n\n```\n## blah ~ blah.blah\n## attr(,\"class\")\n## [1] \"formula\"\n## attr(,\".Environment\")\n## \n```\n\n Hm. You can kind of see that it is trying to print out the same formula, but there’s a bunch of ugly low-level details that have also turned up on screen. This is because the `print.default()` method doesn’t know anything about formulas, and doesn’t know that it’s supposed to be hiding the obnoxious internal gibberish that R produces sometimes. \nAt this stage, this is about as much as we need to know about generic functions and their methods. In fact, you can get through the entire book without learning any more about them than this, so it’s probably a good idea to end this discussion here.\n\n## 4.12 Getting help\n\nThe very last topic I want to mention in this chapter is where to go to find help. Obviously, I’ve tried to make this book as helpful as possible, but it’s not even close to being a comprehensive guide, and there’s thousands of things it doesn’t cover. So where should you go for help?\n\n### 4.12.1 How to read the help documentation\n\nI have somewhat mixed feelings about the help documentation in R. On the plus side, there’s a lot of it, and it’s very thorough. On the minus side, there’s a lot of it, and it’s very thorough. There’s so much help documentation that it sometimes doesn’t help, and most of it is written with an advanced user in mind. Often it feels like most of the help files work on the assumption that the reader already understands everything about R except for the specific topic that it’s providing help for. What that means is that, once you’ve been using R for a long time and are beginning to get a feel for how to use it, the help documentation is awesome. These days, I find myself really liking the help files (most of them anyway). But when I first started using R I found it very dense.\n To some extent, there’s not much I can do to help you with this. You just have to work at it yourself; once you’re moving away from being a pure beginner and are becoming a skilled user, you’ll start finding the help documentation more and more helpful. In the meantime, I’ll help as much as I can by trying to explain to you what you’re looking at when you open a help file. To that end, let’s look at the help documentation for the `load()` function. To do so, I type either of the following: \n\n```\n?load \nhelp(\"load\")\n```\n\nWhen I do that, R goes looking for the help file for the “load” topic. If it finds one, Rstudio takes it and displays it in the help panel. Alternatively, you can try a fuzzy search for a help topic\n\n```\n??load \nhelp.search(\"load\")\n```\n\n `load` ” topic as our example. Firstly, at the very top we see this: \n\nload {base} | R Documentation |\n | --- | --- |\n\n# Reload Saved Datasets\n\n# Description\n\n Reload datasets written with the function `save` . \n\n# Usage\n\n > load(file, envir = parent.frame(), verbose = FALSE)\n In this instance, the usage section is actually pretty readable. It’s telling you that there are two arguments to the `load()` function: the first one is called `file` , and the second one is called `envir` . It’s also telling you that there is a default value for the envir argument; so if the user doesn’t specify what the value of envir should be, then R will assume that \n\n. In contrast, the file argument has no default value at all, so the user must specify a value for it. So in one sense, this section is very straightforward. The problem, of course, is that you don’t know what the `parent.frame()` function actually does, so it’s hard for you to know what the \n\nbit is all about. What you could do is then go look up the help documents for the `parent.frame()` function (and sometimes that’s actually a good idea), but often you’ll find that the help documents for those functions are just as dense (if not more dense) than the help file that you’re currently reading. As an alternative, my general approach when faced with something like this is to skim over it, see if I can make any sense of it. If so, great. If not, I find that the best thing to do is ignore it. In fact, the first time I read the help file for the load() function, I had no idea what any of the `envir` related stuff was about. But fortunately I didn’t have to: the default setting here (i.e., \n\n) is actually the thing you want in about 99% of cases, so it’s safe to ignore it. \nBasically, what I’m trying to say is: don’t let the scary, incomprehensible parts of the help file intimidate you. Especially because there’s often some parts of the help file that will make sense. Of course, I guarantee you that sometimes this strategy will lead you to make mistakes… often embarrassing mistakes. But it’s still better than getting paralysed with fear.\nSo, let’s continue on. The next part of the help documentation discusses each of the arguments, and what they’re supposed to do: \n\n# Arguments\n\nfilea (readable binary-mode) connection or a character string giving the name of the file to load (when tilde expansion is done).envirthe environment where the data should be loaded.verboseshould item names be printed during loading?\n Okay, so what this is telling us is that the `file` argument needs to be a string (i.e., text data) which tells R the name of the file to load. It also seems to be hinting that there’s other possibilities too (e.g., a “binary mode connection”), and you probably aren’t quite sure what “tilde expansion” means64. But overall, the meaning is pretty clear. Turning to the `envir` argument, it’s now a little clearer what the Usage section was babbling about. The `envir` argument specifies the name of an environment (see Section 4.3 if you’ve forgotten what environments are) into which R should place the variables when it loads the file. Almost always, this is a no-brainer: you want R to load the data into the same damn environment in which you’re invoking the `load()` command. That is, if you’re typing `load()` at the R prompt, then you want the data to be loaded into your workspace (i.e., the global environment). But if you’re writing your own function that needs to load some data, you want the data to be loaded inside that function’s private workspace. And in fact, that’s exactly what the `parent.frame()` thing is all about. It’s telling the `load()` function to send the data to the same place that the `load()` command itself was coming from. As it turns out, if we’d just ignored the envir bit we would have been totally safe. Which is nice to know. \n\n# Details\n\n `load` can load R objects saved in the current or any earlier format. It can read a compressed file (see `save` ) directly from a file or from a suitable connection (including a call to `url` ). A not-open connection will be opened in mode `“rb”` and closed after use. Any connection other than a `gzfile` or `gzcon` connection will be wrapped in `gzcon` to allow compressed saves to be handled: note that this leaves the connection in an altered state (in particular, binary-only), and that it needs to be closed explicitly (it will not be garbage-collected). \nOnly R objects saved in the current format (used since R 1.4.0) can be read from a connection. If no input is available on a connection a warning will be given, but any input not in the current format will result in a error.\n Loading from an earlier version will give a warning about the ‘magic number’: magic numbers `1971:1977` are from R < 0.99.0, and `RD[ABX]1` from R 0.99.0 to R 1.3.1. These are all obsolete, and you are strongly recommended to re-save such files in a current format. The `verbose` argument is mainly intended for debugging. If it is `TRUE` , then as objects from the file are loaded, their names will be printed to the console. If `verbose` is set to an integer value greater than one, additional names corresponding to attributes and other parts of individual objects will also be printed. Larger values will print names to a greater depth. \nObjects can be saved with references to namespaces, usually as part of the environment of a function or formula. Such objects can be loaded even if the namespace is not available: it is replaced by a reference to the global environment with a warning. The warning identifies the first object with such a reference (but there may be more than one).\n\nThen it tells you what the output value of the function is:\n\n# Value\n\nA character vector of the names of objects created, invisibly.\n This is usually a bit more interesting, but since the `load()` function is mainly used to load variables into the workspace rather than to return a value, it’s no surprise that this doesn’t do much or say much. Moving on, we sometimes see a few additional sections in the help file, which can be different depending on what the function is: \n\n# Warning\n\n Saved R objects are binary files, even those saved with `ascii = TRUE` , so ensure that they are transferred without conversion of end of line markers. `load` tries to detect such a conversion and gives an informative error message. `load()` replaces all existing objects with the same names in the current environment (typically your workspace, `.GlobalEnv` ) and hence potentially overwrites important data. It is considerably safer to use `envir =` to load into a different environment, or to `attach(file)` which `load()` s into a new entry in the `search` path. \n\n# Note\n\n `file` can be a UTF-8-encoded filepath that cannot be translated to the current locale. \nYeah, yeah. Warning, warning, blah blah blah. Towards the bottom of the help file, we see something like this, which suggests a bunch of related topics that you might want to look at. These can be quite helpful:\n\n# See Also\n\n `save` , `download.file` ; further `attach` as wrapper for `load()` . For other interfaces to the underlying serialization format, see `unserialize` and `readRDS` . `load` ” help file: \n\n# Examples\n\n > ## save all data xx <- pi # to ensure there is some data save(list = ls(all = TRUE), file= \"all.rda\") rm(xx) ## restore the saved values to the current environment local({ load(\"all.rda\") ls() }) xx <- exp(1:3) ## restore the saved values to the user's workspace load(\"all.rda\") ## which is here *equivalent* to ## load(\"all.rda\", .GlobalEnv) ## This however annihilates all objects in .GlobalEnv with the same names ! xx # no longer exp(1:3) rm(xx) attach(\"all.rda\") # safer and will warn about masked objects w/ same name in .GlobalEnv ls(pos = 2) ## also typically need to cleanup the search path: detach(\"file:all.rda\") ## clean up (the example): unlink(\"all.rda\") ## Not run: con <- url(\"http://some.where.net/R/data/example.rda\") ## print the value to see what objects were created. print(load(con)) close(con) # url() always opens the connection ## End(Not run)\n As you can see, they’re pretty dense, and not at all obvious to the novice user. However, they do provide good examples of the various different things that you can do with the `load()` function, so it’s not a bad idea to have a look at them, and to try not to find them too intimidating. \n\n### 4.12.2 Other resources\n\n* The Rseek website (www.rseek.org). One thing that I really find annoying about the R help documentation is that it’s hard to search properly. When coupled with the fact that the documentation is dense and highly technical, it’s often a better idea to search or ask online for answers to your questions. With that in mind, the Rseek website is great: it’s an R specific search engine. I find it really useful, and it’s almost always my first port of call when I’m looking around.\n * The R-help mailing list (see http://www.r-project.org/mail.html for details). This is the official R help mailing list. It can be very helpful, but it’s very important that you do your homework before posting a question. The list gets a lot of traffic. While the people on the list try as hard as they can to answer questions, they do so for free, and you really don’t want to know how much money they could charge on an hourly rate if they wanted to apply market rates. In short, they are doing you a favour, so be polite. Don’t waste their time asking questions that can be easily answered by a quick search on Rseek (it’s rude), make sure your question is clear, and all of the relevant information is included. In short, read the posting guidelines carefully (http://www.r-project.org/posting-guide.html), and make use of the\n `help.request()` function that R provides to check that you’re actually doing what you’re expected. \n\n## 4.13 Summary\n\nThis chapter continued where Chapter 3 left off. The focus was still primarily on introducing basic R concepts, but this time at least you can see how those concepts are related to data analysis:\n\n* Installing, loading and updating packages. Knowing how to extend the functionality of R by installing and using packages is critical to becoming an effective R user\n * Getting around. Section 4.3 talked about how to manage your workspace and how to keep it tidy. Similarly, Section 4.4 talked about how to get R to interact with the rest of the file system.\n * Loading and saving data. Finally, we encountered actual data files. Loading and saving data is obviously a crucial skill, one we discussed in Section 4.5.\n * Useful things to know about variables. In particular, we talked about special values, element names and classes.\n * More complex types of variables. R has a number of important variable types that will be useful when analysing real data. I talked about factors in Section 4.7, data frames in Section 4.8, lists in Section 4.9 and formulas in Section 4.10.\n * Generic functions. How is it that some function seem to be able to do lots of different things? Section 4.11 tells you how.\n * Getting help. Assuming that you’re not looking for counselling, Section 4.12 covers several possibilities. If you are looking for counselling, well, this book really can’t help you there. Sorry.\n\nTaken together, Chapters 3 and 4 provide enough of a background that you can finally get started doing some statistics! Yes, there’s a lot more R concepts that you ought to know (and we’ll talk about some of them in Chapters7 and8), but I think that we’ve talked quite enough about programming for the moment. It’s time to see how your experience with programming can be used to do some data analysis…\n\nNotice that I used\n `print(keeper)` rather than just typing `keeper` . Later on in the text I’ll sometimes use the `print()` function to display things because I think it helps make clear what I’m doing, but in practice people rarely do this.↩ * \n\nMore precisely, there are 5000 or so packages on CRAN, the Comprehensive R Archive Network.↩\n * \n\nBasically, the reason is that there are 5000 packages, and probably about 4000 authors of packages, and no-one really knows what all of them do. Keeping the installation separate from the loading minimizes the chances that two packages will interact with each other in a nasty way.↩\n * \n\nIf you’re using the command line, you can get the same information by typing\n `library()` at the command line.↩ * \n\nThe logit function a simple mathematical function that happens not to have been included in the basic R distribution.↩\n * \n\nTip for advanced users. You can get R to use the one from the\n `car` package by using `car::logit()` as your command rather than `logit()` , since the `car::` part tells R explicitly which package to use. See also `:::` if you’re especially keen to force R to use functions it otherwise wouldn’t, but take care, since `:::` can be dangerous.↩ * \n\nIt is not very difficult.↩\n * \n\nThis would be especially annoying if you’re reading an electronic copy of the book because the text displayed by the\n `who()` function is searchable, whereas text shown in a screen shot isn’t!↩ * \n\nMind you, all that means is that it’s been removed from the workspace. If you’ve got the data saved to file somewhere, then that file is perfectly safe.↩\n * \n\nWell, the partition, technically.↩\n * \n\nOne additional thing worth calling your attention to is the\n `file.choose()` function. Suppose you want to load a file and you don’t quite remember where it is, but would like to browse for it. Typing `file.choose()` at the command line will open a window in which you can browse to find the file; when you click on the file you want, R will print out the full path to that file. This is kind of handy.↩ * \n\nNotably those with .rda, .Rd, .Rhistory, .rdb and .rdx extensions↩\n * \n\nIn a lot of books you’ll see the\n `read.table()` function used for this purpose instead of `read.csv()` . They’re more or less identical functions, with the same arguments and everything. They differ only in the default values.↩ * \n\nNote that I didn’t to this in my earlier example when loading the .Rdata↩\n * \n\nA word of warning: what you don’t want to do is use the “File” menu. If you look in the “File” menu you will see “Save” and “Save As…” options, but they don’t save the workspace. Those options are used for dealing with scripts, and so they’ll produce\n `.R` files. We won’t get to those until Chapter 8.↩ * \n\nOr functions. But let’s ignore functions for the moment.↩\n * \n\nActually, I don’t think I ever use this in practice. I don’t know why I bother to talk about it in the book anymore.↩\n * \n\nTaking all the usual caveats that attach to IQ measurement as a given, of course.↩\n * \n\nOr, more precisely, we don’t know how to measure it. Arguably, a rock has zero intelligence. But it doesn’t make sense to say that the IQ of a rock is 0 in the same way that we can say that the average human has an IQ of 100. And without knowing what the IQ value is that corresponds to a literal absence of any capacity to think, reason or learn, then we really can’t multiply or divide IQ scores and expect a meaningful answer.↩\n * \n\nOnce again, this is an example of coercing a variable from one class to another. I’ll talk about coercion in more detail in Section 7.10.↩\n * \n\nNote that, when I write out the formula, R doesn’t check to see if the\n `out` and `pred` variables actually exist: it’s only later on when you try to use the formula for something that this happens.↩ * \n\nFor readers with a programming background: what I’m describing is the very basics of how S3 methods work. However, you should be aware that R has two entirely distinct systems for doing object oriented programming, known as S3 and S4. Of the two, S3 is simpler and more informal, whereas S4 supports all the stuff that you might expect of a fully object oriented language. Most of the generics we’ll run into in this book use the S3 system, which is convenient for me because I’m still trying to figure out S4.↩\n * \n\nIt’s extremely simple, by the way. We discussed it in Section 4.4, though I didn’t call it by that name. Tilde expansion is the thing where R recognises that, in the context of specifying a file location, the tilde symbol ~ corresponds to the user home directory (e.g., /Users/dan/).↩\n\nDate: 2010-09-24\nCategories: \nTags: \n\n# Chapter 5 Descriptive statistics\n\n Any time that you get a new data set to look at, one of the first tasks that you have to do is find ways of summarising the data in a compact, easily understood fashion. This is what descriptive statistics (as opposed to inferential statistics) is all about. In fact, to many people the term “statistics” is synonymous with descriptive statistics. It is this topic that we’ll consider in this chapter, but before going into any details, let’s take a moment to get a sense of why we need descriptive statistics. To do this, let’s load the `aflsmall.Rdata` file, and use the `who()` function in the `lsr` package to see what variables are stored in the file: \n\n```\nload( \"./data/aflsmall.Rdata\" )\nlibrary(lsr)\nwho()\n```\n\n There are two variables here, `afl.finalists` and `afl.margins` . We’ll focus a bit on these two variables in this chapter, so I’d better tell you what they are. Unlike most of data sets in this book, these are actually real data, relating to the Australian Football League (AFL)65 The `afl.margins` variable contains the winning margin (number of points) for all 176 home and away games played during the 2010 season. The `afl.finalists` variable contains the names of all 400 teams that played in all 200 finals matches played during the period 1987 to 2010. Let’s have a look at the `afl.margins` variable: `print(afl.margins)` \n\n This output doesn’t make it easy to get a sense of what the data are actually saying. Just “looking at the data” isn’t a terribly effective way of understanding data. In order to get some idea about what’s going on, we need to calculate some descriptive statistics (this chapter) and draw some nice pictures (Chapter 6. Since the descriptive statistics are the easier of the two topics, I’ll start with those, but nevertheless I’ll show you a histogram of the `afl.margins` data, since it should help you get a sense of what the data we’re trying to describe actually look like. But for what it’s worth, this histogram – which is shown in Figure 5.1 – was generated using the `hist()` function. We’ll talk a lot more about how to draw histograms in Section 6.3. For now, it’s enough to look at the histogram and note that it provides a fairly interpretable representation of the `afl.margins` data. \n\n## 5.1 Measures of central tendency\n\nDrawing pictures of the data, as I did in Figure 5.1 is an excellent way to convey the “gist” of what the data is trying to tell you, it’s often extremely useful to try to condense the data into a few simple “summary” statistics. In most situations, the first thing that you’ll want to calculate is a measure of central tendency. That is, you’d like to know something about the “average” or “middle” of your data lies. The two most commonly used measures are the mean, median and mode; occasionally people will also report a trimmed mean. I’ll explain each of these in turn, and then discuss when each of them is useful.\n\n### 5.1.1 The mean\n\nThe mean of a set of observations is just a normal, old-fashioned average: add all of the values up, and then divide by the total number of values. The first five AFL margins were 56, 31, 56, 8 and 32, so the mean of these observations is just: \\[ \\frac{56 + 31 + 56 + 8 + 32}{5} = \\frac{183}{5} = 36.60 \\] Of course, this definition of the mean isn’t news to anyone: averages (i.e., means) are used so often in everyday life that this is pretty familiar stuff. However, since the concept of a mean is something that everyone already understands, I’ll use this as an excuse to start introducing some of the mathematical notation that statisticians use to describe this calculation, and talk about how the calculations would be done in R.\n The first piece of notation to introduce is \\(N\\), which we’ll use to refer to the number of observations that we’re averaging (in this case \\(N = 5\\)). Next, we need to attach a label to the observations themselves. It’s traditional to use \\(X\\) for this, and to use subscripts to indicate which observation we’re actually talking about. That is, we’ll use \\(X_1\\) to refer to the first observation, \\(X_2\\) to refer to the second observation, and so on, all the way up to \\(X_N\\) for the last one. Or, to say the same thing in a slightly more abstract way, we use \\(X_i\\) to refer to the \\(i\\)-th observation. Just to make sure we’re clear on the notation, the following table lists the 5 observations in the `afl.margins` variable, along with the mathematical symbol used to refer to it, and the actual value that the observation corresponds to: \n\nOkay, now let’s try to write a formula for the mean. By tradition, we use \\(\\bar{X}\\) as the notation for the mean. So the calculation for the mean could be expressed using the following formula: \\[ \\bar{X} = \\frac{X_1 + X_2 + ... + X_{N-1} + X_N}{N} \\] This formula is entirely correct, but it’s terribly long, so we make use of the summation symbol \\(\\scriptstyle\\sum\\) to shorten it.66 If I want to add up the first five observations, I could write out the sum the long way, \\(X_1 + X_2 + X_3 + X_4 +X_5\\) or I could use the summation symbol to shorten it to this: \\[ \\sum_{i=1}^5 X_i \\] Taken literally, this could be read as “the sum, taken over all \\(i\\) values from 1 to 5, of the value \\(X_i\\)”. But basically, what it means is “add up the first five observations”. In any case, we can use this notation to write out the formula for the mean, which looks like this: \\[ \\bar{X} = \\frac{1}{N} \\sum_{i=1}^N X_i \\]\n\nIn all honesty, I can’t imagine that all this mathematical notation helps clarify the concept of the mean at all. In fact, it’s really just a fancy way of writing out the same thing I said in words: add all the values up, and then divide by the total number of items. However, that’s not really the reason I went into all that detail. My goal was to try to make sure that everyone reading this book is clear on the notation that we’ll be using throughout the book: \\(\\bar{X}\\) for the mean, \\(\\scriptstyle\\sum\\) for the idea of summation, \\(X_i\\) for the \\(i\\)th observation, and \\(N\\) for the total number of observations. We’re going to be re-using these symbols a fair bit, so it’s important that you understand them well enough to be able to “read” the equations, and to be able to see that it’s just saying “add up lots of things and then divide by another thing”.\n\n### 5.1.2 Calculating the mean in R\n\nOkay that’s the maths, how do we get the magic computing box to do the work for us? If you really wanted to, you could do this calculation directly in R. For the first 5 AFL scores, do this just by typing it in as if R were a calculator…\n\n```\n(56 + 31 + 56 + 8 + 32) / 5\n```\n\n `## [1] 36.6` … in which case R outputs the answer 36.6, just as if it were a calculator. However, that’s not the only way to do the calculations, and when the number of observations starts to become large, it’s easily the most tedious. Besides, in almost every real world scenario, you’ve already got the actual numbers stored in a variable of some kind, just like we have with the `afl.margins` variable. Under those circumstances, what you want is a function that will just add up all the values stored in a numeric vector. That’s what the `sum()` function does. If we want to add up all 176 winning margins in the data set, we can do so using the following command:67 `sum( afl.margins )` `## [1] 6213` \nIf we only want the sum of the first five observations, then we can use square brackets to pull out only the first five elements of the vector. So the command would now be:\n\n```\nsum( afl.margins[1:5] )\n```\n\n `## [1] 183` \nTo calculate the mean, we now tell R to divide the output of this summation by five, so the command that we need to type now becomes the following:\n\n```\nsum( afl.margins[1:5] ) / 5\n```\n\n `## [1] 36.6` Although it’s pretty easy to calculate the mean using the `sum()` function, we can do it in an even easier way, since R also provides us with the `mean()` function. To calculate the mean for all 176 games, we would use the following command: \n\n `## [1] 35.30114` However, since `x` is the first argument to the function, I could have omitted the argument name. In any case, just to show you that there’s nothing funny going on, here’s what we would do to calculate the mean for the first five observations: \n\n```\nmean( afl.margins[1:5] )\n```\n\n `## [1] 36.6` \nAs you can see, this gives exactly the same answers as the previous calculations.\n\n### 5.1.3 The median\n\n The second measure of central tendency that people use a lot is the median, and it’s even easier to describe than the mean. The median of a set of observations is just the middle value. As before let’s imagine we were interested only in the first 5 AFL winning margins: 56, 31, 56, 8 and 32. To figure out the median, we sort these numbers into ascending order: \\[ 8, 31, \\mathbf{32}, 56, 56 \\] From inspection, it’s obvious that the median value of these 5 observations is 32, since that’s the middle one in the sorted list (I’ve put it in bold to make it even more obvious). Easy stuff. But what should we do if we were interested in the first 6 games rather than the first 5? Since the sixth game in the season had a winning margin of 14 points, our sorted list is now \\[ 8, 14, \\mathbf{31}, \\mathbf{32}, 56, 56 \\] and there are two middle numbers, 31 and 32. The median is defined as the average of those two numbers, which is of course 31.5. As before, it’s very tedious to do this by hand when you’ve got lots of numbers. To illustrate this, here’s what happens when you use R to sort all 176 winning margins. First, I’ll use the `sort()` function (discussed in Chapter 7) to display the winning margins in increasing numerical order: \n\nThe middle values are 30 and 31, so the median winning margin for 2010 was 30.5 points. In real life, of course, no-one actually calculates the median by sorting the data and then looking for the middle value. In real life, we use the median command:\n\n `## [1] 30.5` \nwhich outputs the median value of 30.5.\n\n### 5.1.4 Mean or median? What’s the difference?\n\nKnowing how to calculate means and medians is only a part of the story. You also need to understand what each one is saying about the data, and what that implies for when you should use each one. This is illustrated in Figure 5.2 the mean is kind of like the “centre of gravity” of the data set, whereas the median is the “middle value” in the data. What this implies, as far as which one you should use, depends a little on what type of data you’ve got and what you’re trying to achieve. As a rough guide:\n\n* If your data are nominal scale, you probably shouldn’t be using either the mean or the median. Both the mean and the median rely on the idea that the numbers assigned to values are meaningful. If the numbering scheme is arbitrary, then it’s probably best to use the mode (Section 5.1.7) instead.\n * If your data are ordinal scale, you’re more likely to want to use the median than the mean. The median only makes use of the order information in your data (i.e., which numbers are bigger), but doesn’t depend on the precise numbers involved. That’s exactly the situation that applies when your data are ordinal scale. The mean, on the other hand, makes use of the precise numeric values assigned to the observations, so it’s not really appropriate for ordinal data.\n * For interval and ratio scale data, either one is generally acceptable. Which one you pick depends a bit on what you’re trying to achieve. The mean has the advantage that it uses all the information in the data (which is useful when you don’t have a lot of data), but it’s very sensitive to extreme values, as we’ll see in Section 5.1.6.\n\nLet’s expand on that last part a little. One consequence is that there’s systematic differences between the mean and the median when the histogram is asymmetric (skewed; see Section 5.3). This is illustrated in Figure 5.2 notice that the median (right hand side) is located closer to the “body” of the histogram, whereas the mean (left hand side) gets dragged towards the “tail” (where the extreme values are). To give a concrete example, suppose Bob (income $50,000), Kate (income $60,000) and Jane (income $65,000) are sitting at a table: the average income at the table is $58,333 and the median income is $60,000. Then Bill sits down with them (income $100,000,000). The average income has now jumped to $25,043,750 but the median rises only to $62,500. If you’re interested in looking at the overall income at the table, the mean might be the right answer; but if you’re interested in what counts as a typical income at the table, the median would be a better choice here.\n\n### 5.1.5 A real life example\n\nTo try to get a sense of why you need to pay attention to the differences between the mean and the median, let’s consider a real life example. Since I tend to mock journalists for their poor scientific and statistical knowledge, I should give credit where credit is due. This is from an excellent article on the ABC news website68 24 September, 2010:\n Senior Commonwealth Bank executives have travelled the world in the past couple of weeks with a presentation showing how Australian house prices, and the key price to income ratios, compare favourably with similar countries. “Housing affordability has actually been going sideways for the last five to six years,” said , the chief economist of the bank’s trading arm, CommSec.\n\nThis probably comes as a huge surprise to anyone with a mortgage, or who wants a mortgage, or pays rent, or isn’t completely oblivious to what’s been going on in the Australian housing market over the last several years. Back to the article:\n CBA has waged its war against what it believes are housing doomsayers with graphs, numbers and international comparisons. In its presentation, the bank rejects arguments that Australia’s housing is relatively expensive compared to incomes. It says Australia’s house price to household income ratio of 5.6 in the major cities, and 4.3 nationwide, is comparable to many other developed nations. It says San Francisco and New York have ratios of 7, Auckland’s is 6.7, and Vancouver comes in at 9.3.\n\nMore excellent news! Except, the article goes on to make the observation that…\n Many analysts say that has led the bank to use misleading figures and comparisons. If you go to page four of CBA’s presentation and read the source information at the bottom of the graph and table, you would notice there is an additional source on the international comparison – Demographia. However, if the Commonwealth Bank had also used Demographia’s analysis of Australia’s house price to income ratio, it would have come up with a figure closer to 9 rather than 5.6 or 4.3\n\nThat’s, um, a rather serious discrepancy. One group of people say 9, another says 4-5. Should we just split the difference, and say the truth lies somewhere in between? Absolutely not: this is a situation where there is a right answer and a wrong answer. Demographia are correct, and the Commonwealth Bank is incorrect. As the article points out\n [An] obvious problem with the Commonwealth Bank’s domestic price to income figures is they compare average incomes with median house prices (unlike the Demographia figures that compare median incomes to median prices). The median is the mid-point, effectively cutting out the highs and lows, and that means the average is generally higher when it comes to incomes and asset prices, because it includes the earnings of Australia’s wealthiest people. To put it another way: the Commonwealth Bank’s figures count ’ multi-million dollar pay packet on the income side, but not his (no doubt) very expensive house in the property price figures, thus understating the house price to income ratio for middle-income Australians.\n\nCouldn’t have put it better myself. The way that Demographia calculated the ratio is the right thing to do. The way that the Bank did it is incorrect. As for why an extremely quantitatively sophisticated organisation such as a major bank made such an elementary mistake, well… I can’t say for sure, since I have no special insight into their thinking, but the article itself does happen to mention the following facts, which may or may not be relevant:\n [As] Australia’s largest home lender, the Commonwealth Bank has one of the biggest vested interests in house prices rising. It effectively owns a massive swathe of Australian housing as security for its home loans as well as many small business loans.\n\nMy, my.\n\n### 5.1.6 Trimmed mean\n\nOne of the fundamental rules of applied statistics is that the data are messy. Real life is never simple, and so the data sets that you obtain are never as straightforward as the statistical theory says.69 This can have awkward consequences. To illustrate, consider this rather strange looking data set: \\[ -100,2,3,4,5,6,7,8,9,10 \\] If you were to observe this in a real life data set, you’d probably suspect that something funny was going on with the \\(-100\\) value. It’s probably an outlier, a value that doesn’t really belong with the others. You might consider removing it from the data set entirely, and in this particular case I’d probably agree with that course of action. In real life, however, you don’t always get such cut-and-dried examples. For instance, you might get this instead: \\[ -15,2,3,4,5,6,7,8,9,12 \\] The \\(-15\\) looks a bit suspicious, but not anywhere near as much as that \\(-100\\) did. In this case, it’s a little trickier. It might be a legitimate observation, it might not.\n\nWhen faced with a situation where some of the most extreme-valued observations might not be quite trustworthy, the mean is not necessarily a good measure of central tendency. It is highly sensitive to one or two extreme values, and is thus not considered to be a robust measure. One remedy that we’ve seen is to use the median. A more general solution is to use a “trimmed mean”. To calculate a trimmed mean, what you do is “discard” the most extreme examples on both ends (i.e., the largest and the smallest), and then take the mean of everything else. The goal is to preserve the best characteristics of the mean and the median: just like a median, you aren’t highly influenced by extreme outliers, but like the mean, you “use” more than one of the observations. Generally, we describe a trimmed mean in terms of the percentage of observation on either side that are discarded. So, for instance, a 10% trimmed mean discards the largest 10% of the observations and the smallest 10% of the observations, and then takes the mean of the remaining 80% of the observations. Not surprisingly, the 0% trimmed mean is just the regular mean, and the 50% trimmed mean is the median. In that sense, trimmed means provide a whole family of central tendency measures that span the range from the mean to the median.\n For our toy example above, we have 10 observations, and so a 10% trimmed mean is calculated by ignoring the largest value (i.e., `12` ) and the smallest value (i.e., `-15` ) and taking the mean of the remaining values. First, let’s enter the data \n\n```\ndataset <- c( -15,2,3,4,5,6,7,8,9,12 )\n```\n\nNext, let’s calculate means and medians:\n `mean( x = dataset )` `## [1] 4.1` \n\n `## [1] 5.5` \nThat’s a fairly substantial difference, but I’m tempted to think that the mean is being influenced a bit too much by the extreme values at either end of the data set, especially the \\(-15\\) one. So let’s just try trimming the mean a bit. If I take a 10% trimmed mean, we’ll drop the extreme values on either side, and take the mean of the rest:\n\n```\nmean( x = dataset, trim = .1)\n```\n\n `## [1] 5.5` which in this case gives exactly the same answer as the median. Note that, to get a 10% trimmed mean you write `trim = .1` , not `trim = 10` . In any case, let’s finish up by calculating the 5% trimmed mean for the `afl.margins` data, \n\n```\nmean( x = afl.margins, trim = .05)\n```\n\n `## [1] 33.75` \n\n### 5.1.7 Mode\n\n The mode of a sample is very simple: it is the value that occurs most frequently. To illustrate the mode using the AFL data, let’s examine a different aspect to the data set. Who has played in the most finals? The `afl.finalists` variable is a factor that contains the name of every team that played in any AFL final from 1987-2010, so let’s have a look at it. To do this we will use the `head()` command. `head()` is useful when you’re working with a data.frame with a lot of rows since you can use it to tell you how many rows to return. There have been a lot of finals in this period so printing afl.finalists using `print(afl.finalists)` will just fill us the screen. The command below tells R we just want the first 25 rows of the data.frame. \n\n```\nhead(afl.finalists, 25)\n```\n\n```\n## [1] Hawthorn Melbourne Carlton Melbourne Hawthorn \n## [6] Carlton Melbourne Carlton Hawthorn Melbourne \n## [11] Melbourne Hawthorn Melbourne Essendon Hawthorn \n## [16] Geelong Geelong Hawthorn Collingwood Melbourne \n## [21] Collingwood West Coast Collingwood Essendon Collingwood\n## 17 Levels: Adelaide Brisbane Carlton Collingwood Essendon ... Western Bulldogs\n```\n\n There are actually 400 entries (aren’t you glad we didn’t print them all?). We could read through all 400, and count the number of occasions on which each team name appears in our list of finalists, thereby producing a frequency table. However, that would be mindless and boring: exactly the sort of task that computers are great at. So let’s use the `table()` function (discussed in more detail in Section 7.1) to do this task for us: \n\n```\ntable( afl.finalists )\n```\n\n Now that we have our frequency table, we can just look at it and see that, over the 24 years for which we have data, Geelong has played in more finals than any other team. Thus, the mode of the `finalists` data is `\"Geelong\"` . The core packages in R don’t have a function for calculating the mode70. However, I’ve included a function in the `lsr` package that does this. The function is called `modeOf()` , and here’s how you use it: \n\n```\nmodeOf( x = afl.finalists )\n```\n\n `## [1] \"Geelong\"` There’s also a function called `maxFreq()` that tells you what the modal frequency is. If we apply this function to our finalists data, we obtain the following: \n\n```\nmaxFreq( x = afl.finalists )\n```\n\n `## [1] 39` \nTaken together, we observe that Geelong (39 finals) played in more finals than any other team during the 1987-2010 period.\n One last point to make with respect to the mode. While it’s generally true that the mode is most often calculated when you have nominal scale data (because means and medians are useless for those sorts of variables), there are some situations in which you really do want to know the mode of an ordinal, interval or ratio scale variable. For instance, let’s go back to thinking about our `afl.margins` variable. This variable is clearly ratio scale (if it’s not clear to you, it may help to re-read Section 2.2), and so in most situations the mean or the median is the measure of central tendency that you want. But consider this scenario… a friend of yours is offering a bet. They pick a football game at random, and (without knowing who is playing) you have to guess the exact margin. If you guess correctly, you win $50. If you don’t, you lose $1. There are no consolation prizes for “almost” getting the right answer. You have to guess exactly the right margin71 For this bet, the mean and the median are completely useless to you. It is the mode that you should bet on. So, we calculate this modal value \n\n```\nmodeOf( x = afl.margins )\n```\n\n `## [1] 3` \n\n```\nmaxFreq( x = afl.margins )\n```\n\n `## [1] 8` \nSo the 2010 data suggest you should bet on a 3 point margin, and since this was observed in 8 of the 176 game (4.5% of games) the odds are firmly in your favour.\n\n## 5.2 Measures of variability\n\n The statistics that we’ve discussed so far all relate to central tendency. That is, they all talk about which values are “in the middle” or “popular” in the data. However, central tendency is not the only type of summary statistic that we want to calculate. The second thing that we really want is a measure of the variability of the data. That is, how “spread out” are the data? How “far” away from the mean or median do the observed values tend to be? For now, let’s assume that the data are interval or ratio scale, so we’ll continue to use the `afl.margins` data. We’ll use this data to discuss several different measures of spread, each with different strengths and weaknesses. \n\n### 5.2.1 Range\n\n The range of a variable is very simple: it’s the biggest value minus the smallest value. For the AFL winning margins data, the maximum value is 116, and the minimum value is 0. We can calculate these values in R using the `max()` and `min()` functions: `max( afl.margins )` `## [1] 116` `min( afl.margins )` `## [1] 0` where I’ve omitted the output because it’s not interesting. The other possibility is to use the `range()` function; which outputs both the minimum value and the maximum value in a vector, like this: `range( afl.margins )` `## [1] 0 116` \nAlthough the range is the simplest way to quantify the notion of “variability”, it’s one of the worst. Recall from our discussion of the mean that we want our summary measure to be robust. If the data set has one or two extremely bad values in it, we’d like our statistics not to be unduly influenced by these cases. If we look once again at our toy example of a data set containing very extreme outliers… \\[ -100,2,3,4,5,6,7,8,9,10 \\] … it is clear that the range is not robust, since this has a range of 110, but if the outlier were removed we would have a range of only 8.\n\n### 5.2.2 Interquartile range\n\n The interquartile range (IQR) is like the range, but instead of calculating the difference between the biggest and smallest value, it calculates the difference between the 25th quantile and the 75th quantile. Probably you already know what a quantile is (they’re more commonly called percentiles), but if not: the 10th percentile of a data set is the smallest number \\(x\\) such that 10% of the data is less than \\(x\\). In fact, we’ve already come across the idea: the median of a data set is its 50th quantile / percentile! R actually provides you with a way of calculating quantiles, using the (surprise, surprise) `quantile()` function. Let’s use it to calculate the median AFL winning margin: \n\n```\nquantile( x = afl.margins, probs = .5)\n```\n\n```\n## 50% \n## 30.5\n```\n\n And not surprisingly, this agrees with the answer that we saw earlier with the `median()` function. Now, we can actually input lots of quantiles at once, by specifying a vector for the `probs` argument. So lets do that, and get the 25th and 75th percentile: \n\n```\nquantile( x = afl.margins, probs = c(.25,.75) )\n```\n\n```\n## 25% 75% \n## 12.75 50.50\n```\n\n And, by noting that \\(50.5 - 12.75 = 37.75\\), we can see that the interquartile range for the 2010 AFL winning margins data is 37.75. Of course, that seems like too much work to do all that typing, so R has a built in function called `IQR()` that we can use: \n\n```\nIQR( x = afl.margins )\n```\n\n `## [1] 37.75` \nWhile it’s obvious how to interpret the range, it’s a little less obvious how to interpret the IQR. The simplest way to think about it is like this: the interquartile range is the range spanned by the “middle half” of the data. That is, one quarter of the data falls below the 25th percentile, one quarter of the data is above the 75th percentile, leaving the “middle half” of the data lying in between the two. And the IQR is the range covered by that middle half.\n\n### 5.2.3 Mean absolute deviation\n\nThe two measures we’ve looked at so far, the range and the interquartile range, both rely on the idea that we can measure the spread of the data by looking at the quantiles of the data. However, this isn’t the only way to think about the problem. A different approach is to select a meaningful reference point (usually the mean or the median) and then report the “typical” deviations from that reference point. What do we mean by “typical” deviation? Usually, the mean or median value of these deviations! In practice, this leads to two different measures, the “mean absolute deviation (from the mean)” and the “median absolute deviation (from the median)”. From what I’ve read, the measure based on the median seems to be used in statistics, and does seem to be the better of the two, but to be honest I don’t think I’ve seen it used much in psychology. The measure based on the mean does occasionally show up in psychology though. In this section I’ll talk about the first one, and I’ll come back to talk about the second one later.\n Since the previous paragraph might sound a little abstract, let’s go through the mean absolute deviation from the mean a little more slowly. One useful thing about this measure is that the name actually tells you exactly how to calculate it. Let’s think about our AFL winning margins data, and once again we’ll start by pretending that there’s only 5 games in total, with winning margins of 56, 31, 56, 8 and 32. Since our calculations rely on an examination of the deviation from some reference point (in this case the mean), the first thing we need to calculate is the mean, \\(\\bar{X}\\). For these five observations, our mean is \\(\\bar{X} = 36.6\\). The next step is to convert each of our observations \\(X_i\\) into a deviation score. We do this by calculating the difference between the observation \\(X_i\\) and the mean \\(\\bar{X}\\). That is, the deviation score is defined to be \\(X_i - \\bar{X}\\). For the first observation in our sample, this is equal to \\(56 - 36.6 = 19.4\\). Okay, that’s simple enough. The next step in the process is to convert these deviations to absolute deviations. As we discussed earlier when talking about the `abs()` function in R (Section 3.5), we do this by converting any negative values to positive ones. Mathematically, we would denote the absolute value of \\(-3\\) as \\(|-3|\\), and so we say that \\(|-3| = 3\\). We use the absolute value function here because we don’t really care whether the value is higher than the mean or lower than the mean, we’re just interested in how close it is to the mean. To help make this process as obvious as possible, the table below shows these calculations for all five observations: \n\nNow that we have calculated the absolute deviation score for every observation in the data set, all that we have to do to calculate the mean of these scores. Let’s do that: \\[ \\frac{19.4 + 5.6 + 19.4 + 28.6 + 4.6}{5} = 15.52 \\] And we’re done. The mean absolute deviation for these five scores is 15.52.\n\nHowever, while our calculations for this little example are at an end, we do have a couple of things left to talk about. Firstly, we should really try to write down a proper mathematical formula. But in order do to this I need some mathematical notation to refer to the mean absolute deviation. Irritatingly, “mean absolute deviation” and “median absolute deviation” have the same acronym (MAD), which leads to a certain amount of ambiguity, and since R tends to use MAD to refer to the median absolute deviation, I’d better come up with something different for the mean absolute deviation. Sigh. What I’ll do is use AAD instead, short for average absolute deviation. Now that we have some unambiguous notation, here’s the formula that describes what we just calculated: \\[ \\mbox{}(X) = \\frac{1}{N} \\sum_{i = 1}^N |X_i - \\bar{X}| \\]\n\nThe last thing we need to talk about is how to calculate AAD in R. One possibility would be to do everything using low level commands, laboriously following the same steps that I used when describing the calculations above. However, that’s pretty tedious. You’d end up with a series of commands that might look like this:\n\n```\nX <- c(56, 31,56,8,32) # enter the data\nX.bar <- mean( X ) # step 1. the mean of the data\nAD <- abs( X - X.bar ) # step 2. the absolute deviations from the mean\nAAD <- mean( AD ) # step 3. the mean absolute deviations\nprint( AAD ) # print the results\n```\n\n `## [1] 15.52` Each of those commands is pretty simple, but there’s just too many of them. And because I find that to be too much typing, the `lsr` package has a very simple function called `aad()` that does the calculations for you. If we apply the `aad()` function to our data, we get this: \n\n```\nlibrary(lsr)\naad( X )\n```\n\n `## [1] 15.52` \nNo suprises there.\n\n### 5.2.4 Variance\n\nAlthough the mean absolute deviation measure has its uses, it’s not the best measure of variability to use. From a purely mathematical perspective, there are some solid reasons to prefer squared deviations rather than absolute deviations. If we do that, we obtain a measure is called the variance, which has a lot of really nice statistical properties that I’m going to ignore,72(X)$ and \\(\\mbox{Var}(Y)\\) respectively. Now imagine I want to define a new variable \\(Z\\) that is the sum of the two, \\(Z = X+Y\\). As it turns out, the variance of \\(Z\\) is equal to \\(\\mbox{Var}(X) + \\mbox{Var}(Y)\\). This is a very useful property, but it’s not true of the other measures that I talk about in this section.] and one massive psychological flaw that I’m going to make a big deal out of in a moment. The variance of a data set \\(X\\) is sometimes written as \\(\\mbox{Var}(X)\\), but it’s more commonly denoted \\(s^2\\) (the reason for this will become clearer shortly). The formula that we use to calculate the variance of a set of observations is as follows: \\[ \\mbox{Var}(X) = \\frac{1}{N} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^2 \\] \\[\\mbox{Var}(X) = \\frac{\\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^2}{N}\\] As you can see, it’s basically the same formula that we used to calculate the mean absolute deviation, except that instead of using “absolute deviations” we use “squared deviations”. It is for this reason that the variance is sometimes referred to as the “mean square deviation”.\n\nNow that we’ve got the basic idea, let’s have a look at a concrete example. Once again, let’s use the first five AFL games as our data. If we follow the same approach that we took last time, we end up with the following table:\n\nNotation [English] | \\(i\\) [which game] | \\(X_i\\) [value] | \\(X_i - \\bar{X}\\) [deviation from mean] | \\((X_i - \\bar{X})^2\\) [absolute deviation] |\n | --- | --- | --- | --- | --- |\n1 | 56 | 19.4 | 376.36 |\n2 | 31 | -5.6 | 31.36 |\n3 | 56 | 19.4 | 376.36 |\n4 | 8 | -28.6 | 817.96 |\n5 | 32 | -4.6 | 21.16 |\n\nThat last column contains all of our squared deviations, so all we have to do is average them. If we do that by typing all the numbers into R by hand…\n\n```\n( 376.36 + 31.36 + 376.36 + 817.96 + 21.16 ) / 5\n```\n\n `## [1] 324.64` \n… we end up with a variance of 324.64. Exciting, isn’t it? For the moment, let’s ignore the burning question that you’re all probably thinking (i.e., what the heck does a variance of 324.64 actually mean?) and instead talk a bit more about how to do the calculations in R, because this will reveal something very weird.\n As always, we want to avoid having to type in a whole lot of numbers ourselves. And as it happens, we have the vector `X` lying around, which we created in the previous section. With this in mind, we can calculate the variance of `X` by using the following command, \n\n```\nmean( (X - mean(X) )^2)\n```\n\n `## [1] 324.64` and as usual we get the same answer as the one that we got when we did everything by hand. However, I still think that this is too much typing. Fortunately, R has a built in function called `var()` which does calculate variances. So we could also do this… `var(X)` `## [1] 405.8` \nand you get the same… no, wait… you get a completely different answer. That’s just weird. Is R broken? Is this a typo? Is Dan an idiot?\n As it happens, the answer is no.73 It’s not a typo, and R is not making a mistake. To get a feel for what’s happening, let’s stop using the tiny data set containing only 5 data points, and switch to the full set of 176 games that we’ve got stored in our `afl.margins` vector. First, let’s calculate the variance by using the formula that I described above: \n\n```\nmean( (afl.margins - mean(afl.margins) )^2)\n```\n\n `## [1] 675.9718` Now let’s use the `var()` function: `var( afl.margins )` `## [1] 679.8345` Hm. These two numbers are very similar this time. That seems like too much of a coincidence to be a mistake. And of course it isn’t a mistake. In fact, it’s very simple to explain what R is doing here, but slightly trickier to explain why R is doing it. So let’s start with the “what”. What R is doing is evaluating a slightly different formula to the one I showed you above. Instead of averaging the squared deviations, which requires you to divide by the number of data points \\(N\\), R has chosen to divide by \\(N-1\\). In other words, the formula that R is using is this one\\[ \\frac{1}{N-1} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^2 \\] It’s easy enough to verify that this is what’s happening, as the following command illustrates: \n\n```\nsum( (X-mean(X))^2 ) / 4\n```\n\n `## [1] 405.8` This is the same answer that R gave us originally when we calculated `var(X)` originally. So that’s the what. The real question is why R is dividing by \\(N-1\\) and not by \\(N\\). After all, the variance is supposed to be the mean squared deviation, right? So shouldn’t we be dividing by \\(N\\), the actual number of observations in the sample? Well, yes, we should. However, as we’ll discuss in Chapter 10, there’s a subtle distinction between “describing a sample” and “making guesses about the population from which the sample came”. Up to this point, it’s been a distinction without a difference. Regardless of whether you’re describing a sample or drawing inferences about the population, the mean is calculated exactly the same way. Not so for the variance, or the standard deviation, or for many other measures besides. What I outlined to you initially (i.e., take the actual average, and thus divide by \\(N\\)) assumes that you literally intend to calculate the variance of the sample. Most of the time, however, you’re not terribly interested in the sample in and of itself. Rather, the sample exists to tell you something about the world. If so, you’re actually starting to move away from calculating a “sample statistic”, and towards the idea of estimating a “population parameter”. However, I’m getting ahead of myself. For now, let’s just take it on faith that R knows what it’s doing, and we’ll revisit the question later on when we talk about estimation in Chapter 10. \nOkay, one last thing. This section so far has read a bit like a mystery novel. I’ve shown you how to calculate the variance, described the weird “\\(N-1\\)” thing that R does and hinted at the reason why it’s there, but I haven’t mentioned the single most important thing… how do you interpret the variance? Descriptive statistics are supposed to describe things, after all, and right now the variance is really just a gibberish number. Unfortunately, the reason why I haven’t given you the human-friendly interpretation of the variance is that there really isn’t one. This is the most serious problem with the variance. Although it has some elegant mathematical properties that suggest that it really is a fundamental quantity for expressing variation, it’s completely useless if you want to communicate with an actual human… variances are completely uninterpretable in terms of the original variable! All the numbers have been squared, and they don’t mean anything anymore. This is a huge issue. For instance, according to the table I presented earlier, the margin in game 1 was “376.36 points-squared higher than the average margin”. This is exactly as stupid as it sounds; and so when we calculate a variance of 324.64, we’re in the same situation. I’ve watched a lot of footy games, and never has anyone referred to “points squared”. It’s not a real unit of measurement, and since the variance is expressed in terms of this gibberish unit, it is totally meaningless to a human.\n\n### 5.2.5 Standard deviation\n\n Okay, suppose that you like the idea of using the variance because of those nice mathematical properties that I haven’t talked about, but – since you’re a human and not a robot – you’d like to have a measure that is expressed in the same units as the data itself (i.e., points, not points-squared). What should you do? The solution to the problem is obvious: take the square root of the variance, known as the standard deviation, also called the “root mean squared deviation”, or RMSD. This solves out problem fairly neatly: while nobody has a clue what “a variance of 324.68 points-squared” really means, it’s much easier to understand “a standard deviation of 18.01 points”, since it’s expressed in the original units. It is traditional to refer to the standard deviation of a sample of data as \\(s\\), though “sd” and “std dev.” are also used at times. Because the standard deviation is equal to the square root of the variance, you probably won’t be surprised to see that the formula is: \\[ s = \\sqrt{ \\frac{1}{N} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^2 } \\] and the R function that we use to calculate it is `sd()` . However, as you might have guessed from our discussion of the variance, what R actually calculates is slightly different to the formula given above. Just like the we saw with the variance, what R calculates is a version that divides by \\(N-1\\) rather than \\(N\\). For reasons that will make sense when we return to this topic in Chapter@refch:estimation I’ll refer to this new quantity as \\(\\hat\\sigma\\) (read as: “sigma hat”), and the formula for this is \\[\n\\hat\\sigma = \\sqrt{ \\frac{1}{N-1} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^2 }\n\\] With that in mind, calculating standard deviations in R is simple: `sd( afl.margins )` `## [1] 26.07364` \nInterpreting standard deviations is slightly more complex. Because the standard deviation is derived from the variance, and the variance is a quantity that has little to no meaning that makes sense to us humans, the standard deviation doesn’t have a simple interpretation. As a consequence, most of us just rely on a simple rule of thumb: in general, you should expect 68% of the data to fall within 1 standard deviation of the mean, 95% of the data to fall within 2 standard deviation of the mean, and 99.7% of the data to fall within 3 standard deviations of the mean. This rule tends to work pretty well most of the time, but it’s not exact: it’s actually calculated based on an assumption that the histogram is symmetric and “bell shaped.”74 As you can tell from looking at the AFL winning margins histogram in Figure 5.1, this isn’t exactly true of our data! Even so, the rule is approximately correct. As it turns out, 65.3% of the AFL margins data fall within one standard deviation of the mean. This is shown visually in Figure 5.3.\n\n### 5.2.6 Median absolute deviation\n\nThe last measure of variability that I want to talk about is the median absolute deviation (MAD). The basic idea behind MAD is very simple, and is pretty much identical to the idea behind the mean absolute deviation (Section 5.2.3). The difference is that you use the median everywhere. If we were to frame this idea as a pair of R commands, they would look like this:\n\n```\n# mean absolute deviation from the mean:\nmean( abs(afl.margins - mean(afl.margins)) )\n```\n\n `## [1] 21.10124` \n\n```\n# *median* absolute deviation from the *median*:\nmedian( abs(afl.margins - median(afl.margins)) )\n```\n\n `## [1] 19.5` \nThis has a straightforward interpretation: every observation in the data set lies some distance away from the typical value (the median). So the MAD is an attempt to describe a typical deviation from a typical value in the data set. It wouldn’t be unreasonable to interpret the MAD value of 19.5 for our AFL data by saying something like this:\n The median winning margin in 2010 was 30.5, indicating that a typical game involved a winning margin of about 30 points. However, there was a fair amount of variation from game to game: the MAD value was 19.5, indicating that a typical winning margin would differ from this median value by about 19-20 points.\n As you’d expect, R has a built in function for calculating MAD, and you will be shocked no doubt to hear that it’s called `mad()` . However, it’s a little bit more complicated than the functions that we’ve been using previously. If you want to use it to calculate MAD in the exact same way that I have described it above, the command that you need to use specifies two arguments: the data set itself `x` , and a `constant` that I’ll explain in a moment. For our purposes, the constant is 1, so our command becomes \n\n```\nmad( x = afl.margins, constant = 1 )\n```\n\n `## [1] 19.5` Apart from the weirdness of having to type that `constant = 1` part, this is pretty straightforward. Okay, so what exactly is this `constant = 1` argument? I won’t go into all the details here, but here’s the gist. Although the “raw” MAD value that I’ve described above is completely interpretable on its own terms, that’s not actually how it’s used in a lot of real world contexts. Instead, what happens a lot is that the researcher actually wants to calculate the standard deviation. However, in the same way that the mean is very sensitive to extreme values, the standard deviation is vulnerable to the exact same issue. So, in much the same way that people sometimes use the median as a “robust” way of calculating “something that is like the mean”, it’s not uncommon to use MAD as a method for calculating “something that is like the standard deviation”. Unfortunately, the raw MAD value doesn’t do this. Our raw MAD value is 19.5, and our standard deviation was 26.07. However, what some clever person has shown is that, under certain assumptions75, you can multiply the raw MAD value by 1.4826 and obtain a number that is directly comparable to the standard deviation. As a consequence, the default value of `constant` is 1.4826, and so when you use the `mad()` command without manually setting a value, here’s what you get: `mad( afl.margins )` `## [1] 28.9107` I should point out, though, that if you want to use this “corrected” MAD value as a robust version of the standard deviation, you really are relying on the assumption that the data are (or at least, are “supposed to be” in some sense) symmetric and basically shaped like a bell curve. That’s really not true for our `afl.margins` data, so in this case I wouldn’t try to use the MAD value this way. \n\n### 5.2.7 Which measure to use?\n\nWe’ve discussed quite a few measures of spread (range, IQR, MAD, variance and standard deviation), and hinted at their strengths and weaknesses. Here’s a quick summary:\n\n* Range. Gives you the full spread of the data. It’s very vulnerable to outliers, and as a consequence it isn’t often used unless you have good reasons to care about the extremes in the data.\n * Interquartile range. Tells you where the “middle half” of the data sits. It’s pretty robust, and complements the median nicely. This is used a lot.\n * Mean absolute deviation. Tells you how far “on average” the observations are from the mean. It’s very interpretable, but has a few minor issues (not discussed here) that make it less attractive to statisticians than the standard deviation. Used sometimes, but not often.\n * Variance. Tells you the average squared deviation from the mean. It’s mathematically elegant, and is probably the “right” way to describe variation around the mean, but it’s completely uninterpretable because it doesn’t use the same units as the data. Almost never used except as a mathematical tool; but it’s buried “under the hood” of a very large number of statistical tools.\n * Standard deviation. This is the square root of the variance. It’s fairly elegant mathematically, and it’s expressed in the same units as the data so it can be interpreted pretty well. In situations where the mean is the measure of central tendency, this is the default. This is by far the most popular measure of variation.\n * Median absolute deviation. The typical (i.e., median) deviation from the median value. In the raw form it’s simple and interpretable; in the corrected form it’s a robust way to estimate the standard deviation, for some kinds of data sets. Not used very often, but it does get reported sometimes.\n\nIn short, the IQR and the standard deviation are easily the two most common measures used to report the variability of the data; but there are situations in which the others are used. I’ve described all of them in this book because there’s a fair chance you’ll run into most of these somewhere.\n\n## 5.3 Skew and kurtosis\n\nThere are two more descriptive statistics that you will sometimes see reported in the psychological literature, known as skew and kurtosis. In practice, neither one is used anywhere near as frequently as the measures of central tendency and variability that we’ve been talking about. Skew is pretty important, so you do see it mentioned a fair bit; but I’ve actually never seen kurtosis reported in a scientific article to date.\n `## [1] -0.9246796` `## [1] -0.00814197` `## [1] 0.9197452` Since it’s the more interesting of the two, let’s start by talking about the skewness. Skewness is basically a measure of asymmetry, and the easiest way to explain it is by drawing some pictures. As Figure 5.4 illustrates, if the data tend to have a lot of extreme small values (i.e., the lower tail is “longer” than the upper tail) and not so many extremely large values (left panel), then we say that the data are negatively skewed. On the other hand, if there are more extremely large values than extremely small ones (right panel) we say that the data are positively skewed. That’s the qualitative idea behind skewness. The actual formula for the skewness of a data set is as follows \\[ \\mbox{skewness}(X) = \\frac{1}{N \\hat{\\sigma}^3} \\sum_{i=1}^N (X_i - \\bar{X})^3 \\] where \\(N\\) is the number of observations, \\(\\bar{X}\\) is the sample mean, and \\(\\hat{\\sigma}\\) is the standard deviation (the “divide by \\(N-1\\)” version, that is). Perhaps more helpfully, it might be useful to point out that the `psych` package contains a `skew()` function that you can use to calculate skewness. So if we wanted to use this function to calculate the skewness of the `afl.margins` data, we’d first need to load the package `library( psych )` \nwhich now makes it possible to use the following command:\n\n```\nskew( x = afl.margins )\n```\n\n `## [1] 0.7671555` \nNot surprisingly, it turns out that the AFL winning margins data is fairly skewed.\n\nThe final measure that is sometimes referred to, though very rarely in practice, is the kurtosis of a data set. Put simply, kurtosis is a measure of the “pointiness” of a data set, as illustrated in Figure 5.5.\n `## [1] -0.9516587` `## [1] 0.006731404` `## [1] 2.068523` \nBy convention, we say that the “normal curve” (black lines) has zero kurtosis, so the pointiness of a data set is assessed relative to this curve. In this Figure, the data on the left are not pointy enough, so the kurtosis is negative and we call the data platykurtic. The data on the right are too pointy, so the kurtosis is positive and we say that the data is leptokurtic. But the data in the middle are just pointy enough, so we say that it is mesokurtic and has kurtosis zero. This is summarised in the table below:\n\ninformal term | technical name | kurtosis value |\n | --- | --- | --- |\ntoo flat | platykurtic | negative |\njust pointy enough | mesokurtic | zero |\ntoo pointy | leptokurtic | positive |\n\n The equation for kurtosis is pretty similar in spirit to the formulas we’ve seen already for the variance and the skewness; except that where the variance involved squared deviations and the skewness involved cubed deviations, the kurtosis involves raising the deviations to the fourth power:76 \\[ \\mbox{kurtosis}(X) = \\frac{1}{N \\hat\\sigma^4} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right)^4 - 3 \\] I know, it’s not terribly interesting to me either. More to the point, the `psych` package has a function called `kurtosi()` that you can use to calculate the kurtosis of your data. For instance, if we were to do this for the AFL margins, \n\n `## [1] 0.02962633` \nwe discover that the AFL winning margins data are just pointy enough.\n\n## 5.4 Getting an overall summary of a variable\n\n Up to this point in the chapter I’ve explained several different summary statistics that are commonly used when analysing data, along with specific functions that you can use in R to calculate each one. However, it’s kind of annoying to have to separately calculate means, medians, standard deviations, skews etc. Wouldn’t it be nice if R had some helpful functions that would do all these tedious calculations at once? Something like `summary()` or `describe()` , perhaps? Why yes, yes it would. So much so that both of these functions exist. The `summary()` function is in the `base` package, so it comes with every installation of R. The `describe()` function is part of the `psych` package, which we loaded earlier in the chapter. \n\n### 5.4.1 “Summarising” a variable\n\n The `summary()` function is an easy thing to use, but a tricky thing to understand in full, since it’s a generic function (see Section 4.11. The basic idea behind the `summary()` function is that it prints out some useful information about whatever object (i.e., variable, as far as we’re concerned) you specify as the `object` argument. As a consequence, the behaviour of the `summary()` function differs quite dramatically depending on the class of the object that you give it. Let’s start by giving it a numeric object: \n\n```\nsummary( object = afl.margins )\n```\n\nFor numeric variables, we get a whole bunch of useful descriptive statistics. It gives us the minimum and maximum values (i.e., the range), the first and third quartiles (25th and 75th percentiles; i.e., the IQR), the mean and the median. In other words, it gives us a pretty good collection of descriptive statistics related to the central tendency and the spread of the data.\n Okay, what about if we feed it a logical vector instead? Let’s say I want to know something about how many “blowouts” there were in the 2010 AFL season. I operationalise the concept of a blowout (see Chapter 2) as a game in which the winning margin exceeds 50 points. Let’s create a logical variable `blowouts` in which the \\(i\\)-th element is `TRUE` if that game was a blowout according to my definition, \n\n```\nblowouts <- afl.margins > 50\nblowouts\n```\n\n```\n## [1] TRUE FALSE TRUE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE\n## [12] TRUE FALSE FALSE TRUE FALSE FALSE FALSE FALSE TRUE FALSE FALSE\n## [23] FALSE FALSE FALSE FALSE FALSE TRUE FALSE TRUE TRUE FALSE FALSE\n## [34] TRUE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE\n## [45] FALSE TRUE TRUE FALSE FALSE FALSE FALSE FALSE TRUE TRUE FALSE\n## [56] TRUE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE\n## [67] TRUE FALSE FALSE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE\n## [78] FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE\n## [89] FALSE FALSE TRUE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE\n## [100] FALSE TRUE FALSE FALSE FALSE TRUE FALSE TRUE TRUE TRUE FALSE\n## [111] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE FALSE\n## [122] TRUE FALSE FALSE FALSE FALSE FALSE FALSE FALSE TRUE TRUE FALSE\n## [133] FALSE TRUE TRUE FALSE FALSE FALSE FALSE FALSE TRUE FALSE TRUE\n## [144] TRUE TRUE TRUE FALSE FALSE FALSE TRUE FALSE FALSE TRUE FALSE\n## [155] TRUE FALSE TRUE FALSE FALSE TRUE FALSE FALSE TRUE FALSE FALSE\n## [166] FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE FALSE\n```\n\n So that’s what the `blowouts` variable looks like. Now let’s ask R for a `summary()` \n\n```\nsummary( object = blowouts )\n```\n\n```\n## Mode FALSE TRUE \n## logical 132 44\n```\n\n In this context, the `summary()` function gives us a count of the number of `TRUE` values, the number of `FALSE` values, and the number of missing values (i.e., the `NA` s). Pretty reasonable behaviour. Next, let’s try to give it a factor. If you recall, I’ve defined the `afl.finalists` vector as a factor, so let’s use that: \n\n```\nsummary( object = afl.finalists )\n```\n\n For factors, we get a frequency table, just like we got when we used the `table()` function. Interestingly, however, if we convert this to a character vector using the `as.character()` function (see Section 7.10, we don’t get the same results: \n\n```\nf2 <- as.character( afl.finalists )\nsummary( object = f2 )\n```\n\n```\n## Length Class Mode \n## 400 character character\n```\n\n This is one of those situations I was referring to in Section 4.7, in which it is helpful to declare your nominal scale variable as a factor rather than a character vector. Because I’ve defined `afl.finalists` as a factor, R knows that it should treat it as a nominal scale variable, and so it gives you a much more detailed (and helpful) summary than it would have if I’d left it as a character vector. \n\n### 5.4.2 “Summarising” a data frame\n\n Okay what about data frames? When you pass a data frame to the `summary()` function, it produces a slightly condensed summary of each variable inside the data frame. To give you a sense of how this can be useful, let’s try this for a new data set, one that you’ve never seen before. The data is stored in the `clinicaltrial.Rdata` file, and we’ll use it a lot in Chapter 14 (you can find a complete description of the data at the start of that chapter). Let’s load it, and see what we’ve got: \n\n```\nload( \"./data/clinicaltrial.Rdata\" )\nwho(TRUE)\n```\n\n```\n## -- Name -- -- Class -- -- Size --\n## clin.trial data.frame 18 x 3 \n## $drug factor 18 \n## $therapy factor 18 \n## $mood.gain numeric 18\n```\n\n There’s a single data frame called `clin.trial` which contains three variables, `drug` , `therapy` and `mood.gain` . Presumably then, this data is from a clinical trial of some kind, in which people were administered different drugs; and the researchers looked to see what the drugs did to their mood. Let’s see if the `summary()` function sheds a little more light on this situation: \n\n```\nsummary( clin.trial )\n```\n\n```\n## drug therapy mood.gain \n## placebo :6 no.therapy:9 Min. :0.1000 \n## anxifree:6 CBT :9 1st Qu.:0.4250 \n## joyzepam:6 Median :0.8500 \n## Mean :0.8833 \n## 3rd Qu.:1.3000 \n## Max. :1.8000\n```\n\n Evidently there were three drugs: a placebo, something called “anxifree” and something called “joyzepam”; and there were 6 people administered each drug. There were 9 people treated using cognitive behavioural therapy (CBT) and 9 people who received no psychological treatment. And we can see from looking at the summary of the `mood.gain` variable that most people did show a mood gain (mean \\(=.88\\)), though without knowing what the scale is here it’s hard to say much more than that. Still, that’s not too bad. Overall, I feel that I learned something from that. \n\n### 5.4.3 “Describing” a data frame\n\n The `describe()` function (in the `psych` package) is a little different, and it’s really only intended to be useful when your data are interval or ratio scale. Unlike the `summary()` function, it calculates the same descriptive statistics for any type of variable you give it. By default, these are: \n\n* \n `var` . This is just an index: 1 for the first variable, 2 for the second variable, and so on. * \n `n` . This is the sample size: more precisely, it’s the number of non-missing values. * \n `mean` . This is the sample mean (Section 5.1.1). * \n `sd` . This is the (bias corrected) standard deviation (Section 5.2.5). * \n `median` . The median (Section 5.1.3). * \n `trimmed` . This is trimmed mean. By default it’s the 10% trimmed mean (Section 5.1.6). * \n `mad` . The median absolute deviation (Section 5.2.6). * \n `min` . The minimum value. * \n `max` . The maximum value. * \n `range` . The range spanned by the data (Section 5.2.1). * \n `skew` . The skewness (Section 5.3). * \n `kurtosis` . The kurtosis (Section 5.3). * \n `se` . The standard error of the mean (Chapter 10). Notice that these descriptive statistics generally only make sense for data that are interval or ratio scale (usually encoded as numeric vectors). For nominal or ordinal variables (usually encoded as factors), most of these descriptive statistics are not all that useful. What the `describe()` function does is convert factors and logical variables to numeric vectors in order to do the calculations. These variables are marked with `*` and most of the time, the descriptive statistics for those variables won’t make much sense. If you try to feed it a data frame that includes a character vector as a variable, it produces an error. With those caveats in mind, let’s use the `describe()` function to have a look at the `clin.trial` data frame. Here’s what we get: \n\n```\ndescribe( x = clin.trial )\n```\n\n```\n## vars n mean sd median trimmed mad min max range skew\n## drug* 1 18 2.00 0.84 2.00 2.00 1.48 1.0 3.0 2.0 0.00\n## therapy* 2 18 1.50 0.51 1.50 1.50 0.74 1.0 2.0 1.0 0.00\n## mood.gain 3 18 0.88 0.53 0.85 0.88 0.67 0.1 1.8 1.7 0.13\n## kurtosis se\n## drug* -1.66 0.20\n## therapy* -2.11 0.12\n## mood.gain -1.44 0.13\n```\n\n As you can see, the output for the asterisked variables is pretty meaningless, and should be ignored. However, for the `mood.gain` variable, there’s a lot of useful information. \n\n## 5.5 Descriptive statistics separately for each group\n\n It is very commonly the case that you find yourself needing to look at descriptive statistics, broken down by some grouping variable. This is pretty easy to do in R, and there are three functions in particular that are worth knowing about: `by()` , `describeBy()` and `aggregate()` . Let’s start with the `describeBy()` function, which is part of the `psych` package. The `describeBy()` function is very similar to the `describe()` function, except that it has an additional argument called `group` which specifies a grouping variable. For instance, let’s say, I want to look at the descriptive statistics for the `clin.trial` data, broken down separately by `therapy` type. The command I would use here is: \n\n```\ndescribeBy( x=clin.trial, group=clin.trial$therapy )\n```\n\n As you can see, the output is essentially identical to the output that the `describe()` function produce, except that the output now gives you means, standard deviations etc separately for the `CBT` group and the `no.therapy` group. Notice that, as before, the output displays asterisks for factor variables, in order to draw your attention to the fact that the descriptive statistics that it has calculated won’t be very meaningful for those variables. Nevertheless, this command has given us some really useful descriptive statistics `mood.gain` variable, broken down as a function of `therapy` . A somewhat more general solution is offered by the `by()` function. There are three arguments that you need to specify when using this function: the `data` argument specifies the data set, the `INDICES` argument specifies the grouping variable, and the `FUN` argument specifies the name of a function that you want to apply separately to each group. To give a sense of how powerful this is, you can reproduce the `describeBy()` function by using a command like this: \n\n```\nby( data=clin.trial, INDICES=clin.trial$therapy, FUN=describe )\n```\n\n This will produce the exact same output as the command shown earlier. However, there’s nothing special about the `describe()` function. You could just as easily use the `by()` function in conjunction with the `summary()` function. For example: \n\n```\nby( data=clin.trial, INDICES=clin.trial$therapy, FUN=summary )\n```\n\n```\n## clin.trial$therapy: no.therapy\n## drug therapy mood.gain \n## placebo :3 no.therapy:9 Min. :0.1000 \n## anxifree:3 CBT :0 1st Qu.:0.3000 \n## joyzepam:3 Median :0.5000 \n## Mean :0.7222 \n## 3rd Qu.:1.3000 \n## Max. :1.7000 \n## -------------------------------------------------------- \n## clin.trial$therapy: CBT\n## drug therapy mood.gain \n## placebo :3 no.therapy:0 Min. :0.300 \n## anxifree:3 CBT :9 1st Qu.:0.800 \n## joyzepam:3 Median :1.100 \n## Mean :1.044 \n## 3rd Qu.:1.300 \n## Max. :1.800\n```\n\n Again, this output is pretty easy to interpret. It’s the output of the `summary()` function, applied separately to `CBT` group and the `no.therapy` group. For the two factors ( `drug` and `therapy` ) it prints out a frequency table, whereas for the numeric variable ( `mood.gain` ) it prints out the range, interquartile range, mean and median. What if you have multiple grouping variables? Suppose, for example, you would like to look at the average mood gain separately for all possible combinations of drug and therapy. It is actually possible to do this using the `by()` and `describeBy()` functions, but I usually find it more convenient to use the `aggregate()` function in this situation. There are again three arguments that you need to specify. The `formula` argument is used to indicate which variable you want to analyse, and which variables are used to specify the groups. For instance, if you want to look at `mood.gain` separately for each possible combination of `drug` and `therapy` , the formula you want is \n\n. The `data` argument is used to specify the data frame containing all the data, and the `FUN` argument is used to indicate what function you want to calculate for each group (e.g., the `mean` ). So, to obtain group means, use this command: \n\n```\naggregate( formula = mood.gain ~ drug + therapy, # mood.gain by drug/therapy combination\n data = clin.trial, # data is in the clin.trial data frame\n FUN = mean # print out group means\n )\n```\n\nor, alternatively, if you want to calculate the standard deviations for each group, you would use the following command (argument names omitted this time):\n\n```\n## drug therapy mood.gain\n## 1 placebo no.therapy 0.2000000\n## 2 anxifree no.therapy 0.2000000\n## 3 joyzepam no.therapy 0.2081666\n## 4 placebo CBT 0.3000000\n## 5 anxifree CBT 0.2081666\n## 6 joyzepam CBT 0.2645751\n```\n\n## 5.6 Standard scores\n\nSuppose my friend is putting together a new questionnaire intended to measure “grumpiness”. The survey has 50 questions, which you can answer in a grumpy way or not. Across a big sample (hypothetically, let’s imagine a million people or so!) the data are fairly normally distributed, with the mean grumpiness score being 17 out of 50 questions answered in a grumpy way, and the standard deviation is 5. In contrast, when I take the questionnaire, I answer 35 out of 50 questions in a grumpy way. So, how grumpy am I? One way to think about would be to say that I have grumpiness of 35/50, so you might say that I’m 70% grumpy. But that’s a bit weird, when you think about it. If my friend had phrased her questions a bit differently, people might have answered them in a different way, so the overall distribution of answers could easily move up or down depending on the precise way in which the questions were asked. So, I’m only 70% grumpy with respect to this set of survey questions. Even if it’s a very good questionnaire, this isn’t very a informative statement.\n\nA simpler way around this is to describe my grumpiness by comparing me to other people. Shockingly, out of my friend’s sample of 1,000,000 people, only 159 people were as grumpy as me (that’s not at all unrealistic, frankly), suggesting that I’m in the top 0.016% of people for grumpiness. This makes much more sense than trying to interpret the raw data. This idea – that we should describe my grumpiness in terms of the overall distribution of the grumpiness of humans – is the qualitative idea that standardisation attempts to get at. One way to do this is to do exactly what I just did, and describe everything in terms of percentiles. However, the problem with doing this is that “it’s lonely at the top”. Suppose that my friend had only collected a sample of 1000 people (still a pretty big sample for the purposes of testing a new questionnaire, I’d like to add), and this time gotten a mean of 16 out of 50 with a standard deviation of 5, let’s say. The problem is that almost certainly, not a single person in that sample would be as grumpy as me.\n However, all is not lost. A different approach is to convert my grumpiness score into a standard score, also referred to as a \\(z\\)-score. The standard score is defined as the number of standard deviations above the mean that my grumpiness score lies. To phrase it in “pseudo-maths” the standard score is calculated like this: \\[ \\mbox{standard score} = \\frac{\\mbox{raw score} - \\mbox{mean}}{\\mbox{standard deviation}} \\] In actual maths, the equation for the \\(z\\)-score is \\[ z_i = \\frac{X_i - \\bar{X}}{\\hat\\sigma} \\] So, going back to the grumpiness data, we can now transform Dan’s raw grumpiness into a standardised grumpiness score.77 If the mean is 17 and the standard deviation is 5 then my standardised grumpiness score would be78 \\[ z = \\frac{35 - 17}{5} = 3.6 \\] To interpret this value, recall the rough heuristic that I provided in Section 5.2.5, in which I noted that 99.7% of values are expected to lie within 3 standard deviations of the mean. So the fact that my grumpiness corresponds to a \\(z\\) score of 3.6 indicates that I’m very grumpy indeed. Later on, in Section 9.5, I’ll introduce a function called `pnorm()` that allows us to be a bit more precise than this. Specifically, it allows us to calculate a theoretical percentile rank for my grumpiness, as follows: `pnorm( 3.6 )` `## [1] 0.9998409` \nAt this stage, this command doesn’t make too much sense, but don’t worry too much about it. It’s not important for now. But the output is fairly straightforward: it suggests that I’m grumpier than 99.98% of people. Sounds about right.\n\nIn addition to allowing you to interpret a raw score in relation to a larger population (and thereby allowing you to make sense of variables that lie on arbitrary scales), standard scores serve a second useful function. Standard scores can be compared to one another in situations where the raw scores can’t. Suppose, for instance, my friend also had another questionnaire that measured extraversion using a 24 items questionnaire. The overall mean for this measure turns out to be 13 with standard deviation 4; and I scored a 2. As you can imagine, it doesn’t make a lot of sense to try to compare my raw score of 2 on the extraversion questionnaire to my raw score of 35 on the grumpiness questionnaire. The raw scores for the two variables are “about” fundamentally different things, so this would be like comparing apples to oranges.\n\nWhat about the standard scores? Well, this is a little different. If we calculate the standard scores, we get \\(z = (35-17)/5 = 3.6\\) for grumpiness and \\(z = (2-13)/4 = -2.75\\) for extraversion. These two numbers can be compared to each other.79 I’m much less extraverted than most people (\\(z = -2.75\\)) and much grumpier than most people (\\(z = 3.6\\)): but the extent of my unusualness is much more extreme for grumpiness (since 3.6 is a bigger number than 2.75). Because each standardised score is a statement about where an observation falls relative to its own population, it is possible to compare standardised scores across completely different variables.\n\n## 5.7 Correlations\n\nUp to this point we have focused entirely on how to construct descriptive statistics for a single variable. What we haven’t done is talked about how to describe the relationships between variables in the data. To do that, we want to talk mostly about the correlation between variables. But first, we need some data.\n\n### 5.7.1 The data\n\n After spending so much time looking at the AFL data, I’m starting to get bored with sports. Instead, let’s turn to a topic close to every parent’s heart: sleep. The following data set is fictitious, but based on real events. Suppose I’m curious to find out how much my infant son’s sleeping habits affect my mood. Let’s say that I can rate my grumpiness very precisely, on a scale from 0 (not at all grumpy) to 100 (grumpy as a very, very grumpy old man). And, lets also assume that I’ve been measuring my grumpiness, my sleeping patterns and my son’s sleeping patterns for quite some time now. Let’s say, for 100 days. And, being a nerd, I’ve saved the data as a file called `parenthood.Rdata` . If we load the data… \n\n```\nload( \"./data/parenthood.Rdata\" )\nwho(TRUE)\n```\n\n```\n## -- Name -- -- Class -- -- Size --\n## parenthood data.frame 100 x 4 \n## $dan.sleep numeric 100 \n## $baby.sleep numeric 100 \n## $dan.grump numeric 100 \n## $day integer 100\n```\n\n … we see that the file contains a single data frame called `parenthood` , which contains four variables `dan.sleep` , `baby.sleep` , `dan.grump` and `day` . If we peek at the data using `head()` out the data, here’s what we get: `head(parenthood,10)` \n\n```\n## dan.sleep baby.sleep dan.grump day\n## 1 7.59 10.18 56 1\n## 2 7.91 11.66 60 2\n## 3 5.14 7.92 82 3\n## 4 7.71 9.61 55 4\n## 5 6.68 9.75 67 5\n## 6 5.99 5.04 72 6\n## 7 8.19 10.45 53 7\n## 8 7.19 8.27 60 8\n## 9 7.40 6.06 60 9\n## 10 6.58 7.09 71 10\n```\n\nNext, I’ll calculate some basic descriptive statistics:\n\n```\ndescribe( parenthood )\n```\n\n```\n## vars n mean sd median trimmed mad min max range\n## dan.sleep 1 100 6.97 1.02 7.03 7.00 1.09 4.84 9.00 4.16\n## baby.sleep 2 100 8.05 2.07 7.95 8.05 2.33 3.25 12.07 8.82\n## dan.grump 3 100 63.71 10.05 62.00 63.16 9.64 41.00 91.00 50.00\n## day 4 100 50.50 29.01 50.50 50.50 37.06 1.00 100.00 99.00\n## skew kurtosis se\n## dan.sleep -0.29 -0.72 0.10\n## baby.sleep -0.02 -0.69 0.21\n## dan.grump 0.43 -0.16 1.00\n## day 0.00 -1.24 2.90\n```\n\nFinally, to give a graphical depiction of what each of the three interesting variables looks like, Figure 5.6 plots histograms.\n\nOne thing to note: just because R can calculate dozens of different statistics doesn’t mean you should report all of them. If I were writing this up for a report, I’d probably pick out those statistics that are of most interest to me (and to my readership), and then put them into a nice, simple table like the one in Table ??.80 Notice that when I put it into a table, I gave everything “human readable” names. This is always good practice. Notice also that I’m not getting enough sleep. This isn’t good practice, but other parents tell me that it’s standard practice.\n\nvariable | min | max | mean | median | std. dev | IQR |\n | --- | --- | --- | --- | --- | --- | --- |\nDan’s grumpiness | 41 | 91 | 63.71 | 62 | 10.05 | 14 |\nDan’s hours slept | 4.84 | 9 | 6.97 | 7.03 | 1.02 | 1.45 |\nDan’s son’s hours slept | 3.25 | 12.07 | 8.05 | 7.95 | 2.07 | 3.21 |\n\n### 5.7.2 The strength and direction of a relationship\n\n We can draw scatterplots to give us a general sense of how closely related two variables are. Ideally though, we might want to say a bit more about it than that. For instance, let’s compare the relationship between `dan.sleep` and `dan.grump` (Figure 5.7 with that between `baby.sleep` and `dan.grump` (Figure 5.8. When looking at these two plots side by side, it’s clear that the relationship is qualitatively the same in both cases: more sleep equals less grump! However, it’s also pretty obvious that the relationship between `dan.sleep` and `dan.grump` is stronger than the relationship between `baby.sleep` and `dan.grump` . The plot on the left is “neater” than the one on the right. What it feels like is that if you want to predict what my mood is, it’d help you a little bit to know how many hours my son slept, but it’d be more helpful to know how many hours I slept. In contrast, let’s consider Figure 5.8 vs. Figure 5.9. If we compare the scatterplot of “ `baby.sleep` v `dan.grump` ” to the scatterplot of “` `baby.sleep` v `dan.sleep` ”, the overall strength of the relationship is the same, but the direction is different. That is, if my son sleeps more, I get more sleep (positive relationship, but if he sleeps more then I get less grumpy (negative relationship). \n\n### 5.7.3 The correlation coefficient\n\nWe can make these ideas a bit more explicit by introducing the idea of a correlation coefficient (or, more specifically, Pearson’s correlation coefficient), which is traditionally denoted by \\(r\\). The correlation coefficient between two variables \\(X\\) and \\(Y\\) (sometimes denoted \\(r_{XY}\\)), which we’ll define more precisely in the next section, is a measure that varies from \\(-1\\) to \\(1\\). When \\(r = -1\\) it means that we have a perfect negative relationship, and when \\(r = 1\\) it means we have a perfect positive relationship. When \\(r = 0\\), there’s no relationship at all. If you look at Figure 5.10, you can see several plots showing what different correlations look like.\n The formula for the Pearson’s correlation coefficient can be written in several different ways. I think the simplest way to write down the formula is to break it into two steps. Firstly, let’s introduce the idea of a covariance. The covariance between two variables \\(X\\) and \\(Y\\) is a generalisation of the notion of the variance; it’s a mathematically simple way of describing the relationship between two variables that isn’t terribly informative to humans: \\[ \\mbox{Cov}(X,Y) = \\frac{1}{N-1} \\sum_{i=1}^N \\left( X_i - \\bar{X} \\right) \\left( Y_i - \\bar{Y} \\right) \\] Because we’re multiplying (i.e., taking the “product” of) a quantity that depends on \\(X\\) by a quantity that depends on \\(Y\\) and then averaging81, you can think of the formula for the covariance as an “average cross product” between \\(X\\) and \\(Y\\). The covariance has the nice property that, if \\(X\\) and \\(Y\\) are entirely unrelated, then the covariance is exactly zero. If the relationship between them is positive (in the sense shown in Figure@reffig:corr) then the covariance is also positive; and if the relationship is negative then the covariance is also negative. In other words, the covariance captures the basic qualitative idea of correlation. Unfortunately, the raw magnitude of the covariance isn’t easy to interpret: it depends on the units in which \\(X\\) and \\(Y\\) are expressed, and worse yet, the actual units that the covariance itself is expressed in are really weird. For instance, if \\(X\\) refers to the `dan.sleep` variable (units: hours) and \\(Y\\) refers to the `dan.grump` variable (units: grumps), then the units for their covariance are “hours \\(\\times\\) grumps”. And I have no freaking idea what that would even mean. \nThe Pearson correlation coefficient \\(r\\) fixes this interpretation problem by standardising the covariance, in pretty much the exact same way that the \\(z\\)-score standardises a raw score: by dividing by the standard deviation. However, because we have two variables that contribute to the covariance, the standardisation only works if we divide by both standard deviations.82 In other words, the correlation between \\(X\\) and \\(Y\\) can be written as follows: \\[ r_{XY} = \\frac{\\mbox{Cov}(X,Y)}{ \\hat{\\sigma}_X \\ \\hat{\\sigma}_Y} \\] By doing this standardisation, not only do we keep all of the nice properties of the covariance discussed earlier, but the actual values of \\(r\\) are on a meaningful scale: \\(r= 1\\) implies a perfect positive relationship, and \\(r = -1\\) implies a perfect negative relationship. I’ll expand a little more on this point later, in Section@refsec:interpretingcorrelations. But before I do, let’s look at how to calculate correlations in R.\n\n### 5.7.4 Calculating correlations in R\n\n Calculating correlations in R can be done using the `cor()` command. The simplest way to use the command is to specify two input arguments `x` and `y` , each one corresponding to one of the variables. The following extract illustrates the basic usage of the function:83 \n\n `## [1] -0.903384` However, the `cor()` function is a bit more powerful than this simple example suggests. For example, you can also calculate a complete “correlation matrix”, between all pairs of variables in the data frame:84 \n\n```\n# correlate all pairs of variables in \"parenthood\":\ncor( x = parenthood )\n```\n\n### 5.7.5 Interpreting a correlation\n\nNaturally, in real life you don’t see many correlations of 1. So how should you interpret a correlation of, say \\(r= .4\\)? The honest answer is that it really depends on what you want to use the data for, and on how strong the correlations in your field tend to be. A friend of mine in engineering once argued that any correlation less than \\(.95\\) is completely useless (I think he was exaggerating, even for engineering). On the other hand there are real cases – even in psychology – where you should really expect correlations that strong. For instance, one of the benchmark data sets used to test theories of how people judge similarities is so clean that any theory that can’t achieve a correlation of at least \\(.9\\) really isn’t deemed to be successful. However, when looking for (say) elementary correlates of intelligence (e.g., inspection time, response time), if you get a correlation above \\(.3\\) you’re doing very very well. In short, the interpretation of a correlation depends a lot on the context. That said, the rough guide in Table ?? is pretty typical.\n\n```\nknitr::kable(\nrbind(\nc(\"-1.0 to -0.9\" ,\"Very strong\", \"Negative\"),\nc(\"-0.9 to -0.7\", \"Strong\", \"Negative\") ,\nc(\"-0.7 to -0.4\", \"Moderate\", \"Negative\") ,\nc(\"-0.4 to -0.2\", \"Weak\", \"Negative\"),\nc(\"-0.2 to 0\",\"Negligible\", \"Negative\") ,\nc(\"0 to 0.2\",\"Negligible\", \"Positive\"),\nc(\"0.2 to 0.4\", \"Weak\", \"Positive\"), \nc(\"0.4 to 0.7\", \"Moderate\", \"Positive\"), \nc(\"0.7 to 0.9\", \"Strong\", \"Positive\"), \nc(\"0.9 to 1.0\", \"Very strong\", \"Positive\")), col.names=c(\"Correlation\", \"Strength\", \"Direction\"),\n booktabs = TRUE)\n```\n\nCorrelation | Strength | Direction |\n | --- | --- | --- |\n-1.0 to -0.9 | Very strong | Negative |\n-0.9 to -0.7 | Strong | Negative |\n-0.7 to -0.4 | Moderate | Negative |\n-0.4 to -0.2 | Weak | Negative |\n-0.2 to 0 | Negligible | Negative |\n0 to 0.2 | Negligible | Positive |\n0.2 to 0.4 | Weak | Positive |\n0.4 to 0.7 | Moderate | Positive |\n0.7 to 0.9 | Strong | Positive |\n0.9 to 1.0 | Very strong | Positive |\n\nHowever, something that can never be stressed enough is that you should always look at the scatterplot before attaching any interpretation to the data. A correlation might not mean what you think it means. The classic illustration of this is “Anscombe’s Quartet” (??? Anscombe1973), which is a collection of four data sets. Each data set has two variables, an \\(X\\) and a \\(Y\\). For all four data sets the mean value for \\(X\\) is 9 and the mean for \\(Y\\) is 7.5. The, standard deviations for all \\(X\\) variables are almost identical, as are those for the the \\(Y\\) variables. And in each case the correlation between \\(X\\) and \\(Y\\) is \\(r = 0.816\\). You can verify this yourself, since the dataset comes distributed with R. The commands would be:\n\n```\ncor( anscombe$x1, anscombe$y1 )\n```\n\n `## [1] 0.8164205` \n\n```\ncor( anscombe$x2, anscombe$y2 )\n```\n\n `## [1] 0.8162365` \nand so on.\nYou’d think that these four data setswould look pretty similar to one another. They do not. If we draw scatterplots of \\(X\\) against \\(Y\\) for all four variables, as shown in Figure 5.11 we see that all four of these are spectacularly different to each other. \nThe lesson here, which so very many people seem to forget in real life is “always graph your raw data”. This will be the focus of Chapter 6.\n\n### 5.7.6 Spearman’s rank correlations\n\nThe Pearson correlation coefficient is useful for a lot of things, but it does have shortcomings. One issue in particular stands out: what it actually measures is the strength of the linear relationship between two variables. In other words, what it gives you is a measure of the extent to which the data all tend to fall on a single, perfectly straight line. Often, this is a pretty good approximation to what we mean when we say “relationship”, and so the Pearson correlation is a good thing to calculation. Sometimes, it isn’t.\n\nOne very common situation where the Pearson correlation isn’t quite the right thing to use arises when an increase in one variable \\(X\\) really is reflected in an increase in another variable \\(Y\\), but the nature of the relationship isn’t necessarily linear. An example of this might be the relationship between effort and reward when studying for an exam. If you put in zero effort (\\(X\\)) into learning a subject, then you should expect a grade of 0% (\\(Y\\)). However, a little bit of effort will cause a massive improvement: just turning up to lectures means that you learn a fair bit, and if you just turn up to classes, and scribble a few things down so your grade might rise to 35%, all without a lot of effort. However, you just don’t get the same effect at the other end of the scale. As everyone knows, it takes a lot more effort to get a grade of 90% than it takes to get a grade of 55%. What this means is that, if I’ve got data looking at study effort and grades, there’s a pretty good chance that Pearson correlations will be misleading.\n To illustrate, consider the data plotted in Figure 5.12, showing the relationship between hours worked and grade received for 10 students taking some class. The curious thing about this – highly fictitious – data set is that increasing your effort always increases your grade. It might be by a lot or it might be by a little, but increasing effort will never decrease your grade. The data are stored in `effort.Rdata` : \n\n```\n> load( \"effort.Rdata\" )\n> who(TRUE)\n -- Name -- -- Class -- -- Size --\n effort data.frame 10 x 2 \n $hours numeric 10 \n $grade numeric 10\n```\n\nThe raw data look like this:\n\n```\n> effort\n hours grade\n1 2 13\n2 76 91\n3 40 79\n4 6 14\n5 16 21\n6 28 74\n7 27 47\n8 59 85\n9 46 84\n10 68 88\n```\n\nIf we run a standard Pearson correlation, it shows a strong relationship between hours worked and grade received,\n\n```\n> cor( effort$hours, effort$grade )\n[1] 0.909402\n```\n\nbut this doesn’t actually capture the observation that increasing hours worked always increases the grade. There’s a sense here in which we want to be able to say that the correlation is perfect but for a somewhat different notion of what a “relationship” is. What we’re looking for is something that captures the fact that there is a perfect ordinal relationship here. That is, if student 1 works more hours than student 2, then we can guarantee that student 1 will get the better grade. That’s not what a correlation of \\(r = .91\\) says at all.\n\nHow should we address this? Actually, it’s really easy: if we’re looking for ordinal relationships, all we have to do is treat the data as if it were ordinal scale! So, instead of measuring effort in terms of “hours worked”, lets rank all 10 of our students in order of hours worked. That is, student 1 did the least work out of anyone (2 hours) so they get the lowest rank (rank = 1). Student 4 was the next laziest, putting in only 6 hours of work in over the whole semester, so they get the next lowest rank (rank = 2). Notice that I’m using “rank =1” to mean “low rank”. Sometimes in everyday language we talk about “rank = 1” to mean “top rank” rather than “bottom rank”. So be careful: you can rank “from smallest value to largest value” (i.e., small equals rank 1) or you can rank “from largest value to smallest value” (i.e., large equals rank 1). In this case, I’m ranking from smallest to largest, because that’s the default way that R does it. But in real life, it’s really easy to forget which way you set things up, so you have to put a bit of effort into remembering!\n\nOkay, so let’s have a look at our students when we rank them from worst to best in terms of effort and reward:\n\nrank (hours worked) | rank (grade received) |\n | --- | --- |\nstudent | 1 | 1 |\nstudent | 2 | 10 |\nstudent | 3 | 6 |\nstudent | 4 | 2 |\nstudent | 5 | 3 |\nstudent | 6 | 5 |\nstudent | 7 | 4 |\nstudent | 8 | 8 |\nstudent | 9 | 7 |\nstudent | 10 | 9 |\n\n Hm. These are identical. The student who put in the most effort got the best grade, the student with the least effort got the worst grade, etc. We can get R to construct these rankings using the `rank()` function, like this: \n\n```\n> hours.rank <- rank( effort$hours ) # rank students by hours worked\n> grade.rank <- rank( effort$grade ) # rank students by grade received\n```\n\nAs the table above shows, these two rankings are identical, so if we now correlate them we get a perfect relationship:\n\n```\n> cor( hours.rank, grade.rank )\n[1] 1\n```\n\n What we’ve just re-invented is Spearman’s rank order correlation, usually denoted \\(\\rho\\) to distinguish it from the Pearson correlation \\(r\\). We can calculate Spearman’s \\(\\rho\\) using R in two different ways. Firstly we could do it the way I just showed, using the `rank()` function to construct the rankings, and then calculate the Pearson correlation on these ranks. However, that’s way too much effort to do every time. It’s much easier to just specify the `method` argument of the `cor()` function. \n\n```\n> cor( effort$hours, effort$grade, method = \"spearman\")\n[1] 1\n```\n\n The default value of the `method` argument is `\"pearson\"` , which is why we didn’t have to specify it earlier on when we were doing Pearson correlations. \n\n### 5.7.7 The\n\n `correlate()` function As we’ve seen, the `cor()` function works pretty well, and handles many of the situations that you might be interested in. One thing that many beginners find frustrating, however, is the fact that it’s not built to handle non-numeric variables. From a statistical perspective, this is perfectly sensible: Pearson and Spearman correlations are only designed to work for numeric variables, so the `cor()` function spits out an error. Here’s what I mean. Suppose you were keeping track of how many `hours` you worked in any given day, and counted how many `tasks` you completed. If you were doing the tasks for money, you might also want to keep track of how much `pay` you got for each job. It would also be sensible to keep track of the `weekday` on which you actually did the work: most of us don’t work as much on Saturdays or Sundays. If you did this for 7 weeks, you might end up with a data set that looks like this one: \n\n```\n> load(\"work.Rdata\")\n\n> who(TRUE)\n -- Name -- -- Class -- -- Size --\n work data.frame 49 x 7 \n $hours numeric 49 \n $tasks numeric 49 \n $pay numeric 49 \n $day integer 49 \n $weekday factor 49 \n $week numeric 49 \n $day.type factor 49 \n\n> head(work)\n hours tasks pay day weekday week day.type\n1 7.2 14 41 1 Tuesday 1 weekday\n2 7.4 11 39 2 Wednesday 1 weekday\n3 6.6 14 13 3 Thursday 1 weekday\n4 6.5 22 47 4 Friday 1 weekday\n5 3.1 5 4 5 Saturday 1 weekend\n6 3.0 7 12 6 Sunday 1 weekend\n```\n\n Obviously, I’d like to know something about how all these variables correlate with one another. I could correlate `hours` with `pay` quite using `cor()` , like so: \n\n```\n> cor(work$hours,work$pay)\n[1] 0.7604283\n```\n\n But what if I wanted a quick and easy way to calculate all pairwise correlations between the numeric variables? I can’t just input the `work` data frame, because it contains two factor variables, `weekday` and `day.type` . If I try this, I get an error: \n\n```\n> cor(work)\nError in cor(work) : 'x' must be numeric\n```\n\n It order to get the correlations that I want using the `cor()` function, is create a new data frame that doesn’t contain the factor variables, and then feed that new data frame into the `cor()` function. It’s not actually very hard to do that, and I’ll talk about how to do it properly in Section@refsec:subsetdataframe. But it would be nice to have some function that is smart enough to just ignore the factor variables. That’s where the `correlate()` function in the `lsr` package can be handy. If you feed it a data frame that contains factors, it knows to ignore them, and returns the pairwise correlations only between the numeric variables: \n\n```\n> correlate(work)\n\n hours tasks pay day weekday week day.type\nhours . 0.800 0.760 -0.049 . 0.018 .\ntasks 0.800 . 0.720 -0.072 . -0.013 .\npay 0.760 0.720 . 0.137 . 0.196 .\nday -0.049 -0.072 0.137 . . 0.990 .\nweekday . . . . . . .\nweek 0.018 -0.013 0.196 0.990 . . .\nday.type . . . . . . .\n```\n\n The output here shows a `.` whenever one of the variables is non-numeric. It also shows a `.` whenever a variable is correlated with itself (it’s not a meaningful thing to do). The `correlate()` function can also do Spearman correlations, by specifying the `corr.method` to use: \n\n```\n> correlate( work, corr.method=\"spearman\" )\n\n hours tasks pay day weekday week day.type\nhours . 0.805 0.745 -0.047 . 0.010 .\ntasks 0.805 . 0.730 -0.068 . -0.008 .\npay 0.745 0.730 . 0.094 . 0.154 .\nday -0.047 -0.068 0.094 . . 0.990 .\nweekday . . . . . . .\nweek 0.010 -0.008 0.154 0.990 . . .\nday.type . . . . . . .\n```\n\n Obviously, there’s no new functionality in the `correlate()` function, and any advanced R user would be perfectly capable of using the `cor()` function to get these numbers out. But if you’re not yet comfortable with extracting a subset of a data frame, the `correlate()` function is for you. \n\n## 5.8 Handling missing values\n\nThere’s one last topic that I want to discuss briefly in this chapter, and that’s the issue of missing data. Real data sets very frequently turn out to have missing values: perhaps someone forgot to fill in a particular survey question, for instance. Missing data can be the source of a lot of tricky issues, most of which I’m going to gloss over. However, at a minimum, you need to understand the basics of handling missing data in R.\n\n### 5.8.1 The single variable case\n\n Let’s start with the simplest case, in which you’re trying to calculate descriptive statistics for a single variable which has missing data. In R, this means that there will be `NA` values in your data vector. Let’s create a variable like that: \n\n```\n> partial <- c(10, 20, NA, 30)\n```\n\n Let’s assume that you want to calculate the mean of this variable. By default, R assumes that you want to calculate the mean using all four elements of this vector, which is probably the safest thing for a dumb automaton to do, but it’s rarely what you actually want. Why not? Well, remember that the basic interpretation of `NA` is “I don’t know what this number is”. This means that `1 + NA = NA` : if I add 1 to some number that I don’t know (i.e., the `NA` ) then the answer is also a number that I don’t know. As a consequence, if you don’t explicitly tell R to ignore the `NA` values, and the data set does have missing values, then the output will itself be a missing value. If I try to calculate the mean of the `partial` vector, without doing anything about the missing value, here’s what happens: \n\n```\n> mean( x = partial )\n[1] NA\n```\n\nTechnically correct, but deeply unhelpful.\n To fix this, all of the descriptive statistics functions that I’ve discussed in this chapter (with the exception of `cor()` which is a special case I’ll discuss below) have an optional argument called `na.rm` , which is shorthand for “remove NA values”. By default, `na.rm = FALSE` , so R does nothing about the missing data problem. Let’s try setting `na.rm = TRUE` and see what happens: When calculating sums and means when missing data are present (i.e., when there are `NA` values) there’s actually an additional argument to the function that you should be aware of. This argument is called `na.rm` , and is a logical value indicating whether R should ignore (or “remove”) the missing data for the purposes of doing the calculations. By default, R assumes that you want to keep the missing values, so unless you say otherwise it will set `na.rm = FALSE` . However, R assumes that `1 + NA = NA` : if I add 1 to some number that I don’t know (i.e., the `NA` ) then the answer is also a number that I don’t know. As a consequence, if you don’t explicitly tell R to ignore the `NA` values, and the data set does have missing values, then the output will itself be a missing value. This is illustrated in the following extract: \n\n```\n> mean( x = partial, na.rm = TRUE )\n[1] 20\n```\n\n Notice that the mean is `20` (i.e., `60 / 3` ) and not `15` . When R ignores a `NA` value, it genuinely ignores it. In effect, the calculation above is identical to what you’d get if you asked for the mean of the three-element vector `c(10, 20, 30)` . As indicated above, this isn’t unique to the `mean()` function. Pretty much all of the other functions that I’ve talked about in this chapter have an `na.rm` argument that indicates whether it should ignore missing values. However, its behaviour is the same for all these functions, so I won’t waste everyone’s time by demonstrating it separately for each one. \n\n### 5.8.2 Missing values in pairwise calculations\n\n I mentioned earlier that the `cor()` function is a special case. It doesn’t have an `na.rm` argument, because the story becomes a lot more complicated when more than one variable is involved. What it does have is an argument called `use` which does roughly the same thing, but you need to think little more carefully about what you want this time. To illustrate the issues, let’s open up a data set that has missing values, `parenthood2.Rdata` . This file contains the same data as the original parenthood data, but with some values deleted. It contains a single data frame, `parenthood2` : \n\n```\n> load( \"parenthood2.Rdata\" )\n> print( parenthood2 )\n dan.sleep baby.sleep dan.grump day\n1 7.59 NA 56 1\n2 7.91 11.66 60 2\n3 5.14 7.92 82 3\n4 7.71 9.61 55 4\n5 6.68 9.75 NA 5\n6 5.99 5.04 72 6\nBLAH BLAH BLAH\n```\n\n If I calculate my descriptive statistics using the `describe()` function \n\n```\n> describe( parenthood2 )\n var n mean sd median trimmed mad min max BLAH\ndan.sleep 1 91 6.98 1.02 7.03 7.02 1.13 4.84 9.00 BLAH\nbaby.sleep 2 89 8.11 2.05 8.20 8.13 2.28 3.25 12.07 BLAH\ndan.grump 3 92 63.15 9.85 61.00 62.66 10.38 41.00 89.00 BLAH\nday 4 100 50.50 29.01 50.50 50.50 37.06 1.00 100.00 BLAH\n```\n\n we can see from the `n` column that there are 9 missing values for `dan.sleep` , 11 missing values for `baby.sleep` and 8 missing values for `dan.grump` .85 Suppose what I would like is a correlation matrix. And let’s also suppose that I don’t bother to tell R how to handle those missing values. Here’s what happens: \n\n```\n> cor( parenthood2 )\n dan.sleep baby.sleep dan.grump day\ndan.sleep 1 NA NA NA\nbaby.sleep NA 1 NA NA\ndan.grump NA NA 1 NA\nday NA NA NA 1\n```\n\n Annoying, but it kind of makes sense. If I don’t know what some of the values of `dan.sleep` and `baby.sleep` actually are, then I can’t possibly know what the correlation between these two variables is either, since the formula for the correlation coefficient makes use of every single observation in the data set. Once again, it makes sense: it’s just not particularly helpful. To make R behave more sensibly in this situation, you need to specify the `use` argument to the `cor()` function. There are several different values that you can specify for this, but the two that we care most about in practice tend to be `\"complete.obs\"` and \n\n```\n\"pairwise.complete.obs\"\n```\n\n. If we specify `use = \"complete.obs\"` , R will completely ignore all cases (i.e., all rows in our `parenthood2` data frame) that have any missing values at all. So, for instance, if you look back at the extract earlier when I used the `head()` function, notice that observation 1 (i.e., day 1) of the `parenthood2` data set is missing the value for `baby.sleep` , but is otherwise complete? Well, if you choose `use = \"complete.obs\"` R will ignore that row completely: that is, even when it’s trying to calculate the correlation between `dan.sleep` and `dan.grump` , observation 1 will be ignored, because the value of `baby.sleep` is missing for that observation. Here’s what we get: \n\n```\n> cor(parenthood2, use = \"complete.obs\")\n dan.sleep baby.sleep dan.grump day\ndan.sleep 1.00000000 0.6394985 -0.89951468 0.06132891\nbaby.sleep 0.63949845 1.0000000 -0.58656066 0.14555814\ndan.grump -0.89951468 -0.5865607 1.00000000 -0.06816586\nday 0.06132891 0.1455581 -0.06816586 1.00000000\n```\n\n The other possibility that we care about, and the one that tends to get used more often in practice, is to set\n\n```\nuse = \"pairwise.complete.obs\"\n```\n\n. When we do that, R only looks at the variables that it’s trying to correlate when determining what to drop. So, for instance, since the only missing value for observation 1 of `parenthood2` is for `baby.sleep` R will only drop observation 1 when `baby.sleep` is one of the variables involved: and so R keeps observation 1 when trying to correlate `dan.sleep` and `dan.grump` . When we do it this way, here’s what we get: \n\n```\n> cor(parenthood2, use = \"pairwise.complete.obs\") \n dan.sleep baby.sleep dan.grump day\ndan.sleep 1.00000000 0.61472303 -0.903442442 -0.076796665\nbaby.sleep 0.61472303 1.00000000 -0.567802669 0.058309485\ndan.grump -0.90344244 -0.56780267 1.000000000 0.005833399\nday -0.07679667 0.05830949 0.005833399 1.000000000\n```\n\n Similar, but not quite the same. It’s also worth noting that the `correlate()` function (in the `lsr` package) automatically uses the “pairwise complete” method: \n\n```\n> correlate(parenthood2)\n\n dan.sleep baby.sleep dan.grump day\ndan.sleep . 0.615 -0.903 -0.077\nbaby.sleep 0.615 . -0.568 0.058\ndan.grump -0.903 -0.568 . 0.006\nday -0.077 0.058 0.006 .\n```\n\nThe two approaches have different strengths and weaknesses. The “pairwise complete” approach has the advantage that it keeps more observations, so you’re making use of more of your data and (as we’ll discuss in tedious detail in Chapter 10 and it improves the reliability of your estimated correlation. On the other hand, it means that every correlation in your correlation matrix is being computed from a slightly different set of observations, which can be awkward when you want to compare the different correlations that you’ve got.\n\nSo which method should you use? It depends a lot on why you think your values are missing, and probably depends a little on how paranoid you are. For instance, if you think that the missing values were “chosen” completely randomly86 then you’ll probably want to use the pairwise method. If you think that missing data are a cue to thinking that the whole observation might be rubbish (e.g., someone just selecting arbitrary responses in your questionnaire), but that there’s no pattern to which observations are “rubbish” then it’s probably safer to keep only those observations that are complete. If you think there’s something systematic going on, in that some observations are more likely to be missing than others, then you have a much trickier problem to solve, and one that is beyond the scope of this book.\n\n## 5.9 Summary\n\nCalculating some basic descriptive statistics is one of the very first things you do when analysing real data, and descriptive statistics are much simpler to understand than inferential statistics, so like every other statistics textbook I’ve started with descriptives. In this chapter, we talked about the following topics:\n\n* Measures of central tendency. Broadly speaking, central tendency measures tell you where the data are. There’s three measures that are typically reported in the literature: the mean, median and mode. (Section 5.1)\n * Measures of variability. In contrast, measures of variability tell you about how “spread out” the data are. The key measures are: range, standard deviation, interquartile reange (Section 5.2)\n * Getting summaries of variables in R. Since this book focuses on doing data analysis in R, we spent a bit of time talking about how descriptive statistics are computed in R. (Section 2.8 and 5.5)\n * Standard scores. The \\(z\\)-score is a slightly unusual beast. It’s not quite a descriptive statistic, and not quite an inference. We talked about it in Section 5.6. Make sure you understand that section: it’ll come up again later.\n * Correlations. Want to know how strong the relationship is between two variables? Calculate a correlation. (Section 5.7)\n * Missing data. Dealing with missing data is one of those frustrating things that data analysts really wish the didn’t have to think about. In real life it can be hard to do well. For the purpose of this book, we only touched on the basics in Section 5.8\n\nIn the next section we’ll move on to a discussion of how to draw pictures! Everyone loves a pretty picture, right? But before we do, I want to end on an important point. A traditional first course in statistics spends only a small proportion of the class on descriptive statistics, maybe one or two lectures at most. The vast majority of the lecturer’s time is spent on inferential statistics, because that’s where all the hard stuff is. That makes sense, but it hides the practical everyday importance of choosing good descriptives. With that in mind…\n\n## 5.10 Epilogue: Good descriptive statistics are descriptive!\n\nThe death of one man is a tragedy. The death of millions is a statistic.\n\n– , Potsdam 1945\n 950,000 – 1,200,000\n\n– Estimate of Soviet repression deaths, 1937-1938 (Ellman 2002)\n\nStalin’s infamous quote about the statistical character death of millions is worth giving some thought. The clear intent of his statement is that the death of an individual touches us personally and its force cannot be denied, but that the deaths of a multitude are incomprehensible, and as a consequence mere statistics, more easily ignored. I’d argue that Stalin was half right. A statistic is an abstraction, a description of events beyond our personal experience, and so hard to visualise. Few if any of us can imagine what the deaths of millions is “really” like, but we can imagine one death, and this gives the lone death its feeling of immediate tragedy, a feeling that is missing from Ellman’s cold statistical description.\n\nYet it is not so simple: without numbers, without counts, without a description of what happened, we have no chance of understanding what really happened, no opportunity event to try to summon the missing feeling. And in truth, as I write this, sitting in comfort on a Saturday morning, half a world and a whole lifetime away from the Gulags, when I put the Ellman estimate next to the Stalin quote a dull dread settles in my stomach and a chill settles over me. The Stalinist repression is something truly beyond my experience, but with a combination of statistical data and those recorded personal histories that have come down to us, it is not entirely beyond my comprehension. Because what Ellman’s numbers tell us is this: over a two year period, Stalinist repression wiped out the equivalent of every man, woman and child currently alive in the city where I live. Each one of those deaths had it’s own story, was it’s own tragedy, and only some of those are known to us now. Even so, with a few carefully chosen statistics, the scale of the atrocity starts to come into focus.\n\nThus it is no small thing to say that the first task of the statistician and the scientist is to summarise the data, to find some collection of numbers that can convey to an audience a sense of what has happened. This is the job of descriptive statistics, but it’s not a job that can be told solely using the numbers. You are a data analyst, not a statistical software package. Part of your job is to take these statistics and turn them into a description. When you analyse data, it is not sufficient to list off a collection of numbers. Always remember that what you’re really trying to do is communicate with a human audience. The numbers are important, but they need to be put together into a meaningful story that your audience can interpret. That means you need to think about framing. You need to think about context. And you need to think about the individual events that your statistics are summarising.\n\nNote for non-Australians: the AFL is an Australian rules football competition. You don’t need to know anything about Australian rules in order to follow this section.↩\n * \n\nThe choice to use \\(\\Sigma\\) to denote summation isn’t arbitrary: it’s the Greek upper case letter sigma, which is the analogue of the letter S in that alphabet. Similarly, there’s an equivalent symbol used to denote the multiplication of lots of numbers: because multiplications are also called “products”, we use the \\(\\Pi\\) symbol for this; the Greek upper case pi, which is the analogue of the letter P.↩\n * \n\nNote that, just as we saw with the combine function\n `c()` and the remove function `rm()` , the `sum()` function has unnamed arguments. I’ll talk about unnamed arguments later in Section 8.4.1, but for now let’s just ignore this detail.↩ * \n\nwww.abc.net.au/news/stories/2010/09/24/3021480.htm↩\n * \n\nOr at least, the basic statistical theory – these days there is a whole subfield of statistics called robust statistics that tries to grapple with the messiness of real data and develop theory that can cope with it.↩\n * \n\nAs we saw earlier, it does have a function called\n `mode()` , but it does something completely different.↩ * \n\nThis is called a “0-1 loss function”, meaning that you either win (1) or you lose (0), with no middle ground.↩\n * \n\nWell, I will very briefly mention the one that I think is coolest, for a very particular definition of “cool”, that is. Variances are additive. Here’s what that means: suppose I have two variables \\(X\\) and \\(Y\\), whose variances are $↩\n * \n\nWith the possible exception of the third question.↩\n * \n\nStrictly, the assumption is that the data are normally distributed, which is an important concept that we’ll discuss more in Chapter 9, and will turn up over and over again later in the book.↩\n * \n\nThe assumption again being that the data are normally-distributed!↩\n * \n\nThe “\\(-3\\)” part is something that statisticians tack on to ensure that the normal curve has kurtosis zero. It looks a bit stupid, just sticking a “-3” at the end of the formula, but there are good mathematical reasons for doing this.↩\n * \n\nI haven’t discussed how to compute \\(z\\)-scores, explicitly, but you can probably guess. For a variable\n `X` , the simplest way is to use a command like \n\n```\n(X - mean(X)) / sd(X)\n```\n\n. There’s also a fancier function called `scale()` that you can use, but it relies on somewhat more complicated R concepts that I haven’t explained yet.↩ * \n\nTechnically, because I’m calculating means and standard deviations from a sample of data, but want to talk about my grumpiness relative to a population, what I’m actually doing is estimating a \\(z\\) score. However, since we haven’t talked about estimation yet (see Chapter 10) I think it’s best to ignore this subtlety, especially as it makes very little difference to our calculations.↩\n * \n\nThough some caution is usually warranted. It’s not always the case that one standard deviation on variable A corresponds to the same “kind” of thing as one standard deviation on variable B. Use common sense when trying to determine whether or not the \\(z\\) scores of two variables can be meaningfully compared.↩\n * \n\nActually, even that table is more than I’d bother with. In practice most people pick one measure of central tendency, and one measure of variability only.↩\n * \n\nJust like we saw with the variance and the standard deviation, in practice we divide by \\(N-1\\) rather than \\(N\\).↩\n * \n\nThis is an oversimplification, but it’ll do for our purposes.↩\n * \n\nIf you are reading this after having already completed Chapter 11 you might be wondering about hypothesis tests for correlations. R has a function called\n `cor.test()` that runs a hypothesis test for a single correlation, and the `psych` package contains a version called `corr.test()` that can run tests for every correlation in a correlation matrix; hypothesis tests for correlations are discussed in more detail in Section 15.6.↩ * \n\nAn alternative usage of\n `cor()` is to correlate one set of variables with another subset of variables. If `X` and `Y` are both data frames with the same number of rows, then `cor(x = X, y = Y)` will produce a correlation matrix that correlates all variables in `X` with all variables in `Y` .↩ * \n\nIt’s worth noting that, even though we have missing data for each of these variables, the output doesn’t contain any\n `NA` values. This is because, while `describe()` also has an `na.rm` argument, the default value for this function is `na.rm = TRUE` .↩ * \n\nThe technical term here is “missing completely at random” (often written MCAR for short). Makes sense, I suppose, but it does sound ungrammatical to me.↩\n\n# Chapter 6 Drawing graphs\n\nAbove all else show the data.\n\n–\n\nVisualising data is one of the most important tasks facing the data analyst. It’s important for two distinct but closely related reasons. Firstly, there’s the matter of drawing “presentation graphics”: displaying your data in a clean, visually appealing fashion makes it easier for your reader to understand what you’re trying to tell them. Equally important, perhaps even more important, is the fact that drawing graphs helps you to understand the data. To that end, it’s important to draw “exploratory graphics” that help you learn about the data as you go about analysing it. These points might seem pretty obvious, but I cannot count the number of times I’ve seen people forget them.\n\nTo give a sense of the importance of this chapter, I want to start with a classic illustration of just how powerful a good graph can be. To that end, Figure 6.1 shows a redrawing of one of the most famous data visualisations of all time: John Snow’s 1854 map of cholera deaths. The map is elegant in its simplicity. In the background we have a street map, which helps orient the viewer. Over the top, we see a large number of small dots, each one representing the location of a cholera case. The larger symbols show the location of water pumps, labelled by name. Even the most casual inspection of the graph makes it very clear that the source of the outbreak is almost certainly the Broad Street pump. Upon viewing this graph, Dr Snow arranged to have the handle removed from the pump, ending the outbreak that had killed over 500 people. Such is the power of a good data visualisation.\n\nThe goals in this chapter are twofold: firstly, to discuss several fairly standard graphs that we use a lot when analysing and presenting data, and secondly, to show you how to create these graphs in R. The graphs themselves tend to be pretty straightforward, so in that respect this chapter is pretty simple. Where people usually struggle is learning how to produce graphs, and especially, learning how to produce good graphs.88 Fortunately, learning how to draw graphs in R is reasonably simple, as long as you’re not too picky about what your graph looks like. What I mean when I say this is that R has a lot of very good graphing functions, and most of the time you can produce a clean, high-quality graphic without having to learn very much about the low-level details of how R handles graphics. Unfortunately, on those occasions when you do want to do something non-standard, or if you need to make highly specific changes to the figure, you actually do need to learn a fair bit about the these details; and those details are both complicated and boring. With that in mind, the structure of this chapter is as follows: I’ll start out by giving you a very quick overview of how graphics work in R. I’ll then discuss several different kinds of graph and how to draw them, as well as showing the basics of how to customise these plots. I’ll then talk in more detail about R graphics, discussing some of those complicated and boring issues. In a future version of this book, I intend to finish this chapter off by talking about what makes a good or a bad graph, but I haven’t yet had the time to write that section.\n\n## 6.1 An overview of R graphics\n\nReduced to its simplest form, you can think of an R graphic as being much like a painting. You start out with an empty canvas. Every time you use a graphics function, it paints some new things onto your canvas. Later on, you can paint more things over the top if you want; but just like painting, you can’t “undo” your strokes. If you make a mistake, you have to throw away your painting and start over. Fortunately, this is way more easy to do when using R than it is when painting a picture in real life: you delete the plot and then type a new set of commands.89 This way of thinking about drawing graphs is referred to as the painter’s model. So far, this probably doesn’t sound particularly complicated, and for the vast majority of graphs you’ll want to draw it’s exactly as simple as it sounds. Much like painting in real life, the headaches usually start when we dig into details. To see why, I’ll expand this “painting metaphor” a bit further just to show you the basics of what’s going on under the hood, but before I do I want to stress that you really don’t need to understand all these complexities in order to draw graphs. I’d been using R for years before I even realised that most of these issues existed! However, I don’t want you to go through the same pain I went through every time I inadvertently discovered one of these things, so here’s a quick overview.\n Firstly, if you want to paint a picture, you need to paint it on something. In real life, you can paint on lots of different things. Painting onto canvas isn’t the same as painting onto paper, and neither one is the same as painting on a wall. In R, the thing that you paint your graphic onto is called a device. For most applications that we’ll look at in this book, this “device” will be a window on your computer. If you’re using Windows as your operating system, then the name for this device is `windows` ; on a Mac it’s called `quartz` because that’s the name of the software that the Mac OS uses to draw pretty pictures; and on Linux/Unix, you’re probably using `X11` . On the other hand, if you’re using Rstudio (regardless of which operating system you’re on), there’s a separate device called `RStudioGD` that forces R to paint inside the “plots” panel in Rstudio. However, from the computers perspective there’s nothing terribly special about drawing pictures on screen: and so R is quite happy to paint pictures directly into a file. R can paint several different types of image files: `jpeg` , `png` , `postscript` , `tiff` and `bmp` files are all among the options that you have available to you. For the most part, these different devices all behave the same way, so you don’t really need to know much about the differences between them when learning how to draw pictures. But, just like real life painting, sometimes the specifics do matter. Unless stated otherwise, you can assume that I’m drawing a picture on screen, using the appropriate device (i.e., `windows` , `quartz` , `X11` or `RStudioGD` ). One the rare occasions where these behave differently from one another, I’ll try to point it out in the text. Secondly, when you paint a picture you need to paint it with something. Maybe you want to do an oil painting, but maybe you want to use watercolour. And, generally speaking, you pretty much have to pick one or the other. The analog to this in R is a “graphics system”. A graphics system defines a collection of very low-level graphics commands about what to draw and where to draw it. Something that surprises most new R users is the discovery that R actually has two completely independent graphics systems, known as traditional graphics (in the `graphics` package) and grid graphics (in the `grid` package).90 Not surprisingly, the traditional graphics system is the older of the two: in fact, it’s actually older than R since it has it’s origins in S, the system from which R is descended. Grid graphics are newer, and in some respects more powerful, so many of the more recent, fancier graphical tools in R make use of grid graphics. However, grid graphics are somewhat more complicated beasts, so most people start out by learning the traditional graphics system. Nevertheless, as long as you don’t want to use any low-level commands yourself, then you don’t really need to care about whether you’re using traditional graphics or grid graphics. However, the moment you do want to tweak your figure by using some low-level commands you do need to care. Because these two different systems are pretty much incompatible with each other, there’s a pretty big divide in R graphics universe. Unless stated otherwise, you can assume that everything I’m saying pertains to traditional graphics. Thirdly, a painting is usually done in a particular style. Maybe it’s a still life, maybe it’s an impressionist piece, or maybe you’re trying to annoy me by pretending that cubism is a legitimate artistic style. Regardless, each artistic style imposes some overarching aesthetic and perhaps even constraints on what can (or should) be painted using that style. In the same vein, R has quite a number of different packages, each of which provide a collection of high-level graphics commands. A single high-level command is capable of drawing an entire graph, complete with a range of customisation options. Most but not all of the high-level commands that I’ll talk about in this book come from the `graphics` package itself, and so belong to the world of traditional graphics. These commands all tend to share a common visual style, although there are a few graphics that I’ll use that come from other packages that differ in style somewhat. On the other side of the great divide, the grid universe relies heavily on two different packages – `lattice` and `ggplots2` – each of which provides a quite different visual style. As you’ve probably guessed, there’s a whole separate bunch of functions that you’d need to learn if you want to use `lattice` graphics or make use of the `ggplots2` . However, for the purposes of this book I’ll restrict myself to talking about the basic `graphics` tools. \nAt this point, I think we’ve covered more than enough background material. The point that I’m trying to make by providing this discussion isn’t to scare you with all these horrible details, but rather to try to convey to you the fact that R doesn’t really provide a single coherent graphics system. Instead, R itself provides a platform, and different people have built different graphical tools using that platform. As a consequence of this fact, there’s two different universes of graphics, and a great multitude of packages that live in them. At this stage you don’t need to understand these complexities, but it’s useful to know that they’re there. But for now, I think we can be happy with a simpler view of things: we’ll draw pictures on screen using the traditional graphics system, and as much as possible we’ll stick to high level commands only.\n\nSo let’s start painting.\n\n## 6.2 An introduction to plotting\n\n Before I discuss any specialised graphics, let’s start by drawing a few very simple graphs just to get a feel for what it’s like to draw pictures using R. To that end, let’s create a small vector `Fibonacci` that contains a few numbers we’d like R to draw for us. Then, we’ll ask R to `plot()` those numbers. The result is Figure 6.2. \n\n```\nFibonacci <- c( 1,1,2,3,5,8,13 )\nplot( Fibonacci )\n```\n\n As you can see, what R has done is plot the values stored in the `Fibonacci` variable on the vertical axis (y-axis) and the corresponding index on the horizontal axis (x-axis). In other words, since the 4th element of the vector has a value of 3, we get a dot plotted at the location (4,3). That’s pretty straightforward, and the image in Figure 6.2 is probably pretty close to what you would have had in mind when I suggested that we plot the `Fibonacci` data. However, there’s quite a lot of customisation options available to you, so we should probably spend a bit of time looking at some of those options. So, be warned: this ends up being a fairly long section, because there’s so many possibilities open to you. Don’t let it overwhelm you though… while all of the options discussed here are handy to know about, you can get by just fine only knowing a few of them. The only reason I’ve included all this stuff right at the beginning is that it ends up making the rest of the chapter a lot more readable! \n\n### 6.2.1 A tedious digression\n\n Before we go into any discussion of customising plots, we need a little more background. The important thing to note when using the `plot()` function, is that it’s another example of a generic function (Section 4.11, much like `print()` and `summary()` , and so its behaviour changes depending on what kind of input you give it. However, the `plot()` function is somewhat fancier than the other two, and its behaviour depends on two arguments, `x` (the first input, which is required) and `y` (which is optional). This makes it (a) extremely powerful once you get the hang of it, and (b) hilariously unpredictable, when you’re not sure what you’re doing. As much as possible, I’ll try to make clear what type of inputs produce what kinds of outputs. For now, however, it’s enough to note that I’m only doing very basic plotting, and as a consequence all of the work is being done by the `plot.default()` function. What kinds of customisations might we be interested in? If you look at the help documentation for the default plotting method (i.e., type `?plot.default` or `help(\"plot.default\")` ) you’ll see a very long list of arguments that you can specify to customise your plot. I’ll talk about several of them in a moment, but first I want to point out something that might seem quite wacky. When you look at all the different options that the help file talks about, you’ll notice that some of the options that it refers to are “proper” arguments to the `plot.default()` function, but it also goes on to mention a bunch of things that look like they’re supposed to be arguments, but they’re not listed in the “Usage” section of the file, and the documentation calls them graphical parameters instead. Even so, it’s usually possible to treat them as if they were arguments of the plotting function. Very odd. In order to stop my readers trying to find a brick and look up my home address, I’d better explain what’s going on; or at least give the basic gist behind it. What exactly is a graphical parameter? Basically, the idea is that there are some characteristics of a plot which are pretty universal: for instance, regardless of what kind of graph you’re drawing, you probably need to specify what colour to use for the plot, right? So you’d expect there to be something like a `col` argument to every single graphics function in R? Well, sort of. In order to avoid having hundreds of arguments for every single function, what R does is refer to a bunch of these “graphical parameters” which are pretty general purpose. Graphical parameters can be changed directly by using the low-level `par()` function, which I discuss briefly in Section 6.7.1 though not in a lot of detail. If you look at the help files for graphical parameters (i.e., type `?par` ) you’ll see that there’s lots of them. Fortunately, (a) the default settings are generally pretty good so you can ignore the majority of the parameters, and (b) as you’ll see as we go through this chapter, you very rarely need to use `par()` directly, because you can “pretend” that graphical parameters are just additional arguments to your high-level function (e.g. `plot.default()` ). In short… yes, R does have these wacky “graphical parameters” which can be quite confusing. But in most basic uses of the plotting functions, you can act as if they were just undocumented additional arguments to your function. \n\n### 6.2.2 Customising the title and the axis labels\n\nOne of the first things that you’ll find yourself wanting to do when customising your plot is to label it better. You might want to specify more appropriate axis labels, add a title or add a subtitle. The arguments that you need to specify to make this happen are:\n\n* \n `main` . A character string containing the title. * \n `sub` . A character string containing the subtitle. * \n `xlab` . A character string containing the x-axis label. * \n `ylab` . A character string containing the y-axis label. \nThese aren’t graphical parameters, they’re arguments to the high-level function. However, because the high-level functions all rely on the same low-level function to do the drawing91 the names of these arguments are identical for pretty much every high-level function I’ve come across. Let’s have a look at what happens when we make use of all these arguments. Here’s the command. The picture that this draws is shown in Figure 6.3.\n\n```\nplot( x = Fibonacci,\n main = \"You specify title using the 'main' argument\",\n sub = \"The subtitle appears here! (Use the 'sub' argument for this)\",\n xlab = \"The x-axis label is 'xlab'\",\n ylab = \"The y-axis label is 'ylab'\" \n )\n```\n\nIt’s more or less as you’d expect. The plot itself is identical to the one we drew in Figure 6.2, except for the fact that we’ve changed the axis labels, and added a title and a subtitle. Even so, there’s a couple of interesting features worth calling your attention to. Firstly, notice that the subtitle is drawn below the plot, which I personally find annoying; as a consequence I almost never use subtitles. You may have a different opinion, of course, but the important thing is that you remember where the subtitle actually goes. Secondly, notice that R has decided to use boldface text and a larger font size for the title. This is one of my most hated default settings in R graphics, since I feel that it draws too much attention to the title. Generally, while I do want my reader to look at the title, I find that the R defaults are a bit overpowering, so I often like to change the settings. To that end, there are a bunch of graphical parameters that you can use to customise the font style:\n\n* Font styles:\n `font.main` , `font.sub` , `font.lab` , `font.axis` . These four parameters control the font style used for the plot title ( `font.main` ), the subtitle ( `font.sub` ), the axis labels ( `font.lab` : note that you can’t specify separate styles for the x-axis and y-axis without using low level commands), and the numbers next to the tick marks on the axis ( `font.axis` ). Somewhat irritatingly, these arguments are numbers instead of meaningful names: a value of 1 corresponds to plain text, 2 means boldface, 3 means italic and 4 means bold italic. * Font colours:\n `col.main` , `col.sub` , `col.lab` , `col.axis` . These parameters do pretty much what the name says: each one specifies a colour in which to type each of the different bits of text. Conveniently, R has a very large number of named colours (type `colours()` to see a list of over 650 colour names that R knows), so you can use the English language name of the colour to select it.92 Thus, the parameter value here string like `\"red\"` , `\"gray25\"` or `\"springgreen4\"` (yes, R really does recognise four different shades of “spring green”). * Font size:\n `cex.main` , `cex.sub` , `cex.lab` , `cex.axis` . Font size is handled in a slightly curious way in R. The “cex” part here is short for “character expansion”, and it’s essentially a magnification value. By default, all of these are set to a value of 1, except for the font title: `cex.main` has a default magnification of 1.2, which is why the title font is 20% bigger than the others. * Font family:\n `family` . This argument specifies a font family to use: the simplest way to use it is to set it to `\"sans\"` , `\"serif\"` , or `\"mono\"` , corresponding to a san serif font, a serif font, or a monospaced font. If you want to, you can give the name of a specific font, but keep in mind that different operating systems use different fonts, so it’s probably safest to keep it simple. Better yet, unless you have some deep objections to the R defaults, just ignore this parameter entirely. That’s what I usually do. \nTo give you a sense of how you can use these parameters to customise your titles, the following command can be used to draw Figure 6.4:\n\n```\nplot( x = Fibonacci, # the data to plot\n main = \"The first 7 Fibonacci numbers\", # the title\n xlab = \"Position in the sequence\", # x-axis label\n ylab = \"The Fibonacci number\", # y-axis \n font.main = 1,\n cex.main = 1,\n font.axis = 2,\n col.lab = \"gray50\" )\n```\n\nAlthough this command is quite long, it’s not complicated: all it does is override a bunch of the default parameter values. The only difficult aspect to this is that you have to remember what each of these parameters is called, and what all the different values are. And in practice I never remember: I have to look up the help documentation every time, or else look it up in this book.\n\n### 6.2.3 Changing the plot type\n\n Adding and customising the titles associated with the plot is one way in which you can play around with what your picture looks like. Another thing that you’ll want to do is customise the appearance of the actual plot! To start with, let’s look at the single most important options that the `plot()` function (or, recalling that we’re dealing with a generic function, in this case the `plot.default()` function, since that’s the one doing all the work) provides for you to use, which is the `type` argument. The type argument specifies the visual style of the plot. The possible values for this are: \n\n* \n `type = \"p\"` . Draw the points only. * \n `type = \"l\"` . Draw a line through the points. * \n `type = \"o\"` . Draw the line over the top of the points. * \n `type = \"b\"` . Draw both points and lines, but don’t overplot. * \n `type = \"h\"` . Draw “histogram-like” vertical bars. * \n `type = \"s\"` . Draw a staircase, going horizontally then vertically. * \n `type = \"S\"` . Draw a Staircase, going vertically then horizontally. * \n `type = \"c\"` . Draw only the connecting lines from the “b” version. * \n `type = \"n\"` . Draw nothing. (Apparently this is useful sometimes?) The simplest way to illustrate what each of these really looks like is just to draw them. To that end, Figure 6.5 shows the same Fibonacci data, drawn using six different `types` of plot. As you can see, by altering the type argument you can get a qualitatively different appearance to your plot. In other words, as far as R is concerned, the only difference between a scatterplot (like the ones we drew in Section 5.7 and a line plot is that you draw a scatterplot by setting `type = \"p\"` and you draw a line plot by setting `type = \"l\"` . However, that doesn’t imply that you should think of them as begin equivalent to each other. As you can see by looking at Figure 6.5, a line plot implies that there is some notion of continuity from one point to the next, whereas a scatterplot does not. \n\n### 6.2.4 Changing other features of the plot\n\nIn Section 6.2.2 we talked about a group of graphical parameters that are related to the formatting of titles, axis labels etc. The second group of parameters I want to discuss are those related to the formatting of the plot itself:\n\n* Colour of the plot:\n `col` . As we saw with the previous colour-related parameters, the simplest way to specify this parameter is using a character string: e.g., `col = \"blue\"` . It’s a pretty straightforward parameter to specify: the only real subtlety is that every high-level function tends to draw a different “thing” as it’s output, and so this parameter gets interpreted a little differently by different functions. However, for the `plot.default()` function it’s pretty simple: the `col` argument refers to the colour of the points and/or lines that get drawn! * Character used to plot points:\n `pch` . The plot character parameter is a number, usually between 1 and 25. What it does is tell R what symbol to use to draw the points that it plots. The simplest way to illustrate what the different values do is with a picture. Figure 6.6 shows the first 25 plotting characters. The default plotting character is a hollow circle (i.e., `pch = 1` ). \n\n* Plot size:\n `cex` . This parameter describes a character expansion factor (i.e., magnification) for the plotted characters. By default `cex=1` , but if you want bigger symbols in your graph you should specify a larger value. * Line type:\n `lty` . The line type parameter describes the kind of line that R draws. It has seven values which you can specify using a number between `0` and `7` , or using a meaningful character string: `\"blank\"` , `\"solid\"` , `\"dashed\"` , `\"dotted\"` , `\"dotdash\"` , `\"longdash\"` , or `\"twodash\"` . Note that the “blank” version (value 0) just means that R doesn’t draw the lines at all. The other six versions are shown in Figure 6.7. \n\n* Line width:\n `lwd` . The last graphical parameter in this category that I want to mention is the line width parameter, which is just a number specifying the width of the line. The default value is 1. Not surprisingly, larger values produce thicker lines and smaller values produce thinner lines. Try playing around with different values of `lwd` to see what happens. \nTo illustrate what you can do by altering these parameters, let’s try the following command, the output is shown in Figure 6.8.\n\n```\nplot( x = Fibonacci,\n type = \"b\",\n col = \"blue\",\n pch = 19,\n cex=5,\n lty=2,\n lwd=4)\n```\n\n### 6.2.5 Changing the appearance of the axes\n\n There are several other possibilities worth discussing. Ignoring graphical parameters for the moment, there’s a few other arguments to the `plot.default()` function that you might want to use. As before, many of these are standard arguments that are used by a lot of high level graphics functions: \n\n* Changing the axis scales:\n `xlim` , `ylim` . Generally R does a pretty good job of figuring out where to set the edges of the plot. However, you can override its choices by setting the `xlim` and `ylim` arguments. For instance, if I decide I want the vertical scale of the plot to run from 0 to 100, then I’d set `ylim = c(0, 100)` . * Suppress labelling:\n `ann` . This is a logical-valued argument that you can use if you don’t want R to include any text for a title, subtitle or axis label. To do so, set `ann = FALSE` . This will stop R from including any text that would normally appear in those places. Note that this will override any of your manual titles. For example, if you try to add a title using the `main` argument, but you also specify `ann = FALSE` , no title will appear. * Suppress axis drawing:\n `axes` . Again, this is a logical valued argument. Suppose you don’t want R to draw any axes at all. To suppress the axes, all you have to do is add `axes = FALSE` . This will remove the axes and the numbering, but not the axis labels (i.e. the `xlab` and `ylab` text). Note that you can get finer grain control over this by specifying the `xaxt` and `yaxt` graphical parameters instead (see below). * Include a framing box:\n `frame.plot` . Suppose you’ve removed the axes by setting `axes = FALSE` , but you still want to have a simple box drawn around the plot; that is, you only wanted to get rid of the numbering and the tick marks, but you want to keep the box. To do that, you set `frame.plot = TRUE` . Note that this list isn’t exhaustive. There are a few other arguments to the `plot.default` function that you can play with if you want to, but those are the ones you are probably most likely to want to use. As always, however, if these aren’t enough options for you, there’s also a number of other graphical parameters that you might want to play with as well. That’s the focus of the next section. In the meantime, here’s a command that makes use of all these different options. The output is shown in Figure 6.9, and it’s pretty much exactly as you’d expect. The axis scales on both the horizontal and vertical dimensions have been expanded, the axes have been suppressed as have the annotations, but I’ve kept a box around the plot. \n\n```\nplot( x = Fibonacci, # the data\n xlim = c(0, 15), # expand the x-scale\n ylim = c(0, 15), # expand the y-scale\n ann = FALSE, # delete all annotations\n axes = FALSE, # delete the axes\n frame.plot = TRUE # but include a framing box\n )\n```\n\nBefore moving on, I should point out that there are several graphical parameters relating to the axes, the box, and the general appearance of the plot which allow finer grain control over the appearance of the axes and the annotations.\n\n* Suppressing the axes individually:\n `xaxt` , `yaxt` . These graphical parameters are basically just fancier versions of the `axes` argument we discussed earlier. If you want to stop R from drawing the vertical axis but you’d like it to keep the horizontal axis, set `yaxt = \"n\"` . I trust that you can figure out how to keep the vertical axis and suppress the horizontal one! * Box type:\n `bty` . In the same way that `xaxt` , `yaxt` are just fancy versions of `axes` , the box type parameter is really just a fancier version of the `frame.plot` argument, allowing you to specify exactly which out of the four borders you want to keep. The way we specify this parameter is a bit stupid, in my opinion: the possible values are `\"o\"` (the default), `\"l\"` , `\"7\"` , `\"c\"` , `\"u\"` , or `\"]\"` , each of which will draw only those edges that the corresponding character suggests. That is, the letter `\"c\"` has a top, a bottom and a left, but is blank on the right hand side, whereas `\"7\"` has a top and a right, but is blank on the left and the bottom. Alternatively a value of `\"n\"` means that no box will be drawn. * Orientation of the axis labels\n `las` . I presume that the name of this parameter is an acronym of label style or something along those lines; but what it actually does is govern the orientation of the text used to label the individual tick marks (i.e., the numbering, not the `xlab` and `ylab` axis labels). There are four possible values for `las` : A value of 0 means that the labels of both axes are printed parallel to the axis itself (the default). A value of 1 means that the text is always horizontal. A value of 2 means that the labelling text is printed at right angles to the axis. Finally, a value of 3 means that the text is always vertical. \nAgain, these aren’t the only possibilities. There are a few other graphical parameters that I haven’t mentioned that you could use to customise the appearance of the axes,93 but that’s probably enough (or more than enough) for now. To give a sense of how you could use these parameters, let’s try the following command. The output is shown in Figure 6.10. As you can see, this isn’t a very useful plot at all. However, it does illustrate the graphical parameters we’re talking about, so I suppose it serves its purpose.\n\n```\nplot( x = Fibonacci, # the data\n xaxt = \"n\", # don't draw the x-axis \n bty = \"]\", # keep bottom, right and top of box only\n las = 1 ) # rotate the text\n```\n\n### 6.2.6 Don’t panic\n\nAt this point, a lot of readers will be probably be thinking something along the lines of, “if there’s this much detail just for drawing a simple plot, how horrible is it going to get when we start looking at more complicated things?” Perhaps, contrary to my earlier pleas for mercy, you’ve found a brick to hurl and are right now leafing through an Adelaide phone book trying to find my address. Well, fear not! And please, put the brick down. In a lot of ways, we’ve gone through the hardest part: we’ve already covered vast majority of the plot customisations that you might want to do. As you’ll see, each of the other high level plotting commands we’ll talk about will only have a smallish number of additional options. Better yet, even though I’ve told you about a billion different ways of tweaking your plot, you don’t usually need them. So in practice, now that you’ve read over it once to get the gist, the majority of the content of this section is stuff you can safely forget: just remember to come back to this section later on when you want to tweak your plot.\n\n## 6.3 Histograms\n\n Now that we’ve tamed (or possibly fled from) the beast that is R graphical parameters, let’s talk more seriously about some real life graphics that you’ll want to draw. We begin with the humble histogram. Histograms are one of the simplest and most useful ways of visualising data. They make most sense when you have an interval or ratio scale (e.g., the `afl.margins` data from Chapter 5 and what you want to do is get an overall impression of the data. Most of you probably know how histograms work, since they’re so widely used, but for the sake of completeness I’ll describe them. All you do is divide up the possible values into bins, and then count the number of observations that fall within each bin. This count is referred to as the frequency of the bin, and is displayed as a bar: in the AFL winning margins data, there are 33 games in which the winning margin was less than 10 points, and it is this fact that is represented by the height of the leftmost bar in Figure 6.11. Drawing this histogram in R is pretty straightforward. The function you need to use is called `hist()` , and it has pretty reasonable default settings. In fact, Figure 6.11 is exactly what you get if you just type this: `hist( afl.margins )` Although this image would need a lot of cleaning up in order to make a good presentation graphic (i.e., one you’d include in a report), it nevertheless does a pretty good job of describing the data. In fact, the big strength of a histogram is that (properly used) it does show the entire spread of the data, so you can get a pretty good sense about what it looks like. The downside to histograms is that they aren’t very compact: unlike some of the other plots I’ll talk about it’s hard to cram 20-30 histograms into a single image without overwhelming the viewer. And of course, if your data are nominal scale (e.g., the `afl.finalists` data) then histograms are useless. The main subtlety that you need to be aware of when drawing histograms is determining where the `breaks` that separate bins should be located, and (relatedly) how many breaks there should be. In Figure 6.11, you can see that R has made pretty sensible choices all by itself: the breaks are located at 0, 10, 20, … 120, which is exactly what I would have done had I been forced to make a choice myself. On the other hand, consider the two histograms in Figure 6.12 and 6.13, which I produced using the following two commands: \n\n```\nhist( x = afl.margins, breaks = 3 )\n```\n\n```\nhist( x = afl.margins, breaks = 0:116 )\n```\n\nIn Figure 6.13, the bins are only 1 point wide. As a result, although the plot is very informative (it displays the entire data set with no loss of information at all!) the plot is very hard to interpret, and feels quite cluttered. On the other hand, the plot in Figure 6.12 has a bin width of 50 points, and has the opposite problem: it’s very easy to “read” this plot, but it doesn’t convey a lot of information. One gets the sense that this histogram is hiding too much. In short, the way in which you specify the breaks has a big effect on what the histogram looks like, so it’s important to make sure you choose the breaks sensibly. In general R does a pretty good job of selecting the breaks on its own, since it makes use of some quite clever tricks that statisticians have devised for automatically selecting the right bins for a histogram, but nevertheless it’s usually a good idea to play around with the breaks a bit to see what happens.\n There is one fairly important thing to add regarding how the `breaks` argument works. There are two different ways you can specify the breaks. You can either specify how many breaks you want (which is what I did for panel b when I typed `breaks = 3` ) and let R figure out where they should go, or you can provide a vector that tells R exactly where the breaks should be placed (which is what I did for panel c when I typed `breaks = 0:116` ). The behaviour of the `hist()` function is slightly different depending on which version you use. If all you do is tell it how many breaks you want, R treats it as a “suggestion” not as a demand. It assumes you want “approximately 3” breaks, but if it doesn’t think that this would look very pretty on screen, it picks a different (but similar) number. It does this for a sensible reason – it tries to make sure that the breaks are located at sensible values (like 10) rather than stupid ones (like 7.224414). And most of the time R is right: usually, when a human researcher says “give me 3 breaks”, he or she really does mean “give me approximately 3 breaks, and don’t put them in stupid places”. However, sometimes R is dead wrong. Sometimes you really do mean “exactly 3 breaks”, and you know precisely where you want them to go. So you need to invoke “real person privilege”, and order R to do what it’s bloody well told. In order to do that, you have to input the full vector that tells R exactly where you want the breaks. If you do that, R will go back to behaving like the nice little obedient calculator that it’s supposed to be. \n\n### 6.3.1 Visual style of your histogram\n\n Okay, so at this point we can draw a basic histogram, and we can alter the number and even the location of the `breaks` . However, the visual style of the histograms shown in Figures 6.11, 6.12, and 6.13 could stand to be improved. We can fix this by making use of some of the other arguments to the `hist()` function. Most of the things you might want to try doing have already been covered in Section 6.2, but there’s a few new things: \n\n* Shading lines:\n `density` , `angle` . You can add diagonal lines to shade the bars: the `density` value is a number indicating how many lines per inch R should draw (the default value of `NULL` means no lines), and the `angle` is a number indicating how many degrees from horizontal the lines should be drawn at (default is `angle = 45` degrees). * Specifics regarding colours:\n `col` , `border` . You can also change the colours: in this instance the `col` parameter sets the colour of the shading (either the shading lines if there are any, or else the colour of the interior of the bars if there are not), and the `border` argument sets the colour of the edges of the bars. * Labelling the bars:\n `labels` . You can also attach labels to each of the bars using the `labels` argument. The simplest way to do this is to set `labels = TRUE` , in which case R will add a number just above each bar, that number being the exact number of observations in the bin. Alternatively, you can choose the labels yourself, by inputting a vector of strings, e.g., \n\n```\nlabels = c(\"label 1\",\"label 2\",\"etc\")\n```\n\n Not surprisingly, this doesn’t exhaust the possibilities. If you type `help(\"hist\")` or `?hist` and have a look at the help documentation for histograms, you’ll see a few more options. A histogram that makes use of the histogram-specific customisations as well as several of the options we discussed in Section 6.2 is shown in Figure 6.14. The R command that I used to draw it is this: \n\n```\nhist( x = afl.margins, \n main = \"2010 AFL margins\", # title of the plot\n xlab = \"Margin\", # set the x-axis label\n density = 10, # draw shading lines: 10 per inch\n angle = 40, # set the angle of the shading lines is 40 degrees\n border = \"gray20\", # set the colour of the borders of the bars\n col = \"gray80\", # set the colour of the shading lines\n labels = TRUE, # add frequency labels to each bar\n ylim = c(0,40) # change the scale of the y-axis\n)\n```\n\nOverall, this is a much nicer histogram than the default ones.\n\n## 6.4 Stem and leaf plots\n\nHistograms are one of the most widely used methods for displaying the observed values for a variable. They’re simple, pretty, and very informative. However, they do take a little bit of effort to draw. Sometimes it can be quite useful to make use of simpler, if less visually appealing, options. One such alternative is the stem and leaf plot. To a first approximation you can think of a stem and leaf plot as a kind of text-based histogram. Stem and leaf plots aren’t used as widely these days as they were 30 years ago, since it’s now just as easy to draw a histogram as it is to draw a stem and leaf plot. Not only that, they don’t work very well for larger data sets. As a consequence you probably won’t have as much of a need to use them yourself, though you may run into them in older publications. These days, the only real world situation where I use them is if I have a small data set with 20-30 data points and I don’t have a computer handy, because it’s pretty easy to quickly sketch a stem and leaf plot by hand.\n With all that as background, lets have a look at stem and leaf plots. The AFL margins data contains 176 observations, which is at the upper end for what you can realistically plot this way. The function in R for drawing stem and leaf plots is called `stem()` and if we ask for a stem and leaf plot of the `afl.margins` data, here’s what we get: `stem( afl.margins )` \n\n```\n## \n## The decimal point is 1 digit(s) to the right of the |\n## \n## 0 | 001111223333333344567788888999999\n## 1 | 0000011122234456666899999\n## 2 | 00011222333445566667788999999\n## 3 | 01223555566666678888899\n## 4 | 012334444477788899\n## 5 | 00002233445556667\n## 6 | 0113455678\n## 7 | 01123556\n## 8 | 122349\n## 9 | 458\n## 10 | 148\n## 11 | 6\n```\n\n The values to the left of the `|` are called stems and the values to the right are called leaves. If you just look at the shape that the leaves make, you can see something that looks a lot like a histogram made out of numbers, just rotated by 90 degrees. But if you know how to read the plot, there’s quite a lot of additional information here. In fact, it’s also giving you the actual values of all of the observations in the data set. To illustrate, let’s have a look at the last line in the stem and leaf plot, namely `11 | 6` . Specifically, let’s compare this to the largest values of the `afl.margins` data set: `max( afl.margins )` `## [1] 116` Hm… `11 | 6` versus `116` . Obviously the stem and leaf plot is trying to tell us that the largest value in the data set is 116. Similarly, when we look at the line that reads `10 | 148` , the way we interpret it to note that the stem and leaf plot is telling us that the data set contains observations with values 101, 104 and 108. Finally, when we see something like \n\n```\n5 | 00002233445556667\n```\n\nthe four `0` s in the the stem and leaf plot are telling us that there are four observations with value 50. \nI won’t talk about them in a lot of detail, but I should point out that some customisation options are available for stem and leaf plots in R. The two arguments that you can use to do this are:\n\n* \n `scale` . Changing the `scale` of the plot (default value is 1), which is analogous to changing the number of breaks in a histogram. Reducing the scale causes R to reduce the number of stem values (i.e., the number of breaks, if this were a histogram) that the plot uses. * \n `width` . The second way that to can customise a stem and leaf plot is to alter the `width` (default value is 80). Changing the width alters the maximum number of leaf values that can be displayed for any given stem. \nHowever, since stem and leaf plots aren’t as important as they used to be, I’ll leave it to the interested reader to investigate these options. Try the following two commands to see what happens:\n\n```\nstem( x = afl.margins, scale = .25 )\nstem( x = afl.margins, width = 20 )\n```\n\n The only other thing to note about stem and leaf plots is the line in which R tells you where the decimal point is. If our data set had included only the numbers .11, .15, .23, .35 and .59 and we’d drawn a stem and leaf plot of these data, then R would move the decimal point: the stem values would be 1,2,3,4 and 5, but R would tell you that the decimal point has moved to the left of the `|` symbol. If you want to see this in action, try the following command: \n\n```\nstem( x = afl.margins / 1000 )\n```\n\nThe stem and leaf plot itself will look identical to the original one we drew, except for the fact that R will tell you that the decimal point has moved.\n\n## 6.5 Boxplots\n\n Another alternative to histograms is a boxplot, sometimes called a “box and whiskers” plot. Like histograms, they’re most suited to interval or ratio scale data. The idea behind a boxplot is to provide a simple visual depiction of the median, the interquartile range, and the range of the data. And because they do so in a fairly compact way, boxplots have become a very popular statistical graphic, especially during the exploratory stage of data analysis when you’re trying to understand the data yourself. Let’s have a look at how they work, again using the `afl.margins` data as our example. Firstly, let’s actually calculate these numbers ourselves using the `summary()` function:94 \n\n```\nsummary( afl.margins )\n```\n\n So how does a boxplot capture these numbers? The easiest way to describe what a boxplot looks like is just to draw one. The function for doing this in R is (surprise, surprise) `boxplot()` . As always there’s a lot of optional arguments that you can specify if you want, but for the most part you can just let R choose the defaults for you. That said, I’m going to override one of the defaults to start with by specifying the `range` option, but for the most part you won’t want to do this (I’ll explain why in a minute). With that as preamble, let’s try the following command: \n\n```\nboxplot( x = afl.margins, range = 100 )\n```\n\nWhat R draws is shown in Figure 6.15, the most basic boxplot possible. When you look at this plot, this is how you should interpret it: the thick line in the middle of the box is the median; the box itself spans the range from the 25th percentile to the 75th percentile; and the “whiskers” cover the full range from the minimum value to the maximum value. This is summarised in the annotated plot in Figure 6.16.\n In practice, this isn’t quite how boxplots usually work. In most applications, the “whiskers” don’t cover the full range from minimum to maximum. Instead, they actually go out to the most extreme data point that doesn’t exceed a certain bound. By default, this value is 1.5 times the interquartile range, corresponding to a `range` value of 1.5. Any observation whose value falls outside this range is plotted as a circle instead of being covered by the whiskers, and is commonly referred to as an outlier. For our AFL margins data, there is one observation (a game with a margin of 116 points) that falls outside this range. As a consequence, the upper whisker is pulled back to the next largest observation (a value of 108), and the observation at 116 is plotted as a circle. This is illustrated in Figure 6.17. Since the default value is `range = 1.5` we can draw this plot using the simple command \n\n```\nboxplot( afl.margins )\n```\n\n### 6.5.1 Visual style of your boxplot\n\n I’ll talk a little more about the relationship between boxplots and outliers in the Section 6.5.2, but before I do let’s take the time to clean this figure up. Boxplots in R are extremely customisable. In addition to the usual range of graphical parameters that you can tweak to make the plot look nice, you can also exercise nearly complete control over every element to the plot. Consider the boxplot in Figure 6.18: in this version of the plot, not only have I added labels ( `xlab` , `ylab` ) and removed the stupid border ( `frame.plot` ), I’ve also dimmed all of the graphical elements of the boxplot except the central bar that plots the median ( `border` ) so as to draw more attention to the median rather than the rest of the boxplot. You’ve seen all these options in previous sections in this chapter, so hopefully those customisations won’t need any further explanation. However, I’ve done two new things as well: I’ve deleted the cross-bars at the top and bottom of the whiskers (known as the “staples” of the plot), and converted the whiskers themselves to solid lines. The arguments that I used to do this are called by the ridiculous names of `staplewex` and `whisklty` ,95 and I’ll explain these in a moment. \nBut first, here’s the actual command I used to draw this figure:\n\n```\nboxplot( x = afl.margins, # the data\n xlab = \"AFL games, 2010\", # x-axis label\n ylab = \"Winning Margin\", # y-axis label\n border = \"grey50\", # dim the border of the box\n frame.plot = FALSE, # don't draw a frame\n staplewex = 0, # don't draw staples\n whisklty = 1 # solid line for whisker \n )\n```\n\nOverall, I think the resulting boxplot is a huge improvement in visual design over the default version. In my opinion at least, there’s a fairly minimalist aesthetic that governs good statistical graphics. Ideally, every visual element that you add to a plot should convey part of the message. If your plot includes things that don’t actually help the reader learn anything new, you should consider removing them. Personally, I can’t see the point of the cross-bars on a standard boxplot, so I’ve deleted them.\n Okay, what commands can we use to customise the boxplot? If you type `?boxplot` and flick through the help documentation, you’ll notice that it does mention `staplewex` as an argument, but there’s no mention of `whisklty` . The reason for this is that the function that handles the drawing is called `bxp()` , so if you type `?bxp` all the gory details appear. Here’s the short summary. In order to understand why these arguments have such stupid names, you need to recognise that they’re put together from two components. The first part of the argument name specifies one part of the box plot: `staple` refers to the staples of the plot (i.e., the cross-bars), and `whisk` refers to the whiskers. The second part of the name specifies a graphical parameter: `wex` is a width parameter, and `lty` is a line type parameter. The parts of the plot you can customise are: \n\n* \n `box` . The box that covers the interquartile range. * \n `med` . The line used to show the median. * \n `whisk` . The vertical lines used to draw the whiskers. * \n `staple` . The cross bars at the ends of the whiskers. * \n `out` . The points used to show the outliers. \nThe actual graphical parameters that you might want to specify are slightly different for each visual element, just because they’re different shapes from each other. As a consequence, the following options are available:\n\n* Width expansion:\n\n```\nboxwex, staplewex, outwex\n```\n\n. These are scaling factors that govern the width of various parts of the plot. The default scaling factor is (usually) 0.8 for the box, and 0.5 for the other two. Note that in the case of the outliers this parameter is meaningless unless you decide to draw lines plotting the outliers rather than use points. * Line type:\n\n```\nboxlty, medlty, whisklty, staplelty, outlty\n```\n\n. These govern the line type for the relevant elements. The values for this are exactly the same as those used for the regular `lty` parameter, with two exceptions. There’s an additional option where you can set `medlty = \"blank\"` to suppress the median line completely (useful if you want to draw a point for the median rather than plot a line). Similarly, by default the outlier line type is set to `outlty = \"blank\"` , because the default behaviour is to draw outliers as points instead of lines. * Line width:\n\n```\nboxlwd, medlwd, whisklwd, staplelwd, outlwd\n```\n\n. These govern the line widths for the relevant elements, and behave the same way as the regular `lwd` parameter. The only thing to note is that the default value for `medlwd` value is three times the value of the others. * Line colour:\n\n```\nboxcol, medcol, whiskcol, staplecol, outcol\n```\n\n. These govern the colour of the lines used to draw the relevant elements. Specify a colour in the same way that you usually do. * Fill colour:\n `boxfill` . What colour should we use to fill the box? * Point character:\n `medpch, outpch` . These behave like the regular `pch` parameter used to select the plot character. Note that you can set `outpch = NA` to stop R from plotting the outliers at all, and you can also set `medpch = NA` to stop it from drawing a character for the median (this is the default!) * Point expansion:\n `medcex, outcex` . Size parameters for the points used to plot medians and outliers. These are only meaningful if the corresponding points are actually plotted. So for the default boxplot, which includes outlier points but uses a line rather than a point to draw the median, only the `outcex` parameter is meaningful. * Background colours:\n `medbg, outbg` . Again, the background colours are only meaningful if the points are actually plotted. \nTaken as a group, these parameters allow you almost complete freedom to select the graphical style for your boxplot that you feel is most appropriate to the data set you’re trying to describe. That said, when you’re first starting out there’s no shame in using the default settings! But if you want to master the art of designing beautiful figures, it helps to try playing around with these parameters to see what works and what doesn’t. Finally, I should mention a few other arguments that you might want to make use of:\n\n* \n `horizontal` . Set this to `TRUE` to display the plot horizontally rather than vertically. * \n `varwidth` . Set this to `TRUE` to get R to scale the width of each box so that the areas are proportional to the number of observations that contribute to the boxplot. This is only useful if you’re drawing multiple boxplots at once (see Section 6.5.3. * \n `show.names` . Set this to `TRUE` to get R to attach labels to the boxplots. * \n `notch` . If you set `notch = TRUE` , R will draw little notches in the sides of each box. If the notches of two boxplots don’t overlap, then there is a “statistically significant” difference between the corresponding medians. If you haven’t read Chapter 11, ignore this argument – we haven’t discussed statistical significance, so this doesn’t mean much to you. I’m mentioning it only because you might want to come back to the topic later on. (see also the `notch.frac` option when you type `?bxp` ). \n\n### 6.5.2 Using box plots to detect outliers\n\n Because the boxplot automatically (unless you change the `range` argument) separates out those observations that lie within a certain range, people often use them as an informal method for detecting outliers: observations that are “suspiciously” distant from the rest of the data. Here’s an example. Suppose that I’d drawn the boxplot for the AFL margins data, and it came up looking like Figure 6.19. It’s pretty clear that something funny is going on with one of the observations. Apparently, there was one game in which the margin was over 300 points! That doesn’t sound right to me. Now that I’ve become suspicious, it’s time to look a bit more closely at the data. One function that can be handy for this is the `which()` function; it takes as input a vector of logicals, and outputs the indices of the `TRUE` cases. This is particularly useful in the current context because it lets me do this: \n\n```\nsuspicious.cases <- afl.margins > 300\nwhich( suspicious.cases )\n```\n\n `## [1] 137` although in real life I probably wouldn’t bother creating the `suspicious.cases` variable: I’d just cut out the middle man and use a command like \n\n```\nwhich( afl.margins > 300 )\n```\n\n. In any case, what this has done is shown me that the outlier corresponds to game 137. Then, I find the recorded margin for that game: `afl.margins[137]` `## [1] 333` Hm. That definitely doesn’t sound right. So then I go back to the original data source (the internet!) and I discover that the actual margin of that game was 33 points. Now it’s pretty clear what happened. Someone must have typed in the wrong number. Easily fixed, just by typing\n\n```\nafl.margins[137] <- 33\n```\n\n. While this might seem like a silly example, I should stress that this kind of thing actually happens a lot. Real world data sets are often riddled with stupid errors, especially when someone had to type something into a computer at some point. In fact, there’s actually a name for this phase of data analysis, since in practice it can waste a huge chunk of our time: data cleaning. It involves searching for typos, missing data and all sorts of other obnoxious errors in raw data files.96 \nWhat about the real data? Does the value of 116 constitute a funny observation not? Possibly. As it turns out the game in question was Fremantle v Hawthorn, and was played in round 21 (the second last home and away round of the season). Fremantle had already qualified for the final series and for them the outcome of the game was irrelevant; and the team decided to rest several of their star players. As a consequence, Fremantle went into the game severely underpowered. In contrast, Hawthorn had started the season very poorly but had ended on a massive winning streak, and for them a win could secure a place in the finals. With the game played on Hawthorn’s home turf97 and with so many unusual factors at play, it is perhaps no surprise that Hawthorn annihilated Fremantle by 116 points. Two weeks later, however, the two teams met again in an elimination final on Fremantle’s home ground, and Fremantle won comfortably by 30 points.98\n\nSo, should we exclude the game from subsequent analyses? If this were a psychology experiment rather than an AFL season, I’d be quite tempted to exclude it because there’s pretty strong evidence that Fremantle weren’t really trying very hard: and to the extent that my research question is based on an assumption that participants are genuinely trying to do the task. On the other hand, in a lot of studies we’re actually interested in seeing the full range of possible behaviour, and that includes situations where people decide not to try very hard: so excluding that observation would be a bad idea. In the context of the AFL data, a similar distinction applies. If I’d been trying to make tips about who would perform well in the finals, I would have (and in fact did) disregard the Round 21 massacre, because it’s way too misleading. On the other hand, if my interest is solely in the home and away season itself, I think it would be a shame to throw away information pertaining to one of the most distinctive (if boring) games of the year. In other words, the decision about whether to include outliers or exclude them depends heavily on why you think the data look they way they do, and what you want to use the data for. Statistical tools can provide an automatic method for suggesting candidates for deletion, but you really need to exercise good judgment here. As I’ve said before, R is a mindless automaton. It doesn’t watch the footy, so it lacks the broader context to make an informed decision. You are not a mindless automaton, so you should exercise judgment: if the outlier looks legitimate to you, then keep it. In any case, I’ll return to the topic again in Section 15.9, so let’s return to our discussion of how to draw boxplots.\n\n### 6.5.3 Drawing multiple boxplots\n\n One last thing. What if you want to draw multiple boxplots at once? Suppose, for instance, I wanted separate boxplots showing the AFL margins not just for 2010, but for every year between 1987 and 2010. To do that, the first thing we’ll have to do is find the data. These are stored in the `aflsmall2.Rdata` file. So let’s load it and take a quick peek at what’s inside: \n\n```\nload( \"aflsmall2.Rdata\" )\nwho( TRUE )\n# -- Name -- -- Class -- -- Size --\n# afl2 data.frame 4296 x 2 \n# $margin numeric 4296 \n# $year numeric 4296\n```\n\n Notice that `afl2` data frame is pretty big. It contains 4296 games, which is far more than I want to see printed out on my computer screen. To that end, R provides you with a few useful functions to print out only a few of the row in the data frame. The first of these is `head()` which prints out the first 6 rows, of the data frame, like this: `head( afl2 )` \n\n```\n## margin year\n## 1 33 1987\n## 2 59 1987\n## 3 45 1987\n## 4 91 1987\n## 5 39 1987\n## 6 1 1987\n```\n\n You can also use the `tail()` function to print out the last 6 rows. The `car` package also provides a handy little function called `some()` which prints out a random subset of the rows. In any case, the important thing is that we have the `afl2` data frame which contains the variables that we’re interested in. What we want to do is have R draw boxplots for the `margin` variable, plotted separately for each separate `year` . The way to do this using the `boxplot()` function is to input a `formula` rather than a variable as the input. In this case, the formula we want is `margin ~ year` . So our boxplot command now looks like this. The result is shown in Figure 6.20.99 \n\n```\nboxplot( formula = margin ~ year,\n data = afl2\n)\n```\n\nEven this, the default version of the plot, gives a sense of why it’s sometimes useful to choose boxplots instead of histograms. Even before taking the time to turn this basic output into something more readable, it’s possible to get a good sense of what the data look like from year to year without getting overwhelmed with too much detail. Now imagine what would have happened if I’d tried to cram 24 histograms into this space: no chance at all that the reader is going to learn anything useful.\n\nThat being said, the default boxplot leaves a great deal to be desired in terms of visual clarity. The outliers are too visually prominent, the dotted lines look messy, and the interesting content (i.e., the behaviour of the median and the interquartile range across years) gets a little obscured. Fortunately, this is easy to fix, since we’ve already covered a lot of tools you can use to customise your output. After playing around with several different versions of the plot, the one I settled on is shown in Figure 6.21. The command I used to produce it is long, but not complicated:\n\n```\nboxplot( formula = margin ~ year, # the formula\n data = afl2, # the data set\n xlab = \"AFL season\", # x axis label\n ylab = \"Winning Margin\", # y axis label\n frame.plot = FALSE, # don't draw a frame\n staplewex = 0, # don't draw staples\n staplecol = \"white\", # (fixes a tiny display issue)\n boxwex = .75, # narrow the boxes slightly\n boxfill = \"grey80\", # lightly shade the boxes\n whisklty = 1, # solid line for whiskers \n whiskcol = \"grey70\", # dim the whiskers\n boxcol = \"grey70\", # dim the box borders\n outcol = \"grey70\", # dim the outliers\n outpch = 20, # outliers as solid dots\n outcex = .5, # shrink the outliers\n medlty = \"blank\", # no line for the medians\n medpch = 20, # instead, draw solid dots\n medlwd = 1.5 # make them larger\n )\n```\n\nOf course, given that the command is that long, you might have guessed that I didn’t spend ages typing all that rubbish in over and over again. Instead, I wrote a script, which I kept tweaking until it produced the figure that I wanted. We’ll talk about scripts later in Section 8.1, but given the length of the command I thought I’d remind you that there’s an easier way of trying out different commands than typing them all in over and over.\n\n## 6.6 Scatterplots\n\n Scatterplots are a simple but effective tool for visualising data. We’ve already seen scatterplots in this chapter, when using the `plot()` function to draw the `Fibonacci` variable as a collection of dots (Section 6.2. However, for the purposes of this section I have a slightly different notion in mind. Instead of just plotting one variable, what I want to do with my scatterplot is display the relationship between two variables, like we saw with the figures in the section on correlation (Section 5.7. It’s this latter application that we usually have in mind when we use the term “scatterplot”. In this kind of plot, each observation corresponds to one dot: the horizontal location of the dot plots the value of the observation on one variable, and the vertical location displays its value on the other variable. In many situations you don’t really have a clear opinions about what the causal relationship is (e.g., does A cause B, or does B cause A, or does some other variable C control both A and B). If that’s the case, it doesn’t really matter which variable you plot on the x-axis and which one you plot on the y-axis. However, in many situations you do have a pretty strong idea which variable you think is most likely to be causal, or at least you have some suspicions in that direction. If so, then it’s conventional to plot the cause variable on the x-axis, and the effect variable on the y-axis. With that in mind, let’s look at how to draw scatterplots in R, using the same `parenthood` data set (i.e. `parenthood.Rdata` ) that I used when introducing the idea of correlations. Suppose my goal is to draw a scatterplot displaying the relationship between the amount of sleep that I get ( `dan.sleep` ) and how grumpy I am the next day ( `dan.grump` ). As you might expect given our earlier use of `plot()` to display the `Fibonacci` data, the function that we use is the `plot()` function, but because it’s a generic function all the hard work is still being done by the `plot.default()` function. In any case, there are two different ways in which we can get the plot that we’re after. The first way is to specify the name of the variable to be plotted on the `x` axis and the variable to be plotted on the `y` axis. When we do it this way, the command looks like this: \n\n```\nplot( x = parenthood$dan.sleep, # data on the x-axis\n y = parenthood$dan.grump # data on the y-axis\n )\n```\n\n The second way do to it is to use a “formula and data frame” format, but I’m going to avoid using it.100 For now, let’s just stick with the `x` and `y` version. If we do this, the result is the very basic scatterplot shown in Figure 6.22. This serves fairly well, but there’s a few customisations that we probably want to make in order to have this work properly. As usual, we want to add some labels, but there’s a few other things we might want to do as well. Firstly, it’s sometimes useful to rescale the plots. In Figure 6.22 R has selected the scales so that the data fall neatly in the middle. But, in this case, we happen to know that the grumpiness measure falls on a scale from 0 to 100, and the hours slept falls on a natural scale between 0 hours and about 12 or so hours (the longest I can sleep in real life). So the command I might use to draw this is: \n\n This command produces the scatterplot in Figure 6.23, or at least very nearly. What it doesn’t do is draw the line through the middle of the points. Sometimes it can be very useful to do this, and I can do so using `lines()` , which is a low level plotting function. Better yet, the arguments that I need to specify are pretty much the exact same ones that I use when calling the `plot()` function. That is, suppose that I want to draw a line that goes from the point (4,93) to the point (9.5,37). Then the `x` locations can be specified by the vector `c(4,9.5)` and the `y` locations correspond to the vector `c(93,37)` . In other words, I use this command: \n\nAnd when I do so, R plots the line over the top of the plot that I drew using the previous command. In most realistic data analysis situations you absolutely don’t want to just guess where the line through the points goes, since there’s about a billion different ways in which you can get R to do a better job. However, it does at least illustrate the basic idea.\n One possibility, if you do want to get R to draw nice clean lines through the data for you, is to use the `scatterplot()` function in the `car` package. Before we can use `scatterplot()` we need to load the package: `library( car )` \nHaving done so, we can now use the function. The command we need is this one:\n\n```\nscatterplot( dan.grump ~ dan.sleep,\n data = parenthood, \n smooth = FALSE\n)\n```\n\n The first two arguments should be familiar: the first input is a formula\n\ntelling R what variables to plot,101 and the second specifies a `data` frame. The third argument `smooth` I’ve set to `FALSE` to stop the `scatterplot()` function from drawing a fancy “smoothed” trendline (since it’s a bit confusing to beginners). The scatterplot itself is shown in Figure 6.24. As you can see, it’s not only drawn the scatterplot, but its also drawn boxplots for each of the two variables, as well as a simple line of best fit showing the relationship between the two variables. \n\n### 6.6.1 More elaborate options\n\nOften you find yourself wanting to look at the relationships between several variables at once. One useful tool for doing so is to produce a scatterplot matrix, analogous to the correlation matrix.\n\n```\ncor( x = parenthood ) # calculate correlation matrix\n```\n\n We can get a the corresponding scatterplot matrix by using the `pairs()` function:102 \n\n```\npairs( x = parenthood ) # draw corresponding scatterplot matrix\n```\n\n The output of the `pairs()` command is shown in Figure 6.25. An alternative way of calling the `pairs()` function, which can be useful in some situations, is to specify the variables to include using a one-sided formula. For instance, this \n\n```\npairs( formula = ~ dan.sleep + baby.sleep + dan.grump,\n data = parenthood\n)\n```\n\n would produce a \\(3 \\times 3\\) scatterplot matrix that only compare `dan.sleep` , `dan.grump` and `baby.sleep` . Obviously, the first version is much easier, but there are cases where you really only want to look at a few of the variables, so it’s nice to use the formula interface. \n\n## 6.7 Bar graphs\n\n Another form of graph that you often want to plot is the bar graph. The main function that you can use in R to draw them is the `barplot()` function.103 And to illustrate the use of the function, I’ll use the `finalists` variable that I introduced in Section 5.1.7. What I want to do is draw a bar graph that displays the number of finals that each team has played in over the time spanned by the `afl` data set. So, let’s start by creating a vector that contains this information. I’ll use the `tabulate()` function to do this (which will be discussed properly in Section 7.1, since it creates a simple numeric vector: \n\n```\nfreq <- tabulate( afl.finalists )\nprint( freq )\n```\n\n```\n## [1] 26 25 26 28 32 0 6 39 27 28 28 17 6 24 26 38 24\n```\n\n This isn’t exactly the prettiest of frequency tables, of course. I’m only doing it this way so that you can see the `barplot()` function in it’s “purest” form: when the input is just an ordinary numeric vector. That being said, I’m obviously going to need the team names to create some labels, so let’s create a variable with those. I’ll do this using the `levels()` function, which outputs the names of all the levels of a factor (see Section 4.7: \n\n```\nteams <- levels( afl.finalists )\nprint( teams )\n```\n\n```\n## [1] \"Adelaide\" \"Brisbane\" \"Carlton\" \n## [4] \"Collingwood\" \"Essendon\" \"Fitzroy\" \n## [7] \"Fremantle\" \"Geelong\" \"Hawthorn\" \n## [10] \"Melbourne\" \"North Melbourne\" \"Port Adelaide\" \n## [13] \"Richmond\" \"\" \"Sydney\" \n## [16] \"West Coast\" \"Western Bulldogs\"\n```\n\n Okay, so now that we have the information we need, let’s draw our bar graph. The main argument that you need to specify for a bar graph is the `height` of the bars, which in our case correspond to the values stored in the `freq` variable: \n\n```\nbarplot( height = freq ) # specifying the argument name\nbarplot( freq ) # the lazier version\n```\n\nEither of these two commands will produce the simple bar graph shown in Figure 6.26.\n As you can see, R has drawn a pretty minimal plot. It doesn’t have any labels, obviously, because we didn’t actually tell the `barplot()` function what the labels are! To do this, we need to specify the `names.arg` argument. The `names.arg` argument needs to be a vector of character strings containing the text that needs to be used as the label for each of the items. In this case, the `teams` vector is exactly what we need, so the command we’re looking for is: \n\n```\nbarplot( height = freq, names.arg = teams )\n```\n\n This is an improvement, but not much of an improvement. R has only included a few of the labels, because it can’t fit them in the plot. This is the same behaviour we saw earlier with the multiple-boxplot graph in Figure 6.20. However, in Figure 6.20 it wasn’t an issue: it’s pretty obvious from inspection that the two unlabelled plots in between 1987 and 1990 must correspond to the data from 1988 and 1989. However, the fact that `barplot()` has omitted the names of every team in between Adelaide and Fitzroy is a lot more problematic. The simplest way to fix this is to rotate the labels, so that the text runs vertically not horizontally. To do this, we need to alter set the `las` parameter, which I discussed briefly in Section 6.2. What I’ll do is tell R to rotate the text so that it’s always perpendicular to the axes (i.e., I’ll set `las = 2` ). When I do that, as per the following command… \n\n```\nbarplot(height = freq, # the frequencies\n names.arg = teams, # the label\n las = 2) # rotate the labels\n```\n\n… the result is the bar graph shown in Figure 6.28. We’ve fixed the problem, but we’ve created a new one: the axis labels don’t quite fit anymore. To fix this, we have to be a bit cleverer again. A simple fix would be to use shorter names rather than the full name of all teams, and in many situations that’s probably the right thing to do. However, at other times you really do need to create a bit more space to add your labels, so I’ll show you how to do that.\n\n### 6.7.1 Changing global settings using par()\n\n Altering the margins to the plot is actually a somewhat more complicated exercise than you might think. In principle it’s a very simple thing to do: the size of the margins is governed by a graphical parameter called `mar` , so all we need to do is alter this parameter. First, let’s look at what the `mar` argument specifies. The `mar` argument is a vector containing four numbers: specifying the amount of space at the bottom, the left, the top and then the right. The units are “number of ‘lines’”. The default value for `mar` is \n\n```\nc(5.1, 4.1, 4.1, 2.1)\n```\n\n, meaning that R leaves 5.1 “lines” empty at the bottom, 4.1 lines on the left and the bottom, and only 2.1 lines on the right. In order to make more room at the bottom, what I need to do is change the first of these numbers. A value of 10.1 should do the trick. So far this doesn’t seem any different to the other graphical parameters that we’ve talked about. However, because of the way that the traditional graphics system in R works, you need to specify what the margins will be before calling your high-level plotting function. Unlike the other cases we’ve see, you can’t treat `mar` as if it were just another argument in your plotting function. Instead, you have to use the `par()` function to change the graphical parameters beforehand, and only then try to draw your figure. In other words, the first thing I would do is this: \n\n```\npar( mar = c( 10.1, 4.1, 4.1, 2.1) )\n```\n\nThere’s no visible output here, but behind the scenes R has changed the graphical parameters associated with the current device (remember, in R terminology all graphics are drawn onto a “device”). Now that this is done, we could use the exact same command as before, but this time you’d see that the labels all fit, because R now leaves twice as much room for the labels at the bottom. However, since I’ve now figured out how to get the labels to display properly, I might as well play around with some of the other options, all of which are things you’ve seen before:\n\n```\nbarplot( height = freq,\n names.arg = teams,\n las=2,\n ylab = \"Number of Finals\",\n main = \"Finals Played, 1987-2010\", \n density = 10,\n angle = 20)\n```\n\n However, one thing to remember about the `par()` function is that it doesn’t just change the graphical parameters for the current plot. Rather, the changes pertain to any subsequent plot that you draw onto the same device. This might be exactly what you want, in which case there’s no problem. But if not, you need to reset the graphical parameters to their original settings. To do this, you can either close the device (e.g., close the window, or click the “Clear All” button in the Plots panel in Rstudio) or you can reset the graphical parameters to their original values, using a command like this: \n\n```\npar( mar = c(5.1, 4.1, 4.1, 2.1) )\n```\n\n## 6.8 Saving image files using R and Rstudio\n\nHold on, you might be thinking. What’s the good of being able to draw pretty pictures in R if I can’t save them and send them to friends to brag about how awesome my data is? How do I save the picture? This is another one of those situations where the easiest thing to do is to use the RStudio tools.\n\nIf you’re running R through Rstudio, then the easiest way to save your image is to click on the “Export” button in the Plot panel (i.e., the area in Rstudio where all the plots have been appearing). When you do that you’ll see a menu that contains the options “Save Plot as PDF” and “Save Plot as Image”. Either version works. Both will bring up dialog boxes that give you a few options that you can play with, but besides that it’s pretty simple.\n\nThis works pretty nicely for most situations. So, unless you’re filled with a burning desire to learn the low level details, feel free to skip the rest of this section.\n\n### 6.8.1 The ugly details (advanced)\n\nAs I say, the menu-based options should be good enough for most people most of the time. However, one day you might want to be a bit more sophisticated, and make use of R’s image writing capabilities at a lower level. In this section I’ll give you a very basic introduction to this. In all honesty, this barely scratches the surface, but it will help a little bit in getting you started if you want to learn the details.\n Okay, as I hinted earlier, whenever you’re drawing pictures in R you’re deemed to be drawing to a device of some kind. There are devices that correspond to a figure drawn on screen, and there are devices that correspond to graphics files that R will produce for you. For the purposes of this section I’ll assume that you’re using the default application in either Windows or Mac OS, not Rstudio. The reason for this is that my experience with the graphical device provided by Rstudio has led me to suspect that it still has a bunch on non-standard (or possibly just undocumented) features, and so I don’t quite trust that it always does what I expect. I’ve no doubt they’ll smooth it out later, but I can honestly say that I don’t quite get what’s going on with the `RStudioGD` device. In any case, we can ask R to list all of the graphics devices that currently exist, simply by using the command `dev.list()` . If there are no figure windows open, then you’ll see this: \n\n```\ndev.list()\n# NULL\n```\n\nwhich just means that R doesn’t have any graphics devices open. However, suppose if you’ve just drawn a histogram and you type the same command, R will now give you a different answer. For instance, if you’re using Windows:\n\n```\nhist( afl.margins )\ndev.list()\n# windows \n# 2\n```\n\n What this means is that there is one graphics device (device 2) that is currently open, and it’s a figure window. If you did the same thing on a Mac, you get basically the same answer, except that the name of the device would be `quartz` rather than `windows` . If you had several graphics windows open (which, incidentally, you can do by using the `dev.new()` command) then you’d see something like this: \n\n```\ndev.list()\n# windows windows windows \n# 2 3 4\n```\n\n Okay, so that’s the basic idea behind graphics devices. The key idea here is that graphics files (like JPEG images etc) are also graphics devices as far as R is concerned. So what you want to do is to copy the contents of one graphics device to another one. There’s a command called `dev.copy()` that does this, but what I’ll explain to you is a simpler one called `dev.print()` . It’s pretty simple: \n\n```\ndev.print( device = jpeg, # what are we printing to?\n filename = \"thisfile.jpg\", # name of the image file\n width = 480, # how many pixels wide should it be\n height = 300 # how many pixels high should it be\n)\n```\n\n This takes the “active” figure window, copies it to a jpeg file (which R treats as a device) and then closes that device. The\n\n```\nfilename = \"thisfile.jpg\"\n```\n\npart tells R what to name the graphics file, and the `width = 480` and `height = 300` arguments tell R to draw an image that is 300 pixels high and 480 pixels wide. If you want a different kind of file, just change the device argument from `jpeg` to something else. R has devices for `png` , `tiff` and `bmp` that all work in exactly the same way as the `jpeg` command, but produce different kinds of files. Actually, for simple cartoonish graphics like this histogram, you’d be better advised to use PNG or TIFF over JPEG. The JPEG format is very good for natural images, but is wasteful for simple line drawings. The information above probably covers most things you might want to. However, if you want more information about what kinds of options you can specify using R, have a look at the help documentation by typing `?jpeg` or `?tiff` or whatever. \n\n## 6.9 Summary\n\nPerhaps I’m a simple minded person, but I love pictures. Every time I write a new scientific paper, one of the first things I do is sit down and think about what the pictures will be. In my head, an article is really just a sequence of pictures, linked together by a story. All the rest of it is just window dressing. What I’m really trying to say here is that the human visual system is a very powerful data analysis tool. Give it the right kind of information and it will supply a human reader with a massive amount of knowledge very quickly. Not for nothing do we have the saying “a picture is worth a thousand words”. With that in mind, I think that this is one of the most important chapters in the book. The topics covered were:\n\n* Basic overview to R graphics. In Section 6.1 we talked about how graphics in R are organised, and then moved on to the basics of how they’re drawn in Section 6.2.\n * Common plots. Much of the chapter was focused on standard graphs that statisticians like to produce: histograms (Section 6.3), stem and leaf plots (Section 6.4), boxplots (Section 6.5), scatterplots (Section 6.6) and bar graphs (Section 6.7).\n * Saving image files. The last part of the chapter talked about how to export your pictures (Section 6.8)\n One final thing to point out. At the start of the chapter I mentioned that R has several completely distinct systems for drawing figures. In this chapter I’ve focused on the traditional graphics system. It’s the easiest one to get started with: you can draw a histogram with a command as simple as `hist(x)` . However, it’s not the most powerful tool for the job, and after a while most R users start looking to shift to fancier systems. One of the most popular graphics systems is provided by the `ggplot2` package (see ), which is loosely based on “The grammar of graphics” (Wilkinson et al. 2006). It’s not for novices: you need to have a pretty good grasp of R before you can start using it, and even then it takes a while to really get the hang of it. But when you’re finally at that stage, it’s worth taking the time to teach yourself, because it’s a much cleaner system. \n\nThe origin of this quote is Tufte’s lovely book The Visual Display of Quantitative Information.↩\n * \n\nI should add that this isn’t unique to R. Like everything in R there’s a pretty steep learning curve to learning how to draw graphs, and like always there’s a massive payoff at the end in terms of the quality of what you can produce. But to be honest, I’ve seen the same problems show up regardless of what system people use. I suspect that the hardest thing to do is to force yourself to take the time to think deeply about what your graphs are doing. I say that in full knowledge that only about half of my graphs turn out as well as they ought to. Understanding what makes a good graph is easy: actually designing a good graph is hard.↩\n * \n\nOr, since you can always use the up and down keys to scroll through your recent command history, you can just pull up your most recent commands and edit them to fix your mistake. It becomes even easier once you start using scripts (Section 8.1, since all you have to do is edit your script and then run it again.↩\n * \n\nOf course, even that is a slightly misleading description, since some R graphics tools make use of external graphical rendering systems like OpenGL (e.g., the\n `rgl` package). I absolutely will not be talking about OpenGL or the like in this book, but as it happens there is one graph in this book that relies on them: Figure 15.6.↩ * \n\nThe low-level function that does this is called\n `title()` in case you ever need to know, and you can type `?title` to find out a bit more detail about what these arguments do.↩ * \n\nOn the off chance that this isn’t enough freedom for you, you can select a colour directly as a “red, green, blue” specification using the\n `rgb()` function, or as a “hue, saturation, value” specification using the `hsv()` function.↩ * \n\nAlso, there’s a low level function called\n `axis()` that allows a lot more control over the appearance of the axes.↩ * \n\nR being what it is, it’s no great surprise that there’s also a\n `fivenum()` function that does much the same thing.↩ * \n\nI realise there’s a kind of logic to the way R names are constructed, but they still sound dumb. When I typed this sentence, all I could think was that it sounded like the name of a kids movie if it had been written by : “The frabjous gambolles of Staplewex and Whisklty” or something along those lines.↩\n * \n\nSometimes it’s convenient to have the boxplot automatically label the outliers for you. The original\n `boxplot()` function doesn’t allow you to do this; however, the `Boxplot()` function in the `car` package does. The design of the `Boxplot()` function is very similar to `boxplot()` . It just adds a few new arguments that allow you to tweak the labelling scheme. I’ll leave it to the reader to check this out.↩ * \n\nSort of. The game was played in Launceston, which is a de facto home away from home for Hawthorn.↩\n * \n\nContrast this situation with the next largest winning margin in the data set, which was Geelong’s 108 point demolition of Richmond in round 6 at their home ground, Kardinia Park. Geelong have been one of the most dominant teams over the last several years, a period during which they strung together an incredible 29-game winning streak at Kardinia Park. Richmond have been useless for several years. This is in no meaningful sense an outlier. Geelong have been winning by these margins (and Richmond losing by them) for quite some time. Frankly I’m surprised that the result wasn’t more lopsided: as happened to Melbourne in 2011 when Geelong won by a modest 186 points.↩\n * \n\nActually, there’s other ways to do this. If the input argument\n `x` is a list object (see Section 4.9, the `boxplot()` function will draw a separate boxplot for each variable in that list. Relatedly, since the `plot()` function – which we’ll discuss shortly – is a generic (see Section 4.11, you might not be surprised to learn that one of its special cases is a boxplot: specifically, if you use `plot()` where the first argument `x` is a factor and the second argument `y` is numeric, then the result will be a boxplot, showing the values in `y` , with a separate boxplot for each level. For instance, something like \n\n```\nplot(x = afl2\\$year, y = afl2\\$margin)\n```\n\nwould work.↩ * \n\nThe reason is that there’s an annoying design flaw in the way the\n `plot()` function handles this situation. The problem is that the `plot.formula()` function uses different names to for the arguments than the `plot()` function expects. As a consequence, you can’t specify the formula argument by name. If you just specify a formula as the first argument without using the name it works fine, because the `plot()` function thinks the formula corresponds to the `x` argument, and the `plot.formula()` function thinks it corresponds to the `formula` argument; and surprisingly, everything works nicely. But the moment that you, the user, tries to be unambiguous about the name, one of those two functions is going to cry.↩ * \n\nYou might be wondering why I haven’t specified the argument name for the formula. The reason is that there’s a bug in how the\n `scatterplot()` function is written: under the hood there’s one function that expects the argument to be named `x` and another one that expects it to be called `formula` . I don’t know why the function was written this way, but it’s not an isolated problem: this particular kind of bug repeats itself in a couple of other functions (you’ll see it again in Chapter 13. The solution in such cases is to omit the argument name: that way, one function “thinks” that you’ve specified `x` and the other one “thinks” you’ve specified `formula` and everything works the way it’s supposed to. It’s not a great state of affairs, I’ll admit, but it sort of works.↩ * \n\nYet again, we could have produced this output using the\n `plot()` function: when the `x` argument is a data frame containing numeric variables only, then the output is a scatterplot matrix. So, once again, what I could have done is just type `plot( parenthood )` .↩ * \n\nOnce again, it’s worth noting the link to the generic\n `plot()` function. If the `x` argument to `plot()` is a factor (and no `y` argument is given), the result is a bar graph. So you could use \n\n```\nplot( afl.finalists )\n```\n\nand get the same output as \n\n```\nbarplot( afl.finalists )\n```\n\n.↩\n\nDate: 2019-01-11\nCategories: \nTags: \n\n# Chapter 7 Pragmatic matters\n\nThe garden of life never seems to confine itself to the plots philosophers have laid out for its convenience. Maybe a few more tractors would do the trick.\n\n–\n\nThis is a somewhat strange chapter, even by my standards. My goal in this chapter is to talk a bit more honestly about the realities of working with data than you’ll see anywhere else in the book. The problem with real world data sets is that they are messy. Very often the data file that you start out with doesn’t have the variables stored in the right format for the analysis you want to do. Sometimes might be a lot of missing values in your data set. Sometimes you only want to analyse a subset of the data. Et cetera. In other words, there’s a lot of data manipulation that you need to do, just to get all your data set into the format that you need it. The purpose of this chapter is to provide a basic introduction to all these pragmatic topics. Although the chapter is motivated by the kinds of practical issues that arise when manipulating real data, I’ll stick with the practice that I’ve adopted through most of the book and rely on very small, toy data sets that illustrate the underlying issue. Because this chapter is essentially a collection of “tricks” and doesn’t tell a single coherent story, it may be useful to start with a list of topics:\n\nAs you can see, the list of topics that the chapter covers is pretty broad, and there’s a lot of content there. Even though this is one of the longest and hardest chapters in the book, I’m really only scratching the surface of several fairly different and important topics. My advice, as usual, is to read through the chapter once and try to follow as much of it as you can. Don’t worry too much if you can’t grasp it all at once, especially the later sections. The rest of the book is only lightly reliant on this chapter, so you can get away with just understanding the basics. However, what you’ll probably find is that later on you’ll need to flick back to this chapter in order to understand some of the concepts that I refer to here.\n\n## 7.1 Tabulating and cross-tabulating data\n\n A very common task when analysing data is the construction of frequency tables, or cross-tabulation of one variable against another. There are several functions that you can use in R for that purpose. In this section I’ll illustrate the use of three functions – `table()` , `xtabs()` and `tabulate()` – though there are other options (e.g., `ftable()` ) available. \n\n### 7.1.1 Creating tables from vectors\n\n Let’s start with a simple example. As the father of a small child, I naturally spend a lot of time watching TV shows like In the Night Garden. In the `nightgarden.Rdata` file, I’ve transcribed a short section of the dialogue. The file contains two variables, `speaker` and `utterance` , and when we take a look at the data, it becomes very clear what happened to my sanity. \n\n```\nlibrary(lsr)\nload(file.path(projecthome,\"data\",\"nightgarden.Rdata\"))\nwho()\n```\n\n```\n## -- Name -- -- Class -- -- Size --\n## afl.finalists factor 400 \n## afl.margins numeric 176 \n## afl2 data.frame 4296 x 2 \n## colour logical 1 \n## d.cor numeric 1 \n## describeImg list 0 \n## effort data.frame 10 x 2 \n## emphCol character 1 \n## emphColLight character 1 \n## emphGrey character 1 \n## eps logical 1 \n## Fibonacci numeric 7 \n## freq integer 17 \n## generateRLineTypes function \n## generateRPointShapes function \n## height numeric 1 \n## old list 66 \n## oneCorPlot function \n## out.0 data.frame 100 x 2 \n## out.1 data.frame 100 x 2 \n## out.2 data.frame 100 x 2 \n## parenthood data.frame 100 x 4 \n## plotOne function \n## projecthome character 1 \n## speaker character 10 \n## suspicious.cases logical 176 \n## teams character 17 \n## utterance character 10 \n## width numeric 1 \n## X1 numeric 11 \n## X2 numeric 11 \n## X3 numeric 11 \n## X4 numeric 11 \n## Y1 numeric 11 \n## Y2 numeric 11 \n## Y3 numeric 11 \n## Y4 numeric 11\n```\n\n `print( speaker )` \n\n```\n## [1] \"upsy-daisy\" \"upsy-daisy\" \"upsy-daisy\" \"upsy-daisy\" \"tombliboo\" \n## [6] \"tombliboo\" \"makka-pakka\" \"makka-pakka\" \"makka-pakka\" \"makka-pakka\"\n```\n\n `print( utterance )` \n\n With these as my data, one task I might find myself needing to do is construct a frequency count of the number of words each character speaks during the show. The `table()` function provides a simple way do to this. The basic usage of the `table()` function is as follows: `table(speaker)` \n\n```\n## speaker\n## makka-pakka tombliboo upsy-daisy \n## 4 2 4\n```\n\n The output here tells us on the first line that what we’re looking at is a tabulation of the `speaker` variable. On the second line it lists all the different speakers that exist in the data, and on the third line it tells you how many times that speaker appears in the data. In other words, it’s a frequency table105 Notice that in the command above I didn’t name the argument, since `table()` is another function that makes use of unnamed arguments. You just type in a list of the variables that you want R to tabulate, and it tabulates them. For instance, if I type in the name of two variables, what I get as the output is a cross-tabulation: \n\n```\ntable(speaker, utterance)\n```\n\nWhen interpreting this table, remember that these are counts: so the fact that the first row and second column corresponds to a value of 2 indicates that Makka-Pakka (row 1) says “onk” (column 2) twice in this data set. As you’d expect, you can produce three way or higher order cross tabulations just by adding more objects to the list of inputs. However, I won’t discuss that in this section.\n\n### 7.1.2 Creating tables from data frames\n\nMost of the time your data are stored in a data frame, not kept as separate variables in the workspace. Let’s create one:\n\n```\nitng <- data.frame( speaker, utterance )\nitng\n```\n\nThere’s a couple of options under these circumstances. Firstly, if you just want to cross-tabulate all of the variables in the data frame, then it’s really easy:\n `table(itng)` \n\n However, it’s often the case that you want to select particular variables from the data frame to tabulate. This is where the `xtabs()` function is useful. In this function, you input a one sided `formula` in order to list all the variables you want to cross-tabulate, and the name of the `data` frame that stores the data: \n\n```\nxtabs( formula = ~ speaker + utterance, data = itng )\n```\n\n Clearly, this is a totally unnecessary command in the context of the `itng` data frame, but in most situations when you’re analysing real data this is actually extremely useful, since your data set will almost certainly contain lots of variables and you’ll only want to tabulate a few of them at a time. \n\n### 7.1.3 Converting a table of counts to a table of proportions\n\n The tabulation commands discussed so far all construct a table of raw frequencies: that is, a count of the total number of cases that satisfy certain conditions. However, often you want your data to be organised in terms of proportions rather than counts. This is where the `prop.table()` function comes in handy. It has two arguments: \n\n* \n `x` . The frequency table that you want to convert. * \n `margin` . Which “dimension” do you want to calculate proportions for. By default, R assumes you want the proportion to be expressed as a fraction of all possible events. See examples for details. \nTo see how this works:\n\n```\nitng.table <- table(itng) # create the table, and assign it to a variable\nitng.table # display the table again, as a reminder\n```\n\n```\nprop.table( x = itng.table ) # express as proportion:\n```\n\n Notice that there were 10 observations in our original data set, so all that R has done here is divide all our raw frequencies by 10. That’s a sensible default, but more often you actually want to calculate the proportions separately by row ( `margin = 1` ) or by column ( `margin = 2` ). Again, this is most clearly seen by looking at examples: \n\nNotice that each row now sums to 1, but that’s not true for each column. What we’re looking at here is the proportions of utterances made by each character. In other words, 50% of Makka-Pakka’s utterances are “pip”, and the other 50% are “onk”. Let’s contrast this with the following command:\n\nNow the columns all sum to 1 but the rows don’t. In this version, what we’re seeing is the proportion of characters associated with each utterance. For instance, whenever the utterance “ee” is made (in this data set), 100% of the time it’s a Tombliboo saying it.\n\n### 7.1.4 Low level tabulation\n\n One final function I want to mention is the `tabulate()` function, since this is actually the low-level function that does most of the hard work. It takes a numeric vector as input, and outputs frequencies as outputs: \n\n```\nsome.data <- c(1,2,3,1,1,3,1,1,2,8,3,1,2,4,2,3,5,2)\ntabulate(some.data)\n```\n\n```\n## [1] 6 5 4 1 1 0 0 1\n```\n\n## 7.2 Transforming and recoding a variable\n\nIt’s not uncommon in real world data analysis to find that one of your variables isn’t quite equivalent to the variable that you really want. For instance, it’s often convenient to take a continuous-valued variable (e.g., age) and break it up into a smallish number of categories (e.g., younger, middle, older). At other times, you may need to convert a numeric variable into a different numeric variable (e.g., you may want to analyse at the absolute value of the original variable). In this section I’ll describe a few key tricks that you can make use of to do this.\n\n### 7.2.1 Creating a transformed variable\n\nThe first trick to discuss is the idea of transforming a variable. Taken literally, anything you do to a variable is a transformation, but in practice what it usually means is that you apply a relatively simple mathematical function to the original variable, in order to create new variable that either (a) provides a better way of describing the thing you’re actually interested in or (b) is more closely in agreement with the assumptions of the statistical tests you want to do. Since – at this stage – I haven’t talked about statistical tests or their assumptions, I’ll show you an example based on the first case.\n To keep the explanation simple, the variable we’ll try to transform ( `likert.raw` ) isn’t inside a data frame, though in real life it almost certainly would be. However, I think it’s useful to start with an example that doesn’t use data frames because it illustrates the fact that you already know how to do variable transformations. To see this, let’s go through an example. Suppose I’ve run a short study in which I ask 10 people a single question: On a scale of 1 (strongly disagree) to 7 (strongly agree), to what extent do you agree with the proposition that “Dinosaurs are awesome”?\n Now let’s load and look at the data. The data file `likert.Rdata` contains a single variable that contains the raw Likert-scale responses: \n\n```\nload(file.path(projecthome,\"data\",\"likert.Rdata\"))\nlikert.raw\n```\n\nHowever, if you think about it, this isn’t the best way to represent these responses. Because of the fairly symmetric way that we set up the response scale, there’s a sense in which the midpoint of the scale should have been coded as 0 (no opinion), and the two endpoints should be \\(+3\\) (strong agree) and \\(-3\\) (strong disagree). By recoding the data in this way, it’s a bit more reflective of how we really think about the responses. The recoding here is trivially easy: we just subtract 4 from the raw scores:\n\n```\nlikert.centred <- likert.raw - 4\nlikert.centred\n```\n\n```\n## [1] -3 3 -1 0 0 0 -2 2 1 1\n```\n\n One reason why it might be useful to have the data in this format is that there are a lot of situations where you might prefer to analyse the strength of the opinion separately from the direction of the opinion. We can do two different transformations on this `likert.centred` variable in order to distinguish between these two different concepts. Firstly, to compute an `opinion.strength` variable, we want to take the absolute value of the centred data (using the `abs()` function that we’ve seen previously), like so: \n\n```\nopinion.strength <- abs( likert.centred )\nopinion.strength\n```\n\n```\n## [1] 3 3 1 0 0 0 2 2 1 1\n```\n\n Secondly, to compute a variable that contains only the direction of the opinion and ignores the strength, we can use the `sign()` function to do this. If you type `?sign` you’ll see that this function is really simple: all negative numbers are converted to \\(-1\\), all positive numbers are converted to \\(1\\) and zero stays as \\(0\\). So, when we apply the `sign()` function we obtain the following: \n\n```\nopinion.dir <- sign( likert.centred )\nopinion.dir\n```\n\n And we’re done. We now have three shiny new variables, all of which are useful transformations of the original `likert.raw` data. All of this should seem pretty familiar to you. The tools that you use to do regular calculations in R (e.g., Chapters 3 and 4) are very much the same ones that you use to transform your variables! To that end, in Section 7.3 I’ll revisit the topic of doing calculations in R because there’s a lot of other functions and operations that are worth knowing about. \nBefore moving on, you might be curious to see what these calculations look like if the data had started out in a data frame. To that end, it may help to note that the following example does all of the calculations using variables inside a data frame, and stores the variables created inside it:\n\n```\ndf <- data.frame( likert.raw ) # create data frame\ndf$likert.centred <- df$likert.raw - 4 # create centred data\ndf$opinion.strength <- abs( df$likert.centred ) # create strength variable\ndf$opinion.dir <- sign( df$likert.centred ) # create direction variable\ndf # print the final data frame:\n```\n\n```\n## likert.raw likert.centred opinion.strength opinion.dir\n## 1 1 -3 3 -1\n## 2 7 3 3 1\n## 3 3 -1 1 -1\n## 4 4 0 0 0\n## 5 4 0 0 0\n## 6 4 0 0 0\n## 7 2 -2 2 -1\n## 8 6 2 2 1\n## 9 5 1 1 1\n## 10 5 1 1 1\n```\n\n In other words, the commands you use are basically ones as before: it’s just that every time you want to read a variable from the data frame or write to the data frame, you use the `$` operator. That’s the easiest way to do it, though I should make note of the fact that people sometimes make use of the `within()` function to do the same thing. However, since (a) I don’t use the `within()` function anywhere else in this book, and (b) the `$` operator works just fine, I won’t discuss it any further. \n\n### 7.2.2 Cutting a numeric variable into categories\n\nOne pragmatic task that arises more often than you’d think is the problem of cutting a numeric variable up into discrete categories. For instance, suppose I’m interested in looking at the age distribution of people at a social gathering:\n\n```\nage <- c( 60,58,24,26,34,42,31,30,33,2,9 )\n```\n\n In some situations it can be quite helpful to group these into a smallish number of categories. For example, we could group the data into three broad categories: young (0-20), adult (21-40) and older (41-60). This is a quite coarse-grained classification, and the labels that I’ve attached only make sense in the context of this data set (e.g., viewed more generally, a 42 year old wouldn’t consider themselves as “older”). We can slice this variable up quite easily using the `cut()` function.106 To make things a little cleaner, I’ll start by creating a variable that defines the boundaries for the categories: \n\n```\nage.breaks <- seq( from = 0, to = 60, by = 20 )\nage.breaks\n```\n\n `## [1] 0 20 40 60` \nand another one for the labels:\n\n```\nage.labels <- c( \"young\", \"adult\", \"older\" )\nage.labels\n```\n\n```\n## [1] \"young\" \"adult\" \"older\"\n```\n\n Note that there are four numbers in the `age.breaks` variable, but only three labels in the `age.labels` variable; I’ve done this because the `cut()` function requires that you specify the edges of the categories rather than the mid-points. In any case, now that we’ve done this, we can use the `cut()` function to assign each observation to one of these three categories. There are several arguments to the `cut()` function, but the three that we need to care about are: \n\n* \n `x` . The variable that needs to be categorised. * \n `breaks` . This is either a vector containing the locations of the breaks separating the categories, or a number indicating how many categories you want. * \n `labels` . The labels attached to the categories. This is optional: if you don’t specify this R will attach a boring label showing the range associated with each category. \nSince we’ve already created variables corresponding to the breaks and the labels, the command we need is just:\n\n```\nage.group <- cut( x = age, # the variable to be categorised\n breaks = age.breaks, # the edges of the categories\n labels = age.labels ) # the labels for the categories\n```\n\n Note that the output variable here is a factor. In order to see what this command has actually done, we could just print out the `age.group` variable, but I think it’s actually more helpful to create a data frame that includes both the original variable and the categorised one, so that you can see the two side by side: \n\n```\ndata.frame(age, age.group)\n```\n\n```\n## age age.group\n## 1 60 older\n## 2 58 older\n## 3 24 adult\n## 4 26 adult\n## 5 34 adult\n## 6 42 older\n## 7 31 adult\n## 8 30 adult\n## 9 33 adult\n## 10 2 young\n## 11 9 young\n```\n\nIt can also be useful to tabulate the output, just to see if you’ve got a nice even division of the sample:\n `table( age.group )` \n\n```\n## age.group\n## young adult older \n## 2 6 3\n```\n\n In the example above, I made all the decisions myself. Much like the `hist()` function that we saw in Chapter 6, if you want to you can delegate a lot of the choices to R. For instance, if you want you can specify the number of categories you want, rather than giving explicit ranges for them, and you can allow R to come up with some labels for the categories. To give you a sense of how this works, have a look at the following example: \n\n```\nage.group2 <- cut( x = age, breaks = 3 )\n```\n\n With this command, I’ve asked for three categories, but let R make the choices for where the boundaries should be. I won’t bother to print out the `age.group2` variable, because it’s not terribly pretty or very interesting. Instead, all of the important information can be extracted by looking at the tabulated data: `table( age.group2 )` \n\n```\n## age.group2\n## (1.94,21.3] (21.3,40.7] (40.7,60.1] \n## 2 6 3\n```\n\nThis output takes a little bit of interpretation, but it’s not complicated. What R has done is determined that the lowest age category should run from 1.94 years up to 21.3 years, the second category should run from 21.3 years to 40.7 years, and so on. The formatting on those labels might look a bit funny to those of you who haven’t studied a lot of maths, but it’s pretty simple. When R describes the first category as corresponding to the range \\((1.94, 21.3]\\) what it’s saying is that the range consists of those numbers that are larger than 1.94 but less than or equal to 21.3. In other words, the weird asymmetric brackets is R s way of telling you that if there happens to be a value that is exactly equal to 21.3, then it belongs to the first category, not the second one. Obviously, this isn’t actually possible since I’ve only specified the ages to the nearest whole number, but R doesn’t know this and so it’s trying to be precise just in case. This notation is actually pretty standard, but I suspect not everyone reading the book will have seen it before. In any case, those labels are pretty ugly, so it’s usually a good idea to specify your own, meaningful labels to the categories.\n Before moving on, I should take a moment to talk a little about the mechanics of the `cut()` function. Notice that R has tried to divide the `age` variable into three roughly equal sized bins. Unless you specify the particular breaks you want, that’s what it will do. But suppose you want to divide the `age` variable into three categories of different size, but with approximately identical numbers of people. How would you do that? Well, if that’s the case, then what you want to do is have the breaks correspond to the 0th, 33rd, 66th and 100th percentiles of the data. One way to do this would be to calculate those values using the `quantiles()` function and then use those quantiles as input to the `cut()` function. That’s pretty easy to do, but it does take a couple of lines to type. So instead, the `lsr` package has a function called `quantileCut()` that does exactly this: \n\n```\nage.group3 <- quantileCut( x = age, n = 3 )\ntable( age.group3 )\n```\n\n```\n## age.group3\n## (1.94,27.3] (27.3,33.7] (33.7,60.1] \n## 4 3 4\n```\n\n Notice the difference in the boundaries that the `quantileCut()` function selects. The first and third categories now span an age range of about 25 years each, whereas the middle category has shrunk to a span of only 6 years. There are some situations where this is genuinely what you want (that’s why I wrote the function!), but in general you should be careful. Usually the numeric variable that you’re trying to cut into categories is already expressed in meaningful units (i.e., it’s interval scale), but if you cut it into unequal bin sizes then it’s often very difficult to attach meaningful interpretations to the resulting categories. More generally, regardless of whether you’re using the original `cut()` function or the `quantileCut()` version, it’s important to take the time to figure out whether or not the resulting categories make any sense at all in terms of your research project. If they don’t make any sense to you as meaningful categories, then any data analysis that uses those categories is likely to be just as meaningless. More generally, in practice I’ve noticed that people have a very strong desire to carve their (continuous and messy) data into a few (discrete and simple) categories; and then run analysis using the categorised data instead of the original one.107 I wouldn’t go so far as to say that this is an inherently bad idea, but it does have some fairly serious drawbacks at times, so I would advise some caution if you are thinking about doing it. \n\n## 7.3 A few more mathematical functions and operations\n\nIn Section 7.2 I discussed the ideas behind variable transformations, and showed that a lot of the transformations that you might want to apply to your data are based on fairly simple mathematical functions and operations, of the kind that we discussed in Chapter 3. In this section I want to return to that discussion, and mention several other mathematical functions and arithmetic operations that I didn’t bother to mention when introducing you to R, but are actually quite useful for a lot of real world data analysis. Table 7.1 gives a brief overview of the various mathematical functions I want to talk about (and some that I already have talked about). Obviously this doesn’t even come close to cataloging the range of possibilities available in R, but it does cover a very wide range of functions that are used in day to day data analysis.\n\nmathematical.function | R.function | example.input | answer |\n | --- | --- | --- | --- |\nsquare root | sqrt() | sqrt(25) | 5 |\nabsolute value | abs() | abs(-23) | 23 |\nlogarithm (base 10) | log10() | log10(1000) | 3 |\nlogarithm (base e) | log() | log(1000) | 6.908 |\nexponentiation | exp() | exp(6.908) | 1000.245 |\nrounding to nearest | round() | round(1.32) | 1 |\nrounding down | floor() | floor(1.32) | 1 |\nrounding up | ceiling() | ceiling(1.32) | 2 |\n\n### 7.3.1 Rounding a number\n\n One very simple transformation that crops up surprisingly often is the need to round a number to the nearest whole number, or to a certain number of significant digits. To start with, let’s assume that we want to round to a whole number. To that end, there are three useful functions in R you want to know about: `round()` , `floor()` and `ceiling()` . The `round()` function just rounds to the nearest whole number. So if you round the number 4.3, it “rounds down” to `4` , like so: `round( x = 4.3 )` `## [1] 4` In contrast, if we want to round the number 4.7, we would round upwards to 5. In everyday life, when someone talks about “rounding”, they usually mean “round to nearest”, so this is the function we use most of the time. However sometimes you have reasons to want to always round up or always round down. If you want to always round down, use the `floor()` function instead; and if you want to force R to round up, then use `ceiling()` . That’s the only difference between the three functions. What if you want to round to a certain number of digits? Let’s suppose you want to round to a fixed number of decimal places, say 2 decimal places. If so, what you need to do is specify the `digits` argument to the `round()` function. That’s pretty straightforward: \n\n `## [1] 0.01` The only subtlety that you need to keep in mind is that sometimes what you want to do is round to 2 significant digits and not to two decimal places. The difference is that, when determining the number of significant digits, zeros don’t count. To see this, let’s apply the `signif()` function instead of the `round()` function: \n\n```\nsignif( x = 0.0123, digits = 2 )\n```\n\n `## [1] 0.012` \nThis time around, we get an answer of 0.012 because the zeros don’t count as significant digits. Quite often scientific journals will ask you to report numbers to two or three significant digits, so it’s useful to remember the distinction.\n\n### 7.3.2 Modulus and integer division\n\noperation | operator | example.input | answer |\n | --- | --- | --- | --- |\ninteger division | %/% | 42 %/% 10 | 4 |\nmodulus | %% | 42 %% 10 | 2 |\n\n Since we’re on the topic of simple calculations, there are two other arithmetic operations that I should mention, since they can come in handy when working with real data. These operations are calculating a modulus and doing integer division. They don’t come up anywhere else in this book, but they are worth knowing about. First, let’s consider integer division. Suppose I have $42 in my wallet, and want to buy some sandwiches, which are selling for $10 each. How many sandwiches can I afford108 to buy? The answer is of course 4. Note that it’s not 4.2, since no shop will sell me one-fifth of a sandwich. That’s integer division. In R we perform integer division by using the `%/%` operator: `42 %/% 10` `## [1] 4` Okay, that’s easy enough. What about the modulus? Basically, a modulus is the remainder after integer division, and it’s calculated using the `%%` operator. For the sake of argument, let’s suppose I buy four overpriced $10 sandwiches. If I started out with $42, how much money do I have left? The answer, as both R and common sense tells us, is $2: `42 %% 10` `## [1] 2` \nSo that’s also pretty easy. There is, however, one subtlety that I need to mention, and this relates to how negative numbers are handled. Firstly, what would happen if I tried to do integer division with a negative number? Let’s have a look:\n `-42 %/% 10` `## [1] -5` This might strike you as counterintuitive: why does `42 %/% 10` produce an answer of `4` , but `-42 %/% 10` gives us an answer of `-5` ? Intuitively you might think that the answer to the second one should be `-4` . The way to think about it is like this. Suppose I owe the sandwich shop $42, but I don’t have any money. How many sandwiches would I have to give them in order to stop them from calling security? The answer109 here is 5, not 4. If I handed them 4 sandwiches, I’d still owe them $2, right? So I actually have to give them 5 sandwiches. And since it’s me giving them the sandwiches, the answer to `-42 %/% 10` is `-5` . As you might expect, the behaviour of the modulus operator has a similar pattern. If I’ve handed 5 sandwiches over to the shop in order to pay off my debt of $42, then they now owe me $8. So the modulus is now: `-42 %% 10` `## [1] 8` \n\n### 7.3.3 Logarithms and exponentials\n\nAs I’ve mentioned earlier, R has an incredible range of mathematical functions built into it, and there really wouldn’t be much point in trying to describe or even list all of them. For the most part, I’ve focused only on those functions that are strictly necessary for this book. However I do want to make an exception for logarithms and exponentials. Although they aren’t needed anywhere else in this book, they are everywhere in statistics more broadly, and not only that, there are a lot of situations in which it is convenient to analyse the logarithm of a variable (i.e., to take a “log-transform” of the variable). I suspect that many (maybe most) readers of this book will have encountered logarithms and exponentials before, but from past experience I know that there’s a substantial proportion of students who take a social science statistics class who haven’t touched logarithms since high school, and would appreciate a bit of a refresher.\n In order to understand logarithms and exponentials, the easiest thing to do is to actually calculate them and see how they relate to other simple calculations. There are three R functions in particular that I want to talk about, namely `log()` , `log10()` and `exp()` . To start with, let’s consider `log10()` , which is known as the “logarithm in base 10”. The trick to understanding a logarithm is to understand that it’s basically the “opposite” of taking a power. Specifically, the logarithm in base 10 is closely related to the powers of 10. So let’s start by noting that 10-cubed is 1000. Mathematically, we would write this: \\[ \n10^3 = 1000\n\\] and in R we’d calculate it by using the command `10^3` . The trick to understanding a logarithm is to recognise that the statement that “10 to the power of 3 is equal to 1000” is equivalent to the statement that “the logarithm (in base 10) of 1000 is equal to 3”. Mathematically, we write this as follows, \\[\n\\log_{10}( 1000 ) = 3\n\\] and if we wanted to do the calculation in R we would type this: `log10( 1000 )` `## [1] 3` \nObviously, since you already know that \\(10^3 = 1000\\) there’s really no point in getting R to tell you that the base-10 logarithm of 1000 is 3. However, most of the time you probably don’t know what right answer is. For instance, I can honestly say that I didn’t know that \\(10^{2.69897} = 500\\), so it’s rather convenient for me that I can use R to calculate the base-10 logarithm of 500:\n `log10( 500 )` `## [1] 2.69897` \nOr at least it would be convenient if I had a pressing need to know the base-10 logarithm of 500.\n Okay, since the `log10()` function is related to the powers of 10, you might expect that there are other logarithms (in bases other than 10) that are related to other powers too. And of course that’s true: there’s not really anything mathematically special about the number 10. You and I happen to find it useful because decimal numbers are built around the number 10, but the big bad world of mathematics scoffs at our decimal numbers. Sadly, the universe doesn’t actually care how we write down numbers. Anyway, the consequence of this cosmic indifference is that there’s nothing particularly special about calculating logarithms in base 10. You could, for instance, calculate your logarithms in base 2, and in fact R does provide a function for doing that, which is (not surprisingly) called `log2()` . Since we know that \\(2^3 = 2 \\times 2 \\times 2 = 8\\), it’s not surprise to see that `log2( 8 )` `## [1] 3` Alternatively, a third type of logarithm – and one we see a lot more of in statistics than either base 10 or base 2 – is called the natural logarithm, and corresponds to the logarithm in base \\(e\\). Since you might one day run into it, I’d better explain what \\(e\\) is. The number \\(e\\), known as Euler’s number, is one of those annoying “irrational” numbers whose decimal expansion is infinitely long, and is considered one of the most important numbers in mathematics. The first few digits of \\(e\\) are: \\[ e = 2.718282 \\] There are quite a few situation in statistics that require us to calculate powers of \\(e\\), though none of them appear in this book. Raising \\(e\\) to the power \\(x\\) is called the exponential of \\(x\\), and so it’s very common to see \\(e^x\\) written as \\(\\exp(x)\\). And so it’s no surprise that R has a function that calculate exponentials, called `exp()` . For instance, suppose I wanted to calculate \\(e^3\\). I could try typing in the value of \\(e\\) manually, like this: `2.718282 ^ 3` `## [1] 20.08554` but it’s much easier to do the same thing using the `exp()` function: `exp( 3 )` `## [1] 20.08554` Anyway, because the number \\(e\\) crops up so often in statistics, the natural logarithm (i.e., logarithm in base \\(e\\)) also tends to turn up. Mathematicians often write it as \\(\\log_e(x)\\) or \\(\\ln(x)\\), or sometimes even just \\(\\log(x)\\). In fact, R works the same way: the `log()` function corresponds to the natural logarithm110 Anyway, as a quick check, let’s calculate the natural logarithm of 20.08554 using R: `log( 20.08554 )` `## [1] 3` \nAnd with that, I think we’ve had quite enough exponentials and logarithms for this book!\n\n## 7.4 Extracting a subset of a vector\n\nOne very important kind of data handling is being able to extract a particular subset of the data. For instance, you might be interested only in analysing the data from one experimental condition, or you may want to look closely at the data from people over 50 years in age. To do this, the first step is getting R to extract the subset of the data corresponding to the observations that you’re interested in. In this section I’ll talk about subsetting as it applies to vectors, extending the discussion from Chapters 3 and 4. In Section 7.5 I’ll go on to talk about how this discussion extends to data frames.\n\n### 7.4.1 Refresher\n\n This section returns to the `nightgarden.Rdata` data set. If you’re reading this whole chapter in one sitting, then you should already have this data set loaded. If not, don’t forget to use the \n\n```\nload(\"nightgarden.Rdata\")\n```\n\ncommand. For this section, let’s ignore the `itng` data frame that we created earlier, and focus instead on the two vectors `speaker` and `utterance` (see Section 7.1 if you’ve forgotten what those vectors look like). Suppose that what I want to do is pull out only those utterances that were made by Makka-Pakka. To that end, I could first use the equality operator to have R tell me which cases correspond to Makka-Pakka speaking: \n\n```\nis.MP.speaking <- speaker == \"makka-pakka\"\nis.MP.speaking\n```\n\n and then use logical indexing to get R to print out those elements of `utterance` for which `is.MP.speaking` is true, like so: \n\n```\nutterance[ is.MP.speaking ]\n```\n\nOr, since I’m lazy, I could collapse it to a single command like so:\n\n```\nutterance[ speaker == \"makka-pakka\" ]\n```\n\n### 7.4.2 Using\n\n `%in%` to match multiple cases A second useful trick to be aware of is the `%in%` operator111. It’s actually very similar to the `==` operator, except that you can supply a collection of acceptable values. For instance, suppose I wanted to keep only those cases when the utterance is either “pip” or “oo”. One simple way do to this is: \n\n```\nutterance %in% c(\"pip\",\"oo\")\n```\n\n What this does if return `TRUE` for those elements of `utterance` that are either `\"pip\"` or `\"oo\"` and returns `FALSE` for all the others. What that means is that if I want a list of all those instances of characters speaking either of these two words, I could do this: \n\n```\nspeaker[ utterance %in% c(\"pip\",\"oo\") ]\n```\n\n```\n## [1] \"upsy-daisy\" \"upsy-daisy\" \"tombliboo\" \"makka-pakka\" \"makka-pakka\"\n```\n\n### 7.4.3 Using negative indices to drop elements\n\n Before moving onto data frames, there’s a couple of other tricks worth mentioning. The first of these is to use negative values as indices. Recall from Section 3.10 that we can use a vector of numbers to extract a set of elements that we would like to keep. For instance, suppose I want to keep only elements 2 and 3 from `utterance` . I could do so like this: `utterance[2:3]` `## [1] \"pip\" \"onk\"` \nBut suppose, on the other hand, that I have discovered that observations 2 and 3 are untrustworthy, and I want to keep everything except those two elements. To that end, R lets you use negative numbers to remove specific values, like so:\n `utterance [ -(2:3) ]` \n\nThe output here corresponds to element 1 of the original vector, followed by elements 4, 5, and so on. When all you want to do is remove a few cases, this is a very handy convention.\n\n### 7.4.4 Splitting a vector by group\n\n One particular example of subsetting that is especially common is the problem of splitting one one variable up into several different variables, one corresponding to each group. For instance, in our In the Night Garden example, I might want to create subsets of the `utterance` variable for every character. One way to do this would be to just repeat the exercise that I went through earlier separately for each character, but that quickly gets annoying. A faster way do it is to use the `split()` function. The arguments are: \n\n* \n `x` . The variable that needs to be split into groups. * \n `f` . The grouping variable. What this function does is output a list (Section 4.9), containing one variable for each group. For instance, I could split up the `utterance` variable by `speaker` using the following command: \n\n```\nspeech.by.char <- split( x = utterance, f = speaker )\nspeech.by.char\n```\n\n```\n## $`makka-pakka`\n## [1] \"pip\" \"pip\" \"onk\" \"onk\"\n## \n## $tombliboo\n## [1] \"ee\" \"oo\"\n## \n## $`upsy-daisy`\n## [1] \"pip\" \"pip\" \"onk\" \"onk\"\n```\n\nOnce you’re starting to become comfortable working with lists and data frames, this output is all you need, since you can work with this list in much the same way that you would work with a data frame. For instance, if you want the first utterance made by Makka-Pakka, all you need to do is type this:\n\n```\nspeech.by.char$`makka-pakka`[1]\n```\n\n `## [1] \"pip\"` Just remember that R does need you to add the quoting characters (i.e. `'` ). Otherwise, there’s nothing particularly new or difficult here. However, sometimes – especially when you’re just starting out – it can be convenient to pull these variables out of the list, and into the workspace. This isn’t too difficult to do, though it can be a little daunting to novices. To that end, I’ve included a function called `importList()` in the `lsr` package that does this.112 First, here’s what you’d have if you had wiped the workspace before the start of this section: `who()` \n\n Now we use the `importList()` function to copy all of the variables within the `speech.by.char` list: \n\n```\nimportList( speech.by.char, ask = FALSE)\n```\n\n Because the `importList()` function is attempting to create new variables based on the names of the elements of the list, it pauses to check that you’re okay with the variable names. The reason it does this is that, if one of the to-be-created variables has the same name as a variable that you already have in your workspace, that variable will end up being overwritten, so it’s a good idea to check. Assuming that you type `y` , it will go on to create the variables. Nothing appears to have happened, but if we look at our workspace now: `who()` \n\n we see that there are three new variables, called `makka.pakka` , `tombliboo` and `upsy.daisy` . Notice that the `importList()` function has converted the original character strings into valid R variable names, so the variable corresponding to `\"makka-pakka\"` is actually `makka.pakka` .113 Nevertheless, even though the names can change, note that each of these variables contains the exact same information as the original elements of the list did. For example: \n\n```\n> makka.pakka\n[1] \"pip\" \"pip\" \"onk\" \"onk\"\n```\n\n## 7.5 Extracting a subset of a data frame\n\n In this section we turn to the question of how to subset a data frame rather than a vector. To that end, the first thing I should point out is that, if all you want to do is subset one of the variables inside the data frame, then as usual the `$` operator is your friend. For instance, suppose I’m working with the `itng` data frame, and what I want to do is create the `speech.by.char` list. I can use the exact same tricks that I used last time, since what I really want to do is `split()` the `itng$utterance` vector, using the `itng$speaker` vector as the grouping variable. However, most of the time what you actually want to do is select several different variables within the data frame (i.e., keep only some of the columns), or maybe a subset of cases (i.e., keep only some of the rows). In order to understand how this works, we need to talk more specifically about data frames and how to subset them. \n\n### 7.5.1 Using the\n\n `subset()` function There are several different ways to subset a data frame in R, some easier than others. I’ll start by discussing the `subset()` function, which is probably the conceptually simplest way do it. For our purposes there are three different arguments that you’ll be most interested in: \n\n* \n `x` . The data frame that you want to subset. * \n `subset` . A vector of logical values indicating which cases (rows) of the data frame you want to keep. By default, all cases will be retained. * \n `select` . This argument indicates which variables (columns) in the data frame you want to keep. This can either be a list of variable names, or a logical vector indicating which ones to keep, or even just a numeric vector containing the relevant column numbers. By default, all variables will be retained. Let’s start with an example in which I use all three of these arguments. Suppose that I want to subset the `itng` data frame, keeping only the utterances made by Makka-Pakka. What that means is that I need to use the `select` argument to pick out the `utterance` variable, and I also need to use the `subset` variable, to pick out the cases when Makka-Pakka is speaking (i.e., \n\n```\nspeaker == \"makka-pakka\"\n```\n\n). Therefore, the command I need to use is this: \n\n```\ndf <- subset( x = itng, # data frame is itng\n subset = speaker == \"makka-pakka\", # keep only Makka-Pakkas speech\n select = utterance ) # keep only the utterance variable\nprint( df )\n```\n\n The variable `df` here is still a data frame, but it only contains one variable (called `utterance` ) and four cases. Notice that the row numbers are actually the same ones from the original data frame. It’s worth taking a moment to briefly explain this. The reason that this happens is that these “row numbers’ are actually row names. When you create a new data frame from scratch R will assign each row a fairly boring row name, which is identical to the row number. However, when you subset the data frame, each row keeps its original row name. This can be quite useful, since – as in the current example – it provides you a visual reminder of what each row in the new data frame corresponds to in the original data frame. However, if it annoys you, you can change the row names using the `rownames()` function.114 In any case, let’s return to the `subset()` function, and look at what happens when we don’t use all three of the arguments. Firstly, suppose that I didn’t bother to specify the `select` argument. Let’s see what happens: \n\n```\nsubset( x = itng,\n subset = speaker == \"makka-pakka\" )\n```\n\n Not surprisingly, R has kept the same cases from the original data set (i.e., rows 7 through 10), but this time it has kept all of the variables from the data frame. Equally unsurprisingly, if I don’t specify the `subset` argument, what we find is that R keeps all of the cases: \n\n```\nsubset( x = itng, \n select = utterance )\n```\n\nAgain, it’s important to note that this output is still a data frame: it’s just a data frame with only a single variable.\n\n### 7.5.2 Using square brackets: I. Rows and columns\n\n Throughout the book so far, whenever I’ve been subsetting a vector I’ve tended use the square brackets `[]` to do so. But in the previous section when I started talking about subsetting a data frame I used the `subset()` function. As a consequence, you might be wondering whether it is possible to use the square brackets to subset a data frame. The answer, of course, is yes. Not only can you use square brackets for this purpose, as you become more familiar with R you’ll find that this is actually much more convenient than using `subset()` . Unfortunately, the use of square brackets for this purpose is somewhat complicated, and can be very confusing to novices. So be warned: this section is more complicated than it feels like it “should” be. With that warning in place, I’ll try to walk you through it slowly. For this section, I’ll use a slightly different data set, namely the `garden` data frame that is stored in the `\"nightgarden2.Rdata\"` file. \n\n```\nload(file.path(projecthome,\"data\",\"nightgarden2.Rdata\"))\ngarden\n```\n\n As you can see, the `garden` data frame contains 3 variables and 5 cases, and this time around I’ve used the `rownames()` function to attach slightly verbose labels to each of the cases. Moreover, let’s assume that what we want to do is to pick out rows 4 and 5 (the two cases when Makka-Pakka is speaking), and columns 1 and 2 (variables `speaker` and `utterance` ). \nHow shall we do this? As usual, there’s more than one way. The first way is based on the observation that, since a data frame is basically a table, every element in the data frame has a row number and a column number. So, if we want to pick out a single element, we have to specify the row number and a column number within the square brackets. By convention, the row number comes first. So, for the data frame above, which has 5 rows and 3 columns, the numerical indexing scheme looks like this:\n\n```\nknitr::kable(data.frame(stringsAsFactors=FALSE, row = c(\"1\",\"2\",\"3\", \"4\", \"5\"), col1 = c(\"[1,1]\", \"[2,1]\", \"[3,1]\", \"[4,1]\", \"[5,1]\"), col2 = c(\"[1,2]\", \"[2,2]\", \"[3,2]\", \"[4,2]\", \"[5,2]\"), col3 = c(\"[1,3]\", \"[2,3]\", \"[3,3]\", \"[4,3]\", \"[5,3]\")))\n```\n\nrow | col1 | col2 | col3 |\n | --- | --- | --- | --- |\n1 | [1,1] | [1,2] | [1,3] |\n2 | [2,1] | [2,2] | [2,3] |\n3 | [3,1] | [3,2] | [3,3] |\n4 | [4,1] | [4,2] | [4,3] |\n5 | [5,1] | [5,2] | [5,3] |\n\n If I want the 3rd case of the 2nd variable, what I would type is `garden[3,2]` , and R would print out some output showing that, this element corresponds to the utterance `\"ee\"` . However, let’s hold off from actually doing that for a moment, because there’s something slightly counterintuitive about the specifics of what R does under those circumstances (see Section 7.5.4). Instead, let’s aim to solve our original problem, which is to pull out two rows (4 and 5) and two columns (1 and 2). This is fairly simple to do, since R allows us to specify multiple rows and multiple columns. So let’s try that: `garden[ 4:5, 1:2 ]` \n\n Clearly, that’s exactly what we asked for: the output here is a data frame containing two variables and two cases. Note that I could have gotten the same answer if I’d used the `c()` function to produce my vectors rather than the `:` operator. That is, the following command is equivalent to the last one: \n\n```\ngarden[ c(4,5), c(1,2) ]\n```\n\n It’s just not as pretty. However, if the columns and rows that you want to keep don’t happen to be next to each other in the original data frame, then you might find that you have to resort to using commands like\n\n```\ngarden[ c(2,4,5), c(1,3) ]\n```\n\nto extract them. A second way to do the same thing is to use the names of the rows and columns. That is, instead of using the row numbers and column numbers, you use the character strings that are used as the labels for the rows and columns. To apply this idea to our `garden` data frame, we would use a command like this: \n\n```\ngarden[ c(\"case.4\", \"case.5\"), c(\"speaker\", \"utterance\") ]\n```\n\n Once again, this produces exactly the same output, so I haven’t bothered to show it. Note that, although this version is more annoying to type than the previous version, it’s a bit easier to read, because it’s often more meaningful to refer to the elements by their names rather than their numbers. Also note that you don’t have to use the same convention for the rows and columns. For instance, I often find that the variable names are meaningful and so I sometimes refer to them by name, whereas the row names are pretty arbitrary so it’s easier to refer to them by number. In fact, that’s more or less exactly what’s happening with the `garden` data frame, so it probably makes more sense to use this as the command: \n\n```\ngarden[ 4:5, c(\"speaker\", \"utterance\") ]\n```\n\nAgain, the output is identical.\n\nFinally, both the rows and columns can be indexed using logicals vectors as well. For example, although I claimed earlier that my goal was to extract cases 4 and 5, it’s pretty obvious that what I really wanted to do was select the cases where Makka-Pakka is speaking. So what I could have done is create a logical vector that indicates which cases correspond to Makka-Pakka speaking:\n\n```\nis.MP.speaking <- garden$speaker == \"makka-pakka\"\nis.MP.speaking\n```\n\n As you can see, the 4th and 5th elements of this vector are `TRUE` while the others are `FALSE` . Now that I’ve constructed this “indicator” variable, what I can do is use this vector to select the rows that I want to keep: \n\n```\ngarden[ is.MP.speaking, c(\"speaker\", \"utterance\") ]\n```\n\nAnd of course the output is, yet again, the same.\n\n### 7.5.3 Using square brackets: II. Some elaborations\n\n There are two fairly useful elaborations on this “rows and columns” approach that I should point out. Firstly, what if you want to keep all of the rows, or all of the columns? To do this, all we have to do is leave the corresponding entry blank, but it is crucial to remember to keep the comma*! For instance, suppose I want to keep all the rows in the `garden` data, but I only want to retain the first two columns. The easiest way do this is to use a command like this: `garden[ , 1:2 ]` \n\nAlternatively, if I want to keep all the columns but only want the last two rows, I use the same trick, but this time I leave the second index blank. So my command becomes:\n `garden[ 4:5, ]` \n\n```\n## speaker utterance line\n## case.4 makka-pakka pip 7\n## case.5 makka-pakka onk 9\n```\n\nThe second elaboration I should note is that it’s still okay to use negative indexes as a way of telling R to delete certain rows or columns. For instance, if I want to delete the 3rd column, then I use this command:\n `garden[ , -3 ]` \n\nwhereas if I want to delete the 3rd row, then I’d use this one:\n `garden[ -3, ]` \n\nSo that’s nice.\n\n### 7.5.4 Using square brackets: III. Understanding “dropping”\n\nAt this point some of you might be wondering why I’ve been so terribly careful to choose my examples in such a way as to ensure that the output always has are multiple rows and multiple columns. The reason for this is that I’ve been trying to hide the somewhat curious “dropping” behaviour that R produces when the output only has a single column. I’ll start by showing you what happens, and then I’ll try to explain it. Firstly, let’s have a look at what happens when the output contains only a single row:\n `garden[ 5, ]` \n\n```\n## speaker utterance line\n## case.5 makka-pakka onk 9\n```\n\nThis is exactly what you’d expect to see: a data frame containing three variables, and only one case per variable. Okay, no problems so far. What happens when you ask for a single column? Suppose, for instance, I try this as a command:\n `garden[ , 3 ]` Based on everything that I’ve shown you so far, you would be well within your rights to expect to see R produce a data frame containing a single variable (i.e., `line` ) and five cases. After all, that is what the `subset()` command does in this situation, and it’s pretty consistent with everything else that I’ve shown you so far about how square brackets work. In other words, you should expect to see this: \n\nHowever, that is emphatically not what happens. What you actually get is this:\n `garden[ , 3 ]` `## [1] 1 2 5 7 9` That output is not a data frame at all! That’s just an ordinary numeric vector containing 5 elements. What’s going on here is that R has “noticed” that the output that we’ve asked for doesn’t really “need” to be wrapped up in a data frame at all, because it only corresponds to a single variable. So what it does is “drop” the output from a data frame containing a single variable, “down” to a simpler output that corresponds to that variable. This behaviour is actually very convenient for day to day usage once you’ve become familiar with it – and I suppose that’s the real reason why R does this – but there’s no escaping the fact that it is deeply confusing to novices. It’s especially confusing because the behaviour appears only for a very specific case: (a) it only works for columns and not for rows, because the columns correspond to variables and the rows do not, and (b) it only applies to the “rows and columns” version of the square brackets, and not to the `subset()` function,115 or to the “just columns” use of the square brackets (next section). As I say, it’s very confusing when you’re just starting out. For what it’s worth, you can suppress this behaviour if you want, by setting `drop = FALSE` when you construct your bracketed expression. That is, you could do something like this: \n\n```\ngarden[ , 3, drop = FALSE ]\n```\n\nI suppose that helps a little bit, in that it gives you some control over the dropping behaviour, but I’m not sure it helps to make things any easier to understand. Anyway, that’s the “dropping” special case. Fun, isn’t it?\n\n### 7.5.5 Using square brackets: IV. Columns only\n\nAs if the weird “dropping” behaviour wasn’t annoying enough, R actually provides a completely different way of using square brackets to index a data frame. Specifically, if you only give a single index, R will assume you want the corresponding columns, not the rows. Do not be fooled by the fact that this second method also uses square brackets: it behaves differently to the “rows and columns” method that I’ve discussed in the last few sections. Again, what I’ll do is show you what happens first, and then I’ll try to explain why it happens afterwards. To that end, let’s start with the following command:\n `garden[ 1:2 ]` \n\n As you can see, the output gives me the first two columns, much as if I’d typed `garden[,1:2]` . It doesn’t give me the first two rows, which is what I’d have gotten if I’d used a command like `garden[1:2,]` . Not only that, if I ask for a single column, R does not drop the output: `garden[3]` \n\nAs I said earlier, the only case where dropping occurs by default is when you use the “row and columns” version of the square brackets, and the output happens to correspond to a single column. However, if you really want to force R to drop the output, you can do so using the “double brackets” notation:\n `garden[[3]]` `## [1] 1 2 5 7 9` Note that R will only allow you to ask for one column at a time using the double brackets. If you try to ask for multiple columns in this way, you get completely different behaviour,116 which may or may not produce an error, but definitely won’t give you the output you’re expecting. The only reason I’m mentioning it at all is that you might run into double brackets when doing further reading, and a lot of books don’t explicitly point out the difference between `[` and `[[` . However, I promise that I won’t be using `[[` anywhere else in this book. Okay, for those few readers that have persevered with this section long enough to get here without having set fire to the book, I should explain why R has these two different systems for subsetting a data frame (i.e., “row and column” versus “just columns”), and why they behave so differently to each other. I’m not 100% sure about this since I’m still reading through some of the old references that describe the early development of R, but I think the answer relates to the fact that data frames are actually a very strange hybrid of two different kinds of thing. At a low level, a data frame is a list (Section 4.9). I can demonstrate this to you by overriding the normal `print()` function117 and forcing R to print out the `garden` data frame using the default print method rather than the special one that is defined only for data frames. Here’s what we get: \n\n```\nprint.default( garden )\n```\n\n```\n## $speaker\n## [1] upsy-daisy upsy-daisy tombliboo makka-pakka makka-pakka\n## Levels: makka-pakka tombliboo upsy-daisy\n## \n## $utterance\n## [1] pip pip ee pip onk\n## Levels: ee onk oo pip\n## \n## $line\n## [1] 1 2 5 7 9\n## \n## attr(,\"class\")\n## [1] \"data.frame\"\n```\n\n Apart from the weird part of the output right at the bottom, this is identical to the print out that you get when you print out a list (see Section 4.9). In other words, a data frame is a list. View from this “list based” perspective, it’s clear what `garden[1]` is: it’s the first variable stored in the list, namely `speaker` . In other words, when you use the “just columns” way of indexing a data frame, using only a single index, R assumes that you’re thinking about the data frame as if it were a list of variables. In fact, when you use the `$` operator you’re taking advantage of the fact that the data frame is secretly a list. \nHowever, a data frame is more than just a list. It’s a very special kind of list where all the variables are of the same length, and the first element in each variable happens to correspond to the first “case” in the data set. That’s why no-one ever wants to see a data frame printed out in the default “list-like” way that I’ve shown in the extract above. In terms of the deeper meaning behind what a data frame is used for, a data frame really does have this rectangular shape to it:\n `print( garden )` \n\n Because of the fact that a data frame is basically a table of data, R provides a second “row and column” method for interacting with the data frame (see Section 7.11.1 for a related example). This method makes much more sense in terms of the high-level table of data interpretation of what a data frame is, and so for the most part it’s this method that people tend to prefer. In fact, throughout the rest of the book I will be sticking to the “row and column” approach (though I will use `$` a lot), and never again referring to the “just columns” approach. However, it does get used a lot in practice, so I think it’s important that this book explain what’s going on. \nAnd now let us never speak of this again.\n\n## 7.6 Sorting, flipping and merging data\n\nIn this section I discuss a few useful operations that I feel are loosely related to one another: sorting a vector, sorting a data frame, binding two or more vectors together into a data frame (or matrix), and flipping a data frame (or matrix) on its side. They’re all fairly straightforward tasks, at least in comparison to some of the more obnoxious data handling problems that turn up in real life.\n\n### 7.6.1 Sorting a numeric or character vector\n\n One thing that you often want to do is sort a variable. If it’s a numeric variable you might want to sort in increasing or decreasing order. If it’s a character vector you might want to sort alphabetically, etc. The `sort()` function provides this capability. \n\n```\nnumbers <- c(2,4,3)\nsort( x = numbers )\n```\n\n `## [1] 2 3 4` \nYou can ask for R to sort in decreasing order rather than increasing:\n\n```\nsort( x = numbers, decreasing = TRUE )\n```\n\n `## [1] 4 3 2` \nAnd you can ask it to sort text data in alphabetical order:\n\n```\ntext <- c(\"aardvark\", \"zebra\", \"swing\")\nsort( text )\n```\n\n That’s pretty straightforward. That being said, it’s important to note that I’m glossing over something here. When you apply `sort()` to a character vector it doesn’t strictly sort into alphabetical order. R actually has a slightly different notion of how characters are ordered (see Section 7.8.5 and Table 7.3), which is more closely related to how computers store text data than to how letters are ordered in the alphabet. However, that’s a topic we’ll discuss later. For now, the only thing I should note is that the `sort()` function doesn’t alter the original variable. Rather, it creates a new, sorted variable as the output. So if I inspect my original `text` variable: `text` \n\nI can see that it has remained unchanged.\n\n### 7.6.2 Sorting a factor\n\n You can also sort factors, but the story here is slightly more subtle because there’s two different ways you can sort a factor: alphabetically (by label) or by factor level. The `sort()` function uses the latter. To illustrate, let’s look at the two different examples. First, let’s create a factor in the usual way: \n\n```\nfac <- factor( text )\nfac\n```\n\nNow let’s sort it:\n `sort(fac)` \n\nThis looks like it’s sorted things into alphabetical order, but that’s only because the factor levels themselves happen to be alphabetically ordered. Suppose I deliberately define the factor levels in a non-alphabetical order:\n\n```\nfac <- factor( text, levels = c(\"zebra\",\"swing\",\"aardvark\") )\nfac\n```\n\n Now what happens when we try to sort `fac` this time? The answer: `sort(fac)` \n\nIt sorts the data into the numerical order implied by the factor levels, not the alphabetical order implied by the labels attached to those levels. Normally you never notice the distinction, because by default the factor levels are assigned in alphabetical order, but it’s important to know the difference:\n\n### 7.6.3 Sorting a data frame\n\n The `sort()` function doesn’t work properly with data frames. If you want to sort a data frame the standard advice that you’ll find online is to use the `order()` function (not described in this book) to determine what order the rows should be sorted, and then use square brackets to do the shuffling. There’s nothing inherently wrong with this advice, I just find it tedious. To that end, the `lsr` package includes a function called `sortFrame()` that you can use to do the sorting. The first argument to the function is named ( `x` ), and should correspond to the data frame that you want sorted. After that, all you do is type a list of the names of the variables that you want to use to do the sorting. For instance, if I type this: \n\n```\nsortFrame( garden, speaker, line)\n```\n\n what R does is first sort by `speaker` (factor level order). Any ties (i.e., data from the same speaker) are then sorted in order of `line` (increasing numerical order). You can use the minus sign to indicate that numerical variables should be sorted in reverse order: \n\n```\nsortFrame( garden, speaker, -line)\n```\n\n As of the current writing, the `sortFrame()` function is under development. I’ve started introducing functionality to allow you to use the `-` sign to non-numeric variables or to make a distinction between sorting factors alphabetically or by factor level. The idea is that you should be able to type in something like this: \n\n and have the output correspond to a sort of the `garden` data frame in reverse alphabetical order (or reverse factor level order) of `speaker` . As things stand right now, this will actually work, and it will produce sensible output: \n\nHowever, I’m not completely convinced that I’ve set this up in the ideal fashion, so this may change a little bit in the future.\n\n### 7.6.4 Binding vectors together\n\nA not-uncommon task that you might find yourself needing to undertake is to combine several vectors. For instance, let’s suppose we have the following two numeric vectors:\n\n```\ncake.1 <- c(100, 80, 0, 0, 0)\ncake.2 <- c(100, 100, 90, 30, 10)\n```\n\n The numbers here might represent the amount of each of the two cakes that are left at five different time points. Apparently the first cake is tastier, since that one gets devoured faster. We’ve already seen one method for combining these vectors: we could use the `data.frame()` function to convert them into a data frame with two variables, like so: \n\n```\ncake.df <- data.frame( cake.1, cake.2 )\ncake.df\n```\n\n```\n## cake.1 cake.2\n## 1 100 100\n## 2 80 100\n## 3 0 90\n## 4 0 30\n## 5 0 10\n```\n\n Two other methods that I want to briefly refer to are the `rbind()` and `cbind()` functions, which will convert the vectors into a matrix. I’ll discuss matrices properly in Section 7.11.1 but the details don’t matter too much for our current purposes. The `cbind()` function (“column bind”) produces a very similar looking output to the data frame example: \n\n```\ncake.mat1 <- cbind( cake.1, cake.2 )\ncake.mat1\n```\n\n```\n## cake.1 cake.2\n## [1,] 100 100\n## [2,] 80 100\n## [3,] 0 90\n## [4,] 0 30\n## [5,] 0 10\n```\n\n but nevertheless it’s important to keep in mind that `cake.mat1` is a matrix rather than a data frame, and so has a few differences from the `cake.df` variable. The `rbind()` function (“row bind”) produces a somewhat different output: it binds the vectors together row-wise rather than column-wise, so the output now looks like this: \n\n```\ncake.mat2 <- rbind( cake.1, cake.2 )\ncake.mat2\n```\n\n```\n## [,1] [,2] [,3] [,4] [,5]\n## cake.1 100 80 0 0 0\n## cake.2 100 100 90 30 10\n```\n\n You can add names to a matrix by using the `rownames()` and `colnames()` functions, and I should also point out that there’s a fancier function in R called `merge()` that supports more complicated “database like” merging of vectors and data frames, but I won’t go into details here. \n\n### 7.6.5 Binding multiple copies of the same vector together\n\n It is sometimes very useful to bind together multiple copies of the same vector. You could do this using the `rbind` and `cbind` functions, using comands like this one \n\n```\nfibonacci <- c( 1,1,2,3,5,8 )\nrbind( fibonacci, fibonacci, fibonacci )\n```\n\n```\n## [,1] [,2] [,3] [,4] [,5] [,6]\n## fibonacci 1 1 2 3 5 8\n## fibonacci 1 1 2 3 5 8\n## fibonacci 1 1 2 3 5 8\n```\n\n but that can be pretty annoying, especially if you needs lots of copies. To make this a little easier, the `lsr` package has two additional functions `rowCopy` and `colCopy` that do the same job, but all you have to do is specify the number of copies that you want, instead of typing the name in over and over again. The two arguments you need to specify are `x` , the vector to be copied, and `times` , indicating how many copies should be created:118 \n\n```\nrowCopy( x = fibonacci, times = 3 )\n```\n\n```\n## [,1] [,2] [,3] [,4] [,5] [,6]\n## [1,] 1 1 2 3 5 8\n## [2,] 1 1 2 3 5 8\n## [3,] 1 1 2 3 5 8\n```\n\n Of course, in practice you don’t need to name the arguments all the time. For instance, here’s an example using the `colCopy()` function with the argument names omitted: \n\n```\ncolCopy( fibonacci, 3 )\n```\n\n```\n## [,1] [,2] [,3]\n## [1,] 1 1 1\n## [2,] 1 1 1\n## [3,] 2 2 2\n## [4,] 3 3 3\n## [5,] 5 5 5\n## [6,] 8 8 8\n```\n\n### 7.6.6 Transposing a matrix or data frame\n\n One of the main reasons that I wanted to discuss the `rbind()` and `cbind()` functions in the same section as the `data.frame()` function is that it immediately raises the question of how to “flip” or transpose a matrix or data frame. Notice that in the last section I was able to produce two different matrices, `cake.mat1` and `cake.mat2` that were basically mirror images of one another. A natural question to ask is whether you can directly transform one into another. The transpose function `t()` allows us to do this in a straightforward fashion. To start with, I’ll show you how to transpose a matrix, and then I’ll move onto talk about data frames. Firstly, let’s load a matrix I prepared earlier, from the `cakes.Rdata` file: \n\n```\nload(file.path(projecthome,\"data\",\"cakes.Rdata\"))\ncakes\n```\n\nAnd just to make sure you believe me that this is actually a matrix:\n `class( cakes )` `## [1] \"matrix\"` \nOkay, now let’s transpose the matrix:\n\n```\ncakes.flipped <- t( cakes )\ncakes.flipped\n```\n\nThe output here is still a matrix:\n\n```\nclass( cakes.flipped )\n```\n\n `## [1] \"matrix\"` At this point you should have two questions: (1) how do we do the same thing for data frames? and (2) why should we care about this? Let’s start with the how question. First, I should note that you can transpose a data frame just fine using the `t()` function, but that has the slightly awkward consequence of converting the output from a data frame to a matrix, which isn’t usually what you want. It’s quite easy to convert the output back again, of course,119 but I hate typing two commands when I can do it with one. To that end, the `lsr` package has a simple “convenience” function called `tFrame()` which does exactly the same thing as `t()` but converts the output to a data frame for you. To illustrate this, let’s transpose the `itng` data frame that we used earlier. Here’s the original data frame: `itng` \n\n and here’s what happens when you transpose it using `tFrame()` : `tFrame( itng )` \n\n```\n## V1 V2 V3 V4 V5 V6\n## speaker upsy-daisy upsy-daisy upsy-daisy upsy-daisy tombliboo tombliboo\n## utterance pip pip onk onk ee oo\n## V7 V8 V9 V10\n## speaker makka-pakka makka-pakka makka-pakka makka-pakka\n## utterance pip pip onk onk\n```\n\n An important point to recognise is that transposing a data frame is not always a sensible thing to do: in fact, I’d go so far as to argue that it’s usually not sensible. It depends a lot on whether the “cases” from your original data frame would make sense as variables, and to think of each of your original “variables” as cases. I think that’s emphatically not true for our `itng` data frame, so I wouldn’t advise doing it in this situation. That being said, sometimes it really is true. For instance, had we originally stored our `cakes` variable as a data frame instead of a matrix, then it would absolutely be sensible to flip the data frame!120 There are some situations where it is useful to flip your data frame, so it’s nice to know that you can do it. Indeed, that’s the main reason why I have spent so much time talking about this topic. A lot of statistical tools make the assumption that the rows of your data frame (or matrix) correspond to observations, and the columns correspond to the variables. That’s not unreasonable, of course, since that is a pretty standard convention. However, think about our `cakes` example here. This is a situation where you might want do an analysis of the different cakes (i.e. cakes as variables, time points as cases), but equally you might want to do an analysis where you think of the times as being the things of interest (i.e., times as variables, cakes as cases). If so, then it’s useful to know how to flip a matrix or data frame around. \n\n## 7.7 Reshaping a data frame\n\n One of the most annoying tasks that you need to undertake on a regular basis is that of reshaping a data frame. Framed in the most general way, reshaping the data means taking the data in whatever format it’s given to you, and converting it to the format you need it. Of course, if we’re going to characterise the problem that broadly, then about half of this chapter can probably be thought of as a kind of reshaping. So we’re going to have to narrow things down a little bit. To that end, I’ll talk about a few different tools that you can use for a few different tasks. In particular, I’ll discuss a couple of easy to use (but limited) functions that I’ve included in the `lsr` package. In future versions of the book I plan to expand this discussion to include some of the more powerful tools that are available in R, but I haven’t had the time to do so yet. \n\n### 7.7.1 Long form and wide form data\n\nThe most common format in which you might obtain data is as a “case by variable” layout, commonly known as the wide form of the data.\n\n```\nload(file.path(projecthome,\"data\",\"repeated.Rdata\"))\nwho()\n```\n\n To get a sense of what I’m talking about, consider an experiment in which we are interested in the different effects that alcohol and and caffeine have on people’s working memory capacity (WMC) and reaction times (RT). We recruit 10 participants, and measure their WMC and RT under three different conditions: a “no drug” condition, in which they are not under the influence of either caffeine or alcohol, a “caffeine” condition, in which they are under the inflence of caffeine, and an “alcohol” condition, in which… well, you can probably guess. Ideally, I suppose, there would be a fourth condition in which both drugs are administered, but for the sake of simplicity let’s ignore that. The `drugs` data frame gives you a sense of what kind of data you might observe in an experiment like this: `drugs` \n\n This is a data set in “wide form”, in which each participant corresponds to a single row. We have two variables that are characteristics of the subject (i.e., their `id` number and their `gender` ) and six variables that refer to one of the two measured variables (WMC or RT) in one of the three testing conditions (alcohol, caffeine or no drug). Because all of the testing conditions (i.e., the three drug types) are applied to all participants, drug type is an example of a within-subject factor. \n\n### 7.7.2 Reshaping data using\n\n `wideToLong()` \nThe “wide form” of this data set is useful for some situations: it is often very useful to have each row correspond to a single subject. However, it is not the only way in which you might want to organise this data. For instance, you might want to have a separate row for each “testing occasion”. That is, “participant 1 under the influence of alcohol” would be one row, and “participant 1 under the influence of caffeine” would be another row. This way of organising the data is generally referred to as the long form of the data. It’s not too difficult to switch between wide and long form, and I’ll explain how it works in a moment; for now, let’s just have a look at what the long form of this data set looks like:\n\n```\n## id gender drug WMC RT\n## 1 1 female alcohol 3.7 488\n## 2 2 female alcohol 6.4 607\n## 3 3 female alcohol 4.6 643\n## 4 4 male alcohol 6.4 684\n## 5 5 female alcohol 4.9 593\n## 6 6 male alcohol 5.4 492\n```\n\n The `drugs.2` data frame that we just created has 30 rows: each of the 10 participants appears in three separate rows, one corresponding to each of the three testing conditions. And instead of having a variable like `WMC_caffeine` that indicates that we were measuring “WMC” in the “caffeine” condition, this information is now recorded in two separate variables, one called `drug` and another called `WMC` . Obviously, the long and wide forms of the data contain the same information, but they represent quite different ways of organising that information. Sometimes you find yourself needing to analyse data in wide form, and sometimes you find that you need long form. So it’s really useful to know how to switch between the two. In the example I gave above, I used a function called `wideToLong()` to do the transformation. The `wideToLong()` function is part of the `lsr` package. The key to understanding this function is that it relies on the variable names to do all the work. Notice that the variable names in the `drugs` data frame follow a very clear scheme. Whenever you have a variable with a name like `WMC_caffeine` you know that the variable being measured is “WMC”, and that the specific condition in which it is being measured is the “caffeine” condition. Similarly, you know that `RT_no.drug` refers to the “RT” variable measured in the “no drug” condition. The measured variable comes first (e.g., `WMC` ), followed by a separator character (in this case the separator is an underscore, `_` ), and then the name of the condition in which it is being measured (e.g., `caffeine` ). There are two different prefixes (i.e, the strings before the separator, `WMC` , `RT` ) which means that there are two separate variables being measured. There are three different suffixes (i.e., the strings after the separtator, `caffeine` , `alcohol` , `no.drug` ) meaning that there are three different levels of the within-subject factor. Finally, notice that the separator string (i.e., `_` ) does not appear anywhere in two of the variables ( `id` , `gender` ), indicating that these are between-subject variables, namely variables that do not vary within participant (e.g., a person’s `gender` is the same regardless of whether they’re under the influence of alcohol, caffeine etc). \nBecause of the fact that the variable naming scheme here is so informative, it’s quite possible to reshape the data frame without any additional input from the user. For example, in this particular case, you could just type the following:\n `wideToLong( drugs )` \n\n```\n## id gender within WMC RT\n## 1 1 female alcohol 3.7 488\n## 2 2 female alcohol 6.4 607\n## 3 3 female alcohol 4.6 643\n## 4 4 male alcohol 6.4 684\n## 5 5 female alcohol 4.9 593\n## 6 6 male alcohol 5.4 492\n## 7 7 male alcohol 7.9 690\n## 8 8 male alcohol 4.1 486\n## 9 9 female alcohol 5.2 686\n## 10 10 female alcohol 6.2 645\n## 11 1 female caffeine 3.7 236\n## 12 2 female caffeine 7.3 376\n## 13 3 female caffeine 7.4 226\n## 14 4 male caffeine 7.8 206\n## 15 5 female caffeine 5.2 262\n## 16 6 male caffeine 6.6 230\n## 17 7 male caffeine 7.9 259\n## 18 8 male caffeine 5.9 230\n## 19 9 female caffeine 6.2 273\n## 20 10 female caffeine 7.4 240\n## 21 1 female no.drug 3.9 371\n## 22 2 female no.drug 7.9 349\n## 23 3 female no.drug 7.3 412\n## 24 4 male no.drug 8.2 252\n## 25 5 female no.drug 7.0 439\n## 26 6 male no.drug 7.2 464\n## 27 7 male no.drug 8.9 327\n## 28 8 male no.drug 4.5 305\n## 29 9 female no.drug 7.2 327\n## 30 10 female no.drug 7.8 498\n```\n\n This is pretty good, actually. The only think it has gotten wrong here is that it doesn’t know what name to assign to the within-subject factor, so instaed of calling it something sensible like `drug` , it has use the unimaginative name `within` . If you want to ensure that the `wideToLong()` function applies a sensible name, you have to specify the `within` argument, which is just a character string that specifies the name of the within-subject factor. So when I used this command earlier, \n\n all I was doing was telling R to use `drug` as the name of the within subject factor. Now, as I was hinting earlier, the `wideToLong()` function is very inflexible. It requires that the variable names all follow this naming scheme that I outlined earlier. If you don’t follow this naming scheme it won’t work.121 The only flexibility that I’ve included here is that you can change the separator character by specifying the `sep` argument. For instance, if you were using variable names of the form `WMC/caffeine` , for instance, you could specify that `sep=\"/\"` , using a command like this \n\n```\ndrugs.2 <- wideToLong( data = drugs, within = \"drug\", sep = \"/\" )\n```\n\nand it would still work.\n\n### 7.7.3 Reshaping data using\n\n `longToWide()` To convert data from long form to wide form, the `lsr` package also includes a function called `longToWide()` . Recall from earlier that the long form of the data (i.e., the `drugs.2` data frame) contains variables named `id` , `gender` , `drug` , `WMC` and `RT` . In order to convert from long form to wide form, all you need to do is indicate which of these variables are measured separately for each condition (i.e., `WMC` and `RT` ), and which variable is the within-subject factor that specifies the condition (i.e., `drug` ). You do this via a two-sided formula, in which the measured variables are on the left hand side, and the within-subject factor is on the ritght hand side. In this case, the formula would be `WMC + RT ~ drug` . So the command that we would use might look like this: \n\n```\nlongToWide( data=drugs.2, formula= WMC+RT ~ drug )\n```\n\nor, if we chose to omit argument names, we could simplify it to this:\n\n```\nlongToWide( drugs.2, WMC+RT ~ drug )\n```\n\n Note that, just like the `wideToLong()` function, the `longToWide()` function allows you to override the default separator character. For instance, if the command I used had been \n\n```\nlongToWide( drugs.2, WMC+RT ~ drug, sep=\"/\" )\n```\n\n```\n## id gender WMC/alcohol RT/alcohol WMC/caffeine RT/caffeine WMC/no.drug\n## 1 1 female 3.7 488 3.7 236 3.9\n## 2 2 female 6.4 607 7.3 376 7.9\n## 3 3 female 4.6 643 7.4 226 7.3\n## 4 4 male 6.4 684 7.8 206 8.2\n## 5 5 female 4.9 593 5.2 262 7.0\n## 6 6 male 5.4 492 6.6 230 7.2\n## 7 7 male 7.9 690 7.9 259 8.9\n## 8 8 male 4.1 486 5.9 230 4.5\n## 9 9 female 5.2 686 6.2 273 7.2\n## 10 10 female 6.2 645 7.4 240 7.8\n## RT/no.drug\n## 1 371\n## 2 349\n## 3 412\n## 4 252\n## 5 439\n## 6 464\n## 7 327\n## 8 305\n## 9 327\n## 10 498\n```\n\n the output would contain variables with names like `RT/alcohol` instead of `RT_alcohol` . \n\n### 7.7.4 Reshaping with multiple within-subject factors\n\n As I mentioned above, the `wideToLong()` and `longToWide()` functions are quite limited in terms of what they can do. However, they do handle a broader range of situations than the one outlined above. Consider the following, fairly simple psychological experiment. I’m interested in the effects of practice on some simple decision making problem. It doesn’t really matter what the problem is, other than to note that I’m interested in two distinct outcome variables. Firstly, I care about people’s accuracy, measured by the proportion of decisions that people make correctly, denoted PC. Secondly, I care about people’s speed, measured by the mean response time taken to make those decisions, denoted MRT. That’s standard in psychological experiments: the speed-accuracy trade-off is pretty ubiquitous, so we generally need to care about both variables. To look at the effects of practice over the long term, I test each participant on two days, `day1` and `day2` , where for the sake of argument I’ll assume that `day1` and `day2` are about a week apart. To look at the effects of practice over the short term, the testing during each day is broken into two “blocks”, `block1` and `block2` , which are about 20 minutes apart. This isn’t the world’s most complicated experiment, but it’s still a fair bit more complicated than the last one. This time around we have two within-subject factors (i.e., `day` and `block` ) and we have two measured variables for each condition (i.e., `PC` and `MRT` ). The `choice` data frame shows what the wide form of this kind of data might look like: `choice` \n\n Notice that this time around we have variable names of the form `MRT/block1/day2` . As before, the first part of the name refers to the measured variable (response time), but there are now two suffixes, one indicating that the testing took place in block 1, and the other indicating that it took place on day 2. And just to complicate matters, it uses `/` as the separator character rather than `_` . Even so, reshaping this data set is pretty easy. The command to do it is, \n\n```\nchoice.2 <- wideToLong( choice, within=c(\"block\",\"day\"), sep=\"/\" )\n```\n\n which is pretty much the exact same command we used last time. The only difference here is that, because there are two within-subject factors, the `within` argument is a vector that contains two names. When we look at the long form data frame that this creates, we get this: `choice.2` \n\n```\n## id gender MRT PC block day\n## 1 1 male 415 79 block1 day1\n## 2 2 male 500 83 block1 day1\n## 3 3 female 478 91 block1 day1\n## 4 4 female 550 75 block1 day1\n## 5 1 male 400 88 block1 day2\n## 6 2 male 490 92 block1 day2\n## 7 3 female 468 98 block1 day2\n## 8 4 female 502 89 block1 day2\n## 9 1 male 455 82 block2 day1\n## 10 2 male 532 86 block2 day1\n## 11 3 female 499 90 block2 day1\n## 12 4 female 602 78 block2 day1\n## 13 1 male 450 93 block2 day2\n## 14 2 male 518 97 block2 day2\n## 15 3 female 474 100 block2 day2\n## 16 4 female 588 95 block2 day2\n```\n\n In this long form data frame we have two between-subject variables ( `id` and `gender` ), two variables that define our within-subject manipulations ( `block` and `day` ), and two more contain the measurements we took ( `MRT` and `PC` ). To convert this back to wide form is equally straightforward. We use the `longToWide()` function, but this time around we need to alter the formula in order to tell it that we have two within-subject factors. The command is now \n\n```\nlongToWide( choice.2, MRT+PC ~ block+day, sep=\"/\" )\n```\n\n and this produces a wide form data set containing the same variables as the original `choice` data frame. \n\n### 7.7.5 What other options are there?\n\n The advantage to the approach described in the previous section is that it solves a quite specific problem (but a commonly encountered one) with a minimum of fuss. The disadvantage is that the tools are quite limited in scope. They allow you to switch your data back and forth between two different formats that are very common in everyday data analysis. However, there a number of other tools that you can use if need be. Just within the core packages distributed with R there is the `reshape()` function, as well as the `stack()` and `unstack()` functions, all of which can be useful under certain circumstances. And there are of course thousands of packages on CRAN that you can use to help you with different tasks. One popular package for this purpose is the `reshape` package, written by (??? for details see Wickham2007). There are two key functions in this package, called `melt()` and `cast()` that are pretty useful for solving a lot of reshaping problems. In a future version of this book I intend to discuss `melt()` and `cast()` in a fair amount of detail. \n\n## 7.8 Working with text\n\n Sometimes your data set is quite text heavy. This can be for a lot of different reasons. Maybe the raw data are actually taken from text sources (e.g., newspaper articles), or maybe your data set contains a lot of free responses to survey questions, in which people can write whatever text they like in response to some query. Or maybe you just need to rejig some of the text used to describe nominal scale variables. Regardless of what the reason is, you’ll probably want to know a little bit about how to handle text in R. Some things you already know how to do: I’ve discussed the use of `nchar()` to calculate the number of characters in a string (Section 3.8.1), and a lot of the general purpose tools that I’ve discussed elsewhere (e.g., the `==` operator) have been applied to text data as well as to numeric data. However, because text data is quite rich, and generally not as well structured as numeric data, R provides a lot of additional tools that are quite specific to text. In this section I discuss only those tools that come as part of the base packages, but there are other possibilities out there: the `stringr` package provides a powerful alternative that is a lot more coherent than the basic tools, and is well worth looking into. \n\n### 7.8.1 Shortening a string\n\nThe first task I want to talk about is how to shorten a character string. For example, suppose that I have a vector that contains the names of several different animals:\n\n```\nanimals <- c( \"cat\", \"dog\", \"kangaroo\", \"whale\" )\n```\n\n It might be useful in some contexts to extract the first three letters of each word. This is often useful when annotating figures, or when creating variable labels: it’s often very inconvenient to use the full name, so you want to shorten it to a short code for space reasons. The `strtrim()` function can be used for this purpose. It has two arguments: `x` is a vector containing the text to be shortened and `width` specifies the number of characters to keep. When applied to the `animals` data, here’s what we get: \n\n```\nstrtrim( x = animals, width = 3 )\n```\n\n```\n## [1] \"cat\" \"dog\" \"kan\" \"wha\"\n```\n\n Note that the only thing that `strtrim()` does is chop off excess characters at the end of a string. It doesn’t insert any whitespace characters to fill them out if the original string is shorter than the `width` argument. For example, if I trim the `animals` data to 4 characters, here’s what I get: \n\n```\nstrtrim( x = animals, width = 4 )\n```\n\n```\n## [1] \"cat\" \"dog\" \"kang\" \"whal\"\n```\n\n The `\"cat\"` and `\"dog\"` strings still only use 3 characters. Okay, but what if you don’t want to start from the first letter? Suppose, for instance, I only wanted to keep the second and third letter of each word. That doesn’t happen quite as often, but there are some situations where you need to do something like that. If that does happen, then the function you need is `substr()` , in which you specify a `start` point and a `stop` point instead of specifying the width. For instance, to keep only the 2nd and 3rd letters of the various `animals` , I can do the following: \n\n```\nsubstr( x = animals, start = 2, stop = 3 )\n```\n\n```\n## [1] \"at\" \"og\" \"an\" \"ha\"\n```\n\n### 7.8.2 Pasting strings together\n\n Much more commonly, you will need either to glue several character strings together or to pull them apart. To glue several strings together, the `paste()` function is very useful. There are three arguments to the `paste()` function: \n\n* \n `...` As usual, the dots “match” up against any number of inputs. In this case, the inputs should be the various different strings you want to paste together. * \n `sep` . This argument should be a string, indicating what characters R should use as separators, in order to keep each of the original strings separate from each other in the pasted output. By default the value is a single space, `sep = \" \"` . This is made a little clearer when we look at the examples. * \n `collapse` . This is an argument indicating whether the `paste()` function should interpret vector inputs as things to be collapsed, or whether a vector of inputs should be converted into a vector of outputs. The default value is `collapse = NULL` which is interpreted as meaning that vectors should not be collapsed. If you want to collapse vectors into as single string, then you should specify a value for `collapse` . Specifically, the value of `collapse` should correspond to the separator character that you want to use for the collapsed inputs. Again, see the examples below for more details. \nThat probably doesn’t make much sense yet, so let’s start with a simple example. First, let’s try to paste two words together, like this:\n\n```\npaste( \"hello\", \"world\" )\n```\n\n `## [1] \"hello world\"` Notice that R has inserted a space between the `\"hello\"` and `\"world\"` . Suppose that’s not what I wanted. Instead, I might want to use `.` as the separator character, or to use no separator at all. To do either of those, I would need to specify `sep = \".\"` or `sep = \"\"` .122 For instance: \n\n```\npaste( \"hello\", \"world\", sep = \".\" )\n```\n\n `## [1] \"hello.world\"` Now let’s consider a slightly more complicated example. Suppose I have two vectors that I want to `paste()` together. Let’s say something like this: \n\n```\nhw <- c( \"hello\", \"world\" )\nng <- c( \"nasty\", \"government\" )\n```\n\n And suppose I want to paste these together. However, if you think about it, this statement is kind of ambiguous. It could mean that I want to do an “element wise” paste, in which all of the first elements get pasted together ( `\"hello nasty\"` ) and all the second elements get pasted together ( `\"world government\"` ). Or, alternatively, I might intend to collapse everything into one big string ( \n\n```\n\"hello nasty world government\"\n```\n\n). By default, the `paste()` function assumes that you want to do an element-wise paste: `paste( hw, ng )` \n\n```\n## [1] \"hello nasty\" \"world government\"\n```\n\n However, there’s nothing stopping you from overriding this default. All you have to do is specify a value for the `collapse` argument, and R will chuck everything into one dirty big string. To give you a sense of exactly how this works, what I’ll do in this next example is specify different values for `sep` and `collapse` : \n\n```\npaste( hw, ng, sep = \".\", collapse = \":::\")\n```\n\n```\n## [1] \"hello.nasty:::world.government\"\n```\n\n### 7.8.3 Splitting strings\n\n At other times you have the opposite problem to the one in the last section: you have a whole lot of text bundled together into a single string that needs to be pulled apart and stored as several different variables. For instance, the data set that you get sent might include a single variable containing someone’s full name, and you need to separate it into first names and last names. To do this in R you can use the `strsplit()` function, and for the sake of argument, let’s assume that the string you want to split up is the following string: \n\n```\nmonkey <- \"It was the best of times. It was the blurst of times.\"\n```\n\n To use the `strsplit()` function to break this apart, there are three arguments that you need to pay particular attention to: \n\n* \n `x` . A vector of character strings containing the data that you want to split. * \n `split` . Depending on the value of the `fixed` argument, this is either a fixed string that specifies a delimiter, or a regular expression that matches against one or more possible delimiters. If you don’t know what regular expressions are (probably most readers of this book), don’t use this option. Just specify a separator string, just like you would for the `paste()` function. * \n `fixed` . Set `fixed = TRUE` if you want to use a fixed delimiter. As noted above, unless you understand regular expressions this is definitely what you want. However, the default value is `fixed = FALSE` , so you have to set it explicitly. \nLet’s look at a simple example:\n\n```\nmonkey.1 <- strsplit( x = monkey, split = \" \", fixed = TRUE )\nmonkey.1\n```\n\n One thing to note in passing is that the output here is a list (you can tell from the `[[1]]` part of the output), whose first and only element is a character vector. This is useful in a lot of ways, since it means that you can input a character vector for `x` and then then have the `strsplit()` function split all of them, but it’s kind of annoying when you only have a single input. To that end, it’s useful to know that you can `unlist()` the output: `unlist( monkey.1 )` \n\n To understand why it’s important to remember to use the `fixed = TRUE` argument, suppose we wanted to split this into two separate sentences. That is, we want to use `split = \".\"` as our delimiter string. As long as we tell R to remember to treat this as a fixed separator character, then we get the right answer: \n\n```\nstrsplit( x = monkey, split = \".\", fixed = TRUE )\n```\n\n```\n## [[1]]\n## [1] \"It was the best of times\" \" It was the blurst of times\"\n```\n\n However, if we don’t do this, then R will assume that when you typed `split = \".\"` you were trying to construct a “regular expression”, and as it happens the character `.` has a special meaning within a regular expression. As a consequence, if you forget to include the `fixed = TRUE` part, you won’t get the answers you’re looking for. \n\n### 7.8.4 Making simple conversions\n\n A slightly different task that comes up quite often is making transformations to text. A simple example of this would be converting text to lower case or upper case, which you can do using the `toupper()` and `tolower()` functions. Both of these functions have a single argument `x` which contains the text that needs to be converted. An example of this is shown below: \n\n```\ntext <- c( \"lIfe\", \"Impact\" )\ntolower( x = text )\n```\n\n```\n## [1] \"life\" \"impact\"\n```\n\n A slightly more powerful way of doing text transformations is to use the `chartr()` function, which allows you to specify a “character by character” substitution. This function contains three arguments, `old` , `new` and `x` . As usual `x` specifies the text that needs to be transformed. The `old` and `new` arguments are strings of the same length, and they specify how `x` is to be converted. Every instance of the first character in `old` is converted to the first character in `new` and so on. For instance, suppose I wanted to convert `\"albino\"` to `\"libido\"` . To do this, I need to convert all of the `\"a\"` characters (all 1 of them) in `\"albino\"` into `\"l\"` characters (i.e., `a` \\(\\rightarrow\\) `l` ). Additionally, I need to make the substitutions `l` \\(\\rightarrow\\) `i` and `n` \\(\\rightarrow\\) `d` . To do so, I would use the following command: \n\n```\nold.text <- \"albino\"\nchartr( old = \"aln\", new = \"lid\", x = old.text )\n```\n\n `## [1] \"libido\"` \n\n### 7.8.5 Applying logical operations to text\n\n In Section 3.9.5 we discussed a very basic text processing tool, namely the ability to use the equality operator `==` to test to see if two strings are identical to each other. However, you can also use other logical operators too. For instance R also allows you to use the `<` and `>` operators to determine which of two strings comes first, alphabetically speaking. Sort of. Actually, it’s a bit more complicated than that, but let’s start with a simple example: `\"cat\" < \"dog\"` `## [1] TRUE` In this case, we see that `\"cat\"` does does come before `\"dog\"` alphabetically, so R judges the statement to be true. However, if we ask R to tell us if `\"cat\"` comes before `\"anteater\"` , `\"cat\" < \"anteater\"` `## [1] FALSE` It tell us that the statement is false. So far, so good. But text data is a bit more complicated than the dictionary suggests. What about `\"cat\"` and `\"CAT\"` ? Which of these comes first? Let’s try it and find out: `\"CAT\" < \"cat\"` `## [1] FALSE` In other words, R assumes that uppercase letters come before lowercase ones. Fair enough. No-one is likely to be surprised by that. What you might find surprising is that R assumes that all uppercase letters come before all lowercase ones. That is, while `\"anteater\" < \"zebra\"` is a true statement, and the uppercase equivalent `\"ANTEATER\" < \"ZEBRA\"` is also true, it is not true to say that `\"anteater\" < \"ZEBRA\"` , as the following extract illustrates: `\"anteater\" < \"ZEBRA\"` `## [1] TRUE` \nThis may seem slightly counterintuitive. With that in mind, it may help to have a quick look Table 7.3, which lists various text characters in the order that R uses.\n\nCharacters |\n | --- |\n! \" # $ % & ’ ( ) * + , - . / 0 1 2 3 4 5 6 7 8 9 : ; < = > ? @ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z [ ] ^ _ ‘ a b c d e f g h i j k l m n o p q r s t u v w x y z } | { |\n\n### 7.8.6 Concatenating and printing with\n\n `cat()` One function that I want to make a point of talking about, even though it’s not quite on topic, is the `cat()` function. The `cat()` function is a of mixture of `paste()` and `print()` . That is, what it does is concatenate strings and then print them out. In your own work you can probably survive without it, since `print()` and `paste()` will actually do what you need, but the `cat()` function is so widely used that I think it’s a good idea to talk about it here. The basic idea behind `cat()` is straightforward. Like `paste()` , it takes several arguments as inputs, which it converts to strings, collapses (using a separator character specified using the `sep` argument), and prints on screen. If you want, you can use the `file` argument to tell R to print the output into a file rather than on screen (I won’t do that here). However, it’s important to note that the `cat()` function collapses vectors first, and then concatenates them. That is, notice that when I use `cat()` to combine `hw` and `ng` , I get a different result than if I’d used `paste()` `cat( hw, ng )` \n\n```\npaste( hw, ng, collapse = \" \" )\n```\n\n```\n## [1] \"hello nasty world government\"\n```\n\n Notice the difference in the ordering of words. There’s a few additional details that I need to mention about `cat()` . Firstly, `cat()` really is a function for printing, and not for creating text strings to store for later. You can’t assign the output to a variable, as the following example illustrates: `x <- cat( hw, ng )` \n\n `x` `## NULL` Despite my attempt to store the output as a variable, `cat()` printed the results on screen anyway, and it turns out that the variable I created doesn’t contain anything at all.123 Secondly, the `cat()` function makes use of a number of “special” characters. I’ll talk more about these in the next section, but I’ll illustrate the basic point now, using the example of `\"\\n\"` which is interpreted as a “new line” character. For instance, compare the behaviour of `print()` and `cat()` when asked to print the string `\"hello\\nworld\"` : \n\n```\nprint( \"hello\\nworld\" ) # print literally:\n```\n\n```\n## [1] \"hello\\nworld\"\n```\n\n```\ncat( \"hello\\nworld\" ) # interpret as newline\n```\n\n```\n## hello\n## world\n```\n\nIn fact, this behaviour is important enough that it deserves a section of its very own…\n\n### 7.8.7 Using escape characters in text\n\n The previous section brings us quite naturally to a fairly fundamental issue when dealing with strings, namely the issue of delimiters and escape characters. Reduced to its most basic form, the problem we have is that R commands are written using text characters, and our strings also consist of text characters. So, suppose I want to type in the word “hello”, and have R encode it as a string. If I were to just type `hello` , R will think that I’m referring to a variable or a function called `hello` rather than interpret it as a string. The solution that R adopts is to require you to enclose your string by delimiter characters, which can be either double quotes or single quotes. So, when I type `\"hello\"` or `'hello'` then R knows that it should treat the text in between the quote marks as a character string. However, this isn’t a complete solution to the problem: after all, `\"` and `'` are themselves perfectly legitimate text characters, and so we might want to include those in our string as well. For instance, suppose I wanted to encode the name “O’Rourke” as a string. It’s not legitimate for me to type `'O'rourke'` because R is too stupid to realise that “O’Rourke” is a real word. So it will interpret the `'O'` part as a complete string, and then will get confused when it reaches the `Rourke'` part. As a consequence, what you get is an error message: \n\n```\n'O'Rourke'\nError: unexpected symbol in \"'O'Rourke\"\n```\n\n To some extent, R offers us a cheap fix to the problem because of the fact that it allows us to use either `\"` or `'` as the delimiter character. Although `'O'rourke'` will make R cry, it is perfectly happy with `\"O'Rourke\"` : `\"O'Rourke\"` `## [1] \"O'Rourke\"` \nThis is a real advantage to having two different delimiter characters. Unfortunately, anyone with even the slightest bit of deviousness to them can see the problem with this. Suppose I’m reading a book that contains the following passage,\n says, “Yay, money!”. It’s a joke, but no-one laughs.\n and I want to enter this as a string. Neither the `'` or `\"` delimiters will solve the problem here, since this string contains both a single quote character and a double quote character. To encode strings like this one, we have to do something a little bit clever. \n\nEscape.sequence | Interpretation |\n | --- | --- |\n | Newline |\n | Horizontal Tab |\n | Vertical Tab |\n | Backspace |\n | Carriage Return |\n | Form feed |\n | Alert sound |\n | Backslash |\n | Single quote |\n | Double quote |\n\n The solution to the problem is to designate an escape character, which in this case is `\\` , the humble backslash. The escape character is a bit of a sacrificial lamb: if you include a backslash character in your string, R will not treat it as a literal character at all. It’s actually used as a way of inserting “special” characters into your string. For instance, if you want to force R to insert actual quote marks into the string, then what you actually type is `\\'` or `\\\"` (these are called escape sequences). So, in order to encode the string discussed earlier, here’s a command I could use: \n\n```\nPJ <- \" says, \\\"Yay, money!\\\". It\\'s a joke, but no-one laughs.\"\n```\n\n Notice that I’ve included the backslashes for both the single quotes and double quotes. That’s actually overkill: since I’ve used `\"` as my delimiter, I only needed to do this for the double quotes. Nevertheless, the command has worked, since I didn’t get an error message. Now let’s see what happens when I print it out: `print( PJ )` \n\n```\n## [1] \" says, \\\"Yay, money!\\\". It's a joke, but no-one laughs.\"\n```\n\n Hm. Why has R printed out the string using `\\\"` ? For the exact same reason that I needed to insert the backslash in the first place. That is, when R prints out the `PJ` string, it has enclosed it with delimiter characters, and it wants to unambiguously show us which of the double quotes are delimiters and which ones are actually part of the string. Fortunately, if this bugs you, you can make it go away by using the `print.noquote()` function, which will just print out the literal string that you encoded in the first place: `print.noquote( PJ )` Typing `cat(PJ)` will produce a similar output. Introducing the escape character solves a lot of problems, since it provides a mechanism by which we can insert all sorts of characters that aren’t on the keyboard. For instance, as far as a computer is concerned, “new line” is actually a text character. It’s the character that is printed whenever you hit the “return” key on your keyboard. If you want to insert a new line character into your string, you can actually do this by including the escape sequence `\\n` . Or, if you want to insert a backslash character, then you can use `\\\\` . A list of the standard escape sequences recognised by R is shown in Table 7.4. A lot of these actually date back to the days of the typewriter (e.g., carriage return), so they might seem a bit counterintuitive to people who’ve never used one. In order to get a sense for what the various escape sequences do, we’ll have to use the `cat()` function, because it’s the only function “dumb” enough to literally print them out: \n\n```\ncat( \"xxxx\\boo\" ) # \\b is a backspace, so it deletes the preceding x\ncat( \"xxxx\\too\" ) # \\t is a tab, so it inserts a tab space\ncat( \"xxxx\\noo\" ) # \\n is a newline character\ncat( \"xxxx\\roo\" ) # \\r returns you to the beginning of the line\n```\n\n And that’s pretty much it. There are a few other escape sequence that R recognises, which you can use to insert arbitrary ASCII or Unicode characters into your string (type `?Quotes` for more details) but I won’t go into details here. \n\n### 7.8.8 Matching and substituting text\n\n Another task that we often want to solve is find all strings that match a certain criterion, and possibly even to make alterations to the text on that basis. There are several functions in R that allow you to do this, three of which I’ll talk about briefly here: `grep()` , `gsub()` and `sub()` . Much like the `substr()` function that I talked about earlier, all three of these functions are intended to be used in conjunction with regular expressions (see Section 7.8.9 but you can also use them in a simpler fashion, since they all allow you to set `fixed = TRUE` , which means we can ignore all this regular expression rubbish and just use simple text matching. So, how do these functions work? Let’s start with the `grep()` function. The purpose of this function is to input a vector of character strings `x` , and to extract all those strings that fit a certain pattern. In our examples, I’ll assume that the `pattern` in question is a literal sequence of characters that the string must contain (that’s what `fixed = TRUE` does). To illustrate this, let’s start with a simple data set, a vector that contains the names of three `beers` . Something like this: \n\n```\nbeers <- c( \"little creatures\", \"sierra nevada\", \"coopers pale\" )\n```\n\n Next, let’s use `grep()` to find out which of these strings contains the substring `\"er\"` . That is, the `pattern` that we need to match is the fixed string `\"er\"` , so the command we need to use is: \n\n `## [1] 2 3` What the output here is telling us is that the second and third elements of `beers` both contain the substring `\"er\"` . Alternatively, however, we might prefer it if `grep()` returned the actual strings themselves. We can do this by specifying `value = TRUE` in our function call. That is, we’d use a command like this: \n\n```\n## [1] \"sierra nevada\" \"coopers pale\"\n```\n\n The other two functions that I wanted to mention in this section are `gsub()` and `sub()` . These are both similar in spirit to `grep()` insofar as what they do is search through the input strings ( `x` ) and find all of the strings that match a `pattern` . However, what these two functions do is replace the pattern with a `replacement` string. The `gsub()` function will replace all instances of the pattern, whereas the `sub()` function just replaces the first instance of it in each string. To illustrate how this works, suppose I want to replace all instances of the letter `\"a\"` with the string `\"BLAH\"` . I can do this to the `beers` data using the `gsub()` function: \n\n```\ngsub( pattern = \"a\", replacement = \"BLAH\", x = beers, fixed = TRUE )\n```\n\n Notice that all three of the `\"a\"` s in `\"sierra nevada\"` have been replaced. In contrast, let’s see what happens when we use the exact same command, but this time using the `sub()` function instead: \n\n```\nsub( pattern = \"a\", replacement = \"BLAH\", x = beers, fixed = TRUE )\n```\n\n Only the first `\"a\"` is changed. \n\n### 7.8.9 Regular expressions (not really)\n\n There’s one last thing I want to talk about regarding text manipulation, and that’s the concept of a regular expression. Throughout this section we’ve often needed to specify `fixed = TRUE` in order to force R to treat some of our strings as actual strings, rather than as regular expressions. So, before moving on, I want to very briefly explain what regular expressions are. I’m not going to talk at all about how they work or how you specify them, because they’re genuinely complicated and not at all relevant to this book. However, they are extremely powerful tools and they’re quite widely used by people who have to work with lots of text data (e.g., people who work with natural language data), and so it’s handy to at least have a vague idea about what they are. The basic idea is quite simple. Suppose I want to extract all strings in my `beers` vector that contain a vowel followed immediately by the letter `\"s\"` . That is, I want to finds the beer names that contain either `\"as\"` , `\"es\"` , `\"is\"` , `\"os\"` or `\"us\"` . One possibility would be to manually specify all of these possibilities and then match against these as fixed strings one at a time, but that’s tedious. The alternative is to try to write out a single “regular” expression that matches all of these. The regular expression that does this124 is `\"[aeiou]s\"` , and you can kind of see what the syntax is doing here. The bracketed expression means “any of the things in the middle”, so the expression as a whole means “any of the things in the middle” (i.e. vowels) followed by the letter `\"s\"` . When applied to our beer names we get this: \n\n```\ngrep( pattern = \"[aeiou]s\", x = beers, value = TRUE )\n```\n\n```\n## [1] \"little creatures\"\n```\n\n So it turns out that only `\"little creatures\"` contains a vowel followed by the letter `\"s\"` . But of course, had the data contained a beer like `\"fosters\"` , that would have matched as well because it contains the string `\"os\"` . However, I deliberately chose not to include it because Fosters is not – in my opinion – a proper beer.125 As you can tell from this example, regular expressions are a neat tool for specifying patterns in text: in this case, “vowel then s”. So they are definitely things worth knowing about if you ever find yourself needing to work with a large body of text. However, since they are fairly complex and not necessary for any of the applications discussed in this book, I won’t talk about them any further. \n\n## 7.9 Reading unusual data files\n\n In this section I’m going to switch topics (again!) and turn to the question of how you can load data from a range of different sources. Throughout this book I’ve assumed that your data are stored as an `.Rdata` file or as a “properly” formatted CSV file. And if so, then the basic tools that I discussed in Section 4.5 should be quite sufficient. However, in real life that’s not a terribly plausible assumption to make, so I’d better talk about some of the other possibilities that you might run into. \n\n### 7.9.1 Loading data from text files\n\n The first thing I should point out is that if your data are saved as a text file but aren’t quite in the proper CSV format, then there’s still a pretty good chance that the `read.csv()` function (or equivalently, `read.table()` ) will be able to open it. You just need to specify a few more of the optional arguments to the function. If you type `?read.csv` you’ll see that the `read.csv()` function actually has several arguments that you can specify. Obviously you need to specify the `file` that you want it to load, but the others all have sensible default values. Nevertheless, you will sometimes need to change them. The ones that I’ve often found myself needing to change are: \n\n* \n `header` . A lot of the time when you’re storing data as a CSV file, the first row actually contains the column names and not data. If that’s not true, you need to set `header = FALSE` . * \n `sep` . As the name “comma separated value” indicates, the values in a row of a CSV file are usually separated by commas. This isn’t universal, however. In Europe the decimal point is typically written as `,` instead of `.` and as a consequence it would be somewhat awkward to use `,` as the separator. Therefore it is not unusual to use `;` over there. At other times, I’ve seen a TAB character used. To handle these cases, we’d need to set `sep = \";\"` or `sep = \"\\t\"` . * \n `quote` . It’s conventional in CSV files to include a quoting character for textual data. As you can see by looking at the `booksales.csv}` file, this is usually a double quote character, `\"` . But sometimes there is no quoting character at all, or you might see a single quote mark `'` used instead. In those cases you’d need to specify `quote = \"\"` or `quote = \"'\"` . * \n `skip` . It’s actually very common to receive CSV files in which the first few rows have nothing to do with the actual data. Instead, they provide a human readable summary of where the data came from, or maybe they include some technical info that doesn’t relate to the data. To tell R to ignore the first (say) three lines, you’d need to set `skip = 3` * \n `na.strings` . Often you’ll get given data with missing values. For one reason or another, some entries in the table are missing. The data file needs to include a “special” string to indicate that the entry is missing. By default R assumes that this string is `NA` , since that’s what it would do, but there’s no universal agreement on what to use in this situation. If the file uses `???` instead, then you’ll need to set `na.strings = \"???\"` . \nIt’s kind of nice to be able to have all these options that you can tinker with. For instance, have a look at the data file shown pictured in Figure 7.1. This file contains almost the same data as the last file (except it doesn’t have a header), and it uses a bunch of wacky features that you don’t normally see in CSV files. In fact, it just so happens that I’m going to have to change all five of those arguments listed above in order to load this file. Here’s how I would do it:\n\n```\ndata <- read.csv( \n file = file.path(projecthome,\"data\",\"booksales2.csv\"), # specify path to the file\n header = FALSE, # variable names in the file?\n skip = 8, # ignore the first 8 lines\n quote = \"*\", # what indicates text data?\n sep = \"\\t\", # what separates different entries?\n na.strings = \"NFI\" ) # what is the code for missing data?\n```\n\nIf I now have a look at the data I’ve loaded, I see that this is what I’ve got:\n `head( data )` \n\n```\n## V1 V2 V3 V4\n## 1 January 31 0 high\n## 2 February 28 100 high\n## 3 March 31 200 low\n## 4 April 30 50 out\n## 5 May 31 NA out\n## 6 June 30 0 high\n```\n\n Because I told R to expect `*` to be used as the quoting character instead of `\"` ; to look for tabs (which we write like this: `\\t` ) instead of commas, and to skip the first 8 lines of the file, it’s basically loaded the right data. However, since `booksales2.csv` doesn’t contain the column names, R has made them up. Showing the kind of imagination I expect from insentient software, R decided to call them `V1` , `V2` , `V3` and `V4` . Finally, because I told it that the file uses “NFI” to denote missing data, R correctly figures out that the sales data for May are actually missing. \nIn real life you’ll rarely see data this stupidly formatted.126\n\n### 7.9.2 Loading data from SPSS (and other statistics packages)\n\n The commands listed above are the main ones we’ll need for data files in this book. But in real life we have many more possibilities. For example, you might want to read data files in from other statistics programs. Since SPSS is probably the most widely used statistics package in psychology, it’s worth briefly showing how to open SPSS data files (file extension `.sav` ). It’s surprisingly easy. The extract below should illustrate how to do so: \n\n```\nlibrary( foreign ) # load the package\nX <- read.spss( file.path(projecthome,\"data\",\"datafile.sav\" )) # create a list containing the data\nX <- as.data.frame( X ) # convert to data frame\n```\n\nIf you wanted to import from an SPSS file to a data frame directly, instead of importing a list and then converting the list to a data frame, you can do that too:\n\n```\nX <- read.spss( file = \"datafile.sav\", to.data.frame = TRUE )\n```\n\n And that’s pretty much it, at least as far as SPSS goes. As far as other statistical software goes, the `foreign` package provides a wealth of possibilities. To open SAS files, check out the `read.ssd()` and `read.xport()` functions. To open data from Minitab, the `read.mtp()` function is what you’re looking for. For Stata, the `read.dta()` function is what you want. For Systat, the `read.systat()` function is what you’re after. \n\n### 7.9.3 Loading Excel files\n\n A different problem is posed by Excel files. Despite years of yelling at people for sending data to me encoded in a proprietary data format, I get sent a lot of Excel files. In general R does a pretty good job of opening them, but it’s bit finicky because Microsoft don’t seem to be terribly fond of people using non-Microsoft products, and go to some lengths to make it tricky. If you get an Excel file, my suggestion would be to open it up in Excel (or better yet, OpenOffice, since that’s free software) and then save the spreadsheet as a CSV file. Once you’ve got the data in that format, you can open it using `read.csv()` . However, if for some reason you’re desperate to open the `.xls` or `.xlsx` file directly, then you can use the `read.xls()` function in the `gdata` package: \n\n```\nlibrary( gdata ) # load the package\nX <- read.xls( \"datafile.xlsx\" ) # create a data frame\n```\n\nThis usually works. And if it doesn’t, you’re probably justified in “suggesting” to the person that sent you the file that they should send you a nice clean CSV file instead.\n\n### 7.9.4 Loading Matlab (& Octave) files\n\n A lot of scientific labs use Matlab as their default platform for scientific computing; or Octave as a free alternative. Opening Matlab data files (file extension `.mat` ) slightly more complicated, and if it wasn’t for the fact that Matlab is so very widespread and is an extremely good platform, I wouldn’t mention it. However, since Matlab is so widely used, I think it’s worth discussing briefly how to get Matlab and R to play nicely together. The way to do this is to install the `R.matlab` package (don’t forget to install the dependencies too). Once you’ve installed and loaded the package, you have access to the `readMat()` function. As any Matlab user will know, the `.mat` files that Matlab produces are workspace files, very much like the `.Rdata` files that R produces. So you can’t import a `.mat` file as a data frame. However, you can import it as a list. So, when we do this: \n\n```\nlibrary( R.matlab ) # load the package\ndata <- readMat( \"matlabfile.mat\" ) # read the data file to a list\n```\n\n The `data` object that gets created will be a list, containing one variable for every variable stored in the Matlab file. It’s fairly straightforward, though there are some subtleties that I’m ignoring. In particular, note that if you don’t have the `Rcompression` package, you can’t open Matlab files above the version 6 format. So, if like me you’ve got a recent version of Matlab, and don’t have the `Rcompression` package, you’ll need to save your files using the `-v6` flag otherwise R can’t open them. Oh, and Octave users? The `foreign` package contains a `read.octave()` command. Just this once, the world makes life easier for you folks than it does for all those cashed-up swanky Matlab bastards. \n\n### 7.9.5 Saving other kinds of data\n\n Given that I talked extensively about how to load data from non-R files, it might be worth briefly mentioning that R is also pretty good at writing data into other file formats besides it’s own native ones. I won’t discuss them in this book, but the `write.csv()` function can write CSV files, and the `write.foreign()` function (in the `foreign` package) can write SPSS, Stata and SAS files. There are also a lot of low level commands that you can use to write very specific information to a file, so if you really, really needed to you could create your own \n\n```\nwrite.obscurefiletype()\n```\n\nfunction, but that’s also a long way beyond the scope of this book. For now, all that I want you to recognise is that this capability is there if you need it. \n\n### 7.9.6 Are we done yet?\n\n Of course not. If I’ve learned nothing else about R it’s that you’re never bloody done. This listing doesn’t even come close to exhausting the possibilities. Databases are supported by the `RODBC` , `DBI` , and `RMySQL` packages among others. You can open webpages using the `RCurl` package. Reading and writing JSON objects is supported through the `rjson` package. And so on. In a sense, the right question is not so much “can R do this?” so much as “whereabouts in the wilds of CRAN is the damn package that does it?” \n\n## 7.10 Coercing data from one class to another\n\nSometimes you want to change the variable class. This can happen for all sorts of reasons. Sometimes when you import data from files, it can come to you in the wrong format: numbers sometimes get imported as text, dates usually get imported as text, and many other possibilities besides. Regardless of how you’ve ended up in this situation, there’s a very good chance that sometimes you’ll want to convert a variable from one class into another one. Or, to use the correct term, you want to coerce the variable from one class into another. Coercion is a little tricky, and so I’ll only discuss the very basics here, using a few simple examples.\n Firstly, let’s suppose we have a variable `x` that is supposed to be representing a number, but the data file that you’ve been given has encoded it as text. Let’s imagine that the variable is something like this: \n\n```\nx <- \"100\" # the variable \nclass(x) # what class is it?\n```\n\n `## [1] \"character\"` Obviously, if I want to do calculations using `x` in its current state, R is going to get very annoyed at me. It thinks that `x` is text, so it’s not going to allow me to try to do mathematics using it! Obviously, we need to coerce `x` from character to numeric. We can do that in a straightforward way by using the `as.numeric()` function: \n\n```\nx <- as.numeric(x) # coerce the variable\nclass(x) # what class is it?\n```\n\n `## [1] \"numeric\"` \n\n```\nx + 1 # hey, addition works!\n```\n\n `## [1] 101` Not surprisingly, we can also convert it back again if we need to. The function that we use to do this is the `as.character()` function: \n\n```\nx <- as.character(x) # coerce back to text\nclass(x) # check the class:\n```\n\n `## [1] \"character\"` However, there’s some fairly obvious limitations: you can’t coerce the string `\"hello world\"` into a number because, well, there’s isn’t a number that corresponds to it. Or, at least, you can’t do anything useful: \n\n```\nas.numeric( \"hello world\" ) # this isn't going to work.\n```\n\n```\n## Warning: NAs introduced by coercion\n```\n\n `## [1] NA` In this case R doesn’t give you an error message; it just gives you a warning, and then says that the data is missing (see Section 4.6.1 for the interpretation of `NA` ). That gives you a feel for how to change between numeric and character data. What about logical data? To cover this briefly, coercing text to logical data is pretty intuitive: you use the `as.logical()` function, and the character strings `\"T\"` , `\"TRUE\"` , `\"True\"` and `\"true\"` all convert to the logical value of `TRUE` . Similarly `\"F\"` , `\"FALSE\"` , `\"False\"` , and `\"false\"` all become `FALSE` . All other strings convert to `NA` . When you go back the other way using `as.character()` , `TRUE` converts to `\"TRUE\"` and `FALSE` converts to `\"FALSE\"` . Converting numbers to logicals – again using `as.logical()` – is straightforward. Following the convention in the study of logic, the number `0` converts to `FALSE` . Everything else is `TRUE` . Going back using `as.numeric()` , `FALSE` converts to `0` and `TRUE` converts to `1` . \n\n## 7.11 Other useful data structures\n\nUp to this point we have encountered several different kinds of variables. At the simplest level, we’ve seen numeric data, logical data and character data. However, we’ve also encountered some more complicated kinds of variables, namely factors, formulas, data frames and lists. We’ll see a few more specialised data structures later on in this book, but there’s a few more generic ones that I want to talk about in passing. None of them are central to the rest of the book (and in fact, the only one we’ll even see anywhere else is the matrix), but they do crop up a fair bit in real life.\n\n### 7.11.1 Matrices\n\n In various different places in this chapter I’ve made reference to an R data structure called a matrix, and mentioned that I’d talk a bit more about matrices later on. That time has come. Much like a data frame, a matrix is basically a big rectangular table of data, and in fact there are quite a few similarities between the two. However, there are also some key differences, so it’s important to talk about matrices in a little detail. Let’s start by using `rbind()` to create a small matrix:127 \n\n```\nrow.1 <- c( 2,3,1 ) # create data for row 1\nrow.2 <- c( 5,6,7 ) # create data for row 2\nM <- rbind( row.1, row.2 ) # row bind them into a matrix\nprint( M ) # and print it out...\n```\n\n```\n## [,1] [,2] [,3]\n## row.1 2 3 1\n## row.2 5 6 7\n```\n\n The variable `M` is a matrix, which we can confirm by using the `class()` function. Notice that, when we bound the two vectors together, R retained the names of the original variables as row names. We could delete these if we wanted by typing `rownames(M)<-NULL` , but I generally prefer having meaningful names attached to my variables, so I’ll keep them. In fact, let’s also add some highly unimaginative column names as well: \n\n```\ncolnames(M) <- c( \"col.1\", \"col.2\", \"col.3\" )\nprint(M)\n```\n\n```\n## col.1 col.2 col.3\n## row.1 2 3 1\n## row.2 5 6 7\n```\n\n You can use square brackets to subset a matrix in much the same way that you can for data frames, again specifying a row index and then a column index. For instance, `M[2,3]` pulls out the entry in the 2nd row and 3rd column of the matrix (i.e., `7` ), whereas `M[2,]` pulls out the entire 2nd row, and `M[,3]` pulls out the entire 3rd column. However, it’s worth noting that when you pull out a column, R will print the results horizontally, not vertically. The reason for this relates to how matrices (and arrays generally) are implemented. The original matrix `M` is treated as a two-dimensional objects, containing 2 rows and 3 columns. However, whenever you pull out a single row or a single column, the result is considered to be one-dimensional. As far as R is concerned there’s no real reason to distinguish between a one-dimensional object printed vertically (a column) and a one-dimensional object printed horizontally (a row), and it prints them all out horizontally.128 There is also a way of using only a single index, but due to the internal structure to how R defines a matrix, it works very differently to what we saw previously with data frames. \nThe single-index approach is illustrated in Table 7.5 but I don’t really want to focus on it since we’ll never really need it for this book, and matrices don’t play anywhere near as large a role in this book as data frames do. The reason for these differences is that for this is that, for both data frames and matrices, the “row and column” version exists to allow the human user to interact with the object in the psychologically meaningful way: since both data frames and matrices are basically just tables of data, it’s the same in each case. However, the single-index version is really a method for you to interact with the object in terms of its internal structure, and the internals for data frames and matrices are quite different.\n\nRow | Col.1 | Col.2 | Col.3 |\n | --- | --- | --- | --- |\nRow 1 | [1,1] | [1,2] | [1,3] |\nRow 2 | [2,1] | [2,2] | [2,3] |\n\nRow | Col.1 | Col.2 | Col.3 |\n | --- | --- | --- | --- |\nRow 1 | 1 | 3 | 5 |\nRow 2 | 2 | 4 | 6 |\n\nThe critical difference between a data frame and a matrix is that, in a data frame, we have this notion that each of the columns corresponds to a different variable: as a consequence, the columns in a data frame can be of different data types. The first column could be numeric, and the second column could contain character strings, and the third column could be logical data. In that sense, there is a fundamental asymmetry build into a data frame, because of the fact that columns represent variables (which can be qualitatively different to each other) and rows represent cases (which cannot). Matrices are intended to be thought of in a different way. At a fundamental level, a matrix really is just one variable: it just happens that this one variable is formatted into rows and columns. If you want a matrix of numeric data, every single element in the matrix must be a number. If you want a matrix of character strings, every single element in the matrix must be a character string. If you try to mix data of different types together, then R will either spit out an error, or quietly coerce the underlying data into a list. If you want to find out what class R secretly thinks the data within the matrix is, you need to do something like this:\n `class( M[1] )` `## [1] \"numeric\"` You can’t type `class(M)` , because all that will happen is R will tell you that `M` is a matrix: we’re not interested in the class of the matrix itself, we want to know what class the underlying data is assumed to be. Anyway, to give you a sense of how R enforces this, let’s try to change one of the elements of our numeric matrix into a character string: \n\n```\nM[1,2] <- \"text\"\nM\n```\n\n```\n## col.1 col.2 col.3\n## row.1 \"2\" \"text\" \"1\" \n## row.2 \"5\" \"6\" \"7\"\n```\n\n It looks as if R has coerced all of the data in our matrix into character strings. And in fact, if we now typed in `class(M[1])` we’d see that this is exactly what has happened. If you alter the contents of one element in a matrix, R will change the underlying data type as necessary. There’s only one more thing I want to talk about regarding matrices. The concept behind a matrix is very much a mathematical one, and in mathematics a matrix is a most definitely a two-dimensional object. However, when doing data analysis, we often have reasons to want to use higher dimensional tables (e.g., sometimes you need to cross-tabulate three variables against each other). You can’t do this with matrices, but you can do it with arrays. An array is just like a matrix, except it can have more than two dimensions if you need it to. In fact, as far as R is concerned a matrix is just a special kind of array, in much the same way that a data frame is a special kind of list. I don’t want to talk about arrays too much, but I will very briefly show you an example of what a 3D array looks like. To that end, let’s cross tabulate the `speaker` and `utterance` variables from the `nightgarden.Rdata` data file, but we’ll add a third variable to the cross-tabs this time, a logical variable which indicates whether or not I was still awake at this point in the show: \n\n```\ndan.awake <- c( TRUE,TRUE,TRUE,TRUE,TRUE,FALSE,FALSE,FALSE,FALSE,FALSE )\n```\n\n Now that we’ve got all three variables in the workspace (assuming you loaded the `nightgarden.Rdata` data earlier in the chapter) we can construct our three way cross-tabulation, using the `table()` function. \n\n```\nxtab.3d <- table( speaker, utterance, dan.awake )\nxtab.3d\n```\n\n```\n## , , dan.awake = FALSE\n## \n## utterance\n## speaker ee onk oo pip\n## makka-pakka 0 2 0 2\n## tombliboo 0 0 1 0\n## upsy-daisy 0 0 0 0\n## \n## , , dan.awake = TRUE\n## \n## utterance\n## speaker ee onk oo pip\n## makka-pakka 0 0 0 0\n## tombliboo 1 0 0 0\n## upsy-daisy 0 2 0 2\n```\n\n Hopefully this output is fairly straightforward: because R can’t print out text in three dimensions, what it does is show a sequence of 2D slices through the 3D table. That is, the\n\n```\n, , dan.awake = FALSE\n```\n\npart indicates that the 2D table that follows below shows the 2D cross-tabulation of `speaker` against `utterance` only for the `dan.awake = FALSE` instances, and so on.129 \n\n### 7.11.2 Ordered factors\n\n One topic that I neglected to mention when discussing factors previously (Section 4.7 is that there are actually two different types of factor in R, unordered factors and ordered factors. An unordered factor corresponds to a nominal scale variable, and all of the factors we’ve discussed so far in this book have been unordered (as will all the factors used anywhere else except in this section). However, it’s often very useful to explicitly tell R that your variable is ordinal scale, and if so you need to declare it to be an ordered factor. For instance, earlier in this chapter we made use of a variable consisting of Likert scale data, which we represented as the `likert.raw` variable: `likert.raw` \n\n We can declare this to be an ordered factor in by using the `factor()` function, and setting `ordered = TRUE` . To illustrate how this works, let’s create an ordered factor called `likert.ordinal` and have a look at it: \n\n```\nlikert.ordinal <- factor( x = likert.raw, # the raw data \n levels = seq(7,1,-1), # strongest agreement is 1, weakest is 7\n ordered = TRUE ) # and it's ordered\nprint( likert.ordinal )\n```\n\n```\n## [1] 1 7 3 4 4 4 2 6 5 5\n## Levels: 7 < 6 < 5 < 4 < 3 < 2 < 1\n```\n\n Notice that when we print out the ordered factor, R explicitly tells us what order the levels come in. Because I wanted to order my levels in terms of increasing strength of agreement, and because a response of 1 corresponded to the strongest agreement and 7 to the strongest disagreement, it was important that I tell R to encode 7 as the lowest value and 1 as the largest. Always check this when creating an ordered factor: it’s very easy to accidentally encode your data “upside down” if you’re not paying attention. In any case, note that we can (and should) attach meaningful names to these factor levels by using the `levels()` function, like this: \n\n```\nlevels( likert.ordinal ) <- c( \"strong.disagree\", \"disagree\", \"weak.disagree\", \n \"neutral\", \"weak.agree\", \"agree\", \"strong.agree\" )\nprint( likert.ordinal )\n```\n\n```\n## [1] strong.agree strong.disagree weak.agree neutral \n## [5] neutral neutral agree disagree \n## [9] weak.disagree weak.disagree \n## 7 Levels: strong.disagree < disagree < weak.disagree < ... < strong.agree\n```\n\nOne nice thing about using ordered factors is that there are a lot of analyses for which R automatically treats ordered factors differently from unordered factors, and generally in a way that is more appropriate for ordinal data. However, since I don’t discuss that in this book, I won’t go into details. Like so many things in this chapter, my main goal here is to make you aware that R has this capability built into it; so if you ever need to start thinking about ordinal scale variables in more detail, you have at least some idea where to start looking!\n\n### 7.11.3 Dates and times\n\n Times and dates are very annoying types of data. To a first approximation we can say that there are 365 days in a year, 24 hours in a day, 60 minutes in an hour and 60 seconds in a minute, but that’s not quite correct. The length of the solar day is not exactly 24 hours, and the length of solar year is not exactly 365 days, so we have a complicated system of corrections that have to be made to keep the time and date system working. On top of that, the measurement of time is usually taken relative to a local time zone, and most (but not all) time zones have both a standard time and a daylight savings time, though the date at which the switch occurs is not at all standardised. So, as a form of data, times and dates suck. Unfortunately, they’re also important. Sometimes it’s possible to avoid having to use any complicated system for dealing with times and dates. Often you just want to know what year something happened in, so you can just use numeric data: in quite a lot of situations something as simple as `this.year <- 2011` works just fine. If you can get away with that for your application, this is probably the best thing to do. However, sometimes you really do need to know the actual date. Or, even worse, the actual time. In this section, I’ll very briefly introduce you to the basics of how R deals with date and time data. As with a lot of things in this chapter, I won’t go into details because I don’t use this kind of data anywhere else in the book. The goal here is to show you the basics of what you need to do if you ever encounter this kind of data in real life. And then we’ll all agree never to speak of it again. To start with, let’s talk about the date. As it happens, modern operating systems are very good at keeping track of the time and date, and can even handle all those annoying timezone issues and daylight savings pretty well. So R takes the quite sensible view that it can just ask the operating system what the date is. We can pull the date using the `Sys.Date()` function: \n\n```\ntoday <- Sys.Date() # ask the operating system for the date\nprint(today) # display the date\n```\n\n `## [1] \"2019-01-11\"` Okay, that seems straightforward. But, it does rather look like `today` is just a character string, doesn’t it? That would be a problem, because dates really do have a numeric character to them, and it would be nice to be able to do basic addition and subtraction to them. Well, fear not. If you type in `class(today)` , R will tell you that the class of the `today` variable is `\"Date\"` . What this means is that, hidden underneath this text string that prints out an actual date, R actually has a numeric representation.130 What that means is that you actually can add and subtract days. For instance, if we add 1 to `today` , R will print out the date for tomorrow: `today + 1` `## [1] \"2019-01-12\"` \nLet’s see what happens when we add 365 days:\n `today + 365` `## [1] \"2020-01-11\"` This is particularly handy if you forget that a year is a leap year since in that case you’d probably get it wrong is doing this in your head. R provides a number of functions for working with dates, but I don’t want to talk about them in any detail. I will, however, make passing mention of the `weekdays()` function which will tell you what day of the week a particular date corresponded to, which is extremely convenient in some situations: `weekdays( today )` `## [1] \"Friday\"` I’ll also point out that you can use the `as.Date()` to convert various different kinds of data into dates. If the data happen to be strings formatted exactly according to the international standard notation (i.e., `yyyy-mm-dd` ) then the conversion is straightforward, because that’s the format that R expects to see by default. You can convert dates from other formats too, but it’s slightly trickier, and beyond the scope of this book. What about times? Well, times are even more annoying, so much so that I don’t intend to talk about them at all in this book, other than to point you in the direction of some vaguely useful things. R itself does provide you with some tools for handling time data, and in fact there are two separate classes of data that are used to represent times, known by the odd names `POSIXct` and `POSIXlt` . You can use these to work with times if you want to, but for most applications you would probably be better off downloading the `chron` package, which provides some much more user friendly tools for working with times and dates. \n\n## 7.12 Miscellaneous topics\n\nTo finish this chapter, I have a few topics to discuss that don’t really fit in with any of the other things in this chapter. They’re all kind of useful things to know about, but they are really just “odd topics” that don’t fit with the other examples. Here goes:\n\n### 7.12.1 The problems with floating point arithmetic\n\nIf I’ve learned nothing else about transfinite arithmetic (and I haven’t) it’s that infinity is a tedious and inconvenient concept. Not only is it annoying and counterintuitive at times, but it has nasty practical consequences. As we were all taught in high school, there are some numbers that cannot be represented as a decimal number of finite length, nor can they be represented as any kind of fraction between two whole numbers; \\(\\sqrt{2}\\), \\(\\pi\\) and \\(e\\), for instance. In everyday life we mostly don’t care about this. I’m perfectly happy to approximate \\(\\pi\\) as 3.14, quite frankly. Sure, this does produce some rounding errors from time to time, and if I’d used a more detailed approximation like 3.1415926535 I’d be less likely to run into those issues, but in all honesty I’ve never needed my calculations to be that precise. In other words, although our pencil and paper calculations cannot represent the number \\(\\pi\\) exactly as a decimal number, we humans are smart enough to realise that we don’t care. Computers, unfortunately, are dumb … and you don’t have to dig too deep in order to run into some very weird issues that arise because they can’t represent numbers perfectly. Here is my favourite example:\n `0.1 + 0.2 == 0.3` `## [1] FALSE` \nObviously, R has made a mistake here, because this is definitely the wrong answer. Your first thought might be that R is broken, and you might be considering switching to some other language. But you can reproduce the same error in dozens of different programming languages, so the issue isn’t specific to R. Your next thought might be that it’s something in the hardware, but you can get the same mistake on any machine. It’s something deeper than that.\n The fundamental issue at hand is floating point arithmetic, which is a fancy way of saying that computers will always round a number to fixed number of significant digits. The exact number of significant digits that the computer stores isn’t important to us:131 what matters is that whenever the number that the computer is trying to store is very long, you get rounding errors. That’s actually what’s happening with our example above. There are teeny tiny rounding errors that have appeared in the computer’s storage of the numbers, and these rounding errors have in turn caused the internal storage of `0.1 + 0.2` to be a tiny bit different from the internal storage of `0.3` . How big are these differences? Let’s ask R: `0.1 + 0.2 - 0.3` `## [1] 5.551115e-17` Very tiny indeed. No sane person would care about differences that small. But R is not a sane person, and the equality operator `==` is very literal minded. It returns a value of `TRUE` only when the two values that it is given are absolutely identical to each other. And in this case they are not. However, this only answers half of the question. The other half of the question is, why are we getting these rounding errors when we’re only using nice simple numbers like 0.1, 0.2 and 0.3? This seems a little counterintuitive. The answer is that, like most programming languages, R doesn’t store numbers using their decimal expansion (i.e., base 10: using digits 0, 1, 2 …, 9). We humans like to write our numbers in base 10 because we have 10 fingers. But computers don’t have fingers, they have transistors; and transistors are built to store 2 numbers not 10. So you can see where this is going: the internal storage of a number in R is based on its binary expansion (i.e., base 2: using digits 0 and 1). And unfortunately, here’s what the binary expansion of 0.1 looks like: \\[\n.1 \\mbox{(decimal)} = .00011001100110011... \\mbox{(binary)} \n\\] and the pattern continues forever. In other words, from the perspective of your computer, which likes to encode numbers in binary,132 0.1 is not a simple number at all. To a computer, 0.1 is actually an infinitely long binary number! As a consequence, the computer can make minor errors when doing calculations here. \nWith any luck you now understand the problem, which ultimately comes down to the twin fact that (1) we usually think in decimal numbers and computers usually compute with binary numbers, and (2) computers are finite machines and can’t store infinitely long numbers. The only questions that remain are when you should care and what you should do about it. Thankfully, you don’t have to care very often: because the rounding errors are small, the only practical situation that I’ve seen this issue arise for is when you want to test whether an arithmetic fact holds exactly numbers are identical (e.g., is someone’s response time equal to exactly \\(2 \\times 0.33\\) seconds?) This is pretty rare in real world data analysis, but just in case it does occur, it’s better to use a test that allows for a small tolerance. That is, if the difference between the two numbers is below a certain threshold value, we deem them to be equal for all practical purposes. For instance, you could do something like this, which asks whether the difference between the two numbers is less than a tolerance of \\(10^{-10}\\)\n\n```\nabs( 0.1 + 0.2 - 0.3 ) < 10^-10\n```\n\n `## [1] TRUE` To deal with this problem, there is a function called `all.equal()` that lets you test for equality but allows a small tolerance for rounding errors: \n\n```\nall.equal( 0.1 + 0.2, 0.3 )\n```\n\n `## [1] TRUE` \n\n### 7.12.2 The recycling rule\n\n There’s one thing that I haven’t mentioned about how vector arithmetic works in R, and that’s the recycling rule. The easiest way to explain it is to give a simple example. Suppose I have two vectors of different length, `x` and `y` , and I want to add them together. It’s not obvious what that actually means, so let’s have a look at what R does: \n\n```\nx <- c( 1,1,1,1,1,1 ) # x is length 6\ny <- c( 0,1 ) # y is length 2\nx + y # now add them:\n```\n\n `## [1] 1 2 1 2 1 2` As you can see from looking at this output, what R has done is “recycle” the value of the shorter vector (in this case `y` ) several times. That is, the first element of `x` is added to the first element of `y` , and the second element of `x` is added to the second element of `y` . However, when R reaches the third element of `x` there isn’t any corresponding element in `y` , so it returns to the beginning: thus, the third element of `x` is added to the first element of `y` . This process continues until R reaches the last element of `x` . And that’s all there is to it really. The same recycling rule also applies for subtraction, multiplication and division. The only other thing I should note is that, if the length of the longer vector isn’t an exact multiple of the length of the shorter one, R still does it, but also gives you a warning message: \n\n```\nx <- c( 1,1,1,1,1 ) # x is length 5\ny <- c( 0,1 ) # y is length 2\nx + y # now add them:\n```\n\n```\n## Warning in x + y: longer object length is not a multiple of shorter object\n## length\n```\n\n `## [1] 1 2 1 2 1` \n\n### 7.12.3 An introduction to environments\n\n In this section I want to ask a slightly different question: what is the workspace exactly? This question seems simple, but there’s a fair bit to it. This section can be skipped if you’re not really interested in the technical details. In the description I gave earlier, I talked about the workspace as an abstract location in which R variables are stored. That’s basically true, but it hides a couple of key details. For example, any time you have R open, it has to store lots of things in the computer’s memory, not just your variables. For example, the `who()` function that I wrote has to be stored in memory somewhere, right? If it weren’t I wouldn’t be able to use it. That’s pretty obvious. But equally obviously it’s not in the workspace either, otherwise you should have seen it! Here’s what’s happening. R needs to keep track of a lot of different things, so what it does is organise them into environments, each of which can contain lots of different variables and functions. Your workspace is one such environment. Every package that you have loaded is another environment. And every time you call a function, R briefly creates a temporary environment in which the function itself can work, which is then deleted after the calculations are complete. So, when I type in `search()` at the command line `search()` \n\n```\n## [1] \".GlobalEnv\" \"package:BayesFactor\" \"package:Matrix\" \n## [4] \"package:coda\" \"package:effects\" \"package:lmtest\" \n## [7] \"package:zoo\" \"package:gplots\" \"package:sciplot\" \n## [10] \"package:HistData\" \"package:MASS\" \"package:lsr\" \n## [13] \"package:psych\" \"package:car\" \"package:carData\" \n## [16] \"tools:rstudio\" \"package:stats\" \"package:graphics\" \n## [19] \"package:grDevices\" \"package:utils\" \"package:datasets\" \n## [22] \"package:methods\" \"Autoloads\" \"package:base\"\n```\n\n what I’m actually looking at is a sequence of environments. The first one, `\".GlobalEnv\"` is the technically-correct name for your workspace. No-one really calls it that: it’s either called the workspace or the global environment. And so when you type in `objects()` or `who()` what you’re really doing is listing the contents of `\".GlobalEnv\"` . But there’s no reason why we can’t look up the contents of these other environments using the `objects()` function (currently `who()` doesn’t support this). You just have to be a bit more explicit in your command. If I wanted to find out what is in the `package:stats` environment (i.e., the environment into which the contents of the `stats` package have been loaded), here’s what I’d get \n\n```\nhead(objects(\"package:stats\"))\n```\n\n```\n## [1] \"acf\" \"acf2AR\" \"add.scope\" \"add1\" \"addmargins\"\n## [6] \"aggregate\"\n```\n\nwhere this time I’ve used head() to hide a lot of output because the stats package contains about 500 functions. In fact, you can actually use the environment panel in Rstudio to browse any of your loaded packages (just click on the text that says “Global Environment” and you’ll see a dropdown menu like the one shown in Figure 7.2). The key thing to understand then, is that you can access any of the R variables and functions that are stored in one of these environments, precisely because those are the environments that you have loaded!133\n\n### 7.12.4 Attaching a data frame\n\n The last thing I want to mention in this section is the `attach()` function, which you often see referred to in introductory R books. Whenever it is introduced, the author of the book usually mentions that the `attach()` function can be used to “attach” the data frame to the search path, so you don’t have to use the `$` operator. That is, if I use the command `attach(df)` to attach my data frame, I no longer need to type `df$variable` , and instead I can just type `variable` . This is true as far as it goes, but it’s very misleading and novice users often get led astray by this description, because it hides a lot of critical details. Here is the very abridged description: when you use the `attach()` function, what R does is create an entirely new environment in the search path, just like when you load a package. Then, what it does is copy all of the variables in your data frame into this new environment. When you do this, however, you end up with two completely different versions of all your variables: one in the original data frame, and one in the new environment. Whenever you make a statement like `df$variable` you’re working with the variable inside the data frame; but when you just type `variable` you’re working with the copy in the new environment. And here’s the part that really upsets new users: changes to one version are not reflected in the other version. As a consequence, it’s really easy for R to end up with different value stored in the two different locations, and you end up really confused as a result. To be fair to the writers of the `attach()` function, the help documentation does actually state all this quite explicitly, and they even give some examples of how this can cause confusion at the bottom of the help page. And I can actually see how it can be very useful to create copies of your data in a separate location (e.g., it lets you make all kinds of modifications and deletions to the data without having to touch the original data frame). However, I don’t think it’s helpful for new users, since it means you have to be very careful to keep track of which copy you’re talking about. As a consequence of all this, for the purpose of this book I’ve decided not to use the `attach()` function. It’s something that you can investigate yourself once you’re feeling a little more confident with R, but I won’t do it here. \n\n## 7.13 Summary\n\nObviously, there’s no real coherence to this chapter. It’s just a grab bag of topics and tricks that can be handy to know about, so the best wrap up I can give here is just to repeat this list:\n\nThere are a number of books out there that extend this discussion. A couple of my favourites are Spector (2008) “Data Manipulation with R” and Teetor (2011) “R Cookbook”.\n\nThe quote comes from Home is the Hangman, published in 1975.↩\n * \n\nAs usual, you can assign this output to a variable. If you type\n\n```\nspeaker.freq <- table(speaker)\n```\n\nat the command prompt R will store the table as a variable. If you then type `class(speaker.freq)` you’ll see that the output is actually of class `table` . The key thing to note about a table object is that it’s basically a matrix (see Section 7.11.1.↩ * \n\nIt’s worth noting that there’s also a more powerful function called\n `recode()` function in the `car` package that I won’t discuss in this book but is worth looking into if you’re looking for a bit more flexibility.↩ * \n\nIf you’ve read further into the book, and are re-reading this section, then a good example of this would be someone choosing to do an ANOVA using\n `age.group3` as the grouping variable, instead of running a regression using `age` as a predictor. There are sometimes good reasons for do this: for instance, if the relationship between `age` and your outcome variable is highly non-linear, and you aren’t comfortable with trying to run non-linear regression! However, unless you really do have a good rationale for doing this, it’s best not to. It tends to introduce all sorts of other problems (e.g., the data will probably violate the normality assumption), and you can lose a lot of power.↩ * \n\nThe real answer is 0: $10 for a sandwich is a total ripoff so I should go next door and buy noodles.↩\n * \n\nAgain, I doubt that’s the right “real world” answer. I suspect that most sandwich shops won’t allow you to pay off your debts to them in sandwiches. But you get the idea.↩\n * \n\nActually, that’s a bit of a lie: the\n `log()` function is more flexible than that, and can be used to calculate logarithms in any base. The `log()` function has a `base` argument that you can specify, which has a default value of \\(e\\). Thus `log10(1000)` is actually equivalent to \n\n```\nlog(x = 1000, base = 10)\n```\n\nIt’s also worth checking out the\n `match()` function↩ * \n\nIt also works on data frames if you ever feel the need to import all of your variables from the data frame into the workspace. This can be useful at times, though it’s not a good idea if you have large data sets or if you’re working with multiple data sets at once. In particular, if you do this, never forget that you now have two copies of all your variables, one in the workspace and another in the data frame.↩\n * \n\nYou can do this yourself using the\n `make.names()` function. In fact, this is itself a handy thing to know about. For example, if you want to convert the names of the variables in the `speech.by.char` list into valid R variable names, you could use a command like this: \n\n```\nnames(speech.by.char) <- make.names(names(speech.by.char))\n```\n\n. However, I won’t go into details here.↩ * \n\nConveniently, if you type\n `rownames(df) <- NULL` R will renumber all the rows from scratch. For the `df` data frame, the labels that currently run from 7 to 10 will be changed to go from 1 to 4.↩ * \n\nActually, you can make the\n `subset()` function behave this way by using the optional `drop` argument, but by default `subset()` does not drop, which is probably more sensible and more intuitive to novice users.↩ * \n\nSpecifically, recursive indexing, a handy tool in some contexts but not something that I want to discuss here.↩\n * \n\nNote for advanced users: both of these functions are just wrappers to the\n `matrix()` function, which is pretty flexible in terms of the ability to convert vectors into matrices. Also, while I’m on this topic, I’ll briefly mention the fact that if you’re a Matlab user and looking for an equivalent of Matlab’s `repmat()` function, I’d suggest checking out the `matlab` package which contains R versions of a lot of handy Matlab functions.↩ * \n\nThe function you need for that is called\n `as.data.frame()` .↩ * \n\nIn truth, I suspect that most of the cases when you can sensibly flip a data frame occur when all of the original variables are measurements of the same type (e.g., all variables are response times), and if so you could easily have chosen to encode your data as a matrix instead of as a data frame. But since people do sometimes prefer to work with data frames, I’ve written the\n `tFrame()` function for the sake of convenience. I don’t really think it’s something that is needed very often.↩ * \n\nThis limitation is deliberate, by the way: if you’re getting to the point where you want to do something more complicated, you should probably start learning how to use\n `reshape()` , `cast()` and `melt()` or some of other the more advanced tools. The `wideToLong()` and `longToWide()` functions are included only to help you out when you’re first starting to use R.↩ * \n\nTo be honest, it does bother me a little that the default value of\n `sep` is a space. Normally when I want to paste strings together I don’t want any separator character, so I’d prefer it if the default were `sep=\"\"` . To that end, it’s worth noting that there’s also a `paste0()` function, which is identical to `paste()` except that it always assumes that `sep=\"\"` . Type `?paste` for more information about this.↩ * \n\nNote that you can capture the output from\n `cat()` if you want to, but you have to be sneaky and use the `capture.output()` function. For example, the command \n\n```\nx <- capture.output(cat(hw,ng))\n```\n\nwould work just fine.↩ * \n\nSigh. For advanced users: R actually supports two different ways of specifying regular expressions. One is the POSIX standard, the other is to use Perl-style regular expressions. The default is generally POSIX. If you understand regular expressions, that probably made sense to you. If not, don’t worry. It’s not important.↩\n * \n\nI thank for this example.↩\n * \n\nIf you’re lucky.↩\n * \n\nYou can also use the\n `matrix()` command itself, but I think the “binding” approach is a little more intuitive.↩ * \n\nThis has some interesting implications for how matrix algebra is implemented in R (which I’ll admit I initially found odd), but that’s a little beyond the scope of this book. However, since there will be a small proportion of readers that do care, I’ll quickly outline the basic thing you need to get used to: when multiplying a matrix by a vector (or one-dimensional array) using the\n `\\%*\\%` operator R will attempt to interpret the vector (or 1D array) as either a row-vector or column-vector, depending on whichever one makes the multiplication work. That is, suppose \\(M\\) is the \\(2\\times3\\) matrix, and \\(v\\) is a \\(1 \\times 3\\) row vector. It is impossible to multiply \\(Mv\\), since the dimensions don’t conform, but you can multiply by the corresponding column vector, \\(Mv^t\\). So, if I set `v <- M[2,]` and then try to calculate `M \\%*\\% v` , which you’d think would fail, it actually works because R treats the one dimensional array as if it were a column vector for the purposes of matrix multiplication. Note that if both objects are one dimensional arrays/vectors, this leads to ambiguity since \\(vv^t\\) (inner product) and \\(v^tv\\) (outer product) yield different answers. In this situation, the `\\%*\\%` operator returns the inner product not the outer product. To understand all the details, check out the help documentation.↩ * \n\nI should note that if you type\n `class(xtab.3d)` you’ll discover that this is a `\"table\"` object rather than an `\"array\"` object. However, this labelling is only skin deep. The underlying data structure here is actually an array. Advanced users may wish to check this using the command \n\n```\nclass(unclass(xtab.3d))\n```\n\n, but it’s not important for our purposes. All I really want to do in this section is show you what the output looks like when you encounter a 3D array.↩ * \n\nDate objects are coded as the number of days that have passed since January 1, 1970.↩\n * \n\nFor advanced users: type\n `?double` for more information.↩ * \n\nOr at least, that’s the default. If all your numbers are integers (whole numbers), then you can explicitly tell R to store them as integers by adding an\n `L` suffix at the end of the number. That is, an assignment like `x <- 2L` tells R to assign `x` a value of 2, and to store it as an integer rather than as a binary expansion. Type `?integer` for more details.↩ * \n\nFor advanced users: that’s a little over simplistic in two respects. First, it’s a terribly imprecise way of talking about scoping. Second, it might give you the impression that all the variables in question are actually loaded into memory. That’s not quite true, since that would be very wasteful of memory. Instead R has a “lazy loading” mechanism, in which what R actually does is create a “promise” to load those objects if they’re actually needed. For details, check out the\n `delayedAssign()` function.↩\n\nDate: 2011-11-22\nCategories: \nTags: \n\n# Chapter 8 Basic programming\n\nMachine dreams hold a special vertigo.\n\n–\n\nUp to this point in the book I’ve tried hard to avoid using the word “programming” too much because – at least in my experience – it’s a word that can cause a lot of fear. For one reason or another, programming (like mathematics and statistics) is often perceived by people on the “outside” as a black art, a magical skill that can be learned only by some kind of super-nerd. I think this is a shame. It’s certainly true that advanced programming is a very specialised skill: several different skills actually, since there’s quite a lot of different kinds of programming out there. However, the basics of programming aren’t all that hard, and you can accomplish a lot of very impressive things just using those basics.\n\nWith that in mind, the goal of this chapter is to discuss a few basic programming concepts and how to apply them in R. However, before I do, I want to make one further attempt to point out just how non-magical programming really is, via one very simple observation: you already know how to do it. Stripped to its essentials, programming is nothing more (and nothing less) than the process of writing out a bunch of instructions that a computer can understand. To phrase this slightly differently, when you write a computer program, you need to write it in a programming language that the computer knows how to interpret. R is one such language. Although I’ve been having you type all your commands at the command prompt, and all the commands in this book so far have been shown as if that’s what I were doing, it’s also quite possible (and as you’ll see shortly, shockingly easy) to write a program using these R commands. In other words, if this is the first time reading this book, then you’re only one short chapter away from being able to legitimately claim that you can program in R, albeit at a beginner’s level.\n\n## 8.1 Scripts\n\n Computer programs come in quite a few different forms: the kind of program that we’re mostly interested in from the perspective of everyday data analysis using R is known as a script. The idea behind a script is that, instead of typing your commands into the R console one at a time, instead you write them all in a text file. Then, once you’ve finished writing them and saved the text file, you can get R to execute all the commands in your file by using the `source()` function. In a moment I’ll show you exactly how this is done, but first I’d better explain why you should care. \n\n### 8.1.1 Why use scripts?\n\nBefore discussing scripting and programming concepts in any more detail, it’s worth stopping to ask why you should bother. After all, if you look at the R commands that I’ve used everywhere else this book, you’ll notice that they’re all formatted as if I were typing them at the command line. Outside this chapter you won’t actually see any scripts. Do not be fooled by this. The reason that I’ve done it that way is purely for pedagogical reasons. My goal in this book is to teach statistics and to teach R. To that end, what I’ve needed to do is chop everything up into tiny little slices: each section tends to focus on one kind of statistical concept, and only a smallish number of R functions. As much as possible, I want you to see what each function does in isolation, one command at a time. By forcing myself to write everything as if it were being typed at the command line, it imposes a kind of discipline on me: it prevents me from piecing together lots of commands into one big script. From a teaching (and learning) perspective I think that’s the right thing to do… but from a data analysis perspective, it is not. When you start analysing real world data sets, you will rapidly find yourself needing to write scripts.\n\nTo understand why scripts are so very useful, it may be helpful to consider the drawbacks to typing commands directly at the command prompt. The approach that we’ve been adopting so far, in which you type commands one at a time, and R sits there patiently in between commands, is referred to as the interactive style. Doing your data analysis this way is rather like having a conversation … a very annoying conversation between you and your data set, in which you and the data aren’t directly speaking to each other, and so you have to rely on R to pass messages back and forth. This approach makes a lot of sense when you’re just trying out a few ideas: maybe you’re trying to figure out what analyses are sensible for your data, or maybe just you’re trying to remember how the various R functions work, so you’re just typing in a few commands until you get the one you want. In other words, the interactive style is very useful as a tool for exploring your data. However, it has a number of drawbacks:\n\nIt’s hard to save your work effectively. You can save the workspace, so that later on you can load any variables you created. You can save your plots as images. And you can even save the history or copy the contents of the R console to a file. Taken together, all these things let you create a reasonably decent record of what you did. But it does leave a lot to be desired. It seems like you ought to be able to save a single file that R could use (in conjunction with your raw data files) and reproduce everything (or at least, everything interesting) that you did during your data analysis.\n * \n\nIt’s annoying to have to go back to the beginning when you make a mistake. Suppose you’ve just spent the last two hours typing in commands. Over the course of this time you’ve created lots of new variables and run lots of analyses. Then suddenly you realise that there was a nasty typo in the first command you typed, so all of your later numbers are wrong. Now you have to fix that first command, and then spend another hour or so combing through the R history to try and recreate what you did.\n * \n\nYou can’t leave notes for yourself. Sure, you can scribble down some notes on a piece of paper, or even save a Word document that summarises what you did. But what you really want to be able to do is write down an English translation of your R commands, preferably right “next to” the commands themselves. That way, you can look back at what you’ve done and actually remember what you were doing. In the simple exercises we’ve engaged in so far, it hasn’t been all that hard to remember what you were doing or why you were doing it, but only because everything we’ve done could be done using only a few commands, and you’ve never been asked to reproduce your analysis six months after you originally did it! When your data analysis starts involving hundreds of variables, and requires quite complicated commands to work, then you really, really need to leave yourself some notes to explain your analysis to, well, yourself.\n * \n\nIt’s nearly impossible to reuse your analyses later, or adapt them to similar problems. Suppose that, sometime in January, you are handed a difficult data analysis problem. After working on it for ages, you figure out some really clever tricks that can be used to solve it. Then, in September, you get handed a really similar problem. You can sort of remember what you did, but not very well. You’d like to have a clean record of what you did last time, how you did it, and why you did it the way you did. Something like that would really help you solve this new problem.\n * \n\nIt’s hard to do anything except the basics. There’s a nasty side effect of these problems. Typos are inevitable. Even the best data analyst in the world makes a lot of mistakes. So the chance that you’ll be able to string together dozens of correct R commands in a row are very small. So unless you have some way around this problem, you’ll never really be able to do anything other than simple analyses.\n * \n\nIt’s difficult to share your work other people. Because you don’t have this nice clean record of what R commands were involved in your analysis, it’s not easy to share your work with other people. Sure, you can send them all the data files you’ve saved, and your history and console logs, and even the little notes you wrote to yourself, but odds are pretty good that no-one else will really understand what’s going on (trust me on this: I’ve been handed lots of random bits of output from people who’ve been analysing their data, and it makes very little sense unless you’ve got the original person who did the work sitting right next to you explaining what you’re looking at)\n Ideally, what you’d like to be able to do is something like this… Suppose you start out with a data set `myrawdata.csv` . What you want is a single document – let’s call it `mydataanalysis.R` – that stores all of the commands that you’ve used in order to do your data analysis. Kind of similar to the R history but much more focused. It would only include the commands that you want to keep for later. Then, later on, instead of typing in all those commands again, you’d just tell R to run all of the commands that are stored in `mydataanalysis.R` . Also, in order to help you make sense of all those commands, what you’d want is the ability to add some notes or comments within the file, so that anyone reading the document for themselves would be able to understand what each of the commands actually does. But these comments wouldn’t get in the way: when you try to get R to run `mydataanalysis.R` it would be smart enough would recognise that these comments are for the benefit of humans, and so it would ignore them. Later on you could tweak a few of the commands inside the file (maybe in a new file called `mynewdatanalaysis.R` ) so that you can adapt an old analysis to be able to handle a new problem. And you could email your friends and colleagues a copy of this file so that they can reproduce your analysis themselves. \nIn other words, what you want is a script.\n\n### 8.1.2 Our first script\n\n Okay then. Since scripts are so terribly awesome, let’s write one. To do this, open up a simple text editing program, like TextEdit (on a Mac) or Notebook (on Windows). Don’t use a fancy word processing program like Microsoft Word or OpenOffice: use the simplest program you can find. Open a new text document, and type some R commands, hitting enter after each command. Let’s try using `x <- \"hello world\"` and `print(x)` as our commands. Then save the document as `hello.R` , and remember to save it as a plain text file: don’t save it as a word document or a rich text file. Just a boring old plain text file. Also, when it asks you where to save the file, save it to whatever folder you’re using as your working directory in R. At this point, you should be looking at something like Figure 8.1. And if so, you have now successfully written your first R program. Because I don’t want to take screenshots for every single script, I’m going to present scripts using extracts formatted as follows: \n\n```\n## --- hello.R\nx <- \"hello world\"\nprint(x)\n```\n\nThe line at the top is the filename, and not part of the script itself. Below that, you can see the two R commands that make up the script itself. Next to each command I’ve included the line numbers. You don’t actually type these into your script, but a lot of text editors (including the one built into Rstudio that I’ll show you in a moment) will show line numbers, since it’s a very useful convention that allows you to say things like “line 1 of the script creates a new variable, and line 2 prints it out”.\n So how do we run the script? Assuming that the `hello.R` file has been saved to your working directory, then you can run the script using the following command: `source( \"hello.R\" )` If the script file is saved in a different directory, then you need to specify the path to the file, in exactly the same way that you would have to when loading a data file using `load()` . In any case, when you type this command, R opens up the script file: it then reads each command in the file in the same order that they appear in the file, and executes those commands in that order. The simple script that I’ve shown above contains two commands. The first one creates a variable `x` and the second one prints it on screen. So, when we run the script, this is what we see on screen: \n\n```\nsource(file.path(projecthome,\"scripts\",\"hello.R\"))\n```\n\n `## [1] \"hello world\"` If we inspect the workspace using a command like `who()` or `objects()` , we discover that R has created the new variable `x` within the workspace, and not surprisingly `x` is a character string containing the text `\"hello world\"` . And just like that, you’ve written your first program R. It really is that simple. \n\n### 8.1.3 Using Rstudio to write scripts\n\n In the example above I assumed that you were writing your scripts using a simple text editor. However, it’s usually more convenient to use a text editor that is specifically designed to help you write scripts. There’s a lot of these out there, and experienced programmers will all have their own personal favourites. For our purposes, however, we can just use the one built into Rstudio. To create new script file in R studio, go to the “File” menu, select the “New” option, and then click on “R script”. This will open a new window within the “source” panel. Then you can type the commands you want (or code as it is generally called when you’re typing the commands into a script file) and save it when you’re done. The nice thing about using Rstudio to do this is that it automatically changes the colour of the text to indicate which parts of the code are comments and which are parts are actual R commands (these colours are called syntax highlighting, but they’re not actually part of the file – it’s just Rstudio trying to be helpful. To see an example of this, let’s open up our `hello.R` script in Rstudio. To do this, go to the “File” menu again, and select “Open…”. Once you’ve opened the file, you should be looking at something like Figure 8.2. As you can see (if you’re looking at this book in colour) the character string “hello world” is highlighted in green. Using Rstudio for your text editor is convenient for other reasons too. Notice in the top right hand corner of Figure 8.2 there’s a little button that reads “Source”? If you click on that, Rstudio will construct the relevant `source()` command for you, and send it straight to the R console. So you don’t even have to type in the `source()` command, which actually I think is a great thing, because it really bugs me having to type all those extra keystrokes every time I want to run my script. Anyway, Rstudio provide several other convenient little tools to help make scripting easier, but I won’t discuss them here.135 \n\n### 8.1.4 Commenting your script\n\nWhen writing up your data analysis as a script, one thing that is generally a good idea is to include a lot of comments in the code. That way, if someone else tries to read it (or if you come back to it several days, weeks, months or years later) they can figure out what’s going on. As a beginner, I think it’s especially useful to comment thoroughly, partly because it gets you into the habit of commenting the code, and partly because the simple act of typing in an explanation of what the code does will help you keep it clear in your own mind what you’re trying to achieve. To illustrate this idea, consider the following script:\n\n```\n## --- itngscript.R\n# A script to analyse nightgarden.Rdata_\n# author: _\n# date: 22/11/2011_\n\n# Create a cross tabulation and print it out:\ncat( \"tabulating data...\\n\" )\nitng.table <- table( speaker, utterance )\nprint( itng.table )\n```\n\nYou’ll notice that I’ve gone a bit overboard with my commenting: at the top of the script I’ve explained the purpose of the script, who wrote it, and when it was written. Then, throughout the script file itself I’ve added a lot of comments explaining what each section of the code actually does. In real life people don’t tend to comment this thoroughly, but the basic idea is a very good one: you really do want your script to explain itself. Nevertheless, as you’d expect R completely ignores all of the commented parts. When we run this script, this is what we see on screen:\n\n```\n## --- itngscript.R\n# A script to analyse nightgarden.Rdata\n# author: \n# date: 22/11/2011\n\n```\n## loading data from nightgarden.Rdata...\n```\n\n```\nload(file.path(projecthome,\"data\",\"nightgarden.Rdata\"))\n\n# Create a cross tabulation and print it out:\ncat( \"tabulating data...\\n\" )\n```\n\n```\n## tabulating data...\n```\n\n```\nitng.table <- table( speaker, utterance )\nprint( itng.table )\n```\n\n Even here, notice that the script announces its behaviour. The first two lines of the output tell us a lot about what the script is actually doing behind the scenes (the code do to this corresponds to the two `cat()` commands on lines 8 and 12 of the script). It’s usually a pretty good idea to do this, since it helps ensure that the output makes sense when the script is executed. \n\n### 8.1.5 Differences between scripts and the command line\n\n For the most part, commands that you insert into a script behave in exactly the same way as they would if you typed the same thing in at the command line. The one major exception to this is that if you want a variable to be printed on screen, you need to explicitly tell R to print it. You can’t just type the name of the variable. For example, our original `hello.R` script produced visible output. The following script does not: \n\n```\n## --- silenthello.R\nx <- \"hello world\"\nx\n```\n\n It does still create the variable `x` when you `source()` the script, but it won’t print anything on screen. However, apart from the fact that scripts don’t use “auto-printing” as it’s called, there aren’t a lot of differences in the underlying mechanics. There are a few stylistic differences though. For instance, if you want to load a package at the command line, you would generally use the `library()` function. If you want do to it from a script, it’s conventional to use `require()` instead. The two commands are basically identical, the only difference being that if the package doesn’t exist, `require()` produces a warning whereas `library()` gives you an error. Stylistically, what this means is that if the `require()` command fails in your script, R will boldly continue on and try to execute the rest of the script. Often that’s what you’d like to see happen, so it’s better to use `require()` . Clearly, however, you can get by just fine using the `library()` command for everyday usage. \n\n### 8.1.6 Done!\n\nAt this point, you’ve learned the basics of scripting. You are now officially allowed to say that you can program in R, though you probably shouldn’t say it too loudly. There’s a lot more to learn, but nevertheless, if you can write scripts like these then what you are doing is in fact basic programming. The rest of this chapter is devoted to introducing some of the key commands that you need in order to make your programs more powerful; and to help you get used to thinking in terms of scripts, for the rest of this chapter I’ll write up most of my extracts as scripts.\n\n## 8.2 Loops\n\n The description I gave earlier for how a script works was a tiny bit of a lie. Specifically, it’s not necessarily the case that R starts at the top of the file and runs straight through to the end of the file. For all the scripts that we’ve seen so far that’s exactly what happens, and unless you insert some commands to explicitly alter how the script runs, that is what will always happen. However, you actually have quite a lot of flexibility in this respect. Depending on how you write the script, you can have R repeat several commands, or skip over different commands, and so on. This topic is referred to as flow control, and the first concept to discuss in this respect is the idea of a loop. The basic idea is very simple: a loop is a block of code (i.e., a sequence of commands) that R will execute over and over again until some termination criterion is met. Looping is a very powerful idea. There are three different ways to construct a loop in R, based on the `while` , `for` and `repeat` functions. I’ll only discuss the first two in this book. \n\n### 8.2.1 The\n\n `while` loop A `while` loop is a simple thing. The basic format of the loop looks like this: \n\n The code corresponding to CONDITION needs to produce a logical value, either `TRUE` or `FALSE` . Whenever R encounters a `while` statement, it checks to see if the CONDITION is `TRUE` . If it is, then R goes on to execute all of the commands inside the curly brackets, proceeding from top to bottom as usual. However, when it gets to the bottom of those statements, it moves back up to the `while` statement. Then, like the mindless automaton it is, it checks to see if the CONDITION is `TRUE` . If it is, then R goes on to execute all … well, you get the idea. This continues endlessly until at some point the CONDITION turns out to be `FALSE` . Once that happens, R jumps to the bottom of the loop (i.e., to the `}` character), and then continues on with whatever commands appear next in the script. To start with, let’s keep things simple, and use a `while` loop to calculate the smallest multiple of 17 that is greater than or equal to 1000. This is a very silly example since you can actually calculate it using simple arithmetic operations, but the point here isn’t to do something novel. The point is to show how to write a `while` loop. Here’s the script: \n\n```\n## --- whileexample.R\nx <- 0\nwhile ( x < 1000 ) {\n x <- x + 17\n}\nprint( x )\n```\n\n When we run this script, R starts at the top and creates a new variable called `x` and assigns it a value of 0. It then moves down to the loop, and “notices” that the condition here is `x < 1000` . Since the current value of `x` is zero, the condition is true, so it enters the body of the loop (inside the curly braces). There’s only one command here136 which instructs R to increase the value of `x` by 17. R then returns to the top of the loop, and rechecks the condition. The value of `x` is now 17, but that’s still less than 1000, so the loop continues. This cycle will continue for a total of 59 iterations, until finally `x` reaches a value of 1003 (i.e., \\(59 \\times 17 = 1003\\)). At this point, the loop stops, and R finally reaches line 5 of the script, prints out the value of `x` on screen, and then halts. Let’s watch: \n\n```\nsource(file.path(projecthome,\"scripts\",\"whileexample.R\"))\n```\n\n `## [1] 1003` \nTruly fascinating stuff.\n\n### 8.2.2 The\n\n `for` loop The `for` loop is also pretty simple, though not quite as simple as the `while` loop. The basic format of this loop goes like this: \n\n```\nfor ( VAR in VECTOR ) {\n STATEMENT1\n STATEMENT2\n ETC\n }\n```\n\n In a `for` loop, R runs a fixed number of iterations. We have a VECTOR which has several elements, each one corresponding to a possible value of the variable VAR. In the first iteration of the loop, VAR is given a value corresponding to the first element of VECTOR; in the second iteration of the loop VAR gets a value corresponding to the second value in VECTOR; and so on. Once we’ve exhausted all of the values in VECTOR, the loop terminates and the flow of the program continues down the script. \nOnce again, let’s use some very simple examples. Firstly, here is a program that just prints out the word “hello” three times and then stops:\n\n```\n## --- forexample.R\nfor ( i in 1:3 ) {\n print( \"hello\" )\n}\n```\n\n This is the simplest example of a `for` loop. The vector of possible values for the `i` variable just corresponds to the numbers from 1 to 3. Not only that, the body of the loop doesn’t actually depend on `i` at all. Not surprisingly, here’s what happens when we run it: \n\n```\nsource(file.path(projecthome,\"scripts\",\"forexample.R\"))\n```\n\n```\n## [1] \"hello\"\n## [1] \"hello\"\n## [1] \"hello\"\n```\n\n However, there’s nothing that stops you from using something non-numeric as the vector of possible values, as the following example illustrates. This time around, we’ll use a character vector to control our loop, which in this case will be a vector of `words` . And what we’ll do in the loop is get R to convert the word to upper case letters, calculate the length of the word, and print it out. Here’s the script: \n\n```\n## --- forexample2.R\n\n#the words_\nwords <- c(\"it\",\"was\",\"the\",\"dirty\",\"end\",\"of\",\"winter\")\n\n#loop over the words_\nfor ( w in words ) {\n\n w.length <- nchar( w ) # calculate the number of letters_\n W <- toupper( w ) # convert the word to upper case letters_\n msg <- paste( W, \"has\", w.length, \"letters\" ) # a message to print_\n print( msg ) # print it_\n\n}\n```\n\nAnd here’s the output:\n\n```\nsource(file.path(projecthome,\"scripts\",\"forexample2.R\"))\n```\n\n```\n## [1] \"IT has 2 letters\"\n## [1] \"WAS has 3 letters\"\n## [1] \"THE has 3 letters\"\n## [1] \"DIRTY has 5 letters\"\n## [1] \"END has 3 letters\"\n## [1] \"OF has 2 letters\"\n## [1] \"WINTER has 6 letters\"\n```\n\nAgain, pretty straightforward I hope.\n\n### 8.2.3 A more realistic example of a loop\n\nTo give you a sense of how you can use a loop in a more complex situation, let’s write a simple script to simulate the progression of a mortgage. Suppose we have a nice young couple who borrow $300000 from the bank, at an annual interest rate of 5%. The mortgage is a 30 year loan, so they need to pay it off within 360 months total. Our happy couple decide to set their monthly mortgage payment at $1600 per month. Will they pay off the loan in time or not? Only time will tell.137 Or, alternatively, we could simulate the whole process and get R to tell us. The script to run this is a fair bit more complicated.\n\n```\n## --- mortgage.R\n\n# set up\nmonth <- 0 # count the number of months\nbalance <- 300000 # initial mortgage balance\npayments <- 1600 # monthly payments\ninterest <- 0.05 # 5% interest rate per year\ntotal.paid <- 0 # track what you've paid the bank\n\n# convert annual interest to a monthly multiplier\nmonthly.multiplier <- (1+interest) ^ (1/12)\n\n# keep looping until the loan is paid off...\nwhile ( balance > 0 ) {\n\n # do the calculations for this month\n month <- month + 1 # one more month\n balance <- balance * monthly.multiplier # add the interest\n balance <- balance - payments # make the payments\n total.paid <- total.paid + payments # track the total paid\n\n # print the results on screen\n cat( \"month\", month, \": balance\", round(balance), \"\\n\")\n\n} # end of loop\n\n# print the total payments at the end\ncat(\"total payments made\", total.paid, \"\\n\" )\n```\n\n To explain what’s going on, let’s go through it carefully. In the first block of code (under `#set up` ) all we’re doing is specifying all the variables that define the problem. The loan starts with a `balance` of $300,000 owed to the bank on `month` zero, and at that point in time the `total.paid` money is nothing. The couple is making monthly `payments` of $1600, at an annual `interest` rate of 5%. Next, we convert the annual percentage interest into a monthly multiplier. That is, the number that you have to multiply the current balance by each month in order to produce an annual interest rate of 5%. An annual interest rate of 5% implies that, if no payments were made over 12 months the balance would end up being \\(1.05\\) times what it was originally, so the annual multiplier is \\(1.05\\). To calculate the monthly multiplier, we need to calculate the 12th root of 1.05 (i.e., raise 1.05 to the power of 1/12). We store this value in as the `monthly.multiplier` variable, which as it happens corresponds to a value of about 1.004. All of which is a rather long winded way of saying that the annual interest rate of 5% corresponds to a monthly interest rate of about 0.4%. Anyway… all of that is really just setting the stage. It’s not the interesting part of the script. The interesting part (such as it is) is the loop. The `while` statement on tells R that it needs to keep looping until the `balance` reaches zero (or less, since it might be that the final payment of $1600 pushes the balance below zero). Then, inside the body of the loop, we have two different blocks of code. In the first bit, we do all the number crunching. Firstly we increase the value `month` by 1. Next, the bank charges the interest, so the `balance` goes up. Then, the couple makes their monthly payment and the `balance` goes down. Finally, we keep track of the total amount of money that the couple has paid so far, by adding the `payments` to the running tally. After having done all this number crunching, we tell R to issue the couple with a very terse monthly statement, which just indicates how many months they’ve been paying the loan and how much money they still owe the bank. Which is rather rude of us really. I’ve grown attached to this couple and I really feel they deserve better than that. But, that’s banks for you. In any case, the key thing here is the tension between the increase in `balance` on and the decrease. As long as the decrease is bigger, then the balance will eventually drop to zero and the loop will eventually terminate. If not, the loop will continue forever! This is actually very bad programming on my part: I really should have included something to force R to stop if this goes on too long. However, I haven’t shown you how to evaluate “if” statements yet, so we’ll just have to hope that the author of the book has rigged the example so that the code actually runs. Hm. I wonder what the odds of that are? Anyway, assuming that the loop does eventually terminate, there’s one last line of code that prints out the total amount of money that the couple handed over to the bank over the lifetime of the loan. \nNow that I’ve explained everything in the script in tedious detail, let’s run it and see what happens:\n\n```\nsource(file.path(projecthome,\"scripts\",\"mortgage.R\"))\n```\n\n```\n## month 1 : balance 299622 \n## month 2 : balance 299243 \n## month 3 : balance 298862 \n## month 4 : balance 298480 \n## month 5 : balance 298096 \n## month 6 : balance 297710 \n## month 7 : balance 297323 \n## month 8 : balance 296934 \n## month 9 : balance 296544 \n## month 10 : balance 296152 \n## month 11 : balance 295759 \n## month 12 : balance 295364 \n## month 13 : balance 294967 \n## month 14 : balance 294569 \n## month 15 : balance 294169 \n## month 16 : balance 293768 \n## month 17 : balance 293364 \n## month 18 : balance 292960 \n## month 19 : balance 292553 \n## month 20 : balance 292145 \n## month 21 : balance 291735 \n## month 22 : balance 291324 \n## month 23 : balance 290911 \n## month 24 : balance 290496 \n## month 25 : balance 290079 \n## month 26 : balance 289661 \n## month 27 : balance 289241 \n## month 28 : balance 288820 \n## month 29 : balance 288396 \n## month 30 : balance 287971 \n## month 31 : balance 287545 \n## month 32 : balance 287116 \n## month 33 : balance 286686 \n## month 34 : balance 286254 \n## month 35 : balance 285820 \n## month 36 : balance 285385 \n## month 37 : balance 284947 \n## month 38 : balance 284508 \n## month 39 : balance 284067 \n## month 40 : balance 283625 \n## month 41 : balance 283180 \n## month 42 : balance 282734 \n## month 43 : balance 282286 \n## month 44 : balance 281836 \n## month 45 : balance 281384 \n## month 46 : balance 280930 \n## month 47 : balance 280475 \n## month 48 : balance 280018 \n## month 49 : balance 279559 \n## month 50 : balance 279098 \n## month 51 : balance 278635 \n## month 52 : balance 278170 \n## month 53 : balance 277703 \n## month 54 : balance 277234 \n## month 55 : balance 276764 \n## month 56 : balance 276292 \n## month 57 : balance 275817 \n## month 58 : balance 275341 \n## month 59 : balance 274863 \n## month 60 : balance 274382 \n## month 61 : balance 273900 \n## month 62 : balance 273416 \n## month 63 : balance 272930 \n## month 64 : balance 272442 \n## month 65 : balance 271952 \n## month 66 : balance 271460 \n## month 67 : balance 270966 \n## month 68 : balance 270470 \n## month 69 : balance 269972 \n## month 70 : balance 269472 \n## month 71 : balance 268970 \n## month 72 : balance 268465 \n## month 73 : balance 267959 \n## month 74 : balance 267451 \n## month 75 : balance 266941 \n## month 76 : balance 266428 \n## month 77 : balance 265914 \n## month 78 : balance 265397 \n## month 79 : balance 264878 \n## month 80 : balance 264357 \n## month 81 : balance 263834 \n## month 82 : balance 263309 \n## month 83 : balance 262782 \n## month 84 : balance 262253 \n## month 85 : balance 261721 \n## month 86 : balance 261187 \n## month 87 : balance 260651 \n## month 88 : balance 260113 \n## month 89 : balance 259573 \n## month 90 : balance 259031 \n## month 91 : balance 258486 \n## month 92 : balance 257939 \n## month 93 : balance 257390 \n## month 94 : balance 256839 \n## month 95 : balance 256285 \n## month 96 : balance 255729 \n## month 97 : balance 255171 \n## month 98 : balance 254611 \n## month 99 : balance 254048 \n## month 100 : balance 253483 \n## month 101 : balance 252916 \n## month 102 : balance 252346 \n## month 103 : balance 251774 \n## month 104 : balance 251200 \n## month 105 : balance 250623 \n## month 106 : balance 250044 \n## month 107 : balance 249463 \n## month 108 : balance 248879 \n## month 109 : balance 248293 \n## month 110 : balance 247705 \n## month 111 : balance 247114 \n## month 112 : balance 246521 \n## month 113 : balance 245925 \n## month 114 : balance 245327 \n## month 115 : balance 244727 \n## month 116 : balance 244124 \n## month 117 : balance 243518 \n## month 118 : balance 242911 \n## month 119 : balance 242300 \n## month 120 : balance 241687 \n## month 121 : balance 241072 \n## month 122 : balance 240454 \n## month 123 : balance 239834 \n## month 124 : balance 239211 \n## month 125 : balance 238585 \n## month 126 : balance 237958 \n## month 127 : balance 237327 \n## month 128 : balance 236694 \n## month 129 : balance 236058 \n## month 130 : balance 235420 \n## month 131 : balance 234779 \n## month 132 : balance 234136 \n## month 133 : balance 233489 \n## month 134 : balance 232841 \n## month 135 : balance 232189 \n## month 136 : balance 231535 \n## month 137 : balance 230879 \n## month 138 : balance 230219 \n## month 139 : balance 229557 \n## month 140 : balance 228892 \n## month 141 : balance 228225 \n## month 142 : balance 227555 \n## month 143 : balance 226882 \n## month 144 : balance 226206 \n## month 145 : balance 225528 \n## month 146 : balance 224847 \n## month 147 : balance 224163 \n## month 148 : balance 223476 \n## month 149 : balance 222786 \n## month 150 : balance 222094 \n## month 151 : balance 221399 \n## month 152 : balance 220701 \n## month 153 : balance 220000 \n## month 154 : balance 219296 \n## month 155 : balance 218590 \n## month 156 : balance 217880 \n## month 157 : balance 217168 \n## month 158 : balance 216453 \n## month 159 : balance 215735 \n## month 160 : balance 215014 \n## month 161 : balance 214290 \n## month 162 : balance 213563 \n## month 163 : balance 212833 \n## month 164 : balance 212100 \n## month 165 : balance 211364 \n## month 166 : balance 210625 \n## month 167 : balance 209883 \n## month 168 : balance 209138 \n## month 169 : balance 208390 \n## month 170 : balance 207639 \n## month 171 : balance 206885 \n## month 172 : balance 206128 \n## month 173 : balance 205368 \n## month 174 : balance 204605 \n## month 175 : balance 203838 \n## month 176 : balance 203069 \n## month 177 : balance 202296 \n## month 178 : balance 201520 \n## month 179 : balance 200741 \n## month 180 : balance 199959 \n## month 181 : balance 199174 \n## month 182 : balance 198385 \n## month 183 : balance 197593 \n## month 184 : balance 196798 \n## month 185 : balance 196000 \n## month 186 : balance 195199 \n## month 187 : balance 194394 \n## month 188 : balance 193586 \n## month 189 : balance 192775 \n## month 190 : balance 191960 \n## month 191 : balance 191142 \n## month 192 : balance 190321 \n## month 193 : balance 189496 \n## month 194 : balance 188668 \n## month 195 : balance 187837 \n## month 196 : balance 187002 \n## month 197 : balance 186164 \n## month 198 : balance 185323 \n## month 199 : balance 184478 \n## month 200 : balance 183629 \n## month 201 : balance 182777 \n## month 202 : balance 181922 \n## month 203 : balance 181063 \n## month 204 : balance 180201 \n## month 205 : balance 179335 \n## month 206 : balance 178466 \n## month 207 : balance 177593 \n## month 208 : balance 176716 \n## month 209 : balance 175836 \n## month 210 : balance 174953 \n## month 211 : balance 174065 \n## month 212 : balance 173175 \n## month 213 : balance 172280 \n## month 214 : balance 171382 \n## month 215 : balance 170480 \n## month 216 : balance 169575 \n## month 217 : balance 168666 \n## month 218 : balance 167753 \n## month 219 : balance 166836 \n## month 220 : balance 165916 \n## month 221 : balance 164992 \n## month 222 : balance 164064 \n## month 223 : balance 163133 \n## month 224 : balance 162197 \n## month 225 : balance 161258 \n## month 226 : balance 160315 \n## month 227 : balance 159368 \n## month 228 : balance 158417 \n## month 229 : balance 157463 \n## month 230 : balance 156504 \n## month 231 : balance 155542 \n## month 232 : balance 154576 \n## month 233 : balance 153605 \n## month 234 : balance 152631 \n## month 235 : balance 151653 \n## month 236 : balance 150671 \n## month 237 : balance 149685 \n## month 238 : balance 148695 \n## month 239 : balance 147700 \n## month 240 : balance 146702 \n## month 241 : balance 145700 \n## month 242 : balance 144693 \n## month 243 : balance 143683 \n## month 244 : balance 142668 \n## month 245 : balance 141650 \n## month 246 : balance 140627 \n## month 247 : balance 139600 \n## month 248 : balance 138568 \n## month 249 : balance 137533 \n## month 250 : balance 136493 \n## month 251 : balance 135449 \n## month 252 : balance 134401 \n## month 253 : balance 133349 \n## month 254 : balance 132292 \n## month 255 : balance 131231 \n## month 256 : balance 130166 \n## month 257 : balance 129096 \n## month 258 : balance 128022 \n## month 259 : balance 126943 \n## month 260 : balance 125861 \n## month 261 : balance 124773 \n## month 262 : balance 123682 \n## month 263 : balance 122586 \n## month 264 : balance 121485 \n## month 265 : balance 120380 \n## month 266 : balance 119270 \n## month 267 : balance 118156 \n## month 268 : balance 117038 \n## month 269 : balance 115915 \n## month 270 : balance 114787 \n## month 271 : balance 113654 \n## month 272 : balance 112518 \n## month 273 : balance 111376 \n## month 274 : balance 110230 \n## month 275 : balance 109079 \n## month 276 : balance 107923 \n## month 277 : balance 106763 \n## month 278 : balance 105598 \n## month 279 : balance 104428 \n## month 280 : balance 103254 \n## month 281 : balance 102074 \n## month 282 : balance 100890 \n## month 283 : balance 99701 \n## month 284 : balance 98507 \n## month 285 : balance 97309 \n## month 286 : balance 96105 \n## month 287 : balance 94897 \n## month 288 : balance 93683 \n## month 289 : balance 92465 \n## month 290 : balance 91242 \n## month 291 : balance 90013 \n## month 292 : balance 88780 \n## month 293 : balance 87542 \n## month 294 : balance 86298 \n## month 295 : balance 85050 \n## month 296 : balance 83797 \n## month 297 : balance 82538 \n## month 298 : balance 81274 \n## month 299 : balance 80005 \n## month 300 : balance 78731 \n## month 301 : balance 77452 \n## month 302 : balance 76168 \n## month 303 : balance 74878 \n## month 304 : balance 73583 \n## month 305 : balance 72283 \n## month 306 : balance 70977 \n## month 307 : balance 69666 \n## month 308 : balance 68350 \n## month 309 : balance 67029 \n## month 310 : balance 65702 \n## month 311 : balance 64369 \n## month 312 : balance 63032 \n## month 313 : balance 61688 \n## month 314 : balance 60340 \n## month 315 : balance 58986 \n## month 316 : balance 57626 \n## month 317 : balance 56261 \n## month 318 : balance 54890 \n## month 319 : balance 53514 \n## month 320 : balance 52132 \n## month 321 : balance 50744 \n## month 322 : balance 49351 \n## month 323 : balance 47952 \n## month 324 : balance 46547 \n## month 325 : balance 45137 \n## month 326 : balance 43721 \n## month 327 : balance 42299 \n## month 328 : balance 40871 \n## month 329 : balance 39438 \n## month 330 : balance 37998 \n## month 331 : balance 36553 \n## month 332 : balance 35102 \n## month 333 : balance 33645 \n## month 334 : balance 32182 \n## month 335 : balance 30713 \n## month 336 : balance 29238 \n## month 337 : balance 27758 \n## month 338 : balance 26271 \n## month 339 : balance 24778 \n## month 340 : balance 23279 \n## month 341 : balance 21773 \n## month 342 : balance 20262 \n## month 343 : balance 18745 \n## month 344 : balance 17221 \n## month 345 : balance 15691 \n## month 346 : balance 14155 \n## month 347 : balance 12613 \n## month 348 : balance 11064 \n## month 349 : balance 9509 \n## month 350 : balance 7948 \n## month 351 : balance 6380 \n## month 352 : balance 4806 \n## month 353 : balance 3226 \n## month 354 : balance 1639 \n## month 355 : balance 46 \n## month 356 : balance -1554 \n## total payments made 569600\n```\n\nSo our nice young couple have paid off their $300,000 loan in just 4 months shy of the 30 year term of their loan, at a bargain basement price of $568,046 (since 569600 - 1554 = 568046). A happy ending!\n\n## 8.3 Conditional statements\n\n A second kind of flow control that programming languages provide is the ability to evaluate conditional statements. Unlike loops, which can repeat over and over again, a conditional statement only executes once, but it can switch between different possible commands depending on a CONDITION that is specified by the programmer. The power of these commands is that they allow the program itself to make choices, and in particular, to make different choices depending on the context in which the program is run. The most prominent of example of a conditional statement is the `if` statement, and the accompanying `else` statement. The basic format of an `if` statement in R is as follows: \n\n And the execution of the statement is pretty straightforward. If the CONDITION is true, then R will execute the statements contained in the curly braces. If the CONDITION is false, then it dose not. If you want to, you can extend the `if` statement to include an `else` statement as well, leading to the following syntax: \n\n```\nif ( CONDITION ) {\n STATEMENT1\n STATEMENT2\n ETC\n } else {\n STATEMENT3\n STATEMENT4\n ETC\n }\n```\n\nAs you’d expect, the interpretation of this version is similar. If the CONDITION is true, then the contents of the first block of code (i.e., STATEMENT1, STATEMENT2, ETC) are executed; but if it is false, then the contents of the second block of code (i.e., STATEMENT3, STATEMENT4, ETC) are executed instead.\n To give you a feel for how you can use `if` and `else` to do something useful, the example that I’ll show you is a script that prints out a different message depending on what day of the week you run it. We can do this making use of some of the tools that we discussed in Section 7.11.3. Here’s the script: \n\n```\n## --- ifelseexample.R\n# find out what day it is...\ntoday <- Sys.Date() # pull the date from the system clock\nday <- weekdays( today ) # what day of the week it is_\n\n# now make a choice depending on the day...\nif ( day == \"Monday\" ) {\n print( \"I don't like Mondays\" )\n} else {\n print( \"I'm a happy little automaton\" )\n}\n```\n\nSince today happens to be a Friday, when I run the script here’s what happens:\n\n```\nsource(file.path(projecthome,\"scripts\",\"ifelseexample.R\"))\n```\n\n There are other ways of making conditional statements in R. In particular, the `ifelse()` function and the `switch()` functions can be very useful in different contexts. However, my main aim in this chapter is to briefly cover the very basics, so I’ll move on. \n\n## 8.4 Writing functions\n\nIn this section I want to talk about functions again. Functions were introduced in Section 3.5, but you’ve learned a lot about R since then, so we can talk about them in more detail. In particular, I want to show you how to create your own. To stick with the same basic framework that I used to describe loops and conditionals, here’s the syntax that you use to create a function:\n\n```\nFNAME <- function ( ARG1, ARG2, ETC ) {\n STATEMENT1\n STATEMENT2\n ETC\n return( VALUE )\n }\n```\n\nWhat this does is create a function with the name FNAME, which has arguments ARG1, ARG2 and so forth. Whenever the function is called, R executes the statements in the curly braces, and then outputs the contents of VALUE to the user. Note, however, that R does not execute the commands inside the function in the workspace. Instead, what it does is create a temporary local environment: all the internal statements in the body of the function are executed there, so they remain invisible to the user. Only the final results in the VALUE are returned to the workspace.\n To give a simple example of this, let’s create a function called `quadruple()` which multiplies its inputs by four. In keeping with the approach taken in the rest of the chapter, I’ll use a script to do this: \n\n```\n## --- functionexample.R\nquadruple <- function(x) {\n y <- x*4\n return(y)\n}\n```\n\nWhen we run this script, as follows\n\n```\nsource(file.path(projecthome,\"scripts\",\"functionexample.R\"))\n```\n\n nothing appears to have happened, but there is a new object created in the workspace called `quadruple` . Not surprisingly, if we ask R to tell us what kind of object it is, it tells us that it is a function: `class( quadruple )` `## [1] \"function\"` And now that we’ve created the `quadruple()` function, we can call it just like any other function And if I want to store the output as a variable, I can do this: \n\n```\nmy.var <- quadruple(10)\nprint(my.var)\n```\n\n `## [1] 40` An important thing to recognise here is that the two internal variables that the `quadruple()` function makes use of, `x` and `y` , stay internal. That is, if we inspect the contents of the workspace, \n\n```\nlibrary(lsr)\nwho()\n```\n\n```\n## -- Name -- -- Class -- -- Size --\n## afl.finalists factor 400 \n## afl.margins numeric 176 \n## afl2 data.frame 4296 x 2 \n## age numeric 11 \n## age.breaks numeric 4 \n## age.group factor 11 \n## age.group2 factor 11 \n## age.group3 factor 11 \n## age.labels character 3 \n## animals character 4 \n## balance numeric 1 \n## beers character 3 \n## cake.1 numeric 5 \n## cake.2 numeric 5 \n## cake.df data.frame 5 x 2 \n## cake.mat1 matrix 5 x 2 \n## cake.mat2 matrix 2 x 5 \n## cakes matrix 4 x 5 \n## cakes.flipped matrix 5 x 4 \n## choice data.frame 4 x 10 \n## choice.2 data.frame 16 x 6 \n## colour logical 1 \n## d.cor numeric 1 \n## dan.awake logical 10 \n## data data.frame 12 x 4 \n## day character 1 \n## describeImg list 0 \n## df data.frame 4 x 1 \n## drugs data.frame 10 x 8 \n## drugs.2 data.frame 30 x 5 \n## effort data.frame 10 x 2 \n## emphCol character 1 \n## emphColLight character 1 \n## emphGrey character 1 \n## eps logical 1 \n## fac factor 3 \n## fibonacci numeric 6 \n## Fibonacci numeric 7 \n## freq integer 17 \n## garden data.frame 5 x 3 \n## generateRLineTypes function \n## generateRPointShapes function \n## height numeric 1 \n## hw character 2 \n## i integer 1 \n## interest numeric 1 \n## is.MP.speaking logical 5 \n## itng data.frame 10 x 2 \n## itng.table table 3 x 4 \n## likert.centred numeric 10 \n## likert.ordinal ordered 10 \n## likert.raw numeric 10 \n## M matrix 2 x 3 \n## makka.pakka character 4 \n## monkey character 1 \n## monkey.1 list 1 \n## month numeric 1 \n## monthly.multiplier numeric 1 \n## msg character 1 \n## my.var numeric 1 \n## ng character 2 \n## numbers numeric 3 \n## old list 66 \n## old.text character 1 \n## oneCorPlot function \n## opinion.dir numeric 10 \n## opinion.strength numeric 10 \n## out.0 data.frame 100 x 2 \n## out.1 data.frame 100 x 2 \n## out.2 data.frame 100 x 2 \n## parenthood data.frame 100 x 4 \n## payments numeric 1 \n## PJ character 1 \n## plotOne function \n## projecthome character 1 \n## quadruple function \n## row.1 numeric 3 \n## row.2 numeric 3 \n## some.data numeric 18 \n## speaker character 10 \n## speech.by.char list 3 \n## suspicious.cases logical 176 \n## teams character 17 \n## text character 2 \n## today Date 1 \n## tombliboo character 2 \n## total.paid numeric 1 \n## upsy.daisy character 4 \n## utterance character 10 \n## w character 1 \n## W character 1 \n## w.length integer 1 \n## width numeric 1 \n## words character 7 \n## x numeric 1 \n## X1 numeric 11 \n## X2 numeric 11 \n## X3 numeric 11 \n## X4 numeric 11 \n## xtab.3d table 3 x 4 x 2 \n## y numeric 2 \n## Y1 numeric 11 \n## Y2 numeric 11 \n## Y3 numeric 11 \n## Y4 numeric 11\n```\n\n we see everything in our workspace from this chapter including the `quadruple()` function itself, as well as the `my.var` variable that we just created. \nNow that we know how to create our own functions in R, it’s probably a good idea to talk a little more about some of the other properties of functions that I’ve been glossing over. To start with, let’s take this opportunity to type the name of the function at the command line without the parentheses:\n `quadruple` \n\n```\n## function(x) {\n## y <- x*4\n## return(y)\n## }\n```\n\n As you can see, when you type the name of a function at the command line, R prints out the underlying source code that we used to define the function in the first place. In the case of the `quadruple()` function, this is quite helpful to us – we can read this code and actually see what the function does. For other functions, this is less helpful, as we saw back in Section 3.5 when we tried typing `citation` rather than `citation()` . \n\n### 8.4.1 Function arguments revisited\n\n Okay, now that we are starting to get a sense for how functions are constructed, let’s have a look at two, slightly more complicated functions that I’ve created. The source code for these functions is contained within the `functionexample2.R` and `functionexample3.R` scripts. Let’s start by looking at the first one: \n\n```\n## --- functionexample2.R\npow <- function( x, y = 1) {\n out <- x^y # raise x to the power y\n return( out )\n}\n```\n\n and if we type\n\n```\nsource(\"functionexample2.R\")\n```\n\nto load the `pow()` function into our workspace, then we can make use of it. As you can see from looking at the code for this function, it has two arguments `x` and `y` , and all it does is raise `x` to the power of `y` . For instance, this command `pow(x=3, y=2)` `## [1] 9` calculates the value of \\(3^2\\). The interesting thing about this function isn’t what it does, since R already has has perfectly good mechanisms for calculating powers. Rather, notice that when I defined the function, I specified `y=1` when listing the arguments? That’s the default value for `y` . So if we enter a command without specifying a value for `y` , then the function assumes that we want `y=1` : `pow( x=3 )` `## [1] 3` However, since I didn’t specify any default value for `x` when I defined the `pow()` function, we always need to input a value for `x` . If we don’t R will spit out an error message. So now you know how to specify default values for an argument. The other thing I should point out while I’m on this topic is the use of the `...` argument. The `...` argument is a special construct in R which is only used within functions. It is used as a way of matching against multiple user inputs: in other words, `...` is used as a mechanism to allow the user to enter as many inputs as they like. I won’t talk at all about the low-level details of how this works at all, but I will show you a simple example of a function that makes use of it. To that end, consider the following script: \n\n```\n## --- functionexample3.R\ndoubleMax <- function( ... ) { \n max.val <- max( ... ) # find the largest value in ... \n out <- 2 * max.val # double it\n return( out )\n}\n```\n\n When we type\n\n```\nsource(\"functionexample3.R\")\n```\n\n, R creates the `doubleMax()` function. You can type in as many inputs as you like. The `doubleMax()` function identifies the largest value in the inputs, by passing all the user inputs to the `max()` function, and then doubles it. For example: `doubleMax( 1,2,5 )` `## [1] 10` \n\n### 8.4.2 There’s more to functions than this\n\nThere’s a lot of other details to functions that I’ve hidden in my description in this chapter. Experienced programmers will wonder exactly how the “scoping rules” work in R,138 or want to know how to use a function to create variables in other environments139, or if function objects can be assigned as elements of a list140 and probably hundreds of other things besides. However, I don’t want to have this discussion get too cluttered with details, so I think it’s best – at least for the purposes of the current book – to stop here.\n\n## 8.5 Implicit loops\n\n There’s one last topic I want to discuss in this chapter. In addition to providing the explicit looping structures via `while` and `for` , R also provides a collection of functions for implicit loops. What I mean by this is that these are functions that carry out operations very similar to those that you’d normally use a loop for. However, instead of typing out the whole loop, the whole thing is done with a single command. The main reason why this can be handy is that – due to the way that R is written – these implicit looping functions are usually about to do the same calculations much faster than the corresponding explicit loops. In most applications that beginners might want to undertake, this probably isn’t very important, since most beginners tend to start out working with fairly small data sets and don’t usually need to undertake extremely time consuming number crunching. However, because you often see these functions referred to in other contexts, it may be useful to very briefly discuss a few of them. The first and simplest of these functions is `sapply()` . The two most important arguments to this function are `X` , which specifies a vector containing the data, and `FUN` , which specifies the name of a function that should be applied to each element of the data vector. The following example illustrates the basics of how it works: \n\n```\nwords <- c(\"along\", \"the\", \"loom\", \"of\", \"the\", \"land\")\nsapply( X = words, FUN = nchar )\n```\n\n```\n## along the loom of the land \n## 5 3 4 2 3 4\n```\n\n Notice how similar this is to the second example of a `for` loop in Section 8.2.2. The `sapply()` function has implicitly looped over the elements of `words` , and for each such element applied the `nchar()` function to calculate the number of letters in the corresponding word. The second of these functions is `tapply()` , which has three key arguments. As before `X` specifies the data, and `FUN` specifies a function. However, there is also an `INDEX` argument which specifies a grouping variable.141 What the `tapply()` function does is loop over all of the different values that appear in the `INDEX` variable. Each such value defines a group: the `tapply()` function constructs the subset of `X` that corresponds to that group, and then applies the function `FUN` to that subset of the data. This probably sounds a little abstract, so let’s consider a specific example, using the `nightgarden.Rdata` file that we used in Chapter 7. \n\n```\ngender <- c( \"male\",\"male\",\"female\",\"female\",\"male\" )\nage <- c( 10,12,9,11,13 )\ntapply( X = age, INDEX = gender, FUN = mean )\n```\n\n```\n## female male \n## 10.00000 11.66667\n```\n\n In this extract, what we’re doing is using `gender` to define two different groups of people, and using their `ages` as the data. We then calculate the `mean()` of the ages, separately for the males and the females. A closely related function is `by()` . It actually does the same thing as `tapply()` , but the output is formatted a bit differently. This time around the three arguments are called `data` , `INDICES` and `FUN` , but they’re pretty much the same thing. An example of how to use the `by()` function is shown in the following extract: \n\n```\nby( data = age, INDICES = gender, FUN = mean )\n```\n\n```\n## gender: female\n## [1] 10\n## -------------------------------------------------------- \n## gender: male\n## [1] 11.66667\n```\n\n The `tapply()` and `by()` functions are quite handy things to know about, and are pretty widely used. However, although I do make passing reference to the `tapply()` later on, I don’t make much use of them in this book. Before moving on, I should mention that there are several other functions that work along similar lines, and have suspiciously similar names: `lapply` , `mapply` , `apply` , `vapply` , `rapply` and `eapply` . However, none of these come up anywhere else in this book, so all I wanted to do here is draw your attention to the fact that they exist. \n\n## 8.6 Summary\n\nIn this chapter I talked about several key programming concepts, things that you should know about if you want to start converting your simple scripts into full fledged programs:\n\n* Writing and using scripts (Section 8.1).\n * Using loops (Section 8.2) and implicit loops (Section 8.5).\n * Making conditional statements (Section 8.3)\n * Writing your own functions (Section 8.4)\n\nAs always, there are lots of things I’m ignoring in this chapter. It takes a lot of work to become a proper programmer, just as it takes a lot of work to be a proper psychologist or a proper statistician, and this book is certainly not going to provide you with all the tools you need to make that step. However, you’d be amazed at how much you can achieve using only the tools that I’ve covered up to this point. Loops, conditionals and functions are very powerful things, especially when combined with the various tools discussed in Chapters 3, 4 and 7. Believe it or not, you’re off to a pretty good start just by having made it to this point. If you want to keep going, there are (as always!) several other books you might want to look at. One that I’ve read and enjoyed is “A first course in statistical programming with R” Braun and Murdoch (2007), but quite a few people have suggested to me that “The art of programming with R” Matloff and Matloff (2011) is worth the effort too.\n\nThe quote comes from Count Zero (1986)↩\n * \n\nOkay, I lied. Sue me. One of the coolest features of Rstudio is the support for R Markdown, which lets you embed R code inside a Markdown document, and you can automatically publish your R Markdown to the web on Rstudio’s servers. If you’re the kind of nerd interested in this sort of thing, it’s really nice. And, yes, since I’m also that kind of nerd, of course I’m aware that iPython notebooks do the same thing and that R just nicked their idea. So what? It’s still cool. And anyway, this book isn’t called Learning Statistics with Python now, is it? Hm. Maybe I should write a Python version…↩\n * \n\nAs an aside: if there’s only a single command that you want to include inside your loop, then you don’t actually need to bother including the curly braces at all. However, until you’re comfortable programming in R I’d advise always using them, even when you don’t have to.↩\n * \n\nOkay, fine. This example is still a bit ridiculous, in three respects. Firstly, the bank absolutely will not let the couple pay less than the amount required to terminate the loan in 30 years. Secondly, a constant interest rate of 30 years is hilarious. Thirdly, you can solve this much more efficiently than through brute force simulation. However, we’re not exactly in the business of being realistic or efficient here.↩\n * \n\nLexical scope.↩\n * \n\nThe\n `assign()` function.↩ * \n\nYes.↩\n * \n\nOr a list of such variables.↩\n\n# Part IV. Statistical theory\n\nPart IV of the book is by far the most theoretical one, focusing as it does on the theory of statistical inference. Over the next three chapters my goal is to give you an introduction to probability theory (Chapter 9, sampling and estimation (Chapter 10 and statistical hypothesis testing (Chapter 11). Before we get started though, I want to say something about the big picture. Statistical inference is primarily about learning from data. The goal is no longer merely to describe our data, but to use the data to draw conclusions about the world. To motivate the discussion, I want to spend a bit of time talking about a philosophical puzzle known as the riddle of induction, because it speaks to an issue that will pop up over and over again throughout the book: statistical inference relies on assumptions. This sounds like a bad thing. In everyday life people say things like ``you should never make assumptions’’, and psychology classes often talk about assumptions and biases as bad things that we should try to avoid. From bitter personal experience I have learned never to say such things around philosophers.\n\n## On the limits of logical reasoning\n\nThe whole art of war consists in getting at what is on the other side of the hill, or, in other words, in learning what we do not know from what we do.\n\n– , 1st Duke of Wellington\n\nI am told that quote above came about as a consequence of a carriage ride across the countryside.142. He and his companion, , were playing a guessing game, each trying to predict what would be on the other side of each hill. In every case it turned out that Wellesley was right and Croker was wrong. Many years later when Wellesley was asked about the game, he explained that “the whole art of war consists in getting at what is on the other side of the hill”. Indeed, war is not special in this respect. All of life is a guessing game of one form or another, and getting by on a day to day basis requires us to make good guesses. So let’s play a guessing game of our own.\n Suppose you and I are observing the Wellesley-Croker competition, and after every three hills you and I have to predict who will win the next one, Wellesley or Croker. Let’s say that `W` refers to a Wellesley victory and `C` refers to a Croker victory. After three hills, our data set looks like this: `WWW` You: Three in a row doesn’t mean much. I suppose Wellesley might be better at this than Croker, but it might just be luck. Still, I’m a bit of a gambler. I’ll bet on Wellesley.\n Me: I agree that three in a row isn’t informative, and I see no reason to prefer Wellesley’s guesses over Croker’s. I can’t justify betting at this stage. Sorry. No bet for me.\n\nYour gamble paid off: three more hills go by, and Wellesley wins all three. Going into the next round of our game the score is 1-0 in favour of you, and our data set looks like this:\n `WWW WWW` \nI’ve organised the data into blocks of three so that you can see which batch corresponds to the observations that we had available at each step in our little side game. After seeing this new batch, our conversation continues:\n You: Six wins in a row for . This is starting to feel a bit suspicious. I’m still not certain, but I reckon that he’s going to win the next one too.\n Me: I guess I don’t see that. Sure, I agree that Wellesley has won six in a row, but I don’t see any logical reason why that means he’ll win the seventh one. No bet.\n You: Do your really think so? Fair enough, but my bet worked out last time, and I’m okay with my choice.\n\nFor a second time you were right, and for a second time I was wrong. Wellesley wins the next three hills, extending his winning record against Croker to 9-0. The data set available to us is now this:\n `WWW WWW WWW` \nAnd our conversation goes like this:\n You: Okay, this is pretty obvious. Wellesley is way better at this game. We both agree he’s going to win the next hill, right?\n Me: Is there really any logical evidence for that? Before we started this game, there were lots of possibilities for the first 10 outcomes, and I had no idea which one to expect.\n `WWW WWW WWW W` was one possibility, but so was `WCC CWC WWC C` and `WWW WWW WWW C` or even `CCC CCC CCC C` . Because I had no idea what would happen so I’d have said they were all equally likely. I assume you would have too, right? I mean, that’s what it means to say you have “no idea”, isn’t it? You: I suppose so.\n Me: Well then, the observations we’ve made logically rule out all possibilities except two:\n `WWW WWW WWW C` or `WWW WWW WWW W` . Both of these are perfectly consistent with the evidence we’ve encountered so far, aren’t they? You: Yes, of course they are. Where are you going with this?\n Me: So what’s changed then? At the start of our game, you’d have agreed with me that these are equally plausible, and none of the evidence that we’ve encountered has discriminated between these two possibilities. Therefore, both of these possibilities remain equally plausible, and I see no logical reason to prefer one over the other. So yes, while I agree with you that Wellesley’s run of 9 wins in a row is remarkable, I can’t think of a good reason to think he’ll win the 10th hill. No bet.\n You: I see your point, but I’m still willing to chance it. I’m betting on Wellesley.\n\nWellesley’s winning streak continues for the next three hills. The score in the Wellesley-Croker game is now 12-0, and the score in our game is now 3-0. As we approach the fourth round of our game, our data set is this:\n `WWW WWW WWW WWW` \nand the conversation continues:\n You: Oh yeah! Three more wins for Wellesley and another victory for me. Admit it, I was right about him! I guess we’re both betting on Wellesley this time around, right?\n Me: I don’t know what to think. I feel like we’re in the same situation we were in last round, and nothing much has changed. There are only two legitimate possibilities for a sequence of 13 hills that haven’t already been ruled out,\n `WWW WWW WWW WWW C` and `WWW WWW WWW WWW W` . It’s just like I said last time: if all possible outcomes were equally sensible before the game started, shouldn’t these two be equally sensible now given that our observations don’t rule out either one? I agree that it feels like Wellesley is on an amazing winning streak, but where’s the logical evidence that the streak will continue? You: I think you’re being unreasonable. Why not take a look at our scorecard, if you need evidence? You’re the expert on statistics and you’ve been using this fancy logical analysis, but the fact is you’re losing. I’m just relying on common sense and I’m winning. Maybe you should switch strategies.\n Me: Hm, that is a good point and I don’t want to lose the game, but I’m afraid I don’t see any logical evidence that your strategy is better than mine. It seems to me that if there were someone else watching our game, what they’d have observed is a run of three wins to you. Their data would look like this:\n `YYY` . Logically, I don’t see that this is any different to our first round of watching Wellesley and Croker. Three wins to you doesn’t seem like a lot of evidence, and I see no reason to think that your strategy is working out any better than mine. If I didn’t think that `WWW` was good evidence then for Wellesley being better than Croker at their game, surely I have no reason now to think that `YYY` is good evidence that you’re better at ours? You: Okay, now I think you’re being a jerk.\n Me: I don’t see the logical evidence for that.\n\n## Learning without making assumptions is a myth\n\nThere are lots of different ways in which we could dissect this dialogue, but since this is a statistics book pitched at psychologists, and not an introduction to the philosophy and psychology of reasoning, I’ll keep it brief. What I’ve described above is sometimes referred to as the riddle of induction: it seems entirely reasonable to think that a 12-0 winning record by Wellesley is pretty strong evidence that he will win the 13th game, but it is not easy to provide a proper logical justification for this belief. On the contrary, despite the obviousness of the answer, it’s not actually possible to justify betting on Wellesley without relying on some assumption that you don’t have any logical justification for.\n\nThe riddle of induction is most associated with the philosophical work of and more recently , but you can find examples of the problem popping up in fields as diverse literature (Lewis Carroll) and machine learning (the “no free lunch” theorem). There really is something weird about trying to “learn what we do not know from what we do”. The critical point is that assumptions and biases are unavoidable if you want to learn anything about the world. There is no escape from this, and it is just as true for statistical inference as it is for human reasoning. In the dialogue, I was taking aim at your perfectly sensible inferences as a human being, but the common sense reasoning that you relied on in is no different to what a statistician would have done. Your “common sense” half of the dialog relied an implicit assumption that there exists some difference in skill between Wellesley and Croker, and what you were doing was trying to work out what that difference in skill level would be. My “logical analysis” rejects that assumption entirely. All I was willing to accept is that there are sequences of wins and losses, and that I did not know which sequences would be observed. Throughout the dialogue, I kept insisting that all logically possible data sets were equally plausible at the start of the Wellesely-Croker game, and the only way in which I ever revised my beliefs was to eliminate those possibilities that were factually inconsistent with the observations.\n\nThat sounds perfectly sensible on its own terms. In fact, it even sounds like the hallmark of good deductive reasoning. Like , my approach was to rule out that which is impossible, in the hope that what would be left is the truth. Yet as we saw, ruling out the impossible never led me to make a prediction. On its own terms, everything I said in my half of the dialogue was entirely correct. An inability to make any predictions is the logical consequence of making “no assumptions”. In the end I lost our game, because you did make some assumptions and those assumptions turned out to be right. Skill is a real thing, and because you believed in the existence of skill you were able to learn that Wellesley had more of it than Croker. Had you relied on a less sensible assumption to drive your learning, you might not have won the game.\n\nUltimately there are two things you should take away from this. Firstly, as I’ve said, you cannot avoid making assumptions if you want to learn anything from your data. But secondly, once you realise that assumptions are necessary, it becomes important to make sure you make the right ones! A data analysis that relies on few assumptions is not necessarily better than one that makes many assumptions: it all depends on whether those assumptions are good ones for your data. As we go through the rest of this book I’ll often point out the assumptions that underpin a particular tool, and how you can check whether those assumptions are sensible.\n\nDate: 2010-10-30\nCategories: \nTags: \n\n# Chapter 9 Introduction to probability\n\n[God] has afforded us only the twilight … of Probability.\n\n– \n\nUp to this point in the book, we’ve discussed some of the key ideas in experimental design, and we’ve talked a little about how you can summarise a data set. To a lot of people, this is all there is to statistics: it’s about calculating averages, collecting all the numbers, drawing pictures, and putting them all in a report somewhere. Kind of like stamp collecting, but with numbers. However, statistics covers much more than that. In fact, descriptive statistics is one of the smallest parts of statistics, and one of the least powerful. The bigger and more useful part of statistics is that it provides tools that let you make inferences about data.\n\nOnce you start thinking about statistics in these terms – that statistics is there to help us draw inferences from data – you start seeing examples of it everywhere. For instance, here’s a tiny extract from a newspaper article in the Sydney Morning Herald (30 Oct 2010):\n “I have a tough job,” the Premier said in response to a poll which found her government is now the most unpopular Labor administration in polling history, with a primary vote of just 23 per cent.\n\nThis kind of remark is entirely unremarkable in the papers or in everyday life, but let’s have a think about what it entails. A polling company has conducted a survey, usually a pretty big one because they can afford it. I’m too lazy to track down the original survey, so let’s just imagine that they called 1000 NSW voters at random, and 230 (23%) of those claimed that they intended to vote for the ALP. For the 2010 Federal election, the Australian Electoral Commission reported 4,610,795 enrolled voters in NSW; so the opinions of the remaining 4,609,795 voters (about 99.98% of voters) remain unknown to us. Even assuming that no-one lied to the polling company the only thing we can say with 100% confidence is that the true ALP primary vote is somewhere between 230/4610795 (about 0.005%) and 4610025/4610795 (about 99.83%). So, on what basis is it legitimate for the polling company, the newspaper, and the readership to conclude that the ALP primary vote is only about 23%?\n\nThe answer to the question is pretty obvious: if I call 1000 people at random, and 230 of them say they intend to vote for the ALP, then it seems very unlikely that these are the only 230 people out of the entire voting public who actually intend to do so. In other words, we assume that the data collected by the polling company is pretty representative of the population at large. But how representative? Would we be surprised to discover that the true ALP primary vote is actually 24%? 29%? 37%? At this point everyday intuition starts to break down a bit. No-one would be surprised by 24%, and everybody would be surprised by 37%, but it’s a bit hard to say whether 29% is plausible. We need some more powerful tools than just looking at the numbers and guessing.\n\nInferential statistics provides the tools that we need to answer these sorts of questions, and since these kinds of questions lie at the heart of the scientific enterprise, they take up the lions share of every introductory course on statistics and research methods. However, the theory of statistical inference is built on top of probability theory. And it is to probability theory that we must now turn. This discussion of probability theory is basically background: there’s not a lot of statistics per se in this chapter, and you don’t need to understand this material in as much depth as the other chapters in this part of the book. Nevertheless, because probability theory does underpin so much of statistics, it’s worth covering some of the basics.\n\n## 9.1 How are probability and statistics different?\n\nBefore we start talking about probability theory, it’s helpful to spend a moment thinking about the relationship between probability and statistics. The two disciplines are closely related but they’re not identical. Probability theory is “the doctrine of chances”. It’s a branch of mathematics that tells you how often different kinds of events will happen. For example, all of these questions are things you can answer using probability theory:\n\n* What are the chances of a fair coin coming up heads 10 times in a row?\n * If I roll two six sided dice, how likely is it that I’ll roll two sixes?\n * How likely is it that five cards drawn from a perfectly shuffled deck will all be hearts?\n * What are the chances that I’ll win the lottery?\n\nNotice that all of these questions have something in common. In each case the “truth of the world” is known, and my question relates to the “what kind of events” will happen. In the first question I know that the coin is fair, so there’s a 50% chance that any individual coin flip will come up heads. In the second question, I know that the chance of rolling a 6 on a single die is 1 in 6. In the third question I know that the deck is shuffled properly. And in the fourth question, I know that the lottery follows specific rules. You get the idea. The critical point is that probabilistic questions start with a known model of the world, and we use that model to do some calculations. The underlying model can be quite simple. For instance, in the coin flipping example, we can write down the model like this: \\[ P(\\mbox{heads}) = 0.5 \\] which you can read as “the probability of heads is 0.5”. As we’ll see later, in the same way that percentages are numbers that range from 0% to 100%, probabilities are just numbers that range from 0 to 1. When using this probability model to answer the first question, I don’t actually know exactly what’s going to happen. Maybe I’ll get 10 heads, like the question says. But maybe I’ll get three heads. That’s the key thing: in probability theory, the model is known, but the data are not.\n\nSo that’s probability. What about statistics? Statistical questions work the other way around. In statistics, we do not know the truth about the world. All we have is the data, and it is from the data that we want to learn the truth about the world. Statistical questions tend to look more like these:\n\n* If my friend flips a coin 10 times and gets 10 heads, are they playing a trick on me?\n * If five cards off the top of the deck are all hearts, how likely is it that the deck was shuffled? - If the lottery commissioner’s spouse wins the lottery, how likely is it that the lottery was rigged?\n\nThis time around, the only thing we have are data. What I know is that I saw my friend flip the coin 10 times and it came up heads every time. And what I want to infer is whether or not I should conclude that what I just saw was actually a fair coin being flipped 10 times in a row, or whether I should suspect that my friend is playing a trick on me. The data I have look like this:\n\n```\nH H H H H H H H H H H\n```\n\nand what I’m trying to do is work out which “model of the world” I should put my trust in. If the coin is fair, then the model I should adopt is one that says that the probability of heads is 0.5; that is, \\(P(\\mbox{heads}) = 0.5\\). If the coin is not fair, then I should conclude that the probability of heads is not 0.5, which we would write as \\(P(\\mbox{heads}) \\neq 0.5\\). In other words, the statistical inference problem is to figure out which of these probability models is right. Clearly, the statistical question isn’t the same as the probability question, but they’re deeply connected to one another. Because of this, a good introduction to statistical theory will start with a discussion of what probability is and how it works.\n\n## 9.2 What does probability mean?\n\nLet’s start with the first of these questions. What is “probability”? It might seem surprising to you, but while statisticians and mathematicians (mostly) agree on what the rules of probability are, there’s much less of a consensus on what the word really means. It seems weird because we’re all very comfortable using words like “chance”, “likely”, “possible” and “probable”, and it doesn’t seem like it should be a very difficult question to answer. If you had to explain “probability” to a five year old, you could do a pretty good job. But if you’ve ever had that experience in real life, you might walk away from the conversation feeling like you didn’t quite get it right, and that (like many everyday concepts) it turns out that you don’t really know what it’s all about.\n\nSo I’ll have a go at it. Let’s suppose I want to bet on a soccer game between two teams of robots, Arduino Arsenal and C Milan. After thinking about it, I decide that there is an 80% probability that Arduino Arsenal winning. What do I mean by that? Here are three possibilities…\n\n* They’re robot teams, so I can make them play over and over again, and if I did that, Arduino Arsenal would win 8 out of every 10 games on average.\n * For any given game, I would only agree that betting on this game is only “fair” if a $1 bet on C Milan gives a $5 payoff (i.e. I get my $1 back plus a $4 reward for being correct), as would a $4 bet on Arduino Arsenal (i.e., my $4 bet plus a $1 reward).\n * My subjective “belief” or “confidence” in an Arduino Arsenal victory is four times as strong as my belief in a C Milan victory.\n\nEach of these seems sensible. However they’re not identical, and not every statistician would endorse all of them. The reason is that there are different statistical ideologies (yes, really!) and depending on which one you subscribe to, you might say that some of those statements are meaningless or irrelevant. In this section, I give a brief introduction the two main approaches that exist in the literature. These are by no means the only approaches, but they’re the two big ones.\n\n### 9.2.1 The frequentist view\n\nThe first of the two major approaches to probability, and the more dominant one in statistics, is referred to as the frequentist view, and it defines probability as a long-run frequency. Suppose we were to try flipping a fair coin, over and over again. By definition, this is a coin that has \\(P(H) = 0.5\\). What might we observe? One possibility is that the first 20 flips might look like this:\n\n```\nT,H,H,H,H,T,T,H,H,H,H,T,H,H,T,T,T,T,T,H\n```\n\nIn this case 11 of these 20 coin flips (55%) came up heads. Now suppose that I’d been keeping a running tally of the number of heads (which I’ll call \\(N_H\\)) that I’ve seen, across the first \\(N\\) flips, and calculate the proportion of heads \\(N_H / N\\) every time. Here’s what I’d get (I did literally flip coins to produce this!):\n\nnumber.of.flips | number.of.heads | proportion |\n | --- | --- | --- |\n1 | 0 | 0.00 |\n2 | 1 | 0.50 |\n3 | 2 | 0.67 |\n4 | 3 | 0.75 |\n5 | 4 | 0.80 |\n6 | 4 | 0.67 |\n7 | 4 | 0.57 |\n8 | 5 | 0.63 |\n9 | 6 | 0.67 |\n10 | 7 | 0.70 |\n11 | 8 | 0.73 |\n12 | 8 | 0.67 |\n13 | 9 | 0.69 |\n14 | 10 | 0.71 |\n15 | 10 | 0.67 |\n16 | 10 | 0.63 |\n17 | 10 | 0.59 |\n18 | 10 | 0.56 |\n19 | 10 | 0.53 |\n20 | 11 | 0.55 |\n\nNotice that at the start of the sequence, the proportion of heads fluctuates wildly, starting at .00 and rising as high as .80. Later on, one gets the impression that it dampens out a bit, with more and more of the values actually being pretty close to the “right” answer of .50. This is the frequentist definition of probability in a nutshell: flip a fair coin over and over again, and as \\(N\\) grows large (approaches infinity, denoted \\(N\\rightarrow \\infty\\)), the proportion of heads will converge to 50%. There are some subtle technicalities that the mathematicians care about, but qualitatively speaking, that’s how the frequentists define probability. Unfortunately, I don’t have an infinite number of coins, or the infinite patience required to flip a coin an infinite number of times. However, I do have a computer, and computers excel at mindless repetitive tasks. So I asked my computer to simulate flipping a coin 1000 times, and then drew a picture of what happens to the proportion \\(N_H / N\\) as \\(N\\) increases. Actually, I did it four times, just to make sure it wasn’t a fluke. The results are shown in Figure 9.1. As you can see, the proportion of observed heads eventually stops fluctuating, and settles down; when it does, the number at which it finally settles is the true probability of heads.\n\nThe frequentist definition of probability has some desirable characteristics. Firstly, it is objective: the probability of an event is necessarily grounded in the world. The only way that probability statements can make sense is if they refer to (a sequence of) events that occur in the physical universe.143 Secondly, it is unambiguous: any two people watching the same sequence of events unfold, trying to calculate the probability of an event, must inevitably come up with the same answer. However, it also has undesirable characteristics. Firstly, infinite sequences don’t exist in the physical world. Suppose you picked up a coin from your pocket and started to flip it. Every time it lands, it impacts on the ground. Each impact wears the coin down a bit; eventually, the coin will be destroyed. So, one might ask whether it really makes sense to pretend that an “infinite” sequence of coin flips is even a meaningful concept, or an objective one. We can’t say that an “infinite sequence” of events is a real thing in the physical universe, because the physical universe doesn’t allow infinite anything. More seriously, the frequentist definition has a narrow scope. There are lots of things out there that human beings are happy to assign probability to in everyday language, but cannot (even in theory) be mapped onto a hypothetical sequence of events. For instance, if a meteorologist comes on TV and says, “the probability of rain in Adelaide on 2 November 2048 is 60%” we humans are happy to accept this. But it’s not clear how to define this in frequentist terms. There’s only one city of Adelaide, and only 2 November 2048. There’s no infinite sequence of events here, just a once-off thing. Frequentist probability genuinely forbids us from making probability statements about a single event. From the frequentist perspective, it will either rain tomorrow or it will not; there is no “probability” that attaches to a single non-repeatable event. Now, it should be said that there are some very clever tricks that frequentists can use to get around this. One possibility is that what the meteorologist means is something like this: “There is a category of days for which I predict a 60% chance of rain; if we look only across those days for which I make this prediction, then on 60% of those days it will actually rain”. It’s very weird and counterintuitive to think of it this way, but you do see frequentists do this sometimes. And it will come up later in this book (see Section 10.5).\n\n### 9.2.2 The Bayesian view\n\nThe Bayesian view of probability is often called the subjectivist view, and it is a minority view among statisticians, but one that has been steadily gaining traction for the last several decades. There are many flavours of Bayesianism, making hard to say exactly what “the” Bayesian view is. The most common way of thinking about subjective probability is to define the probability of an event as the degree of belief that an intelligent and rational agent assigns to that truth of that event. From that perspective, probabilities don’t exist in the world, but rather in the thoughts and assumptions of people and other intelligent beings. However, in order for this approach to work, we need some way of operationalising “degree of belief”. One way that you can do this is to formalise it in terms of “rational gambling”, though there are many other ways. Suppose that I believe that there’s a 60% probability of rain tomorrow. If someone offers me a bet: if it rains tomorrow, then I win $5, but if it doesn’t rain then I lose $5. Clearly, from my perspective, this is a pretty good bet. On the other hand, if I think that the probability of rain is only 40%, then it’s a bad bet to take. Thus, we can operationalise the notion of a “subjective probability” in terms of what bets I’m willing to accept.\n\nWhat are the advantages and disadvantages to the Bayesian approach? The main advantage is that it allows you to assign probabilities to any event you want to. You don’t need to be limited to those events that are repeatable. The main disadvantage (to many people) is that we can’t be purely objective – specifying a probability requires us to specify an entity that has the relevant degree of belief. This entity might be a human, an alien, a robot, or even a statistician, but there has to be an intelligent agent out there that believes in things. To many people this is uncomfortable: it seems to make probability arbitrary. While the Bayesian approach does require that the agent in question be rational (i.e., obey the rules of probability), it does allow everyone to have their own beliefs; I can believe the coin is fair and you don’t have to, even though we’re both rational. The frequentist view doesn’t allow any two observers to attribute different probabilities to the same event: when that happens, then at least one of them must be wrong. The Bayesian view does not prevent this from occurring. Two observers with different background knowledge can legitimately hold different beliefs about the same event. In short, where the frequentist view is sometimes considered to be too narrow (forbids lots of things that that we want to assign probabilities to), the Bayesian view is sometimes thought to be too broad (allows too many differences between observers).\n\n### 9.2.3 What’s the difference? And who is right?\n\nNow that you’ve seen each of these two views independently, it’s useful to make sure you can compare the two. Go back to the hypothetical robot soccer game at the start of the section. What do you think a frequentist and a Bayesian would say about these three statements? Which statement would a frequentist say is the correct definition of probability? Which one would a Bayesian do? Would some of these statements be meaningless to a frequentist or a Bayesian? If you’ve understood the two perspectives, you should have some sense of how to answer those questions.\n\nOkay, assuming you understand the different, you might be wondering which of them is right? Honestly, I don’t know that there is a right answer. As far as I can tell there’s nothing mathematically incorrect about the way frequentists think about sequences of events, and there’s nothing mathematically incorrect about the way that Bayesians define the beliefs of a rational agent. In fact, when you dig down into the details, Bayesians and frequentists actually agree about a lot of things. Many frequentist methods lead to decisions that Bayesians agree a rational agent would make. Many Bayesian methods have very good frequentist properties.\n\nFor the most part, I’m a pragmatist so I’ll use any statistical method that I trust. As it turns out, that makes me prefer Bayesian methods, for reasons I’ll explain towards the end of the book, but I’m not fundamentally opposed to frequentist methods. Not everyone is quite so relaxed. For instance, consider , one of the towering figures of 20th century statistics and a vehement opponent to all things Bayesian, whose paper on the mathematical foundations of statistics referred to Bayesian probability as “an impenetrable jungle [that] arrests progress towards precision of statistical concepts” Fisher (1922b). Or the psychologist , who suggests that relying on frequentist methods could turn you into “a potent but sterile intellectual rake who leaves in his merry path a long train of ravished maidens but no viable scientific offspring” Meehl (1967). The history of statistics, as you might gather, is not devoid of entertainment.\n\nIn any case, while I personally prefer the Bayesian view, the majority of statistical analyses are based on the frequentist approach. My reasoning is pragmatic: the goal of this book is to cover roughly the same territory as a typical undergraduate stats class in psychology, and if you want to understand the statistical tools used by most psychologists, you’ll need a good grasp of frequentist methods. I promise you that this isn’t wasted effort. Even if you end up wanting to switch to the Bayesian perspective, you really should read through at least one book on the “orthodox” frequentist view. And since R is the most widely used statistical language for Bayesians, you might as well read a book that uses R. Besides, I won’t completely ignore the Bayesian perspective. Every now and then I’ll add some commentary from a Bayesian point of view, and I’ll revisit the topic in more depth in Chapter 17.\n\n## 9.3 Basic probability theory\n\nIdeological arguments between Bayesians and frequentists notwithstanding, it turns out that people mostly agree on the rules that probabilities should obey. There are lots of different ways of arriving at these rules. The most commonly used approach is based on the work of , one of the great Soviet mathematicians of the 20th century. I won’t go into a lot of detail, but I’ll try to give you a bit of a sense of how it works. And in order to do so, I’m going to have to talk about my pants.\n\n### 9.3.1 Introducing probability distributions\n\nOne of the disturbing truths about my life is that I only own 5 pairs of pants: three pairs of jeans, the bottom half of a suit, and a pair of tracksuit pants. Even sadder, I’ve given them names: I call them \\(X_1\\), \\(X_2\\), \\(X_3\\), \\(X_4\\) and \\(X_5\\). I really do: that’s why they call me Mister Imaginative. Now, on any given day, I pick out exactly one of pair of pants to wear. Not even I’m so stupid as to try to wear two pairs of pants, and thanks to years of training I never go outside without wearing pants anymore. If I were to describe this situation using the language of probability theory, I would refer to each pair of pants (i.e., each \\(X\\)) as an elementary event. The key characteristic of elementary events is that every time we make an observation (e.g., every time I put on a pair of pants), then the outcome will be one and only one of these events. Like I said, these days I always wear exactly one pair of pants, so my pants satisfy this constraint. Similarly, the set of all possible events is called a sample space. Granted, some people would call it a “wardrobe”, but that’s because they’re refusing to think about my pants in probabilistic terms. Sad.\n\nOkay, now that we have a sample space (a wardrobe), which is built from lots of possible elementary events (pants), what we want to do is assign a probability of one of these elementary events. For an event \\(X\\), the probability of that event \\(P(X)\\) is a number that lies between 0 and 1. The bigger the value of \\(P(X)\\), the more likely the event is to occur. So, for example, if \\(P(X) = 0\\), it means the event \\(X\\) is impossible (i.e., I never wear those pants). On the other hand, if \\(P(X) = 1\\) it means that event \\(X\\) is certain to occur (i.e., I always wear those pants). For probability values in the middle, it means that I sometimes wear those pants. For instance, if \\(P(X) = 0.5\\) it means that I wear those pants half of the time.\n\nAt this point, we’re almost done. The last thing we need to recognise is that “something always happens”. Every time I put on pants, I really do end up wearing pants (crazy, right?). What this somewhat trite statement means, in probabilistic terms, is that the probabilities of the elementary events need to add up to 1. This is known as the law of total probability, not that any of us really care. More importantly, if these requirements are satisfied, then what we have is a probability distribution. For example, this is an example of a probability distribution\n\nWhich.pants | Blue.jeans | Grey.jeans | Black.jeans | Black.suit | Blue.tracksuit |\n | --- | --- | --- | --- | --- | --- |\nLabel | \\(X_1\\) | \\(X_2\\) | \\(X_3\\) | \\(X_4\\) | \\(X_5\\) |\nProbability | \\(P(X_1) = .5\\) | \\(P(X_2) = .3\\) | \\(P(X_3) = .1\\) | \\(P(X_4) = 0\\) | \\(P(X_5) = .1\\) |\n\nEach of the events has a probability that lies between 0 and 1, and if we add up the probability of all events, they sum to 1. Awesome. We can even draw a nice bar graph (see Section 6.7) to visualise this distribution, as shown in Figure 9.2. And at this point, we’ve all achieved something. You’ve learned what a probability distribution is, and I’ve finally managed to find a way to create a graph that focuses entirely on my pants. Everyone wins!\n\nThe only other thing that I need to point out is that probability theory allows you to talk about non elementary events as well as elementary ones. The easiest way to illustrate the concept is with an example. In the pants example, it’s perfectly legitimate to refer to the probability that I wear jeans. In this scenario, the “Dan wears jeans” event said to have happened as long as the elementary event that actually did occur is one of the appropriate ones; in this case “blue jeans”, “black jeans” or “grey jeans”. In mathematical terms, we defined the “jeans” event \\(E\\) to correspond to the set of elementary events \\((X_1, X_2, X_3)\\). If any of these elementary events occurs, then \\(E\\) is also said to have occurred. Having decided to write down the definition of the \\(E\\) this way, it’s pretty straightforward to state what the probability \\(P(E)\\) is: we just add everything up. In this particular case \\[ P(E) = P(X_1) + P(X_2) + P(X_3) \\] and, since the probabilities of blue, grey and black jeans respectively are .5, .3 and .1, the probability that I wear jeans is equal to .9.\n\nAt this point you might be thinking that this is all terribly obvious and simple and you’d be right. All we’ve really done is wrap some basic mathematics around a few common sense intuitions. However, from these simple beginnings it’s possible to construct some extremely powerful mathematical tools. I’m definitely not going to go into the details in this book, but what I will do is list – in Table 9.1 – some of the other rules that probabilities satisfy. These rules can be derived from the simple assumptions that I’ve outlined above, but since we don’t actually use these rules for anything in this book, I won’t do so here.\n\nEnglish | Notation | NANA | Formula |\n | --- | --- | --- | --- |\nNot \\(A\\) | \\(P(\\neg A)\\) | = | \\(1-P(A)\\) |\n\\(A\\) or \\(B\\) | \\(P(A \\cup B)\\) | = | \\(P(A) + P(B) - P(A \\cap B)\\) |\n\\(A\\) and \\(B\\) | \\(P(A \\cap B)\\) | = | \\(P(A|B) P(B)\\) |\n\n## 9.4 The binomial distribution\n\nAs you might imagine, probability distributions vary enormously, and there’s an enormous range of distributions out there. However, they aren’t all equally important. In fact, the vast majority of the content in this book relies on one of five distributions: the binomial distribution, the normal distribution, the \\(t\\) distribution, the \\(\\chi^2\\) (“chi-square”) distribution and the \\(F\\) distribution. Given this, what I’ll do over the next few sections is provide a brief introduction to all five of these, paying special attention to the binomial and the normal. I’ll start with the binomial distribution, since it’s the simplest of the five.\n\n### 9.4.1 Introducing the binomial\n\nThe theory of probability originated in the attempt to describe how games of chance work, so it seems fitting that our discussion of the binomial distribution should involve a discussion of rolling dice and flipping coins. Let’s imagine a simple “experiment”: in my hot little hand I’m holding 20 identical six-sided dice. On one face of each die there’s a picture of a skull; the other five faces are all blank. If I proceed to roll all 20 dice, what’s the probability that I’ll get exactly 4 skulls? Assuming that the dice are fair, we know that the chance of any one die coming up skulls is 1 in 6; to say this another way, the skull probability for a single die is approximately \\(.167\\). This is enough information to answer our question, so let’s have a look at how it’s done.\n\nAs usual, we’ll want to introduce some names and some notation. We’ll let \\(N\\) denote the number of dice rolls in our experiment; which is often referred to as the size parameter of our binomial distribution. We’ll also use \\(\\theta\\) to refer to the the probability that a single die comes up skulls, a quantity that is usually called the success probability of the binomial.144 Finally, we’ll use \\(X\\) to refer to the results of our experiment, namely the number of skulls I get when I roll the dice. Since the actual value of \\(X\\) is due to chance, we refer to it as a random variable. In any case, now that we have all this terminology and notation, we can use it to state the problem a little more precisely. The quantity that we want to calculate is the probability that \\(X = 4\\) given that we know that \\(\\theta = .167\\) and \\(N=20\\). The general “form” of the thing I’m interested in calculating could be written as \\[ P(X \\ | \\ \\theta, N) \\] and we’re interested in the special case where \\(X=4\\), \\(\\theta = .167\\) and \\(N=20\\). There’s only one more piece of notation I want to refer to before moving on to discuss the solution to the problem. If I want to say that \\(X\\) is generated randomly from a binomial distribution with parameters \\(\\theta\\) and \\(N\\), the notation I would use is as follows: \\[ X \\sim \\mbox{Binomial}(\\theta, N) \\]\n\nYeah, yeah. I know what you’re thinking: notation, notation, notation. Really, who cares? Very few readers of this book are here for the notation, so I should probably move on and talk about how to use the binomial distribution. I’ve included the formula for the binomial distribution in Table 9.2, since some readers may want to play with it themselves, but since most people probably don’t care that much and because we don’t need the formula in this book, I won’t talk about it in any detail. Instead, I just want to show you what the binomial distribution looks like. To that end, Figure 9.3 plots the binomial probabilities for all possible values of \\(X\\) for our dice rolling experiment, from \\(X=0\\) (no skulls) all the way up to \\(X=20\\) (all skulls). Note that this is basically a bar chart, and is no different to the “pants probability” plot I drew in Figure 9.2. On the horizontal axis we have all the possible events, and on the vertical axis we can read off the probability of each of those events. So, the probability of rolling 4 skulls out of 20 times is about 0.20 (the actual answer is 0.2022036, as we’ll see in a moment). In other words, you’d expect that to happen about 20% of the times you repeated this experiment.\n\nBinomial | Normal |\n | --- | --- |\n\\(P(X | \\theta, N) = \\displaystyle\\frac{N!}{X! (N-X)!} \\theta^X (1-\\theta)^{N-X}\\) | \\(p(X | \\mu, \\sigma) = \\displaystyle\\frac{1}{\\sqrt{2\\pi}\\sigma} \\exp \\left( -\\frac{(X - \\mu)^2}{2\\sigma^2} \\right)\\) |\n\n### 9.4.2 Working with the binomial distribution in R\n\n Although some people find it handy to know the formulas in Table 9.2, most people just want to know how to use the distributions without worrying too much about the maths. To that end, R has a function called `dbinom()` that calculates binomial probabilities for us. The main arguments to the function are \n\n* \n `x` . This is a number, or vector of numbers, specifying the outcomes whose probability you’re trying to calculate. * \n `size` . This is a number telling R the size of the experiment. * \n `prob` . This is the success probability for any one trial in the experiment. So, in order to calculate the probability of getting `x = 4` skulls, from an experiment of `size = 20` trials, in which the probability of getting a skull on any one trial is `prob = 1/6` … well, the command I would use is simply this: \n\n```\ndbinom( x = 4, size = 20, prob = 1/6 )\n```\n\n `## [1] 0.2022036` \nTo give you a feel for how the binomial distribution changes when we alter the values of \\(\\theta\\) and \\(N\\), let’s suppose that instead of rolling dice, I’m actually flipping coins. This time around, my experiment involves flipping a fair coin repeatedly, and the outcome that I’m interested in is the number of heads that I observe. In this scenario, the success probability is now \\(\\theta = 1/2\\). Suppose I were to flip the coin \\(N=20\\) times. In this example, I’ve changed the success probability, but kept the size of the experiment the same. What does this do to our binomial distribution? Well, as Figure 9.4 shows, the main effect of this is to shift the whole distribution, as you’d expect. Okay, what if we flipped a coin \\(N=100\\) times? Well, in that case, we get Figure 9.5. The distribution stays roughly in the middle, but there’s a bit more variability in the possible outcomes.\n At this point, I should probably explain the name of the `dbinom()` function. Obviously, the “binom” part comes from the fact that we’re working with the binomial distribution, but the “d” prefix is probably a bit of a mystery. In this section I’ll give a partial explanation: specifically, I’ll explain why there is a prefix. As for why it’s a “d” specifically, you’ll have to wait until the next section. What’s going on here is that R actually provides four functions in relation to the binomial distribution. These four functions are `dbinom()` , `pbinom()` , `rbinom()` and `qbinom()` , and each one calculates a different quantity of interest. Not only that, R does the same thing for every probability distribution that it implements. No matter what distribution you’re talking about, there’s a `d` function, a `p` function, a `q` function and a `r` function. This is illustrated in Table 9.3, using the binomial distribution and the normal distribution as examples. \n\nWhat.it.does | Prefix | Normal.distribution | Binomial.distribution |\n | --- | --- | --- | --- |\nprobability (density) of | d | dnorm() | dbinom() |\ncumulative probability of | p | dnorm() | pbinom() |\ngenerate random number from | r | rnorm() | rbinom() |\nq qnorm() qbinom() | q | qnorm() | qbinom( |\n\n Let’s have a look at what all four functions do. Firstly, all four versions of the function require you to specify the `size` and `prob` arguments: no matter what you’re trying to get R to calculate, it needs to know what the parameters are. However, they differ in terms of what the other argument is, and what the output is. So let’s look at them one at a time. \n\n* The\n `d` form we’ve already seen: you specify a particular outcome `x` , and the output is the probability of obtaining exactly that outcome. (the “d” is short for density, but ignore that for now). * The\n `p` form calculates the cumulative probability. You specify a particular quantile `q` , and it tells you the probability of obtaining an outcome smaller than or equal to `q` . * The\n `q` form calculates the quantiles of the distribution. You specify a probability value `p` , and gives you the corresponding percentile. That is, the value of the variable for which there’s a probability `p` of obtaining an outcome lower than that value. * The\n `r` form is a random number generator: specifically, it generates `n` random outcomes from the distribution. This is a little abstract, so let’s look at some concrete examples. Again, we’ve already covered `dbinom()` so let’s focus on the other three versions. We’ll start with `pbinom()` , and we’ll go back to the skull-dice example. Again, I’m rolling 20 dice, and each die has a 1 in 6 chance of coming up skulls. Suppose, however, that I want to know the probability of rolling 4 or fewer skulls. If I wanted to, I could use the `dbinom()` function to calculate the exact probability of rolling 0 skulls, 1 skull, 2 skulls, 3 skulls and 4 skulls and then add these up, but there’s a faster way. Instead, I can calculate this using the `pbinom()` function. Here’s the command: \n\n```\npbinom( q= 4, size = 20, prob = 1/6)\n```\n\n `## [1] 0.7687492` \nIn other words, there is a 76.9% chance that I will roll 4 or fewer skulls. Or, to put it another way, R is telling us that a value of 4 is actually the 76.9th percentile of this binomial distribution.\n Next, let’s consider the `qbinom()` function. Let’s say I want to calculate the 75th percentile of the binomial distribution. If we’re sticking with our skulls example, I would use the following command to do this: \n\n```\nqbinom( p = 0.75, size = 20, prob = 1/6)\n```\n\n `## [1] 4` Hm. There’s something odd going on here. Let’s think this through. What the `qbinom()` function appears to be telling us is that the 75th percentile of the binomial distribution is 4, even though we saw from the `pbinom()` function that 4 is actually the 76.9th percentile. And it’s definitely the `pbinom()` function that is correct. I promise. The weirdness here comes from the fact that our binomial distribution doesn’t really have a 75th percentile. Not really. Why not? Well, there’s a 56.7% chance of rolling 3 or fewer skulls (you can type `pbinom(3, 20, 1/6)` to confirm this if you want), and a 76.9% chance of rolling 4 or fewer skulls. So there’s a sense in which the 75th percentile should lie “in between” 3 and 4 skulls. But that makes no sense at all! You can’t roll 20 dice and get 3.9 of them come up skulls. This issue can be handled in different ways: you could report an in between value (or interpolated value, to use the technical name) like 3.9, you could round down (to 3) or you could round up (to 4). The `qbinom()` function rounds upwards: if you ask for a percentile that doesn’t actually exist (like the 75th in this example), R finds the smallest value for which the the percentile rank is at least what you asked for. In this case, since the “true” 75th percentile (whatever that would mean) lies somewhere between 3 and 4 skulls, R rounds up and gives you an answer of 4. This subtlety is tedious, I admit, but thankfully it’s only an issue for discrete distributions like the binomial (see Section 2.2.5 for a discussion of continuous versus discrete). The other distributions that I’ll talk about (normal, \\(t\\), \\(\\chi^2\\) and \\(F\\)) are all continuous, and so R can always return an exact quantile whenever you ask for it. Finally, we have the random number generator. To use the `rbinom()` function, you specify how many times R should “simulate” the experiment using the `n` argument, and it will generate random outcomes from the binomial distribution. So, for instance, suppose I were to repeat my die rolling experiment 100 times. I could get R to simulate the results of these experiments by using the following command: \n\n```\nrbinom( n = 100, size = 20, prob = 1/6 )\n```\n\n```\n## [1] 3 2 9 2 4 4 3 7 1 0 1 5 3 5 4 3 3 2 3 1 4 3 2 3 2 0 4 2 4 4 6 1 3 4 7\n## [36] 5 4 4 3 4 2 3 1 3 3 4 6 6 2 5 9 1 5 2 3 4 1 3 4 3 4 4 4 4 2 1 3 2 6 3\n## [71] 2 4 6 4 4 2 4 1 5 4 2 4 8 3 3 2 3 5 5 3 1 2 3 4 6 2 2 2 1 2\n```\n\nAs you can see, these numbers are pretty much what you’d expect given the distribution shown in Figure 9.3. Most of the time I roll somewhere between 1 to 5 skulls. There are a lot of subtleties associated with random number generation using a computer,145 but for the purposes of this book we don’t need to worry too much about them.\n\n## 9.5 The normal distribution\n\n While the binomial distribution is conceptually the simplest distribution to understand, it’s not the most important one. That particular honour goes to the normal distribution, which is also referred to as “the bell curve” or a “Gaussian distribution”. A normal distribution is described using two parameters, the mean of the distribution \\(\\mu\\) and the standard deviation of the distribution \\(\\sigma\\). The notation that we sometimes use to say that a variable \\(X\\) is normally distributed is as follows: \\[ X \\sim \\mbox{Normal}(\\mu,\\sigma) \\] Of course, that’s just notation. It doesn’t tell us anything interesting about the normal distribution itself. As was the case with the binomial distribution, I have included the formula for the normal distribution in this book, because I think it’s important enough that everyone who learns statistics should at least look at it, but since this is an introductory text I don’t want to focus on it, so I’ve tucked it away in Table 9.2. Similarly, the R functions for the normal distribution are `dnorm()` , `pnorm()` , `qnorm()` and `rnorm()` . However, they behave in pretty much exactly the same way as the corresponding functions for the binomial distribution, so there’s not a lot that you need to know. The only thing that I should point out is that the argument names for the parameters are `mean` and `sd` . In pretty much every other respect, there’s nothing else to add. \nInstead of focusing on the maths, let’s try to get a sense for what it means for a variable to be normally distributed. To that end, have a look at Figure 9.6, which plots a normal distribution with mean \\(\\mu = 0\\) and standard deviation \\(\\sigma = 1\\). You can see where the name “bell curve” comes from: it looks a bit like a bell. Notice that, unlike the plots that I drew to illustrate the binomial distribution, the picture of the normal distribution in Figure 9.6 shows a smooth curve instead of “histogram-like” bars. This isn’t an arbitrary choice: the normal distribution is continuous, whereas the binomial is discrete. For instance, in the die rolling example from the last section, it was possible to get 3 skulls or 4 skulls, but impossible to get 3.9 skulls. The figures that I drew in the previous section reflected this fact: in Figure 9.3, for instance, there’s a bar located at \\(X=3\\) and another one at \\(X=4\\), but there’s nothing in between. Continuous quantities don’t have this constraint. For instance, suppose we’re talking about the weather. The temperature on a pleasant Spring day could be 23 degrees, 24 degrees, 23.9 degrees, or anything in between since temperature is a continuous variable, and so a normal distribution might be quite appropriate for describing Spring temperatures.146\n\nWith this in mind, let’s see if we can’t get an intuition for how the normal distribution works. Firstly, let’s have a look at what happens when we play around with the parameters of the distribution. To that end, Figure 9.7 plots normal distributions that have different means, but have the same standard deviation. As you might expect, all of these distributions have the same “width”. The only difference between them is that they’ve been shifted to the left or to the right. In every other respect they’re identical. In contrast, if we increase the standard deviation while keeping the mean constant, the peak of the distribution stays in the same place, but the distribution gets wider, as you can see in Figure 9.8. Notice, though, that when we widen the distribution, the height of the peak shrinks. This has to happen: in the same way that the heights of the bars that we used to draw a discrete binomial distribution have to sum to 1, the total area under the curve for the normal distribution must equal 1.\n\nBefore moving on, I want to point out one important characteristic of the normal distribution. Irrespective of what the actual mean and standard deviation are, 68.3% of the area falls within 1 standard deviation of the mean. Similarly, 95.4% of the distribution falls within 2 standard deviations of the mean, and 99.7% of the distribution is within 3 standard deviations. This idea is illustrated in Figures 9.9 and 9.10.\n\n### 9.5.1 Probability density\n\nThere’s something I’ve been trying to hide throughout my discussion of the normal distribution, something that some introductory textbooks omit completely. They might be right to do so: this “thing” that I’m hiding is weird and counterintuitive even by the admittedly distorted standards that apply in statistics. Fortunately, it’s not something that you need to understand at a deep level in order to do basic statistics: rather, it’s something that starts to become important later on when you move beyond the basics. So, if it doesn’t make complete sense, don’t worry: try to make sure that you follow the gist of it.\n Throughout my discussion of the normal distribution, there’s been one or two things that don’t quite make sense. Perhaps you noticed that the \\(y\\)-axis in these figures is labelled “Probability Density” rather than density. Maybe you noticed that I used \\(p(X)\\) instead of \\(P(X)\\) when giving the formula for the normal distribution. Maybe you’re wondering why R uses the “d” prefix for functions like `dnorm()` . And maybe, just maybe, you’ve been playing around with the `dnorm()` function, and you accidentally typed in a command like this: \n\n```\ndnorm( x = 1, mean = 1, sd = 0.1 )\n```\n\n `## [1] 3.989423` And if you’ve done the last part, you’re probably very confused. I’ve asked R to calculate the probability that `x = 1` , for a normally distributed variable with `mean = 1` and standard deviation `sd = 0.1` ; and it tells me that the probability is 3.99. But, as we discussed earlier, probabilities can’t be larger than 1. So either I’ve made a mistake, or that’s not a probability. \nAs it turns out, the second answer is correct. What we’ve calculated here isn’t actually a probability: it’s something else. To understand what that something is, you have to spend a little time thinking about what it really means to say that \\(X\\) is a continuous variable. Let’s say we’re talking about the temperature outside. The thermometer tells me it’s 23 degrees, but I know that’s not really true. It’s not exactly 23 degrees. Maybe it’s 23.1 degrees, I think to myself. But I know that that’s not really true either, because it might actually be 23.09 degrees. But, I know that… well, you get the idea. The tricky thing with genuinely continuous quantities is that you never really know exactly what they are.\n\nNow think about what this implies when we talk about probabilities. Suppose that tomorrow’s maximum temperature is sampled from a normal distribution with mean 23 and standard deviation 1. What’s the probability that the temperature will be exactly 23 degrees? The answer is “zero”, or possibly, “a number so close to zero that it might as well be zero”. Why is this? It’s like trying to throw a dart at an infinitely small dart board: no matter how good your aim, you’ll never hit it. In real life you’ll never get a value of exactly 23. It’ll always be something like 23.1 or 22.99998 or something. In other words, it’s completely meaningless to talk about the probability that the temperature is exactly 23 degrees. However, in everyday language, if I told you that it was 23 degrees outside and it turned out to be 22.9998 degrees, you probably wouldn’t call me a liar. Because in everyday language, “23 degrees” usually means something like “somewhere between 22.5 and 23.5 degrees”. And while it doesn’t feel very meaningful to ask about the probability that the temperature is exactly 23 degrees, it does seem sensible to ask about the probability that the temperature lies between 22.5 and 23.5, or between 20 and 30, or any other range of temperatures.\n\nThe point of this discussion is to make clear that, when we’re talking about continuous distributions, it’s not meaningful to talk about the probability of a specific value. However, what we can talk about is the probability that the value lies within a particular range of values. To find out the probability associated with a particular range, what you need to do is calculate the “area under the curve”. We’ve seen this concept already: in Figure 9.9, the shaded areas shown depict genuine probabilities (e.g., in the left hand panel of Figure 9.9 it shows the probability of observing a value that falls within 1 standard deviation of the mean).\n Okay, so that explains part of the story. I’ve explained a little bit about how continuous probability distributions should be interpreted (i.e., area under the curve is the key thing), but I haven’t actually explained what the `dnorm()` function actually calculates. Equivalently, what does the formula for \\(p(x)\\) that I described earlier actually mean? Obviously, \\(p(x)\\) doesn’t describe a probability, but what is it? The name for this quantity \\(p(x)\\) is a probability density, and in terms of the plots we’ve been drawing, it corresponds to the height of the curve. The densities themselves aren’t meaningful in and of themselves: but they’re “rigged” to ensure that the area under the curve is always interpretable as genuine probabilities. To be honest, that’s about as much as you really need to know for now.147 \n\n## 9.6 Other useful distributions\n\nThe normal distribution is the distribution that statistics makes most use of (for reasons to be discussed shortly), and the binomial distribution is a very useful one for lots of purposes. But the world of statistics is filled with probability distributions, some of which we’ll run into in passing. In particular, the three that will appear in this book are the \\(t\\) distribution, the \\(\\chi^2\\) distribution and the \\(F\\) distribution. I won’t give formulas for any of these, or talk about them in too much detail, but I will show you some pictures.\n\n* The \\(t\\) distribution is a continuous distribution that looks very similar to a normal distribution, but has heavier tails: see Figure 9.11. This distribution tends to arise in situations where you think that the data actually follow a normal distribution, but you don’t know the mean or standard deviation. As you might expect, the relevant R functions are\n `dt()` , `pt()` , `qt()` and `rt()` , and we’ll run into this distribution again in Chapter 13. \n\n* The \\(\\chi^2\\) distribution is another distribution that turns up in lots of different places. The situation in which we’ll see it is when doing categorical data analysis (Chapter 12), but it’s one of those things that actually pops up all over the place. When you dig into the maths (and who doesn’t love doing that?), it turns out that the main reason why the \\(\\chi^2\\) distribution turns up all over the place is that, if you have a bunch of variables that are normally distributed, square their values and then add them up (a procedure referred to as taking a “sum of squares”), this sum has a \\(\\chi^2\\) distribution. You’d be amazed how often this fact turns out to be useful. Anyway, here’s what a \\(\\chi^2\\) distribution looks like: Figure 9.12. Once again, the R commands for this one are pretty predictable:\n `dchisq()` , `pchisq()` , `qchisq()` , `rchisq()` . \n\n* The \\(F\\) distribution looks a bit like a \\(\\chi^2\\) distribution, and it arises whenever you need to compare two \\(\\chi^2\\) distributions to one another. Admittedly, this doesn’t exactly sound like something that any sane person would want to do, but it turns out to be very important in real world data analysis. Remember when I said that \\(\\chi^2\\) turns out to be the key distribution when we’re taking a “sum of squares”? Well, what that means is if you want to compare two different “sums of squares”, you’re probably talking about something that has an \\(F\\) distribution. Of course, as yet I still haven’t given you an example of anything that involves a sum of squares, but I will… in Chapter 14. And that’s where we’ll run into the \\(F\\) distribution. Oh, and here’s a picture: Figure 9.13. And of course we can get R to do things with \\(F\\) distributions just by using the commands\n `df()` , `pf()` , `qf()` and `rf()` . Because these distributions are all tightly related to the normal distribution and to each other, and because they are will turn out to be the important distributions when doing inferential statistics later in this book, I think it’s useful to do a little demonstration using R, just to “convince ourselves” that these distributions really are related to each other in the way that they’re supposed to be. First, we’ll use the `rnorm()` function to generate 1000 normally-distributed observations: \n\n```\nnormal.a <- rnorm( n=1000, mean=0, sd=1 ) \nprint(head(normal.a))\n```\n\n```\n## [1] 0.002520116 -1.759249354 -0.055968257 0.879791922 1.166488549\n## [6] 0.789723465\n```\n\n So the `normal.a` variable contains 1000 numbers that are normally distributed, and have mean 0 and standard deviation 1, and the actual print out of these numbers goes on for rather a long time. Note that, because the default parameters of the `rnorm()` function are `mean=0` and `sd=1` , I could have shortened the command to `rnorm( n=1000 )` . In any case, what we can do is use the `hist()` function to draw a histogram of the data, like so: `hist( normal.a )` \nIf you do this, you should see something similar to Figure 9.14. Your plot won’t look quite as pretty as the one in the figure, of course, because I’ve played around with all the formatting (see Chapter 6), and I’ve also plotted the true distribution of the data as a solid black line (i.e., a normal distribution with mean 0 and standard deviation 1) so that you can compare the data that we just generated to the true distribution.\n In the previous example all I did was generate lots of normally distributed observations using `rnorm()` and then compared those to the true probability distribution in the figure (using `dnorm()` to generate the black line in the figure, but I didn’t show the commmands for that). Now let’s try something trickier. We’ll try to generate some observations that follow a chi-square distribution with 3 degrees of freedom, but instead of using `rchisq()` , we’ll start with variables that are normally distributed, and see if we can exploit the known relationships between normal and chi-square distributions to do the work. As I mentioned earlier, a chi-square distribution with \\(k\\) degrees of freedom is what you get when you take \\(k\\) normally-distributed variables (with mean 0 and standard deviation 1), square them, and add them up. Since we want a chi-square distribution with 3 degrees of freedom, we’ll need to supplement our `normal.a` data with two more sets of normally-distributed observations, imaginatively named `normal.b` and `normal.c` : \n\n```\nnormal.b <- rnorm( n=1000 ) # another set of normally distributed data\nnormal.c <- rnorm( n=1000 ) # and another!\n```\n\nNow that we’ve done that, the theory says we should square these and add them together, like this\n\n```\nchi.sq.3 <- (normal.a)^2 + (normal.b)^2 + (normal.c)^2\n```\n\n and the resulting `chi.sq.3` variable should contain 1000 observations that follow a chi-square distribution with 3 degrees of freedom. You can use the `hist()` function to have a look at these observations yourself, using a command like this, `hist( chi.sq.3 )` and you should obtain a result that looks pretty similar to the chi-square plot in Figure 9.14. Once again, the plot that I’ve drawn is a little fancier: in addition to the histogram of `chi.sq.3` , I’ve also plotted a chi-square distribution with 3 degrees of freedom. It’s pretty clear that – even though I used `rnorm()` to do all the work rather than `rchisq()` – the observations stored in the `chi.sq.3` variable really do follow a chi-square distribution. Admittedly, this probably doesn’t seem all that interesting right now, but later on when we start encountering the chi-square distribution in Chapter 12, it will be useful to understand the fact that these distributions are related to one another. \nWe can extend this demonstration to the \\(t\\) distribution and the \\(F\\) distribution. Earlier, I implied that the \\(t\\) distribution is related to the normal distribution when the standard deviation is unknown. That’s certainly true, and that’s the what we’ll see later on in Chapter 13, but there’s a somewhat more precise relationship between the normal, chi-square and \\(t\\) distributions. Suppose we “scale” our chi-square data by dividing it by the degrees of freedom, like so\n\n```\nscaled.chi.sq.3 <- chi.sq.3 / 3\n```\n\n We then take a set of normally distributed variables and divide them by (the square root of) our scaled chi-square variable which had \\(df=3\\), and the result is a \\(t\\) distribution with 3 degrees of freedom. If we plot the histogram of `t.3` , we end up with something that looks very similar to the t distribution in Figure 9.14. \n\n```\nnormal.d <- rnorm( n=1000 ) # yet another set of normally distributed data\nt.3 <- normal.d / sqrt( scaled.chi.sq.3 ) # divide by square root of scaled chi-square to get t\nhist (t.3)\n```\n\n Similarly, we can obtain an \\(F\\) distribution by taking the ratio between two scaled chi-square distributions. Suppose, for instance, we wanted to generate data from an \\(F\\) distribution with 3 and 20 degrees of freedom. We could do this using `df()` , but we could also do the same thing by generating two chi-square variables, one with 3 degrees of freedom, and the other with 20 degrees of freedom. As the example with `chi.sq.3` illustrates, we can actually do this using `rnorm()` if we really want to, but this time I’ll take a short cut: \n\n```\nchi.sq.20 <- rchisq( 1000, 20) # generate chi square data with df = 20...\nscaled.chi.sq.20 <- chi.sq.20 / 20 # scale the chi square variable...\nF.3.20 <- scaled.chi.sq.3 / scaled.chi.sq.20 # take the ratio of the two chi squares...\nhist( F.3.20 ) # ... and draw a picture\n```\n\n The resulting `F.3.20` variable does in fact store variables that follow an \\(F\\) distribution with 3 and 20 degrees of freedom. This is illustrated in Figure 9.14, which plots the histgram of the observations stored in `F.3.20` against the true \\(F\\) distribution with \\(df_1 = 3\\) and \\(df_2 = 20\\). Again, they match. \nOkay, time to wrap this section up. We’ve seen three new distributions: \\(\\chi^2\\), \\(t\\) and \\(F\\). They’re all continuous distributions, and they’re all closely related to the normal distribution. I’ve talked a little bit about the precise nature of this relationship, and shown you some R commands that illustrate this relationship. The key thing for our purposes, however, is not that you have a deep understanding of all these different distributions, nor that you remember the precise relationships between them. The main thing is that you grasp the basic idea that these distributions are all deeply related to one another, and to the normal distribution. Later on in this book, we’re going to run into data that are normally distributed, or at least assumed to be normally distributed. What I want you to understand right now is that, if you make the assumption that your data are normally distributed, you shouldn’t be surprised to see \\(\\chi^2\\), \\(t\\) and \\(F\\) distributions popping up all over the place when you start trying to do your data analysis.\n\n## 9.7 Summary\n\nIn this chapter we’ve talked about probability. We’ve talked what probability means, and why statisticians can’t agree on what it means. We talked about the rules that probabilities have to obey. And we introduced the idea of a probability distribution, and spent a good chunk of the chapter talking about some of the more important probability distributions that statisticians work with. The section by section breakdown looks like this:\n\n* Probability theory versus statistics (Section 9.1)\n * Frequentist versus Bayesian views of probability (Section 9.2)\n * Basics of probability theory (Section 9.3)\n * Binomial distribution (Section 9.4), normal distribution (Section 9.5), and others (Section 9.6)\n\nAs you’d expect, my coverage is by no means exhaustive. Probability theory is a large branch of mathematics in its own right, entirely separate from its application to statistics and data analysis. As such, there are thousands of books written on the subject and universities generally offer multiple classes devoted entirely to probability theory. Even the “simpler” task of documenting standard probability distributions is a big topic. I’ve described five standard probability distributions in this chapter, but sitting on my bookshelf I have a 45-chapter book called “Statistical Distributions” Evans, Hastings, and Peacock (2011) that lists a lot more than that. Fortunately for you, very little of this is necessary. You’re unlikely to need to know dozens of statistical distributions when you go out and do real world data analysis, and you definitely won’t need them for this book, but it never hurts to know that there’s other possibilities out there.\n\nPicking up on that last point, there’s a sense in which this whole chapter is something of a digression. Many undergraduate psychology classes on statistics skim over this content very quickly (I know mine did), and even the more advanced classes will often “forget” to revisit the basic foundations of the field. Most academic psychologists would not know the difference between probability and density, and until recently very few would have been aware of the difference between Bayesian and frequentist probability. However, I think it’s important to understand these things before moving onto the applications. For example, there are a lot of rules about what you’re “allowed” to say when doing statistical inference, and many of these can seem arbitrary and weird. However, they start to make sense if you understand that there is this Bayesian/frequentist distinction. Similarly, in Chapter 13 we’re going to talk about something called the \\(t\\)-test, and if you really want to have a grasp of the mechanics of the \\(t\\)-test it really helps to have a sense of what a \\(t\\)-distribution actually looks like. You get the idea, I hope.\n\n. 1922b. “On the Mathematical Foundation of Theoretical Statistics.” Philosophical Transactions of the Royal Society A 222: 309–68.\n\nThis doesn’t mean that frequentists can’t make hypothetical statements, of course; it’s just that if you want to make a statement about probability, then it must be possible to redescribe that statement in terms of a sequence of potentially observable events, and the relative frequencies of different outcomes that appear within that sequence.↩\n * \n\nNote that the term “success” is pretty arbitrary, and doesn’t actually imply that the outcome is something to be desired. If \\(\\theta\\) referred to the probability that any one passenger gets injured in a bus crash, I’d still call it the success probability, but that doesn’t mean I want people to get hurt in bus crashes!↩\n * \n\nSince computers are deterministic machines, they can’t actually produce truly random behaviour. Instead, what they do is take advantage of various mathematical functions that share a lot of similarities with true randomness. What this means is that any random numbers generated on a computer are pseudorandom, and the quality of those numbers depends on the specific method used. By default R uses the “Mersenne twister” method. In any case, you can find out more by typing\n `?Random` , but as usual the R help files are fairly dense.↩ * \n\nIn practice, the normal distribution is so handy that people tend to use it even when the variable isn’t actually continuous. As long as there are enough categories (e.g., Likert scale responses to a questionnaire), it’s pretty standard practice to use the normal distribution as an approximation. This works out much better in practice than you’d think.↩\n * \n\nFor those readers who know a little calculus, I’ll give a slightly more precise explanation. In the same way that probabilities are non-negative numbers that must sum to 1, probability densities are non-negative numbers that must integrate to 1 (where the integral is taken across all possible values of \\(X\\)). To calculate the probability that \\(X\\) falls between \\(a\\) and \\(b\\) we calculate the definite integral of the density function over the corresponding range, \\(\\int_a^b p(x) \\ dx\\). If you don’t remember or never learned calculus, don’t worry about this. It’s not needed for this book.↩\n\n# Chapter 10 Estimating unknown quantities from a sample\n\nAt the start of the last chapter I highlighted the critical distinction between descriptive statistics and inferential statistics. As discussed in Chapter 5, the role of descriptive statistics is to concisely summarise what we do know. In contrast, the purpose of inferential statistics is to “learn what we do not know from what we do”. Now that we have a foundation in probability theory, we are in a good position to think about the problem of statistical inference. What kinds of things would we like to learn about? And how do we learn them? These are the questions that lie at the heart of inferential statistics, and they are traditionally divided into two “big ideas”: estimation and hypothesis testing. The goal in this chapter is to introduce the first of these big ideas, estimation theory, but I’m going to witter on about sampling theory first because estimation theory doesn’t make sense until you understand sampling. As a consequence, this chapter divides naturally into two parts Sections 10.1 through 10.3 are focused on sampling theory, and Sections 10.4 and 10.5 make use of sampling theory to discuss how statisticians think about estimation.\n\n## 10.1 Samples, populations and sampling\n\nIn the prelude to Part I discussed the riddle of induction, and highlighted the fact that all learning requires you to make assumptions. Accepting that this is true, our first task to come up with some fairly general assumptions about data that make sense. This is where sampling theory comes in. If probability theory is the foundations upon which all statistical theory builds, sampling theory is the frame around which you can build the rest of the house. Sampling theory plays a huge role in specifying the assumptions upon which your statistical inferences rely. And in order to talk about “making inferences” the way statisticians think about it, we need to be a bit more explicit about what it is that we’re drawing inferences from (the sample) and what it is that we’re drawing inferences about (the population).\n\nIn almost every situation of interest, what we have available to us as researchers is a sample of data. We might have run experiment with some number of participants; a polling company might have phoned some number of people to ask questions about voting intentions; etc. Regardless: the data set available to us is finite, and incomplete. We can’t possibly get every person in the world to do our experiment; a polling company doesn’t have the time or the money to ring up every voter in the country etc. In our earlier discussion of descriptive statistics (Chapter 5, this sample was the only thing we were interested in. Our only goal was to find ways of describing, summarising and graphing that sample. This is about to change.\n\n### 10.1.1 Defining a population\n\nA sample is a concrete thing. You can open up a data file, and there’s the data from your sample. A population, on the other hand, is a more abstract idea. It refers to the set of all possible people, or all possible observations, that you want to draw conclusions about, and is generally much bigger than the sample. In an ideal world, the researcher would begin the study with a clear idea of what the population of interest is, since the process of designing a study and testing hypotheses about the data that it produces does depend on the population about which you want to make statements. However, that doesn’t always happen in practice: usually the researcher has a fairly vague idea of what the population is and designs the study as best he/she can on that basis.\n\nSometimes it’s easy to state the population of interest. For instance, in the “polling company” example that opened the chapter, the population consisted of all voters enrolled at the a time of the study – millions of people. The sample was a set of 1000 people who all belong to that population. In most situations the situation is much less simple. In a typical a psychological experiment, determining the population of interest is a bit more complicated. Suppose I run an experiment using 100 undergraduate students as my participants. My goal, as a cognitive scientist, is to try to learn something about how the mind works. So, which of the following would count as “the population”:\n\n* All of the undergraduate psychology students at the University of Adelaide?\n * Undergraduate psychology students in general, anywhere in the world?\n * Australians currently living?\n * Australians of similar ages to my sample?\n * Anyone currently alive?\n * Any human being, past, present or future?\n * Any biological organism with a sufficient degree of intelligence operating in a terrestrial environment?\n * Any intelligent being?\n\nEach of these defines a real group of mind-possessing entities, all of which might be of interest to me as a cognitive scientist, and it’s not at all clear which one ought to be the true population of interest. As another example, consider the Wellesley-Croker game that we discussed in the prelude. The sample here is a specific sequence of 12 wins and 0 losses for Wellesley. What is the population?\n\n* All outcomes until Wellesley and Croker arrived at their destination?\n * All outcomes if Wellesley and Croker had played the game for the rest of their lives?\n * All outcomes if Wellseley and Croker lived forever and played the game until the world ran out of hills?\n * All outcomes if we created an infinite set of parallel universes and the Wellesely/Croker pair made guesses about the same 12 hills in each universe?\n\nAgain, it’s not obvious what the population is.\n\n### 10.1.2 Simple random samples\n\nIrrespective of how I define the population, the critical point is that the sample is a subset of the population, and our goal is to use our knowledge of the sample to draw inferences about the properties of the population. The relationship between the two depends on the procedure by which the sample was selected. This procedure is referred to as a sampling method, and it is important to understand why it matters.\n\nTo keep things simple, let’s imagine that we have a bag containing 10 chips. Each chip has a unique letter printed on it, so we can distinguish between the 10 chips. The chips come in two colours, black and white. This set of chips is the population of interest, and it is depicted graphically on the left of Figure 10.1. As you can see from looking at the picture, there are 4 black chips and 6 white chips, but of course in real life we wouldn’t know that unless we looked in the bag. Now imagine you run the following “experiment”: you shake up the bag, close your eyes, and pull out 4 chips without putting any of them back into the bag. First out comes the \\(a\\) chip (black), then the \\(c\\) chip (white), then \\(j\\) (white) and then finally \\(b\\) (black). If you wanted, you could then put all the chips back in the bag and repeat the experiment, as depicted on the right hand side of Figure 10.1. Each time you get different results, but the procedure is identical in each case. The fact that the same procedure can lead to different results each time, we refer to it as a random process.148 However, because we shook the bag before pulling any chips out, it seems reasonable to think that every chip has the same chance of being selected. A procedure in which every member of the population has the same chance of being selected is called a simple random sample. The fact that we did not put the chips back in the bag after pulling them out means that you can’t observe the same thing twice, and in such cases the observations are said to have been sampled without replacement.\n\nTo help make sure you understand the importance of the sampling procedure, consider an alternative way in which the experiment could have been run. Suppose that my 5-year old son had opened the bag, and decided to pull out four black chips without putting any of them back in the bag. This biased sampling scheme is depicted in Figure 10.2. Now consider the evidentiary value of seeing 4 black chips and 0 white chips. Clearly, it depends on the sampling scheme, does it not? If you know that the sampling scheme is biased to select only black chips, then a sample that consists of only black chips doesn’t tell you very much about the population! For this reason, statisticians really like it when a data set can be considered a simple random sample, because it makes the data analysis much easier.\n\nA third procedure is worth mentioning. This time around we close our eyes, shake the bag, and pull out a chip. This time, however, we record the observation and then put the chip back in the bag. Again we close our eyes, shake the bag, and pull out a chip. We then repeat this procedure until we have 4 chips. Data sets generated in this way are still simple random samples, but because we put the chips back in the bag immediately after drawing them it is referred to as a sample with replacement. The difference between this situation and the first one is that it is possible to observe the same population member multiple times, as illustrated in Figure 10.3.\n\nIn my experience, most psychology experiments tend to be sampling without replacement, because the same person is not allowed to participate in the experiment twice. However, most statistical theory is based on the assumption that the data arise from a simple random sample with replacement. In real life, this very rarely matters. If the population of interest is large (e.g., has more than 10 entities!) the difference between sampling with- and without- replacement is too small to be concerned with. The difference between simple random samples and biased samples, on the other hand, is not such an easy thing to dismiss.\n\n### 10.1.3 Most samples are not simple random samples\n\nAs you can see from looking at the list of possible populations that I showed above, it is almost impossible to obtain a simple random sample from most populations of interest. When I run experiments, I’d consider it a minor miracle if my participants turned out to be a random sampling of the undergraduate psychology students at Adelaide university, even though this is by far the narrowest population that I might want to generalise to. A thorough discussion of other types of sampling schemes is beyond the scope of this book, but to give you a sense of what’s out there I’ll list a few of the more important ones:\n\n* Stratified sampling. Suppose your population is (or can be) divided into several different subpopulations, or strata. Perhaps you’re running a study at several different sites, for example. Instead of trying to sample randomly from the population as a whole, you instead try to collect a separate random sample from each of the strata. Stratified sampling is sometimes easier to do than simple random sampling, especially when the population is already divided into the distinct strata. It can also be more efficient that simple random sampling, especially when some of the subpopulations are rare. For instance, when studying schizophrenia it would be much better to divide the population into two149 strata (schizophrenic and not-schizophrenic), and then sample an equal number of people from each group. If you selected people randomly, you would get so few schizophrenic people in the sample that your study would be useless. This specific kind of of stratified sampling is referred to as oversampling because it makes a deliberate attempt to over-represent rare groups.\n * Snowball sampling is a technique that is especially useful when sampling from a “hidden” or hard to access population, and is especially common in social sciences. For instance, suppose the researchers want to conduct an opinion poll among transgender people. The research team might only have contact details for a few trans folks, so the survey starts by asking them to participate (stage 1). At the end of the survey, the participants are asked to provide contact details for other people who might want to participate. In stage 2, those new contacts are surveyed. The process continues until the researchers have sufficient data. The big advantage to snowball sampling is that it gets you data in situations that might otherwise be impossible to get any. On the statistical side, the main disadvantage is that the sample is highly non-random, and non-random in ways that are difficult to address. On the real life side, the disadvantage is that the procedure can be unethical if not handled well, because hidden populations are often hidden for a reason. I chose transgender people as an example here to highlight this: if you weren’t careful you might end up outing people who don’t want to be outed (very, very bad form), and even if you don’t make that mistake it can still be intrusive to use people’s social networks to study them. It’s certainly very hard to get people’s informed consent before contacting them, yet in many cases the simple act of contacting them and saying “hey we want to study you” can be hurtful. Social networks are complex things, and just because you can use them to get data doesn’t always mean you should.\n * Convenience sampling is more or less what it sounds like. The samples are chosen in a way that is convenient to the researcher, and not selected at random from the population of interest. Snowball sampling is one type of convenience sampling, but there are many others. A common example in psychology are studies that rely on undergraduate psychology students. These samples are generally non-random in two respects: firstly, reliance on undergraduate psychology students automatically means that your data are restricted to a single subpopulation. Secondly, the students usually get to pick which studies they participate in, so the sample is a self selected subset of psychology students not a randomly selected subset. In real life, most studies are convenience samples of one form or another. This is sometimes a severe limitation, but not always.\n\n### 10.1.4 How much does it matter if you don’t have a simple random sample?\n\nOkay, so real world data collection tends not to involve nice simple random samples. Does that matter? A little thought should make it clear to you that it can matter if your data are not a simple random sample: just think about the difference between Figures 10.1 and 10.2. However, it’s not quite as bad as it sounds. Some types of biased samples are entirely unproblematic. For instance, when using a stratified sampling technique you actually know what the bias is because you created it deliberately, often to increase the effectiveness of your study, and there are statistical techniques that you can use to adjust for the biases you’ve introduced (not covered in this book!). So in those situations it’s not a problem.\n\nMore generally though, it’s important to remember that random sampling is a means to an end, not the end in itself. Let’s assume you’ve relied on a convenience sample, and as such you can assume it’s biased. A bias in your sampling method is only a problem if it causes you to draw the wrong conclusions. When viewed from that perspective, I’d argue that we don’t need the sample to be randomly generated in every respect: we only need it to be random with respect to the psychologically-relevant phenomenon of interest. Suppose I’m doing a study looking at working memory capacity. In study 1, I actually have the ability to sample randomly from all human beings currently alive, with one exception: I can only sample people born on a Monday. In study 2, I am able to sample randomly from the Australian population. I want to generalise my results to the population of all living humans. Which study is better? The answer, obviously, is study 1. Why? Because we have no reason to think that being “born on a Monday” has any interesting relationship to working memory capacity. In contrast, I can think of several reasons why “being Australian” might matter. Australia is a wealthy, industrialised country with a very well-developed education system. People growing up in that system will have had life experiences much more similar to the experiences of the people who designed the tests for working memory capacity. This shared experience might easily translate into similar beliefs about how to “take a test”, a shared assumption about how psychological experimentation works, and so on. These things might actually matter. For instance, “test taking” style might have taught the Australian participants how to direct their attention exclusively on fairly abstract test materials relative to people that haven’t grown up in a similar environment; leading to a misleading picture of what working memory capacity is.\n\nThere are two points hidden in this discussion. Firstly, when designing your own studies, it’s important to think about what population you care about, and try hard to sample in a way that is appropriate to that population. In practice, you’re usually forced to put up with a “sample of convenience” (e.g., psychology lecturers sample psychology students because that’s the least expensive way to collect data, and our coffers aren’t exactly overflowing with gold), but if so you should at least spend some time thinking about what the dangers of this practice might be.\n\nSecondly, if you’re going to criticise someone else’s study because they’ve used a sample of convenience rather than laboriously sampling randomly from the entire human population, at least have the courtesy to offer a specific theory as to how this might have distorted the results. Remember, everyone in science is aware of this issue, and does what they can to alleviate it. Merely pointing out that “the study only included people from group BLAH” is entirely unhelpful, and borders on being insulting to the researchers, who are of course aware of the issue. They just don’t happen to be in possession of the infinite supply of time and money required to construct the perfect sample. In short, if you want to offer a responsible critique of the sampling process, then be helpful. Rehashing the blindingly obvious truisms that I’ve been rambling on about in this section isn’t helpful.\n\n### 10.1.5 Population parameters and sample statistics\n\nOkay. Setting aside the thorny methodological issues associated with obtaining a random sample and my rather unfortunate tendency to rant about lazy methodological criticism, let’s consider a slightly different issue. Up to this point we have been talking about populations the way a scientist might. To a psychologist, a population might be a group of people. To an ecologist, a population might be a group of bears. In most cases the populations that scientists care about are concrete things that actually exist in the real world. Statisticians, however, are a funny lot. On the one hand, they are interested in real world data and real science in the same way that scientists are. On the other hand, they also operate in the realm of pure abstraction in the way that mathematicians do. As a consequence, statistical theory tends to be a bit abstract in how a population is defined. In much the same way that psychological researchers operationalise our abstract theoretical ideas in terms of concrete measurements (Section 2.1, statisticians operationalise the concept of a “population” in terms of mathematical objects that they know how to work with. You’ve already come across these objects in Chapter 9: they’re called probability distributions.\n\nThe idea is quite simple. Let’s say we’re talking about IQ scores. To a psychologist, the population of interest is a group of actual humans who have IQ scores. A statistician “simplifies” this by operationally defining the population as the probability distribution depicted in Figure ??. IQ tests are designed so that the average IQ is 100, the standard deviation of IQ scores is 15, and the distribution of IQ scores is normal. These values are referred to as the population parameters because they are characteristics of the entire population. That is, we say that the population mean \\(\\mu\\) is 100, and the population standard deviation \\(\\sigma\\) is 15.\n\n```\n## [1] \"n= 100 mean= 100.553253383599 sd= 15.0078396909989\"\n```\n\n```\n## [1] \"n= 10000 mean= 99.9199456720285 sd= 14.9487011509218\"\n```\n\nNow suppose I run an experiment. I select 100 people at random and administer an IQ test, giving me a simple random sample from the population. My sample would consist of a collection of numbers like this:\n\n```\n106 101 98 80 74 ... 107 72 100\n```\n\nEach of these IQ scores is sampled from a normal distribution with mean 100 and standard deviation 15. So if I plot a histogram of the sample, I get something like the one shown in Figure 10.4b. As you can see, the histogram is roughly the right shape, but it’s a very crude approximation to the true population distribution shown in Figure 10.4a. When I calculate the mean of my sample, I get a number that is fairly close to the population mean 100 but not identical. In this case, it turns out that the people in my sample have a mean IQ of 98.5, and the standard deviation of their IQ scores is 15.9. These sample statistics are properties of my data set, and although they are fairly similar to the true population values, they are not the same. In general, sample statistics are the things you can calculate from your data set, and the population parameters are the things you want to learn about. Later on in this chapter I’ll talk about how you can estimate population parameters using your sample statistics (Section 10.4 and how to work out how confident you are in your estimates (Section 10.5 but before we get to that there’s a few more ideas in sampling theory that you need to know about.\n\n## 10.2 The law of large numbers\n\nIn the previous section I showed you the results of one fictitious IQ experiment with a sample size of \\(N=100\\). The results were somewhat encouraging: the true population mean is 100, and the sample mean of 98.5 is a pretty reasonable approximation to it. In many scientific studies that level of precision is perfectly acceptable, but in other situations you need to be a lot more precise. If we want our sample statistics to be much closer to the population parameters, what can we do about it?\n The obvious answer is to collect more data. Suppose that we ran a much larger experiment, this time measuring the IQs of 10,000 people. We can simulate the results of this experiment using R. In Section 9.5 I introduced the `rnorm()` function, which generates random numbers sampled from a normal distribution. For an experiment with a sample size of `n = 10000` , and a population with `mean = 100` and `sd = 15` , R produces our fake IQ data using these commands: \n\n```\nIQ <- rnorm(n = 10000, mean = 100, sd = 15) # generate IQ scores\nIQ <- round(IQ) # IQs are whole numbers!\nprint(head(IQ))\n```\n\n```\n## [1] 113 104 64 89 101 100\n```\n\n I can compute the mean IQ using the command `mean(IQ)` and the standard deviation using the command `sd(IQ)` , and I can draw a histgram using `hist()` . The histogram of this much larger sample is shown in Figure 10.4c. Even a moment’s inspections makes clear that the larger sample is a much better approximation to the true population distribution than the smaller one. This is reflected in the sample statistics: the mean IQ for the larger sample turns out to be 99.9, and the standard deviation is 15.1. These values are now very close to the true population. \nI feel a bit silly saying this, but the thing I want you to take away from this is that large samples generally give you better information. I feel silly saying it because it’s so bloody obvious that it shouldn’t need to be said. In fact, it’s such an obvious point that when – one of the founders of probability theory – formalised this idea back in 1713, he was kind of a jerk about it. Here’s how he described the fact that we all share this intuition:\n For even the most stupid of men, by some instinct of nature, by himself and without any instruction (which is a remarkable thing), is convinced that the more observations have been made, the less danger there is of wandering from one’s goal Stigler (1986)\n\nOkay, so the passage comes across as a bit condescending (not to mention sexist), but his main point is correct: it really does feel obvious that more data will give you better answers. The question is, why is this so? Not surprisingly, this intuition that we all share turns out to be correct, and statisticians refer to it as the law of large numbers. The law of large numbers is a mathematical law that applies to many different sample statistics, but the simplest way to think about it is as a law about averages. The sample mean is the most obvious example of a statistic that relies on averaging (because that’s what the mean is… an average), so let’s look at that. When applied to the sample mean, what the law of large numbers states is that as the sample gets larger, the sample mean tends to get closer to the true population mean. Or, to say it a little bit more precisely, as the sample size “approaches” infinity (written as \\(N \\rightarrow \\infty\\)) the sample mean approaches the population mean (\\(\\bar{X} \\rightarrow \\mu\\)).150\n\nI don’t intend to subject you to a proof that the law of large numbers is true, but it’s one of the most important tools for statistical theory. The law of large numbers is the thing we can use to justify our belief that collecting more and more data will eventually lead us to the truth. For any particular data set, the sample statistics that we calculate from it will be wrong, but the law of large numbers tells us that if we keep collecting more data those sample statistics will tend to get closer and closer to the true population parameters.\n\n## 10.3 Sampling distributions and the central limit theorem\n\nThe law of large numbers is a very powerful tool, but it’s not going to be good enough to answer all our questions. Among other things, all it gives us is a “long run guarantee”. In the long run, if we were somehow able to collect an infinite amount of data, then the law of large numbers guarantees that our sample statistics will be correct. But as famously argued in economics, a long run guarantee is of little use in real life:\n [The] long run is a misleading guide to current affairs. In the long run we are all dead. Economists set themselves too easy, too useless a task, if in tempestuous seasons they can only tell us, that when the storm is long past, the ocean is flat again. Keynes (1923)\n\nAs in economics, so too in psychology and statistics. It is not enough to know that we will eventually arrive at the right answer when calculating the sample mean. Knowing that an infinitely large data set will tell me the exact value of the population mean is cold comfort when my actual data set has a sample size of \\(N=100\\). In real life, then, we must know something about the behaviour of the sample mean when it is calculated from a more modest data set!\n\n### 10.3.1 Sampling distribution of the mean\n\n With this in mind, let’s abandon the idea that our studies will have sample sizes of 10000, and consider a very modest experiment indeed. This time around we’ll sample \\(N=5\\) people and measure their IQ scores. As before, I can simulate this experiment in R using the `rnorm()` function: \n\n```\n> IQ.1 <- round( rnorm(n=5, mean=100, sd=15 ))\n> IQ.1\n[1] 90 82 94 99 110\n```\n\nThe mean IQ in this sample turns out to be exactly 95. Not surprisingly, this is much less accurate than the previous experiment. Now imagine that I decided to replicate the experiment. That is, I repeat the procedure as closely as possible: I randomly sample 5 new people and measure their IQ. Again, R allows me to simulate the results of this procedure:\n\n```\n> IQ.2 <- round( rnorm(n=5, mean=100, sd=15 ))\n> IQ.2\n[1] 78 88 111 111 117\n```\n\nThis time around, the mean IQ in my sample is 101. If I repeat the experiment 10 times I obtain the results shown in Table ??, and as you can see the sample mean varies from one replication to the next.\n\nNANA | Person.1 | Person.2 | Person.3 | Person.4 | Person.5 | Sample.Mean | caption |\n | --- | --- | --- | --- | --- | --- | --- | --- |\nReplication 1 | 90 | 82 | 94 | 99 | 110 | 95.0 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 2 | 78 | 88 | 111 | 111 | 117 | 101.0 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 3 | 111 | 122 | 91 | 98 | 86 | 101.6 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 4 | 98 | 96 | 119 | 99 | 107 | 103.8 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 5 | 105 | 113 | 103 | 103 | 98 | 104.4 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 6 | 81 | 89 | 93 | 85 | 114 | 92.4 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 7 | 100 | 93 | 108 | 98 | 133 | 106.4 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 8 | 107 | 100 | 105 | 117 | 85 | 102.8 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 9 | 86 | 119 | 108 | 73 | 116 | 100.4 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\nReplication 10 | 95 | 126 | 112 | 120 | 76 | 105.8 | Ten replications of the IQ experiment, each with a sample size of \\(N=5\\). |\n\nNow suppose that I decided to keep going in this fashion, replicating this “five IQ scores” experiment over and over again. Every time I replicate the experiment I write down the sample mean. Over time, I’d be amassing a new data set, in which every experiment generates a single data point. The first 10 observations from my data set are the sample means listed in Table ??, so my data set starts out like this:\n\n```\n95.0 101.0 101.6 103.8 104.4 ...\n```\n\nWhat if I continued like this for 10,000 replications, and then drew a histogram? Using the magical powers of R that’s exactly what I did, and you can see the results in Figure 10.5. As this picture illustrates, the average of 5 IQ scores is usually between 90 and 110. But more importantly, what it highlights is that if we replicate an experiment over and over again, what we end up with is a distribution of sample means! This distribution has a special name in statistics: it’s called the sampling distribution of the mean.\n\nSampling distributions are another important theoretical idea in statistics, and they’re crucial for understanding the behaviour of small samples. For instance, when I ran the very first “five IQ scores” experiment, the sample mean turned out to be 95. What the sampling distribution in Figure 10.5 tells us, though, is that the “five IQ scores” experiment is not very accurate. If I repeat the experiment, the sampling distribution tells me that I can expect to see a sample mean anywhere between 80 and 120.\n\n### 10.3.2 Sampling distributions exist for any sample statistic!\n\nOne thing to keep in mind when thinking about sampling distributions is that any sample statistic you might care to calculate has a sampling distribution. For example, suppose that each time I replicated the “five IQ scores” experiment I wrote down the largest IQ score in the experiment. This would give me a data set that started out like this:\n\n```\n110 117 122 119 113 ...\n```\n\nDoing this over and over again would give me a very different sampling distribution, namely the sampling distribution of the maximum. The sampling distribution of the maximum of 5 IQ scores is shown in Figure 10.6. Not surprisingly, if you pick 5 people at random and then find the person with the highest IQ score, they’re going to have an above average IQ. Most of the time you’ll end up with someone whose IQ is measured in the 100 to 140 range.\n\n### 10.3.3 The central limit theorem\n\nAn illustration of the how sampling distribution of the mean depends on sample size. In each panel, I generated 10,000 samples of IQ data, and calculated the mean IQ observed within each of these data sets. The histograms in these plots show the distribution of these means (i.e., the sampling distribution of the mean). Each individual IQ score was drawn from a normal distribution with mean 100 and standard deviation 15, which is shown as the solid black line).\n\nAt this point I hope you have a pretty good sense of what sampling distributions are, and in particular what the sampling distribution of the mean is. In this section I want to talk about how the sampling distribution of the mean changes as a function of sample size. Intuitively, you already know part of the answer: if you only have a few observations, the sample mean is likely to be quite inaccurate: if you replicate a small experiment and recalculate the mean you’ll get a very different answer. In other words, the sampling distribution is quite wide. If you replicate a large experiment and recalculate the sample mean you’ll probably get the same answer you got last time, so the sampling distribution will be very narrow. You can see this visually in Figures 10.7, 10.8 and 10.9: the bigger the sample size, the narrower the sampling distribution gets. We can quantify this effect by calculating the standard deviation of the sampling distribution, which is referred to as the standard error. The standard error of a statistic is often denoted SE, and since we’re usually interested in the standard error of the sample mean, we often use the acronym SEM. As you can see just by looking at the picture, as the sample size \\(N\\) increases, the SEM decreases.\n\nOkay, so that’s one part of the story. However, there’s something I’ve been glossing over so far. All my examples up to this point have been based on the “IQ scores” experiments, and because IQ scores are roughly normally distributed, I’ve assumed that the population distribution is normal. What if it isn’t normal? What happens to the sampling distribution of the mean? The remarkable thing is this: no matter what shape your population distribution is, as \\(N\\) increases the sampling distribution of the mean starts to look more like a normal distribution. To give you a sense of this, I ran some simulations using R. To do this, I started with the “ramped” distribution shown in the histogram in Figure 10.10. As you can see by comparing the triangular shaped histogram to the bell curve plotted by the black line, the population distribution doesn’t look very much like a normal distribution at all. Next, I used R to simulate the results of a large number of experiments. In each experiment I took \\(N=2\\) samples from this distribution, and then calculated the sample mean. Figure ?? plots the histogram of these sample means (i.e., the sampling distribution of the mean for \\(N=2\\)). This time, the histogram produces a \\(\\cap\\)-shaped distribution: it’s still not normal, but it’s a lot closer to the black line than the population distribution in Figure ??. When I increase the sample size to \\(N=4\\), the sampling distribution of the mean is very close to normal (Figure ??, and by the time we reach a sample size of \\(N=8\\) it’s almost perfectly normal. In other words, as long as your sample size isn’t tiny, the sampling distribution of the mean will be approximately normal no matter what your population distribution looks like!\n\n```\n# needed for printing\n width <- 6\n height <- 6 \n\n # parameters of the beta\n a <- 2\n b <- 1\n\n # mean and standard deviation of the beta\n s <- sqrt( a*b / (a+b)^2 / (a+b+1) )\n m <- a / (a+b)\n\n # define function to draw a plot\n plotOne <- function(n,N=50000) {\n\n # generate N random sample means of size n\n X <- matrix(rbeta(n*N,a,b),n,N)\n X <- colMeans(X)\n\n # plot the data\n hist( X, breaks=seq(0,1,.025), border=\"white\", freq=FALSE,\n col=ifelse(colour,emphColLight,emphGrey),\n xlab=\"Sample Mean\", ylab=\"\", xlim=c(0,1.2),\n main=paste(\"Sample Size =\",n), axes=FALSE,\n font.main=1, ylim=c(0,5)\n )\n box()\n axis(1)\n #axis(2)\n\n # plot the theoretical distribution\n lines( x <- seq(0,1.2,.01), dnorm(x,m,s/sqrt(n)), \n lwd=2, col=\"black\", type=\"l\"\n )\n }\n\n for( i in c(1,2,4,8)) {\n plotOne(i)}\n```\n\nOn the basis of these figures, it seems like we have evidence for all of the following claims about the sampling distribution of the mean:\n\n* The mean of the sampling distribution is the same as the mean of the population\n * The standard deviation of the sampling distribution (i.e., the standard error) gets smaller as the sample size increases\n * The shape of the sampling distribution becomes normal as the sample size increases\n\nAs it happens, not only are all of these statements true, there is a very famous theorem in statistics that proves all three of them, known as the central limit theorem. Among other things, the central limit theorem tells us that if the population distribution has mean \\(\\mu\\) and standard deviation \\(\\sigma\\), then the sampling distribution of the mean also has mean \\(\\mu\\), and the standard error of the mean is \\[ \\mbox{SEM} = \\frac{\\sigma}{ \\sqrt{N} } \\] Because we divide the population standard devation \\(\\sigma\\) by the square root of the sample size \\(N\\), the SEM gets smaller as the sample size increases. It also tells us that the shape of the sampling distribution becomes normal.151\n\nThis result is useful for all sorts of things. It tells us why large experiments are more reliable than small ones, and because it gives us an explicit formula for the standard error it tells us how much more reliable a large experiment is. It tells us why the normal distribution is, well, normal. In real experiments, many of the things that we want to measure are actually averages of lots of different quantities (e.g., arguably, “general” intelligence as measured by IQ is an average of a large number of “specific” skills and abilities), and when that happens, the averaged quantity should follow a normal distribution. Because of this mathematical law, the normal distribution pops up over and over again in real data.\n\n## 10.4 Estimating population parameters\n\nIn all the IQ examples in the previous sections, we actually knew the population parameters ahead of time. As every undergraduate gets taught in their very first lecture on the measurement of intelligence, IQ scores are defined to have mean 100 and standard deviation 15. However, this is a bit of a lie. How do we know that IQ scores have a true population mean of 100? Well, we know this because the people who designed the tests have administered them to very large samples, and have then “rigged” the scoring rules so that their sample has mean 100. That’s not a bad thing of course: it’s an important part of designing a psychological measurement. However, it’s important to keep in mind that this theoretical mean of 100 only attaches to the population that the test designers used to design the tests. Good test designers will actually go to some lengths to provide “test norms” that can apply to lots of different populations (e.g., different age groups, nationalities etc).\n\nThis is very handy, but of course almost every research project of interest involves looking at a different population of people to those used in the test norms. For instance, suppose you wanted to measure the effect of low level lead poisoning on cognitive functioning in Port Pirie, a South Australian industrial town with a lead smelter. Perhaps you decide that you want to compare IQ scores among people in Port Pirie to a comparable sample in Whyalla, a South Australian industrial town with a steel refinery.152 Regardless of which town you’re thinking about, it doesn’t make a lot of sense simply to assume that the true population mean IQ is 100. No-one has, to my knowledge, produced sensible norming data that can automatically be applied to South Australian industrial towns. We’re going to have to estimate the population parameters from a sample of data. So how do we do this?\n\n### 10.4.1 Estimating the population mean\n\nSuppose we go to Port Pirie and 100 of the locals are kind enough to sit through an IQ test. The average IQ score among these people turns out to be \\(\\bar{X}=98.5\\). So what is the true mean IQ for the entire population of Port Pirie? Obviously, we don’t know the answer to that question. It could be \\(97.2\\), but if could also be \\(103.5\\). Our sampling isn’t exhaustive so we cannot give a definitive answer. Nevertheless if I was forced at gunpoint to give a “best guess” I’d have to say \\(98.5\\). That’s the essence of statistical estimation: giving a best guess.\n\nIn this example, estimating the unknown poulation parameter is straightforward. I calculate the sample mean, and I use that as my estimate of the population mean. It’s pretty simple, and in the next section I’ll explain the statistical justification for this intuitive answer. However, for the moment what I want to do is make sure you recognise that the sample statistic and the estimate of the population parameter are conceptually different things. A sample statistic is a description of your data, whereas the estimate is a guess about the population. With that in mind, statisticians often different notation to refer to them. For instance, if true population mean is denoted \\(\\mu\\), then we would use \\(\\hat\\mu\\) to refer to our estimate of the population mean. In contrast, the sample mean is denoted \\(\\bar{X}\\) or sometimes \\(m\\). However, in simple random samples, the estimate of the population mean is identical to the sample mean: if I observe a sample mean of \\(\\bar{X} = 98.5\\), then my estimate of the population mean is also \\(\\hat\\mu = 98.5\\). To help keep the notation clear, here’s a handy table:\n\n```\nknitr::kable(data.frame(stringsAsFactors=FALSE,\n Symbol = c(\"$\\\\bar{X}$\", \"$\\\\mu$\", \"$\\\\hat{\\\\mu}$\"),\n What.is.it = c(\"Sample mean\", \"True population mean\",\n \"Estimate of the population mean\"),\n Do.we.know.what.it.is = c(\"Yes calculated from the raw data\",\n \"Almost never known for sure\",\n \"Yes identical to the sample mean\")))\n```\n\nSymbol | What.is.it | Do.we.know.what.it.is |\n | --- | --- | --- |\n\\(\\bar{X}\\) | Sample mean | Yes calculated from the raw data |\n\\(\\mu\\) | True population mean | Almost never known for sure |\n\\(\\hat{\\mu}\\) | Estimate of the population mean | Yes identical to the sample mean |\n\n### 10.4.2 Estimating the population standard deviation\n\nSo far, estimation seems pretty simple, and you might be wondering why I forced you to read through all that stuff about sampling theory. In the case of the mean, our estimate of the population parameter (i.e. \\(\\hat\\mu\\)) turned out to identical to the corresponding sample statistic (i.e. \\(\\bar{X}\\)). However, that’s not always true. To see this, let’s have a think about how to construct an estimate of the population standard deviation, which we’ll denote \\(\\hat\\sigma\\). What shall we use as our estimate in this case? Your first thought might be that we could do the same thing we did when estimating the mean, and just use the sample statistic as our estimate. That’s almost the right thing to do, but not quite.\n\nHere’s why. Suppose I have a sample that contains a single observation. For this example, it helps to consider a sample where you have no intutions at all about what the true population values might be, so let’s use something completely fictitious. Suppose the observation in question measures the cromulence of my shoes. It turns out that my shoes have a cromulence of 20. So here’s my sample:\n `20` \nThis is a perfectly legitimate sample, even if it does have a sample size of \\(N=1\\). It has a sample mean of 20, and because every observation in this sample is equal to the sample mean (obviously!) it has a sample standard deviation of 0. As a description of the sample this seems quite right: the sample contains a single observation and therefore there is no variation observed within the sample. A sample standard deviation of \\(s = 0\\) is the right answer here. But as an estimate of the population standard deviation, it feels completely insane, right? Admittedly, you and I don’t know anything at all about what “cromulence” is, but we know something about data: the only reason that we don’t see any variability in the sample is that the sample is too small to display any variation! So, if you have a sample size of \\(N=1\\), it feels like the right answer is just to say “no idea at all”.\n\nNotice that you don’t have the same intuition when it comes to the sample mean and the population mean. If forced to make a best guess about the population mean, it doesn’t feel completely insane to guess that the population mean is 20. Sure, you probably wouldn’t feel very confident in that guess, because you have only the one observation to work with, but it’s still the best guess you can make.\n\nLet’s extend this example a little. Suppose I now make a second observation. My data set now has \\(N=2\\) observations of the cromulence of shoes, and the complete sample now looks like this:\n `20, 22` \nThis time around, our sample is just large enough for us to be able to observe some variability: two observations is the bare minimum number needed for any variability to be observed! For our new data set, the sample mean is \\(\\bar{X}=21\\), and the sample standard deviation is \\(s=1\\). What intuitions do we have about the population? Again, as far as the population mean goes, the best guess we can possibly make is the sample mean: if forced to guess, we’d probably guess that the population mean cromulence is 21. What about the standard deviation? This is a little more complicated. The sample standard deviation is only based on two observations, and if you’re at all like me you probably have the intuition that, with only two observations, we haven’t given the population “enough of a chance” to reveal its true variability to us. It’s not just that we suspect that the estimate is wrong: after all, with only two observations we expect it to be wrong to some degree. The worry is that the error is systematic. Specifically, we suspect that the sample standard deviation is likely to be smaller than the population standard deviation.\n This intuition feels right, but it would be nice to demonstrate this somehow. There are in fact mathematical proofs that confirm this intuition, but unless you have the right mathematical background they don’t help very much. Instead, what I’ll do is use R to simulate the results of some experiments. With that in mind, let’s return to our IQ studies. Suppose the true population mean IQ is 100 and the standard deviation is 15. I can use the `rnorm()` function to generate the the results of an experiment in which I measure \\(N=2\\) IQ scores, and calculate the sample standard deviation. If I do this over and over again, and plot a histogram of these sample standard deviations, what I have is the sampling distribution of the standard deviation. I’ve plotted this distribution in Figure 10.11. Even though the true population standard deviation is 15, the average of the sample standard deviations is only 8.5. Notice that this is a very different result to what we found in Figure 10.8 when we plotted the sampling distribution of the mean. If you look at that sampling distribution, what you see is that the population mean is 100, and the average of the sample means is also 100. \nNow let’s extend the simulation. Instead of restricting ourselves to the situation where we have a sample size of \\(N=2\\), let’s repeat the exercise for sample sizes from 1 to 10. If we plot the average sample mean and average sample standard deviation as a function of sample size, you get the results shown in Figure 10.12. On the left hand side (panel a), I’ve plotted the average sample mean and on the right hand side (panel b), I’ve plotted the average standard deviation. The two plots are quite different: on average, the average sample mean is equal to the population mean. It is an unbiased estimator, which is essentially the reason why your best estimate for the population mean is the sample mean.153 The plot on the right is quite different: on average, the sample standard deviation \\(s\\) is smaller than the population standard deviation \\(\\sigma\\). It is a biased estimator. In other words, if we want to make a “best guess” \\(\\hat\\sigma\\) about the value of the population standard deviation \\(\\sigma\\), we should make sure our guess is a little bit larger than the sample standard deviation \\(s\\).\n The fix to this systematic bias turns out to be very simple. Here’s how it works. Before tackling the standard deviation, let’s look at the variance. If you recall from Section 5.2, the sample variance is defined to be the average of the squared deviations from the sample mean. That is: \\[ s^2 = \\frac{1}{N} \\sum_{i=1}^N (X_i - \\bar{X})^2 \\] The sample variance \\(s^2\\) is a biased estimator of the population variance \\(\\sigma^2\\). But as it turns out, we only need to make a tiny tweak to transform this into an unbiased estimator. All we have to do is divide by \\(N-1\\) rather than by \\(N\\). If we do that, we obtain the following formula: \\[ \\hat\\sigma^2 = \\frac{1}{N-1} \\sum_{i=1}^N (X_i - \\bar{X})^2 \\] This is an unbiased estimator of the population variance \\(\\sigma\\). Moreover, this finally answers the question we raised in Section 5.2. Why did R give us slightly different answers when we used the `var()` function? Because the `var()` function calculates \\(\\hat\\sigma^2\\) not \\(s^2\\), that’s why. A similar story applies for the standard deviation. If we divide by \\(N-1\\) rather than \\(N\\), our estimate of the population standard deviation becomes: \\[\n\\hat\\sigma = \\sqrt{\\frac{1}{N-1} \\sum_{i=1}^N (X_i - \\bar{X})^2} \n\\] and when we use R’s built in standard deviation function `sd()` , what it’s doing is calculating \\(\\hat\\sigma\\), not \\(s\\).154 \nOne final point: in practice, a lot of people tend to refer to \\(\\hat{\\sigma}\\) (i.e., the formula where we divide by \\(N-1\\)) as the sample standard deviation. Technically, this is incorrect: the sample standard deviation should be equal to \\(s\\) (i.e., the formula where we divide by \\(N\\)). These aren’t the same thing, either conceptually or numerically. One is a property of the sample, the other is an estimated characteristic of the population. However, in almost every real life application, what we actually care about is the estimate of the population parameter, and so people always report \\(\\hat\\sigma\\) rather than \\(s\\). This is the right number to report, of course, it’s that people tend to get a little bit imprecise about terminology when they write it up, because “sample standard deviation” is shorter than “estimated population standard deviation”. It’s no big deal, and in practice I do the same thing everyone else does. Nevertheless, I think it’s important to keep the two concepts separate: it’s never a good idea to confuse “known properties of your sample” with “guesses about the population from which it came”. The moment you start thinking that \\(s\\) and \\(\\hat\\sigma\\) are the same thing, you start doing exactly that.\n\nTo finish this section off, here’s another couple of tables to help keep things clear:\n\n```\nknitr::kable(data.frame(stringsAsFactors=FALSE,\n Symbol = c(\"$s$\", \"$\\\\sigma$\", \"$\\\\hat{\\\\sigma}$\", \"$s^2$\",\n \"$\\\\sigma^2$\", \"$\\\\hat{\\\\sigma}^2$\"),\n What.is.it = c(\"Sample standard deviation\",\n \"Population standard deviation\",\n \"Estimate of the population standard deviation\", \"Sample variance\",\n \"Population variance\",\n \"Estimate of the population variance\"),\n Do.we.know.what.it.is = c(\"Yes - calculated from the raw data\",\n \"Almost never known for sure\",\n \"Yes - but not the same as the sample standard deviation\",\n \"Yes - calculated from the raw data\",\n \"Almost never known for sure\",\n \"Yes - but not the same as the sample variance\")\n))\n```\n\nSymbol | What.is.it | Do.we.know.what.it.is |\n | --- | --- | --- |\n\\(s\\) | Sample standard deviation | Yes - calculated from the raw data |\n\\(\\sigma\\) | Population standard deviation | Almost never known for sure |\n\\(\\hat{\\sigma}\\) | Estimate of the population standard deviation | Yes - but not the same as the sample standard deviation |\n\\(s^2\\) | Sample variance | Yes - calculated from the raw data |\n\\(\\sigma^2\\) | Population variance | Almost never known for sure |\n\\(\\hat{\\sigma}^2\\) | Estimate of the population variance | Yes - but not the same as the sample variance |\n\n## 10.5 Estimating a confidence interval\n\nStatistics means never having to say you’re certain – Unknown origin155 but I’ve never found the original source.\n\nUp to this point in this chapter, I’ve outlined the basics of sampling theory which statisticians rely on to make guesses about population parameters on the basis of a sample of data. As this discussion illustrates, one of the reasons we need all this sampling theory is that every data set leaves us with a some of uncertainty, so our estimates are never going to be perfectly accurate. The thing that has been missing from this discussion is an attempt to quantify the amount of uncertainty that attaches to our estimate. It’s not enough to be able guess that, say, the mean IQ of undergraduate psychology students is 115 (yes, I just made that number up). We also want to be able to say something that expresses the degree of certainty that we have in our guess. For example, it would be nice to be able to say that there is a 95% chance that the true mean lies between 109 and 121. The name for this is a confidence interval for the mean.\n Armed with an understanding of sampling distributions, constructing a confidence interval for the mean is actually pretty easy. Here’s how it works. Suppose the true population mean is \\(\\mu\\) and the standard deviation is \\(\\sigma\\). I’ve just finished running my study that has \\(N\\) participants, and the mean IQ among those participants is \\(\\bar{X}\\). We know from our discussion of the central limit theorem (Section 10.3.3 that the sampling distribution of the mean is approximately normal. We also know from our discussion of the normal distribution Section 9.5 that there is a 95% chance that a normally-distributed quantity will fall within two standard deviations of the true mean. To be more precise, we can use the `qnorm()` function to compute the 2.5th and 97.5th percentiles of the normal distribution \n\n```\nqnorm( p = c(.025, .975) )\n```\n\n```\n## [1] -1.959964 1.959964\n```\n\n Okay, so I lied earlier on. The more correct answer is that 95% chance that a normally-distributed quantity will fall within 1.96 standard deviations of the true mean. Next, recall that the standard deviation of the sampling distribution is referred to as the standard error, and the standard error of the mean is written as SEM. When we put all these pieces together, we learn that there is a 95% probability that the sample mean \\(\\bar{X}\\) that we have actually observed lies within 1.96 standard errors of the population mean. Mathematically, we write this as: \\[ \\mu - \\left( 1.96 \\times \\mbox{SEM} \\right) \\ \\leq \\ \\bar{X}\\ \\leq \\ \\mu + \\left( 1.96 \\times \\mbox{SEM} \\right) \\] where the SEM is equal to \\(\\sigma / \\sqrt{N}\\), and we can be 95% confident that this is true. However, that’s not answering the question that we’re actually interested in. The equation above tells us what we should expect about the sample mean, given that we know what the population parameters are. What we want is to have this work the other way around: we want to know what we should believe about the population parameters, given that we have observed a particular sample. However, it’s not too difficult to do this. Using a little high school algebra, a sneaky way to rewrite our equation is like this: \\[ \\bar{X} - \\left( 1.96 \\times \\mbox{SEM} \\right) \\ \\leq \\ \\mu \\ \\leq \\ \\bar{X} + \\left( 1.96 \\times \\mbox{SEM}\\right) \\] What this is telling is is that the range of values has a 95% probability of containing the population mean \\(\\mu\\). We refer to this range as a 95% confidence interval, denoted \\(\\mbox{CI}_{95}\\). In short, as long as \\(N\\) is sufficiently large – large enough for us to believe that the sampling distribution of the mean is normal – then we can write this as our formula for the 95% confidence interval: \\[ \\mbox{CI}_{95} = \\bar{X} \\pm \\left( 1.96 \\times \\frac{\\sigma}{\\sqrt{N}} \\right) \\] Of course, there’s nothing special about the number 1.96: it just happens to be the multiplier you need to use if you want a 95% confidence interval. If I’d wanted a 70% confidence interval, I could have used the `qnorm()` function to calculate the 15th and 85th quantiles: \n\n```\nqnorm( p = c(.15, .85) )\n```\n\n```\n## [1] -1.036433 1.036433\n```\n\nand so the formula for \\(\\mbox{CI}_{70}\\) would be the same as the formula for \\(\\mbox{CI}_{95}\\) except that we’d use 1.04 as our magic number rather than 1.96.\n\n### 10.5.1 A slight mistake in the formula\n\n As usual, I lied. The formula that I’ve given above for the 95% confidence interval is approximately correct, but I glossed over an important detail in the discussion. Notice my formula requires you to use the standard error of the mean, SEM, which in turn requires you to use the true population standard deviation \\(\\sigma\\). Yet, in Section @ref(pointestimates I stressed the fact that we don’t actually know the true population parameters. Because we don’t know the true value of \\(\\sigma\\), we have to use an estimate of the population standard deviation \\(\\hat{\\sigma}\\) instead. This is pretty straightforward to do, but this has the consequence that we need to use the quantiles of the \\(t\\)-distribution rather than the normal distribution to calculate our magic number; and the answer depends on the sample size. When \\(N\\) is very large, we get pretty much the same value using `qt()` that we would if we used `qnorm()` … \n\n `## [1] 1.960201` \nBut when \\(N\\) is small, we get a much bigger number when we use the \\(t\\) distribution:\n\n `## [1] 2.262157` \nThere’s nothing too mysterious about what’s happening here. Bigger values mean that the confidence interval is wider, indicating that we’re more uncertain about what the true value of \\(\\mu\\) actually is. When we use the \\(t\\) distribution instead of the normal distribution, we get bigger numbers, indicating that we have more uncertainty. And why do we have that extra uncertainty? Well, because our estimate of the population standard deviation \\(\\hat\\sigma\\) might be wrong! If it’s wrong, it implies that we’re a bit less sure about what our sampling distribution of the mean actually looks like… and this uncertainty ends up getting reflected in a wider confidence interval.\n\n### 10.5.2 Interpreting a confidence interval\n\nThe hardest thing about confidence intervals is understanding what they mean. Whenever people first encounter confidence intervals, the first instinct is almost always to say that “there is a 95% probabaility that the true mean lies inside the confidence interval”. It’s simple, and it seems to capture the common sense idea of what it means to say that I am “95% confident”. Unfortunately, it’s not quite right. The intuitive definition relies very heavily on your own personal beliefs about the value of the population mean. I say that I am 95% confident because those are my beliefs. In everyday life that’s perfectly okay, but if you remember back to Section 9.2, you’ll notice that talking about personal belief and confidence is a Bayesian idea. Personally (speaking as a Bayesian) I have no problem with the idea that the phrase “95% probability” is allowed to refer to a personal belief. However, confidence intervals are not Bayesian tools. Like everything else in this chapter, confidence intervals are frequentist tools, and if you are going to use frequentist methods then it’s not appropriate to attach a Bayesian interpretation to them. If you use frequentist methods, you must adopt frequentist interpretations!\n\nOkay, so if that’s not the right answer, what is? Remember what we said about frequentist probability: the only way we are allowed to make “probability statements” is to talk about a sequence of events, and to count up the frequencies of different kinds of events. From that perspective, the interpretation of a 95% confidence interval must have something to do with replication. Specifically: if we replicated the experiment over and over again and computed a 95% confidence interval for each replication, then 95% of those intervals would contain the true mean. More generally, 95% of all confidence intervals constructed using this procedure should contain the true population mean. This idea is illustrated in Figure 10.13, which shows 50 confidence intervals constructed for a “measure 10 IQ scores” experiment (top panel) and another 50 confidence intervals for a “measure 25 IQ scores” experiment (bottom panel). A bit fortuitously, across the 100 replications that I simulated, it turned out that exactly 95 of them contained the true mean.\n\nThe critical difference here is that the Bayesian claim makes a probability statement about the population mean (i.e., it refers to our uncertainty about the population mean), which is not allowed under the frequentist interpretation of probability because you can’t “replicate” a population! In the frequentist claim, the population mean is fixed and no probabilistic claims can be made about it. Confidence intervals, however, are repeatable so we can replicate experiments. Therefore a frequentist is allowed to talk about the probability that the confidence interval (a random variable) contains the true mean; but is not allowed to talk about the probability that the true population mean (not a repeatable event) falls within the confidence interval.\n\nI know that this seems a little pedantic, but it does matter. It matters because the difference in interpretation leads to a difference in the mathematics. There is a Bayesian alternative to confidence intervals, known as credible intervals. In most situations credible intervals are quite similar to confidence intervals, but in other cases they are drastically different. As promised, though, I’ll talk more about the Bayesian perspective in Chapter 17.\n\n### 10.5.3 Calculating confidence intervals in R\n\n As far as I can tell, the core packages in R don’t include a simple function for calculating confidence intervals for the mean. They do include a lot of complicated, extremely powerful functions that can be used to calculate confidence intervals associated with lots of different things, such as the `confint()` function that we’ll use in Chapter 15. But I figure that when you’re first learning statistics, it might be useful to start with something simpler. As a consequence, the `lsr` package includes a function called `ciMean()` which you can use to calculate your confidence intervals. There are two arguments that you might want to specify:156 \n\n* \n `x` . This should be a numeric vector containing the data. * \n `conf` . This should be a number, specifying the confidence level. By default, `conf = .95` , since 95% confidence intervals are the de facto standard in psychology. So, for example, if I load the `afl24.Rdata` file, calculate the confidence interval associated with the mean attendance: \n\n```\n> ciMean( x = afl$attendance )\n 2.5% 97.5% \n31597.32 32593.12\n```\n\nHopefully that’s fairly clear.\n\n### 10.5.4 Plotting confidence intervals in R\n\n There’s several different ways you can draw graphs that show confidence intervals as error bars. I’ll show three versions here, but this certainly doesn’t exhaust the possibilities. In doing so, what I’m assuming is that you want to draw is a plot showing the means and confidence intervals for one variable, broken down by different levels of a second variable. For instance, in our `afl` data that we discussed earlier, we might be interested in plotting the average `attendance` by `year` . I’ll do this using two different functions, `bargraph.CI()` and `lineplot.CI()` (both of which are in the `sciplot` package). Assuming that you’ve installed these packages on your system (see Section 4.2 if you’ve forgotten how to do this), you’ll need to load them. You’ll also need to load the `lsr` package, because we’ll make use of the `ciMean()` function to actually calculate the confidence intervals \n\n```\nload( file.path(projecthome, \"data/afl24.Rdata\" )) # contains the \"afl\" data frame\nlibrary( sciplot ) # bargraph.CI() and lineplot.CI() functions\nlibrary( lsr ) # ciMean() function\n```\n\n Here’s how to plot the means and confidence intervals drawn using `bargraph.CI()` . \n\n We can use the same arguments when calling the `lineplot.CI()` function: \n\n## 10.6 Summary\n\nIn this chapter I’ve covered two main topics. The first half of the chapter talks about sampling theory, and the second half talks about how we can use sampling theory to construct estimates of the population parameters. The section breakdown looks like this:\n\n* Basic ideas about samples, sampling and populations (Section 10.1)\n * Statistical theory of sampling: the law of large numbers (Section 10.2), sampling distributions and the central limit theorem (Section 10.3).\n * Estimating means and standard deviations (Section 10.4)\n * Estimating a confidence interval (Section 10.5)\n\nAs always, there’s a lot of topics related to sampling and estimation that aren’t covered in this chapter, but for an introductory psychology class this is fairly comprehensive I think. For most applied researchers you won’t need much more theory than this. One big question that I haven’t touched on in this chapter is what you do when you don’t have a simple random sample. There is a lot of statistical theory you can draw on to handle this situation, but it’s well beyond the scope of this book.\n\nThe proper mathematical definition of randomness is extraordinarily technical, and way beyond the scope of this book. We’ll be non-technical here and say that a process has an element of randomness to it whenever it is possible to repeat the process and get different answers each time.↩\n * \n\nNothing in life is that simple: there’s not an obvious division of people into binary categories like “schizophrenic” and “not schizophrenic”. But this isn’t a clinical psychology text, so please forgive me a few simplifications here and there.↩\n * \n\nTechnically, the law of large numbers pertains to any sample statistic that can be described as an average of independent quantities. That’s certainly true for the sample mean. However, it’s also possible to write many other sample statistics as averages of one form or another. The variance of a sample, for instance, can be rewritten as a kind of average and so is subject to the law of large numbers. The minimum value of a sample, however, cannot be written as an average of anything and is therefore not governed by the law of large numbers.↩\n * \n\nAs usual, I’m being a bit sloppy here. The central limit theorem is a bit more general than this section implies. Like most introductory stats texts, I’ve discussed one situation where the central limit theorem holds: when you’re taking an average across lots of independent events drawn from the same distribution. However, the central limit theorem is much broader than this. There’s a whole class of things called “\\(U\\)-statistics” for instance, all of which satisfy the central limit theorem and therefore become normally distributed for large sample sizes. The mean is one such statistic, but it’s not the only one.↩\n * \n\nPlease note that if you were actually interested in this question, you would need to be a lot more careful than I’m being here. You can’t just compare IQ scores in Whyalla to Port Pirie and assume that any differences are due to lead poisoning. Even if it were true that the only differences between the two towns corresponded to the different refineries (and it isn’t, not by a long shot), you need to account for the fact that people already believe that lead pollution causes cognitive deficits: if you recall back to Chapter 2, this means that there are different demand effects for the Port Pirie sample than for the Whyalla sample. In other words, you might end up with an illusory group difference in your data, caused by the fact that people think that there is a real difference. I find it pretty implausible to think that the locals wouldn’t be well aware of what you were trying to do if a bunch of researchers turned up in Port Pirie with lab coats and IQ tests, and even less plausible to think that a lot of people would be pretty resentful of you for doing it. Those people won’t be as co-operative in the tests. Other people in Port Pirie might be more motivated to do well because they don’t want their home town to look bad. The motivational effects that would apply in Whyalla are likely to be weaker, because people don’t have any concept of “iron ore poisoning” in the same way that they have a concept for “lead poisoning”. Psychology is hard.↩\n * \n\nI should note that I’m hiding something here. Unbiasedness is a desirable characteristic for an estimator, but there are other things that matter besides bias. However, it’s beyond the scope of this book to discuss this in any detail. I just want to draw your attention to the fact that there’s some hidden complexity here.↩\n * \n\nOkay, I’m hiding something else here. In a bizarre and counterintuitive twist, since \\(\\hat\\sigma^2\\) is an unbiased estimator of \\(\\sigma^2\\), you’d assume that taking the square root would be fine, and \\(\\hat\\sigma\\) would be an unbiased estimator of \\(\\sigma\\). Right? Weirdly, it’s not. There’s actually a subtle, tiny bias in \\(\\hat\\sigma\\). This is just bizarre: \\(\\hat\\sigma^2\\) is and unbiased estimate of the population variance \\(\\sigma^2\\), but when you take the square root, it turns out that \\(\\hat\\sigma\\) is a biased estimator of the population standard deviation \\(\\sigma\\). Weird, weird, weird, right? So, why is \\(\\hat\\sigma\\) biased? The technical answer is “because non-linear transformations (e.g., the square root) don’t commute with expectation”, but that just sounds like gibberish to everyone who hasn’t taken a course in mathematical statistics. Fortunately, it doesn’t matter for practical purposes. The bias is small, and in real life everyone uses \\(\\hat\\sigma\\) and it works just fine. Sometimes mathematics is just annoying.↩\n * \n\nThis quote appears on a great many t-shirts and websites, and even gets a mention in a few academic papers (e.g., \\url{http://www.amstat.org/publications/jse/v10n3/friedman.html↩\n * \n\nAs of the current writing, these are the only arguments to the function. However, I am planning to add a bit more functionality to\n `ciMean()` . However, regardless of what those future changes might look like, the `x` and `conf` arguments will remain the same, and the commands used in this book will still work.↩\n\nDate: 2006-01-01\nCategories: \nTags: \n\n# Chapter 11 Hypothesis testing\n\nThe process of induction is the process of assuming the simplest law that can be made to harmonize with our experience. This process, however, has no logical foundation but only a psychological one. It is clear that there are no grounds for believing that the simplest course of events will really happen. It is an hypothesis that the sun will rise tomorrow: and this means that we do not know whether it will rise.\n\n– \n\nIn the last chapter, I discussed the ideas behind estimation, which is one of the two “big ideas” in inferential statistics. It’s now time to turn out attention to the other big idea, which is hypothesis testing. In its most abstract form, hypothesis testing really a very simple idea: the researcher has some theory about the world, and wants to determine whether or not the data actually support that theory. However, the details are messy, and most people find the theory of hypothesis testing to be the most frustrating part of statistics. The structure of the chapter is as follows. Firstly, I’ll describe how hypothesis testing works, in a fair amount of detail, using a simple running example to show you how a hypothesis test is “built”. I’ll try to avoid being too dogmatic while doing so, and focus instead on the underlying logic of the testing procedure.158 Afterwards, I’ll spend a bit of time talking about the various dogmas, rules and heresies that surround the theory of hypothesis testing.\n\n## 11.1 A menagerie of hypotheses\n\nEventually we all succumb to madness. For me, that day will arrive once I’m finally promoted to full professor. Safely ensconced in my ivory tower, happily protected by tenure, I will finally be able to take leave of my senses (so to speak), and indulge in that most thoroughly unproductive line of psychological research: the search for extrasensory perception (ESP).159\n\nLet’s suppose that this glorious day has come. My first study is a simple one, in which I seek to test whether clairvoyance exists. Each participant sits down at a table, and is shown a card by an experimenter. The card is black on one side and white on the other. The experimenter takes the card away, and places it on a table in an adjacent room. The card is placed black side up or white side up completely at random, with the randomisation occurring only after the experimenter has left the room with the participant. A second experimenter comes in and asks the participant which side of the card is now facing upwards. It’s purely a one-shot experiment. Each person sees only one card, and gives only one answer; and at no stage is the participant actually in contact with someone who knows the right answer. My data set, therefore, is very simple. I have asked the question of \\(N\\) people, and some number \\(X\\) of these people have given the correct response. To make things concrete, let’s suppose that I have tested \\(N = 100\\) people, and \\(X = 62\\) of these got the answer right… a surprisingly large number, sure, but is it large enough for me to feel safe in claiming I’ve found evidence for ESP? This is the situation where hypothesis testing comes in useful. However, before we talk about how to test hypotheses, we need to be clear about what we mean by hypotheses.\n\n### 11.1.1 Research hypotheses versus statistical hypotheses\n\nThe first distinction that you need to keep clear in your mind is between research hypotheses and statistical hypotheses. In my ESP study, my overall scientific goal is to demonstrate that clairvoyance exists. In this situation, I have a clear research goal: I am hoping to discover evidence for ESP. In other situations I might actually be a lot more neutral than that, so I might say that my research goal is to determine whether or not clairvoyance exists. Regardless of how I want to portray myself, the basic point that I’m trying to convey here is that a research hypothesis involves making a substantive, testable scientific claim… if you are a psychologist, then your research hypotheses are fundamentally about psychological constructs. Any of the following would count as research hypotheses:\n\n* Listening to music reduces your ability to pay attention to other things. This is a claim about the causal relationship between two psychologically meaningful concepts (listening to music and paying attention to things), so it’s a perfectly reasonable research hypothesis.\n * Intelligence is related to personality. Like the last one, this is a relational claim about two psychological constructs (intelligence and personality), but the claim is weaker: correlational not causal.\n * Intelligence is* speed of information processing. This hypothesis has a quite different character: it’s not actually a relational claim at all. It’s an ontological claim about the fundamental character of intelligence (and I’m pretty sure it’s wrong). It’s worth expanding on this one actually: It’s usually easier to think about how to construct experiments to test research hypotheses of the form “does X affect Y?” than it is to address claims like “what is X?” And in practice, what usually happens is that you find ways of testing relational claims that follow from your ontological ones. For instance, if I believe that intelligence is* speed of information processing in the brain, my experiments will often involve looking for relationships between measures of intelligence and measures of speed. As a consequence, most everyday research questions do tend to be relational in nature, but they’re almost always motivated by deeper ontological questions about the state of nature.\n\nNotice that in practice, my research hypotheses could overlap a lot. My ultimate goal in the ESP experiment might be to test an ontological claim like “ESP exists”, but I might operationally restrict myself to a narrower hypothesis like “Some people can `see’ objects in a clairvoyant fashion”. That said, there are some things that really don’t count as proper research hypotheses in any meaningful sense:\n\n* Love is a battlefield. This is too vague to be testable. While it’s okay for a research hypothesis to have a degree of vagueness to it, it has to be possible to operationalise your theoretical ideas. Maybe I’m just not creative enough to see it, but I can’t see how this can be converted into any concrete research design. If that’s true, then this isn’t a scientific research hypothesis, it’s a pop song. That doesn’t mean it’s not interesting – a lot of deep questions that humans have fall into this category. Maybe one day science will be able to construct testable theories of love, or to test to see if God exists, and so on; but right now we can’t, and I wouldn’t bet on ever seeing a satisfying scientific approach to either.\n * The first rule of tautology club is the first rule of tautology club. This is not a substantive claim of any kind. It’s true by definition. No conceivable state of nature could possibly be inconsistent with this claim. As such, we say that this is an unfalsifiable hypothesis, and as such it is outside the domain of science. Whatever else you do in science, your claims must have the possibility of being wrong.\n * More people in my experiment will say “yes” than “no”. This one fails as a research hypothesis because it’s a claim about the data set, not about the psychology (unless of course your actual research question is whether people have some kind of “yes” bias!). As we’ll see shortly, this hypothesis is starting to sound more like a statistical hypothesis than a research hypothesis.\n\nAs you can see, research hypotheses can be somewhat messy at times; and ultimately they are scientific claims. Statistical hypotheses are neither of these two things. Statistical hypotheses must be mathematically precise, and they must correspond to specific claims about the characteristics of the data generating mechanism (i.e., the “population”). Even so, the intent is that statistical hypotheses bear a clear relationship to the substantive research hypotheses that you care about! For instance, in my ESP study my research hypothesis is that some people are able to see through walls or whatever. What I want to do is to “map” this onto a statement about how the data were generated. So let’s think about what that statement would be. The quantity that I’m interested in within the experiment is \\(P(\\mbox{\"correct\"})\\), the true-but-unknown probability with which the participants in my experiment answer the question correctly. Let’s use the Greek letter \\(\\theta\\) (theta) to refer to this probability. Here are four different statistical hypotheses:\n\n* If ESP doesn’t exist and if my experiment is well designed, then my participants are just guessing. So I should expect them to get it right half of the time and so my statistical hypothesis is that the true probability of choosing correctly is \\(\\theta = 0.5\\).\n * Alternatively, suppose ESP does exist and participants can see the card. If that’s true, people will perform better than chance. The statistical hypotheis would be that \\(\\theta > 0.5\\).\n * A third possibility is that ESP does exist, but the colours are all reversed and people don’t realise it (okay, that’s wacky, but you never know…). If that’s how it works then you’d expect people’s performance to be below chance. This would correspond to a statistical hypothesis that \\(\\theta < 0.5\\).\n * Finally, suppose ESP exists, but I have no idea whether people are seeing the right colour or the wrong one. In that case, the only claim I could make about the data would be that the probability of making the correct answer is not equal to 50. This corresponds to the statistical hypothesis that \\(\\theta \\neq 0.5\\).\n\nAll of these are legitimate examples of a statistical hypothesis because they are statements about a population parameter and are meaningfully related to my experiment.\n\nWhat this discussion makes clear, I hope, is that when attempting to construct a statistical hypothesis test the researcher actually has two quite distinct hypotheses to consider. First, he or she has a research hypothesis (a claim about psychology), and this corresponds to a statistical hypothesis (a claim about the data generating population). In my ESP example, these might be\n\nDan.s.research.hypothesis | Dan.s.statistical.hypothesis |\n | --- | --- |\nESP.exists | \\(\\theta \\neq 0.5\\) |\n\nAnd the key thing to recognise is this: a statistical hypothesis test is a test of the statistical hypothesis, not the research hypothesis. If your study is badly designed, then the link between your research hypothesis and your statistical hypothesis is broken. To give a silly example, suppose that my ESP study was conducted in a situation where the participant can actually see the card reflected in a window; if that happens, I would be able to find very strong evidence that \\(\\theta \\neq 0.5\\), but this would tell us nothing about whether “ESP exists”.\n\n### 11.1.2 Null hypotheses and alternative hypotheses\n\nSo far, so good. I have a research hypothesis that corresponds to what I want to believe about the world, and I can map it onto a statistical hypothesis that corresponds to what I want to believe about how the data were generated. It’s at this point that things get somewhat counterintuitive for a lot of people. Because what I’m about to do is invent a new statistical hypothesis (the “null” hypothesis, \\(H_0\\)) that corresponds to the exact opposite of what I want to believe, and then focus exclusively on that, almost to the neglect of the thing I’m actually interested in (which is now called the “alternative” hypothesis, \\(H_1\\)). In our ESP example, the null hypothesis is that \\(\\theta = 0.5\\), since that’s what we’d expect if ESP didn’t exist. My hope, of course, is that ESP is totally real, and so the alternative to this null hypothesis is \\(\\theta \\neq 0.5\\). In essence, what we’re doing here is dividing up the possible values of \\(\\theta\\) into two groups: those values that I really hope aren’t true (the null), and those values that I’d be happy with if they turn out to be right (the alternative). Having done so, the important thing to recognise is that the goal of a hypothesis test is not to show that the alternative hypothesis is (probably) true; the goal is to show that the null hypothesis is (probably) false. Most people find this pretty weird.\n\nThe best way to think about it, in my experience, is to imagine that a hypothesis test is a criminal trial160… the trial of the null hypothesis. The null hypothesis is the defendant, the researcher is the prosecutor, and the statistical test itself is the judge. Just like a criminal trial, there is a presumption of innocence: the null hypothesis is deemed to be true unless you, the researcher, can prove beyond a reasonable doubt that it is false. You are free to design your experiment however you like (within reason, obviously!), and your goal when doing so is to maximise the chance that the data will yield a conviction… for the crime of being false. The catch is that the statistical test sets the rules of the trial, and those rules are designed to protect the null hypothesis – specifically to ensure that if the null hypothesis is actually true, the chances of a false conviction are guaranteed to be low. This is pretty important: after all, the null hypothesis doesn’t get a lawyer. And given that the researcher is trying desperately to prove it to be false, someone has to protect it.\n\n## 11.2 Two types of errors\n\nBefore going into details about how a statistical test is constructed, it’s useful to understand the philosophy behind it. I hinted at it when pointing out the similarity between a null hypothesis test and a criminal trial, but I should now be explicit. Ideally, we would like to construct our test so that we never make any errors. Unfortunately, since the world is messy, this is never possible. Sometimes you’re just really unlucky: for instance, suppose you flip a coin 10 times in a row and it comes up heads all 10 times. That feels like very strong evidence that the coin is biased (and it is!), but of course there’s a 1 in 1024 chance that this would happen even if the coin was totally fair. In other words, in real life we always have to accept that there’s a chance that we did the wrong thing. As a consequence, the goal behind statistical hypothesis testing is not to eliminate errors, but to minimise them.\n\nAt this point, we need to be a bit more precise about what we mean by “errors”. Firstly, let’s state the obvious: it is either the case that the null hypothesis is true, or it is false; and our test will either reject the null hypothesis or retain it.161 So, as the table below illustrates, after we run the test and make our choice, one of four things might have happened:\n\nretain \\(H_0\\) | retain \\(H_0\\) |\n | --- | --- |\n\\(H_0\\) is true | correct decision | error (type I) |\n\\(H_0\\) is false | error (type II) | correct decision |\n\nAs a consequence there are actually two different types of error here. If we reject a null hypothesis that is actually true, then we have made a type I error. On the other hand, if we retain the null hypothesis when it is in fact false, then we have made a type II error.\n\nRemember how I said that statistical testing was kind of like a criminal trial? Well, I meant it. A criminal trial requires that you establish “beyond a reasonable doubt” that the defendant did it. All of the evidentiary rules are (in theory, at least) designed to ensure that there’s (almost) no chance of wrongfully convicting an innocent defendant. The trial is designed to protect the rights of a defendant: as the English jurist famously said, it is “better that ten guilty persons escape than that one innocent suffer.” In other words, a criminal trial doesn’t treat the two types of error in the same way~… punishing the innocent is deemed to be much worse than letting the guilty go free. A statistical test is pretty much the same: the single most important design principle of the test is to control the probability of a type I error, to keep it below some fixed probability. This probability, which is denoted \\(\\alpha\\), is called the significance level of the test (or sometimes, the size of the test). And I’ll say it again, because it is so central to the whole set-up~… a hypothesis test is said to have significance level \\(\\alpha\\) if the type I error rate is no larger than \\(\\alpha\\).\n\nSo, what about the type II error rate? Well, we’d also like to keep those under control too, and we denote this probability by \\(\\beta\\). However, it’s much more common to refer to the power of the test, which is the probability with which we reject a null hypothesis when it really is false, which is \\(1-\\beta\\). To help keep this straight, here’s the same table again, but with the relevant numbers added:\n\nretain \\(H_0\\) | reject \\(H_0\\) |\n | --- | --- |\n\\(H_0\\) is true | \\(1-\\alpha\\) (probability of correct retention) | \\(\\alpha\\) (type I error rate) |\n\\(H_0\\) is false | \\(\\beta\\) (type II error rate) | \\(1-\\beta\\) (power of the test) |\n\nA “powerful” hypothesis test is one that has a small value of \\(\\beta\\), while still keeping \\(\\alpha\\) fixed at some (small) desired level. By convention, scientists make use of three different \\(\\alpha\\) levels: \\(.05\\), \\(.01\\) and \\(.001\\). Notice the asymmetry here~… the tests are designed to ensure that the \\(\\alpha\\) level is kept small, but there’s no corresponding guarantee regarding \\(\\beta\\). We’d certainly like the type II error rate to be small, and we try to design tests that keep it small, but this is very much secondary to the overwhelming need to control the type I error rate. As Blackstone might have said if he were a statistician, it is “better to retain 10 false null hypotheses than to reject a single true one”. To be honest, I don’t know that I agree with this philosophy – there are situations where I think it makes sense, and situations where I think it doesn’t – but that’s neither here nor there. It’s how the tests are built.\n\n## 11.3 Test statistics and sampling distributions\n\nAt this point we need to start talking specifics about how a hypothesis test is constructed. To that end, let’s return to the ESP example. Let’s ignore the actual data that we obtained, for the moment, and think about the structure of the experiment. Regardless of what the actual numbers are, the form of the data is that \\(X\\) out of \\(N\\) people correctly identified the colour of the hidden card. Moreover, let’s suppose for the moment that the null hypothesis really is true: ESP doesn’t exist, and the true probability that anyone picks the correct colour is exactly \\(\\theta = 0.5\\). What would we expect the data to look like? Well, obviously, we’d expect the proportion of people who make the correct response to be pretty close to 50%. Or, to phrase this in more mathematical terms, we’d say that \\(X/N\\) is approximately \\(0.5\\). Of course, we wouldn’t expect this fraction to be exactly 0.5: if, for example we tested \\(N=100\\) people, and \\(X = 53\\) of them got the question right, we’d probably be forced to concede that the data are quite consistent with the null hypothesis. On the other hand, if \\(X = 99\\) of our participants got the question right, then we’d feel pretty confident that the null hypothesis is wrong. Similarly, if only \\(X=3\\) people got the answer right, we’d be similarly confident that the null was wrong. Let’s be a little more technical about this: we have a quantity \\(X\\) that we can calculate by looking at our data; after looking at the value of \\(X\\), we make a decision about whether to believe that the null hypothesis is correct, or to reject the null hypothesis in favour of the alternative. The name for this thing that we calculate to guide our choices is a test statistic.\n\nHaving chosen a test statistic, the next step is to state precisely which values of the test statistic would cause is to reject the null hypothesis, and which values would cause us to keep it. In order to do so, we need to determine what the sampling distribution of the test statistic would be if the null hypothesis were actually true (we talked about sampling distributions earlier in Section 10.3.1). Why do we need this? Because this distribution tells us exactly what values of \\(X\\) our null hypothesis would lead us to expect. And therefore, we can use this distribution as a tool for assessing how closely the null hypothesis agrees with our data.\n\nHow do we actually determine the sampling distribution of the test statistic? For a lot of hypothesis tests this step is actually quite complicated, and later on in the book you’ll see me being slightly evasive about it for some of the tests (some of them I don’t even understand myself). However, sometimes it’s very easy. And, fortunately for us, our ESP example provides us with one of the easiest cases. Our population parameter \\(\\theta\\) is just the overall probability that people respond correctly when asked the question, and our test statistic \\(X\\) is the count of the number of people who did so, out of a sample size of \\(N\\). We’ve seen a distribution like this before, in Section 9.4: that’s exactly what the binomial distribution describes! So, to use the notation and terminology that I introduced in that section, we would say that the null hypothesis predicts that \\(X\\) is binomially distributed, which is written \\[ X \\sim \\mbox{Binomial}(\\theta,N) \\] Since the null hypothesis states that \\(\\theta = 0.5\\) and our experiment has \\(N=100\\) people, we have the sampling distribution we need. This sampling distribution is plotted in Figure 11.1. No surprises really: the null hypothesis says that \\(X=50\\) is the most likely outcome, and it says that we’re almost certain to see somewhere between 40 and 60 correct responses.\n\n## 11.4 Making decisions\n\nOkay, we’re very close to being finished. We’ve constructed a test statistic (\\(X\\)), and we chose this test statistic in such a way that we’re pretty confident that if \\(X\\) is close to \\(N/2\\) then we should retain the null, and if not we should reject it. The question that remains is this: exactly which values of the test statistic should we associate with the null hypothesis, and which exactly values go with the alternative hypothesis? In my ESP study, for example, I’ve observed a value of \\(X=62\\). What decision should I make? Should I choose to believe the null hypothesis, or the alternative hypothesis?\n\n### 11.4.1 Critical regions and critical values\n\nTo answer this question, we need to introduce the concept of a critical region for the test statistic \\(X\\). The critical region of the test corresponds to those values of \\(X\\) that would lead us to reject null hypothesis (which is why the critical region is also sometimes called the rejection region). How do we find this critical region? Well, let’s consider what we know:\n\n* \\(X\\) should be very big or very small in order to reject the null hypothesis.\n * If the null hypothesis is true, the sampling distribution of \\(X\\) is Binomial\\((0.5, N)\\).\n * If \\(\\alpha =.05\\), the critical region must cover 5% of this sampling distribution.\n\nIt’s important to make sure you understand this last point: the critical region corresponds to those values of \\(X\\) for which we would reject the null hypothesis, and the sampling distribution in question describes the probability that we would obtain a particular value of \\(X\\) if the null hypothesis were actually true. Now, let’s suppose that we chose a critical region that covers 20% of the sampling distribution, and suppose that the null hypothesis is actually true. What would be the probability of incorrectly rejecting the null? The answer is of course 20%. And therefore, we would have built a test that had an \\(\\alpha\\) level of \\(0.2\\). If we want \\(\\alpha = .05\\), the critical region is only allowed to cover 5% of the sampling distribution of our test statistic.\n\nAs it turns out, those three things uniquely solve the problem: our critical region consists of the most extreme values, known as the tails of the distribution. This is illustrated in Figure 11.2. As it turns out, if we want \\(\\alpha = .05\\), then our critical regions correspond to \\(X \\leq 40\\) and \\(X \\geq 60\\).162 That is, if the number of people saying “true” is between 41 and 59, then we should retain the null hypothesis. If the number is between 0 to 40 or between 60 to 100, then we should reject the null hypothesis. The numbers 40 and 60 are often referred to as the critical values, since they define the edges of the critical region.\n\nAt this point, our hypothesis test is essentially complete: (1) we choose an \\(\\alpha\\) level (e.g., \\(\\alpha = .05\\), (2) come up with some test statistic (e.g., \\(X\\)) that does a good job (in some meaningful sense) of comparing \\(H_0\\) to \\(H_1\\), (3) figure out the sampling distribution of the test statistic on the assumption that the null hypothesis is true (in this case, binomial) and then (4) calculate the critical region that produces an appropriate \\(\\alpha\\) level (0-40 and 60-100). All that we have to do now is calculate the value of the test statistic for the real data (e.g., \\(X = 62\\)) and then compare it to the critical values to make our decision. Since 62 is greater than the critical value of 60, we would reject the null hypothesis. Or, to phrase it slightly differently, we say that the test has produced a significant result.\n\n### 11.4.2 A note on statistical “significance”\n\nLike other occult techniques of divination, the statistical method has a private jargon deliberately contrived to obscure its methods from non-practitioners.\n\n– Attributed to 163\n\nA very brief digression is in order at this point, regarding the word “significant”. The concept of statistical significance is actually a very simple one, but has a very unfortunate name. If the data allow us to reject the null hypothesis, we say that “the result is statistically significant”, which is often shortened to “the result is significant”. This terminology is rather old, and dates back to a time when “significant” just meant something like “indicated”, rather than its modern meaning, which is much closer to “important”. As a result, a lot of modern readers get very confused when they start learning statistics, because they think that a “significant result” must be an important one. It doesn’t mean that at all. All that “statistically significant” means is that the data allowed us to reject a null hypothesis. Whether or not the result is actually important in the real world is a very different question, and depends on all sorts of other things.\n\n### 11.4.3 The difference between one sided and two sided tests\n\nThere’s one more thing I want to point out about the hypothesis test that I’ve just constructed. If we take a moment to think about the statistical hypotheses I’ve been using, \\[ \\begin{array}{cc} H_0 : & \\theta = .5 \\\\ H_1 : & \\theta \\neq .5 \\end{array} \\] we notice that the alternative hypothesis covers both the possibility that \\(\\theta < .5\\) and the possibility that \\(\\theta > .5\\). This makes sense if I really think that ESP could produce better-than-chance performance or worse-than-chance performance (and there are some people who think that). In statistical language, this is an example of a two-sided test. It’s called this because the alternative hypothesis covers the area on both “sides” of the null hypothesis, and as a consequence the critical region of the test covers both tails of the sampling distribution (2.5% on either side if \\(\\alpha =.05\\)), as illustrated earlier in Figure 11.2.\n\nHowever, that’s not the only possibility. It might be the case, for example, that I’m only willing to believe in ESP if it produces better than chance performance. If so, then my alternative hypothesis would only covers the possibility that \\(\\theta > .5\\), and as a consequence the null hypothesis now becomes \\(\\theta \\leq .5\\): \\[ \\begin{array}{cc} H_0 : & \\theta \\leq .5 \\\\ H_1 : & \\theta > .5 \\end{array} \\] When this happens, we have what’s called a one-sided test, and when this happens the critical region only covers one tail of the sampling distribution. This is illustrated in Figure 11.3.\n\n## 11.5 The \\(p\\) value of a test\n\nIn one sense, our hypothesis test is complete; we’ve constructed a test statistic, figured out its sampling distribution if the null hypothesis is true, and then constructed the critical region for the test. Nevertheless, I’ve actually omitted the most important number of all: the \\(p\\) value. It is to this topic that we now turn. There are two somewhat different ways of interpreting a \\(p\\) value, one proposed by and the other by . Both versions are legitimate, though they reflect very different ways of thinking about hypothesis tests. Most introductory textbooks tend to give Fisher’s version only, but I think that’s a bit of a shame. To my mind, Neyman’s version is cleaner, and actually better reflects the logic of the null hypothesis test. You might disagree though, so I’ve included both. I’ll start with Neyman’s version…\n\n### 11.5.1 A softer view of decision making\n\nOne problem with the hypothesis testing procedure that I’ve described is that it makes no distinction at all between a result this “barely significant” and those that are “highly significant”. For instance, in my ESP study the data I obtained only just fell inside the critical region - so I did get a significant effect, but was a pretty near thing. In contrast, suppose that I’d run a study in which \\(X=97\\) out of my \\(N=100\\) participants got the answer right. This would obviously be significant too, but my a much larger margin; there’s really no ambiguity about this at all. The procedure that I described makes no distinction between the two. If I adopt the standard convention of allowing \\(\\alpha = .05\\) as my acceptable Type I error rate, then both of these are significant results.\n\nThis is where the \\(p\\) value comes in handy. To understand how it works, let’s suppose that we ran lots of hypothesis tests on the same data set: but with a different value of \\(\\alpha\\) in each case. When we do that for my original ESP data, what we’d get is something like this\n\nValue of \\(\\alpha\\) | Reject the null? |\n | --- | --- |\n0.05 | Yes |\n0.04 | Yes |\n0.03 | Yes |\n0.02 | No |\n0.01 | No |\n\nWhen we test ESP data (\\(X=62\\) successes out of \\(N=100\\) observations) using \\(\\alpha\\) levels of .03 and above, we’d always find ourselves rejecting the null hypothesis. For \\(\\alpha\\) levels of .02 and below, we always end up retaining the null hypothesis. Therefore, somewhere between .02 and .03 there must be a smallest value of \\(\\alpha\\) that would allow us to reject the null hypothesis for this data. This is the \\(p\\) value; as it turns out the ESP data has \\(p = .021\\). In short:\n \\(p\\) is defined to be the smallest Type I error rate (\\(\\alpha\\)) that you have to be willing to tolerate if you want to reject the null hypothesis.\n\nIf it turns out that \\(p\\) describes an error rate that you find intolerable, then you must retain the null. If you’re comfortable with an error rate equal to \\(p\\), then it’s okay to reject the null hypothesis in favour of your preferred alternative.\n\nIn effect, \\(p\\) is a summary of all the possible hypothesis tests that you could have run, taken across all possible \\(\\alpha\\) values. And as a consequence it has the effect of “softening” our decision process. For those tests in which \\(p \\leq \\alpha\\) you would have rejected the null hypothesis, whereas for those tests in which \\(p > \\alpha\\) you would have retained the null. In my ESP study I obtained \\(X=62\\), and as a consequence I’ve ended up with \\(p = .021\\). So the error rate I have to tolerate is 2.1%. In contrast, suppose my experiment had yielded \\(X=97\\). What happens to my \\(p\\) value now? This time it’s shrunk to \\(p = 1.36 \\times 10^{-25}\\), which is a tiny, tiny164 Type I error rate. For this second case I would be able to reject the null hypothesis with a lot more confidence, because I only have to be “willing” to tolerate a type I error rate of about 1 in 10 trillion trillion in order to justify my decision to reject.\n\n### 11.5.2 The probability of extreme data\n\nThe second definition of the \\(p\\)-value comes from , and it’s actually this one that you tend to see in most introductory statistics textbooks. Notice how, when I constructed the critical region, it corresponded to the tails (i.e., extreme values) of the sampling distribution? That’s not a coincidence: almost all “good” tests have this characteristic (good in the sense of minimising our type II error rate, \\(\\beta\\)). The reason for that is that a good critical region almost always corresponds to those values of the test statistic that are least likely to be observed if the null hypothesis is true. If this rule is true, then we can define the \\(p\\)-value as the probability that we would have observed a test statistic that is at least as extreme as the one we actually did get. In other words, if the data are extremely implausible according to the null hypothesis, then the null hypothesis is probably wrong.\n\n### 11.5.3 A common mistake\n\nOkay, so you can see that there are two rather different but legitimate ways to interpret the \\(p\\) value, one based on Neyman’s approach to hypothesis testing and the other based on Fisher’s. Unfortunately, there is a third explanation that people sometimes give, especially when they’re first learning statistics, and it is absolutely and completely wrong. This mistaken approach is to refer to the \\(p\\) value as “the probability that the null hypothesis is true”. It’s an intuitively appealing way to think, but it’s wrong in two key respects: (1) null hypothesis testing is a frequentist tool, and the frequentist approach to probability does not allow you to assign probabilities to the null hypothesis… according to this view of probability, the null hypothesis is either true or it is not; it cannot have a “5% chance” of being true. (2) even within the Bayesian approach, which does let you assign probabilities to hypotheses, the \\(p\\) value would not correspond to the probability that the null is true; this interpretation is entirely inconsistent with the mathematics of how the \\(p\\) value is calculated. Put bluntly, despite the intuitive appeal of thinking this way, there is no justification for interpreting a \\(p\\) value this way. Never do it.\n\n## 11.6 Reporting the results of a hypothesis test\n\nWhen writing up the results of a hypothesis test, there’s usually several pieces of information that you need to report, but it varies a fair bit from test to test. Throughout the rest of the book I’ll spend a little time talking about how to report the results of different tests (see Section 12.1.9 for a particularly detailed example), so that you can get a feel for how it’s usually done. However, regardless of what test you’re doing, the one thing that you always have to do is say something about the \\(p\\) value, and whether or not the outcome was significant.\n\nThe fact that you have to do this is unsurprising; it’s the whole point of doing the test. What might be surprising is the fact that there is some contention over exactly how you’re supposed to do it. Leaving aside those people who completely disagree with the entire framework underpinning null hypothesis testing, there’s a certain amount of tension that exists regarding whether or not to report the exact \\(p\\) value that you obtained, or if you should state only that \\(p < \\alpha\\) for a significance level that you chose in advance (e.g., \\(p<.05\\)).\n\n### 11.6.1 The issue\n\nTo see why this is an issue, the key thing to recognise is that \\(p\\) values are terribly convenient. In practice, the fact that we can compute a \\(p\\) value means that we don’t actually have to specify any \\(\\alpha\\) level at all in order to run the test. Instead, what you can do is calculate your \\(p\\) value and interpret it directly: if you get \\(p = .062\\), then it means that you’d have to be willing to tolerate a Type I error rate of 6.2% to justify rejecting the null. If you personally find 6.2% intolerable, then you retain the null. Therefore, the argument goes, why don’t we just report the actual \\(p\\) value and let the reader make up their own minds about what an acceptable Type I error rate is? This approach has the big advantage of “softening” the decision making process – in fact, if you accept the Neyman definition of the \\(p\\) value, that’s the whole point of the \\(p\\) value. We no longer have a fixed significance level of \\(\\alpha = .05\\) as a bright line separating “accept” from “reject” decisions; and this removes the rather pathological problem of being forced to treat \\(p = .051\\) in a fundamentally different way to \\(p = .049\\).\n\nThis flexibility is both the advantage and the disadvantage to the \\(p\\) value. The reason why a lot of people don’t like the idea of reporting an exact \\(p\\) value is that it gives the researcher a bit too much freedom. In particular, it lets you change your mind about what error tolerance you’re willing to put up with after you look at the data. For instance, consider my ESP experiment. Suppose I ran my test, and ended up with a \\(p\\) value of .09. Should I accept or reject? Now, to be honest, I haven’t yet bothered to think about what level of Type I error I’m “really” willing to accept. I don’t have an opinion on that topic. But I do have an opinion about whether or not ESP exists, and I definitely have an opinion about whether my research should be published in a reputable scientific journal. And amazingly, now that I’ve looked at the data I’m starting to think that a 9% error rate isn’t so bad, especially when compared to how annoying it would be to have to admit to the world that my experiment has failed. So, to avoid looking like I just made it up after the fact, I now say that my \\(\\alpha\\) is .1: a 10% type I error rate isn’t too bad, and at that level my test is significant! I win.\n\nIn other words, the worry here is that I might have the best of intentions, and be the most honest of people, but the temptation to just “shade” things a little bit here and there is really, really strong. As anyone who has ever run an experiment can attest, it’s a long and difficult process, and you often get very attached to your hypotheses. It’s hard to let go and admit the experiment didn’t find what you wanted it to find. And that’s the danger here. If we use the “raw” \\(p\\)-value, people will start interpreting the data in terms of what they want to believe, not what the data are actually saying… and if we allow that, well, why are we bothering to do science at all? Why not let everyone believe whatever they like about anything, regardless of what the facts are? Okay, that’s a bit extreme, but that’s where the worry comes from. According to this view, you really must specify your \\(\\alpha\\) value in advance, and then only report whether the test was significant or not. It’s the only way to keep ourselves honest.\n\n### 11.6.2 Two proposed solutions\n\nIn practice, it’s pretty rare for a researcher to specify a single \\(\\alpha\\) level ahead of time. Instead, the convention is that scientists rely on three standard significance levels: .05, .01 and .001. When reporting your results, you indicate which (if any) of these significance levels allow you to reject the null hypothesis. This is summarised in Table 11.1. This allows us to soften the decision rule a little bit, since \\(p<.01\\) implies that the data meet a stronger evidentiary standard than \\(p<.05\\) would. Nevertheless, since these levels are fixed in advance by convention, it does prevent people choosing their \\(\\alpha\\) level after looking at the data.\n\nUsual notation | Signif. stars | Signif. stars | The null is… |\n | --- | --- | --- | --- |\n\\(p>.05\\) | NA | The test wasn’t significant | Retained |\n\\(p<.05\\) | * | The test was significant at $= .05 but not at \\(\\alpha =.01\\) or \\(\\alpha = .001\\).$ | Rejected |\n\\(p<.01\\) | ** | The test was significant at \\(\\alpha = .05\\) and \\(\\alpha = .01\\) but not at $= .001 | Rejected |\n\\(p<.001\\) | *** | The test was significant at all levels | Rejected |\n\nNevertheless, quite a lot of people still prefer to report exact \\(p\\) values. To many people, the advantage of allowing the reader to make up their own mind about how to interpret \\(p = .06\\) outweighs any disadvantages. In practice, however, even among those researchers who prefer exact \\(p\\) values it is quite common to just write \\(p<.001\\) instead of reporting an exact value for small \\(p\\). This is in part because a lot of software doesn’t actually print out the \\(p\\) value when it’s that small (e.g., SPSS just writes \\(p = .000\\) whenever \\(p<.001\\)), and in part because a very small \\(p\\) value can be kind of misleading. The human mind sees a number like .0000000001 and it’s hard to suppress the gut feeling that the evidence in favour of the alternative hypothesis is a near certainty. In practice however, this is usually wrong. Life is a big, messy, complicated thing: and every statistical test ever invented relies on simplifications, approximations and assumptions. As a consequence, it’s probably not reasonable to walk away from any statistical analysis with a feeling of confidence stronger than \\(p<.001\\) implies. In other words, \\(p<.001\\) is really code for “as far as this test is concerned, the evidence is overwhelming.”\n\nIn light of all this, you might be wondering exactly what you should do. There’s a fair bit of contradictory advice on the topic, with some people arguing that you should report the exact \\(p\\) value, and other people arguing that you should use the tiered approach illustrated in Table 11.1. As a result, the best advice I can give is to suggest that you look at papers/reports written in your field and see what the convention seems to be. If there doesn’t seem to be any consistent pattern, then use whichever method you prefer.\n\n## 11.7 Running the hypothesis test in practice\n\n At this point some of you might be wondering if this is a “real” hypothesis test, or just a toy example that I made up. It’s real. In the previous discussion I built the test from first principles, thinking that it was the simplest possible problem that you might ever encounter in real life. However, this test already exists: it’s called the binomial test, and it’s implemented by an R function called `binom.test()` . To test the null hypothesis that the response probability is one-half `p = .5` ,165 using data in which `x = 62` of `n = 100` people made the correct response, here’s how to do it in R: \n\n```\nbinom.test( x=62, n=100, p=.5 )\n```\n\n```\n## \n## Exact binomial test\n## \n## data: 62 and 100\n## number of successes = 62, number of trials = 100, p-value =\n## 0.02098\n## alternative hypothesis: true probability of success is not equal to 0.5\n## 95 percent confidence interval:\n## 0.5174607 0.7152325\n## sample estimates:\n## probability of success \n## 0.62\n```\n\nRight now, this output looks pretty unfamiliar to you, but you can see that it’s telling you more or less the right things. Specifically, the \\(p\\)-value of 0.02 is less than the usual choice of \\(\\alpha = .05\\), so you can reject the null. We’ll talk a lot more about how to read this sort of output as we go along; and after a while you’ll hopefully find it quite easy to read and understand. For now, however, I just wanted to make the point that R contains a whole lot of functions corresponding to different kinds of hypothesis test. And while I’ll usually spend quite a lot of time explaining the logic behind how the tests are built, every time I discuss a hypothesis test the discussion will end with me showing you a fairly simple R command that you can use to run the test in practice.\n\n## 11.8 Effect size, sample size and power\n\nIn previous sections I’ve emphasised the fact that the major design principle behind statistical hypothesis testing is that we try to control our Type I error rate. When we fix \\(\\alpha = .05\\) we are attempting to ensure that only 5% of true null hypotheses are incorrectly rejected. However, this doesn’t mean that we don’t care about Type II errors. In fact, from the researcher’s perspective, the error of failing to reject the null when it is actually false is an extremely annoying one. With that in mind, a secondary goal of hypothesis testing is to try to minimise \\(\\beta\\), the Type II error rate, although we don’t usually talk in terms of minimising Type II errors. Instead, we talk about maximising the power of the test. Since power is defined as \\(1-\\beta\\), this is the same thing.\n\n### 11.8.1 The power function\n\nLet’s take a moment to think about what a Type II error actually is. A Type II error occurs when the alternative hypothesis is true, but we are nevertheless unable to reject the null hypothesis. Ideally, we’d be able to calculate a single number \\(\\beta\\) that tells us the Type II error rate, in the same way that we can set \\(\\alpha = .05\\) for the Type I error rate. Unfortunately, this is a lot trickier to do. To see this, notice that in my ESP study the alternative hypothesis actually corresponds to lots of possible values of \\(\\theta\\). In fact, the alternative hypothesis corresponds to every value of \\(\\theta\\) except 0.5. Let’s suppose that the true probability of someone choosing the correct response is 55% (i.e., \\(\\theta = .55\\)). If so, then the true sampling distribution for \\(X\\) is not the same one that the null hypothesis predicts: the most likely value for \\(X\\) is now 55 out of 100. Not only that, the whole sampling distribution has now shifted, as shown in Figure 11.4. The critical regions, of course, do not change: by definition, the critical regions are based on what the null hypothesis predicts. What we’re seeing in this figure is the fact that when the null hypothesis is wrong, a much larger proportion of the sampling distribution distribution falls in the critical region. And of course that’s what should happen: the probability of rejecting the null hypothesis is larger when the null hypothesis is actually false! However \\(\\theta = .55\\) is not the only possibility consistent with the alternative hypothesis. Let’s instead suppose that the true value of \\(\\theta\\) is actually 0.7. What happens to the sampling distribution when this occurs? The answer, shown in Figure 11.5, is that almost the entirety of the sampling distribution has now moved into the critical region. Therefore, if \\(\\theta = 0.7\\) the probability of us correctly rejecting the null hypothesis (i.e., the power of the test) is much larger than if \\(\\theta = 0.55\\). In short, while \\(\\theta = .55\\) and \\(\\theta = .70\\) are both part of the alternative hypothesis, the Type II error rate is different.\n\nWhat all this means is that the power of a test (i.e., \\(1-\\beta\\)) depends on the true value of \\(\\theta\\). To illustrate this, I’ve calculated the expected probability of rejecting the null hypothesis for all values of \\(\\theta\\), and plotted it in Figure 11.6. This plot describes what is usually called the power function of the test. It’s a nice summary of how good the test is, because it actually tells you the power (\\(1-\\beta\\)) for all possible values of \\(\\theta\\). As you can see, when the true value of \\(\\theta\\) is very close to 0.5, the power of the test drops very sharply, but when it is further away, the power is large.\n\n### 11.8.2 Effect size\n\nSince all models are wrong the scientist must be alert to what is importantly wrong. It is inappropriate to be concerned with mice when there are tigers abroad\n\n– George Box 1976\n\nThe plot shown in Figure 11.6 captures a fairly basic point about hypothesis testing. If the true state of the world is very different from what the null hypothesis predicts, then your power will be very high; but if the true state of the world is similar to the null (but not identical) then the power of the test is going to be very low. Therefore, it’s useful to be able to have some way of quantifying how “similar” the true state of the world is to the null hypothesis. A statistic that does this is called a measure of effect size (e.g. Cohen 1988; Ellis 2010). Effect size is defined slightly differently in different contexts,166 (and so this section just talks in general terms) but the qualitative idea that it tries to capture is always the same: how big is the difference between the true population parameters, and the parameter values that are assumed by the null hypothesis? In our ESP example, if we let \\(\\theta_0 = 0.5\\) denote the value assumed by the null hypothesis, and let \\(\\theta\\) denote the true value, then a simple measure of effect size could be something like the difference between the true value and null (i.e., \\(\\theta - \\theta_0\\)), or possibly just the magnitude of this difference, \\(\\mbox{abs}(\\theta - \\theta_0)\\).\n\nbig effect size | small effect size |\n | --- | --- |\nsignificant result | difference is real, and of practical importance | difference is real, but might not be interesting |\nnon-significant result | no effect observed | no effect observed |\n\nWhy calculate effect size? Let’s assume that you’ve run your experiment, collected the data, and gotten a significant effect when you ran your hypothesis test. Isn’t it enough just to say that you’ve gotten a significant effect? Surely that’s the point of hypothesis testing? Well, sort of. Yes, the point of doing a hypothesis test is to try to demonstrate that the null hypothesis is wrong, but that’s hardly the only thing we’re interested in. If the null hypothesis claimed that \\(\\theta = .5\\), and we show that it’s wrong, we’ve only really told half of the story. Rejecting the null hypothesis implies that we believe that \\(\\theta \\neq .5\\), but there’s a big difference between \\(\\theta = .51\\) and \\(\\theta = .8\\). If we find that \\(\\theta = .8\\), then not only have we found that the null hypothesis is wrong, it appears to be very wrong. On the other hand, suppose we’ve successfully rejected the null hypothesis, but it looks like the true value of \\(\\theta\\) is only .51 (this would only be possible with a large study). Sure, the null hypothesis is wrong, but it’s not at all clear that we actually care, because the effect size is so small. In the context of my ESP study we might still care, since any demonstration of real psychic powers would actually be pretty cool167, but in other contexts a 1% difference isn’t very interesting, even if it is a real difference. For instance, suppose we’re looking at differences in high school exam scores between males and females, and it turns out that the female scores are 1% higher on average than the males. If I’ve got data from thousands of students, then this difference will almost certainly be statistically significant, but regardless of how small the \\(p\\) value is it’s just not very interesting. You’d hardly want to go around proclaiming a crisis in boys education on the basis of such a tiny difference would you? It’s for this reason that it is becoming more standard (slowly, but surely) to report some kind of standard measure of effect size along with the the results of the hypothesis test. The hypothesis test itself tells you whether you should believe that the effect you have observed is real (i.e., not just due to chance); the effect size tells you whether or not you should care.\n\n### 11.8.3 Increasing the power of your study\n\nNot surprisingly, scientists are fairly obsessed with maximising the power of their experiments. We want our experiments to work, and so we want to maximise the chance of rejecting the null hypothesis if it is false (and of course we usually want to believe that it is false!) As we’ve seen, one factor that influences power is the effect size. So the first thing you can do to increase your power is to increase the effect size. In practice, what this means is that you want to design your study in such a way that the effect size gets magnified. For instance, in my ESP study I might believe that psychic powers work best in a quiet, darkened room; with fewer distractions to cloud the mind. Therefore I would try to conduct my experiments in just such an environment: if I can strengthen people’s ESP abilities somehow, then the true value of \\(\\theta\\) will go up168 and therefore my effect size will be larger. In short, clever experimental design is one way to boost power; because it can alter the effect size.\n\nUnfortunately, it’s often the case that even with the best of experimental designs you may have only a small effect. Perhaps, for example, ESP really does exist, but even under the best of conditions it’s very very weak. Under those circumstances, your best bet for increasing power is to increase the sample size. In general, the more observations that you have available, the more likely it is that you can discriminate between two hypotheses. If I ran my ESP experiment with 10 participants, and 7 of them correctly guessed the colour of the hidden card, you wouldn’t be terribly impressed. But if I ran it with 10,000 participants and 7,000 of them got the answer right, you would be much more likely to think I had discovered something. In other words, power increases with the sample size. This is illustrated in Figure 11.7, which shows the power of the test for a true parameter of \\(\\theta = 0.7\\), for all sample sizes \\(N\\) from 1 to 100, where I’m assuming that the null hypothesis predicts that \\(\\theta_0 = 0.5\\).\n\n```\n## [1] 0.00000000 0.00000000 0.00000000 0.00000000 0.00000000 0.11837800\n## [7] 0.08257300 0.05771362 0.19643626 0.14945203 0.11303734 0.25302172\n## [13] 0.20255096 0.16086106 0.29695959 0.24588947 0.38879291 0.33269435\n## [19] 0.28223844 0.41641377 0.36272868 0.31341925 0.43996501 0.38859619\n## [25] 0.51186665 0.46049782 0.41129777 0.52752694 0.47870819 0.58881596\n## [31] 0.54162450 0.49507894 0.59933871 0.55446069 0.65155826 0.60907715\n## [37] 0.69828554 0.65867614 0.61815357 0.70325017 0.66542910 0.74296156\n## [43] 0.70807163 0.77808343 0.74621569 0.71275488 0.78009449 0.74946571\n## [49] 0.81000236 0.78219322 0.83626633 0.81119597 0.78435605 0.83676444\n## [55] 0.81250680 0.85920268 0.83741123 0.87881491 0.85934395 0.83818214\n## [61] 0.87858194 0.85962510 0.89539581 0.87849413 0.91004390 0.89503851\n## [67] 0.92276845 0.90949768 0.89480727 0.92209753 0.90907263 0.93304809\n## [73] 0.92153987 0.94254237 0.93240638 0.92108426 0.94185449 0.93185881\n## [79] 0.95005094 0.94125189 0.95714694 0.94942195 0.96327866 0.95651332\n## [85] 0.94886329 0.96265653 0.95594208 0.96796884 0.96208909 0.97255504\n## [91] 0.96741721 0.97650832 0.97202770 0.97991117 0.97601093 0.97153910\n## [97] 0.97944717 0.97554675 0.98240749 0.97901142\n```\n\nBecause power is important, whenever you’re contemplating running an experiment it would be pretty useful to know how much power you’re likely to have. It’s never possible to know for sure, since you can’t possibly know what your effect size is. However, it’s often (well, sometimes) possible to guess how big it should be. If so, you can guess what sample size you need! This idea is called power analysis, and if it’s feasible to do it, then it’s very helpful, since it can tell you something about whether you have enough time or money to be able to run the experiment successfully. It’s increasingly common to see people arguing that power analysis should be a required part of experimental design, so it’s worth knowing about. I don’t discuss power analysis in this book, however. This is partly for a boring reason and partly for a substantive one. The boring reason is that I haven’t had time to write about power analysis yet. The substantive one is that I’m still a little suspicious of power analysis. Speaking as a researcher, I have very rarely found myself in a position to be able to do one – it’s either the case that (a) my experiment is a bit non-standard and I don’t know how to define effect size properly, (b) I literally have so little idea about what the effect size will be that I wouldn’t know how to interpret the answers. Not only that, after extensive conversations with someone who does stats consulting for a living (my wife, as it happens), I can’t help but notice that in practice the only time anyone ever asks her for a power analysis is when she’s helping someone write a grant application. In other words, the only time any scientist ever seems to want a power analysis in real life is when they’re being forced to do it by bureaucratic process. It’s not part of anyone’s day to day work. In short, I’ve always been of the view that while power is an important concept, power analysis is not as useful as people make it sound, except in the rare cases where (a) someone has figured out how to calculate power for your actual experimental design and (b) you have a pretty good idea what the effect size is likely to be. Maybe other people have had better experiences than me, but I’ve personally never been in a situation where both (a) and (b) were true. Maybe I’ll be convinced otherwise in the future, and probably a future version of this book would include a more detailed discussion of power analysis, but for now this is about as much as I’m comfortable saying about the topic.\n\n## 11.9 Some issues to consider\n\nWhat I’ve described to you in this chapter is the orthodox framework for null hypothesis significance testing (NHST). Understanding how NHST works is an absolute necessity, since it has been the dominant approach to inferential statistics ever since it came to prominence in the early 20th century. It’s what the vast majority of working scientists rely on for their data analysis, so even if you hate it you need to know it. However, the approach is not without problems. There are a number of quirks in the framework, historical oddities in how it came to be, theoretical disputes over whether or not the framework is right, and a lot of practical traps for the unwary. I’m not going to go into a lot of detail on this topic, but I think it’s worth briefly discussing a few of these issues.\n\n### 11.9.1 Neyman versus Fisher\n\nThe first thing you should be aware of is that orthodox NHST is actually a mash-up of two rather different approaches to hypothesis testing, one proposed by and the other proposed by (for a historical summary see Lehmann 2011). The history is messy because Fisher and Neyman were real people whose opinions changed over time, and at no point did either of them offer “the definitive statement” of how we should interpret their work many decades later. That said, here’s a quick summary of what I take these two approaches to be.\n\nFirst, let’s talk about Fisher’s approach. As far as I can tell, Fisher assumed that you only had the one hypothesis (the null), and what you want to do is find out if the null hypothesis is inconsistent with the data. From his perspective, what you should do is check to see if the data are “sufficiently unlikely” according to the null. In fact, if you remember back to our earlier discussion, that’s how Fisher defines the \\(p\\)-value. According to Fisher, if the null hypothesis provided a very poor account of the data, you could safely reject it. But, since you don’t have any other hypotheses to compare it to, there’s no way of “accepting the alternative” because you don’t necessarily have an explicitly stated alternative. That’s more or less all that there was to it.\n\nIn contrast, Neyman thought that the point of hypothesis testing was as a guide to action, and his approach was somewhat more formal than Fisher’s. His view was that there are multiple things that you could do (accept the null or accept the alternative) and the point of the test was to tell you which one the data support. From this perspective, it is critical to specify your alternative hypothesis properly. If you don’t know what the alternative hypothesis is, then you don’t know how powerful the test is, or even which action makes sense. His framework genuinely requires a competition between different hypotheses. For Neyman, the \\(p\\) value didn’t directly measure the probability of the data (or data more extreme) under the null, it was more of an abstract description about which “possible tests” were telling you to accept the null, and which “possible tests” were telling you to accept the alternative.\n\nAs you can see, what we have today is an odd mishmash of the two. We talk about having both a null hypothesis and an alternative (Neyman), but usually169 define the \\(p\\) value in terms of exreme data (Fisher), but we still have \\(\\alpha\\) values (Neyman). Some of the statistical tests have explicitly specified alternatives (Neyman) but others are quite vague about it (Fisher). And, according to some people at least, we’re not allowed to talk about accepting the alternative (Fisher). It’s a mess: but I hope this at least explains why it’s a mess.\n\n### 11.9.2 Bayesians versus frequentists\n\nEarlier on in this chapter I was quite emphatic about the fact that you cannot interpret the \\(p\\) value as the probability that the null hypothesis is true. NHST is fundamentally a frequentist tool (see Chapter 9) and as such it does not allow you to assign probabilities to hypotheses: the null hypothesis is either true or it is not. The Bayesian approach to statistics interprets probability as a degree of belief, so it’s totally okay to say that there is a 10% chance that the null hypothesis is true: that’s just a reflection of the degree of confidence that you have in this hypothesis. You aren’t allowed to do this within the frequentist approach. Remember, if you’re a frequentist, a probability can only be defined in terms of what happens after a large number of independent replications (i.e., a long run frequency). If this is your interpretation of probability, talking about the “probability” that the null hypothesis is true is complete gibberish: a null hypothesis is either true or it is false. There’s no way you can talk about a long run frequency for this statement. To talk about “the probability of the null hypothesis” is as meaningless as “the colour of freedom”. It doesn’t have one!\n\nMost importantly, this isn’t a purely ideological matter. If you decide that you are a Bayesian and that you’re okay with making probability statements about hypotheses, you have to follow the Bayesian rules for calculating those probabilities. I’ll talk more about this in Chapter 17, but for now what I want to point out to you is the \\(p\\) value is a terrible approximation to the probability that \\(H_0\\) is true. If what you want to know is the probability of the null, then the \\(p\\) value is not what you’re looking for!\n\n### 11.9.3 Traps\n\nAs you can see, the theory behind hypothesis testing is a mess, and even now there are arguments in statistics about how it “should” work. However, disagreements among statisticians are not our real concern here. Our real concern is practical data analysis. And while the “orthodox” approach to null hypothesis significance testing has many drawbacks, even an unrepentant Bayesian like myself would agree that they can be useful if used responsibly. Most of the time they give sensible answers, and you can use them to learn interesting things. Setting aside the various ideologies and historical confusions that we’ve discussed, the fact remains that the biggest danger in all of statistics is thoughtlessness. I don’t mean stupidity, here: I literally mean thoughtlessness. The rush to interpret a result without spending time thinking through what each test actually says about the data, and checking whether that’s consistent with how you’ve interpreted it. That’s where the biggest trap lies.\n\nTo give an example of this, consider the following example (see Gelman and Stern 2006). Suppose I’m running my ESP study, and I’ve decided to analyse the data separately for the male participants and the female participants. Of the male participants, 33 out of 50 guessed the colour of the card correctly. This is a significant effect (\\(p = .03\\)). Of the female participants, 29 out of 50 guessed correctly. This is not a significant effect (\\(p = .32\\)). Upon observing this, it is extremely tempting for people to start wondering why there is a difference between males and females in terms of their psychic abilities. However, this is wrong. If you think about it, we haven’t actually run a test that explicitly compares males to females. All we have done is compare males to chance (binomial test was significant) and compared females to chance (binomial test was non significant). If we want to argue that there is a real difference between the males and the females, we should probably run a test of the null hypothesis that there is no difference! We can do that using a different hypothesis test,170 but when we do that it turns out that we have no evidence that males and females are significantly different (\\(p = .54\\)). Now do you think that there’s anything fundamentally different between the two groups? Of course not. What’s happened here is that the data from both groups (male and female) are pretty borderline: by pure chance, one of them happened to end up on the magic side of the \\(p = .05\\) line, and the other one didn’t. That doesn’t actually imply that males and females are different. This mistake is so common that you should always be wary of it: the difference between significant and not-significant is not evidence of a real difference – if you want to say that there’s a difference between two groups, then you have to test for that difference!\n\nThe example above is just that: an example. I’ve singled it out because it’s such a common one, but the bigger picture is that data analysis can be tricky to get right. Think about what it is you want to test, why you want to test it, and whether or not the answers that your test gives could possibly make any sense in the real world.\n\n## 11.10 Summary\n\nNull hypothesis testing is one of the most ubiquitous elements to statistical theory. The vast majority of scientific papers report the results of some hypothesis test or another. As a consequence it is almost impossible to get by in science without having at least a cursory understanding of what a \\(p\\)-value means, making this one of the most important chapters in the book. As usual, I’ll end the chapter with a quick recap of the key ideas that we’ve talked about:\n\n* Research hypotheses and statistical hypotheses. Null and alternative hypotheses. (Section 11.1).\n * Type 1 and Type 2 errors (Section 11.2)\n * Test statistics and sampling distributions (Section 11.3)\n * Hypothesis testing as a decision making process (Section 11.4)\n * \\(p\\)-values as “soft” decisions (Section 11.5)\n * Writing up the results of a hypothesis test (Section 11.6)\n * Effect size and power (Section 11.8)\n * A few issues to consider regarding hypothesis testing (Section 11.9)\n\nLater in the book, in Chapter 17, I’ll revisit the theory of null hypothesis tests from a Bayesian perspective, and introduce a number of new tools that you can use if you aren’t particularly fond of the orthodox approach. But for now, though, we’re done with the abstract statistical theory, and we can start discussing specific data analysis tools.\n\nThe quote comes from Wittgenstein’s (1922) text, Tractatus Logico-Philosphicus.↩\n * \n\nA technical note. The description below differs subtly from the standard description given in a lot of introductory texts. The orthodox theory of null hypothesis testing emerged from the work of and in the early 20th century; but Fisher and Neyman actually had very different views about how it should work. The standard treatment of hypothesis testing that most texts use is a hybrid of the two approaches. The treatment here is a little more Neyman-style than the orthodox view, especially as regards the meaning of the \\(p\\) value.↩\n * \n\nMy apologies to anyone who actually believes in this stuff, but on my reading of the literature on ESP, it’s just not reasonable to think this is real. To be fair, though, some of the studies are rigorously designed; so it’s actually an interesting area for thinking about psychological research design. And of course it’s a free country, so you can spend your own time and effort proving me wrong if you like, but I wouldn’t think that’s a terribly practical use of your intellect.↩\n * \n\nThis analogy only works if you’re from an adversarial legal system like UK/US/Australia. As I understand these things, the French inquisitorial system is quite different.↩\n * \n\nAn aside regarding the language you use to talk about hypothesis testing. Firstly, one thing you really want to avoid is the word “prove”: a statistical test really doesn’t prove that a hypothesis is true or false. Proof implies certainty, and as the saying goes, statistics means never having to say you’re certain. On that point almost everyone would agree. However, beyond that there’s a fair amount of confusion. Some people argue that you’re only allowed to make statements like “rejected the null”, “failed to reject the null”, or possibly “retained the null”. According to this line of thinking, you can’t say things like “accept the alternative” or “accept the null”. Personally I think this is too strong: in my opinion, this conflates null hypothesis testing with Karl Popper’s falsificationist view of the scientific process. While there are similarities between falsificationism and null hypothesis testing, they aren’t equivalent. However, while I personally think it’s fine to talk about accepting a hypothesis (on the proviso that “acceptance” doesn’t actually mean that it’s necessarily true, especially in the case of the null hypothesis), many people will disagree. And more to the point, you should be aware that this particular weirdness exists, so that you’re not caught unawares by it when writing up your own results.↩\n * \n\nStrictly speaking, the test I just constructed has \\(\\alpha = .057\\), which is a bit too generous. However, if I’d chosen 39 and 61 to be the boundaries for the critical region, then the critical region only covers 3.5% of the distribution. I figured that it makes more sense to use 40 and 60 as my critical values, and be willing to tolerate a 5.7% type I error rate, since that’s as close as I can get to a value of \\(\\alpha = .05\\).↩\n * \n\nThe internet seems fairly convinced that Ashley said this, though I can’t for the life of me find anyone willing to give a source for the claim.↩\n * \n\nThat’s \\(p = .000000000000000000000000136\\) for folks that don’t like scientific notation!↩\n * \n\nNote that the\n `p` here has nothing to do with a \\(p\\) value. The `p` argument in the `binom.test()` function corresponds to the probability of making a correct response, according to the null hypothesis. In other words, it’s the \\(\\theta\\) value.↩ * \n\nThere’s an R package called\n `compute.es` that can be used for calculating a very broad range of effect size measures; but for the purposes of the current book we won’t need it: all of the effect size measures that I’ll talk about here have functions in the `lsr` package↩ * \n\nAlthough in practice a very small effect size is worrying, because even very minor methodological flaws might be responsible for the effect; and in practice no experiment is perfect, so there are always methodological issues to worry about.↩\n * \n\nNotice that the true population parameter \\(\\theta\\) doesn’t necessarily correspond to an immutable fact of nature. In this context \\(\\theta\\) is just the true probability that people would correctly guess the colour of the card in the other room. As such the population parameter can be influenced by all sorts of things. Of course, this is all on the assumption that ESP actually exists!↩\n * \n\nAlthough this book describes both Neyman’s and Fisher’s definition of the \\(p\\) value, most don’t. Most introductory textbooks will only give you the Fisher version.↩\n * \n\nIn this case, the Pearson chi-square test of independence (Chapter 12;\n `chisq.test()` in R) is what we use; see also the `prop.test()` function.↩\n\nDate: 2010-01-12\nCategories: \nTags: \n\n# Chapter 12 Categorical data analysis\n\nNow that we’ve got the basic theory behind hypothesis testing, it’s time to start looking at specific tests that are commonly used in psychology. So where should we start? Not every textbook agrees on where to start, but I’m going to start with “\\(\\chi^2\\) tests” (this chapter) and “\\(t\\)-tests” (Chapter 13). Both of these tools are very frequently used in scientific practice, and while they’re not as powerful as “analysis of variance” (Chapter 14) and “regression” (Chapter 15) they’re much easier to understand.\n\nThe term “categorical data” is just another name for “nominal scale data”. It’s nothing that we haven’t already discussed, it’s just that in the context of data analysis people tend to use the term “categorical data” rather than “nominal scale data”. I don’t know why. In any case, categorical data analysis refers to a collection of tools that you can use when your data are nominal scale. However, there are a lot of different tools that can be used for categorical data analysis, and this chapter only covers a few of the more common ones.\n\n## 12.1 The \\(\\chi^2\\) goodness-of-fit test\n\nThe \\(\\chi^2\\) goodness-of-fit test is one of the oldest hypothesis tests around: it was invented by around the turn of the century (Pearson 1900), with some corrections made later by (Fisher 1922a). To introduce the statistical problem that it addresses, let’s start with some psychology…\n\n### 12.1.1 The cards data\n\n Over the years, there have been a lot of studies showing that humans have a lot of difficulties in simulating randomness. Try as we might to “act” random, we think in terms of patterns and structure, and so when asked to “do something at random”, what people actually do is anything but random. As a consequence, the study of human randomness (or non-randomness, as the case may be) opens up a lot of deep psychological questions about how we think about the world. With this in mind, let’s consider a very simple study. Suppose I asked people to imagine a shuffled deck of cards, and mentally pick one card from this imaginary deck “at random”. After they’ve chosen one card, I ask them to mentally select a second one. For both choices, what we’re going to look at is the suit (hearts, clubs, spades or diamonds) that people chose. After asking, say, \\(N=200\\) people to do this, I’d like to look at the data and figure out whether or not the cards that people pretended to select were really random. The data are contained in the `randomness.Rdata` file, which contains a single data frame called `cards` . Let’s take a look: \n\n```\nlibrary( lsr )\nload( file.path(projecthome, \"data/randomness.Rdata\" ))\nstr(cards)\n```\n\n```\n## 'data.frame': 200 obs. of 3 variables:\n## $ id : Factor w/ 200 levels \"subj1\",\"subj10\",..: 1 112 124 135 146 157 168 179 190 2 ...\n## $ choice_1: Factor w/ 4 levels \"clubs\",\"diamonds\",..: 4 2 3 4 3 1 3 2 4 2 ...\n## $ choice_2: Factor w/ 4 levels \"clubs\",\"diamonds\",..: 1 1 1 1 4 3 2 1 1 4 ...\n```\n\n As you can see, the `cards` data frame contains three variables, an `id` variable that assigns a unique identifier to each participant, and the two variables `choice_1` and `choice_2` that indicate the card suits that people chose. Here’s the first few entries in the data frame: `head( cards )` \n\n```\n## id choice_1 choice_2\n## 1 subj1 spades clubs\n## 2 subj2 diamonds clubs\n## 3 subj3 hearts clubs\n## 4 subj4 spades clubs\n## 5 subj5 hearts spades\n## 6 subj6 clubs hearts\n```\n\n For the moment, let’s just focus on the first choice that people made. We’ll use the `table()` function to count the number of times that we observed people choosing each suit. I’ll save the table to a variable called `observed` , for reasons that will become clear very soon: \n\n```\nobserved <- table( cards$choice_1 )\nobserved\n```\n\nThat little frequency table is quite helpful. Looking at it, there’s a bit of a hint that people might be more likely to select hearts than clubs, but it’s not completely obvious just from looking at it whether that’s really true, or if this is just due to chance. So we’ll probably have to do some kind of statistical analysis to find out, which is what I’m going to talk about in the next section.\n Excellent. From this point on, we’ll treat this table as the data that we’re looking to analyse. However, since I’m going to have to talk about this data in mathematical terms (sorry!) it might be a good idea to be clear about what the notation is. In R, if I wanted to pull out the number of people that selected diamonds, I could do it by name by typing `observed[\"diamonds\"]` but, since `\"diamonds\"` is second element of the `observed` vector, it’s equally effective to refer to it as `observed[2]` . The mathematical notation for this is pretty similar, except that we shorten the human-readable word “observed” to the letter \\(O\\), and we use subscripts rather than brackets: so the second observation in our table is written as `observed[2]` in R, and is written as \\(O_2\\) in maths. The relationship between the English descriptions, the R commands, and the mathematical symbols are illustrated below: \n\nlabel | index \\(i\\) | math. symbol | R command | the value |\n | --- | --- | --- | --- | --- |\nclubs \\(\\clubsuit\\) | 1 | \\(O_1\\) | | 35 |\ndiamonds \\(\\diamondsuit\\) | 2 | \\(O_2\\) | | 51 |\nhearts \\(\\heartsuit\\) | 3 | \\(O_3\\) | | 64 |\nspades \\(\\spadesuit\\) | 4 | \\(O_4\\) | | 50 |\n\n Hopefully that’s pretty clear. It’s also worth nothing that mathematicians prefer to talk about things in general rather than specific things, so you’ll also see the notation \\(O_i\\), which refers to the number of observations that fall within the \\(i\\)-th category (where \\(i\\) could be 1, 2, 3 or 4). Finally, if we want to refer to the set of all observed frequencies, statisticians group all of observed values into a vector, which I’ll refer to as \\(O\\). \\[ O = (O_1, O_2, O_3, O_4) \\] Again, there’s nothing new or interesting here: it’s just notation. If I say that \\(O~=~(35, 51, 64, 50)\\) all I’m doing is describing the table of observed frequencies (i.e., `observed` ), but I’m referring to it using mathematical notation, rather than by referring to an R variable. \n\n### 12.1.2 The null hypothesis and the alternative hypothesis\n\nAs the last section indicated, our research hypothesis is that “people don’t choose cards randomly”. What we’re going to want to do now is translate this into some statistical hypotheses, and construct a statistical test of those hypotheses. The test that I’m going to describe to you is Pearson’s \\(\\chi^2\\) goodness of fit test, and as is so often the case, we have to begin by carefully constructing our null hypothesis. In this case, it’s pretty easy. First, let’s state the null hypothesis in words:\n\n\\(H_0\\) |\n | --- |\nAll four suits are chosen with equal probability |\n\nNow, because this is statistics, we have to be able to say the same thing in a mathematical way. To do this, let’s use the notation \\(P_j\\) to refer to the true probability that the \\(j\\)-th suit is chosen. If the null hypothesis is true, then each of the four suits has a 25% chance of being selected: in other words, our null hypothesis claims that \\(P_1 = .25\\), \\(P_2 = .25\\), \\(P_3 = .25\\) and finally that \\(P_4 = .25\\). However, in the same way that we can group our observed frequencies into a vector \\(O\\) that summarises the entire data set, we can use \\(P\\) to refer to the probabilities that correspond to our null hypothesis. So if I let the vector \\(P = (P_1, P_2, P_3, P_4)\\) refer to the collection of probabilities that describe our null hypothesis, then we have\n\n\\[ H_0: {P} = (.25, .25, .25, .25) \\]\n\nIn this particular instance, our null hypothesis corresponds to a vector of probabilities \\(P\\) in which all of the probabilities are equal to one another. But this doesn’t have to be the case. For instance, if the experimental task was for people to imagine they were drawing from a deck that had twice as many clubs as any other suit, then the null hypothesis would correspond to something like \\(P = (.4, .2, .2, .2)\\). As long as the probabilities are all positive numbers, and they all sum to 1, them it’s a perfectly legitimate choice for the null hypothesis. However, the most common use of the goodness of fit test is to test a null hypothesis that all of the categories are equally likely, so we’ll stick to that for our example.\n\nWhat about our alternative hypothesis, \\(H_1\\)? All we’re really interested in is demonstrating that the probabilities involved aren’t all identical (that is, people’s choices weren’t completely random). As a consequence, the “human friendly” versions of our hypotheses look like this:\n\n\\(H_0\\) | \\(H_1\\) |\n | --- | --- |\nAll four suits are chosen with equal probability | At least one of the suit-choice probabilities isn’t .25 |\n\nand the “mathematician friendly” version is\n\n\\(H_0\\) | \\(H_1\\) |\n | --- | --- |\n\\(P = (.25, .25, .25, .25)\\) | \\(P \\neq (.25,.25,.25,.25)\\) |\n\n Conveniently, the mathematical version of the hypotheses looks quite similar to an R command defining a vector. So maybe what I should do is store the \\(P\\) vector in R as well, since we’re almost certainly going to need it later. And because I’m so imaginative, I’ll call this R vector `probabilities` , \n\n```\nprobabilities <- c(clubs = .25, diamonds = .25, hearts = .25, spades = .25) \nprobabilities\n```\n\n```\n## clubs diamonds hearts spades \n## 0.25 0.25 0.25 0.25\n```\n\n### 12.1.3 The “goodness of fit” test statistic\n\n At this point, we have our observed frequencies \\(O\\) and a collection of probabilities \\(P\\) corresponding the null hypothesis that we want to test. We’ve stored these in R as the corresponding variables `observed` and `probabilities` . What we now want to do is construct a test of the null hypothesis. As always, if we want to test \\(H_0\\) against \\(H_1\\), we’re going to need a test statistic. The basic trick that a goodness of fit test uses is to construct a test statistic that measures how “close” the data are to the null hypothesis. If the data don’t resemble what you’d “expect” to see if the null hypothesis were true, then it probably isn’t true. Okay, if the null hypothesis were true, what would we expect to see? Or, to use the correct terminology, what are the expected frequencies. There are \\(N=200\\) observations, and (if the null is true) the probability of any one of them choosing a heart is \\(P_3 = .25\\), so I guess we’re expecting \\(200 \\times .25 = 50\\) hearts, right? Or, more specifically, if we let \\(E_i\\) refer to “the number of category \\(i\\) responses that we’re expecting if the null is true”, then \\[\nE_i = N \\times P_i\n\\] This is pretty easy to calculate in R: \n\n```\nN <- 200 # sample size\nexpected <- N * probabilities # expected frequencies\nexpected\n```\n\n```\n## clubs diamonds hearts spades \n## 50 50 50 50\n```\n\nNone of which is very surprising: if there are 200 observation that can fall into four categories, and we think that all four categories are equally likely, then on average we’d expect to see 50 observations in each category, right?\n\nNow, how do we translate this into a test statistic? Clearly, what we want to do is compare the expected number of observations in each category (\\(E_i\\)) with the observed number of observations in that category (\\(O_i\\)). And on the basis of this comparison, we ought to be able to come up with a good test statistic. To start with, let’s calculate the difference between what the null hypothesis expected us to find and what we actually did find. That is, we calculate the “observed minus expected” difference score, \\(O_i - E_i\\). This is illustrated in the following table.\n\n\\(\\clubsuit\\) | \\(\\diamondsuit\\) | \\(\\heartsuit\\) | \\(\\spadesuit\\) |\n | --- | --- | --- | --- |\nexpected frequency | \\(E_i\\) | 50 | 50 | 50 | 50 |\nobserved frequency | \\(O_i\\) | 35 | 51 | 64 | 50 |\ndifference score | \\(O_i - E_i\\) | -15 | 1 | 14 | 0 |\n\n The same calculations can be done in R, using our `expected` and `observed` variables: `observed - expected` \n\n```\n## \n## clubs diamonds hearts spades \n## -15 1 14 0\n```\n\nRegardless of whether we do the calculations by hand or whether we do them in R, it’s clear that people chose more hearts and fewer clubs than the null hypothesis predicted. However, a moment’s thought suggests that these raw differences aren’t quite what we’re looking for. Intuitively, it feels like it’s just as bad when the null hypothesis predicts too few observations (which is what happened with hearts) as it is when it predicts too many (which is what happened with clubs). So it’s a bit weird that we have a negative number for clubs and a positive number for heards. One easy way to fix this is to square everything, so that we now calculate the squared differences, \\((E_i - O_i)^2\\). As before, we could do this by hand, but it’s easier to do it in R…\n\n```\n(observed - expected)^2\n```\n\n```\n## \n## clubs diamonds hearts spades \n## 225 1 196 0\n```\n\nNow we’re making progress. What we’ve got now is a collection of numbers that are big whenever the null hypothesis makes a bad prediction (clubs and hearts), but are small whenever it makes a good one (diamonds and spades). Next, for some technical reasons that I’ll explain in a moment, let’s also divide all these numbers by the expected frequency \\(E_i\\), so we’re actually calculating \\(\\frac{(E_i-O_i)^2}{E_i}\\). Since \\(E_i = 50\\) for all categories in our example, it’s not a very interesting calculation, but let’s do it anyway. The R command becomes:\n\n```\n## \n## clubs diamonds hearts spades \n## 4.50 0.02 3.92 0.00\n```\n\nIn effect, what we’ve got here are four different “error” scores, each one telling us how big a “mistake” the null hypothesis made when we tried to use it to predict our observed frequencies. So, in order to convert this into a useful test statistic, one thing we could do is just add these numbers up. The result is called the goodness of fit statistic, conventionally referred to either as \\(X^2\\) or GOF. We can calculate it using this command in R\n\n `## [1] 8.44` \nThe formula for this statistic looks remarkably similar to the R command. If we let \\(k\\) refer to the total number of categories (i.e., \\(k=4\\) for our cards data), then the \\(X^2\\) statistic is given by: \\[ X^2 = \\sum_{i=1}^k \\frac{(O_i - E_i)^2}{E_i} \\] Intuitively, it’s clear that if \\(X^2\\) is small, then the observed data \\(O_i\\) are very close to what the null hypothesis predicted \\(E_i\\), so we’re going to need a large \\(X^2\\) statistic in order to reject the null. As we’ve seen from our calculations, in our cards data set we’ve got a value of \\(X^2 = 8.44\\). So now the question becomes, is this a big enough value to reject the null?\n\n### 12.1.4 The sampling distribution of the GOF statistic (advanced)\n\nTo determine whether or not a particular value of \\(X^2\\) is large enough to justify rejecting the null hypothesis, we’re going to need to figure out what the sampling distribution for \\(X^2\\) would be if the null hypothesis were true. So that’s what I’m going to do in this section. I’ll show you in a fair amount of detail how this sampling distribution is constructed, and then – in the next section – use it to build up a hypothesis test. If you want to cut to the chase and are willing to take it on faith that the sampling distribution is a chi-squared (\\(\\chi^2\\)) distribution with \\(k-1\\) degrees of freedom, you can skip the rest of this section. However, if you want to understand why the goodness of fit test works the way it does, read on…\n\nOkay, let’s suppose that the null hypothesis is actually true. If so, then the true probability that an observation falls in the \\(i\\)-th category is \\(P_i\\) – after all, that’s pretty much the definition of our null hypothesis. Let’s think about what this actually means. If you think about it, this is kind of like saying that “nature” makes the decision about whether or not the observation ends up in category \\(i\\) by flipping a weighted coin (i.e., one where the probability of getting a head is \\(P_j\\)). And therefore, we can think of our observed frequency \\(O_i\\) by imagining that nature flipped \\(N\\) of these coins (one for each observation in the data set)… and exactly \\(O_i\\) of them came up heads. Obviously, this is a pretty weird way to think about the experiment. But what it does (I hope) is remind you that we’ve actually seen this scenario before. It’s exactly the same set up that gave rise to the binomial distribution in Section 9.4. In other words, if the null hypothesis is true, then it follows that our observed frequencies were generated by sampling from a binomial distribution: \\[ O_i \\sim \\mbox{Binomial}(P_i, N) \\] Now, if you remember from our discussion of the central limit theorem (Section 10.3.3), the binomial distribution starts to look pretty much identical to the normal distribution, especially when \\(N\\) is large and when \\(P_i\\) isn’t too close to 0 or 1. In other words as long as \\(N \\times P_i\\) is large enough – or, to put it another way, when the expected frequency \\(E_i\\) is large enough – the theoretical distribution of \\(O_i\\) is approximately normal. Better yet, if \\(O_i\\) is normally distributed, then so is \\((O_i - E_i)/\\sqrt{E_i}\\) … since \\(E_i\\) is a fixed value, subtracting off \\(E_i\\) and dividing by \\(\\sqrt{E_i}\\) changes the mean and standard deviation of the normal distribution; but that’s all it does. Okay, so now let’s have a look at what our goodness of fit statistic actually is. What we’re doing is taking a bunch of things that are normally-distributed, squaring them, and adding them up. Wait. We’ve seen that before too! As we discussed in Section 9.6, when you take a bunch of things that have a standard normal distribution (i.e., mean 0 and standard deviation 1), square them, then add them up, then the resulting quantity has a chi-square distribution. So now we know that the null hypothesis predicts that the sampling distribution of the goodness of fit statistic is a chi-square distribution. Cool.\n\nThere’s one last detail to talk about, namely the degrees of freedom. If you remember back to Section 9.6, I said that if the number of things you’re adding up is \\(k\\), then the degrees of freedom for the resulting chi-square distribution is \\(k\\). Yet, what I said at the start of this section is that the actual degrees of freedom for the chi-square goodness of fit test is \\(k-1\\). What’s up with that? The answer here is that what we’re supposed to be looking at is the number of genuinely independent things that are getting added together. And, as I’ll go on to talk about in the next section, even though there’s \\(k\\) things that we’re adding, only \\(k-1\\) of them are truly independent; and so the degrees of freedom is actually only \\(k-1\\). That’s the topic of the next section.171\n\n### 12.1.5 Degrees of freedom\n\nWhen I introduced the chi-square distribution in Section 9.6, I was a bit vague about what “degrees of freedom” actually means. Obviously, it matters: looking Figure 12.1 you can see that if we change the degrees of freedom, then the chi-square distribution changes shape quite substantially. But what exactly is it? Again, when I introduced the distribution and explained its relationship to the normal distribution, I did offer an answer… it’s the number of “normally distributed variables” that I’m squaring and adding together. But, for most people, that’s kind of abstract, and not entirely helpful. What we really need to do is try to understand degrees of freedom in terms of our data. So here goes.\n\nThe basic idea behind degrees of freedom is quite simple: you calculate it by counting up the number of distinct “quantities” that are used to describe your data; and then subtracting off all of the “constraints” that those data must satisfy.172 This is a bit vague, so let’s use our cards data as a concrete example. We describe out data using four numbers, \\(O_1\\), \\(O_2\\), \\(O_3\\) and \\(O_4\\) corresponding to the observed frequencies of the four different categories (hearts, clubs, diamonds, spades). These four numbers are the random outcomes of our experiment. But, my experiment actually has a fixed constraint built into it: the sample size \\(N\\).173 That is, if we know how many people chose hearts, how many chose diamonds and how many chose clubs; then we’d be able to figure out exactly how many chose spades. In other words, although our data are described using four numbers, they only actually correspond to \\(4-1 = 3\\) degrees of freedom. A slightly different way of thinking about it is to notice that there are four probabilities that we’re interested in (again, corresponding to the four different categories), but these probabilities must sum to one, which imposes a constraint. Therefore, the degrees of freedom is \\(4-1 = 3\\). Regardless of whether you want to think about it in terms of the observed frequencies or in terms of the probabilities, the answer is the same. In general, when running the chi-square goodness of fit test for an experiment involving \\(k\\) groups, then the degrees of freedom will be \\(k-1\\).\n\n### 12.1.6 Testing the null hypothesis\n\nThe final step in the process of constructing our hypothesis test is to figure out what the rejection region is. That is, what values of \\(X^2\\) would lead is to reject the null hypothesis. As we saw earlier, large values of \\(X^2\\) imply that the null hypothesis has done a poor job of predicting the data from our experiment, whereas small values of \\(X^2\\) imply that it’s actually done pretty well. Therefore, a pretty sensible strategy would be to say there is some critical value, such that if \\(X^2\\) is bigger than the critical value we reject the null; but if \\(X^2\\) is smaller than this value we retain the null. In other words, to use the language we introduced in Chapter @ref(hypothesistesting the chi-squared goodness of fit test is always a one-sided test. Right, so all we have to do is figure out what this critical value is. And it’s pretty straightforward. If we want our test to have significance level of \\(\\alpha = .05\\) (that is, we are willing to tolerate a Type I error rate of 5%), then we have to choose our critical value so that there is only a 5% chance that \\(X^2\\) could get to be that big if the null hypothesis is true. That is to say, we want the 95th percentile of the sampling distribution. This is illustrated in Figure 12.2.\n Ah, but – I hear you ask – how do I calculate the 95th percentile of a chi-squared distribution with \\(k-1\\) degrees of freedom? If only R had some function, called… oh, I don’t know, `qchisq()` … that would let you calculate this percentile (see Chapter 9 if you’ve forgotten). Like this… \n\n```\nqchisq( p = .95, df = 3 )\n```\n\n `## [1] 7.814728` So if our \\(X^2\\) statistic is bigger than 7.81 or so, then we can reject the null hypothesis. Since we actually calculated that before (i.e., \\(X^2 = 8.44\\)) we can reject the null. If we want an exact \\(p\\)-value, we can calculate it using the `pchisq()` function: \n\n```\npchisq( q = 8.44, df = 3, lower.tail = FALSE )\n```\n\n `## [1] 0.03774185` This is hopefully pretty straightforward, as long as you recall that the “ `p` ” form of the probability distribution functions in R always calculates the probability of getting a value of less than the value you entered (in this case 8.44). We want the opposite: the probability of getting a value of 8.44 or more. That’s why I told R to use the upper tail, not the lower tail. That said, it’s usually easier to calculate the \\(p\\)-value this way: \n\n```\n1-pchisq( q = 8.44, df = 3 )\n```\n\n `## [1] 0.03774185` \nSo, in this case we would reject the null hypothesis, since \\(p < .05\\). And that’s it, basically. You now know “Pearson’s \\(\\chi^2\\) test for the goodness of fit”. Lucky you.\n\n### 12.1.7 Doing the test in R\n\n Gosh darn it. Although we did manage to do everything in R as we were going through that little example, it does rather feel as if we’re typing too many things into the magic computing box. And I hate typing. Not surprisingly, R provides a function that will do all of these calculations for you. In fact, there are several different ways of doing it. The one that most people use is the `chisq.test()` function, which comes with every installation of R. I’ll show you how to use the `chisq.test()` function later on (in Section @ref(chisq.test), but to start out with I’m going to show you the `goodnessOfFitTest()` function in the `lsr` package, because it produces output that I think is easier for beginners to understand. It’s pretty straightforward: our raw data are stored in the variable `cards$choice_1` , right? If you want to test the null hypothesis that all four suits are equally likely, then (assuming you have the `lsr` package loaded) all you have to do is type this: \n\n```\ngoodnessOfFitTest( cards$choice_1 )\n```\n\nR then runs the test, and prints several lines of text. I’ll go through the output line by line, so that you can make sure that you understand what you’re looking at. The first two lines are just telling you things you already know:\n\n```\nChi-square test against specified probabilities\n\nData variable: cards$choice_1\n```\n\nThe first line tells us what kind of hypothesis test we ran, and the second line tells us the name of the variable that we ran it on. After that comes a statement of what the null and alternative hypotheses are:\n\n```\nHypotheses: \n null: true probabilities are as specified\n alternative: true probabilities differ from those specified\n```\n\n For a beginner, it’s kind of handy to have this as part of the output: it’s a nice reminder of what your null and alternative hypotheses are. Don’t get used to seeing this though. The vast majority of hypothesis tests in R aren’t so kind to novices. Most R functions are written on the assumption that you already understand the statistical tool that you’re using, so they don’t bother to include an explicit statement of the null and alternative hypothesis. The only reason that `goodnessOfFitTest()` actually does give you this is that I wrote it with novices in mind. \nThe next part of the output shows you the comparison between the observed frequencies and the expected frequencies:\n\n```\nDescriptives: \n observed freq. expected freq. specified prob.\nclubs 35 50 0.25\ndiamonds 51 50 0.25\nhearts 64 50 0.25\nspades 50 50 0.25\n```\n\nThe first column shows what the observed frequencies were, the second column shows the expected frequencies according to the null hypothesis, and the third column shows you what the probabilities actually were according to the null. For novice users, I think this is helpful: you can look at this part of the output and check that it makes sense: if it doesn’t you might have typed something incorrecrtly.\n\nThe last part of the output is the “important” stuff: it’s the result of the hypothesis test itself. There are three key numbers that need to be reported: the value of the \\(X^2\\) statistic, the degrees of freedom, and the \\(p\\)-value:\n\n```\nTest results: \n X-squared statistic: 8.44 \n degrees of freedom: 3 \n p-value: 0.038\n```\n\nNotice that these are the same numbers that we came up with when doing the calculations the long way.\n\n### 12.1.8 Specifying a different null hypothesis\n\nAt this point you might be wondering what to do if you want to run a goodness of fit test, but your null hypothesis is not that all categories are equally likely. For instance, let’s suppose that someone had made the theoretical prediction that people should choose red cards 60% of the time, and black cards 40% of the time (I’ve no idea why you’d predict that), but had no other preferences. If that were the case, the null hypothesis would be to expect 30% of the choices to be hearts, 30% to be diamonds, 20% to be spades and 20% to be clubs. This seems like a silly theory to me, and it’s pretty easy to test it using our data. All we need to do is specify the probabilities associated with the null hypothesis. We create a vector like this:\n\n```\nnullProbs <- c(clubs = .2, diamonds = .3, hearts = .3, spades = .2)\nnullProbs\n```\n\n```\n## clubs diamonds hearts spades \n## 0.2 0.3 0.3 0.2\n```\n\n Now that we have an explicitly specified null hypothesis, we include it in our command. This time round I’ll use the argument names properly. The data variable corresponds to the argument `x` , and the probabilities according to the null hypothesis correspond to the argument `p` . So our command is: \n\n```\ngoodnessOfFitTest( x = cards$choice_1, p = nullProbs )\n```\n\nAs you can see the null hypothesis and the expected frequencies are different to what they were last time. As a consequence our \\(X^2\\) test statistic is different, and our \\(p\\)-value is different too. Annoyingly, the \\(p\\)-value is .192, so we can’t reject the null hypothesis. Sadly, despite the fact that the null hypothesis corresponds to a very silly theory, these data don’t provide enough evidence against it.\n\n### 12.1.9 How to report the results of the test\n\nSo now you know how the test works, and you know how to do the test using a wonderful magic computing box. The next thing you need to know is how to write up the results. After all, there’s no point in designing and running an experiment and then analysing the data if you don’t tell anyone about it! So let’s now talk about what you need to do when reporting your analysis. Let’s stick with our card-suits example. If I wanted to write this result up for a paper or something, the conventional way to report this would be to write something like this:\n Of the 200 participants in the experiment, 64 selected hearts for their first choice, 51 selected diamonds, 50 selected spades, and 35 selected clubs. A chi-square goodness of fit test was conducted to test whether the choice probabilities were identical for all four suits. The results were significant (\\(\\chi^2(3) = 8.44, p<.05\\)), suggesting that people did not select suits purely at random.\n\nThis is pretty straightforward, and hopefully it seems pretty unremarkable. That said, there’s a few things that you should note about this description:\n\n* The statistical test is preceded by the descriptive statistics. That is, I told the reader something about what the data look like before going on to do the test. In general, this is good practice: always remember that your reader doesn’t know your data anywhere near as well as you do. So unless you describe it to them properly, the statistical tests won’t make any sense to them, and they’ll get frustrated and cry.\n * The description tells you what the null hypothesis being tested is. To be honest, writers don’t always do this, but it’s often a good idea in those situations where some ambiguity exists; or when you can’t rely on your readership being intimately familiar with the statistical tools that you’re using. Quite often the reader might not know (or remember) all the details of the test that your using, so it’s a kind of politeness to “remind” them! As far as the goodness of fit test goes, you can usually rely on a scientific audience knowing how it works (since it’s covered in most intro stats classes). However, it’s still a good idea to be explicit about stating the null hypothesis (briefly!) because the null hypothesis can be different depending on what you’re using the test for. For instance, in the cards example my null hypothesis was that all the four suit probabilities were identical (i.e., \\(P_1 = P_2 = P_3 = P_4 = 0.25\\)), but there’s nothing special about that hypothesis. I could just as easily have tested the null hypothesis that \\(P_1 = 0.7\\) and \\(P_2 = P_3 = P_4 = 0.1\\) using a goodness of fit test. So it’s helpful to the reader if you explain to them what your null hypothesis was. Also, notice that I described the null hypothesis in words, not in maths. That’s perfectly acceptable. You can describe it in maths if you like, but since most readers find words easier to read than symbols, most writers tend to describe the null using words if they can.\n * A “stat block” is included. When reporting the results of the test itself, I didn’t just say that the result was significant, I included a “stat block” (i.e., the dense mathematical-looking part in the parentheses), which reports all the “raw” statistical data. For the chi-square goodness of fit test, the information that gets reported is the test statistic (that the goodness of fit statistic was 8.44), the information about the distribution used in the test (\\(\\chi^2\\) with 3 degrees of freedom, which is usually shortened to \\(\\chi^2(3)\\)), and then the information about whether the result was significant (in this case \\(p<.05\\)). The particular information that needs to go into the stat block is different for every test, and so each time I introduce a new test I’ll show you what the stat block should look like.174 However the general principle is that you should always provide enough information so that the reader could check the test results themselves if they really wanted to.\n * The results are interpreted. In addition to indicating that the result was significant, I provided an interpretation of the result (i.e., that people didn’t choose randomly). This is also a kindness to the reader, because it tells them something about what they should believe about what’s going on in your data. If you don’t include something like this, it’s really hard for your reader to understand what’s going on.175\n\nAs with everything else, your overriding concern should be that you explain things to your reader. Always remember that the point of reporting your results is to communicate to another human being. I cannot tell you just how many times I’ve seen the results section of a report or a thesis or even a scientific article that is just gibberish, because the writer has focused solely on making sure they’ve included all the numbers, and forgotten to actually communicate with the human reader.\n\n### 12.1.10 A comment on statistical notation (advanced)\n\nSatan delights equally in statistics and in quoting scripture\n\n– \n\nIf you’ve been reading very closely, and are as much of a mathematical pedant as I am, there is one thing about the way I wrote up the chi-square test in the last section that might be bugging you a little bit. There’s something that feels a bit wrong with writing “\\(\\chi^2(3) = 8.44\\)”, you might be thinking. After all, it’s the goodness of fit statistic that is equal to 8.44, so shouldn’t I have written \\(X^2 = 8.44\\) or maybe GOF\\(=8.44\\)? This seems to be conflating the sampling distribution (i.e., \\(\\chi^2\\) with \\(df = 3\\)) with the test statistic (i.e., \\(X^2\\)). Odds are you figured it was a typo, since \\(\\chi\\) and \\(X\\) look pretty similar. Oddly, it’s not. Writing \\(\\chi^2(3) = 8.44\\) is essentially a highly condensed way of writing “the sampling distribution of the test statistic is \\(\\chi^2(3)\\), and the value of the test statistic is 8.44”.\n\nIn one sense, this is kind of stupid. There are lots of different test statistics out there that turn out to have a chi-square sampling distribution: the \\(X^2\\) statistic that we’ve used for our goodness of fit test is only one of many (albeit one of the most commonly encountered ones). In a sensible, perfectly organised world, we’d always have a separate name for the test statistic and the sampling distribution: that way, the stat block itself would tell you exactly what it was that the researcher had calculated. Sometimes this happens. For instance, the test statistic used in the Pearson goodness of fit test is written \\(X^2\\); but there’s a closely related test known as the \\(G\\)-test176 , in which the test statistic is written as \\(G\\). As it happens, the Pearson goodness of fit test and the \\(G\\)-test both test the same null hypothesis; and the sampling distribution is exactly the same (i.e., chi-square with \\(k-1\\) degrees of freedom). If I’d done a \\(G\\)-test for the cards data rather than a goodness of fit test, then I’d have ended up with a test statistic of \\(G = 8.65\\), which is slightly different from the \\(X^2 = 8.44\\) value that I got earlier; and produces a slightly smaller \\(p\\)-value of \\(p = .034\\). Suppose that the convention was to report the test statistic, then the sampling distribution, and then the \\(p\\)-value. If that were true, then these two situations would produce different stat blocks: my original result would be written \\(X^2 = 8.44, \\chi^2(3), p = .038\\), whereas the new version using the \\(G\\)-test would be written as \\(G = 8.65, \\chi^2(3), p = .034\\). However, using the condensed reporting standard, the original result is written \\(\\chi^2(3) = 8.44, p = .038\\), and the new one is written \\(\\chi^2(3) = 8.65, p = .034\\), and so it’s actually unclear which test I actually ran.\n\nSo why don’t we live in a world in which the contents of the stat block uniquely specifies what tests were ran? The deep reason is that life is messy. We (as users of statistical tools) want it to be nice and neat and organised… we want it to be designed, as if it were a product. But that’s not how life works: statistics is an intellectual discipline just as much as any other one, and as such it’s a massively distributed, partly-collaborative and partly-competitive project that no-one really understands completely. The things that you and I use as data analysis tools weren’t created by an Act of the Gods of Statistics; they were invented by lots of different people, published as papers in academic journals, implemented, corrected and modified by lots of other people, and then explained to students in textbooks by someone else. As a consequence, there’s a lot of test statistics that don’t even have names; and as a consequence they’re just given the same name as the corresponding sampling distribution. As we’ll see later, any test statistic that follows a \\(\\chi^2\\) distribution is commonly called a “chi-square statistic”; anything that follows a \\(t\\)-distribution is called a “\\(t\\)-statistic” and so on. But, as the \\(X^2\\) versus \\(G\\) example illustrates, two different things with the same sampling distribution are still, well, different.\n\nAs a consequence, it’s sometimes a good idea to be clear about what the actual test was that you ran, especially if you’re doing something unusual. If you just say “chi-square test”, it’s not actually clear what test you’re talking about. Although, since the two most common chi-square tests are the goodness of fit test and the independence test (Section 12.2), most readers with stats training can probably guess. Nevertheless, it’s something to be aware of.\n\n## 12.2 The \\(\\chi^2\\) test of independence (or association)\n\nGUARDBOT1: | Halt! |\n | --- | --- |\nGUARDBOT2: | Be you robot or human? |\nLEELA: | Robot…we be. |\nFRY: | Uh, yup! Just two robots out roboting it up! Eh? |\nGUARDBOT1: | Administer the test. |\nGUARDBOT2: | Which of the following would you most prefer? A: A puppy, B: A pretty flower from your sweetie, or C: A large properly-formatted data file? |\nGUARDBOT1: | Choose! |\n\n– Futurama, “Fear of a Bot Planet\n The other day I was watching an animated documentary examining the quaint customs of the natives of the planet Chapek 9. Apparently, in order to gain access to their capital city, a visitor must prove that they’re a robot, not a human. In order to determine whether or not visitor is human, they ask whether the visitor prefers puppies, flowers or large, properly formatted data files. “Pretty clever,” I thought to myself “but what if humans and robots have the same preferences? That probably wouldn’t be a very good test then, would it?” As it happens, I got my hands on the testing data that the civil authorities of Chapek 9 used to check this. It turns out that what they did was very simple… they found a bunch of robots and a bunch of humans and asked them what they preferred. I saved their data in a file called `chapek9.Rdata` , which I can now load and have a quick look at: \n\n```\nload( file.path(projecthome, \"data/chapek9.Rdata\" ))\nstr(chapek9)\n```\n\n```\n## 'data.frame': 180 obs. of 2 variables:\n## $ species: Factor w/ 2 levels \"robot\",\"human\": 1 2 2 2 1 2 2 1 2 1 ...\n## $ choice : Factor w/ 3 levels \"puppy\",\"flower\",..: 2 3 3 3 3 2 3 3 1 2 ...\n```\n\n Okay, so we have a single data frame called `chapek9` , which contains two factors, `species` and `choice` . As always, it’s nice to have a quick look at the data, `head(chapek9)` \n\n and then take a `summary()` , `summary(chapek9)` \n\n```\n## species choice \n## robot:87 puppy : 28 \n## human:93 flower: 43 \n## data :109\n```\n\n In total there are 180 entries in the data frame, one for each person (counting both robots and humans as “people”) who was asked to make a choice. Specifically, there’s 93 humans and 87 robots; and overwhelmingly the preferred choice is the data file. However, these summaries don’t address the question we’re interested in. To do that, we need a more detailed description of the data. What we want to do is look at the `choices` broken down by `species` . That is, we need to cross-tabulate the data (see Section 7.1). There’s quite a few ways to do this, as we’ve seen, but since our data are stored in a data frame, it’s convenient to use the `xtabs()` function. \n\n```\nchapekFrequencies <- xtabs( ~ choice + species, data = chapek9)\nchapekFrequencies\n```\n\nThat’s more or less what we’re after. So, if we add the row and column totals (which is convenient for the purposes of explaining the statistical tests), we would have a table like this,\n\nRobot | Human | Total |\n | --- | --- | --- |\nPuppy | 13 | 15 | 28 |\nFlower | 30 | 13 | 43 |\nData file | 44 | 65 | 109 |\nTotal | 87 | 93 | 180 |\nwhich actual | ly would | be a nice | way to report the descriptive statistics for this data set. In any case, it’s quite clear that the vast majority of the humans chose the data file, whereas the robots tended to be a lot more even in their preferences. Leaving aside the question of why the humans might be more likely to choose the data file for the moment (which does seem quite odd, admittedly), our first order of business is to determine if the discrepancy between human choices and robot choices in the data set is statistically significant. |\n\n### 12.2.1 Constructing our hypothesis test\n\nHow do we analyse this data? Specifically, since my research hypothesis is that “humans and robots answer the question in different ways”, how can I construct a test of the null hypothesis that “humans and robots answer the question the same way”? As before, we begin by establishing some notation to describe the data:\n\nRobot | Human | Total |\n | --- | --- | --- |\nPuppy | \\(O_{11}\\) | \\(O_{12}\\) | \\(R_{1}\\) |\nFlower | \\(O_{21}\\) | \\(O_{22}\\) | \\(R_{2}\\) |\nData file | \\(O_{31}\\) | \\(O_{32}\\) | \\(R_{3}\\) |\nTotal | \\(C_{1}\\) | \\(C_{2}\\) | \\(N\\) |\n\nIn this notation we say that \\(O_{ij}\\) is a count (observed frequency) of the number of respondents that are of species \\(j\\) (robots or human) who gave answer \\(i\\) (puppy, flower or data) when asked to make a choice. The total number of observations is written \\(N\\), as usual. Finally, I’ve used \\(R_i\\) to denote the row totals (e.g., \\(R_1\\) is the total number of people who chose the flower), and \\(C_j\\) to denote the column totals (e.g., \\(C_1\\) is the total number of robots).177\n\nSo now let’s think about what the null hypothesis says. If robots and humans are responding in the same way to the question, it means that the probability that “a robot says puppy” is the same as the probability that “a human says puppy”, and so on for the other two possibilities. So, if we use \\(P_{ij}\\) to denote “the probability that a member of species \\(j\\) gives response \\(i\\)” then our null hypothesis is that:\n\n\\(H_0\\): | All of the following are true: |\n | --- | --- |\n\\(P_{11} = P_{12}\\) (same probability of saying puppy) |\n\\(P_{21} = P_{22}\\) (same probability of saying flower) and |\n\\(P_{31} = P_{32}\\) (same probability of saying data). |\n\nAnd actually, since the null hypothesis is claiming that the true choice probabilities don’t depend on the species of the person making the choice, we can let \\(P_i\\) refer to this probability: e.g., \\(P_1\\) is the true probability of choosing the puppy.\n\nNext, in much the same way that we did with the goodness of fit test, what we need to do is calculate the expected frequencies. That is, for each of the observed counts \\(O_{ij}\\), we need to figure out what the null hypothesis would tell us to expect. Let’s denote this expected frequency by \\(E_{ij}\\). This time, it’s a little bit trickier. If there are a total of \\(C_j\\) people that belong to species \\(j\\), and the true probability of anyone (regardless of species) choosing option \\(i\\) is \\(P_i\\), then the expected frequency is just: \\[ E_{ij} = C_j \\times P_i \\] Now, this is all very well and good, but we have a problem. Unlike the situation we had with the goodness of fit test, the null hypothesis doesn’t actually specify a particular value for \\(P_i\\). It’s something we have to estimate (Chapter 10) from the data! Fortunately, this is pretty easy to do. If 28 out of 180 people selected the flowers, then a natural estimate for the probability of choosing flowers is \\(28/180\\), which is approximately \\(.16\\). If we phrase this in mathematical terms, what we’re saying is that our estimate for the probability of choosing option \\(i\\) is just the row total divided by the total sample size: \\[ \\hat{P}_i = \\frac{R_i}{N} \\] Therefore, our expected frequency can be written as the product (i.e. multiplication) of the row total and the column total, divided by the total number of observations:178 \\[ E_{ij} = \\frac{R_i \\times C_j}{N} \\] Now that we’ve figured out how to calculate the expected frequencies, it’s straightforward to define a test statistic; following the exact same strategy that we used in the goodness of fit test. In fact, it’s pretty much the same statistic. For a contingency table with \\(r\\) rows and \\(c\\) columns, the equation that defines our \\(X^2\\) statistic is \\[ X^2 = \\sum_{i=1}^r \\sum_{j=1}^c \\frac{({E}_{ij} - O_{ij})^2}{{E}_{ij}} \\] The only difference is that I have to include two summation sign (i.e., \\(\\sum\\)) to indicate that we’re summing over both rows and columns. As before, large values of \\(X^2\\) indicate that the null hypothesis provides a poor description of the data, whereas small values of \\(X^2\\) suggest that it does a good job of accounting for the data. Therefore, just like last time, we want to reject the null hypothesis if \\(X^2\\) is too large.\n\nNot surprisingly, this statistic is \\(\\chi^2\\) distributed. All we need to do is figure out how many degrees of freedom are involved, which actually isn’t too hard. As I mentioned before, you can (usually) think of the degrees of freedom as being equal to the number of data points that you’re analysing, minus the number of constraints. A contingency table with \\(r\\) rows and \\(c\\) columns contains a total of \\(r \\times c\\) observed frequencies, so that’s the total number of observations. What about the constraints? Here, it’s slightly trickier. The answer is always the same \\[ df = (r-1)(c-1) \\] but the explanation for why the degrees of freedom takes this value is different depending on the experimental design. For the sake of argument, let’s suppose that we had honestly intended to survey exactly 87 robots and 93 humans (column totals fixed by the experimenter), but left the row totals free to vary (row totals are random variables). Let’s think about the constraints that apply here. Well, since we deliberately fixed the column totals by Act of Experimenter, we have \\(c\\) constraints right there. But, there’s actually more to it than that. Remember how our null hypothesis had some free parameters (i.e., we had to estimate the \\(P_i\\) values)? Those matter too. I won’t explain why in this book, but every free parameter in the null hypothesis is rather like an additional constraint. So, how many of those are there? Well, since these probabilities have to sum to 1, there’s only \\(r-1\\) of these. So our total degrees of freedom is: \\[ \\begin{array}{rcl} df &=& \\mbox{(number of observations)} - \\mbox{(number of constraints)} \\\\ &=& (rc) - (c + (r-1)) \\\\ &=& rc - c - r + 1 \\\\ &=& (r - 1)(c - 1) \\end{array} \\] Alternatively, suppose that the only thing that the experimenter fixed was the total sample size \\(N\\). That is, we quizzed the first 180 people that we saw, and it just turned out that 87 were robots and 93 were humans. This time around our reasoning would be slightly different, but would still lead is to the same answer. Our null hypothesis still has \\(r-1\\) free parameters corresponding to the choice probabilities, but it now also has \\(c-1\\) free parameters corresponding to the species probabilities, because we’d also have to estimate the probability that a randomly sampled person turns out to be a robot.179 Finally, since we did actually fix the total number of observations \\(N\\), that’s one more constraint. So now we have, \\(rc\\) observations, and \\((c-1) + (r-1) + 1\\) constraints. What does that give? \\[ \\begin{array}{rcl} df &=& \\mbox{(number of observations)} - \\mbox{(number of constraints)} \\\\ &=& rc - ( (c-1) + (r-1) + 1) \\\\ &=& rc - c - r + 1 \\\\ &=& (r - 1)(c - 1) \\end{array} \\] Amazing.\n\n### 12.2.2 Doing the test in R\n\n Okay, now that we know how the test works, let’s have a look at how it’s done in R. As tempting as it is to lead you through the tedious calculations so that you’re forced to learn it the long way, I figure there’s no point. I already showed you how to do it the long way for the goodness of fit test in the last section, and since the test of independence isn’t conceptually any different, you won’t learn anything new by doing it the long way. So instead, I’ll go straight to showing you the easy way. As always, R lets you do it multiple ways. There’s the `chisq.test()` function, which I’ll talk about in Section @ref(chisq.test, but first I want to use the `associationTest()` function in the `lsr` package, which I think is easier on beginners. It works in the exact same way as the `xtabs()` function. Recall that, in order to produce the contingency table, we used this command: \n\n```\nxtabs( formula = ~choice+species, data = chapek9 )\n```\n\n The `associationTest()` function has exactly the same structure: it needs a `formula` that specifies which variables you’re cross-tabulating, and the name of a `data` frame that contains those variables. So the command is just this: \n\n```\nassociationTest( formula = ~choice+species, data = chapek9 )\n```\n\nJust like we did with the goodness of fit test, I’ll go through it line by line. The first two lines are, once again, just reminding you what kind of test you ran and what variables were used:\n\n```\nChi-square test of categorical association\n\nVariables: choice, species\n```\n\nNext, it tells you what the null and alternative hypotheses are (and again, I want to remind you not to get used to seeing these hypotheses written out so explicitly):\n\n```\nHypotheses: \n null: variables are independent of one another\n alternative: some contingency exists between variables\n```\n\nNext, it shows you the observed contingency table that is being tested:\n\nand it also shows you what the expected frequencies would be if the null hypothesis were true:\n\n```\nExpected contingency table under the null hypothesis:\n species\nchoice robot human\n puppy 13.5 14.5\n flower 20.8 22.2\n data 52.7 56.3\n```\n\nThe next part describes the results of the hypothesis test itself:\n\n```\nTest results: \n X-squared statistic: 10.722 \n degrees of freedom: 2 \n p-value: 0.005\n```\n\nAnd finally, it reports a measure of effect size:\n\n```\nOther information: \n estimated effect size (Cramer's v): 0.244\n```\n\nYou can ignore this bit for now. I’ll talk about it in just a moment.\n\nThis output gives us enough information to write up the result:\n Pearson’s \\(\\chi^2\\) revealed a significant association between species and choice (\\(\\chi^2(2) = 10.7, p < .01\\)): robots appeared to be more likely to say that they prefer flowers, but the humans were more likely to say they prefer data.\n\nNotice that, once again, I provided a little bit of interpretation to help the human reader understand what’s going on with the data. Later on in my discussion section, I’d provide a bit more context. To illustrate the difference, here’s what I’d probably say later on:\n The fact that humans appeared to have a stronger preference for raw data files than robots is somewhat counterintuitive. However, in context it makes some sense: the civil authority on Chapek 9 has an unfortunate tendency to kill and dissect humans when they are identified. As such it seems most likely that the human participants did not respond honestly to the question, so as to avoid potentially undesirable consequences. This should be considered to be a substantial methodological weakness.\n\nThis could be classified as a rather extreme example of a reactivity effect, I suppose. Obviously, in this case the problem is severe enough that the study is more or less worthless as a tool for understanding the difference preferences among humans and robots. However, I hope this illustrates the difference between getting a statistically significant result (our null hypothesis is rejected in favour of the alternative), and finding something of scientific value (the data tell us nothing of interest about our research hypothesis due to a big methodological flaw).\n\n### 12.2.3 Postscript\n\nI later found out the data were made up, and I’d been watching cartoons instead of doing work.\n\n## 12.3 The continuity correction\n\nOkay, time for a little bit of a digression. I’ve been lying to you a little bit so far. There’s a tiny change that you need to make to your calculations whenever you only have 1 degree of freedom. It’s called the “continuity correction”, or sometimes the Yates correction. Remember what I pointed out earlier: the \\(\\chi^2\\) test is based on an approximation, specifically on the assumption that binomial distribution starts to look like a normal distribution for large \\(N\\). One problem with this is that it often doesn’t quite work, especially when you’ve only got 1 degree of freedom (e.g., when you’re doing a test of independence on a \\(2 \\times 2\\) contingency table). The main reason for this is that the true sampling distribution for the \\(X^2\\) statistic is actually discrete (because you’re dealing with categorical data!) but the \\(\\chi^2\\) distribution is continuous. This can introduce systematic problems. Specifically, when \\(N\\) is small and when \\(df=1\\), the goodness of fit statistic tends to be “too big”, meaning that you actually have a bigger \\(\\alpha\\) value than you think (or, equivalently, the \\(p\\) values are a bit too small). Yates (1934) suggested a simple fix, in which you redefine the goodness of fit statistic as: \\[ X^2 = \\sum_{i} \\frac{(|E_i - O_i| - 0.5)^2}{E_i} \\] Basically, he just subtracts off 0.5 everywhere. As far as I can tell from reading Yates’ paper, the correction is basically a hack. It’s not derived from any principled theory: rather, it’s based on an examination of the behaviour of the test, and observing that the corrected version seems to work better. I feel obliged to explain this because you will sometimes see R (or any other software for that matter) introduce this correction, so it’s kind of useful to know what they’re about. You’ll know when it happens, because the R output will explicitly say that it has used a “continuity correction” or “Yates’ correction”.\n\n## 12.4 Effect size\n\nAs we discussed earlier (Section 11.8), it’s becoming commonplace to ask researchers to report some measure of effect size. So, let’s suppose that you’ve run your chi-square test, which turns out to be significant. So you now know that there is some association between your variables (independence test) or some deviation from the specified probabilities (goodness of fit test). Now you want to report a measure of effect size. That is, given that there is an association/deviation, how strong is it?\n\nThere are several different measures that you can choose to report, and several different tools that you can use to calculate them. I won’t discuss all of them,180 but will instead focus on the most commonly reported measures of effect size.\n\nBy default, the two measures that people tend to report most frequently are the \\(\\phi\\) statistic and the somewhat superior version, known as Cram'er’s \\(V\\). Mathematically, they’re very simple. To calculate the \\(\\phi\\) statistic, you just divide your \\(X^2\\) value by the sample size, and take the square root: \\[ \\phi = \\sqrt{\\frac{X^2}{N}} \\] The idea here is that the \\(\\phi\\) statistic is supposed to range between 0 (no at all association) and 1 (perfect association), but it doesn’t always do this when your contingency table is bigger than \\(2 \\times 2\\), which is a total pain. For bigger tables it’s actually possible to obtain \\(\\phi>1\\), which is pretty unsatisfactory. So, to correct for this, people usually prefer to report the \\(V\\) statistic proposed by Cramér (1946). It’s a pretty simple adjustment to \\(\\phi\\). If you’ve got a contingency table with \\(r\\) rows and \\(c\\) columns, then define \\(k = \\min(r,c)\\) to be the smaller of the two values. If so, then Cram'er’s \\(V\\) statistic is \\[ V = \\sqrt{\\frac{X^2}{N(k-1)}} \\] And you’re done. This seems to be a fairly popular measure, presumably because it’s easy to calculate, and it gives answers that aren’t completely silly: you know that \\(V\\) really does range from 0 (no at all association) to 1 (perfect association).\n Calculating \\(V\\) or \\(\\phi\\) is obviously pretty straightforward. So much so that the core packages in R don’t seem to have functions to do it, though other packages do. To save you the time and effort of finding one, I’ve included one in the `lsr` package, called `cramersV()` . It takes a contingency table as input, and prints out the measure of effect size: \n\n```\ncramersV( chapekFrequencies )\n```\n\n `## [1] 0.244058` However, if you’re using the `associationTest()` function to do your analysis, then you won’t actually need to use this at all, because it reports the Cram'er’s \\(V\\) statistic as part of the output. \n\n## 12.5 Assumptions of the test(s)\n\nAll statistical tests make assumptions, and it’s usually a good idea to check that those assumptions are met. For the chi-square tests discussed so far in this chapter, the assumptions are:\n\n* Expected frequencies are sufficiently large. Remember how in the previous section we saw that the \\(\\chi^2\\) sampling distribution emerges because the binomial distribution is pretty similar to a normal distribution? Well, like we discussed in Chapter 9 this is only true when the number of observations is sufficiently large. What that means in practice is that all of the expected frequencies need to be reasonably big. How big is reasonably big? Opinions differ, but the default assumption seems to be that you generally would like to see all your expected frequencies larger than about 5, though for larger tables you would probably be okay if at least 80% of the the expected frequencies are above 5 and none of them are below 1. However, from what I’ve been able to discover , these seem to have been proposed as rough guidelines, not hard and fast rules; and they seem to be somewhat conservative [Larntz1978].\n * Data are independent of one another. One somewhat hidden assumption of the chi-square test is that you have to genuinely believe that the observations are independent. Here’s what I mean. Suppose I’m interested in proportion of babies born at a particular hospital that are boys. I walk around the maternity wards, and observe 20 girls and only 10 boys. Seems like a pretty convincing difference, right? But later on, it turns out that I’d actually walked into the same ward 10 times, and in fact I’d only seen 2 girls and 1 boy. Not as convincing, is it? My original 30 observations were massively non-independent… and were only in fact equivalent to 3 independent observations. Obviously this is an extreme (and extremely silly) example, but it illustrates the basic issue. Non-independence “stuffs things up”. Sometimes it causes you to falsely reject the null, as the silly hospital example illustrats, but it can go the other way too. To give a slightly less stupid example, let’s consider what would happen if I’d done the cards experiment slightly differently: instead of asking 200 people to try to imagine sampling one card at random, suppose I asked 50 people to select 4 cards. One possibility would be that everyone selects one heart, one club, one diamond and one spade (in keeping with the “representativeness heuristic”; Tversky & Kahneman 1974). This is highly non-random behaviour from people, but in this case, I would get an observed frequency of 50 four all four suits. For this example, the fact that the observations are non-independent (because the four cards that you pick will be related to each other) actually leads to the opposite effect… falsely retaining the null.\n\nIf you happen to find yourself in a situation where independence is violated, it may be possible to use the McNemar test (which we’ll discuss) or the Cochran test (which we won’t). Similarly, if your expected cell counts are too small, check out the Fisher exact test. It is to these topics that we now turn.\n\n## 12.6 The most typical way to do chi-square tests in R\n\n When discussing how to do a chi-square goodness of fit test (Section 12.1.7) and the chi-square test of independence (Section 12.2.2), I introduced you to two separate functions in the `lsr` package. We ran our goodness of fit tests using the `goodnessOfFitTest()` function, and our tests of independence (or association) using the `associationTest()` function. And both of those functions produced quite detailed output, showing you the relevant descriptive statistics, printing out explicit reminders of what the hypotheses are, and so on. When you’re first starting out, it can be very handy to be given this sort of guidance. However, once you start becoming a bit more proficient in statistics and in R it can start to get very tiresome. A real statistician hardly needs to be told what the null and alternative hypotheses for a chi-square test are, and if an advanced R user wants the descriptive statistics to be printed out, they know how to produce them! For this reason, the basic `chisq.test()` function in R is a lot more terse in its output, and because the mathematics that underpin the goodness of fit test and the test of independence is basically the same in each case, it can run either test depending on what kind of input it is given. First, here’s the goodness of fit test. Suppose you have the frequency table `observed` that we used earlier, `observed` \n\n If you want to run the goodness of fit test against the hypothesis that all four suits are equally likely to appear, then all you need to do is input this frequenct table to the `chisq.test()` function: \n\n```\nchisq.test( x = observed )\n```\n\n Notice that the output is very compressed in comparison to the `goodnessOfFitTest()` function. It doesn’t bother to give you any descriptive statistics, it doesn’t tell you what null hypothesis is being tested, and so on. And as long as you already understand the test, that’s not a problem. Once you start getting familiar with R and with statistics, you’ll probably find that you prefer this simple output rather than the rather lengthy output that `goodnessOfFitTest()` produces. Anyway, if you want to change the null hypothesis, it’s exactly the same as before, just specify the probabilities using the `p` argument. For instance: \n\n```\nchisq.test( x = observed, p = c(.2, .3, .3, .2) )\n```\n\n Again, these are the same numbers that the `goodnessOfFitTest()` function reports at the end of the output. It just hasn’t included any of the other details. What about a test of independence? As it turns out, the `chisq.test()` function is pretty clever.181 If you input a cross-tabulation rather than a simple frequency table, it realises that you’re asking for a test of independence and not a goodness of fit test. Recall that we already have this cross-tabulation stored as the `chapekFrequencies` variable: `chapekFrequencies` \n\n To get the test of independence, all we have to do is feed this frequency table into the `chisq.test()` function like so: \n\n```\nchisq.test( chapekFrequencies )\n```\n\n```\n## \n## Pearson's Chi-squared test\n## \n## data: chapekFrequencies\n## X-squared = 10.722, df = 2, p-value = 0.004697\n```\n\n Again, the numbers are the same as last time, it’s just that the output is very terse and doesn’t really explain what’s going on in the rather tedious way that `associationTest()` does. As before, my intuition is that when you’re just getting started it’s easier to use something like `associationTest()` because it shows you more detail about what’s going on, but later on you’ll probably find that `chisq.test()` is more convenient. \n\n## 12.7 The Fisher exact test\n\n What should you do if your cell counts are too small, but you’d still like to test the null hypothesis that the two variables are independent? One answer would be “collect more data”, but that’s far too glib: there are a lot of situations in which it would be either infeasible or unethical do that. If so, statisticians have a kind of moral obligation to provide scientists with better tests. In this instance, Fisher (1922) kindly provided the right answer to the question. To illustrate the basic idea, let’s suppose that we’re analysing data from a field experiment, looking at the emotional status of people who have been accused of witchcraft; some of whom are currently being burned at the stake.182 Unfortunately for the scientist (but rather fortunately for the general populace), it’s actually quite hard to find people in the process of being set on fire, so the cell counts are awfully small in some cases. The `salem.Rdata` file illustrates the point: \n\n```\nload( file.path(projecthome, \"data/salem.Rdata\"))\n\nsalem.tabs <- table( trial )\nprint( salem.tabs )\n```\n\n```\n## on.fire\n## happy FALSE TRUE\n## FALSE 3 3\n## TRUE 10 0\n```\n\nLooking at this data, you’d be hard pressed not to suspect that people not on fire are more likely to be happy than people on fire. However, the chi-square test makes this very hard to test because of the small sample size. If I try to do so, R gives me a warning message:\n\n```\nchisq.test( salem.tabs )\n```\n\n```\n## Warning in chisq.test(salem.tabs): Chi-squared approximation may be\n## incorrect\n```\n\n```\n## \n## Pearson's Chi-squared test with Yates' continuity correction\n## \n## data: salem.tabs\n## X-squared = 3.3094, df = 1, p-value = 0.06888\n```\n\nSpeaking as someone who doesn’t want to be set on fire, I’d really like to be able to get a better answer than this. This is where Fisher’s exact test comes in very handy.\n\nThe Fisher exact test works somewhat differently to the chi-square test (or in fact any of the other hypothesis tests that I talk about in this book) insofar as it doesn’t have a test statistic; it calculates the \\(p\\)-value “directly”. I’ll explain the basics of how the test works for a \\(2 \\times 2\\) contingency table, though the test works fine for larger tables. As before, let’s have some notation:\n\nHappy | Sad | Total |\n | --- | --- | --- |\nSet on fire | \\(O_{11}\\) | \\(O_{12}\\) | \\(R_{1}\\) |\nNot set on fire | \\(O_{21}\\) | \\(O_{22}\\) | \\(R_{2}\\) |\nTotal | \\(C_{1}\\) | \\(C_{2}\\) | \\(N\\) |\n\nIn order to construct the test Fisher treats both the row and column totals (\\(R_1\\), \\(R_2\\), \\(C_1\\) and \\(C_2\\)) are known, fixed quantities; and then calculates the probability that we would have obtained the observed frequencies that we did (\\(O_{11}\\), \\(O_{12}\\), \\(O_{21}\\) and \\(O_{22}\\)) given those totals. In the notation that we developed in Chapter 9 this is written: \\[ P(O_{11}, O_{12}, O_{21}, O_{22} \\ | \\ R_1, R_2, C_1, C_2) \\] and as you might imagine, it’s a slightly tricky exercise to figure out what this probability is, but it turns out that this probability is described by a distribution known as the hypergeometric distribution.183 Now that we know this, what we have to do to calculate our \\(p\\)-value is calculate the probability of observing this particular table or a table that is “more extreme”.184 Back in the 1920s, computing this sum was daunting even in the simplest of situations, but these days it’s pretty easy as long as the tables aren’t too big and the sample size isn’t too large. The conceptually tricky issue is to figure out what it means to say that one contingency table is more “extreme” than another. The easiest solution is to say that the table with the lowest probability is the most extreme. This then gives us the \\(p\\)-value.\n The implementation of the test in R is via the `fisher.test()` function. Here’s how it is used: \n\n```\nfisher.test( salem.tabs )\n```\n\n```\n## \n## Fisher's Exact Test for Count Data\n## \n## data: salem.tabs\n## p-value = 0.03571\n## alternative hypothesis: true odds ratio is not equal to 1\n## 95 percent confidence interval:\n## 0.000000 1.202913\n## sample estimates:\n## odds ratio \n## 0\n```\n\nThis is a bit more output than we got from some of our earlier tests. The main thing we’re interested in here is the \\(p\\)-value, which in this case is small enough (\\(p=.036\\)) to justify rejecting the null hypothesis that people on fire are just as happy as people not on fire.\n\n## 12.8 The McNemar test\n\nSuppose you’ve been hired to work for the Australian Generic Political Party (AGPP), and part of your job is to find out how effective the AGPP political advertisements are. So, what you do, is you put together a sample of \\(N=100\\) people, and ask them to watch the AGPP ads. Before they see anything, you ask them if they intend to vote for the AGPP; and then after showing the ads, you ask them again, to see if anyone has changed their minds. Obviously, if you’re any good at your job, you’d also do a whole lot of other things too, but let’s consider just this one simple experiment. One way to describe your data is via the following contingency table:\n\nBefore | After | Total |\n | --- | --- | --- |\nYes | 30 | 10 | 40 |\nNo | 70 | 90 | 160 |\nTotal | 100 | 100 | 200 |\n\nAt first pass, you might think that this situation lends itself to the Pearson \\(\\chi^2\\) test of independence (as per Section 12.2). However, a little bit of thought reveals that we’ve got a problem: we have 100 participants, but 200 observations. This is because each person has provided us with an answer in both the before column and the after column. What this means is that the 200 observations aren’t independent of each other: if voter A says “yes” the first time and voter B says “no”, then you’d expect that voter A is more likely to say “yes” the second time than voter B! The consequence of this is that the usual \\(\\chi^2\\) test won’t give trustworthy answers due to the violation of the independence assumption. Now, if this were a really uncommon situation, I wouldn’t be bothering to waste your time talking about it. But it’s not uncommon at all: this is a standard repeated measures design, and none of the tests we’ve considered so far can handle it. Eek.\n\nThe solution to the problem was published by McNemar (1947). The trick is to start by tabulating your data in a slightly different way:\n\nBefore: Yes | Before: No | Total |\n | --- | --- | --- |\nAfter: Yes | 5 | 5 | 10 |\nAfter: No | 25 | 65 | 90 |\nTotal | 30 | 70 | 100 |\n\nThis is exactly the same data, but it’s been rewritten so that each of our 100 participants appears in only one cell. Because we’ve written our data this way, the independence assumption is now satisfied, and this is a contingency table that we can use to construct an \\(X^2\\) goodness of fit statistic. However, as we’ll see, we need to do it in a slightly nonstandard way. To see what’s going on, it helps to label the entries in our table a little differently:\n\nBefore: Yes | Before: No | Total |\n | --- | --- | --- |\nAfter: Yes | \\(a\\) | \\(b\\) | \\(a+b\\) |\nAfter: No | \\(c\\) | \\(d\\) | \\(c+d\\) |\nTotal | \\(a+c\\) | \\(b+d\\) | \\(n\\) |\n\nNext, let’s think about what our null hypothesis is: it’s that the “before” test and the “after” test have the same proportion of people saying “Yes, I will vote for AGPP”. Because of the way that we have rewritten the data, it means that we’re now testing the hypothesis that the row totals and column totals come from the same distribution. Thus, the null hypothesis in McNemar’s test is that we have “marginal homogeneity”. That is, the row totals and column totals have the same distribution: \\(P_a + P_b = P_a + P_c\\), and similarly that \\(P_c + P_d = P_b + P_d\\). Notice that this means that the null hypothesis actually simplifies to \\(P_b = P_c\\). In other words, as far as the McNemar test is concerned, it’s only the off-diagonal entries in this table (i.e., \\(b\\) and \\(c\\)) that matter! After noticing this, the McNemar test of marginal homogeneity is no different to a usual \\(\\chi^2\\) test. After applying the Yates correction, our test statistic becomes: \\[ X^2 = \\frac{(|b-c| - 0.5)^2}{b+c} \\] or, to revert to the notation that we used earlier in this chapter: \\[ X^2 = \\frac{(|O_{12}-O_{21}| - 0.5)^2}{O_{12} + O_{21}} \\] and this statistic has an (approximately) \\(\\chi^2\\) distribution with \\(df=1\\). However, remember that – just like the other \\(\\chi^2\\) tests – it’s only an approximation, so you need to have reasonably large expected cell counts for it to work.\n\n### 12.8.1 Doing the McNemar test in R\n\n Now that you know what the McNemar test is all about, lets actually run one. The `agpp.Rdata` file contains the raw data that I discussed previously, so let’s have a look at it: \n\n```\nload(file.path(projecthome, \"data/agpp.Rdata\"))\nstr(agpp)\n```\n\n```\n## 'data.frame': 100 obs. of 3 variables:\n## $ id : Factor w/ 100 levels \"subj.1\",\"subj.10\",..: 1 13 24 35 46 57 68 79 90 2 ...\n## $ response_before: Factor w/ 2 levels \"no\",\"yes\": 1 2 2 2 1 1 1 1 1 1 ...\n## $ response_after : Factor w/ 2 levels \"no\",\"yes\": 2 1 1 1 1 1 1 2 1 1 ...\n```\n\n The `agpp` data frame contains three variables, an `id` variable that labels each participant in the data set (we’ll see why that’s useful in a moment), a `response_before` variable that records the person’s answer when they were asked the question the first time, and a `response_after` variable that shows the answer that they gave when asked the same question a second time. As usual, here’s the first 6 entries: `head(agpp)` \n\n```\n## id response_before response_after\n## 1 subj.1 no yes\n## 2 subj.2 yes no\n## 3 subj.3 yes no\n## 4 subj.4 yes no\n## 5 subj.5 no no\n## 6 subj.6 no no\n```\n\nand here’s a summary:\n `summary(agpp)` \n\n```\n## id response_before response_after\n## subj.1 : 1 no :70 no :90 \n## subj.10 : 1 yes:30 yes:10 \n## subj.100: 1 \n## subj.11 : 1 \n## subj.12 : 1 \n## subj.13 : 1 \n## (Other) :94\n```\n\n Notice that each participant appears only once in this data frame. When we tabulate this data frame using `xtabs()` , we get the appropriate table: \n\n```\nright.table <- xtabs( ~ response_before + response_after, data = agpp)\nprint( right.table )\n```\n\n```\n## response_after\n## response_before no yes\n## no 65 5\n## yes 25 5\n```\n\n and from there, we can run the McNemar test by using the `mcnemar.test()` function: \n\n```\nmcnemar.test( right.table )\n```\n\n```\n## \n## McNemar's Chi-squared test with continuity correction\n## \n## data: right.table\n## McNemar's chi-squared = 12.033, df = 1, p-value = 0.0005226\n```\n\nAnd we’re done. We’ve just run a McNemar’s test to determine if people were just as likely to vote AGPP after the ads as they were before hand. The test was significant (\\(\\chi^2(1) = 12.04, p<.001\\)), suggesting that they were not. And in fact, it looks like the ads had a negative effect: people were less likely to vote AGPP after seeing the ads. Which makes a lot of sense when you consider the quality of a typical political advertisement.\n\n## 12.9 What’s the difference between McNemar and independence?\n\n Let’s go all the way back to the beginning of the chapter, and look at the `cards` data set again. If you recall, the actual experimental design that I described involved people making two choices. Because we have information about the first choice and the second choice that everyone made, we can construct the following contingency table that cross-tabulates the first choice against the second choice. \n\n```\ncardChoices <- xtabs( ~ choice_1 + choice_2, data = cards )\ncardChoices\n```\n\n```\n## choice_2\n## choice_1 clubs diamonds hearts spades\n## clubs 10 9 10 6\n## diamonds 20 4 13 14\n## hearts 20 18 3 23\n## spades 18 13 15 4\n```\n\nSuppose I wanted to know whether the choice you make the second time is dependent on the choice you made the first time. This is where a test of independence is useful, and what we’re trying to do is see if there’s some relationship between the rows and columns of this table. Here’s the result:\n\n```\nchisq.test( cardChoices )\n```\n\n```\n## \n## Pearson's Chi-squared test\n## \n## data: cardChoices\n## X-squared = 29.237, df = 9, p-value = 0.0005909\n```\n\n Alternatively, suppose I wanted to know if on average, the frequencies of suit choices were different the second time than the first time. In that situation, what I’m really trying to see if the row totals in `cardChoices` (i.e., the frequencies for `choice_1` ) are different from the column totals (i.e., the frequencies for `choice_2` ). That’s when you use the McNemar test: \n\n```\nmcnemar.test( cardChoices )\n```\n\n```\n## \n## McNemar's Chi-squared test\n## \n## data: cardChoices\n## McNemar's chi-squared = 16.033, df = 6, p-value = 0.01358\n```\n\nNotice that the results are different! These aren’t the same test.\n\n## 12.10 Summary\n\nThe key ideas discussed in this chapter are:\n\n* The chi-square goodness of fit test (Section 12.1) is used when you have a table of observed frequencies of different categories; and the null hypothesis gives you a set of “known” probabilities to compare them to. You can either use the\n `goodnessOfFitTest()` function in the `lsr` package to run this test, or the `chisq.test()` function. * The chi-square test of independence (Section 12.2) is used when you have a contingency table (cross-tabulation) of two categorical variables. The null hypothesis is that there is no relationship/association between the variables. You can either use the\n `associationTest()` function in the `lsr` package, or you can use `chisq.test()` . * Effect size for a contingency table can be measured in several ways (Section 12.4). In particular we noted the Cramer’s \\(V\\) statistic, which can be calculated using\n `cramersV()` . This is also part of the output produced by `associationTest()` . * Both versions of the Pearson test rely on two assumptions: that the expected frequencies are sufficiently large, and that the observations are independent (Section 12.5). The Fisher exact test (Section 12.7) can be used when the expected frequencies are small,\n\n```\nfisher.test(x = contingency.table)\n```\n\n. The McNemar test (Section 12.8) can be used for some kinds of violations of independence, \n\n```\nmcnemar.test(x = contingency.table)\n```\n\n. \nIf you’re interested in learning more about categorical data analysis, a good first choice would be Agresti (1996) which, as the title suggests, provides an Introduction to Categorical Data Analysis. If the introductory book isn’t enough for you (or can’t solve the problem you’re working on) you could consider Agresti (2002), Categorical Data Analysis. The latter is a more advanced text, so it’s probably not wise to jump straight from this book to that one.\n\n. 2002. Categorical Data Analysis. 2nd ed. Hoboken, NJ: Wiley.\n\nI should point out that this issue does complicate the story somewhat: I’m not going to cover it in this book, but there’s a sneaky trick that you can do to rewrite the equation for the goodness of fit statistic as a sum over \\(k-1\\) independent things. When we do so we get the “proper” sampling distribution, which is chi-square with \\(k-1\\) degrees of freedom. In fact, in order to get the maths to work out properly, you actually have to rewrite things that way. But it’s beyond the scope of an introductory book to show the maths in that much detail: all I wanted to do is give you a sense of why the goodness of fit statistic is associated with the chi-squared distribution.↩\n * \n\nI feel obliged to point out that this is an over-simplification. It works nicely for quite a few situations; but every now and then we’ll come across degrees of freedom values that aren’t whole numbers. Don’t let this worry you too much – when you come across this, just remind yourself that “degrees of freedom” is actually a bit of a messy concept, and that the nice simple story that I’m telling you here isn’t the whole story. For an introductory class, it’s usually best to stick to the simple story: but I figure it’s best to warn you to expect this simple story to fall apart. If I didn’t give you this warning, you might start getting confused when you see \\(df = 3.4\\) or something; and (incorrectly) thinking that you had misunderstood something that I’ve taught you, rather than (correctly) realising that there’s something that I haven’t told you.↩\n * \n\nIn practice, the sample size isn’t always fixed… e.g., we might run the experiment over a fixed period of time, and the number of people participating depends on how many people show up. That doesn’t matter for the current purposes.↩\n * \n\nWell, sort of. The conventions for how statistics should be reported tend to differ somewhat from discipline to discipline; I’ve tended to stick with how things are done in psychology, since that’s what I do. But the general principle of providing enough information to the reader to allow them to check your results is pretty universal, I think.↩\n * \n\nTo some people, this advice might sound odd, or at least in conflict with the “usual” advice on how to write a technical report. Very typically, students are told that the “results” section of a report is for describing the data and reporting statistical analysis; and the “discussion” section is for providing interpretation. That’s true as far as it goes, but I think people often interpret it way too literally. The way I usually approach it is to provide a quick and simple interpretation of the data in the results section, so that my reader understands what the data are telling us. Then, in the discussion, I try to tell a bigger story; about how my results fit with the rest of the scientific literature. In short; don’t let the “interpretation goes in the discussion” advice turn your results section into incomprehensible garbage. Being understood by your reader is much more important.↩\n * \n\nComplicating matters, the \\(G\\)-test is a special case of a whole class of tests that are known as likelihood ratio tests. I don’t cover LRTs in this book, but they are quite handy things to know about.↩\n * \n\nA technical note. The way I’ve described the test pretends that the column totals are fixed (i.e., the researcher intended to survey 87 robots and 93 humans) and the row totals are random (i.e., it just turned out that 28 people chose the puppy). To use the terminology from my mathematical statistics textbook (Hogg, McKean, and Craig 2005) I should technically refer to this situation as a chi-square test of homogeneity; and reserve the term chi-square test of independence for the situation where both the row and column totals are random outcomes of the experiment. In the initial drafts of this book that’s exactly what I did. However, it turns out that these two tests are identical; and so I’ve collapsed them together.↩\n * \n\nTechnically, \\(E_{ij}\\) here is an estimate, so I should probably write it \\(\\hat{E}_{ij}\\). But since no-one else does, I won’t either.↩\n * \n\nA problem many of us worry about in real life.↩\n * \n\nThough I do feel that it’s worth mentioning the\n `assocstats()` function in the `vcd` package. If you install and load the `vcd` package, then a command like \n\n```\nassocstats( chapekFrequencies )\n```\n\nwill run the \\(\\chi^2\\) test as well as the likelihood ratio test (not discussed here); and then report three different measures of effect size: \\(\\phi^2\\), Cram'er’s \\(V\\), and the contingency coefficient (not discussed here)↩ * \n\nNot really.↩\n * \n\nThis example is based on a joke article published in the Journal of Irreproducible Results.↩\n * \n\nThe R functions for this distribution are\n `dhyper()` , `phyper()` , `qhyper()` and `rhyper()` , though you don’t need them for this book, and I haven’t given you enough information to use these to perform the Fisher exact test the long way.↩ * \n\nNot surprisingly, the Fisher exact test is motivated by Fisher’s interpretation of a \\(p\\)-value, not Neyman’s!↩\n\nDate: 2013-10-01\nCategories: \nTags: \n\n# Chapter 13 Comparing two means\n\nIn the previous chapter we covered the situation when your outcome variable is nominal scale and your predictor variable185 is also nominal scale. Lots of real world situations have that character, and so you’ll find that chi-square tests in particular are quite widely used. However, you’re much more likely to find yourself in a situation where your outcome variable is interval scale or higher, and what you’re interested in is whether the average value of the outcome variable is higher in one group or another. For instance, a psychologist might want to know if anxiety levels are higher among parents than non-parents, or if working memory capacity is reduced by listening to music (relative to not listening to music). In a medical context, we might want to know if a new drug increases or decreases blood pressure. An agricultural scientist might want to know whether adding phosphorus to Australian native plants will kill them.186 In all these situations, our outcome variable is a fairly continuous, interval or ratio scale variable; and our predictor is a binary “grouping” variable. In other words, we want to compare the means of the two groups.\n\nThe standard answer to the problem of comparing means is to use a \\(t\\)-test, of which there are several varieties depending on exactly what question you want to solve. As a consequence, the majority of this chapter focuses on different types of \\(t\\)-test: one sample \\(t\\)-tests are discussed in Section 13.2, independent samples \\(t\\)-tests are discussed in Sections 13.3 and 13.4, and paired samples \\(t\\)-tests are discussed in Section 13.5. After that, we’ll talk a bit about Cohen’s \\(d\\), which is the standard measure of effect size for a \\(t\\)-test (Section 13.8). The later sections of the chapter focus on the assumptions of the \\(t\\)-tests, and possible remedies if they are violated. However, before discussing any of these useful things, we’ll start with a discussion of the \\(z\\)-test.\n\n## 13.1 The one-sample \\(z\\)-test\n\nIn this section I’ll describe one of the most useless tests in all of statistics: the \\(z\\)-test. Seriously – this test is almost never used in real life. Its only real purpose is that, when teaching statistics, it’s a very convenient stepping stone along the way towards the \\(t\\)-test, which is probably the most (over)used tool in all statistics.\n\n### 13.1.1 The inference problem that the test addresses\n\n To introduce the idea behind the \\(z\\)-test, let’s use a simple example. A friend of mine, Dr Zeppo, grades his introductory statistics class on a curve. Let’s suppose that the average grade in his class is 67.5, and the standard deviation is 9.5. Of his many hundreds of students, it turns out that 20 of them also take psychology classes. Out of curiosity, I find myself wondering: do the psychology students tend to get the same grades as everyone else (i.e., mean 67.5) or do they tend to score higher or lower? He emails me the `zeppo.Rdata` file, which I use to pull up the `grades` of those students, \n\n```\nload( file.path(projecthome, \"data/zeppo.Rdata\" )) \nprint( grades )\n```\n\n```\n## [1] 50 60 60 64 66 66 67 69 70 74 76 76 77 79 79 79 81 82 82 89\n```\n\nand calculate the mean:\n `mean( grades )` `## [1] 72.3` \nHm. It might be that the psychology students are scoring a bit higher than normal: that sample mean of \\(\\bar{X} = 72.3\\) is a fair bit higher than the hypothesised population mean of \\(\\mu = 67.5\\), but on the other hand, a sample size of \\(N = 20\\) isn’t all that big. Maybe it’s pure chance.\n\nTo answer the question, it helps to be able to write down what it is that I think I know. Firstly, I know that the sample mean is \\(\\bar{X} = 72.3\\). If I’m willing to assume that the psychology students have the same standard deviation as the rest of the class then I can say that the population standard deviation is \\(\\sigma = 9.5\\). I’ll also assume that since Dr Zeppo is grading to a curve, the psychology student grades are normally distributed.\n\nNext, it helps to be clear about what I want to learn from the data. In this case, my research hypothesis relates to the population mean \\(\\mu\\) for the psychology student grades, which is unknown. Specifically, I want to know if \\(\\mu = 67.5\\) or not. Given that this is what I know, can we devise a hypothesis test to solve our problem? The data, along with the hypothesised distribution from which they are thought to arise, are shown in Figure 13.1. Not entirely obvious what the right answer is, is it? For this, we are going to need some statistics.\n\n### 13.1.2 Constructing the hypothesis test\n\nThe first step in constructing a hypothesis test is to be clear about what the null and alternative hypotheses are. This isn’t too hard to do. Our null hypothesis, \\(H_0\\), is that the true population mean \\(\\mu\\) for psychology student grades is 67.5%; and our alternative hypothesis is that the population mean isn’t 67.5%. If we write this in mathematical notation, these hypotheses become, \\[ \\begin{array}{ll} H_0: & \\mu = 67.5 \\\\ H_1: & \\mu \\neq 67.5 \\end{array} \\] though to be honest this notation doesn’t add much to our understanding of the problem, it’s just a compact way of writing down what we’re trying to learn from the data. The null hypotheses \\(H_0\\) and the alternative hypothesis \\(H_1\\) for our test are both illustrated in Figure 13.2. In addition to providing us with these hypotheses, the scenario outlined above provides us with a fair amount of background knowledge that might be useful. Specifically, there are two special pieces of information that we can add:\n\n1 The psychology grades are normally distributed. 1 The true standard deviation of these scores \\(\\sigma\\) is known to be 9.5.\n\nFor the moment, we’ll act as if these are absolutely trustworthy facts. In real life, this kind of absolutely trustworthy background knowledge doesn’t exist, and so if we want to rely on these facts we’ll just have make the assumption that these things are true. However, since these assumptions may or may not be warranted, we might need to check them. For now though, we’ll keep things simple.\n\nThe next step is to figure out what we would be a good choice for a diagnostic test statistic; something that would help us discriminate between \\(H_0\\) and \\(H_1\\). Given that the hypotheses all refer to the population mean \\(\\mu\\), you’d feel pretty confident that the sample mean \\(\\bar{X}\\) would be a pretty useful place to start. What we could do, is look at the difference between the sample mean \\(\\bar{X}\\) and the value that the null hypothesis predicts for the population mean. In our example, that would mean we calculate \\(\\bar{X} - 67.5\\). More generally, if we let \\(\\mu_0\\) refer to the value that the null hypothesis claims is our population mean, then we’d want to calculate \\[ \\bar{X} - \\mu_0 \\] If this quantity equals or is very close to 0, things are looking good for the null hypothesis. If this quantity is a long way away from 0, then it’s looking less likely that the null hypothesis is worth retaining. But how far away from zero should it be for us to reject \\(H_0\\)?\n\nTo figure that out, we need to be a bit more sneaky, and we’ll need to rely on those two pieces of background knowledge that I wrote down previously, namely that the raw data are normally distributed, and we know the value of the population standard deviation \\(\\sigma\\). If the null hypothesis is actually true, and the true mean is \\(\\mu_0\\), then these facts together mean that we know the complete population distribution of the data: a normal distribution with mean \\(\\mu_0\\) and standard deviation \\(\\sigma\\). Adopting the notation from Section 9.5, a statistician might write this as: \\[ X \\sim \\mbox{Normal}(\\mu_0,\\sigma^2) \\]\n\nOkay, if that’s true, then what can we say about the distribution of \\(\\bar{X}\\)? Well, as we discussed earlier (see Section 10.3.3), the sampling distribution of the mean \\(\\bar{X}\\) is also normal, and has mean \\(\\mu\\). But the standard deviation of this sampling distribution \\(\\mbox{SE}({\\bar{X}})\\), which is called the standard error of the mean, is \\[ \\mbox{SE}({\\bar{X}}) = \\frac{\\sigma}{\\sqrt{N}} \\] In other words, if the null hypothesis is true then the sampling distribution of the mean can be written as follows: \\[ \\bar{X} \\sim \\mbox{Normal}(\\mu_0,\\mbox{SE}({\\bar{X}})) \\] Now comes the trick. What we can do is convert the sample mean \\(\\bar{X}\\) into a standard score (Section 5.6). This is conventionally written as \\(z\\), but for now I’m going to refer to it as \\(z_{\\bar{X}}\\). (The reason for using this expanded notation is to help you remember that we’re calculating standardised version of a sample mean, not a standardised version of a single observation, which is what a \\(z\\)-score usually refers to). When we do so, the \\(z\\)-score for our sample mean is \\[ z_{\\bar{X}} = \\frac{\\bar{X} - \\mu_0}{\\mbox{SE}({\\bar{X}})} \\] or, equivalently \\[ z_{\\bar{X}} = \\frac{\\bar{X} - \\mu_0}{\\sigma / \\sqrt{N}} \\] This \\(z\\)-score is our test statistic. The nice thing about using this as our test statistic is that like all \\(z\\)-scores, it has a standard normal distribution: \\[ z_{\\bar{X}} \\sim \\mbox{Normal}(0,1) \\] (again, see Section 5.6 if you’ve forgotten why this is true). In other words, regardless of what scale the original data are on, the \\(z\\)-statistic iteself always has the same interpretation: it’s equal to the number of standard errors that separate the observed sample mean \\(\\bar{X}\\) from the population mean \\(\\mu_0\\) predicted by the null hypothesis. Better yet, regardless of what the population parameters for the raw scores actually are, the 5% critical regions for \\(z\\)-test are always the same, as illustrated in Figures 13.4 and 13.3. And what this meant, way back in the days where people did all their statistics by hand, is that someone could publish a table like this:\n\ndesired \\(\\alpha\\) level | two-sided test | one-sided test |\n | --- | --- | --- |\n.1 | 1.644854 | 1.281552 |\n.05 | 1.959964 | 1.644854 |\n.01 | 2.575829 | 2.326348 |\n.001 | 3.290527 | 3.090232 |\n\nwhich in turn meant that researchers could calculate their \\(z\\)-statistic by hand, and then look up the critical value in a text book. That was an incredibly handy thing to be able to do back then, but it’s kind of unnecessary these days, since it’s trivially easy to do it with software like R.\n\n### 13.1.3 A worked example using R\n\n Now, as I mentioned earlier, the \\(z\\)-test is almost never used in practice. It’s so rarely used in real life that the basic installation of R doesn’t have a built in function for it. However, the test is so incredibly simple that it’s really easy to do one manually. Let’s go back to the data from Dr Zeppo’s class. Having loaded the `grades` data, the first thing I need to do is calculate the sample mean: \n\n```\nsample.mean <- mean( grades )\nprint( sample.mean )\n```\n\n `## [1] 72.3` \nThen, I create variables corresponding to known population standard deviation (\\(\\sigma = 9.5\\)), and the value of the population mean that the null hypothesis specifies (\\(\\mu_0 = 67.5\\)):\n\n```\nmu.null <- 67.5\nsd.true <- 9.5\n```\n\n Let’s also create a variable for the sample size. We could count up the number of observations ourselves, and type `N <- 20` at the command prompt, but counting is tedious and repetitive. Let’s get R to do the tedious repetitive bit by using the `length()` function, which tells us how many elements there are in a vector: \n\n```\nN <- length( grades )\nprint( N )\n```\n\n `## [1] 20` \nNext, let’s calculate the (true) standard error of the mean:\n\n```\nsem.true <- sd.true / sqrt(N)\nprint(sem.true)\n```\n\n `## [1] 2.124265` \nAnd finally, we calculate our \\(z\\)-score:\n\n```\nz.score <- (sample.mean - mu.null) / sem.true\nprint( z.score )\n```\n\n `## [1] 2.259606` \nAt this point, we would traditionally look up the value 2.26 in our table of critical values. Our original hypothesis was two-sided (we didn’t really have any theory about whether psych students would be better or worse at statistics than other students) so our hypothesis test is two-sided (or two-tailed) also. Looking at the little table that I showed earlier, we can see that 2.26 is bigger than the critical value of 1.96 that would be required to be significant at \\(\\alpha = .05\\), but smaller than the value of 2.58 that would be required to be significant at a level of \\(\\alpha = .01\\). Therefore, we can conclude that we have a significant effect, which we might write up by saying something like this:\n With a mean grade of 73.2 in the sample of psychology students, and assuming a true population standard deviation of 9.5, we can conclude that the psychology students have significantly different statistics scores to the class average (\\(z = 2.26\\), \\(N=20\\), \\(p<.05\\)).\n However, what if want an exact \\(p\\)-value? Well, back in the day, the tables of critical values were huge, and so you could look up your actual \\(z\\)-value, and find the smallest value of \\(\\alpha\\) for which your data would be significant (which, as discussed earlier, is the very definition of a \\(p\\)-value). However, looking things up in books is tedious, and typing things into computers is awesome. So let’s do it using R instead. Now, notice that the \\(\\alpha\\) level of a \\(z\\)-test (or any other test, for that matter) defines the total area “under the curve” for the critical region, right? That is, if we set \\(\\alpha = .05\\) for a two-sided test, then the critical region is set up such that the area under the curve for the critical region is \\(.05\\). And, for the \\(z\\)-test, the critical value of 1.96 is chosen that way because the area in the lower tail (i.e., below \\(-1.96\\)) is exactly \\(.025\\) and the area under the upper tail (i.e., above \\(1.96\\)) is exactly \\(.025\\). So, since our observed \\(z\\)-statistic is \\(2.26\\), why not calculate the area under the curve below \\(-2.26\\) or above \\(2.26\\)? In R we can calculate this using the `pnorm()` function. For the upper tail: \n\n```\nupper.area <- pnorm( q = z.score, lower.tail = FALSE )\nprint( upper.area )\n```\n\n `## [1] 0.01192287` The `lower.tail = FALSE` is me telling R to calculate the area under the curve from 2.26 and upwards. If I’d told it that `lower.tail = TRUE` , then R would calculate the area from 2.26 and below, and it would give me an answer 0.9880771. Alternatively, to calculate the area from \\(-2.26\\) and below, we get \n\n```\nlower.area <- pnorm( q = -z.score, lower.tail = TRUE )\nprint( lower.area )\n```\n\n `## [1] 0.01192287` \nThus we get our \\(p\\)-value:\n\n```\np.value <- lower.area + upper.area\nprint( p.value )\n```\n\n `## [1] 0.02384574` \n\n### 13.1.4 Assumptions of the \\(z\\)-test\n\nAs I’ve said before, all statistical tests make assumptions. Some tests make reasonable assumptions, while other tests do not. The test I’ve just described – the one sample \\(z\\)-test – makes three basic assumptions. These are:\n\n* Normality. As usually described, the \\(z\\)-test assumes that the true population distribution is normal.187 is often pretty reasonable, and not only that, it’s an assumption that we can check if we feel worried about it (see Section 13.9).\n * Independence. The second assumption of the test is that the observations in your data set are not correlated with each other, or related to each other in some funny way. This isn’t as easy to check statistically: it relies a bit on good experimetal design. An obvious (and stupid) example of something that violates this assumption is a data set where you “copy” the same observation over and over again in your data file: so you end up with a massive “sample size”, consisting of only one genuine observation. More realistically, you have to ask yourself if it’s really plausible to imagine that each observation is a completely random sample from the population that you’re interested in. In practice, this assumption is never met; but we try our best to design studies that minimise the problems of correlated data.\n * Known standard deviation. The third assumption of the \\(z\\)-test is that the true standard deviation of the population is known to the researcher. This is just stupid. In no real world data analysis problem do you know the standard deviation \\(\\sigma\\) of some population, but are completely ignorant about the mean \\(\\mu\\). In other words, this assumption is always wrong.\n\nIn view of the stupidity of assuming that \\(\\sigma\\) is known, let’s see if we can live without it. This takes us out of the dreary domain of the \\(z\\)-test, and into the magical kingdom of the \\(t\\)-test, with unicorns and fairies and leprechauns, and um…\n\n## 13.2 The one-sample \\(t\\)-test\n\nAfter some thought, I decided that it might not be safe to assume that the psychology student grades necessarily have the same standard deviation as the other students in Dr Zeppo’s class. After all, if I’m entertaining the hypothesis that they don’t have the same mean, then why should I believe that they absolutely have the same standard deviation? In view of this, I should really stop assuming that I know the true value of \\(\\sigma\\). This violates the assumptions of my \\(z\\)-test, so in one sense I’m back to square one. However, it’s not like I’m completely bereft of options. After all, I’ve still got my raw data, and those raw data give me an estimate of the population standard deviation:\n `sd( grades )` `## [1] 9.520615` \nIn other words, while I can’t say that I know that \\(\\sigma = 9.5\\), I can say that \\(\\hat\\sigma = 9.52\\).\n\nOkay, cool. The obvious thing that you might think to do is run a \\(z\\)-test, but using the estimated standard deviation of 9.52 instead of relying on my assumption that the true standard deviation is 9.5. So, we could just type this new number into R and out would come the answer. And you probably wouldn’t be surprised to hear that this would still give us a significant result. This approach is close, but it’s not quite correct. Because we are now relying on an estimate of the population standard deviation, we need to make some adjustment for the fact that we have some uncertainty about what the true population standard deviation actually is. Maybe our data are just a fluke … maybe the true population standard deviation is 11, for instance. But if that were actually true, and we ran the \\(z\\)-test assuming \\(\\sigma=11\\), then the result would end up being non-significant. That’s a problem, and it’s one we’re going to have to address.\n\n### 13.2.1 Introducing the \\(t\\)-test\n\nThis ambiguity is annoying, and it was resolved in 1908 by a guy called (Student 1908), who was working as a chemist for the Guinness brewery at the time (see Box 1987). Because Guinness took a dim view of its employees publishing statistical analysis (apparently they felt it was a trade secret), he published the work under the pseudonym “A Student”, and to this day, the full name of the \\(t\\)-test is actually Student’s \\(t\\)-test. The key thing that Gosset figured out is how we should accommodate the fact that we aren’t completely sure what the true standard deviation is.188 The answer is that it subtly changes the sampling distribution. In the \\(t\\)-test, our test statistic (now called a \\(t\\)-statistic) is calculated in exactly the same way I mentioned above. If our null hypothesis is that the true mean is \\(\\mu\\), but our sample has mean \\(\\bar{X}\\) and our estimate of the population standard deviation is \\(\\hat{\\sigma}\\), then our \\(t\\) statistic is: \\[ t = \\frac{\\bar{X} - \\mu}{\\hat{\\sigma}/\\sqrt{N} } \\] The only thing that has changed in the equation is that instead of using the known true value \\(\\sigma\\), we use the estimate \\(\\hat{\\sigma}\\). And if this estimate has been constructed from \\(N\\) observations, then the sampling distribution turns into a \\(t\\)-distribution with \\(N-1\\) degrees of freedom (df). The \\(t\\) distribution is very similar to the normal distribution, but has “heavier” tails, as discussed earlier in Section 9.6 and illustrated in Figure 13.5. Notice, though, that as df gets larger, the \\(t\\)-distribution starts to look identical to the standard normal distribution. This is as it should be: if you have a sample size of \\(N = 70,000,000\\) then your “estimate” of the standard deviation would be pretty much perfect, right? So, you should expect that for large \\(N\\), the \\(t\\)-test would behave exactly the same way as a \\(z\\)-test. And that’s exactly what happens!\n\n### 13.2.2 Doing the test in R\n\n As you might expect, the mechanics of the \\(t\\)-test are almost identical to the mechanics of the \\(z\\)-test. So there’s not much point in going through the tedious exercise of showing you how to do the calculations using low level commands: it’s pretty much identical to the calculations that we did earlier, except that we use the estimated standard deviation (i.e., something like `se.est <- sd(grades)` ), and then we test our hypothesis using the \\(t\\) distribution rather than the normal distribution (i.e. we use `pt()` rather than `pnorm()` . And so instead of going through the calculations in tedious detail for a second time, I’ll jump straight to showing you how \\(t\\)-tests are actually done in practice. The situation with \\(t\\)-tests is very similar to the one we encountered with chi-squared tests in Chapter 12. R comes with one function called `t.test()` that is very flexible (it can run lots of different kinds of \\(t\\)-tests) and is somewhat terse (the output is quite compressed). Later on in the chapter I’ll show you how to use the `t.test()` function (Section 13.7), but to start out with I’m going to rely on some simpler functions in the `lsr` package. Just like last time, what I’ve done is written a few simpler functions, each of which does only one thing. So, if you want to run a one-sample \\(t\\)-test, use the `oneSampleTTest()` function! It’s pretty straightforward to use: all you need to do is specify `x` , the variable containing the data, and `mu` , the true population mean according to the null hypothesis. All you need to type is this: \n\n```\nlibrary(lsr)\n\noneSampleTTest( x=grades, mu=67.5 )\n```\n\n```\n## \n## One sample t-test \n## \n## Data variable: grades \n## \n## Descriptive statistics: \n## grades\n## mean 72.300\n## std dev. 9.521\n## \n## Hypotheses: \n## null: population mean equals 67.5 \n## alternative: population mean not equal to 67.5 \n## \n## Test results: \n## t-statistic: 2.255 \n## degrees of freedom: 19 \n## p-value: 0.036 \n## \n## Other information: \n## two-sided 95% confidence interval: [67.844, 76.756] \n## estimated effect size (Cohen's d): 0.504\n```\n\nEasy enough. Now lets go through the output. Just like we saw in the last chapter, I’ve written the functions so that the output is pretty verbose. It tries to describe in a lot of detail what its actually done:\n\n```\nOne sample t-test \n\nData variable: grades \n\nDescriptive statistics: \n grades\n mean 72.300\n std dev. 9.521\n\nHypotheses: \n null: population mean equals 67.5 \n alternative: population mean not equal to 67.5 \n\nTest results: \n t-statistic: 2.255 \n degrees of freedom: 19 \n p-value: 0.036 \n\nOther information: \n two-sided 95% confidence interval: [67.844, 76.756] \n estimated effect size (Cohen's d): 0.504\n```\n\nReading this output from top to bottom, you can see it’s trying to lead you through the data analysis process. The first two lines tell you what kind of test was run and what data were used. It then gives you some basic information about the sample: specifically, the sample mean and standard deviation of the data. It then moves towards the inferential statistics part. It starts by telling you what the null and alternative hypotheses were, and then it reports the results of the test: the \\(t\\)-statistic, the degrees of freedom, and the \\(p\\)-value. Finally, it reports two other things you might care about: the confidence interval for the mean, and a measure of effect size (we’ll talk more about effect sizes later).\n\nSo that seems straightforward enough. Now what do we do with this output? Well, since we’re pretending that we actually care about my toy example, we’re overjoyed to discover that the result is statistically significant (i.e. \\(p\\) value below .05). We could report the result by saying something like this:\n With a mean grade of 72.3, the psychology students scored slightly higher than the average grade of 67.5 (\\(t(19) = 2.25\\), \\(p<.05\\)); the 95% confidence interval is [67.8, 76.8].\n\nwhere \\(t(19)\\) is shorthand notation for a \\(t\\)-statistic that has 19 degrees of freedom. That said, it’s often the case that people don’t report the confidence interval, or do so using a much more compressed form than I’ve done here. For instance, it’s not uncommon to see the confidence interval included as part of the stat block, like this:\n \\(t(19) = 2.25\\), \\(p<.05\\), CI\\(_{95} = [67.8, 76.8]\\)\n\nWith that much jargon crammed into half a line, you know it must be really smart.189\n\n### 13.2.3 Assumptions of the one sample \\(t\\)-test\n\nOkay, so what assumptions does the one-sample \\(t\\)-test make? Well, since the \\(t\\)-test is basically a \\(z\\)-test with the assumption of known standard deviation removed, you shouldn’t be surprised to see that it makes the same assumptions as the \\(z\\)-test, minus the one about the known standard deviation. That is\n\n* Normality. We’re still assuming that the the population distribution is normal^[A technical comment… in the same way that we can weaken the assumptions of the \\(z\\)-test so that we’re only talking about the sampling distribution, we can weaken the \\(t\\) test assumptions so that we don’t have to assume normality of the population. However, for the \\(t\\)-test, it’s trickier to do this. As before, we can replace the assumption of population normality with an assumption that the sampling distribution of \\(\\bar{X}\\) is normal. However, remember that we’re also relying on a sample estimate of the standard deviation; and so we also require the sampling distribution of \\(\\hat{\\sigma}\\) to be chi-square. That makes things nastier, and this version is rarely used in practice: fortunately, if the population is normal, then both of these two assumptions are met., and as noted earlier, there are standard tools that you can use to check to see if this assumption is met (Section 13.9), and other tests you can do in it’s place if this assumption is violated (Section 13.10).\n * Independence. Once again, we have to assume that the observations in our sample are generated independently of one another. See the earlier discussion about the \\(z\\)-test for specifics (Section 13.1.4).\n\nOverall, these two assumptions aren’t terribly unreasonable, and as a consequence the one-sample \\(t\\)-test is pretty widely used in practice as a way of comparing a sample mean against a hypothesised population mean.\n\n## 13.3 The independent samples \\(t\\)-test (Student test)\n\nAlthough the one sample \\(t\\)-test has its uses, it’s not the most typical example of a \\(t\\)-test190. A much more common situation arises when you’ve got two different groups of observations. In psychology, this tends to correspond to two different groups of participants, where each group corresponds to a different condition in your study. For each person in the study, you measure some outcome variable of interest, and the research question that you’re asking is whether or not the two groups have the same population mean. This is the situation that the independent samples \\(t\\)-test is designed for.\n\n### 13.3.1 The data\n\n Suppose we have 33 students taking Dr Harpo’s statistics lectures, and Dr Harpo doesn’t grade to a curve. Actually, Dr Harpo’s grading is a bit of a mystery, so we don’t really know anything about what the average grade is for the class as a whole. There are two tutors for the class, Anastasia and Bernadette. There are \\(N_1 = 15\\) students in Anastasia’s tutorials, and \\(N_2 = 18\\) in Bernadette’s tutorials. The research question I’m interested in is whether Anastasia or Bernadette is a better tutor, or if it doesn’t make much of a difference. Dr Harpo emails me the course grades, in the `harpo.Rdata` file. As usual, I’ll load the file and have a look at what variables it contains: \n\n```\nload (file.path(projecthome, \"data/harpo.Rdata\" ))\nstr(harpo)\n```\n\n```\n## 'data.frame': 33 obs. of 2 variables:\n## $ grade: num 65 72 66 74 73 71 66 76 69 79 ...\n## $ tutor: Factor w/ 2 levels \"Anastasia\",\"Bernadette\": 1 2 2 1 1 2 2 2 2 2 ...\n```\n\n As we can see, there’s a single data frame with two variables, `grade` and `tutor` . The `grade` variable is a numeric vector, containing the grades for all \\(N = 33\\) students taking Dr Harpo’s class; the `tutor` variable is a factor that indicates who each student’s tutor was. The first six observations in this data set are shown below: `head( harpo )` \n\n We can calculate means and standard deviations, using the `mean()` and `sd()` functions. Rather than show the R output, here’s a nice little summary table: \n\nmean | std dev | N |\n | --- | --- | --- |\nAnastasia’s students | 74.53 | 9.00 | 15 |\nBernadette’s students | 69.06 | 5.77 | 18 |\n\nTo give you a more detailed sense of what’s going on here, I’ve plotted histograms showing the distribution of grades for both tutors (Figure 13.6 and 13.7). Inspection of these histograms suggests that the students in Anastasia’s class may be getting slightly better grades on average, though they also seem a little more variable.\n\nHere is a simpler plot showing the means and corresponding confidence intervals for both groups of students (Figure 13.8).\n\n### 13.3.2 Introducing the test\n\nThe independent samples \\(t\\)-test comes in two different forms, Student’s and Welch’s. The original Student \\(t\\)-test – which is the one I’ll describe in this section – is the simpler of the two, but relies on much more restrictive assumptions than the Welch \\(t\\)-test. Assuming for the moment that you want to run a two-sided test, the goal is to determine whether two “independent samples” of data are drawn from populations with the same mean (the null hypothesis) or different means (the alternative hypothesis). When we say “independent” samples, what we really mean here is that there’s no special relationship between observations in the two samples. This probably doesn’t make a lot of sense right now, but it will be clearer when we come to talk about the paired samples \\(t\\)-test later on. For now, let’s just point out that if we have an experimental design where participants are randomly allocated to one of two groups, and we want to compare the two groups’ mean performance on some outcome measure, then an independent samples \\(t\\)-test (rather than a paired samples \\(t\\)-test) is what we’re after.\n\nOkay, so let’s let \\(\\mu_1\\) denote the true population mean for group 1 (e.g., Anastasia’s students), and \\(\\mu_2\\) will be the true population mean for group 2 (e.g., Bernadette’s students),191 and as usual we’ll let \\(\\bar{X}_1\\) and \\(\\bar{X}_2\\) denote the observed sample means for both of these groups. Our null hypothesis states that the two population means are identical (\\(\\mu_1 = \\mu_2\\)) and the alternative to this is that they are not (\\(\\mu_1 \\neq \\mu_2\\)). Written in mathematical-ese, this is… \\[ \\begin{array}{ll} H_0: & \\mu_1 = \\mu_2 \\\\ H_1: & \\mu_1 \\neq \\mu_2 \\end{array} \\]\n\nTo construct a hypothesis test that handles this scenario, we start by noting that if the null hypothesis is true, then the difference between the population means is exactly zero, \\(\\mu_1 - \\mu_2 = 0\\) As a consequence, a diagnostic test statistic will be based on the difference between the two sample means. Because if the null hypothesis is true, then we’d expect \\[ \\bar{X}_1 - \\bar{X}_2 \\] to be pretty close to zero. However, just like we saw with our one-sample tests (i.e., the one-sample \\(z\\)-test and the one-sample \\(t\\)-test) we have to be precise about exactly how close to zero this difference should be. And the solution to the problem is more or less the same one: we calculate a standard error estimate (SE), just like last time, and then divide the difference between means by this estimate. So our \\(t\\)-statistic will be of the form \\[ t = \\frac{\\bar{X}_1 - \\bar{X}_2}{\\mbox{SE}} \\] We just need to figure out what this standard error estimate actually is. This is a bit trickier than was the case for either of the two tests we’ve looked at so far, so we need to go through it a lot more carefully to understand how it works.\n\n### 13.3.3 A “pooled estimate” of the standard deviation\n\nIn the original “Student \\(t\\)-test”, we make the assumption that the two groups have the same population standard deviation: that is, regardless of whether the population means are the same, we assume that the population standard deviations are identical, \\(\\sigma_1 = \\sigma_2\\). Since we’re assuming that the two standard deviations are the same, we drop the subscripts and refer to both of them as \\(\\sigma\\). How should we estimate this? How should we construct a single estimate of a standard deviation when we have two samples? The answer is, basically, we average them. Well, sort of. Actually, what we do is take a weighed average of the variance estimates, which we use as our pooled estimate of the variance. The weight assigned to each sample is equal to the number of observations in that sample, minus 1. Mathematically, we can write this as \\[ \\begin{array}{rcl} w_1 &=& N_1 - 1\\\\ w_2 &=& N_2 - 1 \\end{array} \\] Now that we’ve assigned weights to each sample, we calculate the pooled estimate of the variance by taking the weighted average of the two variance estimates, \\({\\hat\\sigma_1}^2\\) and \\({\\hat\\sigma_2}^2\\) \\[ \\hat\\sigma^2_p = \\frac{w_1 {\\hat\\sigma_1}^2 + w_2 {\\hat\\sigma_2}^2}{w_1 + w_2} \\] Finally, we convert the pooled variance estimate to a pooled standard deviation estimate, by taking the square root. This gives us the following formula for \\(\\hat\\sigma_p\\), \\[ \\hat\\sigma_p = \\sqrt{\\frac{w_1 {\\hat\\sigma_1}^2 + w_2 {\\hat\\sigma_2}^2}{w_1 + w_2}} \\] And if you mentally substitute \\(w_1 = N_1 -1\\) and \\(w_2 = N_2 -1\\) into this equation you get a very ugly looking formula; a very ugly formula that actually seems to be the “standard” way of describing the pooled standard deviation estimate. It’s not my favourite way of thinking about pooled standard deviations, however.192\n\n### 13.3.4 The same pooled estimate, described differently\n\nI prefer to think about it like this. Our data set actually corresponds to a set of \\(N\\) observations, which are sorted into two groups. So let’s use the notation \\(X_{ik}\\) to refer to the grade received by the \\(i\\)-th student in the \\(k\\)-th tutorial group: that is, \\(X_{11}\\) is the grade received by the first student in Anastasia’s class, \\(X_{21}\\) is her second student, and so on. And we have two separate group means \\(\\bar{X}_1\\) and \\(\\bar{X}_2\\), which we could “generically” refer to using the notation \\(\\bar{X}_k\\), i.e., the mean grade for the \\(k\\)-th tutorial group. So far, so good. Now, since every single student falls into one of the two tutorials, and so we can describe their deviation from the group mean as the difference \\[ X_{ik} - \\bar{X}_k \\] So why not just use these deviations (i.e., the extent to which each student’s grade differs from the mean grade in their tutorial?) Remember, a variance is just the average of a bunch of squared deviations, so let’s do that. Mathematically, we could write it like this: \\[ \\frac{\\sum_{ik} \\left( X_{ik} - \\bar{X}_k \\right)^2}{N} \\] where the notation “\\(\\sum_{ik}\\)” is a lazy way of saying “calculate a sum by looking at all students in all tutorials”, since each “\\(ik\\)” corresponds to one student.193 But, as we saw in Chapter 10, calculating the variance by dividing by \\(N\\) produces a biased estimate of the population variance. And previously, we needed to divide by \\(N-1\\) to fix this. However, as I mentioned at the time, the reason why this bias exists is because the variance estimate relies on the sample mean; and to the extent that the sample mean isn’t equal to the population mean, it can systematically bias our estimate of the variance. But this time we’re relying on two sample means! Does this mean that we’ve got more bias? Yes, yes it does. And does this mean we now need to divide by \\(N-2\\) instead of \\(N-1\\), in order to calculate our pooled variance estimate? Why, yes… \\[ \\hat\\sigma^2_p = \\frac{\\sum_{ik} \\left( X_{ik} - \\bar{X}_k \\right)^2}{N -2} \\] Oh, and if you take the square root of this then you get \\(\\hat{\\sigma}_p\\), the pooled standard deviation estimate. In other words, the pooled standard deviation calculation is nothing special: it’s not terribly different to the regular standard deviation calculation.\n\n### 13.3.5 Completing the test\n\nRegardless of which way you want to think about it, we now have our pooled estimate of the standard deviation. From now on, I’ll drop the silly \\(p\\) subscript, and just refer to this estimate as \\(\\hat\\sigma\\). Great. Let’s now go back to thinking about the bloody hypothesis test, shall we? Our whole reason for calculating this pooled estimate was that we knew it would be helpful when calculating our standard error estimate. But, standard error of what? In the one-sample \\(t\\)-test, it was the standard error of the sample mean, \\(\\mbox{SE}({\\bar{X}})\\), and since \\(\\mbox{SE}({\\bar{X}}) = \\sigma / \\sqrt{N}\\) that’s what the denominator of our \\(t\\)-statistic looked like. This time around, however, we have two sample means. And what we’re interested in, specifically, is the the difference between the two \\(\\bar{X}_1 - \\bar{X}_2\\). As a consequence, the standard error that we need to divide by is in fact the standard error of the difference between means. As long as the two variables really do have the same standard deviation, then our estimate for the standard error is \\[ \\mbox{SE}({\\bar{X}_1 - \\bar{X}_2}) = \\hat\\sigma \\sqrt{\\frac{1}{N_1} + \\frac{1}{N_2}} \\] and our \\(t\\)-statistic is therefore \\[ t = \\frac{\\bar{X}_1 - \\bar{X}_2}{\\mbox{SE}({\\bar{X}_1 - \\bar{X}_2})} \\] Just as we saw with our one-sample test, the sampling distribution of this \\(t\\)-statistic is a \\(t\\)-distribution (shocking, isn’t it?) as long as the null hypothesis is true, and all of the assumptions of the test are met. The degrees of freedom, however, is slightly different. As usual, we can think of the degrees of freedom to be equal to the number of data points minus the number of constraints. In this case, we have \\(N\\) observations (\\(N_1\\) in sample 1, and \\(N_2\\) in sample 2), and 2 constraints (the sample means). So the total degrees of freedom for this test are \\(N-2\\).\n\n### 13.3.6 Doing the test in R\n\n Not surprisingly, you can run an independent samples \\(t\\)-test using the `t.test()` function (Section 13.7), but once again I’m going to start with a somewhat simpler function in the `lsr` package. That function is unimaginatively called \n\n. First, recall that our data look like this: `head( harpo )` \n\n The outcome variable for our test is the student `grade` , and the groups are defined in terms of the `tutor` for each class. So you probably won’t be too surprised to see that we’re going to describe the test that we want in terms of an R formula that reads like this `grade ~ tutor` . The specific command that we need is: \n\n```\nindependentSamplesTTest( \n formula = grade ~ tutor, # formula specifying outcome and group variables\n data = harpo, # data frame that contains the variables\n var.equal = TRUE # assume that the two groups have the same variance\n )\n```\n\n The first two arguments should be familiar to you. The first one is the formula that tells R what variables to use and the second one tells R the name of the data frame that stores those variables. The third argument is not so obvious. By saying `var.equal = TRUE` , what we’re really doing is telling R to use the Student independent samples \\(t\\)-test. More on this later. For now, lets ignore that bit and look at the output: \nThe output has a very familiar form. First, it tells you what test was run, and it tells you the names of the variables that you used. The second part of the output reports the sample means and standard deviations for both groups (i.e., both tutorial groups). The third section of the output states the null hypothesis and the alternative hypothesis in a fairly explicit form. It then reports the test results: just like last time, the test results consist of a \\(t\\)-statistic, the degrees of freedom, and the \\(p\\)-value. The final section reports two things: it gives you a confidence interval, and an effect size. I’ll talk about effect sizes later. The confidence interval, however, I should talk about now.\n It’s pretty important to be clear on what this confidence interval actually refers to: it is a confidence interval for the difference between the group means. In our example, Anastasia’s students had an average grade of 74.5, and Bernadette’s students had an average grade of 69.1, so the difference between the two sample means is 5.4. But of course the difference between population means might be bigger or smaller than this. The confidence interval reported by the\n\nfunction tells you that there’s a 95% chance that the true difference between means lies between 0.2 and 10.8. \nIn any case, the difference between the two groups is significant (just barely), so we might write up the result using text like this:\n The mean grade in Anastasia’s class was 74.5% (std dev = 9.0), whereas the mean in Bernadette’s class was 69.1% (std dev = 5.8). A Student’s independent samples \\(t\\)-test showed that this 5.4% difference was significant (\\(t(31) = 2.1\\), \\(p<.05\\), \\(CI_{95} = [0.2, 10.8]\\), \\(d = .74\\)), suggesting that a genuine difference in learning outcomes has occurred.\n\nNotice that I’ve included the confidence interval and the effect size in the stat block. People don’t always do this. At a bare minimum, you’d expect to see the \\(t\\)-statistic, the degrees of freedom and the \\(p\\) value. So you should include something like this at a minimum: \\(t(31) = 2.1\\), \\(p<.05\\). If statisticians had their way, everyone would also report the confidence interval and probably the effect size measure too, because they are useful things to know. But real life doesn’t always work the way statisticians want it to: you should make a judgment based on whether you think it will help your readers, and (if you’re writing a scientific paper) the editorial standard for the journal in question. Some journals expect you to report effect sizes, others don’t. Within some scientific communities it is standard practice to report confidence intervals, in other it is not. You’ll need to figure out what your audience expects. But, just for the sake of clarity, if you’re taking my class: my default position is that it’s usually worth includng the effect size, but don’t worry about the confidence interval unless the assignment asks you to or implies that you should.\n\n### 13.3.7 Positive and negative \\(t\\) values\n\nBefore moving on to talk about the assumptions of the \\(t\\)-test, there’s one additional point I want to make about the use of \\(t\\)-tests in practice. The first one relates to the sign of the \\(t\\)-statistic (that is, whether it is a positive number or a negative one). One very common worry that students have when they start running their first \\(t\\)-test is that they often end up with negative values for the \\(t\\)-statistic, and don’t know how to interpret it. In fact, it’s not at all uncommon for two people working independently to end up with R outputs that are almost identical, except that one person has a negative \\(t\\) values and the other one has a positive \\(t\\) value. Assuming that you’re running a two-sided test, then the \\(p\\)-values will be identical. On closer inspection, the students will notice that the confidence intervals also have the opposite signs. This is perfectly okay: whenever this happens, what you’ll find is that the two versions of the R output arise from slightly different ways of running the \\(t\\)-test. What’s happening here is very simple. The \\(t\\)-statistic that R is calculating here is always of the form \\[ t = \\frac{\\mbox{(mean 1)} -\\mbox{(mean 2)}}{ \\mbox{(SE)}} \\] If “mean 1” is larger than “mean 2” the \\(t\\) statistic will be positive, whereas if “mean 2” is larger then the \\(t\\) statistic will be negative. Similarly, the confidence interval that R reports is the confidence interval for the difference “(mean 1) minus (mean 2)”, which will be the reverse of what you’d get if you were calculating the confidence interval for the difference “(mean 2) minus (mean 1)”.\n\nOkay, that’s pretty straightforward when you think about it, but now consider our \\(t\\)-test comparing Anastasia’s class to Bernadette’s class. Which one should we call “mean 1” and which one should we call “mean 2”. It’s arbitrary. However, you really do need to designate one of them as “mean 1” and the other one as “mean 2”. Not surprisingly, the way that R handles this is also pretty arbitrary. In earlier versions of the book I used to try to explain it, but after a while I gave up, because it’s not really all that important, and to be honest I can never remember myself. Whenever I get a significant \\(t\\)-test result, and I want to figure out which mean is the larger one, I don’t try to figure it out by looking at the \\(t\\)-statistic. Why would I bother doing that? It’s foolish. It’s easier just look at the actual group means, since the R output actually shows them!\n\nHere’s the important thing. Because it really doesn’t matter what R printed out, I usually try to report the \\(t\\)-statistic in such a way that the numbers match up with the text. Here’s what I mean… suppose that what I want to write in my report is “Anastasia’s class had higher grades than Bernadette’s class”. The phrasing here implies that Anastasia’s group comes first, so it makes sense to report the \\(t\\)-statistic as if Anastasia’s class corresponded to group 1. If so, I would write\n (I wouldn’t actually emphasise the word “higher” in real life, I’m just doing it to emphasise the point that “higher” corresponds to positive \\(t\\) values). On the other hand, suppose the phrasing I wanted to use has Bernadette’s class listed first. If so, it makes more sense to treat her class as group 1, and if so, the write up looks like this:\n Because I’m talking about one group having “lower” scores this time around, it is more sensible to use the negative form of the \\(t\\)-statistic. It just makes it read more cleanly.\n\nOne last thing: please note that you can’t do this for other types of test statistics. It works for \\(t\\)-tests, but it wouldn’t be meaningful for chi-square testsm \\(F\\)-tests or indeed for most of the tests I talk about in this book. So don’t overgeneralise this advice! I’m really just talking about \\(t\\)-tests here and nothing else!\n\n### 13.3.8 Assumptions of the test\n\nAs always, our hypothesis test relies on some assumptions. So what are they? For the Student t-test there are three assumptions, some of which we saw previously in the context of the one sample \\(t\\)-test (see Section 13.2.3):\n\n* Normality. Like the one-sample \\(t\\)-test, it is assumed that the data are normally distributed. Specifically, we assume that both groups are normally distributed. In Section 13.9 we’ll discuss how to test for normality, and in Section 13.10 we’ll discuss possible solutions.\n * Independence. Once again, it is assumed that the observations are independently sampled. In the context of the Student test this has two aspects to it. Firstly, we assume that the observations within each sample are independent of one another (exactly the same as for the one-sample test). However, we also assume that there are no cross-sample dependencies. If, for instance, it turns out that you included some participants in both experimental conditions of your study (e.g., by accidentally allowing the same person to sign up to different conditions), then there are some cross sample dependencies that you’d need to take into account.\n * Homogeneity of variance (also called “homoscedasticity”). The third assumption is that the population standard deviation is the same in both groups. You can test this assumption using the Levene test, which I’ll talk about later on in the book (Section 14.7). However, there’s a very simple remedy for this assumption, which I’ll talk about in the next section.\n\n## 13.4 The independent samples \\(t\\)-test (Welch test)\n\nThe biggest problem with using the Student test in practice is the third assumption listed in the previous section: it assumes that both groups have the same standard deviation. This is rarely true in real life: if two samples don’t have the same means, why should we expect them to have the same standard deviation? There’s really no reason to expect this assumption to be true. We’ll talk a little bit about how you can check this assumption later on because it does crop up in a few different places, not just the \\(t\\)-test. But right now I’ll talk about a different form of the \\(t\\)-test (Welch 1947) that does not rely on this assumption. A graphical illustration of what the Welch \\(t\\) test assumes about the data is shown in Figure 13.10, to provide a contrast with the Student test version in Figure 13.9. I’ll admit it’s a bit odd to talk about the cure before talking about the diagnosis, but as it happens the Welch test is the default \\(t\\)-test in R, so this is probably the best place to discuss it.\n\nThe Welch test is very similar to the Student test. For example, the \\(t\\)-statistic that we use in the Welch test is calculated in much the same way as it is for the Student test. That is, we take the difference between the sample means, and then divide it by some estimate of the standard error of that difference: \\[ t = \\frac{\\bar{X}_1 - \\bar{X}_2}{\\mbox{SE}({\\bar{X}_1 - \\bar{X}_2})} \\] The main difference is that the standard error calculations are different. If the two populations have different standard deviations, then it’s a complete nonsense to try to calculate a pooled standard deviation estimate, because you’re averaging apples and oranges.194 But you can still estimate the standard error of the difference between sample means; it just ends up looking different: \\[ \\mbox{SE}({\\bar{X}_1 - \\bar{X}_2}) = \\sqrt{ \\frac{{\\hat{\\sigma}_1}^2}{N_1} + \\frac{{\\hat{\\sigma}_2}^2}{N_2} } \\] The reason why it’s calculated this way is beyond the scope of this book. What matters for our purposes is that the \\(t\\)-statistic that comes out of the Welch test is actually somewhat different to the one that comes from the Student test.\n\nThe second difference between Welch and Student is that the degrees of freedom are calculated in a very different way. In the Welch test, the “degrees of freedom” doesn’t have to be a whole number any more, and it doesn’t correspond all that closely to the “number of data points minus the number of constraints” heuristic that I’ve been using up to this point. The degrees of freedom are, in fact… \\[ \\mbox{df} = \\frac{ ({\\hat{\\sigma}_1}^2 / N_1 + {\\hat{\\sigma}_2}^2 / N_2)^2 }{ ({\\hat{\\sigma}_1}^2 / N_1)^2 / (N_1 -1 ) + ({\\hat{\\sigma}_2}^2 / N_2)^2 / (N_2 -1 ) } \\] … which is all pretty straightforward and obvious, right? Well, perhaps not. It doesn’t really matter for out purposes. What matters is that you’ll see that the “df” value that pops out of a Welch test tends to be a little bit smaller than the one used for the Student test, and it doesn’t have to be a whole number.\n\n### 13.4.1 Doing the test in R\n\n To run a Welch test in R is pretty easy. All you have to do is not bother telling R to assume equal variances. That is, you take the command we used to run a Student’s \\(t\\)-test and drop the `var.equal = TRUE` bit. So the command for a Welch test becomes: \n\n```\nindependentSamplesTTest( \n formula = grade ~ tutor, # formula specifying outcome and group variables\n data = harpo # data frame that contains the variables\n )\n```\n\nNot too difficult, right? Not surprisingly, the output has exactly the same format as it did last time too:\n\nThe very first line is different, because it’s telling you that its run a Welch test rather than a Student test, and of course all the numbers are a bit different. But I hope that the interpretation of this output should be fairly obvious. You read the output in the same way that you would for the Student test. You’ve got your descriptive statistics, the hypotheses, the test results and some other information. So that’s all pretty easy.\n\nExcept, except… our result isn’t significant anymore. When we ran the Student test, we did get a significant effect; but the Welch test on the same data set is not (\\(t(23.03) = 2.03\\), \\(p = .054\\)). What does this mean? Should we panic? Is the sky burning? Probably not. The fact that one test is significant and the other isn’t doesn’t itself mean very much, especially since I kind of rigged the data so that this would happen. As a general rule, it’s not a good idea to go out of your way to try to interpret or explain the difference between a \\(p\\)-value of .049 and a \\(p\\)-value of .051. If this sort of thing happens in real life, the difference in these \\(p\\)-values is almost certainly due to chance. What does matter is that you take a little bit of care in thinking about what test you use. The Student test and the Welch test have different strengths and weaknesses. If the two populations really do have equal variances, then the Student test is slightly more powerful (lower Type II error rate) than the Welch test. However, if they don’t have the same variances, then the assumptions of the Student test are violated and you may not be able to trust it: you might end up with a higher Type I error rate. So it’s a trade off. However, in real life, I tend to prefer the Welch test; because almost no-one actually believes that the population variances are identical.\n\n### 13.4.2 Assumptions of the test\n\nThe assumptions of the Welch test are very similar to those made by the Student \\(t\\)-test (see Section 13.3.8), except that the Welch test does not assume homogeneity of variance. This leaves only the assumption of normality, and the assumption of independence. The specifics of these assumptions are the same for the Welch test as for the Student test.\n\n## 13.5 The paired-samples \\(t\\)-test\n\nRegardless of whether we’re talking about the Student test or the Welch test, an independent samples \\(t\\)-test is intended to be used in a situation where you have two samples that are, well, independent of one another. This situation arises naturally when participants are assigned randomly to one of two experimental conditions, but it provides a very poor approximation to other sorts of research designs. In particular, a repeated measures design – in which each participant is measured (with respect to the same outcome variable) in both experimental conditions – is not suited for analysis using independent samples \\(t\\)-tests. For example, we might be interested in whether listening to music reduces people’s working memory capacity. To that end, we could measure each person’s working memory capacity in two conditions: with music, and without music. In an experimental design such as this one,195 each participant appears in both groups. This requires us to approach the problem in a different way; by using the paired samples \\(t\\)-test.\n\n### 13.5.1 The data\n\n The data set that we’ll use this time comes from Dr Chico’s class.196 In her class, students take two major tests, one early in the semester and one later in the semester. To hear her tell it, she runs a very hard class, one that most students find very challenging; but she argues that by setting hard assessments, students are encouraged to work harder. Her theory is that the first test is a bit of a “wake up call” for students: when they realise how hard her class really is, they’ll work harder for the second test and get a better mark. Is she right? To test this, let’s have a look at the `chico.Rdata` file: \n\n```\nload( file.path(projecthome, \"data/chico.Rdata\" ))\nstr(chico)\n```\n\n```\n## 'data.frame': 20 obs. of 3 variables:\n## $ id : Factor w/ 20 levels \"student1\",\"student10\",..: 1 12 14 15 16 17 18 19 20 2 ...\n## $ grade_test1: num 42.9 51.8 71.7 51.6 63.5 58 59.8 50.8 62.5 61.9 ...\n## $ grade_test2: num 44.6 54 72.3 53.4 63.8 59.3 60.8 51.6 64.3 63.2 ...\n```\n\n The data frame `chico` contains three variables: an `id` variable that identifies each student in the class, the `grade_test1` variable that records the student grade for the first test, and the `grade_test2` variable that has the grades for the second test. Here’s the first six students: `head( chico )` \n\nAt a glance, it does seem like the class is a hard one (most grades are between 50% and 60%), but it does look like there’s an improvement from the first test to the second one. If we take a quick look at the descriptive statistics\n\n```\nlibrary( psych )\ndescribe( chico )\n```\n\n```\n## vars n mean sd median trimmed mad min max range skew\n## id* 1 20 10.50 5.92 10.5 10.50 7.41 1.0 20.0 19.0 0.00\n## grade_test1 2 20 56.98 6.62 57.7 56.92 7.71 42.9 71.7 28.8 0.05\n## grade_test2 3 20 58.38 6.41 59.7 58.35 6.45 44.6 72.3 27.7 -0.05\n## kurtosis se\n## id* -1.38 1.32\n## grade_test1 -0.35 1.48\n## grade_test2 -0.39 1.43\n```\n\n we see that this impression seems to be supported. Across all 20 students197 the mean grade for the first test is 57%, but this rises to 58% for the second test. Although, given that the standard deviations are 6.6% and 6.4% respectively, it’s starting to feel like maybe the improvement is just illusory; maybe just random variation. This impression is reinforced when you see the means and confidence intervals plotted in Figure 13.11. If we were to rely on this plot alone, we’d come to the same conclusion that we got from looking at the descriptive statistics that the `describe()` function produced. Looking at how wide those confidence intervals are, we’d be tempted to think that the apparent improvement in student performance is pure chance. \nNevertheless, this impression is wrong. To see why, take a look at the scatterplot of the grades for test 1 against the grades for test 2. shown in Figure 13.12.\n In this plot, each dot corresponds to the two grades for a given student: if their grade for test 1 (\\(x\\) co-ordinate) equals their grade for test 2 (\\(y\\) co-ordinate), then the dot falls on the line. Points falling above the line are the students that performed better on the second test. Critically, almost all of the data points fall above the diagonal line: almost all of the students do seem to have improved their grade, if only by a small amount. This suggests that we should be looking at the improvement made by each student from one test to the next, and treating that as our raw data. To do this, we’ll need to create a new variable for the `improvement` that each student makes, and add it to the `chico` data frame. The easiest way to do this is as follows: \n\n```\nchico$improvement <- chico$grade_test2 - chico$grade_test1\n```\n\n Notice that I assigned the output to a variable called `chico$improvement` . That has the effect of creating a new variable called `improvement` inside the `chico` data frame. So now when I look at the `chico` data frame, I get an output that looks like this: `head( chico )` \n\n Now that we’ve created and stored this `improvement` variable, we can draw a histogram showing the distribution of these improvement scores (using the `hist()` function), shown in Figure 13.13. When we look at histogram, it’s very clear that there is a real improvement here. The vast majority of the students scored higher on the test 2 than on test 1, reflected in the fact that almost the entire histogram is above zero. In fact, if we use `ciMean()` to compute a confidence interval for the population mean of this new variable, \n\n```\nciMean( x = chico$improvement )\n```\n\n```\n## 2.5% 97.5%\n## [1,] 0.9508686 1.859131\n```\n\nwe see that it is 95% certain that the true (population-wide) average improvement would lie between 0.95% and 1.86%. So you can see, qualitatively, what’s going on: there is a real “within student” improvement (everyone improves by about 1%), but it is very small when set against the quite large “between student” differences (student grades vary by about 20% or so).\n\n### 13.5.2 What is the paired samples \\(t\\)-test?\n\n In light of the previous exploration, let’s think about how to construct an appropriate \\(t\\) test. One possibility would be to try to run an independent samples \\(t\\)-test using `grade_test1` and `grade_test2` as the variables of interest. However, this is clearly the wrong thing to do: the independent samples \\(t\\)-test assumes that there is no particular relationship between the two samples. Yet clearly that’s not true in this case, because of the repeated measures structure to the data. To use the language that I introduced in the last section, if we were to try to do an independent samples \\(t\\)-test, we would be conflating the within subject differences (which is what we’re interested in testing) with the between subject variability (which we are not). The solution to the problem is obvious, I hope, since we already did all the hard work in the previous section. Instead of running an independent samples \\(t\\)-test on `grade_test1` and `grade_test2` , we run a one-sample \\(t\\)-test on the within-subject difference variable, `improvement` . To formalise this slightly, if \\(X_{i1}\\) is the score that the \\(i\\)-th participant obtained on the first variable, and \\(X_{i2}\\) is the score that the same person obtained on the second one, then the difference score is: \\[\nD_{i} = X_{i1} - X_{i2} \n\\] Notice that the difference scores is variable 1 minus variable 2 and not the other way around, so if we want improvement to correspond to a positive valued difference, we actually want “test 2” to be our “variable 1”. Equally, we would say that \\(\\mu_D = \\mu_1 - \\mu_2\\) is the population mean for this difference variable. So, to convert this to a hypothesis test, our null hypothesis is that this mean difference is zero; the alternative hypothesis is that it is not: \\[\n\\begin{array}{ll}\nH_0: & \\mu_D = 0 \\\\\nH_1: & \\mu_D \\neq 0\n\\end{array}\n\\] (this is assuming we’re talking about a two-sided test here). This is more or less identical to the way we described the hypotheses for the one-sample \\(t\\)-test: the only difference is that the specific value that the null hypothesis predicts is 0. And so our \\(t\\)-statistic is defined in more or less the same way too. If we let \\(\\bar{D}\\) denote the mean of the difference scores, then \\[\nt = \\frac{\\bar{D}}{\\mbox{SE}({\\bar{D}})}\n\\] which is \\[\nt = \\frac{\\bar{D}}{\\hat\\sigma_D / \\sqrt{N}}\n\\] where \\(\\hat\\sigma_D\\) is the standard deviation of the difference scores. Since this is just an ordinary, one-sample \\(t\\)-test, with nothing special about it, the degrees of freedom are still \\(N-1\\). And that’s it: the paired samples \\(t\\)-test really isn’t a new test at all: it’s a one-sample \\(t\\)-test, but applied to the difference between two variables. It’s actually very simple; the only reason it merits a discussion as long as the one we’ve just gone through is that you need to be able to recognise when a paired samples test is appropriate, and to understand why it’s better than an independent samples \\(t\\) test. \n\n### 13.5.3 Doing the test in R, part 1\n\n How do you do a paired samples \\(t\\)-test in R. One possibility is to follow the process I outlined above: create a “difference” variable and then run a one sample \\(t\\)-test on that. Since we’ve already created a variable called `chico$improvement` , let’s do that: \n\n```\noneSampleTTest( chico$improvement, mu=0 )\n```\n\n```\n## \n## One sample t-test \n## \n## Data variable: chico$improvement \n## \n## Descriptive statistics: \n## improvement\n## mean 1.405\n## std dev. 0.970\n## \n## Hypotheses: \n## null: population mean equals 0 \n## alternative: population mean not equal to 0 \n## \n## Test results: \n## t-statistic: 6.475 \n## degrees of freedom: 19 \n## p-value: <.001 \n## \n## Other information: \n## two-sided 95% confidence interval: [0.951, 1.859] \n## estimated effect size (Cohen's d): 1.448\n```\n\n The output here is (obviously) formatted exactly the same was as it was the last time we used the `oneSampleTTest()` function (Section 13.2), and it confirms our intuition. There’s an average improvement of 1.4% from test 1 to test 2, and this is significantly different from 0 (\\(t(19)=6.48, p<.001\\)). However, suppose you’re lazy and you don’t want to go to all the effort of creating a new variable. Or perhaps you just want to keep the difference between one-sample and paired-samples tests clear in your head. If so, you can use the `pairedSamplesTTest()` function, also in the `lsr` package. Let’s assume that your data organised like they are in the `chico` data frame, where there are two separate variables, one for each measurement. The way to run the test is to input a one-sided formula, just like you did when running a test of association using the `associationTest()` function in Chapter 12. For the `chico` data frame, the formula that you need would be \n\n```\n~ grade_time2 + grade_time1\n```\n\n. As usual, you’ll also need to input the name of the data frame too. So the command just looks like this: \n\n```\npairedSamplesTTest( \n formula = ~ grade_test2 + grade_test1, # one-sided formula listing the two variables\n data = chico # data frame containing the two variables \n )\n```\n\n```\n## \n## Paired samples t-test \n## \n## Variables: grade_test2 , grade_test1 \n## \n## Descriptive statistics: \n## grade_test2 grade_test1 difference\n## mean 58.385 56.980 1.405\n## std dev. 6.406 6.616 0.970\n## \n## Hypotheses: \n## null: population means equal for both measurements\n## alternative: different population means for each measurement\n## \n## Test results: \n## t-statistic: 6.475 \n## degrees of freedom: 19 \n## p-value: <.001 \n## \n## Other information: \n## two-sided 95% confidence interval: [0.951, 1.859] \n## estimated effect size (Cohen's d): 1.448\n```\n\nThe numbers are identical to those that come from the one sample test, which of course they have to be given that the paired samples \\(t\\)-test is just a one sample test under the hood. However, the output is a bit more detailed:\n\nThis time around the descriptive statistics block shows you the means and standard deviations for the original variables, as well as for the difference variable (notice that it always defines the difference as the first listed variable mines the second listed one). The null hypothesis and the alternative hypothesis are now framed in terms of the original variables rather than the difference score, but you should keep in mind that in a paired samples test it’s still the difference score being tested. The statistical information at the bottom about the test result is of course the same as before.\n\n### 13.5.4 Doing the test in R, part 2\n\n The paired samples \\(t\\)-test is a little different from the other \\(t\\)-tests, because it is used in repeated measures designs. For the `chico` data, every student is “measured” twice, once for the first test, and again for the second test. Back in Section 7.7 I talked about the fact that repeated measures data can be expressed in two standard ways, known as wide form and long form. The `chico` data frame is in wide form: every row corresponds to a unique person. I’ve shown you the data in that form first because that’s the form that you’re most used to seeing, and it’s also the format that you’re most likely to receive data in. However, the majority of tools in R for dealing with repeated measures data expect to receive data in long form. The paired samples \\(t\\)-test is a bit of an exception that way. As you make the transition from a novice user to an advanced one, you’re going to have to get comfortable with long form data, and switching between the two forms. To that end, I want to show you how to apply the `pairedSamplesTTest()` function to long form data. First, let’s use the `wideToLong()` function to create a long form version of the `chico` data frame. If you’ve forgotten how the `wideToLong()` function works, it might be worth your while quickly re-reading Section 7.7. Assuming that you’ve done so, or that you’re already comfortable with data reshaping, I’ll use it to create a new data frame called `chico2` : \n\n```\nchico2 <- wideToLong( chico, within=\"time\" )\nhead( chico2 )\n```\n\n As you can see, this has created a new data frame containing three variables: an `id` variable indicating which person provided the data, a `time` variable indicating which test the data refers to (i.e., test 1 or test 2), and a `grade` variable that records what score the person got on that test. Notice that this data frame is in long form: every row corresponds to a unique measurement. Because every person provides two observations (test 1 and test 2), there are two rows for every person. To see this a little more clearly, I’ll use the `sortFrame()` function to sort the rows of `chico2` by `id` variable (see Section 7.6.3). \n\n```\nchico2 <- sortFrame( chico2, id )\nhead( chico2 )\n```\n\n```\n## id improvement time grade\n## 1 student1 1.7 test1 42.9\n## 21 student1 1.7 test2 44.6\n## 10 student10 1.3 test1 61.9\n## 30 student10 1.3 test2 63.2\n## 11 student11 1.4 test1 50.4\n## 31 student11 1.4 test2 51.8\n```\n\nAs you can see, there are two rows for “student1”: one showing their grade on the first test, the other showing their grade on the second test.198\n Okay, suppose that we were given the `chico2` data frame to analyse. How would we run our paired samples \\(t\\)-test now? One possibility would be to use the `longToWide()` function (Section 7.7) to force the data back into wide form, and do the same thing that we did previously. But that’s sort of defeating the point, and besides, there’s an easier way. Let’s think about what how the `chico2` data frame is structured: there are three variables here, and they all matter. The outcome measure is stored as the `grade` , and we effectively have two “groups” of measurements (test 1 and test 2) that are defined by the `time` points at which a test is given. Finally, because we want to keep track of which measurements should be paired together, we need to know which student obtained each grade, which is what the `id` variable gives us. So, when your data are presented to you in long form, we would want specify a two-sided formula and a data frame, in the same way that we do for an independent samples \\(t\\)-test: the formula specifies the outcome variable and the groups, so in this case it would be `grade ~ time` , and the data frame is `chico2` . However, we also need to tell it the id variable, which in this case is boringly called `id` . So our command is: \n\n```\npairedSamplesTTest( \n formula = grade ~ time, # two sided formula: outcome ~ group\n data = chico2, # data frame\n id = \"id\" # name of the id variable\n )\n```\n\n```\n## \n## Paired samples t-test \n## \n## Outcome variable: grade \n## Grouping variable: time \n## ID variable: id \n## \n## Descriptive statistics: \n## test1 test2 difference\n## mean 56.980 58.385 -1.405\n## std dev. 6.616 6.406 0.970\n## \n## Hypotheses: \n## null: population means equal for both measurements\n## alternative: different population means for each measurement\n## \n## Test results: \n## t-statistic: -6.475 \n## degrees of freedom: 19 \n## p-value: <.001 \n## \n## Other information: \n## two-sided 95% confidence interval: [-1.859, -0.951] \n## estimated effect size (Cohen's d): 1.448\n```\n\n Note that the name of the id variable is `\"id\"` and not `id` . Note that the `id` variable must be a factor. As of the current writing, you do need to include the quote marks, because the `pairedSamplesTTest()` function is expecting a character string that specifies the name of a variable. If I ever find the time I’ll try to relax this constraint. As you can see, it’s a bit more detailed than the output from `oneSampleTTest()` . It gives you the descriptive statistics for the original variables, states the null hypothesis in a fashion that is a bit more appropriate for a repeated measures design, and then reports all the nuts and bolts from the hypothesis test itself. Not surprisingly the numbers the same as the ones that we saw last time. One final comment about the `pairedSamplesTTest()` function. One of the reasons I designed it to be able handle long form and wide form data is that I want you to be get comfortable thinking about repeated measures data in both formats, and also to become familiar with the different ways in which R functions tend to specify models and tests for repeated measures data. With that last point in mind, I want to highlight a slightly different way of thinking about what the paired samples \\(t\\)-test is doing. There’s a sense in which what you’re really trying to do is look at how the outcome variable ( `grade` ) is related to the grouping variable ( `time` ), after taking account of the fact that there are individual differences between people ( `id` ). So there’s a sense in which `id` is actually a second predictor: you’re trying to predict the `grade` on the basis of the `time` and the `id` . With that in mind, the `pairedSamplesTTest()` function lets you specify a formula like this one `grade ~ time + (id)` This formula tells R everything it needs to know: the variable on the left ( `grade` ) is the outcome variable, the bracketed term on the right ( `id` ) is the id variable, and the other term on the right is the grouping variable ( `time` ). If you specify your formula that way, then you only need to specify the `formula` and the `data` frame, and so you can get away with using a command as simple as this one: \n\n```\npairedSamplesTTest( \n formula = grade ~ time + (id),\n data = chico2\n )\n```\n\nor you can drop the argument names and just do this:\n\n```\n> pairedSamplesTTest( grade ~ time + (id), chico2 )\n```\n\n These commands will produce the same output as the last one, I personally find this format a lot more elegant. That being said, the main reason for allowing you to write your formulas that way is that they’re quite similar to the way that mixed models (fancy pants repeated measures analyses) are specified in the `lme4` package. This book doesn’t talk about mixed models (yet!), but if you go on to learn more statistics you’ll find them pretty hard to avoid, so I’ve tried to lay a little bit of the groundwork here. \n\n## 13.6 One sided tests\n\n When introducing the theory of null hypothesis tests, I mentioned that there are some situations when it’s appropriate to specify a one-sided test (see Section 11.4.3). So far, all of the \\(t\\)-tests have been two-sided tests. For instance, when we specified a one sample \\(t\\)-test for the grades in Dr Zeppo’s class, the null hypothesis was that the true mean was 67.5%. The alternative hypothesis was that the true mean was greater than or less than 67.5%. Suppose we were only interested in finding out if the true mean is greater than 67.5%, and have no interest whatsoever in testing to find out if the true mean is lower than 67.5%. If so, our null hypothesis would be that the true mean is 67.5% or less, and the alternative hypothesis would be that the true mean is greater than 67.5%. The `oneSampleTTest()` function lets you do this, by specifying the `one.sided` argument. If you set `one.sided=\"greater\"` , it means that you’re testing to see if the true mean is larger than `mu` . If you set `one.sided=\"less\"` , then you’re testing to see if the true mean is smaller than `mu` . Here’s how it would work for Dr Zeppo’s class: \n\n```\noneSampleTTest( x=grades, mu=67.5, one.sided=\"greater\" )\n```\n\n```\n## \n## One sample t-test \n## \n## Data variable: grades \n## \n## Descriptive statistics: \n## grades\n## mean 72.300\n## std dev. 9.521\n## \n## Hypotheses: \n## null: population mean less than or equal to 67.5 \n## alternative: population mean greater than 67.5 \n## \n## Test results: \n## t-statistic: 2.255 \n## degrees of freedom: 19 \n## p-value: 0.018 \n## \n## Other information: \n## one-sided 95% confidence interval: [68.619, Inf] \n## estimated effect size (Cohen's d): 0.504\n```\n\n Notice that there are a few changes from the output that we saw last time. Most important is the fact that the null and alternative hypotheses have changed, to reflect the different test. The second thing to note is that, although the \\(t\\)-statistic and degrees of freedom have not changed, the \\(p\\)-value has. This is because the one-sided test has a different rejection region from the two-sided test. If you’ve forgotten why this is and what it means, you may find it helpful to read back over Chapter 11, and Section 11.4.3 in particular. The third thing to note is that the confidence interval is different too: it now reports a “one-sided” confidence interval rather than a two-sided one. In a two-sided confidence interval, we’re trying to find numbers \\(a\\) and \\(b\\) such that we’re 95% confident that the true mean lies between \\(a\\) and \\(b\\). In a one-sided confidence interval, we’re trying to find a single number \\(a\\) such that we’re 95% confident that the true mean is greater than \\(a\\) (or less than \\(a\\) if you set `one.sided=\"less\"` ). So that’s how to do a one-sided one sample \\(t\\)-test. However, all versions of the \\(t\\)-test can be one-sided. For an independent samples \\(t\\) test, you could have a one-sided test if you’re only interestd in testing to see if group A has higher scores than group B, but have no interest in finding out if group B has higher scores than group A. Let’s suppose that, for Dr Harpo’s class, you wanted to see if Anastasia’s students had higher grades than Bernadette’s. The\n\nfunction lets you do this, again by specifying the `one.sided` argument. However, this time around you need to specify the name of the group that you’re expecting to have the higher score. In our case, we’d write \n\n```\none.sided = \"Anastasia\"\n```\n\n. So the command would be: \n\n```\nindependentSamplesTTest( \n formula = grade ~ tutor, \n data = harpo, \n one.sided = \"Anastasia\"\n )\n```\n\n```\n## \n## Welch's independent samples t-test \n## \n## Outcome variable: grade \n## Grouping variable: tutor \n## \n## Descriptive statistics: \n## \n## mean 74.533 69.056\n## std dev. 8.999 5.775\n## \n## Hypotheses: \n## null: population means are equal, or smaller for group 'Anastasia' \n## alternative: population mean is larger for group 'Anastasia' \n## \n## Test results: \n## t-statistic: 2.034 \n## degrees of freedom: 23.025 \n## p-value: 0.027 \n## \n## Other information: \n## one-sided 95% confidence interval: [0.863, Inf] \n## estimated effect size (Cohen's d): 0.724\n```\n\nAgain, the output changes in a predictable way. The definition of the null and alternative hypotheses has changed, the \\(p\\)-value has changed, and it now reports a one-sided confidence interval rather than a two-sided one.\n What about the paired samples \\(t\\)-test? Suppose we wanted to test the hypothesis that grades go up from test 1 to test 2 in Dr Zeppo’s class, and are not prepared to consider the idea that the grades go down. Again, we can use the `one.sided` argument to specify the one-sided test, and it works the same way it does for the independent samples \\(t\\)-test. You need to specify the name of the group whose scores are expected to be larger under the alternative hypothesis. If your data are in wide form, as they are in the `chico` data frame, you’d use this command: \n\n```\npairedSamplesTTest( \n formula = ~ grade_test2 + grade_test1, \n data = chico, \n one.sided = \"grade_test2\" \n )\n```\n\n```\n## \n## Paired samples t-test \n## \n## Variables: grade_test2 , grade_test1 \n## \n## Descriptive statistics: \n## grade_test2 grade_test1 difference\n## mean 58.385 56.980 1.405\n## std dev. 6.406 6.616 0.970\n## \n## Hypotheses: \n## null: population means are equal, or smaller for measurement 'grade_test2' \n## alternative: population mean is larger for measurement 'grade_test2' \n## \n## Test results: \n## t-statistic: 6.475 \n## degrees of freedom: 19 \n## p-value: <.001 \n## \n## Other information: \n## one-sided 95% confidence interval: [1.03, Inf] \n## estimated effect size (Cohen's d): 1.448\n```\n\n Yet again, the output changes in a predictable way. The hypotheses have changed, the \\(p\\)-value has changed, and the confidence interval is now one-sided. If your data are in long form, as they are in the `chico2` data frame, it still works the same way. Either of the following commands would work, \n\n```\n> pairedSamplesTTest( \n formula = grade ~ time, \n data = chico2, \n id = \"id\", \n one.sided = \"test2\" \n )\n\n> pairedSamplesTTest( \n formula = grade ~ time + (id), \n data = chico2, \n one.sided = \"test2\" \n )\n```\n\nand would produce the same answer as the output shown above.\n\n## 13.7 Using the t.test() function\n\n In this chapter, we’ve talked about three different kinds of \\(t\\)-test: the one sample test, the independent samples test (Student’s and Welch’s), and the paired samples test. In order to run these different tests, I’ve shown you three different functions: `oneSampleTTest()` , \n\nand `pairedSamplesTTest()` . I wrote these as three different functions for two reasons. Firstly, I thought it made sense to have separate functions for each test, in order to help make it clear to beginners that there are different tests. Secondly, I wanted to show you some functions that produced “verbose” output, to help you see what hypotheses are being tested and so on. However, once you’ve started to become familiar with \\(t\\)-tests and with using R, you might find it easier to use the `t.test()` function. It’s one function, but it can run all four of the different \\(t\\)-tests that we’ve talked about. Here’s how it works. Firstly, suppose you want to run a one sample \\(t\\)-test. To run the test on the `grades` data from Dr Zeppo’s class (Section 13.2), we’d use a command like this: \n\n```\nt.test( x = grades, mu = 67.5 )\n```\n\n```\n## \n## One Sample t-test\n## \n## data: grades\n## t = 2.2547, df = 19, p-value = 0.03615\n## alternative hypothesis: true mean is not equal to 67.5\n## 95 percent confidence interval:\n## 67.84422 76.75578\n## sample estimates:\n## mean of x \n## 72.3\n```\n\n The input is the same as for the `oneSampleTTest()` : we specify the sample data using the argument `x` , and the value against which it is to be tested using the argument `mu` . The output is a lot more compressed. \nAs you can see, it still has all the information you need. It tells you what type of test it ran and the data it tested it on. It gives you the \\(t\\)-statistic, the degrees of freedom and the \\(p\\)-value. And so on. There’s nothing wrong with this output, but in my experience it can be a little confusing when you’re just starting to learn statistics, because it’s a little disorganised. Once you know what you’re looking at though, it’s pretty easy to read off the relevant information.\n What about independent samples \\(t\\)-tests? As it happens, the `t.test()` function can be used in much the same way as the \n\nfunction, by specifying a formula, a data frame, and using `var.equal` to indicate whether you want a Student test or a Welch test. If you want to run the Welch test from Section 13.4, then you’d use this command: \n\n```\n## \n## Welch Two Sample t-test\n## \n## data: grade by tutor\n## t = 2.0342, df = 23.025, p-value = 0.05361\n## alternative hypothesis: true difference in means is not equal to 0\n## 95 percent confidence interval:\n## -0.09249349 11.04804904\n## sample estimates:\n## mean in group Anastasia mean in group Bernadette \n## 74.53333 69.05556\n```\n\n If you want to do the Student test, it’s exactly the same except that you need to add an additional argument indicating that `var.equal = TRUE` . This is no different to how it worked in the \n\nfunction. Finally, we come to the paired samples \\(t\\)-test. Somewhat surprisingly, given that most R functions for dealing with repeated measures data require data to be in long form, the `t.test()` function isn’t really set up to handle data in long form. Instead it expects to be given two separate variables, `x` and `y` , and you need to specify `paired=TRUE` . And on top of that, you’d better make sure that the first element of `x` and the first element of `y` actually correspond to the same person! Because it doesn’t ask for an “id” variable. I don’t know why. So, in order to run the paired samples \\(t\\) test on the data from Dr Chico’s class, we’d use this command: \n\n```\nt.test( x = chico$grade_test2, # variable 1 is the \"test2\" scores\n y = chico$grade_test1, # variable 2 is the \"test1\" scores\n paired = TRUE # paired test\n )\n```\n\n```\n## \n## Paired t-test\n## \n## data: chico$grade_test2 and chico$grade_test1\n## t = 6.4754, df = 19, p-value = 3.321e-06\n## alternative hypothesis: true difference in means is not equal to 0\n## 95 percent confidence interval:\n## 0.9508686 1.8591314\n## sample estimates:\n## mean of the differences \n## 1.405\n```\n\nYet again, these are the same numbers that we saw in Section 13.5. Feel free to check.\n\n## 13.8 Effect size\n\n The most commonly used measure of effect size for a \\(t\\)-test is Cohen’s \\(d\\) (Cohen 1988). It’s a very simple measure in principle, with quite a few wrinkles when you start digging into the details. Cohen himself defined it primarily in the context of an independent samples \\(t\\)-test, specifically the Student test. In that context, a natural way of defining the effect size is to divide the difference between the means by an estimate of the standard deviation. In other words, we’re looking to calculate something along the lines of this: \\[ d = \\frac{\\mbox{(mean 1)} - \\mbox{(mean 2)}}{\\mbox{std dev}} \\] and he suggested a rough guide for interpreting \\(d\\) in Table ??. You’d think that this would be pretty unambiguous, but it’s not; largely because Cohen wasn’t too specific on what he thought should be used as the measure of the standard deviation (in his defence, he was trying to make a broader point in his book, not nitpick about tiny details). As discussed by McGrath and Meyer (2006), there are several different version in common usage, and each author tends to adopt slightly different notation. For the sake of simplicity (as opposed to accuracy) I’ll use \\(d\\) to refer to any statistic that you calculate from the sample, and use \\(\\delta\\) to refer to a theoretical population effect. Obviously, that does mean that there are several different things all called \\(d\\). The `cohensD()` function in the `lsr` package uses the `method` argument to distinguish between them, so that’s what I’ll do in the text. My suspicion is that the only time that you would want Cohen’s \\(d\\) is when you’re running a \\(t\\)-test, and if you’re using the `oneSampleTTest` , \n\n```\nindependentSamplesTTest\n```\n\nand `pairedSamplesTTest()` functions to run your \\(t\\)-tests, then you don’t need to learn any new commands, because they automatically produce an estimate of Cohen’s \\(d\\) as part of the output. However, if you’re using `t.test()` then you’ll need to use the `cohensD()` function (also in the `lsr` package) to do the calculations. \n\n\\(d\\)-value | rough interpretation |\n | --- | --- |\nabout 0.2 | small effect |\nabout 0.5 | moderate effect |\nabout 0.8 | large effect |\n\n### 13.8.1 Cohen’s \\(d\\) from one sample\n\n The simplest situation to consider is the one corresponding to a one-sample \\(t\\)-test. In this case, the one sample mean \\(\\bar{X}\\) and one (hypothesised) population mean \\(\\mu_o\\) to compare it to. Not only that, there’s really only one sensible way to estimate the population standard deviation: we just use our usual estimate \\(\\hat{\\sigma}\\). Therefore, we end up with the following as the only way to calculate \\(d\\), \\[ d = \\frac{\\bar{X} - \\mu_0}{\\hat{\\sigma}} \\] When writing the `cohensD()` function, I’ve made some attempt to make it work in a similar way to `t.test()` . As a consequence, `cohensD()` can calculate your effect size regardless of which type of \\(t\\)-test you performed. If what you want is a measure of Cohen’s \\(d\\) to accompany a one-sample \\(t\\)-test, there’s only two arguments that you need to care about. These are: \n\n* \n `x` . A numeric vector containing the sample data. * \n `mu` . The mean against which the mean of `x` is compared (default value is `mu = 0` ). We don’t need to specify what `method` to use, because there’s only one version of \\(d\\) that makes sense in this context. So, in order to compute an effect size for the data from Dr Zeppo’s class (Section 13.2), we’d type something like this: \n\n```\ncohensD( x = grades, # data are stored in the grades vector\n mu = 67.5 # compare students to a mean of 67.5\n )\n```\n\n `## [1] 0.5041691` and, just so that you can see that there’s nothing fancy going on, the command below shows you how to calculate it if there weren’t no fancypants `cohensD()` function available: \n\n```\n( mean(grades) - 67.5 ) / sd(grades)\n```\n\n `## [1] 0.5041691` \nYep, same number. Overall, then, the psychology students in Dr Zeppo’s class are achieving grades (mean = 72.3%) that are about .5 standard deviations higher than the level that you’d expect (67.5%) if they were performing at the same level as other students. Judged against Cohen’s rough guide, this is a moderate effect size.\n\n### 13.8.2 Cohen’s \\(d\\) from a Student \\(t\\) test\n\n The majority of discussions of Cohen’s \\(d\\) focus on a situation that is analogous to Student’s independent samples \\(t\\) test, and it’s in this context that the story becomes messier, since there are several different versions of \\(d\\) that you might want to use in this situation, and you can use the `method` argument to the `cohensD()` function to pick the one you want. To understand why there are multiple versions of \\(d\\), it helps to take the time to write down a formula that corresponds to the true population effect size \\(\\delta\\). It’s pretty straightforward, \\[\n\\delta = \\frac{\\mu_1 - \\mu_2}{\\sigma}\n\\] where, as usual, \\(\\mu_1\\) and \\(\\mu_2\\) are the population means corresponding to group 1 and group 2 respectively, and \\(\\sigma\\) is the standard deviation (the same for both populations). The obvious way to estimate \\(\\delta\\) is to do exactly the same thing that we did in the \\(t\\)-test itself: use the sample means as the top line, and a pooled standard deviation estimate for the bottom line: \\[\nd = \\frac{\\bar{X}_1 - \\bar{X}_2}{\\hat{\\sigma}_p}\n\\] where \\(\\hat\\sigma_p\\) is the exact same pooled standard deviation measure that appears in the \\(t\\)-test. This is the most commonly used version of Cohen’s \\(d\\) when applied to the outcome of a Student \\(t\\)-test ,and is sometimes referred to as Hedges’ \\(g\\) statistic (Hedges 1981). It corresponds to `method = \"pooled\"` in the `cohensD()` function, and it’s the default. However, there are other possibilities, which I’ll briefly describe. Firstly, you may have reason to want to use only one of the two groups as the basis for calculating the standard deviation. This approach (often called Glass’ \\(\\Delta\\)) only makes most sense when you have good reason to treat one of the two groups as a purer reflection of “natural variation” than the other. This can happen if, for instance, one of the two groups is a control group. If that’s what you want, then use `method = \"x.sd\"` or `method = \"y.sd\"` when using `cohensD()` . Secondly, recall that in the usual calculation of the pooled standard deviation we divide by \\(N-2\\) to correct for the bias in the sample variance; in one version of Cohen’s \\(d\\) this correction is omitted. Instead, we divide by \\(N\\). This version ( `method = \"raw\"` ) makes sense primarily when you’re trying to calculate the effect size in the sample; rather than estimating an effect size in the population. Finally, there is a version based on Hedges and Olkin (1985), who point out there is a small bias in the usual (pooled) estimation for Cohen’s \\(d\\). Thus they introduce a small correction ( `method = \"corrected\"` ), by multiplying the usual value of \\(d\\) by \\((N-3)/(N-2.25)\\). In any case, ignoring all those variations that you could make use of if you wanted, let’s have a look at how to calculate the default version. In particular, suppose we look at the data from Dr Harpo’s class (the `harpo` data frame). The command that we want to use is very similar to the relevant `t.test()` command, but also specifies a `method` \n\n```\ncohensD( formula = grade ~ tutor, # outcome ~ group\n data = harpo, # data frame \n method = \"pooled\" # which version to calculate?\n)\n```\n\n `## [1] 0.7395614` This is the version of Cohen’s \\(d\\) that gets reported by the\n\nfunction whenever it runs a Student \\(t\\)-test. \n\n### 13.8.3 Cohen’s \\(d\\) from a Welch test\n\n Suppose the situation you’re in is more like the Welch test: you still have two independent samples, but you no longer believe that the corresponding populations have equal variances. When this happens, we have to redefine what we mean by the population effect size. I’ll refer to this new measure as \\(\\delta^\\prime\\), so as to keep it distinct from the measure \\(\\delta\\) which we defined previously. What Cohen (1988) suggests is that we could define our new population effect size by averaging the two population variances. What this means is that we get: \\[ \\delta^\\prime = \\frac{\\mu_1 - \\mu_2}{\\sigma^\\prime} \\] where \\[ \\sigma^\\prime = \\sqrt{\\displaystyle{\\frac{ {\\sigma_1}^2 + {\\sigma_2}^2}{2}}} \\] This seems quite reasonable, but notice that none of the measures that we’ve discussed so far are attempting to estimate this new quantity. It might just be my own ignorance of the topic, but I’m only aware of one version of Cohen’s \\(d\\) that actually estimates the unequal-variance effect size \\(\\delta^\\prime\\) rather than the equal-variance effect size \\(\\delta\\). All we do to calculate \\(d\\) for this version ( `method = \"unequal\"` ) is substitute the sample means \\(\\bar{X}_1\\) and \\(\\bar{X}_2\\) and the corrected sample standard deviations \\(\\hat{\\sigma}_1\\) and \\(\\hat{\\sigma}_2\\) into the equation for \\(\\delta^\\prime\\). This gives us the following equation for \\(d\\), \\[\nd = \\frac{\\bar{X}_1 - \\bar{X}_2}{\\sqrt{\\displaystyle{\\frac{ {\\hat\\sigma_1}^2 + {\\hat\\sigma_2}^2}{2}}}}\n\\] as our estimate of the effect size. There’s nothing particularly difficult about calculating this version in R, since all we have to do is change the `method` argument: \n\n```\ncohensD( formula = grade ~ tutor, \n data = harpo,\n method = \"unequal\" \n )\n```\n\n `## [1] 0.7244995` This is the version of Cohen’s \\(d\\) that gets reported by the\n\nfunction whenever it runs a Welch \\(t\\)-test. \n\n### 13.8.4 Cohen’s \\(d\\) from a paired-samples test\n\n Finally, what should we do for a paired samples \\(t\\)-test? In this case, the answer depends on what it is you’re trying to do. If you want to measure your effect sizes relative to the distribution of difference scores, the measure of \\(d\\) that you calculate is just ( `method = \"paired\"` ) \\[\nd = \\frac{\\bar{D}}{\\hat{\\sigma}_D}\n\\] where \\(\\hat{\\sigma}_D\\) is the estimate of the standard deviation of the differences. The calculation here is pretty straightforward \n\n```\ncohensD( x = chico$grade_test2, \n y = chico$grade_test1,\n method = \"paired\" \n )\n```\n\n `## [1] 1.447952` This is the version of Cohen’s \\(d\\) that gets reported by the `pairedSamplesTTest()` function. The only wrinkle is figuring out whether this is the measure you want or not. To the extent that you care about the practical consequences of your research, you often want to measure the effect size relative to the original variables, not the difference scores (e.g., the 1% improvement in Dr Chico’s class is pretty small when measured against the amount of between-student variation in grades), in which case you use the same versions of Cohen’s \\(d\\) that you would use for a Student or Welch test. For instance, when we do that for Dr Chico’s class, \n\n```\ncohensD( x = chico$grade_test2, \n y = chico$grade_test1,\n method = \"pooled\" \n )\n```\n\n `## [1] 0.2157646` \nwhat we see is that the overall effect size is quite small, when assessed on the scale of the original variables.\n\n## 13.9 Checking the normality of a sample\n\nAll of the tests that we have discussed so far in this chapter have assumed that the data are normally distributed. This assumption is often quite reasonable, because the central limit theorem (Section 10.3.3) does tend to ensure that many real world quantities are normally distributed: any time that you suspect that your variable is actually an average of lots of different things, there’s a pretty good chance that it will be normally distributed; or at least close enough to normal that you can get away with using \\(t\\)-tests. However, life doesn’t come with guarantees; and besides, there are lots of ways in which you can end up with variables that are highly non-normal. For example, any time you think that your variable is actually the minimum of lots of different things, there’s a very good chance it will end up quite skewed. In psychology, response time (RT) data is a good example of this. If you suppose that there are lots of things that could trigger a response from a human participant, then the actual response will occur the first time one of these trigger events occurs.199 This means that RT data are systematically non-normal. Okay, so if normality is assumed by all the tests, and is mostly but not always satisfied (at least approximately) by real world data, how can we check the normality of a sample? In this section I discuss two methods: QQ plots, and the Shapiro-Wilk test.\n\n### 13.9.1 QQ plots\n\n```\n## Normally Distributed Data \n## skew= -0.02936155 \n## kurtosis= -0.06035938 \n## \n## Shapiro-Wilk normality test\n## \n## data: data\n## W = 0.99108, p-value = 0.7515\n```\n\nThe Shapiro-Wilk statistic associated with the data in Figures 13.14 and 13.15 is \\(W = .99\\), indicating that no significant departures from normality were detected (\\(p = .73\\)). As you can see, these data form a pretty straight line; which is no surprise given that we sampled them from a normal distribution! In contrast, have a look at the two data sets shown in Figures 13.16, 13.17, 13.18, 13.19. Figures 13.16 and 13.17 show the histogram and a QQ plot for a data set that is highly skewed: the QQ plot curves upwards. Figures 13.18 and 13.19 show the same plots for a heavy tailed (i.e., high kurtosis) data set: in this case, the QQ plot flattens in the middle and curves sharply at either end.\n\n```\n## Skewed Data \n## skew= 1.889475 \n## kurtosis= 4.4396 \n## \n## Shapiro-Wilk normality test\n## \n## data: data\n## W = 0.81758, p-value = 8.908e-10\n```\n\nThe skewness of the data in Figures 13.16 and 13.17 is 1.94, and is reflected in a QQ plot that curves upwards. As a consequence, the Shapiro-Wilk statistic is \\(W=.80\\), reflecting a significant departure from normality (\\(p<.001\\)).\n\n```\n## Heavy-Tailed Data \n## skew= -0.05308273 \n## kurtosis= 7.508765 \n## \n## Shapiro-Wilk normality test\n## \n## data: data\n## W = 0.83892, p-value = 4.718e-09\n```\n\nFigures 13.18 and 13.19 shows the same plots for a heavy tailed data set, again consisting of 100 observations. In this case, the heavy tails in the data produce a high kurtosis (2.80), and cause the QQ plot to flatten in the middle, and curve away sharply on either side. The resulting Shapiro-Wilk statistic is \\(W = .93\\), again reflecting significant non-normality (\\(p < .001\\)).\n One way to check whether a sample violates the normality assumption is to draw a “quantile-quantile” plot (QQ plot). This allows you to visually check whether you’re seeing any systematic violations. In a QQ plot, each observation is plotted as a single dot. The x co-ordinate is the theoretical quantile that the observation should fall in, if the data were normally distributed (with mean and variance estimated from the sample) and on the y co-ordinate is the actual quantile of the data within the sample. If the data are normal, the dots should form a straight line. For instance, lets see what happens if we generate data by sampling from a normal distribution, and then drawing a QQ plot using the R function `qqnorm()` . The `qqnorm()` function has a few arguments, but the only one we really need to care about here is `y` , a vector specifying the data whose normality we’re interested in checking. Here’s the R commands: \n\n```\nnormal.data <- rnorm( n = 100 ) # generate N = 100 normally distributed numbers\nhist( x = normal.data ) # draw a histogram of these numbers\n```\n\n```\nqqnorm( y = normal.data ) # draw the QQ plot\n```\n\n### 13.9.2 Shapiro-Wilk tests\n\nAlthough QQ plots provide a nice way to informally check the normality of your data, sometimes you’ll want to do something a bit more formal. And when that moment comes, the Shapiro-Wilk test (Shapiro and Wilk 1965) is probably what you’re looking for.200 As you’d expect, the null hypothesis being tested is that a set of \\(N\\) observations is normally distributed. The test statistic that it calculates is conventionally denoted as \\(W\\), and it’s calculated as follows. First, we sort the observations in order of increasing size, and let \\(X_1\\) be the smallest value in the sample, \\(X_2\\) be the second smallest and so on. Then the value of \\(W\\) is given by \\[ W = \\frac{ \\left( \\sum_{i = 1}^N a_i X_i \\right)^2 }{ \\sum_{i = 1}^N (X_i - \\bar{X})^2} \\] where \\(\\bar{X}\\) is the mean of the observations, and the \\(a_i\\) values are … mumble, mumble … something complicated that is a bit beyond the scope of an introductory text.\n\nBecause it’s a little hard to explain the maths behind the \\(W\\) statistic, a better idea is to give a broad brush description of how it behaves. Unlike most of the test statistics that we’ll encounter in this book, it’s actually small values of \\(W\\) that indicated departure from normality. The \\(W\\) statistic has a maximum value of 1, which arises when the data look “perfectly normal”. The smaller the value of \\(W\\), the less normal the data are. However, the sampling distribution for \\(W\\) – which is not one of the standard ones that I discussed in Chapter 9 and is in fact a complete pain in the arse to work with – does depend on the sample size \\(N\\). To give you a feel for what these sampling distributions look like, I’ve plotted three of them in Figure 13.20. Notice that, as the sample size starts to get large, the sampling distribution becomes very tightly clumped up near \\(W=1\\), and as a consequence, for larger samples \\(W\\) doesn’t have to be very much smaller than 1 in order for the test to be significant.\n To run the test in R, we use the `shapiro.test()` function. It has only a single argument `x` , which is a numeric vector containing the data whose normality needs to be tested. For example, when we apply this function to our `normal.data` , we get the following: \n\n```\nshapiro.test( x = normal.data )\n```\n\n```\n## \n## Shapiro-Wilk normality test\n## \n## data: normal.data\n## W = 0.98654, p-value = 0.4076\n```\n\nSo, not surprisingly, we have no evidence that these data depart from normality. When reporting the results for a Shapiro-Wilk test, you should (as usual) make sure to include the test statistic \\(W\\) and the \\(p\\) value, though given that the sampling distribution depends so heavily on \\(N\\) it would probably be a politeness to include \\(N\\) as well.\n\n## 13.10 Testing non-normal data with Wilcoxon tests\n\nOkay, suppose your data turn out to be pretty substantially non-normal, but you still want to run something like a \\(t\\)-test? This situation occurs a lot in real life: for the AFL winning margins data, for instance, the Shapiro-Wilk test made it very clear that the normality assumption is violated. This is the situation where you want to use Wilcoxon tests.\n\nLike the \\(t\\)-test, the Wilcoxon test comes in two forms, one-sample and two-sample, and they’re used in more or less the exact same situations as the corresponding \\(t\\)-tests. Unlike the \\(t\\)-test, the Wilcoxon test doesn’t assume normality, which is nice. In fact, they don’t make any assumptions about what kind of distribution is involved: in statistical jargon, this makes them nonparametric tests. While avoiding the normality assumption is nice, there’s a drawback: the Wilcoxon test is usually less powerful than the \\(t\\)-test (i.e., higher Type II error rate). I won’t discuss the Wilcoxon tests in as much detail as the \\(t\\)-tests, but I’ll give you a brief overview.\n\n### 13.10.1 Two sample Wilcoxon test\n\n I’ll start by describing the two sample Wilcoxon test (also known as the Mann-Whitney test), since it’s actually simpler than the one sample version. Suppose we’re looking at the scores of 10 people on some test. Since my imagination has now failed me completely, let’s pretend it’s a “test of awesomeness”, and there are two groups of people, “A” and “B”. I’m curious to know which group is more awesome. The data are included in the file `awesome.Rdata` , and like many of the data sets I’ve been using, it contains only a single data frame, in this case called `awesome` . Here’s the data: \n\n```\nload(file.path(projecthome, \"data/awesome.Rdata\"))\nprint( awesome )\n```\n\n```\n## scores group\n## 1 6.4 A\n## 2 10.7 A\n## 3 11.9 A\n## 4 7.3 A\n## 5 10.0 A\n## 6 14.5 B\n## 7 10.4 B\n## 8 12.9 B\n## 9 11.7 B\n## 10 13.0 B\n```\n\nAs long as there are no ties (i.e., people with the exact same awesomeness score), then the test that we want to do is surprisingly simple. All we have to do is construct a table that compares every observation in group \\(A\\) against every observation in group \\(B\\). Whenever the group \\(A\\) datum is larger, we place a check mark in the table:\n\nWe then count up the number of checkmarks. This is our test statistic, \\(W\\).201 The actual sampling distribution for \\(W\\) is somewhat complicated, and I’ll skip the details. For our purposes, it’s sufficient to note that the interpretation of \\(W\\) is qualitatively the same as the interpretation of \\(t\\) or \\(z\\). That is, if we want a two-sided test, then we reject the null hypothesis when \\(W\\) is very large or very small; but if we have a directional (i.e., one-sided) hypothesis, then we only use one or the other.\n The structure of the `wilcox.test()` function should feel very familiar to you by now. When you have your data organised in terms of an outcome variable and a grouping variable, then you use the `formula` and `data` arguments, so your command looks like this: \n\n```\nwilcox.test( formula = scores ~ group, data = awesome)\n```\n\n```\n## \n## Wilcoxon rank sum test\n## \n## data: scores by group\n## W = 3, p-value = 0.05556\n## alternative hypothesis: true location shift is not equal to 0\n```\n\n Just like we saw with the `t.test()` function, there is an `alternative` argument that you can use to switch between two-sided tests and one-sided tests, plus a few other arguments that we don’t need to worry too much about at an introductory level. Similarly, the `wilcox.test()` function allows you to use the `x` and `y` arguments when you have your data stored separately for each group. For instance, suppose we use the data from the `awesome2.Rdata` file: \n\n```\nload( file.path(projecthome, \"data/awesome2.Rdata\" ))\nscore.A\n```\n\n```\n## [1] 6.4 10.7 11.9 7.3 10.0\n```\n\n `score.B` \n\n```\n## [1] 14.5 10.4 12.9 11.7 13.0\n```\n\nWhen your data are organised like this, then you would use a command like this:\n\n```\nwilcox.test( x = score.A, y = score.B )\n```\n\n```\n## \n## Wilcoxon rank sum test\n## \n## data: score.A and score.B\n## W = 3, p-value = 0.05556\n## alternative hypothesis: true location shift is not equal to 0\n```\n\nThe output that R produces is pretty much the same as last time.\n\n### 13.10.2 One sample Wilcoxon test\n\nWhat about the one sample Wilcoxon test (or equivalently, the paired samples Wilcoxon test)? Suppose I’m interested in finding out whether taking a statistics class has any effect on the happiness of students. Here’s my data:\n\n```\nload( file.path(projecthome, \"data/happy.Rdata\" ))\nprint( happiness )\n```\n\n```\n## before after change\n## 1 30 6 -24\n## 2 43 29 -14\n## 3 21 11 -10\n## 4 24 31 7\n## 5 23 17 -6\n## 6 40 2 -38\n## 7 29 31 2\n## 8 56 21 -35\n## 9 38 8 -30\n## 10 16 21 5\n```\n\n What I’ve measured here is the happiness of each student `before` taking the class and `after` taking the class; the `change` score is the difference between the two. Just like we saw with the \\(t\\)-test, there’s no fundamental difference between doing a paired-samples test using `before` and `after` , versus doing a one-sample test using the `change` scores. As before, the simplest way to think about the test is to construct a tabulation. The way to do it this time is to take those change scores that are positive valued, and tabulate them against all the complete sample. What you end up with is a table that looks like this: \nCounting up the tick marks this time, we get a test statistic of \\(V = 7\\). As before, if our test is two sided, then we reject the null hypothesis when \\(V\\) is very large or very small. As far of running it in R goes, it’s pretty much what you’d expect. For the one-sample version, the command you would use is\n\n```\nwilcox.test( x = happiness$change,\n mu = 0\n)\n```\n\nAs this shows, we have a significant effect. Evidently, taking a statistics class does have an effect on your happiness. Switching to a paired samples version of the test won’t give us different answers, of course; but here’s the command to do it:\n\n```\nwilcox.test( x = happiness$after,\n y = happiness$before,\n paired = TRUE \n)\n```\n\n## 13.11 Summary\n\n* A one sample \\(t\\)-test is used to compare a single sample mean against a hypothesised value for the population mean. (Section 13.2)\n * An independent samples \\(t\\)-test is used to compare the means of two groups, and tests the null hypothesis that they have the same mean. It comes in two forms: the Student test (Section 13.3 assumes that the groups have the same standard deviation, the Welch test (Section 13.4) does not.\n * A paired samples \\(t\\)-test is used when you have two scores from each person, and you want to test the null hypothesis that the two scores have the same mean. It is equivalent to taking the difference between the two scores for each person, and then running a one sample \\(t\\)-test on the difference scores. (Section 13.5)\n * Effect size calculations for the difference between means can be calculated via the Cohen’s \\(d\\) statistic. (Section 13.8).\n * You can check the normality of a sample using QQ plots and the Shapiro-Wilk test. (Section 13.9)\n * If your data are non-normal, you can use Wilcoxon tests instead of \\(t\\)-tests. (Section 13.10)\n\nInformal experimentation in my garden suggests that yes, it does. Australian natives are adapted to low phosphorus levels relative to everywhere else on Earth, apparently, so if you’ve bought a house with a bunch of exotics and you want to plant natives, don’t follow my example: keep them separate. Nutrients to European plants are poison to Australian ones. There’s probably a joke in that, but I can’t figure out what it is.↩\n * \n\nActually this is too strong. Strictly speaking the \\(z\\) test only requires that the sampling distribution of the mean be normally distributed; if the population is normal then it necessarily follows that the sampling distribution of the mean is also normal. However, as we saw when talking about the central limit theorem, it’s quite possible (even commonplace) for the sampling distribution to be normal even if the population distribution itself is non-normal. However, in light of the sheer ridiculousness of the assumption that the true standard deviation is known, there really isn’t much point in going into details on this front!↩\n * \n\nWell, sort of. As I understand the history, Gosset only provided a partial solution: the general solution to the problem was provided by .↩\n * \n\nMore seriously, I tend to think the reverse is true: I get very suspicious of technical reports that fill their results sections with nothing except the numbers. It might just be that I’m an arrogant jerk, but I often feel like an author that makes no attempt to explain and interpret their analysis to the reader either doesn’t understand it themselves, or is being a bit lazy. Your readers are smart, but not infinitely patient. Don’t annoy them if you can help it.↩\n * \n\nAlthough it is the simplest, which is why I started with it.↩\n * \n\nA funny question almost always pops up at this point: what the heck is the population being referred to in this case? Is it the set of students actually taking ’s class (all 33 of them)? The set of people who might take the class (an unknown number) of them? Or something else? Does it matter which of these we pick? It’s traditional in an introductory behavioural stats class to mumble a lot at this point, but since I get asked this question every year by my students, I’ll give a brief answer. Technically yes, it does matter: if you change your definition of what the “real world” population actually is, then the sampling distribution of your observed mean \\(\\bar{X}\\) changes too. The \\(t\\)-test relies on an assumption that the observations are sampled at random from an infinitely large population; and to the extent that real life isn’t like that, then the \\(t\\)-test can be wrong. In practice, however, this isn’t usually a big deal: even though the assumption is almost always wrong, it doesn’t lead to a lot of pathological behaviour from the test, so we tend to just ignore it.↩\n * \n\nYes, I have a “favourite” way of thinking about pooled standard deviation estimates. So what?↩\n * \n\nWell, I guess you can average apples and oranges, and what you end up with is a delicious fruit smoothie. But no one really thinks that a fruit smoothie is a very good way to describe the original fruits, do they?↩\n * \n\nThis design is very similar to the one in Section 12.8 that motivated the McNemar test. This should be no surprise. Both are standard repeated measures designs involving two measurements. The only difference is that this time our outcome variable is interval scale (working memory capacity) rather than a binary, nominal scale variable (a yes-or-no question).↩\n * \n\nAt this point we have Drs Harpo, Chico and Zeppo. No prizes for guessing who Dr Groucho is.↩\n * \n\nThis is obviously a class being taught at a very small or very expensive university, or else is a postgraduate class. I’ve never taught an intro stats class with less than 350 students.↩\n * \n\nThe\n `sortFrame()` function sorts factor variables like `id` in alphabetical order, which is why it jumps from “student1” to “student10”↩ * \n\nThis is a massive oversimplification.↩\n * \n\nEither that, or the Kolmogorov-Smirnov test, which is probably more traditional than the Shapiro-Wilk, though most things I’ve read seem to suggest Shapiro-Wilk is the better test of normality; although Kolomogorov-Smirnov is a general purpose test of distributional equivalence, so it can be adapted to handle other kinds of distribution tests; in R it’s implemented via the\n `ks.test()` function.↩ * \n\nActually, there are two different versions of the test statistic; they differ from each other by a constant value. The version that I’ve described is the one that R calculates.↩\n\nDate: 2014-10-01\nCategories: \nTags: \n\n# Chapter 14 Comparing several means (one-way ANOVA)\n\nThis chapter introduces one of the most widely used tools in statistics, known as “the analysis of variance”, which is usually referred to as ANOVA. The basic technique was developed by in the early 20th century, and it is to him that we owe the rather unfortunate terminology. The term ANOVA is a little misleading, in two respects. Firstly, although the name of the technique refers to variances, ANOVA is concerned with investigating differences in means. Secondly, there are several different things out there that are all referred to as ANOVAs, some of which have only a very tenuous connection to one another. Later on in the book we’ll encounter a range of different ANOVA methods that apply in quite different situations, but for the purposes of this chapter we’ll only consider the simplest form of ANOVA, in which we have several different groups of observations, and we’re interested in finding out whether those groups differ in terms of some outcome variable of interest. This is the question that is addressed by a one-way ANOVA.\n\nThe structure of this chapter is as follows: In Section 14.1 I’ll introduce a fictitious data set that we’ll use as a running example throughout the chapter. After introducing the data, I’ll describe the mechanics of how a one-way ANOVA actually works (Section 14.2) and then focus on how you can run one in R (Section 14.3). These two sections are the core of the chapter. The remainder of the chapter discusses a range of important topics that inevitably arise when running an ANOVA, namely how to calculate effect sizes (Section 14.4), post hoc tests and corrections for multiple comparisons (Section 14.5) and the assumptions that ANOVA relies upon (Section 14.6). We’ll also talk about how to check those assumptions and some of the things you can do if the assumptions are violated (Sections 14.7 to 14.10). At the end of the chapter we’ll talk a little about the relationship between ANOVA and other statistical tools (Section 14.11).\n\n## 14.1 An illustrative data set\n\nSuppose you’ve become involved in a clinical trial in which you are testing a new antidepressant drug called Joyzepam. In order to construct a fair test of the drug’s effectiveness, the study involves three separate drugs to be administered. One is a placebo, and the other is an existing antidepressant / anti-anxiety drug called Anxifree. A collection of 18 participants with moderate to severe depression are recruited for your initial testing. Because the drugs are sometimes administered in conjunction with psychological therapy, your study includes 9 people undergoing cognitive behavioural therapy (CBT) and 9 who are not. Participants are randomly assigned (doubly blinded, of course) a treatment, such that there are 3 CBT people and 3 no-therapy people assigned to each of the 3 drugs. A psychologist assesses the mood of each person after a 3 month run with each drug: and the overall improvement in each person’s mood is assessed on a scale ranging from \\(-5\\) to \\(+5\\).\n\nWith that as the study design, let’s now look at what we’ve got in the data file:\n\n```\nload(file.path(projecthome, \"data\", \"clinicaltrial.Rdata\")) # load data\nstr(clin.trial)\n```\n\n```\n## 'data.frame': 18 obs. of 3 variables:\n## $ drug : Factor w/ 3 levels \"placebo\",\"anxifree\",..: 1 1 1 2 2 2 3 3 3 1 ...\n## $ therapy : Factor w/ 2 levels \"no.therapy\",\"CBT\": 1 1 1 1 1 1 1 1 1 2 ...\n## $ mood.gain: num 0.5 0.3 0.1 0.6 0.4 0.2 1.4 1.7 1.3 0.6 ...\n```\n\n So we have a single data frame called `clin.trial` , containing three variables; `drug` , `therapy` and `mood.gain` . Next, let’s print the data frame to get a sense of what the data actually look like. `print( clin.trial )` \n\n```\n## drug therapy mood.gain\n## 1 placebo no.therapy 0.5\n## 2 placebo no.therapy 0.3\n## 3 placebo no.therapy 0.1\n## 4 anxifree no.therapy 0.6\n## 5 anxifree no.therapy 0.4\n## 6 anxifree no.therapy 0.2\n## 7 joyzepam no.therapy 1.4\n## 8 joyzepam no.therapy 1.7\n## 9 joyzepam no.therapy 1.3\n## 10 placebo CBT 0.6\n## 11 placebo CBT 0.9\n## 12 placebo CBT 0.3\n## 13 anxifree CBT 1.1\n## 14 anxifree CBT 0.8\n## 15 anxifree CBT 1.2\n## 16 joyzepam CBT 1.8\n## 17 joyzepam CBT 1.3\n## 18 joyzepam CBT 1.4\n```\n\n For the purposes of this chapter, what we’re really interested in is the effect of `drug` on `mood.gain` . The first thing to do is calculate some descriptive statistics and draw some graphs. In Chapter 5 we discussed a variety of different functions that can be used for this purpose. For instance, we can use the `xtabs()` function to see how many people we have in each group: \n\n```\nxtabs( ~drug, clin.trial )\n```\n\n```\n## drug\n## placebo anxifree joyzepam \n## 6 6 6\n```\n\n Similarly, we can use the `aggregate()` function to calculate means and standard deviations for the `mood.gain` variable broken down by which `drug` was administered: \n\n```\n## drug mood.gain\n## 1 placebo 0.2810694\n## 2 anxifree 0.3920034\n## 3 joyzepam 0.2136976\n```\n\n Finally, we can use `plotmeans()` from the `gplots` package to produce a pretty picture. \n\n```\nlibrary(gplots)\nplotmeans( formula = mood.gain ~ drug, # plot mood.gain by drug\n data = clin.trial, # the data frame\n xlab = \"Drug Administered\", # x-axis label\n ylab = \"Mood Gain\", # y-axis label\n n.label = FALSE # don't display sample size\n )\n```\n\nThe results are shown in Figure 14.1, which plots the average mood gain for all three conditions; error bars show 95% confidence intervals. As the plot makes clear, there is a larger improvement in mood for participants in the Joyzepam group than for either the Anxifree group or the placebo group. The Anxifree group shows a larger mood gain than the control group, but the difference isn’t as large.\n\nThe question that we want to answer is: are these difference “real”, or are they just due to chance?\n\n## 14.2 How ANOVA works\n\nIn order to answer the question posed by our clinical trial data, we’re going to run a one-way ANOVA. As usual, I’m going to start by showing you how to do it the hard way, building the statistical tool from the ground up and showing you how you could do it in R if you didn’t have access to any of the cool built-in ANOVA functions. And, as always, I hope you’ll read it carefully, try to do it the long way once or twice to make sure you really understand how ANOVA works, and then – once you’ve grasped the concept – never ever do it this way again.\n\nThe experimental design that I described in the previous section strongly suggests that we’re interested in comparing the average mood change for the three different drugs. In that sense, we’re talking about an analysis similar to the \\(t\\)-test (Chapter 13, but involving more than two groups. If we let \\(\\mu_P\\) denote the population mean for the mood change induced by the placebo, and let \\(\\mu_A\\) and \\(\\mu_J\\) denote the corresponding means for our two drugs, Anxifree and Joyzepam, then the (somewhat pessimistic) null hypothesis that we want to test is that all three population means are identical: that is, neither of the two drugs is any more effective than a placebo. Mathematically, we write this null hypothesis like this: \\[ \\begin{array}{rcl} H_0 &:& \\mbox{it is true that } \\mu_P = \\mu_A = \\mu_J \\end{array} \\] As a consequence, our alternative hypothesis is that at least one of the three different treatments is different from the others. It’s a little trickier to write this mathematically, because (as we’ll discuss) there are quite a few different ways in which the null hypothesis can be false. So for now we’ll just write the alternative hypothesis like this: \\[ \\begin{array}{rcl} H_1 &:& \\mbox{it is *not* true that } \\mu_P = \\mu_A = \\mu_J \\end{array} \\] This null hypothesis is a lot trickier to test than any of the ones we’ve seen previously. How shall we do it? A sensible guess would be to “do an ANOVA”, since that’s the title of the chapter, but it’s not particularly clear why an “analysis of variances” will help us learn anything useful about the means. In fact, this is one of the biggest conceptual difficulties that people have when first encountering ANOVA. To see how this works, I find it most helpful to start by talking about variances. In fact, what I’m going to do is start by playing some mathematical games with the formula that describes the variance. That is, we’ll start out by playing around with variances, and it will turn out that this gives us a useful tool for investigating means.\n\n### 14.2.1 Two formulas for the variance of \\(Y\\)\n\nFirstly, let’s start by introducing some notation. We’ll use \\(G\\) to refer to the total number of groups. For our data set, there are three drugs, so there are \\(G=3\\) groups. Next, we’ll use \\(N\\) to refer to the total sample size: there are a total of \\(N=18\\) people in our data set. Similarly, let’s use \\(N_k\\) to denote the number of people in the \\(k\\)-th group. In our fake clinical trial, the sample size is \\(N_k = 6\\) for all three groups.202 Finally, we’ll use \\(Y\\) to denote the outcome variable: in our case, \\(Y\\) refers to mood change. Specifically, we’ll use \\(Y_{ik}\\) to refer to the mood change experienced by the \\(i\\)-th member of the \\(k\\)-th group. Similarly, we’ll use \\(\\bar{Y}\\) to be the average mood change, taken across all 18 people in the experiment, and \\(\\bar{Y}_k\\) to refer to the average mood change experienced by the 6 people in group \\(k\\).\n\nExcellent. Now that we’ve got our notation sorted out, we can start writing down formulas. To start with, let’s recall the formula for the variance that we used in Section 5.2, way back in those kinder days when we were just doing descriptive statistics. The sample variance of \\(Y\\) is defined as follows: \\[ \\mbox{Var}(Y) = \\frac{1}{N} \\sum_{k=1}^G \\sum_{i=1}^{N_k} \\left(Y_{ik} - \\bar{Y} \\right)^2 \\] This formula looks pretty much identical to the formula for the variance in Section 5.2. The only difference is that this time around I’ve got two summations here: I’m summing over groups (i.e., values for \\(k\\)) and over the people within the groups (i.e., values for \\(i\\)). This is purely a cosmetic detail: if I’d instead used the notation \\(Y_p\\) to refer to the value of the outcome variable for person \\(p\\) in the sample, then I’d only have a single summation. The only reason that we have a double summation here is that I’ve classified people into groups, and then assigned numbers to people within groups.\n\nA concrete example might be useful here. Let’s consider this table, in which we have a total of \\(N=5\\) people sorted into \\(G=2\\) groups. Arbitrarily, let’s say that the “cool” people are group 1, and the “uncool” people are group 2, and it turns out that we have three cool people (\\(N_1 = 3\\)) and two uncool people (\\(N_2 = 2\\)).\n\nname | person (\\(p\\)) | group | group num (\\(k\\)) | index in group (\\(i\\)) | grumpiness (\\(Y_{ik}\\) or \\(Y_p\\)) |\n | --- | --- | --- | --- | --- | --- |\nAnn | 1 | cool | 1 | 1 | 20 |\nBen | 2 | cool | 1 | 2 | 55 |\nCat | 3 | cool | 1 | 3 | 21 |\nDan | 4 | uncool | 2 | 1 | 91 |\nEgg | 5 | uncool | 2 | 2 | 22 |\n\nNotice that I’ve constructed two different labelling schemes here. We have a “person” variable \\(p\\), so it would be perfectly sensible to refer to \\(Y_p\\) as the grumpiness of the \\(p\\)-th person in the sample. For instance, the table shows that Dan is the four so we’d say \\(p = 4\\). So, when talking about the grumpiness \\(Y\\) of this “Dan” person, whoever he might be, we could refer to his grumpiness by saying that \\(Y_p = 91\\), for person \\(p = 4\\) that is. However, that’s not the only way we could refer to Dan. As an alternative we could note that Dan belongs to the “uncool” group (\\(k = 2\\)), and is in fact the first person listed in the uncool group (\\(i = 1\\)). So it’s equally valid to refer to Dan’s grumpiness by saying that \\(Y_{ik} = 91\\), where \\(k = 2\\) and \\(i = 1\\). In other words, each person \\(p\\) corresponds to a unique \\(ik\\) combination, and so the formula that I gave above is actually identical to our original formula for the variance, which would be \\[ \\mbox{Var}(Y) = \\frac{1}{N} \\sum_{p=1}^N \\left(Y_{p} - \\bar{Y} \\right)^2 \\] In both formulas, all we’re doing is summing over all of the observations in the sample. Most of the time we would just use the simpler \\(Y_p\\) notation: the equation using \\(Y_p\\) is clearly the simpler of the two. However, when doing an ANOVA it’s important to keep track of which participants belong in which groups, and we need to use the \\(Y_{ik}\\) notation to do this.\n\n### 14.2.2 From variances to sums of squares\n\nOkay, now that we’ve got a good grasp on how the variance is calculated, let’s define something called the total sum of squares, which is denoted SS\\(_{tot}\\). This is very simple: instead of averaging the squared deviations, which is what we do when calculating the variance, we just add them up. So the formula for the total sum of squares is almost identical to the formula for the variance: \\[ \\mbox{SS}_{tot} = \\sum_{k=1}^G \\sum_{i=1}^{N_k} \\left(Y_{ik} - \\bar{Y} \\right)^2 \\] When we talk about analysing variances in the context of ANOVA, what we’re really doing is working with the total sums of squares rather than the actual variance. One very nice thing about the total sum of squares is that we can break it up into two different kinds of variation. Firstly, we can talk about the within-group sum of squares, in which we look to see how different each individual person is from their own group mean: \\[ \\mbox{SS}_w = \\sum_{k=1}^G \\sum_{i=1}^{N_k} \\left( Y_{ik} - \\bar{Y}_k \\right)^2 \\] where \\(\\bar{Y}_k\\) is a group mean. In our example, \\(\\bar{Y}_k\\) would be the average mood change experienced by those people given the \\(k\\)-th drug. So, instead of comparing individuals to the average of all people in the experiment, we’re only comparing them to those people in the the same group. As a consequence, you’d expect the value of \\(\\mbox{SS}_w\\) to be smaller than the total sum of squares, because it’s completely ignoring any group differences – that is, the fact that the drugs (if they work) will have different effects on people’s moods.\n\nNext, we can define a third notion of variation which captures only the differences between groups. We do this by looking at the differences between the group means \\(\\bar{Y}_k\\) and grand mean \\(\\bar{Y}\\). In order to quantify the extent of this variation, what we do is calculate the between-group sum of squares: \\[ \\begin{array}{rcl} \\mbox{SS}_{b} &=& \\sum_{k=1}^G \\sum_{i=1}^{N_k} \\left( \\bar{Y}_k - \\bar{Y} \\right)^2 \\\\ &=& \\sum_{k=1}^G N_k \\left( \\bar{Y}_k - \\bar{Y} \\right)^2 \\end{array} \\] It’s not too difficult to show that the total variation among people in the experiment \\(\\mbox{SS}_{tot}\\) is actually the sum of the differences between the groups \\(\\mbox{SS}_b\\) and the variation inside the groups \\(\\mbox{SS}_w\\). That is: \\[ \\mbox{SS}_w + \\mbox{SS}_{b} = \\mbox{SS}_{tot} \\] Yay.\n\nOkay, so what have we found out? We’ve discovered that the total variability associated with the outcome variable (\\(\\mbox{SS}_{tot}\\)) can be mathematically carved up into the sum of “the variation due to the differences in the sample means for the different groups” (\\(\\mbox{SS}_{b}\\)) plus “all the rest of the variation” (\\(\\mbox{SS}_{w}\\)). How does that help me find out whether the groups have different population means? Um. Wait. Hold on a second… now that I think about it, this is exactly what we were looking for. If the null hypothesis is true, then you’d expect all the sample means to be pretty similar to each other, right? And that would imply that you’d expect \\(\\mbox{SS}_{b}\\) to be really small, or at least you’d expect it to be a lot smaller than the “the variation associated with everything else”, \\(\\mbox{SS}_{w}\\). Hm. I detect a hypothesis test coming on…\n\n### 14.2.3 From sums of squares to the \\(F\\)-test\n\nAs we saw in the last section, the qualitative idea behind ANOVA is to compare the two sums of squares values \\(\\mbox{SS}_b\\) and \\(\\mbox{SS}_w\\) to each other: if the between-group variation is \\(\\mbox{SS}_b\\) is large relative to the within-group variation \\(\\mbox{SS}_w\\) then we have reason to suspect that the population means for the different groups aren’t identical to each other. In order to convert this into a workable hypothesis test, there’s a little bit of “fiddling around” needed. What I’ll do is first show you what we do to calculate our test statistic – which is called an \\(F\\) ratio – and then try to give you a feel for why we do it this way.\n\nIn order to convert our SS values into an \\(F\\)-ratio, the first thing we need to calculate is the degrees of freedom associated with the SS\\(_b\\) and SS\\(_w\\) values. As usual, the degrees of freedom corresponds to the number of unique “data points” that contribute to a particular calculation, minus the number of “constraints” that they need to satisfy. For the within-groups variability, what we’re calculating is the variation of the individual observations (\\(N\\) data points) around the group means (\\(G\\) constraints). In contrast, for the between groups variability, we’re interested in the variation of the group means (\\(G\\) data points) around the grand mean (1 constraint). Therefore, the degrees of freedom here are: \\[ \\begin{array}{lcl} \\mbox{df}_b &=& G-1 \\\\ \\mbox{df}_w &=& N-G \\\\ \\end{array} \\] Okay, that seems simple enough. What we do next is convert our summed squares value into a “mean squares” value, which we do by dividing by the degrees of freedom: \\[ \\begin{array}{lcl} \\mbox{MS}_b &=& \\displaystyle\\frac{\\mbox{SS}_b }{ \\mbox{df}_b} \\\\ \\mbox{MS}_w &=& \\displaystyle\\frac{\\mbox{SS}_w }{ \\mbox{df}_w} \\end{array} \\] Finally, we calculate the \\(F\\)-ratio by dividing the between-groups MS by the within-groups MS: \\[ F = \\frac{\\mbox{MS}_b }{ \\mbox{MS}_w } \\] At a very general level, the intuition behind the \\(F\\) statistic is straightforward: bigger values of \\(F\\) means that the between-groups variation is large, relative to the within-groups variation. As a consequence, the larger the value of \\(F\\), the more evidence we have against the null hypothesis. But how large does \\(F\\) have to be in order to actually reject \\(H_0\\)? In order to understand this, you need a slightly deeper understanding of what ANOVA is and what the mean squares values actually are.\n\nThe next section discusses that in a bit of detail, but for readers that aren’t interested in the details of what the test is actually measuring, I’ll cut to the chase. In order to complete our hypothesis test, we need to know the sampling distribution for \\(F\\) if the null hypothesis is true. Not surprisingly, the sampling distribution for the \\(F\\) statistic under the null hypothesis is an \\(F\\) distribution. If you recall back to our discussion of the \\(F\\) distribution in Chapter @ref(probability, the \\(F\\) distribution has two parameters, corresponding to the two degrees of freedom involved: the first one df\\(_1\\) is the between groups degrees of freedom df\\(_b\\), and the second one df\\(_2\\) is the within groups degrees of freedom df\\(_w\\).\n\nA summary of all the key quantities involved in a one-way ANOVA, including the formulas showing how they are calculated, is shown in Table 14.1.\n\ndf | sum of squares | mean squares | \\(F\\) statistic | \\(p\\) value |\n | --- | --- | --- | --- | --- |\nbetween groups | \\(\\mbox{df}_b = G-1\\) | SS\\(_b = \\displaystyle\\sum_{k=1}^G N_k (\\bar{Y}_k - \\bar{Y})^2\\) | \\(\\mbox{MS}_b = \\frac{\\mbox{SS}_b}{\\mbox{df}_b}\\) | \\(F = \\frac{\\mbox{MS}_b }{ \\mbox{MS}_w }\\) | [complicated] |\nwithin groups | \\(\\mbox{df}_w = N-G\\) | SS\\(_w = \\sum_{k=1}^G \\sum_{i = 1}^{N_k} ( {Y}_{ik} - \\bar{Y}_k)^2\\) | \\(\\mbox{MS}_w = \\frac{\\mbox{SS}_w}{\\mbox{df}_w}\\) | - | - |\n\n### 14.2.4 The model for the data and the meaning of \\(F\\) (advanced)\n\nAt a fundamental level, ANOVA is a competition between two different statistical models, \\(H_0\\) and \\(H_1\\). When I described the null and alternative hypotheses at the start of the section, I was a little imprecise about what these models actually are. I’ll remedy that now, though you probably won’t like me for doing so. If you recall, our null hypothesis was that all of the group means are identical to one another. If so, then a natural way to think about the outcome variable \\(Y_{ik}\\) is to describe individual scores in terms of a single population mean \\(\\mu\\), plus the deviation from that population mean. This deviation is usually denoted \\(\\epsilon_{ik}\\) and is traditionally called the error or residual associated with that observation. Be careful though: just like we saw with the word “significant”, the word “error” has a technical meaning in statistics that isn’t quite the same as its everyday English definition. In everyday language, “error” implies a mistake of some kind; in statistics, it doesn’t (or at least, not necessarily). With that in mind, the word “residual” is a better term than the word “error”. In statistics, both words mean “leftover variability”: that is, “stuff” that the model can’t explain. In any case, here’s what the null hypothesis looks like when we write it as a statistical model: \\[ Y_{ik} = \\mu + \\epsilon_{ik} \\] where we make the assumption (discussed later) that the residual values \\(\\epsilon_{ik}\\) are normally distributed, with mean 0 and a standard deviation \\(\\sigma\\) that is the same for all groups. To use the notation that we introduced in Chapter 9 we would write this assumption like this: \\[ \\epsilon_{ik} \\sim \\mbox{Normal}(0, \\sigma^2) \\]\n\nWhat about the alternative hypothesis, \\(H_1\\)? The only difference between the null hypothesis and the alternative hypothesis is that we allow each group to have a different population mean. So, if we let \\(\\mu_k\\) denote the population mean for the \\(k\\)-th group in our experiment, then the statistical model corresponding to \\(H_1\\) is: \\[ Y_{ik} = \\mu_k + \\epsilon_{ik} \\] where, once again, we assume that the error terms are normally distributed with mean 0 and standard deviation \\(\\sigma\\). That is, the alternative hypothesis also assumes that \\[ \\epsilon \\sim \\mbox{Normal}(0, \\sigma^2) \\]\n\nOkay, now that we’ve described the statistical models underpinning \\(H_0\\) and \\(H_1\\) in more detail, it’s now pretty straightforward to say what the mean square values are measuring, and what this means for the interpretation of \\(F\\). I won’t bore you with the proof of this, but it turns out that the within-groups mean square, MS\\(_w\\), can be viewed as an estimator (in the technical sense: Chapter 10 of the error variance \\(\\sigma^2\\). The between-groups mean square MS\\(_b\\) is also an estimator; but what it estimates is the error variance plus a quantity that depends on the true differences among the group means. If we call this quantity \\(Q\\), then we can see that the \\(F\\)-statistic is basically203\n\n\\[ F = \\frac{\\hat{Q} + \\hat\\sigma^2}{\\hat\\sigma^2} \\] where the true value \\(Q=0\\) if the null hypothesis is true, and \\(Q > 0\\) if the alternative hypothesis is true (e.g. ch. 10 Hays 1994). Therefore, at a bare minimum the \\(F\\) value must be larger than 1 to have any chance of rejecting the null hypothesis. Note that this doesn’t mean that it’s impossible to get an \\(F\\)-value less than 1. What it means is that, if the null hypothesis is true the sampling distribution of the \\(F\\) ratio has a mean of 1,204 and so we need to see \\(F\\)-values larger than 1 in order to safely reject the null.\n\nTo be a bit more precise about the sampling distribution, notice that if the null hypothesis is true, both MS\\(_b\\) and MS\\(_w\\) are estimators of the variance of the residuals \\(\\epsilon_{ik}\\). If those residuals are normally distributed, then you might suspect that the estimate of the variance of \\(\\epsilon_{ik}\\) is chi-square distributed… because (as discussed in Section 9.6 that’s what a chi-square distribution is: it’s what you get when you square a bunch of normally-distributed things and add them up. And since the \\(F\\) distribution is (again, by definition) what you get when you take the ratio between two things that are \\(\\chi^2\\) distributed… we have our sampling distribution. Obviously, I’m glossing over a whole lot of stuff when I say this, but in broad terms, this really is where our sampling distribution comes from.\n\n### 14.2.5 A worked example\n\nThe previous discussion was fairly abstract, and a little on the technical side, so I think that at this point it might be useful to see a worked example. For that, let’s go back to the clinical trial data that I introduced at the start of the chapter. The descriptive statistics that we calculated at the beginning tell us our group means: an average mood gain of 0.45 for the placebo, 0.72 for Anxifree, and 1.48 for Joyzepam. With that in mind, let’s party like it’s 1899205 and start doing some pencil and paper calculations. I’ll only do this for the first 5 observations, because it’s not bloody 1899 and I’m very lazy. Let’s start by calculating \\(\\mbox{SS}_{w}\\), the within-group sums of squares. First, let’s draw up a nice table to help us with our calculations…\n\ngroup (\\(k\\)) | outcome (\\(Y_{ik}\\)) |\n | --- | --- |\nplacebo | 0.5 |\nplacebo | 0.3 |\nplacebo | 0.1 |\nanxifree | 0.6 |\nanxifree | 0.4 |\n\n At this stage, the only thing I’ve included in the table is the raw data itself: that is, the grouping variable (i.e., `drug` ) and outcome variable (i.e. `mood.gain` ) for each person. Note that the outcome variable here corresponds to the \\(Y_{ik}\\) value in our equation previously. The next step in the calculation is to write down, for each person in the study, the corresponding group mean; that is, \\(\\bar{Y}_k\\). This is slightly repetitive, but not particularly difficult since we already calculated those group means when doing our descriptive statistics: \n\ngroup (\\(k\\)) | outcome (\\(Y_{ik}\\)) | group mean (\\(\\bar{Y}_k\\)) |\n | --- | --- | --- |\nplacebo | 0.5 | 0.45 |\nplacebo | 0.3 | 0.45 |\nplacebo | 0.1 | 0.45 |\nanxifree | 0.6 | 0.72 |\nanxifree | 0.4 | 0.72 |\n\nNow that we’ve written those down, we need to calculate – again for every person – the deviation from the corresponding group mean. That is, we want to subtract \\(Y_{ik} - \\bar{Y}_k\\). After we’ve done that, we need to square everything. When we do that, here’s what we get:\n\ngroup (\\(k\\)) | outcome (\\(Y_{ik}\\)) | group mean (\\(\\bar{Y}_k\\)) | dev. from group mean (\\(Y_{ik} - \\bar{Y}_{k}\\)) | squared deviation (\\((Y_{ik} - \\bar{Y}_{k})^2\\)) |\n | --- | --- | --- | --- | --- |\nplacebo | 0.5 | 0.45 | 0.05 | 0.0025 |\nplacebo | 0.3 | 0.45 | -0.15 | 0.0225 |\nplacebo | 0.1 | 0.45 | -0.35 | 0.1225 |\nanxifree | 0.6 | 0.72 | -0.12 | 0.0136 |\nanxifree | 0.4 | 0.72 | -0.32 | 0.1003 |\n\nThe last step is equally straightforward. In order to calculate the within-group sum of squares, we just add up the squared deviations across all observations: \\[ \\begin{array}{rcl} \\mbox{SS}_w &=& 0.0025 + 0.0225 + 0.1225 + 0.0136 + 0.1003 \\\\ &=& 0.2614 \\end{array} \\]\n\nOf course, if we actually wanted to get the right answer, we’d need to do this for all 18 observations in the data set, not just the first five. We could continue with the pencil and paper calculations if we wanted to, but it’s pretty tedious. Alternatively, it’s not too hard to get R to do it. Here’s how:\n\n```\noutcome <- clin.trial$mood.gain\ngroup <- clin.trial$drug\ngp.means <- tapply(outcome,group,mean)\ngp.means <- gp.means[group]\ndev.from.gp.means <- outcome - gp.means\nsquared.devs <- dev.from.gp.means ^2\n```\n\n It might not be obvious from inspection what these commands are doing: as a general rule, the human brain seems to just shut down when faced with a big block of programming. However, I strongly suggest that – if you’re like me and tend to find that the mere sight of this code makes you want to look away and see if there’s any beer left in the fridge or a game of footy on the telly – you take a moment and look closely at these commands one at a time. Every single one of these commands is something you’ve seen before somewhere else in the book. There’s nothing novel about them (though I’ll have to admit that the `tapply()` function takes a while to get a handle on), so if you’re not quite sure how these commands work, this might be a good time to try playing around with them yourself, to try to get a sense of what’s happening. On the other hand, if this does seem to make sense, then you won’t be all that surprised at what happens when I wrap these variables in a data frame, and print it out… \n\n```\nY <- data.frame( group, outcome, gp.means,\n dev.from.gp.means, squared.devs )\nprint(Y, digits = 2)\n```\n\n```\n## group outcome gp.means dev.from.gp.means squared.devs\n## 1 placebo 0.5 0.45 0.050 0.0025\n## 2 placebo 0.3 0.45 -0.150 0.0225\n## 3 placebo 0.1 0.45 -0.350 0.1225\n## 4 anxifree 0.6 0.72 -0.117 0.0136\n## 5 anxifree 0.4 0.72 -0.317 0.1003\n## 6 anxifree 0.2 0.72 -0.517 0.2669\n## 7 joyzepam 1.4 1.48 -0.083 0.0069\n## 8 joyzepam 1.7 1.48 0.217 0.0469\n## 9 joyzepam 1.3 1.48 -0.183 0.0336\n## 10 placebo 0.6 0.45 0.150 0.0225\n## 11 placebo 0.9 0.45 0.450 0.2025\n## 12 placebo 0.3 0.45 -0.150 0.0225\n## 13 anxifree 1.1 0.72 0.383 0.1469\n## 14 anxifree 0.8 0.72 0.083 0.0069\n## 15 anxifree 1.2 0.72 0.483 0.2336\n## 16 joyzepam 1.8 1.48 0.317 0.1003\n## 17 joyzepam 1.3 1.48 -0.183 0.0336\n## 18 joyzepam 1.4 1.48 -0.083 0.0069\n```\n\n If you compare this output to the contents of the table I’ve been constructing by hand, you can see that R has done exactly the same calculations that I was doing, and much faster too. So, if we want to finish the calculations of the within-group sum of squares in R, we just ask for the `sum()` of the `squared.devs` variable: \n\n```\nSSw <- sum( squared.devs )\nprint( SSw )\n```\n\n `## [1] 1.391667` Obviously, this isn’t the same as what I calculated, because R used all 18 observations. But if I’d typed\n\n```\nsum( squared.devs[1:5] )\n```\n\ninstead, it would have given the same answer that I got earlier. \nOkay. Now that we’ve calculated the within groups variation, \\(\\mbox{SS}_w\\), it’s time to turn our attention to the between-group sum of squares, \\(\\mbox{SS}_b\\). The calculations for this case are very similar. The main difference is that, instead of calculating the differences between an observation \\(Y_{ik}\\) and a group mean \\(\\bar{Y}_k\\) for all of the observations, we calculate the differences between the group means \\(\\bar{Y}_k\\) and the grand mean \\(\\bar{Y}\\) (in this case 0.88) for all of the groups…\n\ngroup (\\(k\\)) | group mean (\\(\\bar{Y}_k\\)) | grand mean (\\(\\bar{Y}\\)) | deviation (\\(\\bar{Y}_{k} - \\bar{Y}\\)) | squared deviations (\\((\\bar{Y}_{k} - \\bar{Y})^2\\)) |\n | --- | --- | --- | --- | --- |\nplacebo | 0.45 | 0.88 | -0.43 | 0.18 |\nanxifree | 0.72 | 0.88 | -0.16 | 0.03 |\njoyzepam | 1.48 | 0.88 | 0.60 | 0.36 |\n\nHowever, for the between group calculations we need to multiply each of these squared deviations by \\(N_k\\), the number of observations in the group. We do this because every observation in the group (all \\(N_k\\) of them) is associated with a between group difference. So if there are six people in the placebo group, and the placebo group mean differs from the grand mean by 0.19, then the total between group variation associated with these six people is \\(6 \\times 0.16 = 1.14\\). So we have to extend our little table of calculations…\n\ngroup (\\(k\\)) | squared deviations (\\((\\bar{Y}_{k} - \\bar{Y})^2\\)) | sample size (\\(N_k\\)) | weighted squared dev (\\(N_k (\\bar{Y}_{k} - \\bar{Y})^2\\)) |\n | --- | --- | --- | --- |\nplacebo | 0.18 | 6 | 1.11 |\nanxifree | 0.03 | 6 | 0.16 |\njoyzepam | 0.36 | 6 | 2.18 |\n\nAnd so now our between group sum of squares is obtained by summing these “weighted squared deviations” over all three groups in the study: \\[ \\begin{array}{rcl} \\mbox{SS}_{b} &=& 1.11 + 0.16 + 2.18 \\\\ &=& 3.45 \\end{array} \\] As you can see, the between group calculations are a lot shorter, so you probably wouldn’t usually want to bother using R as your calculator. However, if you did decide to do so, here’s one way you could do it:\n\n```\ngp.means <- tapply(outcome,group,mean)\ngrand.mean <- mean(outcome)\ndev.from.grand.mean <- gp.means - grand.mean\nsquared.devs <- dev.from.grand.mean ^2\ngp.sizes <- tapply(outcome,group,length)\nwt.squared.devs <- gp.sizes * squared.devs\n```\n\nAgain, I won’t actually try to explain this code line by line, but – just like last time – there’s nothing in there that we haven’t seen in several places elsewhere in the book, so I’ll leave it as an exercise for you to make sure you understand it. Once again, we can dump all our variables into a data frame so that we can print it out as a nice table:\n\n```\nY <- data.frame( gp.means, grand.mean, dev.from.grand.mean, \n squared.devs, gp.sizes, wt.squared.devs )\nprint(Y, digits = 2)\n```\n\n```\n## gp.means grand.mean dev.from.grand.mean squared.devs gp.sizes\n## placebo 0.45 0.88 -0.43 0.188 6\n## anxifree 0.72 0.88 -0.17 0.028 6\n## joyzepam 1.48 0.88 0.60 0.360 6\n## wt.squared.devs\n## placebo 1.13\n## anxifree 0.17\n## joyzepam 2.16\n```\n\nClearly, these are basically the same numbers that we got before. There are a few tiny differences, but that’s only because the hand-calculated versions have some small errors caused by the fact that I rounded all my numbers to 2 decimal places at each step in the calculations, whereas R only does it at the end (obviously, R s version is more accurate). Anyway, here’s the R command showing the final step:\n\n```\nSSb <- sum( wt.squared.devs )\nprint( SSb )\n```\n\n `## [1] 3.453333` \nwhich is (ignoring the slight differences due to rounding error) the same answer that I got when doing things by hand.\n Now that we’ve calculated our sums of squares values, \\(\\mbox{SS}_b\\) and \\(\\mbox{SS}_w\\), the rest of the ANOVA is pretty painless. The next step is to calculate the degrees of freedom. Since we have \\(G = 3\\) groups and \\(N = 18\\) observations in total, our degrees of freedom can be calculated by simple subtraction: \\[ \\begin{array}{lclcl} \\mbox{df}_b &=& G - 1 &=& 2 \\\\ \\mbox{df}_w &=& N - G &=& 15 \\end{array} \\] Next, since we’ve now calculated the values for the sums of squares and the degrees of freedom, for both the within-groups variability and the between-groups variability, we can obtain the mean square values by dividing one by the other: \\[ \\begin{array}{lclclcl} \\mbox{MS}_b &=& \\displaystyle\\frac{\\mbox{SS}_b }{ \\mbox{df}_b } &=& \\displaystyle\\frac{3.45}{ 2} &=& 1.73 \\end{array} \\] \\[ \\begin{array}{lclclcl} \\mbox{MS}_w &=& \\displaystyle\\frac{\\mbox{SS}_w }{ \\mbox{df}_w } &=& \\displaystyle\\frac{1.39}{15} &=& 0.09 \\end{array} \\] We’re almost done. The mean square values can be used to calculate the \\(F\\)-value, which is the test statistic that we’re interested in. We do this by dividing the between-groups MS value by the and within-groups MS value. \\[ F \\ = \\ \\frac{\\mbox{MS}_b }{ \\mbox{MS}_w } \\ = \\ \\frac{1.73}{0.09} \\ = \\ 18.6 \\] Woohooo! This is terribly exciting, yes? Now that we have our test statistic, the last step is to find out whether the test itself gives us a significant result. As discussed in Chapter @ref(hypothesistesting, what we really ought to do is choose an \\(\\alpha\\) level (i.e., acceptable Type I error rate) ahead of time, construct our rejection region, etc etc. But in practice it’s just easier to directly calculate the \\(p\\)-value. Back in the “old days”, what we’d do is open up a statistics textbook or something and flick to the back section which would actually have a huge lookup table… that’s how we’d “compute” our \\(p\\)-value, because it’s too much effort to do it any other way. However, since we have access to R, I’ll use the `pf()` function to do it instead. Now, remember that I explained earlier that the \\(F\\)-test is always one sided? And that we only reject the null hypothesis for very large \\(F\\)-values? That means we’re only interested in the upper tail of the \\(F\\)-distribution. The command that you’d use here would be this… \n\n```\npf( 18.6, df1 = 2, df2 = 15, lower.tail = FALSE)\n```\n\n `## [1] 8.672727e-05` \nTherefore, our \\(p\\)-value comes to 0.0000867, or \\(8.67 \\times 10^{-5}\\) in scientific notation. So, unless we’re being extremely conservative about our Type I error rate, we’re pretty much guaranteed to reject the null hypothesis.\n\nAt this point, we’re basically done. Having completed our calculations, it’s traditional to organise all these numbers into an ANOVA table like the one in Table@reftab:anovatable. For our clinical trial data, the ANOVA table would look like this:\n\ndf | sum of squares | mean squares | \\(F\\)-statistic | \\(p\\)-value |\n | --- | --- | --- | --- | --- |\nbetween groups | 2 | 3.45 | 1.73 | 18.6 | \\(8.67 \\times 10^{-5}\\) |\nwithin groups | 15 | 1.39 | 0.09 | - | - |\n\nThese days, you’ll probably never have much reason to want to construct one of these tables yourself, but you will find that almost all statistical software (R included) tends to organise the output of an ANOVA into a table like this, so it’s a good idea to get used to reading them. However, although the software will output a full ANOVA table, there’s almost never a good reason to include the whole table in your write up. A pretty standard way of reporting this result would be to write something like this:\n One-way ANOVA showed a significant effect of drug on mood gain (\\(F(2,15) = 18.6, p<.001\\)).\n\nSigh. So much work for one short sentence.\n\n## 14.3 Running an ANOVA in R\n\nI’m pretty sure I know what you’re thinking after reading the last section, especially if you followed my advice and tried typing all the commands in yourself…. doing the ANOVA calculations yourself sucks. There’s quite a lot of calculations that we needed to do along the way, and it would be tedious to have to do this over and over again every time you wanted to do an ANOVA. One possible solution to the problem would be to take all these calculations and turn them into some R functions yourself. You’d still have to do a lot of typing, but at least you’d only have to do it the one time: once you’ve created the functions, you can reuse them over and over again. However, writing your own functions is a lot of work, so this is kind of a last resort. Besides, it’s much better if someone else does all the work for you…\n\n### 14.3.1 Using the\n\n `aov()` function to specify your ANOVA To make life easier for you, R provides a function called `aov()` , which – obviously – is an acronym of “Analysis Of Variance.”206 If you type `?aov` and have a look at the help documentation, you’ll see that there are several arguments to the `aov()` function, but the only two that we’re interested in are `formula` and `data` . As we’ve seen in a few places previously, the `formula` argument is what you use to specify the outcome variable and the grouping variable, and the `data` argument is what you use to specify the data frame that stores these variables. In other words, to do the same ANOVA that I laboriously calculated in the previous section, I’d use a command like this: \n\n```\naov( formula = mood.gain ~ drug, data = clin.trial )\n```\n\nActually, that’s not quite the whole story, as you’ll see as soon as you look at the output from this command, which I’ve hidden for the moment in order to avoid confusing you. Before we go into specifics, I should point out that either of these commands will do the same thing:\n\n```\naov( clin.trial$mood.gain ~ clin.trial$drug ) \naov( mood.gain ~ drug, clin.trial )\n```\n\n In the first command, I didn’t specify a `data` set, and instead relied on the `$` operator to tell R how to find the variables. In the second command, I dropped the argument names, which is okay in this case because `formula` is the first argument to the `aov()` function, and `data` is the second one. Regardless of how I specify the ANOVA, I can assign the output of the `aov()` function to a variable, like this for example: \n\n```\nmy.anova <- aov( mood.gain ~ drug, clin.trial )\n```\n\n This is almost always a good thing to do, because there’s lots of useful things that we can do with the `my.anova` variable. So let’s assume that it’s this last command that I used to specify the ANOVA that I’m trying to run, and as a consequence I have this `my.anova` variable sitting in my workspace, waiting for me to do something with it… \n\n### 14.3.2 Understanding what the\n\n `aov()` function produces Now that we’ve seen how to use the `aov()` function to create `my.anova` we’d better have a look at what this variable actually is. The first thing to do is to check to see what class of variable we’ve created, since it’s kind of interesting in this case. When we do that… `class( my.anova )` `## [1] \"aov\" \"lm\"` … we discover that `my.anova` actually has two classes! The first class tells us that it’s an `aov` (analysis of variance) object, but the second tells us that it’s also an `lm` (linear model) object. Later on, we’ll see that this reflects a pretty deep statistical relationship between ANOVA and regression (Chapter 15 and it means that any function that exists in R for dealing with regressions can also be applied to `aov` objects, which is neat; but I’m getting ahead of myself. For now, I want to note that what we’ve created is an `aov` object, and to also make the point that `aov` objects are actually rather complicated beasts. I won’t be trying to explain everything about them, since it’s way beyond the scope of an introductory statistics subject, but to give you a tiny hint of some of the stuff that R stores inside an `aov` object, let’s ask it to print out the `names()` of all the stored quantities… `names( my.anova )` \n\n```\n## [1] \"coefficients\" \"residuals\" \"effects\" \"rank\" \n## [5] \"fitted.values\" \"assign\" \"qr\" \"df.residual\" \n## [9] \"contrasts\" \"xlevels\" \"call\" \"terms\" \n## [13] \"model\"\n```\n\n As we go through the rest of the book, I hope that a few of these will become a little more obvious to you, but right now that’s going to look pretty damned opaque. That’s okay. You don’t need to know any of the details about it right now, and most of it you don’t need at all… what you do need to understand is that the `aov()` function does a lot of calculations for you, not just the basic ones that I outlined in the previous sections. What this means is that it’s generally a good idea to create a variable like `my.anova` that stores the output of the `aov()` function… because later on, you can use `my.anova` as an input to lots of other functions: those other functions can pull out bits and pieces from the `aov` object, and calculate various other things that you might need. Right then. The simplest thing you can do with an `aov` object is to `print()` it out. When we do that, it shows us a few of the key quantities of interest: `print( my.anova )` \n\n```\n## Call:\n## aov(formula = mood.gain ~ drug, data = clin.trial)\n## \n## Terms:\n## drug Residuals\n## Sum of Squares 3.453333 1.391667\n## Deg. of Freedom 2 15\n## \n## Residual standard error: 0.3045944\n## Estimated effects may be unbalanced\n```\n\n Specificially, it prints out a reminder of the command that you used when you called `aov()` in the first place, shows you the sums of squares values, the degrees of freedom, and a couple of other quantities that we’re not really interested in right now. Notice, however, that R doesn’t use the names “between-group” and “within-group”. Instead, it tries to assign more meaningful names: in our particular example, the between groups variance corresponds to the effect that the `drug` has on the outcome variable; and the within groups variance is corresponds to the “leftover” variability, so it calls that the residuals. If we compare these numbers to the numbers that I calculated by hand in Section 14.2.5, you can see that they’re identical… the between groups sums of squares is \\(\\mbox{SS}_b = 3.45\\), the within groups sums of squares is \\(\\mbox{SS}_w = 1.39\\), and the degrees of freedom are 2 and 15 repectively. \n\n### 14.3.3 Running the hypothesis tests for the ANOVA\n\n Okay, so we’ve verified that `my.anova` seems to be storing a bunch of the numbers that we’re looking for, but the `print()` function didn’t quite give us the output that we really wanted. Where’s the \\(F\\)-value? The \\(p\\)-value? These are the most important numbers in our hypothesis test, but the `print()` function doesn’t provide them. To get those numbers, we need to use a different function. Instead of asking R to `print()` out the `aov` object, we should have asked for a `summary()` of it.207 When we do that… `summary( my.anova )` \n\n… we get all of the key numbers that we calculated earlier. We get the sums of squares, the degrees of freedom, the mean squares, the \\(F\\)-statistic, and the \\(p\\)-value itself. These are all identical to the numbers that we calculated ourselves when doing it the long and tedious way, and it’s even organised into the same kind of ANOVA table that I showed in Table 14.1, and then filled out by hand in Section 14.2.5. The only things that are even slightly different is that some of the row and column names are a bit different.\n\n## 14.4 Effect size\n\n There’s a few different ways you could measure the effect size in an ANOVA, but the most commonly used measures are \\(\\eta^2\\) (eta squared) and partial \\(\\eta^2\\). For a one way analysis of variance they’re identical to each other, so for the moment I’ll just explain \\(\\eta^2\\). The definition of \\(\\eta^2\\) is actually really simple: \\[ \\eta^2 = \\frac{\\mbox{SS}_b}{\\mbox{SS}_{tot}} \\] That’s all it is. So when I look at the ANOVA table above, I see that \\(\\mbox{SS}_b = 3.45\\) and \\(\\mbox{SS}_{tot} = 3.45 + 1.39 = 4.84\\). Thus we get an \\(\\eta^2\\) value of \\[ \\eta^2 = \\frac{3.45}{4.84} = 0.71 \\] The interpretation of \\(\\eta^2\\) is equally straightforward: it refers to the proportion of the variability in the outcome variable ( `mood.gain` ) that can be explained in terms of the predictor ( `drug` ). A value of \\(\\eta^2 = 0\\) means that there is no relationship at all between the two, whereas a value of \\(\\eta^2 = 1\\) means that the relationship is perfect. Better yet, the \\(\\eta^2\\) value is very closely related to a squared correlation (i.e., \\(r^2\\)). So, if you’re trying to figure out whether a particular value of \\(\\eta^2\\) is big or small, it’s sometimes useful to remember that \\[\n\\eta= \\sqrt{\\frac{\\mbox{SS}_b}{\\mbox{SS}_{tot}}}\n\\] can be interpreted as if it referred to the magnitude of a Pearson correlation. So in our drugs example, the \\(\\eta^2\\) value of .71 corresponds to an \\(\\eta\\) value of \\(\\sqrt{.71} = .84\\). If we think about this as being equivalent to a correlation of about .84, we’d conclude that the relationship between `drug` and `mood.gain` is strong. The core packages in R don’t include any functions for calculating \\(\\eta^2\\). However, it’s pretty straightforward to calculate it directly from the numbers in the ANOVA table. In fact, since I’ve already got the `SSw` and `SSb` variables lying around from my earlier calculations, I can do this: \n\n```\nSStot <- SSb + SSw # total sums of squares\neta.squared <- SSb / SStot # eta-squared value\nprint( eta.squared )\n```\n\n `## [1] 0.7127623` However, since it can be tedious to do this the long way (especially when we start running more complicated ANOVAs, such as those in Chapter 16 I’ve included an `etaSquared()` function in the `lsr` package which will do it for you. For now, the only argument you need to care about is `x` , which should be the `aov` object corresponding to your ANOVA. When we do this, what we get as output is this: \n\n```\netaSquared( x = my.anova )\n```\n\n```\n## eta.sq eta.sq.part\n## drug 0.7127623 0.7127623\n```\n\nThe output here shows two different numbers. The first one corresponds to the \\(\\eta^2\\) statistic, precisely as described above. The second one refers to “partial \\(\\eta^2\\)”, which is a somewhat different measure of effect size that I’ll describe later. For the simple ANOVA that we’ve just run, they’re the same number. But this won’t always be true once we start running more complicated ANOVAs.208\n\n## 14.5 Multiple comparisons and post hoc tests\n\nAny time you run an ANOVA with more than two groups, and you end up with a significant effect, the first thing you’ll probably want to ask is which groups are actually different from one another. In our drugs example, our null hypothesis was that all three drugs (placebo, Anxifree and Joyzepam) have the exact same effect on mood. But if you think about it, the null hypothesis is actually claiming three different things all at once here. Specifically, it claims that:\n\n* Your competitor’s drug (Anxifree) is no better than a placebo (i.e., \\(\\mu_A = \\mu_P\\))\n * Your drug (Joyzepam) is no better than a placebo (i.e., \\(\\mu_J = \\mu_P\\))\n * Anxifree and Joyzepam are equally effective (i.e., \\(\\mu_J = \\mu_A\\))\n\nIf any one of those three claims is false, then the null hypothesis is also false. So, now that we’ve rejected our null hypothesis, we’re thinking that at least one of those things isn’t true. But which ones? All three of these propositions are of interest: you certainly want to know if your new drug Joyzepam is better than a placebo, and it would be nice to know how well it stacks up against an existing commercial alternative (i.e., Anxifree). It would even be useful to check the performance of Anxifree against the placebo: even if Anxifree has already been extensively tested against placebos by other researchers, it can still be very useful to check that your study is producing similar results to earlier work.\n\nWhen we characterise the null hypothesis in terms of these three distinct propositions, it becomes clear that there are eight possible “states of the world” that we need to distinguish between:\n\npossibility: | is \\(\\mu_P = \\mu_A\\)? | is \\(\\mu_P = \\mu_J\\)? | is \\(\\mu_A = \\mu_J\\)? | which hypothesis? |\n | --- | --- | --- | --- | --- |\n1 | \\(\\checkmark\\) | \\(\\checkmark\\) | \\(\\checkmark\\) | null |\n2 | \\(\\checkmark\\) | \\(\\checkmark\\) | alternative |\n3 | \\(\\checkmark\\) | \\(\\checkmark\\) | alternative |\n4 | \\(\\checkmark\\) | alternative |\n5 | \\(\\checkmark\\) | \\(\\checkmark\\) | alternative |\n6 | \\(\\checkmark\\) | alternative |\n7 | \\(\\checkmark\\) | alternative |\n8 | alternative |\n\nBy rejecting the null hypothesis, we’ve decided that we don’t believe that #1 is the true state of the world. The next question to ask is, which of the other seven possibilities do we think is right? When faced with this situation, its usually helps to look at the data. For instance, if we look at the plots in Figure 14.1, it’s tempting to conclude that Joyzepam is better than the placebo and better than Anxifree, but there’s no real difference between Anxifree and the placebo. However, if we want to get a clearer answer about this, it might help to run some tests.\n\n### 14.5.1 Running “pairwise” \\(t\\)-tests\n\n How might we go about solving our problem? Given that we’ve got three separate pairs of means (placebo versus Anxifree, placebo versus Joyzepam, and Anxifree versus Joyzepam) to compare, what we could do is run three separate \\(t\\)-tests and see what happens. There’s a couple of ways that we could do this. One method would be to construct new variables corresponding the groups you want to compare (e.g., `anxifree` , `placebo` and `joyzepam` ), and then run a \\(t\\)-test on these new variables: \n\n```\nt.test( anxifree, placebo, var.equal = TRUE ) # Student t-test\n\nanxifree <- with(clin.trial, mood.gain[drug == \"anxifree\"]) # mood change due to anxifree\nplacebo <- with(clin.trial, mood.gain[drug == \"placebo\"]) # mood change due to placebo\n```\n\n or, you could use the `subset` argument in the `t.test()` function to select only those observations corresponding to one of the two groups we’re interested in: \n\n```\nt.test( formula = mood.gain ~ drug, \n data = clin.trial, \n subset = drug %in% c(\"placebo\",\"anxifree\"), \n var.equal = TRUE \n)\n```\n\n See Chapter 7 if you’ve forgotten how the `%in%` operator works. Regardless of which version we do, R will print out the results of the \\(t\\)-test, though I haven’t included that output here. If we go on to do this for all possible pairs of variables, we can look to see which (if any) pairs of groups are significantly different to each other. This “lots of \\(t\\)-tests idea” isn’t a bad strategy, though as we’ll see later on there are some problems with it. However, for the moment our bigger problem is that it’s a pain to have to type in such a long command over and over again: for instance, if your experiment has 10 groups, then you have to run 45 \\(t\\)-tests. That’s way too much typing. To help keep the typing to a minimum, R provides a function called `pairwise.t.test()` that automatically runs all of the \\(t\\)-tests for you. There are three arguments that you need to specify, the outcome variable `x` , the group variable `g` , and the `p.adjust.method` argument, which “adjusts” the \\(p\\)-value in one way or another. I’ll explain \\(p\\)-value adjustment in a moment, but for now we can just set \n\n```\np.adjust.method = \"none\"\n```\n\nsince we’re not doing any adjustments. For our example, here’s what we do: \n\n```\npairwise.t.test( x = clin.trial$mood.gain, # outcome variable\n g = clin.trial$drug, # grouping variable\n p.adjust.method = \"none\" # which correction to use?\n )\n```\n\n One thing that bugs me slightly about the `pairwise.t.test()` function is that you can’t just give it an `aov` object, and have it produce this output. After all, I went to all that trouble earlier of getting R to create the `my.anova` variable and – as we saw in Section 14.3.2 – R has actually stored enough information inside it that I should just be able to get it to run all the pairwise tests using `my.anova` as an input. To that end, I’ve included a `posthocPairwiseT()` function in the `lsr` package that lets you do this. The idea behind this function is that you can just input the `aov` object itself,209 and then get the pairwise tests as an output. As of the current writing, `posthocPairwiseT()` is actually just a simple way of calling `pairwise.t.test()` function, but you should be aware that I intend to make some changes to it later on. Here’s an example: \n\n```\nposthocPairwiseT( x = my.anova, p.adjust.method = \"none\" )\n```\n\n In later versions, I plan to add more functionality (e.g., adjusted confidence intervals), but for now I think it’s at least kind of useful. To see why, let’s suppose you’ve run your ANOVA and stored the results in `my.anova` , and you’re happy using the Holm correction (the default method in `pairwise.t.test()` , which I’ll explain this in a moment). In that case, all you have to do is type this: \n\nand R will output the test results. Much more convenient, I think.\n\n### 14.5.2 Corrections for multiple testing\n\nIn the previous section I hinted that there’s a problem with just running lots and lots of \\(t\\)-tests. The concern is that when running these analyses, what we’re doing is going on a “fishing expedition”: we’re running lots and lots of tests without much theoretical guidance, in the hope that some of them come up significant. This kind of theory-free search for group differences is referred to as post hoc analysis (“post hoc” being Latin for “after this”).210\n\nIt’s okay to run post hoc analyses, but a lot of care is required. For instance, the analysis that I ran in the previous section is actually pretty dangerous: each individual \\(t\\)-test is designed to have a 5% Type I error rate (i.e., \\(\\alpha = .05\\)), and I ran three of these tests. Imagine what would have happened if my ANOVA involved 10 different groups, and I had decided to run 45 “post hoc” \\(t\\)-tests to try to find out which ones were significantly different from each other, you’d expect 2 or 3 of them to come up significant by chance alone. As we saw in Chapter 11, the central organising principle behind null hypothesis testing is that we seek to control our Type I error rate, but now that I’m running lots of \\(t\\)-tests at once, in order to determine the source of my ANOVA results, my actual Type I error rate across this whole family of tests has gotten completely out of control.\n\nThe usual solution to this problem is to introduce an adjustment to the \\(p\\)-value, which aims to control the total error rate across the family of tests (see Shaffer 1995). An adjustment of this form, which is usually (but not always) applied because one is doing post hoc analysis, is often referred to as a correction for multiple comparisons, though it is sometimes referred to as “simultaneous inference”. In any case, there are quite a few different ways of doing this adjustment. I’ll discuss a few of them in this section and in Section 16.8, but you should be aware that there are many other methods out there (see, e.g., Hsu 1996).\n\n### 14.5.3 Bonferroni corrections\n\nThe simplest of these adjustments is called the Bonferroni correction (Dunn 1961), and it’s very very simple indeed. Suppose that my post hoc analysis consists of \\(m\\) separate tests, and I want to ensure that the total probability of making any Type I errors at all is at most \\(\\alpha\\).211 If so, then the Bonferroni correction just says “multiply all your raw \\(p\\)-values by \\(m\\)”. If we let \\(p\\) denote the original \\(p\\)-value, and let \\(p^\\prime_j\\) be the corrected value, then the Bonferroni correction tells that: \\[ p^\\prime = m \\times p \\] And therefore, if you’re using the Bonferroni correction, you would reject the null hypothesis if \\(p^\\prime < \\alpha\\). The logic behind this correction is very straightforward. We’re doing \\(m\\) different tests; so if we arrange it so that each test has a Type I error rate of at most \\(\\alpha / m\\), then the total Type I error rate across these tests cannot be larger than \\(\\alpha\\). That’s pretty simple, so much so that in the original paper, the author writes:\n The method given here is so simple and so general that I am sure it must have been used before this. I do not find it, however, so can only conclude that perhaps its very simplicity has kept statisticians from realizing that it is a very good method in some situations (pp 52-53 Dunn 1961)\n To use the Bonferroni correction in R, you can use the `pairwise.t.test()` function,212 making sure that you set \n\n```\np.adjust.method = \"bonferroni\"\n```\n\n. Alternatively, since the whole reason why we’re doing these pairwise tests in the first place is because we have an ANOVA that we’re trying to understand, it’s probably more convenient to use the `posthocPairwiseT()` function in the `lsr` package, since we can use `my.anova` as the input: \n\n```\nposthocPairwiseT( my.anova, p.adjust.method = \"bonferroni\")\n```\n\n```\n## \n## Pairwise comparisons using t tests with pooled SD \n## \n## data: mood.gain and drug \n## \n## placebo anxifree\n## anxifree 0.4506 - \n## joyzepam 9.1e-05 0.0017 \n## \n## P value adjustment method: bonferroni\n```\n\nIf we compare these three \\(p\\)-values to those that we saw in the previous section when we made no adjustment at all, it is clear that the only thing that R has done is multiply them by 3.\n\n### 14.5.4 Holm corrections\n\nAlthough the Bonferroni correction is the simplest adjustment out there, it’s not usually the best one to use. One method that is often used instead is the Holm correction (Holm 1979). The idea behind the Holm correction is to pretend that you’re doing the tests sequentially; starting with the smallest (raw) \\(p\\)-value and moving onto the largest one. For the \\(j\\)-th largest of the \\(p\\)-values, the adjustment is either \\[ p^\\prime_j = j \\times p_j \\] (i.e., the biggest \\(p\\)-value remains unchanged, the second biggest \\(p\\)-value is doubled, the third biggest \\(p\\)-value is tripled, and so on), or \\[ p^\\prime_j = p^\\prime_{j+1} \\] whichever one is larger. This might sound a little confusing, so let’s go through it a little more slowly. Here’s what the Holm correction does. First, you sort all of your \\(p\\)-values in order, from smallest to largest. For the smallest \\(p\\)-value all you do is multiply it by \\(m\\), and you’re done. However, for all the other ones it’s a two-stage process. For instance, when you move to the second smallest \\(p\\) value, you first multiply it by \\(m-1\\). If this produces a number that is bigger than the adjusted \\(p\\)-value that you got last time, then you keep it. But if it’s smaller than the last one, then you copy the last \\(p\\)-value. To illustrate how this works, consider the table below, which shows the calculations of a Holm correction for a collection of five \\(p\\)-values:\n\nraw \\(p\\) | rank \\(j\\) | \\(p \\times j\\) | Holm \\(p\\) |\n | --- | --- | --- | --- |\n.001 | 5 | .005 | .005 |\n.005 | 4 | .020 | .020 |\n.019 | 3 | .057 | .057 |\n.022 | 2 | .044 | .057 |\n.103 | 1 | .103 | .103 |\n\nHopefully that makes things clear.\n Although it’s a little harder to calculate, the Holm correction has some very nice properties: it’s more powerful than Bonferroni (i.e., it has a lower Type II error rate), but – counterintuitive as it might seem – it has the same Type I error rate. As a consequence, in practice there’s never any reason to use the simpler Bonferroni correction, since it is always outperformed by the slightly more elaborate Holm correction. Because of this, the Holm correction is the default one used by `pairwise.t.test()` and `posthocPairwiseT()` . To run the Holm correction in R, you could specify \n\n```\np.adjust.method = \"Holm\"\n```\n\nif you wanted to, but since it’s the default you can just to do this: \n\n```\n## \n## Pairwise comparisons using t tests with pooled SD \n## \n## data: mood.gain and drug \n## \n## placebo anxifree\n## anxifree 0.1502 - \n## joyzepam 9.1e-05 0.0011 \n## \n## P value adjustment method: holm\n```\n\nAs you can see, the biggest \\(p\\)-value (corresponding to the comparison between Anxifree and the placebo) is unaltered: at a value of \\(.15\\), it is exactly the same as the value we got originally when we applied no correction at all. In contrast, the smallest \\(p\\)-value (Joyzepam versus placebo) has been multiplied by three.\n\n### 14.5.5 Writing up the post hoc test\n\nFinally, having run the post hoc analysis to determine which groups are significantly different to one another, you might write up the result like this:\n Post hoc tests (using the Holm correction to adjust \\(p\\)) indicated that Joyzepam produced a significantly larger mood change than both Anxifree (\\(p = .001\\)) and the placebo (\\(p = 9.1 \\times 10^{-5}\\)). We found no evidence that Anxifree performed better than the placebo (\\(p = .15\\)).\n\nOr, if you don’t like the idea of reporting exact \\(p\\)-values, then you’d change those numbers to \\(p<.01\\), \\(p<.001\\) and \\(p > .05\\) respectively. Either way, the key thing is that you indicate that you used Holm’s correction to adjust the \\(p\\)-values. And of course, I’m assuming that elsewhere in the write up you’ve included the relevant descriptive statistics (i.e., the group means and standard deviations), since these \\(p\\)-values on their own aren’t terribly informative.\n\n## 14.6 Assumptions of one-way ANOVA\n\nLike any statistical test, analysis of variance relies on some assumptions about the data. There are three key assumptions that you need to be aware of: normality, homogeneity of variance and independence. If you remember back to Section 14.2.4 – which I hope you at least skimmed even if you didn’t read the whole thing – I described the statistical models underpinning ANOVA, which I wrote down like this: \\[ \\begin{array}{lrcl} H_0: & Y_{ik} &=& \\mu + \\epsilon_{ik} \\\\ H_1: & Y_{ik} &=& \\mu_k + \\epsilon_{ik} \\end{array} \\] In these equations \\(\\mu\\) refers to a single, grand population mean which is the same for all groups, and \\(\\mu_k\\) is the population mean for the \\(k\\)-th group. Up to this point we’ve been mostly interested in whether our data are best described in terms of a single grand mean (the null hypothesis) or in terms of different group-specific means (the alternative hypothesis). This makes sense, of course: that’s actually the important research question! However, all of our testing procedures have – implicitly – relied on a specific assumption about the residuals, \\(\\epsilon_{ik}\\), namely that \\[ \\epsilon_{ik} \\sim \\mbox{Normal}(0, \\sigma^2) \\] None of the maths works properly without this bit. Or, to be precise, you can still do all the calculations, and you’ll end up with an \\(F\\)-statistic, but you have no guarantee that this \\(F\\)-statistic actually measures what you think it’s measuring, and so any conclusions that you might draw on the basis of the \\(F\\) test might be wrong.\n\nSo, how do we check whether this assumption about the residuals is accurate? Well, as I indicated above, there are three distinct claims buried in this one statement, and we’ll consider them separately.\n\n* Normality. The residuals are assumed to be normally distributed. As we saw in Section 13.9, we can assess this by looking at QQ plots or running a Shapiro-Wilk test. I’ll talk about this in an ANOVA context in Section 14.9.\n * Homogeneity of variance. Notice that we’ve only got the one value for the population standard deviation (i.e., \\(\\sigma\\)), rather than allowing each group to have it’s own value (i.e., \\(\\sigma_k\\)). This is referred to as the homogeneity of variance (sometimes called homoscedasticity) assumption. ANOVA assumes that the population standard deviation is the same for all groups. We’ll talk about this extensively in Section 14.7.\n * Independence. The independence assumption is a little trickier. What it basically means is that, knowing one residual tells you nothing about any other residual. All of the \\(\\epsilon_{ik}\\) values are assumed to have been generated without any “regard for” or “relationship to” any of the other ones. There’s not an obvious or simple way to test for this, but there are some situations that are clear violations of this: for instance, if you have a repeated-measures design, where each participant in your study appears in more than one condition, then independence doesn’t hold; there’s a special relationship between some observations… namely those that correspond to the same person! When that happens, you need to use something like repeated measures ANOVA. I don’t currently talk about repeated measures ANOVA in this book, but it will be included in later versions.\n\n### 14.6.1 How robust is ANOVA?\n\nOne question that people often want to know the answer to is the extent to which you can trust the results of an ANOVA if the assumptions are violated. Or, to use the technical language, how robust is ANOVA to violations of the assumptions. Due to deadline constraints I don’t have the time to discuss this topic. This is a topic I’ll cover in some detail in a later version of the book.\n\n## 14.7 Checking the homogeneity of variance assumption\n\n There’s more than one way to skin a cat, as the saying goes, and more than one way to test the homogeneity of variance assumption, too (though for some reason no-one made a saying out of that). The most commonly used test for this that I’ve seen in the literature is the Levene test (Levene 1960), and the closely related Brown-Forsythe test (Brown and Forsythe 1974), both of which I’ll describe here. Alternatively, you could use the Bartlett test, which is implemented in R via the `bartlett.test()` function, but I’ll leave it as an exercise for the reader to go check that one out if you’re interested. \nLevene’s test is shockingly simple. Suppose we have our outcome variable \\(Y_{ik}\\). All we do is define a new variable, which I’ll call \\(Z_{ik}\\), corresponding to the absolute deviation from the group mean: \\[ Z_{ik} = \\left| Y_{ik} - \\bar{Y}_k \\right| \\] Okay, what good does this do us? Well, let’s take a moment to think about what \\(Z_{ik}\\) actually is, and what we’re trying to test. The value of \\(Z_{ik}\\) is a measure of how the \\(i\\)-th observation in the \\(k\\)-th group deviates from its group mean. And our null hypothesis is that all groups have the same variance; that is, the same overall deviations from the group means! So, the null hypothesis in a Levene’s test is that the population means of \\(Z\\) are identical for all groups. Hm. So what we need now is a statistical test of the null hypothesis that all group means are identical. Where have we seen that before? Oh right, that’s what ANOVA is… and so all that the Levene’s test does is run an ANOVA on the new variable \\(Z_{ik}\\).\n\nWhat about the Brown-Forsythe test? Does that do anything particularly different? Nope. The only change from the Levene’s test is that it constructs the transformed variable \\(Z\\) in a slightly different way, using deviations from the group medians rather than deviations from the group means. That is, for the Brown-Forsythe test, \\[ Z_{ik} = \\left| Y_{ik} - \\mbox{median}_k(Y) \\right| \\] where \\(\\mbox{median}_k(Y)\\) is the median for group \\(k\\). Regardless of whether you’re doing the standard Levene test or the Brown-Forsythe test, the test statistic – which is sometimes denoted \\(F\\), but sometimes written as \\(W\\) – is calculated in exactly the same way that the \\(F\\)-statistic for the regular ANOVA is calculated, just using a \\(Z_{ik}\\) rather than \\(Y_{ik}\\). With that in mind, let’s just move on and look at how to run the test in R.\n\n### 14.7.1 Running the Levene’s test in R\n\n Okay, so how do we run the Levene test? Obviously, since the Levene test is just an ANOVA, it would be easy enough to manually create the transformed variable \\(Z_{ik}\\) and then use the `aov()` function to run an ANOVA on that. However, that’s the tedious way to do it. A better way to do run your Levene’s test is to use the `leveneTest()` function, which is in the `car` package. As usual, we first load the package `library( car )` and now that we have, we can run our Levene test. The main argument that you need to specify is `y` , but you can do this in lots of different ways. Probably the simplest way to do it is actually input the original `aov` object. Since I’ve got the `my.anova` variable stored from my original ANOVA, I can just do this: \n\n```\nleveneTest( my.anova )\n```\n\n If we look at the output, we see that the test is non-significant \\((F_{2,15} = 1.47, p = .26)\\), so it looks like the homogeneity of variance assumption is fine. Remember, although R reports the test statistic as an \\(F\\)-value, it could equally be called \\(W\\), in which case you’d just write \\(W_{2,15} = 1.47\\). Also, note the part of the output that says `center = median` . That’s telling you that, by default, the `leveneTest()` function actually does the Brown-Forsythe test. If you want to use the mean instead, then you need to explicitly set the `center` argument, like this: \n\n```\nleveneTest( y = my.anova, center = mean )\n```\n\nThat being said, in most cases it’s probably best to stick to the default value, since the Brown-Forsythe test is a bit more robust than the original Levene test.\n\n### 14.7.2 Additional comments\n\n Two more quick comments before I move onto a different topic. Firstly, as mentioned above, there are other ways of calling the `leveneTest()` function. Although the vast majority of situations that call for a Levene test involve checking the assumptions of an ANOVA (in which case you probably have a variable like `my.anova` lying around), sometimes you might find yourself wanting to specify the variables directly. Two different ways that you can do this are shown below: \n\n```\nleveneTest(y = mood.gain ~ drug, data = clin.trial) # y is a formula in this case\nleveneTest(y = clin.trial$mood.gain, group = clin.trial$drug) # y is the outcome\n```\n\n Secondly, I did mention that it’s possible to run a Levene test just using the `aov()` function. I don’t want to waste a lot of space on this, but just in case some readers are interested in seeing how this is done, here’s the code that creates the new variables and runs an ANOVA. If you are interested, feel free to run this to verify that it produces the same answers as the Levene test (i.e., with `center = mean` ): \n\n```\nY <- clin.trial $ mood.gain # the original outcome variable, Y\nG <- clin.trial $ drug # the grouping variable, G\ngp.mean <- tapply(Y, G, mean) # calculate group means\nYbar <- gp.mean[G] # group mean associated with each obs\nZ <- abs(Y - Ybar) # the transformed variable, Z\nsummary( aov(Z ~ G) ) # run the ANOVA\n```\n\n```\n## Df Sum Sq Mean Sq F value Pr(>F)\n## G 2 0.0616 0.03080 1.45 0.266\n## Residuals 15 0.3187 0.02125\n```\n\nThat said, I don’t imagine that many people will care about this. Nevertheless, it’s nice to know that you could do it this way if you wanted to. And for those of you who do try it, I think it helps to demystify the test a little bit when you can see – with your own eyes – the way in which Levene’s test relates to ANOVA.\n\n## 14.8 Removing the homogeneity of variance assumption\n\n In our example, the homogeneity of variance assumption turned out to be a pretty safe one: the Levene test came back non-significant, so we probably don’t need to worry. However, in real life we aren’t always that lucky. How do we save our ANOVA when the homogeneity of variance assumption is violated? If you recall from our discussion of \\(t\\)-tests, we’ve seen this problem before. The Student \\(t\\)-test assumes equal variances, so the solution was to use the Welch \\(t\\)-test, which does not. In fact, Welch (1951) also showed how we can solve this problem for ANOVA too (the Welch one-way test). It’s implemented in R using the `oneway.test()` function. The arguments that we’ll need for our example are: \n\n* \n `formula` . This is the model formula, which (as usual) needs to specify the outcome variable on the left hand side and the grouping variable on the right hand side: i.e., something like `outcome ~ group` . * \n `data` . Specifies the data frame containing the variables. * \n `var.equal` . If this is `FALSE` (the default) a Welch one-way test is run. If it is `TRUE` then it just runs a regular ANOVA. The function also has a `subset` argument that lets you analyse only some of the observations and a `na.action` argument that tells it how to handle missing data, but these aren’t necessary for our purposes. So, to run the Welch one-way ANOVA for our example, we would do this: \n\n```\noneway.test(mood.gain ~ drug, data = clin.trial)\n```\n\n```\n## \n## One-way analysis of means (not assuming equal variances)\n## \n## data: mood.gain and drug\n## F = 26.322, num df = 2.0000, denom df = 9.4932, p-value = 0.000134\n```\n\n To understand what’s happening here, let’s compare these numbers to what we got earlier in Section 14.3 when we ran our original ANOVA. To save you the trouble of flicking back, here are those numbers again, this time calculated by setting `var.equal = TRUE` for the `oneway.test()` function: \n\n```\noneway.test(mood.gain ~ drug, data = clin.trial, var.equal = TRUE)\n```\n\n```\n## \n## One-way analysis of means\n## \n## data: mood.gain and drug\n## F = 18.611, num df = 2, denom df = 15, p-value = 8.646e-05\n```\n\nOkay, so originally our ANOVA gave us the result \\(F(2,15) = 18.6\\), whereas the Welch one-way test gave us \\(F(2,9.49) = 26.32\\). In other words, the Welch test has reduced the within-groups degrees of freedom from 15 to 9.49, and the \\(F\\)-value has increased from 18.6 to 26.32.\n\n## 14.9 Checking the normality assumption\n\n Testing the normality assumption is relatively straightforward. We covered most of what you need to know in Section 13.9. The only thing we really need to know how to do is pull out the residuals (i.e., the \\(\\epsilon_{ik}\\) values) so that we can draw our QQ plot and run our Shapiro-Wilk test. First, let’s extract the residuals. R provides a function called `residuals()` that will do this for us. If we pass our `my.anova` to this function, it will return the residuals. So let’s do that: \n\n```\nmy.anova.residuals <- residuals( object = my.anova ) # extract the residuals\n```\n\nWe can print them out too, though it’s not exactly an edifying experience. In fact, given that I’m on the verge of putting myself to sleep just typing this, it might be a good idea to skip that step. Instead, let’s draw some pictures and run ourselves a hypothesis test:\n\n```\nhist( x = my.anova.residuals ) # plot a histogram (similar to Figure @ref{fig:normalityanova}a)\n```\n\n```\nqqnorm( y = my.anova.residuals ) # draw a QQ plot (similar to Figure @ref{fig:normalityanova}b)\n```\n\n```\nshapiro.test( x = my.anova.residuals ) # run Shapiro-Wilk test\n```\n\n```\n## \n## Shapiro-Wilk normality test\n## \n## data: my.anova.residuals\n## W = 0.96019, p-value = 0.6053\n```\n\nThe histogram and QQ plot are both look pretty normal to me.213 This is supported by the results of our Shapiro-Wilk test (\\(W = .96\\), \\(p = .61\\)) which finds no indication that normality is violated.\n\n## 14.10 Removing the normality assumption\n\nNow that we’ve seen how to check for normality, we are led naturally to ask what we can do to address violations of normality. In the context of a one-way ANOVA, the easiest solution is probably to switch to a non-parametric test (i.e., one that doesn’t rely on any particular assumption about the kind of distribution involved). We’ve seen non-parametric tests before, in Chapter 13: when you only have two groups, the Wilcoxon test provides the non-parametric alternative that you need. When you’ve got three or more groups, you can use the Kruskal-Wallis rank sum test (Kruskal and Wallis 1952). So that’s the test we’ll talk about next.\n\n### 14.10.1 The logic behind the Kruskal-Wallis test\n\nThe Kruskal-Wallis test is surprisingly similar to ANOVA, in some ways. In ANOVA, we started with \\(Y_{ik}\\), the value of the outcome variable for the \\(i\\)th person in the \\(k\\)th group. For the Kruskal-Wallis test, what we’ll do is rank order all of these \\(Y_{ik}\\) values, and conduct our analysis on the ranked data. So let’s let \\(R_{ik}\\) refer to the ranking given to the \\(i\\)th member of the \\(k\\)th group. Now, let’s calculate \\(\\bar{R}_k\\), the average rank given to observations in the \\(k\\)th group: \\[ \\bar{R}_k = \\frac{1}{N_K} \\sum_{i} R_{ik} \\] and let’s also calculate \\(\\bar{R}\\), the grand mean rank: \\[ \\bar{R} = \\frac{1}{N} \\sum_{i} \\sum_{k} R_{ik} \\] Now that we’ve done this, we can calculate the squared deviations from the grand mean rank \\(\\bar{R}\\). When we do this for the individual scores – i.e., if we calculate \\((R_{ik} - \\bar{R})^2\\) – what we have is a “nonparametric” measure of how far the \\(ik\\)-th observation deviates from the grand mean rank. When we calculate the squared deviation of the group means from the grand means – i.e., if we calculate \\((\\bar{R}_k - \\bar{R} )^2\\) – then what we have is a nonparametric measure of how much the group deviates from the grand mean rank. With this in mind, let’s follow the same logic that we did with ANOVA, and define our ranked sums of squares measures in much the same way that we did earlier. First, we have our “total ranked sums of squares”: \\[ \\mbox{RSS}_{tot} = \\sum_k \\sum_i ( R_{ik} - \\bar{R} )^2 \\] and we can define the “between groups ranked sums of squares” like this: \\[ \\begin{array}{rcl} \\mbox{RSS}_{b} &=& \\sum_k \\sum_i ( \\bar{R}_k - \\bar{R} )^2 \\\\ &=& \\sum_k N_k ( \\bar{R}_k - \\bar{R} )^2 \\end{array} \\] So, if the null hypothesis is true and there are no true group differences at all, you’d expect the between group rank sums \\(\\mbox{RSS}_{b}\\) to be very small, much smaller than the total rank sums \\(\\mbox{RSS}_{tot}\\). Qualitatively this is very much the same as what we found when we went about constructing the ANOVA \\(F\\)-statistic; but for technical reasons the Kruskal-Wallis test statistic, usually denoted \\(K\\), is constructed in a slightly different way: \\[ K = (N - 1) \\times \\frac{\\mbox{RSS}_b}{\\mbox{RSS}_{tot}} \\] and, if the null hypothesis is true, then the sampling distribution of \\(K\\) is approximately chi-square with \\(G-1\\) degrees of freedom (where \\(G\\) is the number of groups). The larger the value of \\(K\\), the less consistent the data are with null hypothesis, so this is a one-sided test: we reject \\(H_0\\) when \\(K\\) is sufficiently large.\n\n### 14.10.2 Additional details\n\nThe description in the previous section illustrates the logic behind the Kruskal-Wallis test. At a conceptual level, this is the right way to think about how the test works. However, from a purely mathematical perspective it’s needlessly complicated. I won’t show you the derivation, but you can use a bit of algebraic jiggery-pokery214 to show that the equation for \\(K\\) can be rewritten as \\[ K = \\frac{12}{N(N-1)} \\sum_k N_k {\\bar{R}_k}^2 - 3(N+1) \\] It’s this last equation that you sometimes see given for \\(K\\). This is way easier to calculate than the version I described in the previous section, it’s just that it’s totally meaningless to actual humans. It’s probably best to think of \\(K\\) the way I described it earlier… as an analogue of ANOVA based on ranks. But keep in mind that the test statistic that gets calculated ends up with a rather different look to it than the one we used for our original ANOVA.\n\nBut wait, there’s more! Dear lord, why is there always more? The story I’ve told so far is only actually true when there are no ties in the raw data. That is, if there are no two observations that have exactly the same value. If there are ties, then we have to introduce a correction factor to these calculations. At this point I’m assuming that even the most diligent reader has stopped caring (or at least formed the opinion that the tie-correction factor is something that doesn’t require their immediate attention). So I’ll very quickly tell you how it’s calculated, and omit the tedious details about why it’s done this way. Suppose we construct a frequency table for the raw data, and let \\(f_j\\) be the number of observations that have the \\(j\\)-th unique value. This might sound a bit abstract, so here’s the R code showing a concrete example:\n\n```\nf <- table( clin.trial$mood.gain ) # frequency table for mood gain\nprint(f) # we have some ties\n```\n\n```\n## \n## 0.1 0.2 0.3 0.4 0.5 0.6 0.8 0.9 1.1 1.2 1.3 1.4 1.7 1.8 \n## 1 1 2 1 1 2 1 1 1 1 2 2 1 1\n```\n\n Looking at this table, notice that the third entry in the frequency table has a value of \\(2\\). Since this corresponds to a `mood.gain` of 0.3, this table is telling us that two people’s mood increased by 0.3. More to the point, note that we can say that `f[3]` has a value of `2` . Or, in the mathematical notation I introduced above, this is telling us that \\(f_3 = 2\\). Yay. So, now that we know this, the tie correction factor (TCF) is: \\[\n\\mbox{TCF} = 1 - \\frac{\\sum_j {f_j}^3 - f_j}{N^3 - N} \n\\] The tie-corrected value of the Kruskal-Wallis statistic obtained by dividing the value of \\(K\\) by this quantity: it is this tie-corrected version that R calculates. And at long last, we’re actually finished with the theory of the Kruskal-Wallis test. I’m sure you’re all terribly relieved that I’ve cured you of the existential anxiety that naturally arises when you realise that you don’t know how to calculate the tie-correction factor for the Kruskal-Wallis test. Right? \n\n### 14.10.3 How to run the Kruskal-Wallis test in R\n\n Despite the horror that we’ve gone through in trying to understand what the Kruskal-Wallis test actually does, it turns out that running the test is pretty painless, since R has a function called `kruskal.test()` . The function is pretty flexible, and allows you to input your data in a few different ways. Most of the time you’ll have data like the `clin.trial` data set, in which you have your outcome variable `mood.gain` , and a grouping variable `drug` . If so, you can call the `kruskal.test()` function by specifying a formula, and a data frame: \n\n```\nkruskal.test(mood.gain ~ drug, data = clin.trial)\n```\n\n```\n## \n## Kruskal-Wallis rank sum test\n## \n## data: mood.gain by drug\n## Kruskal-Wallis chi-squared = 12.076, df = 2, p-value = 0.002386\n```\n\n A second way of using the `kruskal.test()` function, which you probably won’t have much reason to use, is to directly specify the outcome variable and the grouping variable as separate input arguments, `x` and `g` : \n\n```\nkruskal.test(x = clin.trial$mood.gain, g = clin.trial$drug)\n```\n\n```\n## \n## Kruskal-Wallis rank sum test\n## \n## data: clin.trial$mood.gain and clin.trial$drug\n## Kruskal-Wallis chi-squared = 12.076, df = 2, p-value = 0.002386\n```\n\n This isn’t very interesting, since it’s just plain easier to specify a formula. However, sometimes it can be useful to specify `x` as a list. What I mean is this. Suppose you actually had data as three separate variables, `placebo` , `anxifree` and `joyzepam` . If that’s the format that your data are in, then it’s convenient to know that you can bundle all three together as a list: \n\n```\nmood.gain <- list( placebo, joyzepam, anxifree )\nkruskal.test( x = mood.gain )\n```\n\nAnd again, this would give you exactly the same results as the command we tried originally.\n\n## 14.11 On the relationship between ANOVA and the Student \\(t\\) test\n\n There’s one last thing I want to point out before finishing. It’s something that a lot of people find kind of surprising, but it’s worth knowing about: an ANOVA with two groups is identical to the Student \\(t\\)-test. No, really. It’s not just that they are similar, but they are actually equivalent in every meaningful way. I won’t try to prove that this is always true, but I will show you a single concrete demonstration. Suppose that, instead of running an ANOVA on our `mood.gain ~ drug` model, let’s instead do it using `therapy` as the predictor. If we run this ANOVA, here’s what we get: \n\n```\nsummary( aov( mood.gain ~ therapy, data = clin.trial ))\n```\n\n```\n## Df Sum Sq Mean Sq F value Pr(>F)\n## therapy 1 0.467 0.4672 1.708 0.21\n## Residuals 16 4.378 0.2736\n```\n\nOverall, it looks like there’s no significant effect here at all but, as we’ll see in Chapter @ref(anova2 this is actually a misleading answer! In any case, it’s irrelevant to our current goals: our interest here is in the \\(F\\)-statistic, which is \\(F(1,16) = 1.71\\), and the \\(p\\)-value, which is .21. Since we only have two groups, I didn’t actually need to resort to an ANOVA, I could have just decided to run a Student \\(t\\)-test. So let’s see what happens when I do that:\n\n```\nt.test( mood.gain ~ therapy, data = clin.trial, var.equal = TRUE )\n```\n\n```\n## \n## Two Sample t-test\n## \n## data: mood.gain by therapy\n## t = -1.3068, df = 16, p-value = 0.2098\n## alternative hypothesis: true difference in means is not equal to 0\n## 95 percent confidence interval:\n## -0.8449518 0.2005073\n## sample estimates:\n## mean in group no.therapy mean in group CBT \n## 0.7222222 1.0444444\n```\n\nCuriously, the \\(p\\)-values are identical: once again we obtain a value of \\(p = .21\\). But what about the test statistic? Having run a \\(t\\)-test instead of an ANOVA, we get a somewhat different answer, namely \\(t(16) = -1.3068\\). However, there is a fairly straightforward relationship here. If we square the \\(t\\)-statistic\n `1.3068 ^ 2` `## [1] 1.707726` \nwe get the \\(F\\)-statistic from before.\n\n## 14.12 Summary\n\nThere’s a fair bit covered in this chapter, but there’s still a lot missing. Most obviously, I haven’t yet discussed any analog of the paired samples \\(t\\)-test for more than two groups. There is a way of doing this, known as repeated measures ANOVA, which will appear in a later version of this book. I also haven’t discussed how to run an ANOVA when you are interested in more than one grouping variable, but that will be discussed in a lot of detail in Chapter 16. In terms of what we have discussed, the key topics were:\n\n* The basic logic behind how ANOVA works (Section 14.2) and how to run one in R (Section 14.3).\n * How to compute an effect size for an ANOVA (Section 14.4)\n * Post hoc analysis and corrections for multiple testing (Section 14.5).\n * The assumptions made by ANOVA (Section 14.6).\n * How to check the homogeneity of variance assumption (Section 14.7) and what to do if it is violated (Section 14.8).\n * How to check the normality assumption (Section 14.9 and what to do if it is violated (Section 14.10).\n\nAs with all of the chapters in this book, there are quite a few different sources that I’ve relied upon, but the one stand-out text that I’ve been most heavily influenced by is Sahai and Ageel (2000). It’s not a good book for beginners, but it’s an excellent book for more advanced readers who are interested in understanding the mathematics behind ANOVA.\n\nWhen all groups have the same number of observations, the experimental design is said to be “balanced”. Balance isn’t such a big deal for one-way ANOVA, which is the topic of this chapter. It becomes more important when you start doing more complicated ANOVAs.↩\n * \n\nIn a later versions I’m intending to expand on this. But because I’m writing in a rush, and am already over my deadlines, I’ll just briefly note that if you read ahead to Chapter 16 and look at how the “treatment effect” at level \\(k\\) of a factor is defined in terms of the \\(\\alpha_k\\) values (see Section 16.2), it turns out that \\(Q\\) refers to a weighted mean of the squared treatment effects, \\(Q=(\\sum_{k=1}^G N_k\\alpha_k^2)/(G-1)\\).↩\n * \n\nOr, if we want to be sticklers for accuracy, \\(1 + \\frac{2}{df_2 - 2}\\).↩\n * \n\nOr, to be precise, party like “it’s 1899 and we’ve got no friends and nothing better to do with our time than do some calculations that wouldn’t have made any sense in 1899 because ANOVA didn’t exist until about the 1920s”.↩\n * \n\nActually, it also provides a function called\n `anova()` , but that works a bit differently, so let’s just ignore it for now.↩ * \n\nIt’s worth noting that you can get the same result by using the command\n `anova( my.anova )` .↩ * \n\nA potentially important footnote – I wrote the\n `etaSquared()` function for the `lsr` package as a teaching exercise, but like all the other functions in the `lsr` package it hasn’t been exhaustively tested. As of this writing – `lsr` package version 0.5 – there is at least one known bug in the code. In some cases at least, it doesn’t work (and can give very silly answers) when you set the `weights` on the observations to something other than uniform. That doesn’t matter at all for this book, since those kinds of analyses are well beyond the scope, but I haven’t had a chance to revisit the package in a long time; it would probably be wise to be very cautious with the use of this function in any context other than very simple introductory analyses. Thanks to for finding the bug! (Oh, and while I’m here, there’s an interesting blog post by suggesting that eta-squared itself is perhaps not the best measure of effect size in real world data analysis: http://daniellakens.blogspot.com.au/2015/06/why-you-should-use-omega-squared.html↩ * \n\nI should point out that there are other functions in R for running multiple comparisons, and at least one of them works this way: the\n `TukeyHSD()` function takes an `aov` object as its input, and outputs Tukey’s “honestly significant difference” tests. I talk about Tukey’s HSD in Chapter 16.↩ * \n\nIf you do have some theoretical basis for wanting to investigate some comparisons but not others, it’s a different story. In those circumstances you’re not really running “post hoc” analyses at all: you’re making “planned comparisons”. I do talk about this situation later in the book (Section 16.9), but for now I want to keep things simple.↩\n * \n\nIt’s worth noting in passing that not all adjustment methods try to do this. What I’ve described here is an approach for controlling “family wise Type I error rate”. However, there are other post hoc tests seek to control the “false discovery rate”, which is a somewhat different thing.↩\n * \n\nThere’s also a function called\n `p.adjust()` in which you can input a vector of raw \\(p\\)-values, and it will output a vector of adjusted \\(p\\)-values. This can be handy sometimes. I should also note that more advanced users may wish to consider using some of the tools provided by the `multcomp` package.↩ * \n\nNote that neither of these figures has been tidied up at all: if you want to create nicer looking graphs it’s always a good idea to use the tools from Chapter 6 to help you draw cleaner looking images.↩\n * \n\nA technical term.↩\n\nDate: 2015-10-01\nCategories: \nTags: \n\n# Chapter 15 Linear regression\n\nThe goal in this chapter is to introduce linear regression, the standard tool that statisticians rely on when analysing the relationship between interval scale predictors and interval scale outcomes. Stripped to its bare essentials, linear regression models are basically a slightly fancier version of the Pearson correlation (Section 5.7) though as we’ll see, regression models are much more powerful tools.\n\n## 15.1 What is a linear regression model?\n\n Since the basic ideas in regression are closely tied to correlation, we’ll return to the `parenthood.Rdata` file that we were using to illustrate how correlations work. Recall that, in this data set, we were trying to find out why Dan is so very grumpy all the time, and our working hypothesis was that I’m not getting enough sleep. We drew some scatterplots to help us examine the relationship between the amount of sleep I get, and my grumpiness the following day. The actual scatterplot that we draw is the one shown in Figure 15.1, and as we saw previously this corresponds to a correlation of \\(r=-.90\\), but what we find ourselves secretly imagining is something that looks closer to Figure 15.2. That is, we mentally draw a straight line through the middle of the data. In statistics, this line that we’re drawing is called a regression line. Notice that – since we’re not idiots – the regression line goes through the middle of the data. We don’t find ourselves imagining anything like the rather silly plot shown in Figure 15.3. \nThis is not highly surprising: the line that I’ve drawn in Figure 15.3 doesn’t “fit” the data very well, so it doesn’t make a lot of sense to propose it as a way of summarising the data, right? This is a very simple observation to make, but it turns out to be very powerful when we start trying to wrap just a little bit of maths around it. To do so, let’s start with a refresher of some high school maths. The formula for a straight line is usually written like this: \\[ y = mx + c \\] Or, at least, that’s what it was when I went to high school all those years ago. The two variables are \\(x\\) and \\(y\\), and we have two coefficients, \\(m\\) and \\(c\\). The coefficient \\(m\\) represents the slope of the line, and the coefficient \\(c\\) represents the \\(y\\)-intercept of the line. Digging further back into our decaying memories of high school (sorry, for some of us high school was a long time ago), we remember that the intercept is interpreted as “the value of \\(y\\) that you get when \\(x=0\\)”. Similarly, a slope of \\(m\\) means that if you increase the \\(x\\)-value by 1 unit, then the \\(y\\)-value goes up by \\(m\\) units; a negative slope means that the \\(y\\)-value would go down rather than up. Ah yes, it’s all coming back to me now.\n\nNow that we’ve remembered that, it should come as no surprise to discover that we use the exact same formula to describe a regression line. If \\(Y\\) is the outcome variable (the DV) and \\(X\\) is the predictor variable (the IV), then the formula that describes our regression is written like this: \\[ \\hat{Y_i} = b_1 X_i + b_0 \\] Hm. Looks like the same formula, but there’s some extra frilly bits in this version. Let’s make sure we understand them. Firstly, notice that I’ve written \\(X_i\\) and \\(Y_i\\) rather than just plain old \\(X\\) and \\(Y\\). This is because we want to remember that we’re dealing with actual data. In this equation, \\(X_i\\) is the value of predictor variable for the \\(i\\)th observation (i.e., the number of hours of sleep that I got on day \\(i\\) of my little study), and \\(Y_i\\) is the corresponding value of the outcome variable (i.e., my grumpiness on that day). And although I haven’t said so explicitly in the equation, what we’re assuming is that this formula works for all observations in the data set (i.e., for all \\(i\\)). Secondly, notice that I wrote \\(\\hat{Y}_i\\) and not \\(Y_i\\). This is because we want to make the distinction between the actual data \\(Y_i\\), and the estimate \\(\\hat{Y}_i\\) (i.e., the prediction that our regression line is making). Thirdly, I changed the letters used to describe the coefficients from \\(m\\) and \\(c\\) to \\(b_1\\) and \\(b_0\\). That’s just the way that statisticians like to refer to the coefficients in a regression model. I’ve no idea why they chose \\(b\\), but that’s what they did. In any case \\(b_0\\) always refers to the intercept term, and \\(b_1\\) refers to the slope.\n\nExcellent, excellent. Next, I can’t help but notice that – regardless of whether we’re talking about the good regression line or the bad one – the data don’t fall perfectly on the line. Or, to say it another way, the data \\(Y_i\\) are not identical to the predictions of the regression model \\(\\hat{Y_i}\\). Since statisticians love to attach letters, names and numbers to everything, let’s refer to the difference between the model prediction and that actual data point as a residual, and we’ll refer to it as \\(\\epsilon_i\\).215 Written using mathematics, the residuals are defined as: \\[ \\epsilon_i = Y_i - \\hat{Y}_i \\] which in turn means that we can write down the complete linear regression model as: \\[ Y_i = b_1 X_i + b_0 + \\epsilon_i \\]\n\n## 15.2 Estimating a linear regression model\n\nOkay, now let’s redraw our pictures, but this time I’ll add some lines to show the size of the residual for all observations. When the regression line is good, our residuals (the lengths of the solid black lines) all look pretty small, as shown in Figure 15.4, but when the regression line is a bad one, the residuals are a lot larger, as you can see from looking at Figure 15.5. Hm. Maybe what we “want” in a regression model is small residuals. Yes, that does seem to make sense. In fact, I think I’ll go so far as to say that the “best fitting” regression line is the one that has the smallest residuals. Or, better yet, since statisticians seem to like to take squares of everything why not say that …\n The estimated regression coefficients, \\(\\hat{b}_0\\) and \\(\\hat{b}_1\\) are those that minimise the sum of the squared residuals, which we could either write as \\(\\sum_i (Y_i - \\hat{Y}_i)^2\\) or as \\(\\sum_i {\\epsilon_i}^2\\).\n\nYes, yes that sounds even better. And since I’ve indented it like that, it probably means that this is the right answer. And since this is the right answer, it’s probably worth making a note of the fact that our regression coefficients are estimates (we’re trying to guess the parameters that describe a population!), which is why I’ve added the little hats, so that we get \\(\\hat{b}_0\\) and \\(\\hat{b}_1\\) rather than \\(b_0\\) and \\(b_1\\). Finally, I should also note that – since there’s actually more than one way to estimate a regression model – the more technical name for this estimation process is ordinary least squares (OLS) regression.\n At this point, we now have a concrete definition for what counts as our “best” choice of regression coefficients, \\(\\hat{b}_0\\) and \\(\\hat{b}_1\\). The natural question to ask next is, if our optimal regression coefficients are those that minimise the sum squared residuals, how do we find these wonderful numbers? The actual answer to this question is complicated, and it doesn’t help you understand the logic of regression.216 As a result, this time I’m going to let you off the hook. Instead of showing you how to do it the long and tedious way first, and then “revealing” the wonderful shortcut that R provides you with, let’s cut straight to the chase… and use the `lm()` function (short for “linear model”) to do all the heavy lifting. \n\n### 15.2.1 Using the\n\n `lm()` function The `lm()` function is a fairly complicated one: if you type `?lm` , the help files will reveal that there are a lot of arguments that you can specify, and most of them won’t make a lot of sense to you. At this stage however, there’s really only two of them that you care about, and as it turns out you’ve seen them before: \n\n* \n `formula` . A formula that specifies the regression model. For the simple linear regression models that we’ve talked about so far, in which you have a single predictor variable as well as an intercept term, this formula is of the form `outcome ~ predictor` . However, more complicated formulas are allowed, and we’ll discuss them later. * \n `data` . The data frame containing the variables. As we saw with `aov()` in Chapter 14, the output of the `lm()` function is a fairly complicated object, with quite a lot of technical information buried under the hood. Because this technical information is used by other functions, it’s generally a good idea to create a variable that stores the results of your regression. With this in mind, to run my linear regression, the command I want to use is this: \n\n```\nregression.1 <- lm( formula = dan.grump ~ dan.sleep, \n data = parenthood )\n```\n\n Note that I used\n\nas the formula: in the model that I’m trying to estimate, `dan.grump` is the outcome variable, and `dan.sleep` is the predictor variable. It’s always a good idea to remember which one is which! Anyway, what this does is create an “ `lm` object” (i.e., a variable whose class is `\"lm\"` ) called `regression.1` . Let’s have a look at what happens when we `print()` it out: \n\n```\nprint( regression.1 )\n```\n\nThis looks promising. There’s two separate pieces of information here. Firstly, R is politely reminding us what the command was that we used to specify the model in the first place, which can be helpful. More importantly from our perspective, however, is the second part, in which R gives us the intercept \\(\\hat{b}_0 = 125.96\\) and the slope \\(\\hat{b}_1 = -8.94\\). In other words, the best-fitting regression line that I plotted in Figure 15.2 has this formula: \\[ \\hat{Y}_i = -8.94 \\ X_i + 125.96 \\]\n\n### 15.2.2 Interpreting the estimated model\n\nThe most important thing to be able to understand is how to interpret these coefficients. Let’s start with \\(\\hat{b}_1\\), the slope. If we remember the definition of the slope, a regression coefficient of \\(\\hat{b}_1 = -8.94\\) means that if I increase \\(X_i\\) by 1, then I’m decreasing \\(Y_i\\) by 8.94. That is, each additional hour of sleep that I gain will improve my mood, reducing my grumpiness by 8.94 grumpiness points. What about the intercept? Well, since \\(\\hat{b}_0\\) corresponds to “the expected value of \\(Y_i\\) when \\(X_i\\) equals 0”, it’s pretty straightforward. It implies that if I get zero hours of sleep (\\(X_i =0\\)) then my grumpiness will go off the scale, to an insane value of (\\(Y_i = 125.96\\)). Best to be avoided, I think.\n\n## 15.3 Multiple linear regression\n\n The simple linear regression model that we’ve discussed up to this point assumes that there’s a single predictor variable that you’re interested in, in this case `dan.sleep` . In fact, up to this point, every statistical tool that we’ve talked about has assumed that your analysis uses one predictor variable and one outcome variable. However, in many (perhaps most) research projects you actually have multiple predictors that you want to examine. If so, it would be nice to be able to extend the linear regression framework to be able to include multiple predictors. Perhaps some kind of multiple regression model would be in order? Multiple regression is conceptually very simple. All we do is add more terms to our regression equation. Let’s suppose that we’ve got two variables that we’re interested in; perhaps we want to use both `dan.sleep` and `baby.sleep` to predict the `dan.grump` variable. As before, we let \\(Y_i\\) refer to my grumpiness on the \\(i\\)-th day. But now we have two \\(X\\) variables: the first corresponding to the amount of sleep I got and the second corresponding to the amount of sleep my son got. So we’ll let \\(X_{i1}\\) refer to the hours I slept on the \\(i\\)-th day, and \\(X_{i2}\\) refers to the hours that the baby slept on that day. If so, then we can write our regression model like this: \\[\nY_i = b_2 X_{i2} + b_1 X_{i1} + b_0 + \\epsilon_i\n\\] As before, \\(\\epsilon_i\\) is the residual associated with the \\(i\\)-th observation, \\(\\epsilon_i = {Y}_i - \\hat{Y}_i\\). In this model, we now have three coefficients that need to be estimated: \\(b_0\\) is the intercept, \\(b_1\\) is the coefficient associated with my sleep, and \\(b_2\\) is the coefficient associated with my son’s sleep. However, although the number of coefficients that need to be estimated has changed, the basic idea of how the estimation works is unchanged: our estimated coefficients \\(\\hat{b}_0\\), \\(\\hat{b}_1\\) and \\(\\hat{b}_2\\) are those that minimise the sum squared residuals. \n\n### 15.3.1 Doing it in R\n\n Multiple regression in R is no different to simple regression: all we have to do is specify a more complicated `formula` when using the `lm()` function. For example, if we want to use both `dan.sleep` and `baby.sleep` as predictors in our attempt to explain why I’m so grumpy, then the formula we need is this: \n\n Notice that, just like last time, I haven’t explicitly included any reference to the intercept term in this formula; only the two predictor variables and the outcome. By default, the `lm()` function assumes that the model should include an intercept (though you can get rid of it if you want). In any case, I can create a new regression model – which I’ll call `regression.2` – using the following command: \n\n```\nregression.2 <- lm( formula = dan.grump ~ dan.sleep + baby.sleep, \n data = parenthood )\n```\n\n And just like last time, if we `print()` out this regression model we can see what the estimated regression coefficients are: \n\n The coefficient associated with `dan.sleep` is quite large, suggesting that every hour of sleep I lose makes me a lot grumpier. However, the coefficient for `baby.sleep` is very small, suggesting that it doesn’t really matter how much sleep my son gets; not really. What matters as far as my grumpiness goes is how much sleep I get. To get a sense of what this multiple regression model looks like, Figure 15.6 shows a 3D plot that plots all three variables, along with the regression model itself. \n\n### 15.3.2 Formula for the general case\n\nThe equation that I gave above shows you what a multiple regression model looks like when you include two predictors. Not surprisingly, then, if you want more than two predictors all you have to do is add more \\(X\\) terms and more \\(b\\) coefficients. In other words, if you have \\(K\\) predictor variables in the model then the regression equation looks like this: \\[ Y_i = \\left( \\sum_{k=1}^K b_{k} X_{ik} \\right) + b_0 + \\epsilon_i \\]\n\n## 15.4 Quantifying the fit of the regression model\n\n So we now know how to estimate the coefficients of a linear regression model. The problem is, we don’t yet know if this regression model is any good. For example, the `regression.1` model claims that every hour of sleep will improve my mood by quite a lot, but it might just be rubbish. Remember, the regression model only produces a prediction \\(\\hat{Y}_i\\) about what my mood is like: my actual mood is \\(Y_i\\). If these two are very close, then the regression model has done a good job. If they are very different, then it has done a bad job. \n\n### 15.4.1 The \\(R^2\\) value\n\n Once again, let’s wrap a little bit of mathematics around this. Firstly, we’ve got the sum of the squared residuals: \\[ \\mbox{SS}_{res} = \\sum_i (Y_i - \\hat{Y}_i)^2 \\] which we would hope to be pretty small. Specifically, what we’d like is for it to be very small in comparison to the total variability in the outcome variable, \\[ \\mbox{SS}_{tot} = \\sum_i (Y_i - \\bar{Y})^2 \\] While we’re here, let’s calculate these values in R. Firstly, in order to make my R commands look a bit more similar to the mathematical equations, I’ll create variables `X` and `Y` : \n\n```\nX <- parenthood$dan.sleep # the predictor\nY <- parenthood$dan.grump # the outcome\n```\n\n Now that we’ve done this, let’s calculate the \\(\\hat{Y}\\) values and store them in a variable called `Y.pred` . For the simple model that uses only a single predictor, `regression.1` , we would do the following: \n\n```\nY.pred <- -8.94 * X + 125.97\n```\n\nOkay, now that we’ve got a variable which stores the regression model predictions for how grumpy I will be on any given day, let’s calculate our sum of squared residuals. We would do that using the following command:\n\n```\nSS.resid <- sum( (Y - Y.pred)^2 )\nprint( SS.resid )\n```\n\n `## [1] 1838.722` \nWonderful. A big number that doesn’t mean very much. Still, let’s forge boldly onwards anyway, and calculate the total sum of squares as well. That’s also pretty simple:\n\n```\nSS.tot <- sum( (Y - mean(Y))^2 )\nprint( SS.tot )\n```\n\n `## [1] 9998.59` \nHm. Well, it’s a much bigger number than the last one, so this does suggest that our regression model was making good predictions. But it’s not very interpretable.\n\nPerhaps we can fix this. What we’d like to do is to convert these two fairly meaningless numbers into one number. A nice, interpretable number, which for no particular reason we’ll call \\(R^2\\). What we would like is for the value of \\(R^2\\) to be equal to 1 if the regression model makes no errors in predicting the data. In other words, if it turns out that the residual errors are zero – that is, if \\(\\mbox{SS}_{res} = 0\\) – then we expect \\(R^2 = 1\\). Similarly, if the model is completely useless, we would like \\(R^2\\) to be equal to 0. What do I mean by “useless”? Tempting as it is demand that the regression model move out of the house, cut its hair and get a real job, I’m probably going to have to pick a more practical definition: in this case, all I mean is that the residual sum of squares is no smaller than the total sum of squares, \\(\\mbox{SS}_{res} = \\mbox{SS}_{tot}\\). Wait, why don’t we do exactly that? The formula that provides us with out \\(R^2\\) value is pretty simple to write down, \\[ R^2 = 1 - \\frac{\\mbox{SS}_{res}}{\\mbox{SS}_{tot}} \\] and equally simple to calculate in R:\n\n```\nR.squared <- 1 - (SS.resid / SS.tot)\nprint( R.squared )\n```\n\n `## [1] 0.8161018` The \\(R^2\\) value, sometimes called the coefficient of determination217 has a simple interpretation: it is the proportion of the variance in the outcome variable that can be accounted for by the predictor. So in this case, the fact that we have obtained \\(R^2 = .816\\) means that the predictor ( `my.sleep` ) explains 81.6% of the variance in the outcome ( `my.grump` ). Naturally, you don’t actually need to type in all these commands yourself if you want to obtain the \\(R^2\\) value for your regression model. As we’ll see later on in Section 15.5.3, all you need to do is use the `summary()` function. However, let’s put that to one side for the moment. There’s another property of \\(R^2\\) that I want to point out. \n\n### 15.4.2 The relationship between regression and correlation\n\nAt this point we can revisit my earlier claim that regression, in this very simple form that I’ve discussed so far, is basically the same thing as a correlation. Previously, we used the symbol \\(r\\) to denote a Pearson correlation. Might there be some relationship between the value of the correlation coefficient \\(r\\) and the \\(R^2\\) value from linear regression? Of course there is: the squared correlation \\(r^2\\) is identical to the \\(R^2\\) value for a linear regression with only a single predictor. To illustrate this, here’s the squared correlation:\n\n```\nr <- cor(X, Y) # calculate the correlation\nprint( r^2 ) # print the squared correlation\n```\n\n `## [1] 0.8161027` \nYep, same number. In other words, running a Pearson correlation is more or less equivalent to running a linear regression model that uses only one predictor variable.\n\n### 15.4.3 The adjusted \\(R^2\\) value\n\nOne final thing to point out before moving on. It’s quite common for people to report a slightly different measure of model performance, known as “adjusted \\(R^2\\)”. The motivation behind calculating the adjusted \\(R^2\\) value is the observation that adding more predictors into the model will always call the \\(R^2\\) value to increase (or at least not decrease). The adjusted \\(R^2\\) value introduces a slight change to the calculation, as follows. For a regression model with \\(K\\) predictors, fit to a data set containing \\(N\\) observations, the adjusted \\(R^2\\) is: \\[ \\mbox{adj. } R^2 = 1 - \\left(\\frac{\\mbox{SS}_{res}}{\\mbox{SS}_{tot}} \\times \\frac{N-1}{N-K-1} \\right) \\] This adjustment is an attempt to take the degrees of freedom into account. The big advantage of the adjusted \\(R^2\\) value is that when you add more predictors to the model, the adjusted \\(R^2\\) value will only increase if the new variables improve the model performance more than you’d expect by chance. The big disadvantage is that the adjusted \\(R^2\\) value can’t be interpreted in the elegant way that \\(R^2\\) can. \\(R^2\\) has a simple interpretation as the proportion of variance in the outcome variable that is explained by the regression model; to my knowledge, no equivalent interpretation exists for adjusted \\(R^2\\).\n\nAn obvious question then, is whether you should report \\(R^2\\) or adjusted \\(R^2\\). This is probably a matter of personal preference. If you care more about interpretability, then \\(R^2\\) is better. If you care more about correcting for bias, then adjusted \\(R^2\\) is probably better. Speaking just for myself, I prefer \\(R^2\\): my feeling is that it’s more important to be able to interpret your measure of model performance. Besides, as we’ll see in Section 15.5, if you’re worried that the improvement in \\(R^2\\) that you get by adding a predictor is just due to chance and not because it’s a better model, well, we’ve got hypothesis tests for that.\n\n## 15.5 Hypothesis tests for regression models\n\nSo far we’ve talked about what a regression model is, how the coefficients of a regression model are estimated, and how we quantify the performance of the model (the last of these, incidentally, is basically our measure of effect size). The next thing we need to talk about is hypothesis tests. There are two different (but related) kinds of hypothesis tests that we need to talk about: those in which we test whether the regression model as a whole is performing significantly better than a null model; and those in which we test whether a particular regression coefficient is significantly different from zero.\n\nAt this point, you’re probably groaning internally, thinking that I’m going to introduce a whole new collection of tests. You’re probably sick of hypothesis tests by now, and don’t want to learn any new ones. Me too. I’m so sick of hypothesis tests that I’m going to shamelessly reuse the \\(F\\)-test from Chapter 14 and the \\(t\\)-test from Chapter 13. In fact, all I’m going to do in this section is show you how those tests are imported wholesale into the regression framework.\n\n### 15.5.1 Testing the model as a whole\n\nOkay, suppose you’ve estimated your regression model. The first hypothesis test you might want to try is one in which the null hypothesis that there is no relationship between the predictors and the outcome, and the alternative hypothesis is that the data are distributed in exactly the way that the regression model predicts. Formally, our “null model” corresponds to the fairly trivial “regression” model in which we include 0 predictors, and only include the intercept term \\(b_0\\) \\[ H_0: Y_i = b_0 + \\epsilon_i \\] If our regression model has \\(K\\) predictors, the “alternative model” is described using the usual formula for a multiple regression model: \\[ H_1: Y_i = \\left( \\sum_{k=1}^K b_{k} X_{ik} \\right) + b_0 + \\epsilon_i \\]\n\nHow can we test these two hypotheses against each other? The trick is to understand that just like we did with ANOVA, it’s possible to divide up the total variance \\(\\mbox{SS}_{tot}\\) into the sum of the residual variance \\(\\mbox{SS}_{res}\\) and the regression model variance \\(\\mbox{SS}_{mod}\\). I’ll skip over the technicalities, since we covered most of them in the ANOVA chapter, and just note that: \\[ \\mbox{SS}_{mod} = \\mbox{SS}_{tot} - \\mbox{SS}_{res} \\] And, just like we did with the ANOVA, we can convert the sums of squares in to mean squares by dividing by the degrees of freedom. \\[ \\begin{array}{rcl} \\mbox{MS}_{mod} &=& \\displaystyle\\frac{\\mbox{SS}_{mod} }{df_{mod}} \\\\ \\\\ \\mbox{MS}_{res} &=& \\displaystyle\\frac{\\mbox{SS}_{res} }{df_{res} } \\end{array} \\]\n\nSo, how many degrees of freedom do we have? As you might expect, the \\(df\\) associated with the model is closely tied to the number of predictors that we’ve included. In fact, it turns out that \\(df_{mod} = K\\). For the residuals, the total degrees of freedom is \\(df_{res} = N -K - 1\\).\n\nNow that we’ve got our mean square values, you’re probably going to be entirely unsurprised (possibly even bored) to discover that we can calculate an \\(F\\)-statistic like this: \\[ F = \\frac{\\mbox{MS}_{mod}}{\\mbox{MS}_{res}} \\] and the degrees of freedom associated with this are \\(K\\) and \\(N-K-1\\). This \\(F\\) statistic has exactly the same interpretation as the one we introduced in Chapter 14. Large \\(F\\) values indicate that the null hypothesis is performing poorly in comparison to the alternative hypothesis. And since we already did some tedious “do it the long way” calculations back then, I won’t waste your time repeating them. In a moment I’ll show you how to do the test in R the easy way, but first, let’s have a look at the tests for the individual regression coefficients.\n\n### 15.5.2 Tests for individual coefficients\n\n The \\(F\\)-test that we’ve just introduced is useful for checking that the model as a whole is performing better than chance. This is important: if your regression model doesn’t produce a significant result for the \\(F\\)-test then you probably don’t have a very good regression model (or, quite possibly, you don’t have very good data). However, while failing this test is a pretty strong indicator that the model has problems, passing the test (i.e., rejecting the null) doesn’t imply that the model is good! Why is that, you might be wondering? The answer to that can be found by looking at the coefficients for the `regression.2` model: \n\n I can’t help but notice that the estimated regression coefficient for the `baby.sleep` variable is tiny (0.01), relative to the value that we get for `dan.sleep` (-8.95). Given that these two variables are absolutely on the same scale (they’re both measured in “hours slept”), I find this suspicious. In fact, I’m beginning to suspect that it’s really only the amount of sleep that I get that matters in order to predict my grumpiness. \nOnce again, we can reuse a hypothesis test that we discussed earlier, this time the \\(t\\)-test. The test that we’re interested has a null hypothesis that the true regression coefficient is zero (\\(b = 0\\)), which is to be tested against the alternative hypothesis that it isn’t (\\(b \\neq 0\\)). That is: \\[ \\begin{array}{rl} H_0: & b = 0 \\\\ H_1: & b \\neq 0 \\end{array} \\] How can we test this? Well, if the central limit theorem is kind to us, we might be able to guess that the sampling distribution of \\(\\hat{b}\\), the estimated regression coefficient, is a normal distribution with mean centred on \\(b\\). What that would mean is that if the null hypothesis were true, then the sampling distribution of \\(\\hat{b}\\) has mean zero and unknown standard deviation. Assuming that we can come up with a good estimate for the standard error of the regression coefficient, \\(\\mbox{SE}({\\hat{b}})\\), then we’re in luck. That’s exactly the situation for which we introduced the one-sample \\(t\\) way back in Chapter 13. So let’s define a \\(t\\)-statistic like this, \\[ t = \\frac{\\hat{b}}{\\mbox{SE}({\\hat{b})}} \\] I’ll skip over the reasons why, but our degrees of freedom in this case are \\(df = N- K- 1\\). Irritatingly, the estimate of the standard error of the regression coefficient, \\(\\mbox{SE}({\\hat{b}})\\), is not as easy to calculate as the standard error of the mean that we used for the simpler \\(t\\)-tests in Chapter 13. In fact, the formula is somewhat ugly, and not terribly helpful to look at. For our purposes it’s sufficient to point out that the standard error of the estimated regression coefficient depends on both the predictor and outcome variables, and is somewhat sensitive to violations of the homogeneity of variance assumption (discussed shortly).\n\nIn any case, this \\(t\\)-statistic can be interpreted in the same way as the \\(t\\)-statistics that we discussed in Chapter 13. Assuming that you have a two-sided alternative (i.e., you don’t really care if \\(b >0\\) or \\(b < 0\\)), then it’s the extreme values of \\(t\\) (i.e., a lot less than zero or a lot greater than zero) that suggest that you should reject the null hypothesis.\n\n### 15.5.3 Running the hypothesis tests in R\n\n To compute all of the quantities that we have talked about so far, all you need to do is ask for a `summary()` of your regression model. Since I’ve been using `regression.2` as my example, let’s do that: \n\n```\nsummary( regression.2 )\n```\n\n```\n## \n## Call:\n## lm(formula = dan.grump ~ dan.sleep + baby.sleep, data = parenthood)\n## \n## Residuals:\n## Min 1Q Median 3Q Max \n## -11.0345 -2.2198 -0.4016 2.6775 11.7496 \n## \n## Coefficients:\n## Estimate Std. Error t value Pr(>|t|) \n## (Intercept) 125.96557 3.04095 41.423 <2e-16 ***\n## dan.sleep -8.95025 0.55346 -16.172 <2e-16 ***\n## baby.sleep 0.01052 0.27106 0.039 0.969 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n## \n## Residual standard error: 4.354 on 97 degrees of freedom\n## Multiple R-squared: 0.8161, Adjusted R-squared: 0.8123 \n## F-statistic: 215.2 on 2 and 97 DF, p-value: < 2.2e-16\n```\n\nThe output that this command produces is pretty dense, but we’ve already discussed everything of interest in it, so what I’ll do is go through it line by line. The first line reminds us of what the actual regression model is:\n\n```\nCall:\nlm(formula = dan.grump ~ dan.sleep + baby.sleep, data = parenthood)\n```\n\n You can see why this is handy, since it was a little while back when we actually created the `regression.2` model, and so it’s nice to be reminded of what it was we were doing. The next part provides a quick summary of the residuals (i.e., the \\(\\epsilon_i\\) values), \n\n```\nResiduals:\n Min 1Q Median 3Q Max \n-11.0345 -2.2198 -0.4016 2.6775 11.7496\n```\n\nwhich can be convenient as a quick and dirty check that the model is okay. Remember, we did assume that these residuals were normally distributed, with mean 0. In particular it’s worth quickly checking to see if the median is close to zero, and to see if the first quartile is about the same size as the third quartile. If they look badly off, there’s a good chance that the assumptions of regression are violated. These ones look pretty nice to me, so let’s move on to the interesting stuff. The next part of the R output looks at the coefficients of the regression model:\n\n```\nCoefficients:\n Estimate Std. Error t value Pr(>|t|) \n(Intercept) 125.96557 3.04095 41.423 <2e-16 ***\ndan.sleep -8.95025 0.55346 -16.172 <2e-16 ***\nbaby.sleep 0.01052 0.27106 0.039 0.969 \n---\nSignif. codes: 0 �***� 0.001 �**� 0.01 �*� 0.05 �.� 0.1 � � 1\n```\n\n Each row in this table refers to one of the coefficients in the regression model. The first row is the intercept term, and the later ones look at each of the predictors. The columns give you all of the relevant information. The first column is the actual estimate of \\(b\\) (e.g., 125.96 for the intercept, and -8.9 for the `dan.sleep` predictor). The second column is the standard error estimate \\(\\hat\\sigma_b\\). The third column gives you the \\(t\\)-statistic, and it’s worth noticing that in this table \\(t= \\hat{b}/\\mbox{SE}({\\hat{b}})\\) every time. Finally, the fourth column gives you the actual \\(p\\) value for each of these tests.218 The only thing that the table itself doesn’t list is the degrees of freedom used in the \\(t\\)-test, which is always \\(N-K-1\\) and is listed immediately below, in this line: \n\n```\nResidual standard error: 4.354 on 97 degrees of freedom\n```\n\nThe value of \\(df = 97\\) is equal to \\(N-K-1\\), so that’s what we use for our \\(t\\)-tests. In the final part of the output we have the \\(F\\)-test and the \\(R^2\\) values which assess the performance of the model as a whole\n\n```\nResidual standard error: 4.354 on 97 degrees of freedom\nMultiple R-squared: 0.8161, Adjusted R-squared: 0.8123 \nF-statistic: 215.2 on 2 and 97 DF, p-value: < 2.2e-16\n```\n\n So in this case, the model performs significantly better than you’d expect by chance (\\(F(2,97) = 215.2\\), \\(p<.001\\)), which isn’t all that surprising: the \\(R^2 = .812\\) value indicate that the regression model accounts for 81.2% of the variability in the outcome measure. However, when we look back up at the \\(t\\)-tests for each of the individual coefficients, we have pretty strong evidence that the `baby.sleep` variable has no significant effect; all the work is being done by the `dan.sleep` variable. Taken together, these results suggest that `regression.2` is actually the wrong model for the data: you’d probably be better off dropping the `baby.sleep` predictor entirely. In other words, the `regression.1` model that we started with is the better model. \n\n## 15.6 Testing the significance of a correlation\n\n### 15.6.1 Hypothesis tests for a single correlation\n\n I don’t want to spend too much time on this, but it’s worth very briefly returning to the point I made earlier, that Pearson correlations are basically the same thing as linear regressions with only a single predictor added to the model. What this means is that the hypothesis tests that I just described in a regression context can also be applied to correlation coefficients. To see this, let’s take a `summary()` of the `regression.1` model: \n\n```\nsummary( regression.1 )\n```\n\n```\n## \n## Call:\n## lm(formula = dan.grump ~ dan.sleep, data = parenthood)\n## \n## Residuals:\n## Min 1Q Median 3Q Max \n## -11.025 -2.213 -0.399 2.681 11.750 \n## \n## Coefficients:\n## Estimate Std. Error t value Pr(>|t|) \n## (Intercept) 125.9563 3.0161 41.76 <2e-16 ***\n## dan.sleep -8.9368 0.4285 -20.85 <2e-16 ***\n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n## \n## Residual standard error: 4.332 on 98 degrees of freedom\n## Multiple R-squared: 0.8161, Adjusted R-squared: 0.8142 \n## F-statistic: 434.9 on 1 and 98 DF, p-value: < 2.2e-16\n```\n\n The important thing to note here is the \\(t\\) test associated with the predictor, in which we get a result of \\(t(98) = -20.85\\), \\(p<.001\\). Now let’s compare this to the output of a different function, which goes by the name of `cor.test()` . As you might expect, this function runs a hypothesis test to see if the observed correlation between two variables is significantly different from 0. Let’s have a look: \n\n```\n## \n## Pearson's product-moment correlation\n## \n## data: parenthood$dan.sleep and parenthood$dan.grump\n## t = -20.854, df = 98, p-value < 2.2e-16\n## alternative hypothesis: true correlation is not equal to 0\n## 95 percent confidence interval:\n## -0.9340614 -0.8594714\n## sample estimates:\n## cor \n## -0.903384\n```\n\nAgain, the key thing to note is the line that reports the hypothesis test itself, which seems to be saying that \\(t(98) = -20.85\\), \\(p<.001\\). Hm. Looks like it’s exactly the same test, doesn’t it? And that’s exactly what it is. The test for the significance of a correlation is identical to the \\(t\\) test that we run on a coefficient in a regression model.\n\n### 15.6.2 Hypothesis tests for all pairwise correlations\n\n Okay, one more digression before I return to regression properly. In the previous section I talked about the `cor.test()` function, which lets you run a hypothesis test on a single correlation. The `cor.test()` function is (obviously) an extension of the `cor()` function, which we talked about in Section 5.7. However, the `cor()` function isn’t restricted to computing a single correlation: you can use it to compute all pairwise correlations among the variables in your data set. This leads people to the natural question: can the `cor.test()` function do the same thing? Can we use `cor.test()` to run hypothesis tests for all possible parwise correlations among the variables in a data frame? The answer is no, and there’s a very good reason for this. Testing a single correlation is fine: if you’ve got some reason to be asking “is A related to B?”, then you should absolutely run a test to see if there’s a significant correlation. But if you’ve got variables A, B, C, D and E and you’re thinking about testing the correlations among all possible pairs of these, a statistician would want to ask: what’s your hypothesis? If you’re in the position of wanting to test all possible pairs of variables, then you’re pretty clearly on a fishing expedition, hunting around in search of significant effects when you don’t actually have a clear research hypothesis in mind. This is dangerous, and the authors of `cor.test()` obviously felt that they didn’t want to support that kind of behaviour. On the other hand… a somewhat less hardline view might be to argue we’ve encountered this situation before, back in Section 14.5 when we talked about post hoc tests in ANOVA. When running post hoc tests, we didn’t have any specific comparisons in mind, so what we did was apply a correction (e.g., Bonferroni, Holm, etc) in order to avoid the possibility of an inflated Type I error rate. From this perspective, it’s okay to run hypothesis tests on all your pairwise correlations, but you must treat them as post hoc analyses, and if so you need to apply a correction for multiple comparisons. That’s what the `correlate()` function in the `lsr` package does. When we use the `correlate()` function in Section 5.7 all it did was print out the correlation matrix. But you can get it to output the results of all the pairwise tests as well by specifying `test=TRUE` . Here’s what happens with the `parenthood` data: \n\n```\nlibrary(lsr)\ncorrelate(parenthood, test=TRUE)\n```\n\n```\n## \n## CORRELATIONS\n## ============\n## - correlation type: pearson \n## - correlations shown only when both variables are numeric\n## \n## dan.sleep baby.sleep dan.grump day \n## dan.sleep . 0.628*** -0.903*** -0.098 \n## baby.sleep 0.628*** . -0.566*** -0.010 \n## dan.grump -0.903*** -0.566*** . 0.076 \n## day -0.098 -0.010 0.076 . \n## \n## ---\n## Signif. codes: . = p < .1, * = p<.05, ** = p<.01, *** = p<.001\n## \n## \n## p-VALUES\n## ========\n## - total number of tests run: 6 \n## - correction for multiple testing: holm \n## \n## dan.sleep baby.sleep dan.grump day\n## dan.sleep . 0.000 0.000 0.990\n## baby.sleep 0.000 . 0.000 0.990\n## dan.grump 0.000 0.000 . 0.990\n## day 0.990 0.990 0.990 .\n## \n## \n## SAMPLE SIZES\n## ============\n## \n## dan.sleep baby.sleep dan.grump day\n## dan.sleep 100 100 100 100\n## baby.sleep 100 100 100 100\n## dan.grump 100 100 100 100\n## day 100 100 100 100\n```\n\nThe output here contains three matrices. First it prints out the correlation matrix. Second it prints out a matrix of \\(p\\)-values, using the Holm method219 to correct for multiple comparisons. Finally, it prints out a matrix indicating the sample size (number of pairwise complete cases) that contributed to each correlation.\n So there you have it. If you really desperately want to do pairwise hypothesis tests on your correlations, the `correlate()` function will let you do it. But please, please be careful. I can’t count the number of times I’ve had a student panicking in my office because they’ve run these pairwise correlation tests, and they get one or two significant results that don’t make any sense. For some reason, the moment people see those little significance stars appear, they feel compelled to throw away all common sense and assume that the results must correspond to something real that requires an explanation. In most such cases, my experience has been that the right answer is “it’s a Type I error”. \n\n## 15.7 Regarding regression coefficients\n\nBefore moving on to discuss the assumptions underlying linear regression and what you can do to check if they’re being met, there’s two more topics I want to briefly discuss, both of which relate to the regression coefficients. The first thing to talk about is calculating confidence intervals for the coefficients; after that, I’ll discuss the somewhat murky question of how to determine which of predictor is most important.\n\n### 15.7.1 Confidence intervals for the coefficients\n\n Like any population parameter, the regression coefficients \\(b\\) cannot be estimated with complete precision from a sample of data; that’s part of why we need hypothesis tests. Given this, it’s quite useful to be able to report confidence intervals that capture our uncertainty about the true value of \\(b\\). This is especially useful when the research question focuses heavily on an attempt to find out how strongly variable \\(X\\) is related to variable \\(Y\\), since in those situations the interest is primarily in the regression weight \\(b\\). Fortunately, confidence intervals for the regression weights can be constructed in the usual fashion, \\[ \\mbox{CI}(b) = \\hat{b} \\pm \\left( t_{crit} \\times \\mbox{SE}({\\hat{b})} \\right) \\] where \\(\\mbox{SE}({\\hat{b}})\\) is the standard error of the regression coefficient, and \\(t_{crit}\\) is the relevant critical value of the appropriate \\(t\\) distribution. For instance, if it’s a 95% confidence interval that we want, then the critical value is the 97.5th quantile of a \\(t\\) distribution with \\(N-K-1\\) degrees of freedom. In other words, this is basically the same approach to calculating confidence intervals that we’ve used throughout. To do this in R we can use the `confint()` function. There arguments to this function are \n\n* \n `object` . The regression model ( `lm` object) for which confidence intervals are required. * \n `parm` . A vector indicating which coefficients we should calculate intervals for. This can be either a vector of numbers or (more usefully) a character vector containing variable names. By default, all coefficients are included, so usually you don’t bother specifying this argument. * \n `level` . A number indicating the confidence level that should be used. As is usually the case, the default value is 0.95, so you wouldn’t usually need to specify this argument. So, suppose I want 99% confidence intervals for the coefficients in the `regression.2` model. I could do this using the following command: \n\n```\nconfint( object = regression.2,\n level = .99)\n```\n\n```\n## 0.5 % 99.5 %\n## (Intercept) 117.9755724 133.9555593\n## dan.sleep -10.4044419 -7.4960575\n## baby.sleep -0.7016868 0.7227357\n```\n\nSimple enough.\n\n### 15.7.2 Calculating standardised regression coefficients\n\nOne more thing that you might want to do is to calculate “standardised” regression coefficients, often denoted \\(\\beta\\). The rationale behind standardised coefficients goes like this. In a lot of situations, your variables are on fundamentally different scales. Suppose, for example, my regression model aims to predict people’s IQ scores, using their educational attainment (number of years of education) and their income as predictors. Obviously, educational attainment and income are not on the same scales: the number of years of schooling can only vary by 10s of years, whereas income would vary by 10,000s of dollars (or more). The units of measurement have a big influence on the regression coefficients: the \\(b\\) coefficients only make sense when interpreted in light of the units, both of the predictor variables and the outcome variable. This makes it very difficult to compare the coefficients of different predictors. Yet there are situations where you really do want to make comparisons between different coefficients. Specifically, you might want some kind of standard measure of which predictors have the strongest relationship to the outcome. This is what standardised coefficients aim to do.\n\nThe basic idea is quite simple: the standardised coefficients are the coefficients that you would have obtained if you’d converted all the variables to \\(z\\)-scores before running the regression.220 The idea here is that, by converting all the predictors to \\(z\\)-scores, they all go into the regression on the same scale, thereby removing the problem of having variables on different scales. Regardless of what the original variables were, a \\(\\beta\\) value of 1 means that an increase in the predictor of 1 standard deviation will produce a corresponding 1 standard deviation increase in the outcome variable. Therefore, if variable A has a larger absolute value of \\(\\beta\\) than variable B, it is deemed to have a stronger relationship with the outcome. Or at least that’s the idea: it’s worth being a little cautious here, since this does rely very heavily on the assumption that “a 1 standard deviation change” is fundamentally the same kind of thing for all variables. It’s not always obvious that this is true.\n Leaving aside the interpretation issues, let’s look at how it’s calculated. What you could do is standardise all the variables yourself and then run a regression, but there’s a much simpler way to do it. As it turns out, the \\(\\beta\\) coefficient for a predictor \\(X\\) and outcome \\(Y\\) has a very simple formula, namely \\[ \\beta_X = b_X \\times \\frac{\\sigma_X}{\\sigma_Y} \\] where \\(\\sigma_X\\) is the standard deviation of the predictor, and \\(\\sigma_Y\\) is the standard deviation of the outcome variable \\(Y\\). This makes matters a lot simpler. To make things even simpler, the `lsr` package includes a function `standardCoefs()` that computes the \\(\\beta\\) coefficients. \n\n```\nstandardCoefs( regression.2 )\n```\n\n```\n## b beta\n## dan.sleep -8.95024973 -0.90474809\n## baby.sleep 0.01052447 0.00217223\n```\n\n This clearly shows that the `dan.sleep` variable has a much stronger effect than the `baby.sleep` variable. However, this is a perfect example of a situation where it would probably make sense to use the original coefficients \\(b\\) rather than the standardised coefficients \\(\\beta\\). After all, my sleep and the baby’s sleep are already on the same scale: number of hours slept. Why complicate matters by converting these to \\(z\\)-scores? \n\n## 15.8 Assumptions of regression\n\nThe linear regression model that I’ve been discussing relies on several assumptions. In Section 15.9 we’ll talk a lot more about how to check that these assumptions are being met, but first, let’s have a look at each of them.\n\n* Normality. Like half the models in statistics, standard linear regression relies on an assumption of normality. Specifically, it assumes that the residuals are normally distributed. It’s actually okay if the predictors \\(X\\) and the outcome \\(Y\\) are non-normal, so long as the residuals \\(\\epsilon\\) are normal. See Section 15.9.3.\n * Linearity. A pretty fundamental assumption of the linear regression model is that relationship between \\(X\\) and \\(Y\\) actually be linear! Regardless of whether it’s a simple regression or a multiple regression, we assume that the relatiships involved are linear. See Section 15.9.4.\n * Homogeneity of variance. Strictly speaking, the regression model assumes that each residual \\(\\epsilon_i\\) is generated from a normal distribution with mean 0, and (more importantly for the current purposes) with a standard deviation \\(\\sigma\\) that is the same for every single residual. In practice, it’s impossible to test the assumption that every residual is identically distributed. Instead, what we care about is that the standard deviation of the residual is the same for all values of \\(\\hat{Y}\\), and (if we’re being especially paranoid) all values of every predictor \\(X\\) in the model. See Section 15.9.5.\n * Uncorrelated predictors. The idea here is that, is a multiple regression model, you don’t want your predictors to be too strongly correlated with each other. This isn’t “technically” an assumption of the regression model, but in practice it’s required. Predictors that are too strongly correlated with each other (referred to as “collinearity”) can cause problems when evaluating the model. See Section 15.9.6\n * Residuals are independent of each other. This is really just a “catch all” assumption, to the effect that “there’s nothing else funny going on in the residuals”. If there is something weird (e.g., the residuals all depend heavily on some other unmeasured variable) going on, it might screw things up.\n * No “bad” outliers. Again, not actually a technical assumption of the model (or rather, it’s sort of implied by all the others), but there is an implicit assumption that your regression model isn’t being too strongly influenced by one or two anomalous data points; since this raises questions about the adequacy of the model, and the trustworthiness of the data in some cases. See Section 15.9.2.\n\n## 15.9 Model checking\n\nThe main focus of this section is regression diagnostics, a term that refers to the art of checking that the assumptions of your regression model have been met, figuring out how to fix the model if the assumptions are violated, and generally to check that nothing “funny” is going on. I refer to this as the “art” of model checking with good reason: it’s not easy, and while there are a lot of fairly standardised tools that you can use to diagnose and maybe even cure the problems that ail your model (if there are any, that is!), you really do need to exercise a certain amount of judgment when doing this. It’s easy to get lost in all the details of checking this thing or that thing, and it’s quite exhausting to try to remember what all the different things are. This has the very nasty side effect that a lot of people get frustrated when trying to learn all the tools, so instead they decide not to do any model checking. This is a bit of a worry!\n In this section, I describe several different things you can do to check that your regression model is doing what it’s supposed to. It doesn’t cover the full space of things you could do, but it’s still much more detailed than what I see a lot of people doing in practice; and I don’t usually cover all of this in my intro stats class myself. However, I do think it’s important that you get a sense of what tools are at your disposal, so I’ll try to introduce a bunch of them here. Finally, I should note that this section draws quite heavily from the Fox and Weisberg (2011) text, the book associated with the `car` package. The `car` package is notable for providing some excellent tools for regression diagnostics, and the book itself talks about them in an admirably clear fashion. I don’t want to sound too gushy about it, but I do think that Fox and Weisberg (2011) is well worth reading. \n\n### 15.9.1 Three kinds of residuals\n\nThe majority of regression diagnostics revolve around looking at the residuals, and by now you’ve probably formed a sufficiently pessimistic theory of statistics to be able to guess that – precisely because of the fact that we care a lot about the residuals – there are several different kinds of residual that we might consider. In particular, the following three kinds of residual are referred to in this section: “ordinary residuals”, “standardised residuals”, and “Studentised residuals”. There is a fourth kind that you’ll see referred to in some of the Figures, and that’s the “Pearson residual”: however, for the models that we’re talking about in this chapter, the Pearson residual is identical to the ordinary residual.\n\nThe first and simplest kind of residuals that we care about are ordinary residuals. These are the actual, raw residuals that I’ve been talking about throughout this chapter. The ordinary residual is just the difference between the fitted value \\(\\hat{Y}_i\\) and the observed value \\(Y_i\\). I’ve been using the notation \\(\\epsilon_i\\) to refer to the \\(i\\)-th ordinary residual, and by gum I’m going to stick to it. With this in mind, we have the very simple equation \\[ \\epsilon_i = Y_i - \\hat{Y}_i \\] This is of course what we saw earlier, and unless I specifically refer to some other kind of residual, this is the one I’m talking about. So there’s nothing new here: I just wanted to repeat myself. In any case, you can get R to output a vector of ordinary residuals, you can use a command like this:\n\n```\nresiduals( object = regression.2 )\n```\n\n```\n## 1 2 3 4 5 6 \n## -2.1403095 4.7081942 1.9553640 -2.0602806 0.7194888 -0.4066133 \n## 7 8 9 10 11 12 \n## 0.2269987 -1.7003077 0.2025039 3.8524589 3.9986291 -4.9120150 \n## 13 14 15 16 17 18 \n## 1.2060134 0.4946578 -2.6579276 -0.3966805 3.3538613 1.7261225 \n## 19 20 21 22 23 24 \n## -0.4922551 -5.6405941 -0.4660764 2.7238389 9.3653697 0.2841513 \n## 25 26 27 28 29 30 \n## -0.5037668 -1.4941146 8.1328623 1.9787316 -1.5126726 3.5171148 \n## 31 32 33 34 35 36 \n## -8.9256951 -2.8282946 6.1030349 -7.5460717 4.5572128 -10.6510836 \n## 37 38 39 40 41 42 \n## -5.6931846 6.3096506 -2.1082466 -0.5044253 0.1875576 4.8094841 \n## 43 44 45 46 47 48 \n## -5.4135163 -6.2292842 -4.5725232 -5.3354601 3.9950111 2.1718745 \n## 49 50 51 52 53 54 \n## -3.4766440 0.4834367 6.2839790 2.0109396 -1.5846631 -2.2166613 \n## 55 56 57 58 59 60 \n## 2.2033140 1.9328736 -1.8301204 -1.5401430 2.5298509 -3.3705782 \n## 61 62 63 64 65 66 \n## -2.9380806 0.6590736 -0.5917559 -8.6131971 5.9781035 5.9332979 \n## 67 68 69 70 71 72 \n## -1.2341956 3.0047669 -1.0802468 6.5174672 -3.0155469 2.1176720 \n## 73 74 75 76 77 78 \n## 0.6058757 -2.7237421 -2.2291472 -1.4053822 4.7461491 11.7495569 \n## 79 80 81 82 83 84 \n## 4.7634141 2.6620908 -11.0345292 -0.7588667 1.4558227 -0.4745727 \n## 85 86 87 88 89 90 \n## 8.9091201 -1.1409777 0.7555223 -0.4107130 0.8797237 -1.4095586 \n## 91 92 93 94 95 96 \n## 3.1571385 -3.4205757 -5.7228699 -2.2033958 -3.8647891 0.4982711 \n## 97 98 99 100 \n## -5.5249495 4.1134221 -8.2038533 5.6800859\n```\n\nOne drawback to using ordinary residuals is that they’re always on a different scale, depending on what the outcome variable is and how good the regression model is. That is, Unless you’ve decided to run a regression model without an intercept term, the ordinary residuals will have mean 0; but the variance is different for every regression. In a lot of contexts, especially where you’re only interested in the pattern of the residuals and not their actual values, it’s convenient to estimate the standardised residuals, which are normalised in such a way as to have standard deviation 1. The way we calculate these is to divide the ordinary residual by an estimate of the (population) standard deviation of these residuals. For technical reasons, mumble mumble, the formula for this is: \\[ \\epsilon_{i}^\\prime = \\frac{\\epsilon_i}{\\hat{\\sigma} \\sqrt{1-h_i}} \\] where \\(\\hat\\sigma\\) in this context is the estimated population standard deviation of the ordinary residuals, and \\(h_i\\) is the “hat value” of the \\(i\\)th observation. I haven’t explained hat values to you yet (but have no fear,221 it’s coming shortly), so this won’t make a lot of sense. For now, it’s enough to interpret the standardised residuals as if we’d converted the ordinary residuals to \\(z\\)-scores. In fact, that is more or less the truth, it’s just that we’re being a bit fancier. To get the standardised residuals, the command you want is this:\n\n```\nrstandard( model = regression.2 )\n```\n\n```\n## 1 2 3 4 5 6 \n## -0.49675845 1.10430571 0.46361264 -0.47725357 0.16756281 -0.09488969 \n## 7 8 9 10 11 12 \n## 0.05286626 -0.39260381 0.04739691 0.89033990 0.95851248 -1.13898701 \n## 13 14 15 16 17 18 \n## 0.28047841 0.11519184 -0.61657092 -0.09191865 0.77692937 0.40403495 \n## 19 20 21 22 23 24 \n## -0.11552373 -1.31540412 -0.10819238 0.62951824 2.17129803 0.06586227 \n## 25 26 27 28 29 30 \n## -0.11980449 -0.34704024 1.91121833 0.45686516 -0.34986350 0.81233165 \n## 31 32 33 34 35 36 \n## -2.08659993 -0.66317843 1.42930082 -1.77763064 1.07452436 -2.47385780 \n## 37 38 39 40 41 42 \n## -1.32715114 1.49419658 -0.49115639 -0.11674947 0.04401233 1.11881912 \n## 43 44 45 46 47 48 \n## -1.27081641 -1.46422595 -1.06943700 -1.24659673 0.94152881 0.51069809 \n## 49 50 51 52 53 54 \n## -0.81373349 0.11412178 1.47938594 0.46437962 -0.37157009 -0.51609949 \n## 55 56 57 58 59 60 \n## 0.51800753 0.44813204 -0.42662358 -0.35575611 0.58403297 -0.78022677 \n## 61 62 63 64 65 66 \n## -0.67833325 0.15484699 -0.13760574 -2.05662232 1.40238029 1.37505125 \n## 67 68 69 70 71 72 \n## -0.28964989 0.69497632 -0.24945316 1.50709623 -0.69864682 0.49071427 \n## 73 74 75 76 77 78 \n## 0.14267297 -0.63246560 -0.51972828 -0.32509811 1.10842574 2.72171671 \n## 79 80 81 82 83 84 \n## 1.09975101 0.62057080 -2.55172097 -0.17584803 0.34340064 -0.11158952 \n## 85 86 87 88 89 90 \n## 2.10863391 -0.26386516 0.17624445 -0.09504416 0.20450884 -0.32730740 \n## 91 92 93 94 95 96 \n## 0.73475640 -0.79400855 -1.32768248 -0.51940736 -0.91512580 0.11661226 \n## 97 98 99 100 \n## -1.28069115 0.96332849 -1.90290258 1.31368144\n```\n\nNote that this function uses a different name for the input argument, but it’s still just a linear regression object that the function wants to take as its input here.\n\nThe third kind of residuals are Studentised residuals (also called “jackknifed residuals”) and they’re even fancier than standardised residuals. Again, the idea is to take the ordinary residual and divide it by some quantity in order to estimate some standardised notion of the residual, but the formula for doing the calculations this time is subtly different: \\[ \\epsilon_{i}^* = \\frac{\\epsilon_i}{\\hat{\\sigma}_{(-i)} \\sqrt{1-h_i}} \\] Notice that our estimate of the standard deviation here is written \\(\\hat{\\sigma}_{(-i)}\\). What this corresponds to is the estimate of the residual standard deviation that you would have obtained, if you just deleted the \\(i\\)th observation from the data set. This sounds like the sort of thing that would be a nightmare to calculate, since it seems to be saying that you have to run \\(N\\) new regression models (even a modern computer might grumble a bit at that, especially if you’ve got a large data set). Fortunately, some terribly clever person has shown that this standard deviation estimate is actually given by the following equation: \\[ \\hat\\sigma_{(-i)} = \\hat{\\sigma} \\ \\sqrt{\\frac{N-K-1 - {\\epsilon_{i}^\\prime}^2}{N-K-2}} \\] Isn’t that a pip? Anyway, the command that you would use if you wanted to pull out the Studentised residuals for our regression model is\n\n```\nrstudent( model = regression.2 )\n```\n\n```\n## 1 2 3 4 5 6 \n## -0.49482102 1.10557030 0.46172854 -0.47534555 0.16672097 -0.09440368 \n## 7 8 9 10 11 12 \n## 0.05259381 -0.39088553 0.04715251 0.88938019 0.95810710 -1.14075472 \n## 13 14 15 16 17 18 \n## 0.27914212 0.11460437 -0.61459001 -0.09144760 0.77533036 0.40228555 \n## 19 20 21 22 23 24 \n## -0.11493461 -1.32043609 -0.10763974 0.62754813 2.21456485 0.06552336 \n## 25 26 27 28 29 30 \n## -0.11919416 -0.34546127 1.93818473 0.45499388 -0.34827522 0.81089646 \n## 31 32 33 34 35 36 \n## -2.12403286 -0.66125192 1.43712830 -1.79797263 1.07539064 -2.54258876 \n## 37 38 39 40 41 42 \n## -1.33244515 1.50388257 -0.48922682 -0.11615428 0.04378531 1.12028904 \n## 43 44 45 46 47 48 \n## -1.27490649 -1.47302872 -1.07023828 -1.25020935 0.94097261 0.50874322 \n## 49 50 51 52 53 54 \n## -0.81230544 0.11353962 1.48863006 0.46249410 -0.36991317 -0.51413868 \n## 55 56 57 58 59 60 \n## 0.51604474 0.44627831 -0.42481754 -0.35414868 0.58203894 -0.77864171 \n## 61 62 63 64 65 66 \n## -0.67643392 0.15406579 -0.13690795 -2.09211556 1.40949469 1.38147541 \n## 67 68 69 70 71 72 \n## -0.28827768 0.69311245 -0.24824363 1.51717578 -0.69679156 0.48878534 \n## 73 74 75 76 77 78 \n## 0.14195054 -0.63049841 -0.51776374 -0.32359434 1.10974786 2.81736616 \n## 79 80 81 82 83 84 \n## 1.10095270 0.61859288 -2.62827967 -0.17496714 0.34183379 -0.11101996 \n## 85 86 87 88 89 90 \n## 2.14753375 -0.26259576 0.17536170 -0.09455738 0.20349582 -0.32579584 \n## 91 92 93 94 95 96 \n## 0.73300184 -0.79248469 -1.33298848 -0.51744314 -0.91435205 0.11601774 \n## 97 98 99 100 \n## -1.28498273 0.96296745 -1.92942389 1.31867548\n```\n\n Before moving on, I should point out that you don’t often need to manually extract these residuals yourself, even though they are at the heart of almost all regression diagnostics. That is, the `residuals()` , `rstandard()` and `rstudent()` functions are all useful to know about, but most of the time the various functions that run the diagnostics will take care of these calculations for you. Even so, it’s always nice to know how to actually get hold of these things yourself in case you ever need to do something non-standard. \n\n### 15.9.2 Three kinds of anomalous data\n\n One danger that you can run into with linear regression models is that your analysis might be disproportionately sensitive to a smallish number of “unusual” or “anomalous” observations. I discussed this idea previously in Section 6.5.2 in the context of discussing the outliers that get automatically identified by the `boxplot()` function, but this time we need to be much more precise. In the context of linear regression, there are three conceptually distinct ways in which an observation might be called “anomalous”. All three are interesting, but they have rather different implications for your analysis. \nThe second way in which an observation can be unusual is if it has high leverage: this happens when the observation is very different from all the other observations. This doesn’t necessarily have to correspond to a large residual: if the observation happens to be unusual on all variables in precisely the same way, it can actually lie very close to the regression line. An example of this is shown in Figure 15.8. The leverage of an observation is operationalised in terms of its hat value, usually written \\(h_i\\). The formula for the hat value is rather complicated222 but its interpretation is not: \\(h_i\\) is a measure of the extent to which the \\(i\\)-th observation is “in control” of where the regression line ends up going. You can extract the hat values using the following command:\n\n```\n## 1 2 3 4 5 6 \n## 0.02067452 0.04105320 0.06155445 0.01685226 0.02734865 0.03129943 \n## 7 8 9 10 11 12 \n## 0.02735579 0.01051224 0.03698976 0.01229155 0.08189763 0.01882551 \n## 13 14 15 16 17 18 \n## 0.02462902 0.02718388 0.01964210 0.01748592 0.01691392 0.03712530 \n## 19 20 21 22 23 24 \n## 0.04213891 0.02994643 0.02099435 0.01233280 0.01853370 0.01804801 \n## 25 26 27 28 29 30 \n## 0.06722392 0.02214927 0.04472007 0.01039447 0.01381812 0.01105817 \n## 31 32 33 34 35 36 \n## 0.03468260 0.04048248 0.03814670 0.04934440 0.05107803 0.02208177 \n## 37 38 39 40 41 42 \n## 0.02919013 0.05928178 0.02799695 0.01519967 0.04195751 0.02514137 \n## 43 44 45 46 47 48 \n## 0.04267879 0.04517340 0.03558080 0.03360160 0.05019778 0.04587468 \n## 49 50 51 52 53 54 \n## 0.03701290 0.05331282 0.04814477 0.01072699 0.04047386 0.02681315 \n## 55 56 57 58 59 60 \n## 0.04556787 0.01856997 0.02919045 0.01126069 0.01012683 0.01546412 \n## 61 62 63 64 65 66 \n## 0.01029534 0.04428870 0.02438944 0.07469673 0.04135090 0.01775697 \n## 67 68 69 70 71 72 \n## 0.04217616 0.01384321 0.01069005 0.01340216 0.01716361 0.01751844 \n## 73 74 75 76 77 78 \n## 0.04863314 0.02158623 0.02951418 0.01411915 0.03276064 0.01684599 \n## 79 80 81 82 83 84 \n## 0.01028001 0.02920514 0.01348051 0.01752758 0.05184527 0.04583604 \n## 85 86 87 88 89 90 \n## 0.05825858 0.01359644 0.03054414 0.01487724 0.02381348 0.02159418 \n## 91 92 93 94 95 96 \n## 0.02598661 0.02093288 0.01982480 0.05063492 0.05907629 0.03682026 \n## 97 98 99 100 \n## 0.01817919 0.03811718 0.01945603 0.01373394\n```\n\nIn general, if an observation lies far away from the other ones in terms of the predictor variables, it will have a large hat value (as a rough guide, high leverage is when the hat value is more than 2-3 times the average; and note that the sum of the hat values is constrained to be equal to \\(K+1\\)). High leverage points are also worth looking at in more detail, but they’re much less likely to be a cause for concern unless they are also outliers. % guide from Venables and Ripley.\n\nThis brings us to our third measure of unusualness, the influence of an observation. A high influence observation is an outlier that has high leverage. That is, it is an observation that is very different to all the other ones in some respect, and also lies a long way from the regression line. This is illustrated in Figure 15.9. Notice the contrast to the previous two figures: outliers don’t move the regression line much, and neither do high leverage points. But something that is an outlier and has high leverage… that has a big effect on the regression line.\n\nThat’s why we call these points high influence; and it’s why they’re the biggest worry. We operationalise influence in terms of a measure known as Cook’s distance, \\[ D_i = \\frac{{\\epsilon_i^*}^2 }{K+1} \\times \\frac{h_i}{1-h_i} \\] Notice that this is a multiplication of something that measures the outlier-ness of the observation (the bit on the left), and something that measures the leverage of the observation (the bit on the right). In other words, in order to have a large Cook’s distance, an observation must be a fairly substantial outlier and have high leverage. In a stunning turn of events, you can obtain these values using the following command:\n\n```\ncooks.distance( model = regression.2 )\n```\n\n```\n## 1 2 3 4 5 \n## 1.736512e-03 1.740243e-02 4.699370e-03 1.301417e-03 2.631557e-04 \n## 6 7 8 9 10 \n## 9.697585e-05 2.620181e-05 5.458491e-04 2.876269e-05 3.288277e-03 \n## 11 12 13 14 15 \n## 2.731835e-02 8.296919e-03 6.621479e-04 1.235956e-04 2.538915e-03 \n## 16 17 18 19 20 \n## 5.012283e-05 3.461742e-03 2.098055e-03 1.957050e-04 1.780519e-02 \n## 21 22 23 24 25 \n## 8.367377e-05 1.649478e-03 2.967594e-02 2.657610e-05 3.448032e-04 \n## 26 27 28 29 30 \n## 9.093379e-04 5.699951e-02 7.307943e-04 5.716998e-04 2.459564e-03 \n## 31 32 33 34 35 \n## 5.214331e-02 6.185200e-03 2.700686e-02 5.467345e-02 2.071643e-02 \n## 36 37 38 39 40 \n## 4.606378e-02 1.765312e-02 4.689817e-02 2.316122e-03 7.012530e-05 \n## 41 42 43 44 45 \n## 2.827824e-05 1.076083e-02 2.399931e-02 3.381062e-02 1.406498e-02 \n## 46 47 48 49 50 \n## 1.801086e-02 1.561699e-02 4.179986e-03 8.483514e-03 2.444787e-04 \n## 51 52 53 54 55 \n## 3.689946e-02 7.794472e-04 1.941235e-03 2.446230e-03 4.270361e-03 \n## 56 57 58 59 60 \n## 1.266609e-03 1.824212e-03 4.804705e-04 1.163181e-03 3.187235e-03 \n## 61 62 63 64 65 \n## 1.595512e-03 3.703826e-04 1.577892e-04 1.138165e-01 2.827715e-02 \n## 66 67 68 69 70 \n## 1.139374e-02 1.231422e-03 2.260006e-03 2.241322e-04 1.028479e-02 \n## 71 72 73 74 75 \n## 2.841329e-03 1.431223e-03 3.468538e-04 2.941757e-03 2.738249e-03 \n## 76 77 78 79 80 \n## 5.045357e-04 1.387108e-02 4.230966e-02 4.187440e-03 3.861831e-03 \n## 81 82 83 84 85 \n## 2.965826e-02 1.838888e-04 2.149369e-03 1.993929e-04 9.168733e-02 \n## 86 87 88 89 90 \n## 3.198994e-04 3.262192e-04 4.547383e-05 3.400893e-04 7.881487e-04 \n## 91 92 93 94 95 \n## 4.801204e-03 4.493095e-03 1.188427e-02 4.796360e-03 1.752666e-02 \n## 96 97 98 99 100 \n## 1.732793e-04 1.012302e-02 1.225818e-02 2.394964e-02 8.010508e-03\n```\n\nAs a rough guide, Cook’s distance greater than 1 is often considered large (that’s what I typically use as a quick and dirty rule), though a quick scan of the internet and a few papers suggests that \\(4/N\\) has also been suggested as a possible rule of thumb.\nAs hinted above, you don’t usually need to make use of these functions, since you can have R automatically draw the critical plots.223 For the `regression.2` model, these are the plots showing Cook’s distance (Figure 15.10) and the more detailed breakdown showing the scatter plot of the Studentised residual against leverage (Figure 15.11). To draw these, we can use the `plot()` function. When the main argument `x` to this function is a linear model object, it will draw one of six different plots, each of which is quite useful for doing regression diagnostics. You specify which one you want using the `which` argument (a number between 1 and 6). If you don’t do this then R will draw all six. The two plots of interest to us in this context are generated using the following commands: An obvious question to ask next is, if you do have large values of Cook’s distance, what should you do? As always, there’s no hard and fast rules. Probably the first thing to do is to try running the regression with that point excluded and see what happens to the model performance and to the regression coefficients. If they really are substantially different, it’s time to start digging into your data set and your notes that you no doubt were scribbling as your ran your study; try to figure out why the point is so different. If you start to become convinced that this one data point is badly distorting your results, you might consider excluding it, but that’s less than ideal unless you have a solid explanation for why this particular case is qualitatively different from the others and therefore deserves to be handled separately.224 To give an example, let’s delete the observation from day 64, the observation with the largest Cook’s distance for the `regression.2` model. We can do this using the `subset` argument: \n\n```\nlm( formula = dan.grump ~ dan.sleep + baby.sleep, # same formula\n data = parenthood, # same data frame...\n subset = -64 # ...but observation 64 is deleted\n )\n```\n\n```\n## \n## Call:\n## lm(formula = dan.grump ~ dan.sleep + baby.sleep, data = parenthood, \n## subset = -64)\n## \n## Coefficients:\n## (Intercept) dan.sleep baby.sleep \n## 126.3553 -8.8283 -0.1319\n```\n\nAs you can see, those regression coefficients have barely changed in comparison to the values we got earlier. In other words, we really don’t have any problem as far as anomalous data are concerned.\n\n### 15.9.3 Checking the normality of the residuals\n\nLike many of the statistical tools we’ve discussed in this book, regression models rely on a normality assumption. In this case, we assume that the residuals are normally distributed. The tools for testing this aren’t fundamentally different to those that we discussed earlier in Section 13.9. Firstly, I firmly believe that it never hurts to draw an old fashioned histogram. The command I use might be something like this:\n\n```\nhist( x = residuals( regression.2 ), # data are the residuals\n xlab = \"Value of residual\", # x-axis label\n main = \"\", # no title \n breaks = 20 # lots of breaks\n )\n```\n\nThe resulting plot is shown in Figure 15.12, and as you can see the plot looks pretty damn close to normal, almost unnaturally so.\n I could also run a Shapiro-Wilk test to check, using the `shapiro.test()` function; the \\(W\\) value of .99, at this sample size, is non-significant (\\(p=.84\\)), again suggesting that the normality assumption isn’t in any danger here. As a third measure, we might also want to draw a QQ-plot using the `qqnorm()` function. The QQ plot is an excellent one to draw, and so you might not be surprised to discover that it’s one of the regression plots that we can produce using the `plot()` function: \n\n```\nplot( x = regression.2, which = 2 ) # Figure @ref{fig:regressionplot2}\n```\n\nThe output is shown in Figure 15.13, showing the standardised residuals plotted as a function of their theoretical quantiles according to the regression model. The fact that the output appends the model specification to the picture is nice.\n\n### 15.9.4 Checking the linearity of the relationship\n\n The third thing we might want to test is the linearity of the relationships between the predictors and the outcomes. There’s a few different things that you might want to do in order to check this. Firstly, it never hurts to just plot the relationship between the fitted values \\(\\hat{Y}_i\\) and the observed values \\(Y_i\\) for the outcome variable, as illustrated in Figure 15.14. To draw this we could use the `fitted.values()` function to extract the \\(\\hat{Y_i}\\) values in much the same way that we used the `residuals()` function to extract the \\(\\epsilon_i\\) values. So the commands to draw this figure might look like this: \n\n```\nyhat.2 <- fitted.values( object = regression.2 )\n plot( x = yhat.2, \n y = parenthood$dan.grump,\n xlab = \"Fitted Values\",\n ylab = \"Observed Values\" \n )\n```\n\nOne of the reasons I like to draw these plots is that they give you a kind of “big picture view”. If this plot looks approximately linear, then we’re probably not doing too badly (though that’s not to say that there aren’t problems). However, if you can see big departures from linearity here, then it strongly suggests that you need to make some changes.\n In any case, in order to get a more detailed picture it’s often more informative to look at the relationship between the fitted values and the residuals themselves. Again, we could draw this plot using low level commands, but there’s an easier way. Just `plot()` the regression model, and select `which = 1` : \n\n```\nplot(x = regression.2, which = 1)\n```\n\nThe output is shown in Figure 15.15. As you can see, not only does it draw the scatterplot showing the fitted value against the residuals, it also plots a line through the data that shows the relationship between the two. Ideally, this should be a straight, perfectly horizontal line. There’s some hint of curvature here, but it’s not clear whether or not we be concerned.\n A somewhat more advanced version of the same plot is produced by the `residualPlots()` function in the `car` package. This function not only draws plots comparing the fitted values to the residuals, it does so for each individual predictor. The command is and the resulting plots are shown in Figure 15.16. \n\n```\n## Test stat Pr(>|Test stat|) \n## dan.sleep 2.1604 0.03323 *\n## baby.sleep -0.5445 0.58733 \n## Tukey test 2.1615 0.03066 *\n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nNote that this function also reports the results of a bunch of curvature tests. For a predictor variable \\(X\\) in some regression model, this test is equivalent to adding a new predictor to the model corresponding to \\(X^2\\), and running the \\(t\\)-test on the \\(b\\) coefficient associated with this new predictor. If it comes up significant, it implies that there is some nonlinear relationship between the variable and the residuals.\n\nThe third line here is the Tukey test, which is basically the same test, except that instead of squaring one of the predictors and adding it to the model, you square the fitted-value. In any case, the fact that the curvature tests have come up significant is hinting that the curvature that we can see in Figures 15.15 and 15.16 is genuine;225 although it still bears remembering that the pattern in Figure 15.14 is pretty damn straight: in other words the deviations from linearity are pretty small, and probably not worth worrying about.\n In a lot of cases, the solution to this problem (and many others) is to transform one or more of the variables. We discussed the basics of variable transformation in Sections 7.2 and (mathfunc), but I do want to make special note of one additional possibility that I didn’t mention earlier: the Box-Cox transform. The Box-Cox function is a fairly simple one, but it’s very widely used \\[ f(x,\\lambda) = \\frac{x^\\lambda - 1}{\\lambda} \\] for all values of \\(\\lambda\\) except \\(\\lambda = 0\\). When \\(\\lambda = 0\\) we just take the natural logarithm (i.e., \\(\\ln(x)\\)). You can calculate it using the `boxCox()` function in the `car` package. Better yet, if what you’re trying to do is convert a data to normal, or as normal as possible, there’s the `powerTransform()` function in the `car` package that can estimate the best value of \\(\\lambda\\). Variable transformation is another topic that deserves a fairly detailed treatment, but (again) due to deadline constraints, it will have to wait until a future version of this book. \n\n### 15.9.5 Checking the homogeneity of variance\n\n The regression models that we’ve talked about all make a homogeneity of variance assumption: the variance of the residuals is assumed to be constant. The “default” plot that R provides to help with doing this ( `which = 3` when using `plot()` ) shows a plot of the square root of the size of the residual \\(\\sqrt{|\\epsilon_i|}\\), as a function of the fitted value \\(\\hat{Y}_i\\). We can produce the plot using the following command, \n\n```\nplot(x = regression.2, which = 3)\n```\n\nand the resulting plot is shown in Figure 15.17. Note that this plot actually uses the standardised residuals (i.e., converted to \\(z\\) scores) rather than the raw ones, but it’s immaterial from our point of view. What we’re looking to see here is a straight, horizontal line running through the middle of the plot.\n A slightly more formal approach is to run hypothesis tests. The `car` package provides a function called `ncvTest()` (non-constant variance test) that can be used for this purpose (Cook and Weisberg 1983). I won’t explain the details of how it works, other than to say that the idea is that what you do is run a regression to see if there is a relationship between the squared residuals \\(\\epsilon_i\\) and the fitted values \\(\\hat{Y}_i\\), or possibly to run a regression using all of the original predictors instead of just \\(\\hat{Y}_i\\).226 Using the default settings, the `ncvTest()` looks for a relationship between \\(\\hat{Y}_i\\) and the variance of the residuals, making it a straightforward analogue of Figure 15.17. So if we run it for our model, \n\n```\nncvTest( regression.2 )\n```\n\n```\n## Non-constant Variance Score Test \n## Variance formula: ~ fitted.values \n## Chisquare = 0.09317511, Df = 1, p = 0.76018\n```\n\nWe see that our original impression was right: there’s no violations of homogeneity of variance in this data.\n It’s a bit beyond the scope of this chapter to talk too much about how to deal with violations of homogeneity of variance, but I’ll give you a quick sense of what you need to consider. The main thing to worry about, if homogeneity of variance is violated, is that the standard error estimates associated with the regression coefficients are no longer entirely reliable, and so your \\(t\\) tests for the coefficients aren’t quite right either. A simple fix to the problem is to make use of a “heteroscedasticity corrected covariance matrix” when estimating the standard errors. These are often called sandwich estimators, for reasons that only make sense if you understand the maths at a low level227 have implemented as the default in the `hccm()` function is a tweak on this, proposed by (2000). This version uses \\(\\Sigma = \\mbox{diag}(\\epsilon_i^2/(1-h_i^2))\\), where \\(h_i\\) is the \\(i\\)th hat value. Gosh, regression is fun, isn’t it?] You don’t need to understand what this means (not for an introductory class), but it might help to note that there’s a `hccm()` function in the `car()` package that does it. Better yet, you don’t even need to use it. You can use the `coeftest()` function in the `lmtest` package, but you need the `car` package loaded: \n\n```\nlibrary(lmtest)\nlibrary(car)\ncoeftest( regression.2, vcov= hccm )\n```\n\n```\n## \n## t test of coefficients:\n## \n## Estimate Std. Error t value Pr(>|t|) \n## (Intercept) 125.965566 3.247285 38.7910 <2e-16 ***\n## dan.sleep -8.950250 0.615820 -14.5339 <2e-16 ***\n## baby.sleep 0.010524 0.291565 0.0361 0.9713 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n Not surprisingly, these \\(t\\) tests are pretty much identical to the ones that we saw when we used the\n\n```\nsummary(regression.2)\n```\n\ncommand earlier; because the homogeneity of variance assumption wasn’t violated. But if it had been, we might have seen some more substantial differences. \n\n### 15.9.6 Checking for collinearity\n\n The last kind of regression diagnostic that I’m going to discuss in this chapter is the use of variance inflation factors (VIFs), which are useful for determining whether or not the predictors in your regression model are too highly correlated with each other. There is a variance inflation factor associated with each predictor \\(X_k\\) in the model, and the formula for the \\(k\\)-th VIF is: \\[ \\mbox{VIF}_k = \\frac{1}{1-{R^2_{(-k)}}} \\] where \\(R^2_{(-k)}\\) refers to \\(R\\)-squared value you would get if you ran a regression using \\(X_k\\) as the outcome variable, and all the other \\(X\\) variables as the predictors. The idea here is that \\(R^2_{(-k)}\\) is a very good measure of the extent to which \\(X_k\\) is correlated with all the other variables in the model. Better yet, the square root of the VIF is pretty interpretable: it tells you how much wider the confidence interval for the corresponding coefficient \\(b_k\\) is, relative to what you would have expected if the predictors are all nice and uncorrelated with one another. If you’ve only got two predictors, the VIF values are always going to be the same, as we can see if we use the `vif()` function ( `car` package)… \n\n```\nvif( mod = regression.2 )\n```\n\n```\n## dan.sleep baby.sleep \n## 1.651038 1.651038\n```\n\nAnd since the square root of 1.65 is 1.28, we see that the correlation between our two predictors isn’t causing much of a problem.\n To give a sense of how we could end up with a model that has bigger collinearity problems, suppose I were to run a much less interesting regression model, in which I tried to predict the `day` on which the data were collected, as a function of all the other variables in the data set. To see why this would be a bit of a problem, let’s have a look at the correlation matrix for all four variables: `cor( parenthood )` \n\nWe have some fairly large correlations between some of our predictor variables! When we run the regression model and look at the VIF values, we see that the collinearity is causing a lot of uncertainty about the coefficients. First, run the regression…\n\n```\nregression.3 <- lm( day ~ baby.sleep + dan.sleep + dan.grump, parenthood )\n```\n\nand second, look at the VIFs…\n `vif( regression.3 )` \n\n```\n## baby.sleep dan.sleep dan.grump \n## 1.651064 6.102337 5.437903\n```\n\nYep, that’s some mighty fine collinearity you’ve got there.\n\n## 15.10 Model selection\n\nOne fairly major problem that remains is the problem of “model selection”. That is, if we have a data set that contains several variables, which ones should we include as predictors, and which ones should we not include? In other words, we have a problem of variable selection. In general, model selection is a complex business, but it’s made somewhat simpler if we restrict ourselves to the problem of choosing a subset of the variables that ought to be included in the model. Nevertheless, I’m not going to try covering even this reduced topic in a lot of detail. Instead, I’ll talk about two broad principles that you need to think about; and then discuss one concrete tool that R provides to help you select a subset of variables to include in your model. Firstly, the two principles:\n\n* It’s nice to have an actual substantive basis for your choices. That is, in a lot of situations you the researcher have good reasons to pick out a smallish number of possible regression models that are of theoretical interest; these models will have a sensible interpretation in the context of your field. Never discount the importance of this. Statistics serves the scientific process, not the other way around.\n * To the extent that your choices rely on statistical inference, there is a trade off between simplicity and goodness of fit. As you add more predictors to the model, you make it more complex; each predictor adds a new free parameter (i.e., a new regression coefficient), and each new parameter increases the model’s capacity to “absorb” random variations. So the goodness of fit (e.g., \\(R^2\\)) continues to rise as you add more predictors no matter what. If you want your model to be able to generalise well to new observations, you need to avoid throwing in too many variables.\n\nThis latter principle is often referred to as Ockham’s razor, and is often summarised in terms of the following pithy saying: do not multiply entities beyond necessity. In this context, it means: don’t chuck in a bunch of largely irrelevant predictors just to boost your \\(R^2\\). Hm. Yeah, the original was better.\n In any case, what we need is an actual mathematical criterion that will implement the qualitative principle behind Ockham’s razor in the context of selecting a regression model. As it turns out there are several possibilities. The one that I’ll talk about is the Akaike information criterion (AIC; Akaike 1974) simply because it’s the default one used in the R function `step()` . In the context of a linear regression model (and ignoring terms that don’t depend on the model in any way!), the AIC for a model that has \\(K\\) predictor variables plus an intercept is:228 \\[\n\\mbox{AIC} = \\displaystyle\\frac{\\mbox{SS}_{res}}{\\hat{\\sigma}}^2+ 2K\n\\] The smaller the AIC value, the better the model performance is. If we ignore the low level details, it’s fairly obvious what the AIC does: on the left we have a term that increases as the model predictions get worse; on the right we have a term that increases as the model complexity increases. The best model is the one that fits the data well (low residuals; left hand side) using as few predictors as possible (low \\(K\\); right hand side). In short, this is a simple implementation of Ockham’s razor. \n\n### 15.10.1 Backward elimination\n\n Okay, let’s have a look at the `step()` function at work. In this example I’ll keep it simple and use only the basic backward elimination approach. That is, start with the complete regression model, including all possible predictors. Then, at each “step” we try all possible ways of removing one of the variables, and whichever of these is best (in terms of lowest AIC value) is accepted. This becomes our new regression model; and we then try all possible deletions from the new model, again choosing the option with lowest AIC. This process continues until we end up with a model that has a lower AIC value than any of the other possible models that you could produce by deleting one of its predictors. Let’s see this in action. First, I need to define the model from which the process starts. \n\n```\nfull.model <- lm( formula = dan.grump ~ dan.sleep + baby.sleep + day, \n data = parenthood \n )\n```\n\n That’s nothing terribly new: yet another regression. Booooring. Still, we do need to do it: the `object` argument to the `step()` function will be this regression model. With this in mind, I would call the `step()` function using the following command: \n\n```\nstep( object = full.model, # start at the full model\n direction = \"backward\" # allow it remove predictors but not add them\n )\n```\n\n```\n## Start: AIC=299.08\n## dan.grump ~ dan.sleep + baby.sleep + day\n## \n## Df Sum of Sq RSS AIC\n## - baby.sleep 1 0.1 1837.2 297.08\n## - day 1 1.6 1838.7 297.16\n## 1837.1 299.08\n## - dan.sleep 1 4909.0 6746.1 427.15\n## \n## Step: AIC=297.08\n## dan.grump ~ dan.sleep + day\n## \n## Df Sum of Sq RSS AIC\n## - day 1 1.6 1838.7 295.17\n## 1837.2 297.08\n## - dan.sleep 1 8103.0 9940.1 463.92\n## \n## Step: AIC=295.17\n## dan.grump ~ dan.sleep\n## \n## Df Sum of Sq RSS AIC\n## 1838.7 295.17\n## - dan.sleep 1 8159.9 9998.6 462.50\n```\n\n although in practice I didn’t need to specify `direction` because `\"backward\"` is the default. The output is somewhat lengthy, so I’ll go through it slowly. Firstly, the output reports the AIC value for the current best model: \n\n```\nStart: AIC=299.08\ndan.grump ~ dan.sleep + baby.sleep + day\n```\n\nThat’s our starting point. Since small AIC values are good, we want to see if we can get a value smaller than 299.08 by deleting one of those three predictors. So what R does is try all three possibilities, calculate the AIC values for each one, and then print out a short table with the results:\n\n```\nDf Sum of Sq RSS AIC\n- baby.sleep 1 0.1 1837.2 297.08\n- day 1 1.6 1838.7 297.16\n 1837.1 299.08\n- dan.sleep 1 4909.0 6746.1 427.15\n```\n\n To read this table, it helps to note that the text in the left hand column is telling you what change R made to the regression model. So the line that reads `` is the actual model we started with, and you can see on the right hand side that this still corresponds to an AIC value of 299.08 (obviously). The other three rows in the table correspond to the other three models that it looked at: it tried removing the `baby.sleep` variable, which is indicated by `- baby.sleep` , and this produced an AIC value of 297.08. That was the best of the three moves, so it’s at the top of the table. So, this move is accepted, and now we start again. There are two predictors left in the model, `dan.sleep` and `day` , so it tries deleting those: \n\n```\nStep: AIC=297.08\ndan.grump ~ dan.sleep + day\n\n Df Sum of Sq RSS AIC\n- day 1 1.6 1838.7 295.17\n 1837.2 297.08\n- dan.sleep 1 8103.0 9940.1 463.92\n```\n\n Okay, so what we can see is that removing the `day` variable lowers the AIC value from 297.08 to 295.17. So R decides to keep that change too, and moves on: \n\n Df Sum of Sq RSS AIC\n 1838.7 295.17\n- dan.sleep 1 8159.9 9998.6 462.50\n```\n\n This time around, there’s no further deletions that can actually improve the AIC value. So the `step()` function stops, and prints out the result of the best regression model it could find: \n\n which is (perhaps not all that surprisingly) the `regression.1` model that we started with at the beginning of the chapter. \n\n### 15.10.2 Forward selection\n\n As an alternative, you can also try forward selection. This time around we start with the smallest possible model as our start point, and only consider the possible additions to the model. However, there’s one complication: you also need to tell `step()` what the largest possible model you’re willing to entertain is, using the `scope` argument. The simplest usage is like this: \n\n```\nnull.model <- lm( dan.grump ~ 1, parenthood ) # intercept only.\n step( object = null.model, # start with null.model\n direction = \"forward\", # only consider \"addition\" moves\n scope = dan.grump ~ dan.sleep + baby.sleep + day # largest model allowed\n )\n```\n\n```\n## Start: AIC=462.5\n## dan.grump ~ 1\n## \n## Df Sum of Sq RSS AIC\n## + dan.sleep 1 8159.9 1838.7 295.17\n## + baby.sleep 1 3202.7 6795.9 425.89\n## 9998.6 462.50\n## + day 1 58.5 9940.1 463.92\n## \n## Step: AIC=295.17\n## dan.grump ~ dan.sleep\n## \n## Df Sum of Sq RSS AIC\n## 1838.7 295.17\n## + day 1 1.55760 1837.2 297.08\n## + baby.sleep 1 0.02858 1838.7 297.16\n```\n\n If I do this, the output takes on a similar form, but now it only considers addition ( `+` ) moves rather than deletion ( `-` ) moves: \n\n```\nStart: AIC=462.5\ndan.grump ~ 1\n\n Df Sum of Sq RSS AIC\n+ dan.sleep 1 8159.9 1838.7 295.17\n+ baby.sleep 1 3202.7 6795.9 425.89\n 9998.6 462.50\n+ day 1 58.5 9940.1 463.92\n\n Df Sum of Sq RSS AIC\n 1838.7 295.17\n+ day 1 1.55760 1837.2 297.08\n+ baby.sleep 1 0.02858 1838.7 297.16\n\nAs you can see, it’s found the same model. In general though, forward and backward selection don’t always have to end up in the same place.\n\n### 15.10.3 A caveat\n\n Automated variable selection methods are seductive things, especially when they’re bundled up in (fairly) simple functions like `step()` . They provide an element of objectivity to your model selection, and that’s kind of nice. Unfortunately, they’re sometimes used as an excuse for thoughtlessness. No longer do you have to think carefully about which predictors to add to the model and what the theoretical basis for their inclusion might be… everything is solved by the magic of AIC. And if we start throwing around phrases like Ockham’s razor, well, it sounds like everything is wrapped up in a nice neat little package that no-one can argue with. Or, perhaps not. Firstly, there’s very little agreement on what counts as an appropriate model selection criterion. When I was taught backward elimination as an undergraduate, we used \\(F\\)-tests to do it, because that was the default method used by the software. The default in the `step()` function is AIC, and since this is an introductory text that’s the only method I’ve described, but the AIC is hardly the Word of the Gods of Statistics. It’s an approximation, derived under certain assumptions, and it’s guaranteed to work only for large samples when those assumptions are met. Alter those assumptions and you get a different criterion, like the BIC for instance. Take a different approach again and you get the NML criterion. Decide that you’re a Bayesian and you get model selection based on posterior odds ratios. Then there are a bunch of regression specific tools that I haven’t mentioned. And so on. All of these different methods have strengths and weaknesses, and some are easier to calculate than others (AIC is probably the easiest of the lot, which might account for its popularity). Almost all of them produce the same answers when the answer is “obvious” but there’s a fair amount of disagreement when the model selection problem becomes hard. What does this mean in practice? Well, you could go and spend several years teaching yourself the theory of model selection, learning all the ins and outs of it; so that you could finally decide on what you personally think the right thing to do is. Speaking as someone who actually did that, I wouldn’t recommend it: you’ll probably come out the other side even more confused than when you started. A better strategy is to show a bit of common sense… if you’re staring at the results of a `step()` procedure, and the model that makes sense is close to having the smallest AIC, but is narrowly defeated by a model that doesn’t make any sense… trust your instincts. Statistical model selection is an inexact tool, and as I said at the beginning, interpretability matters. \n\n### 15.10.4 Comparing two regression models\n\n An alternative to using automated model selection procedures is for the researcher to explicitly select two or more regression models to compare to each other. You can do this in a few different ways, depending on what research question you’re trying to answer. Suppose we want to know whether or not the amount of sleep that my son got has any relationship to my grumpiness, over and above what we might expect from the amount of sleep that I got. We also want to make sure that the day on which we took the measurement has no influence on the relationship. That is, we’re interested in the relationship between `baby.sleep` and `dan.grump` , and from that perspective `dan.sleep` and `day` are nuisance variable or covariates that we want to control for. In this situation, what we would like to know is whether \n\n(which I’ll call Model 1, or `M1` ) is a better regression model for these data than \n\n(which I’ll call Model 0, or `M0` ). There are two different ways we can compare these two models, one based on a model selection criterion like AIC, and the other based on an explicit hypothesis test. I’ll show you the AIC based approach first because it’s simpler, and follows naturally from the `step()` function that we saw in the last section. The first thing I need to do is actually run the regressions: \n\n```\nM0 <- lm( dan.grump ~ dan.sleep + day, parenthood )\nM1 <- lm( dan.grump ~ dan.sleep + day + baby.sleep, parenthood )\n```\n\n Now that I have my regression models, I could use the `summary()` function to run various hypothesis tests and other useful statistics, just as we have discussed throughout this chapter. However, since the current focus on model comparison, I’ll skip this step and go straight to the AIC calculations. Conveniently, the `AIC()` function in R lets you input several regression models, and it will spit out the AIC values for each of them:229 `AIC( M0, M1 )` \n\n```\n## df AIC\n## M0 4 582.8681\n## M1 5 584.8646\n```\n\nSince Model 0 has the smaller AIC value, it is judged to be the better model for these data.\n\nA somewhat different approach to the problem comes out of the hypothesis testing framework. Suppose you have two regression models, where one of them (Model 0) contains a subset of the predictors from the other one (Model 1). That is, Model 1 contains all of the predictors included in Model 0, plus one or more additional predictors. When this happens we say that Model 0 is nested within Model 1, or possibly that Model 0 is a submodel of Model 1. Regardless of the terminology what this means is that we can think of Model 0 as a null hypothesis and Model 1 as an alternative hypothesis. And in fact we can construct an \\(F\\) test for this in a fairly straightforward fashion. We can fit both models to the data and obtain a residual sum of squares for both models. I’ll denote these as SS\\(_{res}^{(0)}\\) and SS\\(_{res}^{(1)}\\) respectively. The superscripting here just indicates which model we’re talking about. Then our \\(F\\) statistic is \\[ F = \\frac{(\\mbox{SS}_{res}^{(0)} - \\mbox{SS}_{res}^{(1)})/k}{(\\mbox{SS}_{res}^{(1)})/(N-p-1)} \\] where \\(N\\) is the number of observations, \\(p\\) is the number of predictors in the full model (not including the intercept), and \\(k\\) is the difference in the number of parameters between the two models.230 The degrees of freedom here are \\(k\\) and \\(N-p-1\\). Note that it’s often more convenient to think about the difference between those two SS values as a sum of squares in its own right. That is: \\[ \\mbox{SS}_\\Delta = \\mbox{SS}_{res}^{(0)} - \\mbox{SS}_{res}^{(1)} \\] The reason why this his helpful is that we can express \\(\\mbox{SS}_\\Delta\\) a measure of the extent to which the two models make different predictions about the the outcome variable. Specifically: \\[ \\mbox{SS}_\\Delta = \\sum_{i} \\left( \\hat{y}_i^{(1)} - \\hat{y}_i^{(0)} \\right)^2 \\] where \\(\\hat{y}_i^{(0)}\\) is the fitted value for \\(y_i\\) according to model \\(M_0\\) and \\(\\hat{y}_i^{(1)}\\) is the is the fitted value for \\(y_i\\) according to model \\(M_1\\).\n Okay, so that’s the hypothesis test that we use to compare two regression models to one another. Now, how do we do it in R? The answer is to use the `anova()` function. All we have to do is input the two models that we want to compare (null model first): `anova( M0, M1 )` \n\n```\n## Analysis of Variance Table\n## \n## Model 1: dan.grump ~ dan.sleep + day\n## Model 2: dan.grump ~ dan.sleep + day + baby.sleep\n## Res.Df RSS Df Sum of Sq F Pr(>F)\n## 1 97 1837.2 \n## 2 96 1837.1 1 0.063688 0.0033 0.9541\n```\n\n Note that, just like we saw with the output from the `step()` function, R has used the acronym `RSS` to refer to the residual sum of squares from each model. That is, RSS in this output corresponds to SS\\(_{res}\\) in the formula above. Since we have \\(p>.05\\) we retain the null hypothesis ( `M0` ). This approach to regression, in which we add all of our covariates into a null model, and then add the variables of interest into an alternative model, and then compare the two models in hypothesis testing framework, is often referred to as hierarchical regression. \n\n## 15.11 Summary\n\n* Basic ideas in linear regression and how regression models are estimated (Sections 15.1 and 15.2).\n * Multiple linear regression (Section 15.3).\n * Measuring the overall performance of a regression model using \\(R^2\\) (Section 15.4)\n * Hypothesis tests for regression models (Section 15.5)\n * Calculating confidence intervals for regression coefficients, and standardised coefficients (Section 15.7)\n * The assumptions of regression (Section 15.8) and how to check them (Section 15.9)\n * Selecting a regression model (Section 15.10)\n\nThe \\(\\epsilon\\) symbol is the Greek letter epsilon. It’s traditional to use \\(\\epsilon_i\\) or \\(e_i\\) to denote a residual.↩\n * \n\nOr at least, I’m assuming that it doesn’t help most people. But on the off chance that someone reading this is a proper kung fu master of linear algebra (and to be fair, I always have a few of these people in my intro stats class), it will help you to know that the solution to the estimation problem turns out to be \\(\\hat{b} = (X^TX)^{-1} X^T y\\), where \\(\\hat{b}\\) is a vector containing the estimated regression coefficients, \\(X\\) is the “design matrix” that contains the predictor variables (plus an additional column containing all ones; strictly \\(X\\) is a matrix of the regressors, but I haven’t discussed the distinction yet), and \\(y\\) is a vector containing the outcome variable. For everyone else, this isn’t exactly helpful, and can be downright scary. However, since quite a few things in linear regression can be written in linear algebra terms, you’ll see a bunch of footnotes like this one in this chapter. If you can follow the maths in them, great. If not, ignore it.↩\n * \n\nAnd by “sometimes” I mean “almost never”. In practice everyone just calls it “\\(R\\)-squared”.↩\n * \n\nNote that, although R has done multiple tests here, it hasn’t done a Bonferroni correction or anything. These are standard one-sample \\(t\\)-tests with a two-sided alternative. If you want to make corrections for multiple tests, you need to do that yourself.↩\n * \n\nYou can change the kind of correction it applies by specifying the\n `p.adjust.method` argument.↩ * \n\nStrictly, you standardise all the regressors: that is, every “thing” that has a regression coefficient associated with it in the model. For the regression models that I’ve talked about so far, each predictor variable maps onto exactly one regressor, and vice versa. However, that’s not actually true in general: we’ll see some examples of this in Chapter 16. But for now, we don’t need to care too much about this distinction.↩\n * \n\nOr have no hope, as the case may be.↩\n * \n\nAgain, for the linear algebra fanatics: the “hat matrix” is defined to be that matrix \\(H\\) that converts the vector of observed values \\(y\\) into a vector of fitted values \\(\\hat{y}\\), such that \\(\\hat{y} = H y\\). The name comes from the fact that this is the matrix that “puts a hat on \\(y\\)”. The hat value of the \\(i\\)-th observation is the \\(i\\)-th diagonal element of this matrix (so technically I should be writing it as \\(h_{ii}\\) rather than \\(h_{i}\\)). Oh, and in case you care, here’s how it’s calculated: \\(H = X(X^TX)^{-1} X^T\\). Pretty, isn’t it?↩\n * \n\nThough special mention should be made of the\n `influenceIndexPlot()` and `influencePlot()` functions in the `car` package. These produce somewhat more detailed pictures than the default plots that I’ve shown here. There’s also an `outlierTest()` function that tests to see if any of the Studentised residuals are significantly larger than would be expected by chance.↩ * \n\nAn alternative is to run a “robust regression”; I’ll discuss robust regression in a later version of this book.↩\n * \n\nAnd, if you take the time to check the\n `residualPlots()` for `regression.1` , it’s pretty clear that this isn’t some wacky distortion being caused by the fact that `baby.sleep` is a useless predictor variable. It’s an actual nonlinearity in the relationship between `dan.sleep` and `dan.grump` .↩ * \n\nNote that the underlying mechanics of the test aren’t the same as the ones I’ve described for regressions; the goodness of fit is assessed using what’s known as a score-test not an \\(F\\)-test, and the test statistic is (approximately) \\(\\chi^2\\) distributed if there’s no relationship↩\n * \n\nAgain, a footnote that should be read only by the two readers of this book that love linear algebra (mmmm… I love the smell of matrix computations in the morning; smells like… nerd). In these estimators, the covariance matrix for \\(b\\) is given by \\((X^T X)^{-1}\\ X^T \\Sigma X \\ (X^T X)^{-1}\\). See, it’s a “sandwich”? Assuming you think that \\((X^T X)^{-1} = \\mbox{\"bread\"}\\) and \\(X^T \\Sigma X = \\mbox{\"filling\"}\\), that is. Which of course everyone does, right? In any case, the usual estimator is what you get when you set \\(\\Sigma = \\hat\\sigma^2 I\\). The corrected version that I learned originally uses \\(\\Sigma = \\mbox{diag}(\\epsilon_i^2)\\) (White 1980). However, the version that Fox and Weisberg (2011)↩\n * \n\nNote, however, that the\n `step()` function computes the full version of AIC, including the irrelevant constants that I’ve dropped here. As a consequence this equation won’t correctly describe the AIC values that you see in the outputs here. However, if you calculate the AIC values using my formula for two different regression models and take the difference between them, this will be the same as the differences between AIC values that `step()` reports. In practice, this is all you care about: the actual value of an AIC statistic isn’t very informative, but the differences between two AIC values are useful, since these provide a measure of the extent to which one model outperforms another.↩ * \n\nWhile I’m on this topic I should point out that there is also a function called\n `BIC()` which computes the Bayesian information criterion (BIC) for the models. So you could type `BIC(M0,M1)` and get a very similar output. In fact, while I’m not particularly impressed with either AIC or BIC as model selection methods, if you do find yourself using one of these two, the empirical evidence suggests that BIC is the better criterion of the two. In most simulation studies that I’ve seen, BIC does a much better job of selecting the correct model.↩ * \n\nIt’s worth noting in passing that this same \\(F\\) statistic can be used to test a much broader range of hypotheses than those that I’m mentioning here. Very briefly: notice that the nested model M0 corresponds to the full model M1 when we constrain some of the regression coefficients to zero. It is sometimes useful to construct submodels by placing other kinds of constraints on the regression coefficients. For instance, maybe two different coefficients might have to sum to zero, or something like that. You can construct hypothesis tests for those kind of constraints too, but it is somewhat more complicated and the sampling distribution for \\(F\\) can end up being something known as the non-central \\(F\\) distribution, which is waaaaay beyond the scope of this book! All I want to do is alert you to this possibility.↩\n\nDate: 2015-12-03\nCategories: \nTags: \n\n# Chapter 16 Factorial ANOVA\n\nOver the course of the last few chapters you can probably detect a general trend. We started out looking at tools that you can use to compare two groups to one another, most notably the \\(t\\)-test (Chapter 13). Then, we introduced analysis of variance (ANOVA) as a method for comparing more than two groups (Chapter 14). The chapter on regression (Chapter 15) covered a somewhat different topic, but in doing so it introduced a powerful new idea: building statistical models that have multiple predictor variables used to explain a single outcome variable. For instance, a regression model could be used to predict the number of errors a student makes in a reading comprehension test based on the number of hours they studied for the test, and their score on a standardised IQ test. The goal in this chapter is to import this idea into the ANOVA framework. For instance, suppose we were interested in using the reading comprehension test to measure student achievements in three different schools, and we suspect that girls and boys are developing at different rates (and so would be expected to have different performance on average). Each student is classified in two different ways: on the basis of their gender, and on the basis of their school. What we’d like to do is analyse the reading comprehension scores in terms of both of these grouping variables. The tool for doing so is generically referred to as factorial ANOVA. However, since we have two grouping variables, we sometimes refer to the analysis as a two-way ANOVA, in contrast to the one-way ANOVAs that we ran in Chapter 14.\n\n## 16.1 Factorial ANOVA 1: balanced designs, no interactions\n\n When we discussed analysis of variance in Chapter 14, we assumed a fairly simple experimental design: each person falls into one of several groups, and we want to know whether these groups have different means on some outcome variable. In this section, I’ll discuss a broader class of experimental designs, known as factorial designs, in we have more than one grouping variable. I gave one example of how this kind of design might arise above. Another example appears in Chapter 14, in which we were looking at the effect of different drugs on the `mood.gain` experienced by each person. In that chapter we did find a significant effect of drug, but at the end of the chapter we also ran an analysis to see if there was an effect of therapy. We didn’t find one, but there’s something a bit worrying about trying to run two separate analyses trying to predict the same outcome. Maybe there actually is an effect of therapy on mood gain, but we couldn’t find it because it was being “hidden” by the effect of drug? In other words, we’re going to want to run a single analysis that includes both `drug` and `therapy` as predictors. For this analysis each person is cross-classified by the drug they were given (a factor with 3 levels) and what therapy they received (a factor with 2 levels). We refer to this as a \\(3 \\times 2\\) factorial design. If we cross-tabulate `drug` by `therapy` , using the `xtabs()` function (see Section 7.1), we get the following table:231 \n\n```\nload(file.path(projecthome, \"data\",\"clinicaltrial.Rdata\"))\nxtabs( ~ drug + therapy, clin.trial )\n```\n\n```\n## therapy\n## drug no.therapy CBT\n## placebo 3 3\n## anxifree 3 3\n## joyzepam 3 3\n```\n\nAs you can see, not only do we have participants corresponding to all possible combinations of the two factors, indicating that our design is completely crossed, it turns out that there are an equal number of people in each group. In other words, we have a balanced design. In this section I’ll talk about how to analyse data from balanced designs, since this is the simplest case. The story for unbalanced designs is quite tedious, so we’ll put it to one side for the moment.\n\n### 16.1.1 What hypotheses are we testing?\n\nLike one-way ANOVA, factorial ANOVA is a tool for testing certain types of hypotheses about population means. So a sensible place to start would be to be explicit about what our hypotheses actually are. However, before we can even get to that point, it’s really useful to have some clean and simple notation to describe the population means. Because of the fact that observations are cross-classified in terms of two different factors, there are quite a lot of different means that one might be interested. To see this, let’s start by thinking about all the different sample means that we can calculate for this kind of design. Firstly, there’s the obvious idea that we might be interested in this table of group means:\n\nNow, this output shows a cross-tabulation of the group means for all possible combinations of the two factors (e.g., people who received the placebo and no therapy, people who received the placebo while getting CBT, etc). However, we can also construct tables that ignore one of the two factors. That gives us output that looks like this:\n\n```\naggregate( mood.gain ~ therapy, clin.trial, mean )\n```\n\n```\n## therapy mood.gain\n## 1 no.therapy 0.7222222\n## 2 CBT 1.0444444\n```\n\nBut of course, if we can ignore one factor we can certainly ignore both. That is, we might also be interested in calculating the average mood gain across all 18 participants, regardless of what drug or psychological therapy they were given:\n\n```\nmean( clin.trial$mood.gain )\n```\n\n `## [1] 0.8833333` \nAt this point we have 12 different sample means to keep track of! It is helpful to organise all these numbers into a single table, which would look like this:\n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2, ~V3, ~V4,\n NA, \"no therapy\", \"CBT\", \"total\",\n \"placebo\", \"0.30\", \"0.60\", \"0.45\",\n \"anxifree\", \"0.40\", \"1.03\", \"0.72\",\n \"joyzepam\", \"1.47\", \"1.50\", \"1.48\",\n \"total\", \"0.72\", \"1.04\", \"0.88\"\n ), col.names = c(\"\", \"no therapy\", \"CBT\", \"total\"))\n```\n\nno therapy | CBT | total |\n | --- | --- | --- |\nNA | no therapy | CBT | total |\nplacebo | 0.30 | 0.60 | 0.45 |\nanxifree | 0.40 | 1.03 | 0.72 |\njoyzepam | 1.47 | 1.50 | 1.48 |\ntotal | 0.72 | 1.04 | 0.88 |\n\nNow, each of these different means is of course a sample statistic: it’s a quantity that pertains to the specific observations that we’ve made during our study. What we want to make inferences about are the corresponding population parameters: that is, the true means as they exist within some broader population. Those population means can also be organised into a similar table, but we’ll need a little mathematical notation to do so. As usual, I’ll use the symbol \\(\\mu\\) to denote a population mean. However, because there are lots of different means, I’ll need to use subscripts to distinguish between them.\n Here’s how the notation works. Our table is defined in terms of two factors: each row corresponds to a different level of Factor A (in this case `drug` ), and each column corresponds to a different level of Factor B (in this case `therapy` ). If we let \\(R\\) denote the number of rows in the table, and \\(C\\) denote the number of columns, we can refer to this as an \\(R \\times C\\) factorial ANOVA. In this case \\(R=3\\) and \\(C=2\\). We’ll use lowercase letters to refer to specific rows and columns, so \\(\\mu_{rc}\\) refers to the population mean associated with the \\(r\\)th level of Factor A (i.e. row number \\(r\\)) and the \\(c\\)th level of Factor B (column number \\(c\\)).232 So the population means are now written like this: \n\n \"placebo\", \"$\\\\mu_{11}$\", \"$\\\\mu_{12}$\", \"\",\n \"anxifree\", \"$\\\\mu_{21}$\", \"$\\\\mu_{22}$\", \"\",\n \"joyzepam\", \"$\\\\mu_{31}$\", \"$\\\\mu_{32}$\", \"\",\n \"total\", \"\", \"\", \"\"\n ), col.names = c( \"\", \"no therapy\", \"CBT\", \"total\"))\n```\n\nno therapy | CBT | total |\n | --- | --- | --- |\nplacebo | \\(\\mu_{11}\\) | \\(\\mu_{12}\\) |\nanxifree | \\(\\mu_{21}\\) | \\(\\mu_{22}\\) |\njoyzepam | \\(\\mu_{31}\\) | \\(\\mu_{32}\\) |\ntotal |\n\nOkay, what about the remaining entries? For instance, how should we describe the average mood gain across the entire (hypothetical) population of people who might be given Joyzepam in an experiment like this, regardless of whether they were in CBT? We use the “dot” notation to express this. In the case of Joyzepam, notice that we’re talking about the mean associated with the third row in the table. That is, we’re averaging across two cell means (i.e., \\(\\mu_{31}\\) and \\(\\mu_{32}\\)). The result of this averaging is referred to as a marginal mean, and would be denoted \\(\\mu_{3.}\\) in this case. The marginal mean for CBT corresponds to the population mean associated with the second column in the table, so we use the notation \\(\\mu_{.2}\\) to describe it. The grand mean is denoted \\(\\mu_{..}\\) because it is the mean obtained by averaging (marginalising233) over both. So our full table of population means can be written down like this:\n\n \"placebo\", \"$\\\\mu_{11}$\", \"$\\\\mu_{12}$\", \"$\\\\mu_{1.}$\",\n \"anxifree\", \"$\\\\mu_{21}$\", \"$\\\\mu_{22}$\", \"$\\\\mu_{2.}$\",\n \"joyzepam\", \"$\\\\mu_{31}$\", \"$\\\\mu_{32}$\", \"$\\\\mu_{3.}$\",\n \"total\", \"$\\\\mu_{.1}$\", \"$\\\\mu_{.2}$\", \"$\\\\mu_{..}$\"\n ), col.names=c( NA, \"no therapy\", \"CBT\", \"total\"))\n```\n\nNA | no therapy | CBT | total |\n | --- | --- | --- | --- |\nplacebo | \\(\\mu_{11}\\) | \\(\\mu_{12}\\) | \\(\\mu_{1.}\\) |\nanxifree | \\(\\mu_{21}\\) | \\(\\mu_{22}\\) | \\(\\mu_{2.}\\) |\njoyzepam | \\(\\mu_{31}\\) | \\(\\mu_{32}\\) | \\(\\mu_{3.}\\) |\ntotal | \\(\\mu_{.1}\\) | \\(\\mu_{.2}\\) | \\(\\mu_{..}\\) |\n\nNow that we have this notation, it is straightforward to formulate and express some hypotheses. Let’s suppose that the goal is to find out two things: firstly, does the choice of drug have any effect on mood, and secondly, does CBT have any effect on mood? These aren’t the only hypotheses that we could formulate of course, and we’ll see a really important example of a different kind of hypothesis in Section 16.2, but these are the two simplest hypotheses to test, and so we’ll start there. Consider the first test. If drug has no effect, then we would expect all of the row means to be identical, right? So that’s our null hypothesis. On the other hand, if the drug does matter then we should expect these row means to be different. Formally, we write down our null and alternative hypotheses in terms of the equality of marginal means:\n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2,\n \"Null hypothesis $H_0$:\", \"row means are the same i.e. $\\\\mu_{1.} = \\\\mu_{2.} = \\\\mu_{3.}$\",\n \"Alternative hypothesis $H_1$:\", \"at least one row mean is different.\"\n ), col.names = c(\"\", \"\"))\n```\n\nNull hypothesis \\(H_0\\): | row means are the same i.e. \\(\\mu_{1.} = \\mu_{2.} = \\mu_{3.}\\) |\n | --- | --- |\nAlternative hypothesis \\(H_1\\): | at least one row mean is different. |\n\nIt’s worth noting that these are exactly the same statistical hypotheses that we formed when we ran a one-way ANOVA on these data back in Chapter 14. Back then I used the notation \\(\\mu_P\\) to refer to the mean mood gain for the placebo group, with \\(\\mu_A\\) and \\(\\mu_J\\) corresponding to the group means for the two drugs, and the null hypothesis was \\(\\mu_P = \\mu_A = \\mu_J\\). So we’re actually talking about the same hypothesis: it’s just that the more complicated ANOVA requires more careful notation due to the presence of multiple grouping variables, so we’re now referring to this hypothesis as \\(\\mu_{1.} = \\mu_{2.} = \\mu_{3.}\\). However, as we’ll see shortly, although the hypothesis is identical, the test of that hypothesis is subtly different due to the fact that we’re now acknowledging the existence of the second grouping variable.\n\nSpeaking of the other grouping variable, you won’t be surprised to discover that our second hypothesis test is formulated the same way. However, since we’re talking about the psychological therapy rather than drugs, our null hypothesis now corresponds to the equality of the column means:\n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2,\n \"Null hypothesis $H_0$:\", \"column means are the same, i.e., $\\\\mu_{.1} = \\\\mu_{.2}$\",\n \"Alternative hypothesis $H_1$:\", \"column means are different, i.e., $\\\\mu_{.1} \\\\neq \\\\mu_{.2}\"\n ), col.names = c(\"\", \"\"))\n```\n\nNull hypothesis \\(H_0\\): | column means are the same, i.e., \\(\\mu_{.1} = \\mu_{.2}\\) |\n | --- | --- |\nAlternative hypothesis \\(H_1\\): | column means are different, i.e., ${.1} {.2} |\n\n### 16.1.2 Running the analysis in R\n\nThe null and alternative hypotheses that I described in the last section should seem awfully familiar: they’re basically the same as the hypotheses that we were testing in our simpler one-way ANOVAs in Chapter 14. So you’re probably expecting that the hypothesis tests that are used in factorial ANOVA will be essentially the same as the \\(F\\)-test from Chapter 14. You’re expecting to see references to sums of squares (SS), mean squares (MS), degrees of freedom (df), and finally an \\(F\\)-statistic that we can convert into a \\(p\\)-value, right? Well, you’re absolutely and completely right. So much so that I’m going to depart from my usual approach. Throughout this book, I’ve generally taken the approach of describing the logic (and to an extent the mathematics) that underpins a particular analysis first; and only then introducing the R commands that you’d use to produce the analysis. This time I’m going to do it the other way around, and show you the R commands first. The reason for doing this is that I want to highlight the similarities between the simple one-way ANOVA tool that we discussed in Chapter 14, and the more complicated tools that we’re going to use in this chapter.\n If the data you’re trying to analyse correspond to a balanced factorial design, then running your analysis of variance is easy. To see how easy it is, let’s start by reproducing the original analysis from Chapter 14. In case you’ve forgotten, for that analysis we were using only a single factor (i.e., `drug` ) to predict our outcome variable (i.e., `mood.gain` ), and so this was what we did: \n\n```\nmodel.1 <- aov( mood.gain ~ drug, clin.trial ) \nsummary( model.1 )\n```\n\n Note that this time around I’ve used the name `model.1` as the label for my `aov` object, since I’m planning on creating quite a few other models too. To start with, suppose I’m also curious to find out if `therapy` has a relationship to `mood.gain` . In light of what we’ve seen from our discussion of multiple regression in Chapter 15, you probably won’t be surprised that all we have to do is extend the formula: in other words, if we specify \n\nas our model, we’ll probably get what we’re after: \n\n This output is pretty simple to read too: the first row of the table reports a between-group sum of squares (SS) value associated with the `drug` factor, along with a corresponding between-group \\(df\\) value. It also calculates a mean square value (MS), and \\(F\\)-statistic and a \\(p\\)-value. There is also a row corresponding to the `therapy` factor, and a row corresponding to the residuals (i.e., the within groups variation). Not only are all of the individual quantities pretty familiar, the relationships between these different quantities has remained unchanged: just like we saw with the original one-way ANOVA, note that the mean square value is calculated by dividing SS by the corresponding \\(df\\). That is, it’s still true that \\[ \\mbox{MS} = \\frac{\\mbox{SS}}{df} \\] regardless of whether we’re talking about `drug` , `therapy` or the residuals. To see this, let’s not worry about how the sums of squares values are calculated: instead, let’s take it on faith that R has calculated the SS values correctly, and try to verify that all the rest of the numbers make sense. First, note that for the `drug` factor, we divide \\(3.45\\) by \\(2\\), and end up with a mean square value of \\(1.73\\). For the `therapy` factor, there’s only 1 degree of freedom, so our calculations are even simpler: dividing \\(0.47\\) (the SS value) by 1 gives us an answer of \\(0.47\\) (the MS value). Turning to the \\(F\\) statistics and the \\(p\\) values, notice that we have two of each: one corresponding to the `drug` factor and the other corresponding to the `therapy` factor. Regardless of which one we’re talking about, the \\(F\\) statistic is calculated by dividing the mean square value associated with the factor by the mean square value associated with the residuals. If we use “A” as shorthand notation to refer to the first factor (factor A; in this case `drug` ) and “R” as shorthand notation to refer to the residuals, then the \\(F\\) statistic associated with factor A is denoted \\(F_A\\), and is calculated as follows: \\[\nF_{A} = \\frac{\\mbox{MS}_{A}}{\\mbox{MS}_{R}}\n\\] and an equivalent formula exists for factor B (i.e., `therapy` ). Note that this use of “R” to refer to residuals is a bit awkward, since we also used the letter R to refer to the number of rows in the table, but I’m only going to use “R” to mean residuals in the context of SS\\(_R\\) and MS\\(_R\\), so hopefully this shouldn’t be confusing. Anyway, to apply this formula to the `drugs` factor, we take the mean square of \\(1.73\\) and divide it by the residual mean square value of \\(0.07\\), which gives us an \\(F\\)-statistic of \\(26.15\\). The corresponding calculation for the `therapy` variable would be to divide \\(0.47\\) by \\(0.07\\) which gives \\(7.08\\) as the \\(F\\)-statistic. Not surprisingly, of course, these are the same values that R has reported in the ANOVA table above. The last part of the ANOVA table is the calculation of the \\(p\\) values. Once again, there is nothing new here: for each of our two factors, what we’re trying to do is test the null hypothesis that there is no relationship between the factor and the outcome variable (I’ll be a bit more precise about this later on). To that end, we’ve (apparently) followed a similar strategy that we did in the one way ANOVA, and have calculated an \\(F\\)-statistic for each of these hypotheses. To convert these to \\(p\\) values, all we need to do is note that the that the sampling distribution for the \\(F\\) statistic under the null hypothesis (that the factor in question is irrelevant) is an \\(F\\) distribution: and that two degrees of freedom values are those corresponding to the factor, and those corresponding to the residuals. For the `drug` factor we’re talking about an \\(F\\) distribution with 2 and 14 degrees of freedom (I’ll discuss degrees of freedom in more detail later). In contrast, for the `therapy` factor sampling distribution is \\(F\\) with 1 and 14 degrees of freedom. If we really wanted to, we could calculate the \\(p\\) value ourselves using the `pf()` function (see Section 9.6). Just to prove that there’s nothing funny going on, here’s what that would look like for the `drug` variable: \n\n```\npf( q=26.15, df1=2, df2=14, lower.tail=FALSE )\n```\n\n `## [1] 1.871981e-05` \nAnd as you can see, this is indeed the \\(p\\) value reported in the ANOVA table above.\n At this point, I hope you can see that the ANOVA table for this more complicated analysis corresponding to `model.2` should be read in much the same way as the ANOVA table for the simpler analysis for `model.1` . In short, it’s telling us that the factorial ANOVA for our \\(3 \\times 2\\) design found a significant effect of drug (\\(F_{2,14} = 26.15, p < .001\\)) as well as a significant effect of therapy (\\(F_{1,14} = 7.08, p = .02\\)). Or, to use the more technically correct terminology, we would say that there are two main effects of drug and therapy. At the moment, it probably seems a bit redundant to refer to these as “main” effects: but it actually does make sense. Later on, we’re going to want to talk about the possibility of “interactions” between the two factors, and so we generally make a distinction between main effects and interaction effects. \n\n### 16.1.3 How are the sum of squares calculated?\n\n In the previous section I had two goals: firstly, to show you that the R commands needed to do factorial ANOVA are pretty much the same ones that we used for a one way ANOVA. The only difference is the `formula` argument to the `aov()` function. Secondly, I wanted to show you what the ANOVA table looks like in this case, so that you can see from the outset that the basic logic and structure behind factorial ANOVA is the same as that which underpins one way ANOVA. Try to hold onto that feeling. It’s genuinely true, insofar as factorial ANOVA is built in more or less the same way as the simpler one-way ANOVA model. It’s just that this feeling of familiarity starts to evaporate once you start digging into the details. Traditionally, this comforting sensation is replaced by an urge to murder the the authors of statistics textbooks. \nOkay, let’s start looking at some of those details. The explanation that I gave in the last section illustrates the fact that the hypothesis tests for the main effects (of drug and therapy in this case) are \\(F\\)-tests, but what it doesn’t do is show you how the sum of squares (SS) values are calculated. Nor does it tell you explicitly how to calculate degrees of freedom (\\(df\\) values) though that’s a simple thing by comparison. Let’s assume for now that we have only two predictor variables, Factor A and Factor B. If we use \\(Y\\) to refer to the outcome variable, then we would use \\(Y_{rci}\\) to refer to the outcome associated with the \\(i\\)-th member of group \\(rc\\) (i.e., level/row \\(r\\) for Factor A and level/column \\(c\\) for Factor B). Thus, if we use \\(\\bar{Y}\\) to refer to a sample mean, we can use the same notation as before to refer to group means, marginal means and grand means: that is, \\(\\bar{Y}_{rc}\\) is the sample mean associated with the \\(r\\)th level of Factor A and the \\(c\\)th level of Factor B, \\(\\bar{Y}_{r.}\\) would be the marginal mean for the \\(r\\)th level of Factor A, \\(\\bar{Y}_{.c}\\) would be the marginal mean for the \\(c\\)th level of Factor B, and \\(\\bar{Y}_{..}\\) is the grand mean. In other words, our sample means can be organised into the same table as the population means. For our clinical trial data, that table looks like this:\n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2, ~V3, ~V4,\n \"placebo\", \"$\\\\bar{Y}_{11}$\", \"$\\\\bar{Y}_{12}$\", \"$\\\\bar{Y}_{1.}$\",\n \"anxifree\", \"$\\\\bar{Y}_{21}$\", \"$\\\\bar{Y}_{22}$\", \"$\\\\bar{Y}_{2.}$\",\n \"joyzepam\", \"$\\\\bar{Y}_{31}$\", \"$\\\\bar{Y}_{32}$\", \"$\\\\bar{Y}_{3.}$\",\n \"total\", \"$\\\\bar{Y}_{.1}$\", \"$\\\\bar{Y}_{.2}$\", \"$\\\\bar{Y}_{..}$\"\n ), col.names = c(\"\", \"no therapy\", \"CBT\", \"total\"))\n```\n\nno therapy | CBT | total |\n | --- | --- | --- |\nplacebo | \\(\\bar{Y}_{11}\\) | \\(\\bar{Y}_{12}\\) | \\(\\bar{Y}_{1.}\\) |\nanxifree | \\(\\bar{Y}_{21}\\) | \\(\\bar{Y}_{22}\\) | \\(\\bar{Y}_{2.}\\) |\njoyzepam | \\(\\bar{Y}_{31}\\) | \\(\\bar{Y}_{32}\\) | \\(\\bar{Y}_{3.}\\) |\ntotal | \\(\\bar{Y}_{.1}\\) | \\(\\bar{Y}_{.2}\\) | \\(\\bar{Y}_{..}\\) |\n\n And if we look at the sample means that I showed earlier, we have \\(\\bar{Y}_{11} = 0.30\\), \\(\\bar{Y}_{12} = 0.60\\) etc. In our clinical trial example, the `drugs` factor has 3 levels and the `therapy` factor has 2 levels, and so what we’re trying to run is a \\(3 \\times 2\\) factorial ANOVA. However, we’ll be a little more general and say that Factor A (the row factor) has \\(R\\) levels and Factor B (the column factor) has \\(C\\) levels, and so what we’re runnning here is an \\(R \\times C\\) factorial ANOVA. Now that we’ve got our notation straight, we can compute the sum of squares values for each of the two factors in a relatively familiar way. For Factor A, our between group sum of squares is calculated by assessing the extent to which the (row) marginal means \\(\\bar{Y}_{1.}\\), \\(\\bar{Y}_{2.}\\) etc, are different from the grand mean \\(\\bar{Y}_{..}\\). We do this in the same way that we did for one-way ANOVA: calculate the sum of squared difference between the \\(\\bar{Y}_{i.}\\) values and the \\(\\bar{Y}_{..}\\) values. Specifically, if there are \\(N\\) people in each group, then we calculate this: \\[ \\mbox{SS}_{A} = (N \\times C) \\sum_{r=1}^R \\left( \\bar{Y}_{r.} - \\bar{Y}_{..} \\right)^2 \\] As with one-way ANOVA, the most interesting234 part of this formula is the \\(\\left( \\bar{Y}_{r.} - \\bar{Y}_{..} \\right)^2\\) bit, which corresponds to the squared deviation associated with level \\(r\\). All that this formula does is calculate this squared deviation for all \\(R\\) levels of the factor, add them up, and then multiply the result by \\(N \\times C\\). The reason for this last part is that there are multiple cells in our design that have level \\(r\\) on Factor A: in fact, there are \\(C\\) of them, one corresponding to each possible level of Factor B! For instance, in our toy example, there are two different cells in the design corresponding to the `anxifree` drug: one for people with `no.therapy` , and one for the `CBT` group. Not only that, within each of these cells there are \\(N\\) observations. So, if we want to convert our SS value into a quantity that calculates the between-groups sum of squares on a “per observation” basis, we have to multiply by by \\(N \\times C\\). The formula for factor B is of course the same thing, just with some subscripts shuffled around: \\[\n\\mbox{SS}_{B} = (N \\times R) \\sum_{c=1}^C \\left( \\bar{Y}_{.c} - \\bar{Y}_{..} \\right)^2\n\\] Now that we have these formulas, we can check them against the R output from the earlier section. First, notice that we calculated all the marginal means (i.e., row marginal means \\(\\bar{Y}_{r.}\\) and column marginal means \\(\\bar{Y}_{.c}\\)) earlier using `aggregate()` , and we also calculated the grand mean. Let’s repeat those calculations, but this time we’ll save the results to varibles so that we can use them in subsequent calculations: \n\n```\ndrug.means <- aggregate( mood.gain ~ drug, clin.trial, mean )[,2]\ntherapy.means <- aggregate( mood.gain ~ therapy, clin.trial, mean )[,2]\ngrand.mean <- mean( clin.trial$mood.gain )\n```\n\n Okay, now let’s calculate the sum of squares associated with the main effect of `drug` . There are a total of \\(N=3\\) people in each group, and \\(C=2\\) different types of therapy. Or, to put it another way, there are \\(3 \\times 2 = 6\\) people who received any particular drug. So our calculations are: \n\n```\nSS.drug <- (3*2) * sum( (drug.means - grand.mean)^2 )\nSS.drug\n```\n\n `## [1] 3.453333` \nNot surprisingly, this is the same number that you get when you look up the SS value for the drugs factor in the ANOVA table that I presented earlier. We can repeat the same kind of calculation for the effect of therapy. Again there are \\(N=3\\) people in each group, but since there are \\(R=3\\) different drugs, this time around we note that there are \\(3 \\times 3 = 9\\) people who received CBT, and an additional 9 people who received the placebo. So our calculation is now:\n\n```\nSS.therapy <- (3*3) * sum( (therapy.means - grand.mean)^2 )\nSS.therapy\n```\n\n `## [1] 0.4672222` \nand we are, once again, unsurprised to see that our calculations are identical to the ANOVA output.\n\nSo that’s how you calculate the SS values for the two main effects. These SS values are analogous to the between-group sum of squares values that we calculated when doing one-way ANOVA in Chapter 14. However, it’s not a good idea to think of them as between-groups SS values anymore, just because we have two different grouping variables and it’s easy to get confused. In order to construct an \\(F\\) test, however, we also need to calculate the within-groups sum of squares. In keeping with the terminology that we used in the regression chapter (Chapter 15) and the terminology that R uses when printing out the ANOVA table, I’ll start referring to the within-groups SS value as the residual sum of squares SS\\(_R\\).\n\nThe easiest way to think about the residual SS values in this context, I think, is to think of it as the leftover variation in the outcome variable after you take into account the differences in the marginal means (i.e., after you remove SS\\(_A\\) and SS\\(_B\\)). What I mean by that is we can start by calculating the total sum of squares, which I’ll label SS\\(_T\\). The formula for this is pretty much the same as it was for one-way ANOVA: we take the difference between each observation \\(Y_{rci}\\) and the grand mean \\(\\bar{Y}_{..}\\), square the differences, and add them all up \\[ \\mbox{SS}_T = \\sum_{r=1}^R \\sum_{c=1}^C \\sum_{i=1}^N \\left( Y_{rci} - \\bar{Y}_{..}\\right)^2 \\] The “triple summation” here looks more complicated than it is. In the first two summations, we’re summing across all levels of Factor A (i.e., over all possible rows \\(r\\) in our table), across all levels of Factor B (i.e., all possible columns \\(c\\)). Each \\(rc\\) combination corresponds to a single group, and each group contains \\(N\\) people: so we have to sum across all those people (i.e., all \\(i\\) values) too. In other words, all we’re doing here is summing across all observations in the data set (i.e., all possible \\(rci\\) combinations).\n\nAt this point, we know the total variability of the outcome variable SS\\(_T\\), and we know how much of that variability can be attributed to Factor A (SS\\(_A\\)) and how much of it can be attributed to Factor B (SS\\(_B\\)). The residual sum of squares is thus defined to be the variability in \\(Y\\) that can’t be attributed to either of our two factors. In other words: \\[ \\mbox{SS}_R = \\mbox{SS}_T - (\\mbox{SS}_A + \\mbox{SS}_B) \\] Of course, there is a formula that you can use to calculate the residual SS directly, but I think that it makes more conceptual sense to think of it like this. The whole point of calling it a residual is that it’s the leftover variation, and the formula above makes that clear. I should also note that, in keeping with the terminology used in the regression chapter, it is commonplace to refer to \\(\\mbox{SS}_A + \\mbox{SS}_B\\) as the variance attributable to the “ANOVA model”, denoted SS\\(_M\\), and so we often say that the total sum of squares is equal to the model sum of squares plus the residual sum of squares. Later on in this chapter we’ll see that this isn’t just a surface similarity: ANOVA and regression are actually the same thing under the hood.\n\nIn any case, it’s probably worth taking a moment to check that we can calculate SS\\(_R\\) using this formula, and verify that we do obtain the same answer that R produces in its ANOVA table. The calculations are pretty straightforward. First we calculate the total sum of squares:\n\n```\nSS.tot <- sum( (clin.trial$mood.gain - grand.mean)^2 )\nSS.tot\n```\n\n `## [1] 4.845` \nand then we use it to calculate the residual sum of squares:\n\n```\nSS.res <- SS.tot - (SS.drug + SS.therapy)\nSS.res\n```\n\n `## [1] 0.9244444` \nYet again, we get the same answer.\n\n### 16.1.4 What are our degrees of freedom?\n\n The degrees of freedom are calculated in much the same way as for one-way ANOVA. For any given factor, the degrees of freedom is equal to the number of levels minus 1 (i.e., \\(R-1\\) for the row variable, Factor A, and \\(C-1\\) for the column variable, Factor B). So, for the `drugs` factor we obtain \\(df = 2\\), and for the `therapy` factor we obtain \\(df=1\\). Later on on, when we discuss the interpretation of ANOVA as a regression model (see Section 16.6) I’ll give a clearer statement of how we arrive at this number, but for the moment we can use the simple definition of degrees of freedom, namely that the degrees of freedom equals the number of quantities that are observed, minus the number of constraints. So, for the `drugs` factor, we observe 3 separate group means, but these are constrained by 1 grand mean; and therefore the degrees of freedom is 2. For the residuals, the logic is similar, but not quite the same. The total number of observations in our experiment is 18. The constraints correspond to the 1 grand mean, the 2 additional group means that the `drug` factor introduces, and the 1 additional group mean that the the `therapy` factor introduces, and so our degrees of freedom is 14. As a formula, this is \\(N-1 -(R-1)-(C-1)\\), which simplifies to \\(N-R-C+1\\). \n\n### 16.1.5 Factorial ANOVA versus one-way ANOVAs\n\n Now that we’ve seen how a factorial ANOVA works, it’s worth taking a moment to compare it to the results of the one way analyses, because this will give us a really good sense of why it’s a good idea to run the factorial ANOVA. In Chapter 14 that I ran a one-way ANOVA that looked to see if there are any differences between drugs, and a second one-way ANOVA to see if there were any differences between therapies. As we saw in Section 16.1.1, the null and alternative hypotheses tested by the one-way ANOVAs are in fact identical to the hypotheses tested by the factorial ANOVA. Looking even more carefully at the ANOVA tables, we can see that the sum of squares associated with the factors are identical in the two different analyses (3.45 for `drug` and 0.92 for `therapy` ), as are the degrees of freedom (2 for `drug` , 1 for `therapy` ). But they don’t give the same answers! Most notably, when we ran the one-way ANOVA for `therapy` in Section 14.11 we didn’t find a significant effect (the \\(p\\)-value was 0.21). However, when we look at the main effect of `therapy` within the context of the two-way ANOVA, we do get a significant effect (\\(p=.019\\)). The two analyses are clearly not the same. Why does that happen? The answer lies in understanding how the residuals are calculated. Recall that the whole idea behind an \\(F\\)-test is to compare the variability that can be attributed to a particular factor with the variability that cannot be accounted for (the residuals). If you run a one-way ANOVA for `therapy` , and therefore ignore the effect of `drug` , the ANOVA will end up dumping all of the drug-induced variability into the residuals! This has the effect of making the data look more noisy than they really are, and the effect of `therapy` which is correctly found to be significant in the two-way ANOVA now becomes non-significant. If we ignore something that actually matters (e.g., `drug` ) when trying to assess the contribution of something else (e.g., `therapy` ) then our analysis will be distorted. Of course, it’s perfectly okay to ignore variables that are genuinely irrelevant to the phenomenon of interest: if we had recorded the colour of the walls, and that turned out to be non-significant in a three-way ANOVA (i.e. \n\n), it would be perfectly okay to disregard it and just report the simpler two-way ANOVA that doesn’t include this irrelevant factor. What you shouldn’t do is drop variables that actually make a difference! \n\n### 16.1.6 What kinds of outcomes does this analysis capture?\n\nThe ANOVA model that we’ve been talking about so far covers a range of different patterns that we might observe in our data. For instance, in a two-way ANOVA design, there are four possibilities: (a) only Factor A matters, (b) only Factor B matters, (c) both A and B matter, and (d) neither A nor B matters. An example of each of these four possibilities is plotted in Figure ??.\n\n## 16.2 Factorial ANOVA 2: balanced designs, interactions allowed\n\n The four patterns of data shown in Figure ?? are all quite realistic: there are a great many data sets that produce exactly those patterns. However, they are not the whole story, and the ANOVA model that we have been talking about up to this point is not sufficient to fully account for a table of group means. Why not? Well, so far we have the ability to talk about the idea that drugs can influence mood, and therapy can influence mood, but no way of talking about the possibility of an interaction between the two. An interaction between A and B is said to occur whenever the effect of Factor A is different, depending on which level of Factor B we’re talking about. Several examples of an interaction effect with the context of a 2 x 2 ANOVA are shown in Figure ??. To give a more concrete example, suppose that the operation of Anxifree and Joyzepam is governed quite different physiological mechanisms, and one consequence of this is that while Joyzepam has more or less the same effect on mood regardless of whether one is in therapy, Anxifree is actually much more effective when administered in conjunction with CBT. The ANOVA that we developed in the previous section does not capture this idea. To get some idea of whether an interaction is actually happening here, it helps to plot the various group means. There are quite a few different ways draw these plots in R. One easy way is to use the `interaction.plot()` function, but this function won’t draw error bars for you. A fairly simple function that will include error bars for you is the `lineplot.CI()` function in the `sciplots` package (see Section 10.5.4). The command \n\n```\nlibrary(sciplot)\nlibrary(lsr)\n lineplot.CI( x.factor = clin.trial$drug, \n response = clin.trial$mood.gain,\n group = clin.trial$therapy,\n ci.fun = ciMean,\n xlab = \"drug\",\n ylab = \"mood gain\" )\n```\n\n produces the output is shown in Figure 16.9 (don’t forget that the `ciMean` function is in the `lsr` package, so you need to have `lsr` loaded!). Our main concern relates to the fact that the two lines aren’t parallel. The effect of CBT (difference between solid line and dotted line) when the drug is Joyzepam (right side) appears to be near zero, even smaller than the effect of CBT when a placebo is used (left side). However, when Anxifree is administered, the effect of CBT is larger than the placebo (middle). Is this effect real, or is this just random variation due to chance? Our original ANOVA cannot answer this question, because we make no allowances for the idea that interactions even exist! In this section, we’ll fix this problem. \n\n### 16.2.1 What exactly is an interaction effect?\n\nThe key idea that we’re going to introduce in this section is that of an interaction effect. What that means for our R formulas is that we’ll write down models like so although there are only two factors involved in our model (i.e., `drug` and `therapy` ), there are three distinct terms (i.e., `drug` , `therapy` and `drug:therapy` ). That is, in addition to the main effects of `drug` and `therapy` , we have a new component to the model, which is our interaction term `drug:therapy` . Intuitively, the idea behind an interaction effect is fairly simple: it just means that the effect of Factor A is different, depending on which level of Factor B we’re talking about. But what does that actually mean in terms of our data? Figure ?? depicts several different patterns that, although quite different to each other, would all count as an interaction effect. So it’s not entirely straightforward to translate this qualitative idea into something mathematical that a statistician can work with. As a consequence, the way that the idea of an interaction effect is formalised in terms of null and alternative hypotheses is slightly difficult, and I’m guessing that a lot of readers of this book probably won’t be all that interested. Even so, I’ll try to give the basic idea here. To start with, we need to be a little more explicit about our main effects. Consider the main effect of Factor A ( `drug` in our running example). We originally formulated this in terms of the null hypothesis that the two marginal means \\(\\mu_{r.}\\) are all equal to each other. Obviously, if all of these are equal to each other, then they must also be equal to the grand mean \\(\\mu_{..}\\) as well, right? So what we can do is define the effect of Factor A at level \\(r\\) to be equal to the difference between the marginal mean \\(\\mu_{r.}\\) and the grand mean \\(\\mu_{..}\\). Let’s denote this effect by \\(\\alpha_r\\), and note that \\[ \\alpha_r = \\mu_{r.} - \\mu_{..} \\] Now, by definition all of the \\(\\alpha_r\\) values must sum to zero, for the same reason that the average of the marginal means \\(\\mu_{r.}\\) must be the grand mean \\(\\mu_{..}\\). We can similarly define the effect of Factor B at level \\(i\\) to be the difference between the column marginal mean \\(\\mu_{.c}\\) and the grand mean \\(\\mu_{..}\\) \\[ \\beta_c = \\mu_{.c} - \\mu_{..} \\] and once again, these \\(\\beta_c\\) values must sum to zero. The reason that statisticians sometimes like to talk about the main effects in terms of these \\(\\alpha_r\\) and \\(\\beta_c\\) values is that it allows them to be precise about what it means to say that there is no interaction effect. If there is no interaction at all, then these \\(\\alpha_r\\) and \\(\\beta_c\\) values will perfectly describe the group means \\(\\mu_{rc}\\). Specifically, it means that \\[ \\mu_{rc} = \\mu_{..} + \\alpha_r + \\beta_c \\] That is, there’s nothing special about the group means that you couldn’t predict perfectly by knowing all the marginal means. And that’s our null hypothesis, right there. The alternative hypothesis is that \\[ \\mu_{rc} \\neq \\mu_{..} + \\alpha_r + \\beta_c \\] for at least one group \\(rc\\) in our table. However, statisticians often like to write this slightly differently. They’ll usually define the specific interaction associated with group \\(rc\\) to be some number, awkwardly referred to as \\((\\alpha\\beta)_{rc}\\), and then they will say that the alternative hypothesis is that \\[\\mu_{rc} = \\mu_{..} + \\alpha_r + \\beta_c + (\\alpha\\beta)_{rc}\\] where \\((\\alpha\\beta)_{rc}\\) is non-zero for at least one group. This notation is kind of ugly to look at, but it is handy as we’ll see in the next section when discussing how to calculate the sum of squares. \n\n### 16.2.2 Calculating sums of squares for the interaction\n\nHow should we calculate the sum of squares for the interaction terms, SS\\(_{A:B}\\)? Well, first off, it helps to notice how the previous section defined the interaction effect in terms of the extent to which the actual group means differ from what you’d expect by just looking at the marginal means. Of course, all of those formulas refer to population parameters rather than sample statistics, so we don’t actually know what they are. However, we can estimate them by using sample means in place of population means. So for Factor A, a good way to estimate the main effect at level \\(r\\) as the difference between the sample marginal mean \\(\\bar{Y}_{rc}\\) and the sample grand mean \\(\\bar{Y}_{..}\\). That is, we would use this as our estimate of the effect: \\[ \\hat{\\alpha}_r = \\bar{Y}_{r.} - \\bar{Y}_{..} \\] Similarly, our estimate of the main effect of Factor B at level \\(c\\) can be defined as follows: \\[ \\hat{\\beta}_c = \\bar{Y}_{.c} - \\bar{Y}_{..} \\] Now, if you go back to the formulas that I used to describe the SS values for the two main effects, you’ll notice that these effect terms are exactly the quantities that we were squaring and summing! So what’s the analog of this for interaction terms? The answer to this can be found by first rearranging the formula for the group means \\(\\mu_{rc}\\) under the alternative hypothesis, so that we get this: \\[\\begin{eqnarray*} (\\alpha \\beta)_{rc} &=& \\mu_{rc} - \\mu_{..} - \\alpha_r - \\beta_c \\\\ &=& \\mu_{rc} - \\mu_{..} - (\\mu_{r.} - \\mu_{..}) - (\\mu_{.c} - \\mu_{..}) \\\\ &=& \\mu_{rc} - \\mu_{r.} - \\mu_{.c} + \\mu_{..} \\end{eqnarray*}\\] \nSo, once again, if we substitute our sample statistics in place of the population means, we get the following as our estimate of the interaction effect for group \\(rc\\), which is \\[ \\hat{(\\alpha\\beta)}_{rc} = \\bar{Y}_{rc} - \\bar{Y}_{r.} - \\bar{Y}_{.c} + \\bar{Y}_{..} \\] Now all we have to do is sum all of these estimates across all \\(R\\) levels of Factor A and all \\(C\\) levels of Factor B, and we obtain the following formula for the sum of squares associated with the interaction as a whole: \\[ \\mbox{SS}_{A:B} = N \\sum_{r=1}^R \\sum_{c=1}^C \\left( \\bar{Y}_{rc} - \\bar{Y}_{r.} - \\bar{Y}_{.c} + \\bar{Y}_{..} \\right)^2 \\] where, we multiply by \\(N\\) because there are \\(N\\) observations in each of the groups, and we want our SS values to reflect the variation among observations accounted for by the interaction, not the variation among groups.\n\nNow that we have a formula for calculating SS\\(_{A:B}\\), it’s important to recognise that the interaction term is part of the model (of course), so the total sum of squares associated with the model, SS\\(_M\\) is now equal to the sum of the three relevant SS values, \\(\\mbox{SS}_A + \\mbox{SS}_B + \\mbox{SS}_{A:B}\\). The residual sum of squares \\(\\mbox{SS}_R\\) is still defined as the leftover variation, namely \\(\\mbox{SS}_T - \\mbox{SS}_M\\), but now that we have the interaction term this becomes \\[ \\mbox{SS}_R = \\mbox{SS}_T - (\\mbox{SS}_A + \\mbox{SS}_B + \\mbox{SS}_{A:B}) \\] As a consequence, the residual sum of squares SS\\(_R\\) will be smaller than in our original ANOVA that didn’t include interactions.\n\n### 16.2.3 Degrees of freedom for the interaction\n\nCalculating the degrees of freedom for the interaction is, once again, slightly trickier than the corresponding calculation for the main effects. To start with, let’s think about the ANOVA model as a whole. Once we include interaction effects in the model, we’re allowing every single group has a unique mean, \\(\\mu_{rc}\\). For an \\(R \\times C\\) factorial ANOVA, this means that there are \\(R \\times C\\) quantities of interest in the model, and only the one constraint: all of the group means need to average out to the grand mean. So the model as a whole needs to have \\((R\\times C) - 1\\) degrees of freedom. But the main effect of Factor A has \\(R-1\\) degrees of freedom, and the main effect of Factor B has \\(C-1\\) degrees of freedom. Which means that the degrees of freedom associated with the interaction is \\[\\begin{eqnarray*} df_{A:B} &=& (R\\times C - 1) - (R - 1) - (C -1 ) \\\\ &=& RC - R - C + 1 \\\\ &=& (R-1)(C-1) \\end{eqnarray*}\\] \nwhich is just the product of the degrees of freedom associated with the row factor and the column factor.\n\nWhat about the residual degrees of freedom? Because we’ve added interaction terms, which absorb some degrees of freedom, there are fewer residual degrees of freedom left over. Specifically, note that if the model with interaction has a total of \\((R\\times C) - 1\\), and there are \\(N\\) observations in your data set that are constrained to satisfy 1 grand mean, your residual degrees of freedom now become \\(N-(R \\times C)-1+1\\), or just \\(N-(R \\times C)\\).\n\n### 16.2.4 Running the ANOVA in R\n\n Adding interaction terms to the ANOVA model in R is straightforward. Returning to our running example of the clinical trial, in addition to the main effect terms of `drug` and `therapy` , we include the interaction term `drug:therapy` . So the R command to create the ANOVA model now looks like this: \n\n```\nmodel.3 <- aov( mood.gain ~ drug + therapy + drug:therapy, clin.trial )\n```\n\n However, R allows a convenient shorthand. Instead of typing out all three terms, you can shorten the right hand side of the formula to `drug*therapy` . The `*` operator inside the formula is taken to indicate that you want both main effects and the interaction. So we can also run our ANOVA like this, and get the same answer: \n\n```\nmodel.3 <- aov( mood.gain ~ drug * therapy, clin.trial )\n summary( model.3 )\n```\n\n```\n## Df Sum Sq Mean Sq F value Pr(>F) \n## drug 2 3.453 1.7267 31.714 1.62e-05 ***\n## therapy 1 0.467 0.4672 8.582 0.0126 * \n## drug:therapy 2 0.271 0.1356 2.490 0.1246 \n## Residuals 12 0.653 0.0544 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nAs it turns out, while we do have a significant main effect of drug (\\(F_{2,12} = 31.7, p <.001\\)) and therapy type (\\(F_{1,12} = 8.6, p=.013\\)), there is no significant interaction between the two (\\(F_{2,12} = 2.5, p = 0.125\\)).\n\n### 16.2.5 Interpreting the results\n\n There’s a couple of very important things to consider when interpreting the results of factorial ANOVA. Firstly, there’s the same issue that we had with one-way ANOVA, which is that if you obtain a significant main effect of (say) `drug` , it doesn’t tell you anything about which drugs are different to one another. To find that out, you need to run additional analyses. We’ll talk about some analyses that you can run in Sections 16.7 and ??. The same is true for interaction effects: knowing that there’s a significant interaction doesn’t tell you anything about what kind of interaction exists. Again, you’ll need to run additional analyses. \nSecondly, there’s a very peculiar interpretation issue that arises when you obtain a significant interaction effect but no corresponding main effect. This happens sometimes. For instance, in the crossover interaction shown in Figure ??, this is exactly what you’d find: in this case, neither of the main effects would be significant, but the interaction effect would be. This is a difficult situation to interpret, and people often get a bit confused about it. The general advice that statisticians like to give in this situation is that you shouldn’t pay much attention to the main effects when an interaction is present. The reason they say this is that, although the tests of the main effects are perfectly valid from a mathematical point of view, when there is a significant interaction effect the main effects rarely test interesting hypotheses. Recall from Section 16.1.1 that the null hypothesis for a main effect is that the marginal means are equal to each other, and that a marginal mean is formed by averaging across several different groups. But if you have a significant interaction effect, then you know that the groups that comprise the marginal mean aren’t homogeneous, so it’s not really obvious why you would even care about those marginal means.\n\nHere’s what I mean. Again, let’s stick with a clinical example. Suppose that we had a \\(2 \\times 2\\) design comparing two different treatments for phobias (e.g., systematic desensitisation vs flooding), and two different anxiety reducing drugs (e.g., Anxifree vs Joyzepam). Now suppose what we found was that Anxifree had no effect when desensitisation was the treatment, and Joyzepam had no effect when flooding was the treatment. But both were pretty effective for the other treatment. This is a classic crossover interaction, and what we’d find when running the ANOVA is that there is no main effect of drug, but a significant interaction. Now, what does it actually mean to say that there’s no main effect? Wel, it means that, if we average over the two different psychological treatments, then the average effect of Anxifree and Joyzepam is the same. But why would anyone care about that? When treating someone for phobias, it is never the case that a person can be treated using an “average” of flooding and desensitisation: that doesn’t make a lot of sense. You either get one or the other. For one treatment, one drug is effective; and for the other treatment, the other drug is effective. The interaction is the important thing; the main effect is kind of irrelevant.\n\nThis sort of thing happens a lot: the main effect are tests of marginal means, and when an interaction is present we often find ourselves not being terribly interested in marginal means, because they imply averaging over things that the interaction tells us shouldn’t be averaged! Of course, it’s not always the case that a main effect is meaningless when an interaction is present. Often you can get a big main effect and a very small interaction, in which case you can still say things like “drug A is generally more effective than drug B” (because there was a big effect of drug), but you’d need to modify it a bit by adding that “the difference in effectiveness was different for different psychological treatments”. In any case, the main point here is that whenever you get a significant interaction you should stop and think about what the main effect actually means in this context. Don’t automatically assume that the main effect is interesting.\n\n## 16.3 Effect size, estimated means, and confidence intervals\n\nIn this section I’ll discuss a few additional quantities that you might find yourself wanting to calculate for a factorial ANOVA. The main thing you will probably want to calculate is the effect size for each term in your model, but you may also want to R to give you some estimates for the group means and associated confidence intervals.\n\n### 16.3.1 Effect sizes\n\nThe effect size calculations for a factorial ANOVA is pretty similar to those used in one way ANOVA (see Section 14.4). Specifically, we can use \\(\\eta^2\\) (eta-squared) as simple way to measure how big the overall effect is for any particular term. As before, \\(\\eta^2\\) is defined by dividing the sum of squares associated with that term by the total sum of squares. For instance, to determine the size of the main effect of Factor A, we would use the following formula \\[ \\eta_A^2 = \\frac{\\mbox{SS}_{A}}{\\mbox{SS}_{T}} \\] As before, this can be interpreted in much the same way as \\(R^2\\) in regression.235 It tells you the proportion of variance in the outcome variable that can be accounted for by the main effect of Factor A. It is therefore a number that ranges from 0 (no effect at all) to 1 (accounts for all of the variability in the outcome). Moreover, the sum of all the \\(\\eta^2\\) values, taken across all the terms in the model, will sum to the the total \\(R^2\\) for the ANOVA model. If, for instance, the ANOVA model fits perfectly (i.e., there is no within-groups variability at all!), the \\(\\eta^2\\) values will sum to 1. Of course, that rarely if ever happens in real life.\n\nHowever, when doing a factorial ANOVA, there is a second measure of effect size that people like to report, known as partial \\(\\eta^2\\). The idea behind partial \\(\\eta^2\\) (which is sometimes denoted \\(_p\\eta^2\\) or \\(\\eta^2_p\\)) is that, when measuring the effect size for a particular term (say, the main effect of Factor A), you want to deliberately ignore the other effects in the model (e.g., the main effect of Factor B). That is, you would pretend that the effect of all these other terms is zero, and then calculate what the \\(\\eta^2\\) value would have been. This is actually pretty easy to calculate. All you have to do is remove the sum of squares associated with the other terms from the denominator. In other words, if you want the partial \\(\\eta^2\\) for the main effect of Factor A, the denominator is just the sum of the SS values for Factor A and the residuals: \\[ \\mbox{partial } \\eta^2_A = \\frac{\\mbox{SS}_{A}}{\\mbox{SS}_{A} + \\mbox{SS}_{R}} \\] This will always give you a larger number than \\(\\eta^2\\), which the cynic in me suspects accounts for the popularity of partial \\(\\eta^2\\). And once again you get a number between 0 and 1, where 0 represents no effect. However, it’s slightly trickier to interpret what a large partial \\(\\eta^2\\) value means. In particular, you can’t actually compare the partial \\(\\eta^2\\) values across terms! Suppose, for instance, there is no within-groups variability at all: if so, SS\\(_R = 0\\). What that means is that every term has a partial \\(\\eta^2\\) value of 1. But that doesn’t mean that all terms in your model are equally important, or indeed that they are equally large. All it mean is that all terms in your model have effect sizes that are large relative to the residual variation. It is not comparable across terms.\n To see what I mean by this, it’s useful to see a concrete example. Once again, we’ll use the `etaSquared()` function from the `lsr` package. As before, we input the `aov` object for which we want the \\(\\eta^2\\) calculations performed, and R outputs a matrix showing the effect sizes for each term in the model. First, let’s have a look at the effect sizes for the original ANOVA without the interaction term: \n\n```\netaSquared( model.2 )\n```\n\n```\n## eta.sq eta.sq.part\n## drug 0.7127623 0.7888325\n## therapy 0.0964339 0.3357285\n```\n\n Looking at the \\(\\eta^2\\) values first, we see that `drug` accounts for 71.3% of the variance (i.e. \\(\\eta^2 = 0.713\\)) in `mood.gain` , whereas `therapy` only accounts for 9.6%. This leaves a total of 19.1% of the variation unaccounted for (i.e., the residuals constitute 19.1% of the variation in the outcome). Overall, this implies that we have a very large effect236 of `drug` and a modest effect of `therapy` . Now let’s look at the partial \\(\\eta^2\\) values. Because the effect of `therapy` isn’t all that large, controlling for it doesn’t make much of a difference, so the partial \\(\\eta^2\\) for `drug` doesn’t increase very much, and we obtain a value of \\(_p\\eta^2 = 0.789\\)). In contrast, because the effect of `drug` was very large, controlling for it makes a big difference, and so when we calculate the partial \\(\\eta^2\\) for `therapy` you can see that it rises to \\(_p\\eta^2 = 0.336\\). The question that we have to ask ourselves is, what does these partial \\(\\eta^2\\) values actually mean? The way I generally interpret the partial \\(\\eta^2\\) for the main effect of Factor A is to interpret it as a statement about a hypothetical experiment in which only Factor A was being varied. So, even though in this experiment we varied both A and B, we can easily imagine an experiment in which only Factor A was varied: the partial \\(\\eta^2\\) statistic tells you how much of the variance in the outcome variable you would expect to see accounted for in that experiment. However, it should be noted that this interpretation – like many things associated with main effects – doesn’t make a lot of sense when there is a large and significant interaction effect. \nSpeaking of interaction effects, here’s what we get when we calculate the effect sizes for the model that includes the interaction term. As you can see, the \\(\\eta^2\\) values for the main effects don’t change, but the partial \\(\\eta^2\\) values do:\n\n```\netaSquared( model.3 )\n```\n\n```\n## eta.sq eta.sq.part\n## drug 0.71276230 0.8409091\n## therapy 0.09643390 0.4169559\n## drug:therapy 0.05595689 0.2932692\n```\n\n### 16.3.2 Estimated group means\n\n In many situations you will find yourself wanting to report estimates of all the group means based on the results of your ANOVA, as well as confidence intervals associated with them. You can use the `effect()` function in the `effects` package to do this (don’t forget to install the package if you don’t have it already!). If the ANOVA that you have run is a saturated model (i.e., contains all possible main effects and all possible interaction effects) then the estimates of the group means are actually identical to the sample means, though the confidence intervals will use a pooled estimate of the standard errors, rather than use a separate one for each group. To illustrate this, let’s apply the `effect()` function to our saturated model (i.e., `model.3` ) for the clinical trial data. The `effect()` function contains two arguments we care about: the `term` argument specifies what terms in the model we want the means to be calculated for, and the `mod` argument specifies the model: \n\n```\nlibrary(effects)\n eff <- effect( term = \"drug*therapy\", mod = model.3 )\n eff\n```\n\n```\n## \n## drug*therapy effect\n## therapy\n## drug no.therapy CBT\n## placebo 0.300000 0.600000\n## anxifree 0.400000 1.033333\n## joyzepam 1.466667 1.500000\n```\n\n Notice that these are actually the same numbers we got when computing the sample means earlier (i.e., the `group.means` variable that we computed using `aggregate()` ). One useful thing that we can do using the effect variable `eff` , however, is extract the confidence intervals using the `summary()` function: `summary(eff)` \n\n```\n## \n## drug*therapy effect\n## therapy\n## drug no.therapy CBT\n## placebo 0.300000 0.600000\n## anxifree 0.400000 1.033333\n## joyzepam 1.466667 1.500000\n## \n## Lower 95 Percent Confidence Limits\n## therapy\n## drug no.therapy CBT\n## placebo 0.006481093 0.3064811\n## anxifree 0.106481093 0.7398144\n## joyzepam 1.173147759 1.2064811\n## \n## Upper 95 Percent Confidence Limits\n## therapy\n## drug no.therapy CBT\n## placebo 0.5935189 0.8935189\n## anxifree 0.6935189 1.3268522\n## joyzepam 1.7601856 1.7935189\n```\n\nIn this output, we see that the estimated mean mood gain for the placebo group with no therapy was 0.300, with a 95% confidence interval from 0.006 to 0.594. Note that these are not the same confidence intervals that you would get if you calculated them separately for each group, because of the fact that the ANOVA model assumes homogeneity of variance and therefore uses a pooled estimate of the standard deviation.\n When the model doesn’t contain the interaction term, then the estimated group means will be different from the sample means. Instead of reporting the sample mean, the `effect()` function will calculate the value of the group means that would be expected on the basis of the marginal means (i.e., assuming no interaction). Using the notation we developed earlier, the estimate reported for \\(\\mu_{rc}\\), the mean for level \\(r\\) on the (row) Factor A and level \\(c\\) on the (column) Factor B would be \\(\\mu_{..} + \\alpha_r + \\beta_c\\). If there are genuinely no interactions between the two factors, this is actually a better estimate of the population mean than the raw sample mean would be. The command to obtain these estimates is actually identical to the last one, except that we use `model.2` . When you do this, R will give you a warning message: \n\n```\neff <- effect( \"drug*therapy\", model.2 )\n```\n\n```\n## NOTE: drug:therapy does not appear in the model\n```\n\n but this isn’t anything to worry about. This is R being polite, and letting you know that the estimates it is constructing are based on the assumption that no interactions exist. It kind of makes sense that it would do this: when we use `\"drug*therapy\"` as our input, we’re telling R that we want it to output the estimated group means (rather than marginal means), but the actual input `\"drug*therapy\"` might mean that you want interactions included or you might not. There’s no actual ambiguity here, because the model itself either does or doesn’t have interactions, but the authors of the function thought it sensible to include a warning just to make sure that you’ve specified the actual model you care about. But, assuming that we genuinely don’t believe that there are any interactions, `model.2` is the right model to use, so we can ignore this warning.237 In any case, when we inspect the output, we get the following table of estimated group means: `eff` \n\n```\n## \n## drug*therapy effect\n## therapy\n## drug no.therapy CBT\n## placebo 0.2888889 0.6111111\n## anxifree 0.5555556 0.8777778\n## joyzepam 1.3222222 1.6444444\n```\n\nAs before, we can obtain confidence intervals using the following command:\n `summary( eff )` \n\n```\n## \n## drug*therapy effect\n## therapy\n## drug no.therapy CBT\n## placebo 0.2888889 0.6111111\n## anxifree 0.5555556 0.8777778\n## joyzepam 1.3222222 1.6444444\n## \n## Lower 95 Percent Confidence Limits\n## therapy\n## drug no.therapy CBT\n## placebo 0.02907986 0.3513021\n## anxifree 0.29574653 0.6179687\n## joyzepam 1.06241319 1.3846354\n## \n## Upper 95 Percent Confidence Limits\n## therapy\n## drug no.therapy CBT\n## placebo 0.5486979 0.8709201\n## anxifree 0.8153646 1.1375868\n## joyzepam 1.5820313 1.9042535\n```\n\nbut the output looks pretty much the same as last time, and this book is already way too long, so I won’t include it here.\n\n## 16.4 Assumption checking\n\nAs with one-way ANOVA, the key assumptions of factorial ANOVA are homogeneity of variance (all groups have the same standard deviation), normality of the residuals, and independence of the observations. The first two are things we can test for. The third is something that you need to assess yourself by asking if there are any special relationships between different observations. Additionally, if you aren’t using a saturated model (e.g., if you’ve omitted the interaction terms) then you’re also assuming that the omitted terms aren’t important. Of course, you can check this last one by running an ANOVA with the omitted terms included and see if they’re significant, so that’s pretty easy. What about homogeneity of variance and normality of the residuals? As it turns out, these are pretty easy to check: it’s no different to the checks we did for a one-way ANOVA.\n\n### 16.4.1 Levene test for homogeneity of variance\n\n To test whether the groups have the same variance, we can use the Levene test. The theory behind the Levene test was discussed in Section 14.7, so I won’t discuss it again. Once again, you can use the `leveneTest()` function in the `car` package to do this. This function expects that you have a saturated model (i.e., included all of the relevant terms), because the test is primarily concerned with the within-group variance, and it doesn’t really make a lot of sense to calculate this any way other than with respect to the full model. So we try either of the following commands: \n\nR will spit out the following error:\n\n```\nError in leveneTest.formula(formula(y), data = model.frame(y), ...) : \n Model must be completely crossed formula only.\n```\n\nInstead, if you want to run the Levene test, you need to specify a saturated model. Either of the following two commands would work:238\n\n```\nlibrary(car)\n leveneTest( model.3 )\n```\n\n```\nleveneTest( mood.gain ~ drug * therapy, clin.trial )\n```\n\nThe fact that the Levene test is non-significant means that we can safely assume that the homogeneity of variance assumption is not violated.\n\n### 16.4.2 Normality of residuals\n\n As with one-way ANOVA, we can test for the normality of residuals in a straightforward fashion (see Section 14.9). First, we use the `residuals()` function to extract the residuals from the model itself, and then we can examine those residuals in a few different ways. It’s generally a good idea to examine them graphically, by drawing histograms (i.e., `hist()` function) and QQ plots (i.e., `qqnorm()` function. If you want a formal test for the normality of the residuals, then we can run the Shapiro-Wilk test (i.e., `shapiro.test()` ). If we wanted to check the residuals with respect to `model.2` (i.e., the model with both main effects but no interactions) then we could do the following: \n\n```\nresid <- residuals( model.2 ) # pull the residuals\n hist( resid ) # draw a histogram\n```\n\n```\nqqnorm( resid ) # draw a normal QQ plot\n```\n\n```\nshapiro.test( resid ) # run the Shapiro-Wilk test\n```\n\n```\n## \n## Shapiro-Wilk normality test\n## \n## data: resid\n## W = 0.95635, p-value = 0.5329\n```\n\nI haven’t included the plots (you can draw them yourself if you want to see them), but you can see from the non-significance of the Shapiro-Wilk test that normality isn’t violated here.\n\n## 16.5 The \\(F\\) test as a model comparison\n\n At this point, I want to talk in a little more detail about what the \\(F\\)-tests in an ANOVA are actually doing. In the context of ANOVA, I’ve been referring to the \\(F\\)-test as a way of testing whether a particular term in the model (e.g., main effect of Factor A) is significant. This interpretation is perfectly valid, but it’s not necessarily the most useful way to think about the test. In fact, it’s actually a fairly limiting way of thinking about what the \\(F\\)-test does. Consider the clinical trial data we’ve been working with in this chapter. Suppose I want to see if there are any effects of any kind that involve `therapy` . I’m not fussy: I don’t care if it’s a main effect or an interaction effect.239 One thing I could do is look at the output for `model.3` earlier: in this model we did see a main effect of therapy (\\(p=.013\\)) but we did not see an interaction effect (\\(p=.125\\)). That’s kind of telling us what we want to know, but it’s not quite the same thing. What we really want is a single test that jointly checks the main effect of therapy and the interaction effect. Given the way that I’ve been describing the ANOVA \\(F\\)-test up to this point, you’d be tempted to think that this isn’t possible. On the other hand, if you recall the chapter on regression (in Section 15.10), we were able to use \\(F\\)-tests to make comparisons between a wide variety of regression models. Perhaps something of that sort is possible with ANOVA? And of course, the answer here is yes. The thing that you really need to understand is that the \\(F\\)-test, as it is used in both ANOVA and regression, is really a comparison of two statistical models. One of these models is the full model (alternative hypothesis), and the other model is a simpler model that is missing one or more of the terms that the full model includes (null hypothesis). The null model cannot contain any terms that are not in the full model. In the example I gave above, the full model is `model.3` , and it contains a main effect for therapy, a main effect for drug, and the drug by therapy interaction term. The null model would be `model.1` since it contains only the main effect of drug. \n\n### 16.5.1 The \\(F\\) test comparing two models\n\n Let’s frame this in a slightly more abstract way. We’ll say that our full model can be written as an R formula that contains several different terms, say `Y ~ A + B + C + D` . Our null model only contains some subset of these terms, say `Y ~ A + B` . Some of these terms might be main effect terms, others might be interaction terms. It really doesn’t matter. The only thing that matters here is that we want to treat some of these terms as the “starting point” (i.e. the terms in the null model, `A` and `B` ), and we want to see if including the other terms (i.e., `C` and `D` ) leads to a significant improvement in model performance, over and above what could be achieved by a model that includes only `A` and `B` . In essence, we have null and alternative hypotheses that look like this: \n\nHypothesis | Correct model? | R formula for correct model |\n | --- | --- | --- |\n\nHypothesisCorrect model?R formula for correct modelNull\\(M0\\)Y ~ A + BAlternative\\(M1\\)Y ~ A + B + C + D\n\nIs there a way of making this comparison directly?\n\nTo answer this, let’s go back to fundamentals. As we saw in Chapter 14, the \\(F\\)-test is constructed from two kinds of quantity: sums of squares (SS) and degrees of freedom (df). These two things define a mean square value (MS = SS/df), and we obtain our \\(F\\) statistic by contrasting the MS value associated with “the thing we’re interested in” (the model) with the MS value associated with “everything else” (the residuals). What we want to do is figure out how to talk about the SS value that is associated with the difference between two models. It’s actually not all that hard to do.\n\nLet’s start with the fundamental rule that we used throughout the chapter on regression: \\[ \\mbox{SS}_{T} = \\mbox{SS}_{M} + \\mbox{SS}_{R} \\] That is, the total sums of squares (i.e., the overall variability of the outcome variable) can be decomposed into two parts: the variability associated with the model \\(\\mbox{SS}_{M}\\), and the residual or leftover variability, \\(\\mbox{SS}_{R}\\). However, it’s kind of useful to rearrange this equation slightly, and say that the SS value associated with a model is defined like this… \\[ \\mbox{SS}_{M} = \\mbox{SS}_{T} - \\mbox{SS}_{R} \\] Now, in our scenario, we have two models: the null model (M0) and the full model (M1):\n\n\\[ \\mbox{SS}_{M0} = \\mbox{SS}_{T} - \\mbox{SS}_{R0} \\]\n\n\\[ \\mbox{SS}_{M1} = \\mbox{SS}_{T} - \\mbox{SS}_{R1} \\] Next, let’s think about what it is we actually care about here. What we’re interested in is the difference between the full model and the null model. So, if we want to preserve the idea that what we’re doing is an “analysis of the variance” (ANOVA) in the outcome variable, what we should do is define the SS associated with the difference to be equal to the difference in the SS: \\[ \\begin{array} \\mbox{SS}_{\\Delta} &=& \\mbox{SS}_{M1} - \\mbox{SS}_{M0}\\\\ &=& (\\mbox{SS}_{T} - \\mbox{SS}_{R1}) - (\\mbox{SS}_{T} - \\mbox{SS}_{R0} ) \\\\ &=& \\mbox{SS}_{R0} - \\mbox{SS}_{R1} \\end{array} \\]\n\nNow that we have our degrees of freedom, we can calculate mean squares and \\(F\\) values in the usual way. Specifically, we’re interested in the mean square for the difference between models, and the mean square for the residuals associated with the full model (M1), which are given by \\[ \\begin{array} \\mbox{MS}_{\\Delta} &=& \\frac{\\mbox{SS}_{\\Delta} }{ \\mbox{df}_{\\Delta} } \\\\ \\mbox{MS}_{R1} &=& \\frac{ \\mbox{SS}_{R1} }{ \\mbox{df}_{R1} }\\\\ \\end{array} \\] Finally, taking the ratio of these two gives us our \\(F\\) statistic: \\[ F = \\frac\\mbox{MS}_{\\Delta}\\mbox{MS}_{R1} \\] ### Running the test in R\n At this point, it may help to go back to our concrete example. The null model here is `model.1` , which stipulates that there is a main effect of drug, but no other effects exist. We expressed this via the model formula `mood.gain ~ drug` . The alternative model here is `model.3` , which stipulates that there is a main effect of drug, a main effect of therapy, and an interaction. If we express this in the “long” format, this model corresponds to the formula \n\n, though we often express this using the `*` shorthand. The key thing here is that if we compare `model.1` to `model.3` , we’re lumping the main effect of therapy and the interaction term together. Running this test in R is straightforward: we just input both models to the `anova()` function, and it will run the exact \\(F\\)-test that I outlined above. \n\n```\nanova( model.1, model.3 )\n```\n\n```\n## Analysis of Variance Table\n## \n## Model 1: mood.gain ~ drug\n## Model 2: mood.gain ~ drug * therapy\n## Res.Df RSS Df Sum of Sq F Pr(>F) \n## 1 15 1.39167 \n## 2 12 0.65333 3 0.73833 4.5204 0.02424 *\n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n Let’s see if we can reproduce this \\(F\\)-test ourselves. Firstly, if you go back and look at the ANOVA tables that we printed out for `model.1` and `model.3` you can reassure yourself that the RSS values printed in this table really do correspond to the residual sum of squares associated with these two models. So let’s type them in as variables: \n\n```\nss.res.null <- 1.392\n ss.res.full <- 0.653\n```\n\nNow, following the procedure that I described above, we will say that the “between model” sum of squares, is the difference between these two residual sum of squares values. So, if we do the subtraction, we discover that the sum of squares associated with those terms that appear in the full model but not the null model is:\n\n```\nss.diff <- ss.res.null - ss.res.full \n ss.diff\n```\n\n `## [1] 0.739` Right. Next, as always we need to convert these SS values into MS (mean square) values, which we do by dividing by the degrees of freedom. The degrees of freedom associated with the full-model residuals hasn’t changed from our original ANOVA for `model.3` : it’s the total sample size \\(N\\), minus the total number of groups \\(G\\) that are relevant to the model. We have 18 people in the trial and 6 possible groups (i.e., 2 therapies \\(\\times\\) 3 drugs), so the degrees of freedom here is 12. The degrees of freedom for the null model are calculated similarly. The only difference here is that there are only 3 relevant groups (i.e., 3 drugs), so the degrees of freedom here is 15. And, because the degrees of freedom associated with the difference is equal to the difference in the two degrees of freedom, we arrive at the conclusion that we have \\(15-12 = 3\\) degrees of freedom. Now that we know the degrees of freedom, we can calculate our MS values: \n\n```\nms.res <- ss.res.full / 12\n ms.diff <- ss.diff / 3\n```\n\nOkay, now that we have our two MS values, we can divide one by the other, and obtain an \\(F\\)-statistic …\n\n```\nF.stat <- ms.diff / ms.res \n F.stat\n```\n\n `## [1] 4.526799` … and, just as we had hoped, this turns out to be identical to the \\(F\\)-statistic that the `anova()` function produced earlier. \n\n## 16.6 ANOVA as a linear model\n\nOne of the most important things to understand about ANOVA and regression is that they’re basically the same thing. On the surface of it, you wouldn’t think that this is true: after all, the way that I’ve described them so far suggests that ANOVA is primarily concerned with testing for group differences, and regression is primarily concerned with understanding the correlations between variables. And as far as it goes, that’s perfectly true. But when you look under the hood, so to speak, the underlying mechanics of ANOVA and regression are awfully similar. In fact, if you think about it, you’ve already seen evidence of this. ANOVA and regression both rely heavily on sums of squares (SS), both make use of \\(F\\) tests, and so on. Looking back, it’s hard to escape the feeling that Chapters 14 and 15 were a bit repetitive.\n The reason for this is that ANOVA and regression are both kinds of linear models. In the case of regression, this is kind of obvious. The regression equation that we use to define the relationship between predictors and outcomes is the equation for a straight line, so it’s quite obviously a linear model. And if that wasn’t a big enough clue, the simple fact that the command to run a regression is `lm()` is kind of a hint too. When we use an R formula like \n\n```\noutcome ~ predictor1 + predictor2\n```\n\nwhat we’re really working with is the somewhat uglier linear model: \\[\nY_{p} = b_1 X_{1p} + b_2 X_{2p} + b_0 + \\epsilon_{p}\n\\] where \\(Y_p\\) is the outcome value for the \\(p\\)-th observation (e.g., \\(p\\)-th person), \\(X_{1p}\\) is the value of the first predictor for the \\(p\\)-th observation, \\(X_{2p}\\) is the value of the second predictor for the \\(p\\)-th observation, the \\(b_1\\), \\(b_2\\) and \\(b_0\\) terms are our regression coefficients, and \\(\\epsilon_{p}\\) is the \\(p\\)-th residual. If we ignore the residuals \\(\\epsilon_{p}\\) and just focus on the regression line itself, we get the following formula: \\[\n\\hat{Y}_{p} = b_1 X_{1p} + b_2 X_{2p} + b_0\n\\] where \\(\\hat{Y}_p\\) is the value of \\(Y\\) that the regression line predicts for person \\(p\\), as opposed to the actually-observed value \\(Y_p\\). The thing that isn’t immediately obvious is that we can write ANOVA as a linear model as well. However, it’s actually pretty straightforward to do this. Let’s start with a really simple example: rewriting a \\(2 \\times 2\\) factorial ANOVA as a linear model. \n\n### 16.6.1 Some data\n\n To make things concrete, let’s suppose that our outcome variable is the `grade` that a student receives in my class, a ratio-scale variable corresponding to a mark from 0% to 100%. There are two predictor variables of interest: whether or not the student turned up to lectures (the `attend` variable), and whether or not the student actually read the textbook (the `reading` variable). We’ll say that `attend = 1` if the student attended class, and `attend = 0` if they did not. Similarly, we’ll say that `reading = 1` if the student read the textbook, and `reading = 0` if they did not. Okay, so far that’s simple enough. The next thing we need to do is to wrap some maths around this (sorry!). For the purposes of this example, let \\(Y_p\\) denote the `grade` of the \\(p\\)-th student in the class. This is not quite the same notation that we used earlier in this chapter: previously, we’ve used the notation \\(Y_{rci}\\) to refer to the \\(i\\)-th person in the \\(r\\)-th group for predictor 1 (the row factor) and the \\(c\\)-th group for predictor 2 (the column factor). This extended notation was really handy for describing how the SS values are calculated, but it’s a pain in the current context, so I’ll switch notation here. Now, the \\(Y_p\\) notation is visually simpler than \\(Y_{rci}\\), but it has the shortcoming that it doesn’t actually keep track of the group memberships! That is, if I told you that \\(Y_{0,0,3} = 35\\), you’d immediately know that we’re talking about a student (the 3rd such student, in fact) who didn’t attend the lectures (i.e., `attend = 0` ) and didn’t read the textbook (i.e. `reading = 0` ), and who ended up failing the class ( `grade = 35` ). But if I tell you that \\(Y_p = 35\\) all you know is that the \\(p\\)-th student didn’t get a good grade. We’ve lost some key information here. Of course, it doesn’t take a lot of thought to figure out how to fix this: what we’ll do instead is introduce two new variables \\(X_{1p}\\) and \\(X_{2p}\\) that keep track of this information. In the case of our hypothetical student, we know that \\(X_{1p} = 0\\) (i.e., `attend = 0` ) and \\(X_{2p} = 0\\) (i.e., `reading = 0` ). So the data might look like this: \n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2, ~V3, ~V4,\n \"1\", \"90\", \"1\", \"1\",\n \"2\", \"87\", \"1\", \"1\",\n \"3\", \"75\", \"0\", \"1\",\n \"4\", \"60\", \"1\", \"0\",\n \"5\", \"35\", \"0\", \"0\",\n \"6\", \"50\", \"0\", \"0\",\n \"7\", \"65\", \"1\", \"0\",\n \"8\", \"70\", \"0\", \"1\"), \ncol.names= c(\"person $p$\", \"grade $Y_p$\", \"attendance $X_{1p}$\", \"reading $X_{2p}$\"), align = 'cccc')\n```\n\nperson \\(p\\) | grade \\(Y_p\\) | attendance \\(X_{1p}\\) | reading \\(X_{2p}\\) |\n | --- | --- | --- | --- |\n1 | 90 | 1 | 1 |\n2 | 87 | 1 | 1 |\n3 | 75 | 0 | 1 |\n4 | 60 | 1 | 0 |\n5 | 35 | 0 | 0 |\n6 | 50 | 0 | 0 |\n7 | 65 | 1 | 0 |\n8 | 70 | 0 | 1 |\n\n This isn’t anything particularly special, of course: it’s exactly the format in which we expect to see our data! In other words, if your data have been stored as a data frame in R then you’re probably expecting to see something that looks like the `rtfm.1` data frame: \n\n```\nload(file.path(projecthome, \"data\",\"rtfm.Rdata\"))\nrtfm.1\n```\n\n Well, sort of. I suspect that a few readers are probably frowning a little at this point. Earlier on in the book I emphasised the importance of converting nominal scale variables such as `attend` and `reading` to factors, rather than encoding them as numeric variables. The `rtfm.1` data frame doesn’t do this, but the `rtfm.2` data frame does, and so you might instead be expecting to see data like this: `rtfm.2` \n\n However, for the purposes of this section it’s important that we be able to switch back and forth between these two different ways of thinking about the data. After all, our goal in this section is to look at some of the mathematics that underpins ANOVA, and if we want to do that we need to be able to see the numerical representation of the data (in `rtfm.1` ) as well as the more meaningful factor representation (in `rtfm.2` ). In any case, we can use the `xtabs()` function to confirm that this data set corresponds to a balanced design \n\n```\nxtabs( ~ attend + reading, rtfm.2 )\n```\n\n```\n## reading\n## attend no yes\n## no 2 2\n## yes 2 2\n```\n\n For each possible combination of the `attend` and `reading` variables, we have exactly two students. If we’re interested in calculating the mean `grade` for each of these cells, we can use the `aggregate()` function: \n\n```\naggregate( grade ~ attend + reading, rtfm.2, mean )\n```\n\n```\n## attend reading grade\n## 1 no no 42.5\n## 2 yes no 62.5\n## 3 no yes 72.5\n## 4 yes yes 88.5\n```\n\nLooking at this table, one gets the strong impression that reading the text and attending the class both matter a lot.\n\n### 16.6.2 ANOVA with binary factors as a regression model\n\n Okay, let’s get back to talking about the mathematics. We now have our data expressed in terms of three numeric variables: the continuous variable \\(Y\\), and the two binary variables \\(X_1\\) and \\(X_2\\). What I want to you to recognise is that our 2$$2 factorial ANOVA is exactly equivalent to the regression model \\[ Y_{p} = b_1 X_{1p} + b_2 X_{2p} + b_0 + \\epsilon_p \\] This is, of course, the exact same equation that I used earlier to describe a two-predictor regression model! The only difference is that \\(X_1\\) and \\(X_2\\) are now binary variables (i.e., values can only be 0 or 1), whereas in a regression analysis we expect that \\(X_1\\) and \\(X_2\\) will be continuous. There’s a couple of ways I could try to convince you of this. One possibility would be to do a lengthy mathematical exercise, proving that the two are identical. However, I’m going to go out on a limb and guess that most of the readership of this book will find that to be annoying rather than helpful. Instead, I’ll explain the basic ideas, and then rely on R to show that that ANOVA analyses and regression analyses aren’t just similar, they’re identical for all intents and purposes.240 Let’s start by running this as an ANOVA. To do this, we’ll use the `rtfm.2` data frame, since that’s the one in which I did the proper thing of coding `attend` and `reading` as factors, and I’ll use the `aov()` function to do the analysis. Here’s what we get… \n\n```\nanova.model <- aov( grade ~ attend + reading, data = rtfm.2 )\n summary( anova.model )\n```\n\n```\n## Df Sum Sq Mean Sq F value Pr(>F) \n## attend 1 648 648 21.60 0.00559 ** \n## reading 1 1568 1568 52.27 0.00079 ***\n## Residuals 5 150 30 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nSo, by reading the key numbers off the ANOVA table and the table of means that we presented earlier, we can see that the students obtained a higher grade if they attended class \\((F_{1,5} = 26.1, p = .0056)\\) and if they read the textbook \\((F_{1,5} = 52.3, p = .0008)\\). Let’s make a note of those \\(p\\)-values and those \\(F\\) statistics.\n\n```\nlibrary(effects)\n Effect( c(\"attend\",\"reading\"), anova.model )\n```\n\n```\n## \n## attend*reading effect\n## reading\n## attend no yes\n## no 43.5 71.5\n## yes 61.5 89.5\n```\n\n Now let’s think about the same analysis from a linear regression perspective. In the `rtfm.1` data set, we have encoded `attend` and `reading` as if they were numeric predictors. In this case, this is perfectly acceptable. There really is a sense in which a student who turns up to class (i.e. `attend = 1` ) has in fact done “more attendance” than a student who does not (i.e. `attend = 0` ). So it’s not at all unreasonable to include it as a predictor in a regression model. It’s a little unusual, because the predictor only takes on two possible values, but it doesn’t violate any of the assumptions of linear regression. And it’s easy to interpret. If the regression coefficient for `attend` is greater than 0, it means that students that attend lectures get higher grades; if it’s less than zero, then students attending lectures get lower grades. The same is true for our `reading` variable. Wait a second… why is this true? It’s something that is intuitively obvious to everyone who has taken a few stats classes and is comfortable with the maths, but it isn’t clear to everyone else at first pass. To see why this is true, it helps to look closely at a few specific students. Let’s start by considering the 6th and 7th students in our data set (i.e. \\(p=6\\) and \\(p=7\\)). Neither one has read the textbook, so in both cases we can set `reading = 0` . Or, to say the same thing in our mathematical notation, we observe \\(X_{2,6} = 0\\) and \\(X_{2,7} = 0\\). However, student number 7 did turn up to lectures (i.e., `attend = 1` , \\(X_{1,7} = 1\\)) whereas student number 6 did not (i.e., `attend = 0` , \\(X_{1,6} = 0\\)). Now let’s look at what happens when we insert these numbers into the general formula for our regression line. For student number 6, the regression predicts that \\[\n\\begin{array}\n\\hat{Y}_{6} &=& b_1 X_{1,6} + b_2 X_{2,6} + b_0 \\\\\n&=& (b_1 \\times 0) + ( b_2 \\times 0) + b_0 \\\\\n&=& b_0\n\\end{array}\n\\] So we’re expecting that this student will obtain a grade corresponding to the value of the intercept term \\(b_0\\). What about student 7? This time, when we insert the numbers into the formula for the regression line, we obtain the following: \\[\n\\begin{array}\n\\hat{Y}_{7} &=& b_1 X_{1,7} + b_2 X_{2,7} + b_0 \\\\\n&=& (b_1 \\times 1) + ( b_2 \\times 0) + b_0 \\\\\n&=& b_1 + b_0\n\\end{array}\n\\] Because this student attended class, the predicted grade is equal to the intercept term \\(b_0\\) plus the coefficient associated with the `attend` variable, \\(b_1\\). So, if \\(b_1\\) is greater than zero, we’re expecting that the students who turn up to lectures will get higher grades than those students who don’t. If this coefficient is negative, we’re expecting the opposite: students who turn up at class end up performing much worse. In fact, we can push this a little bit further. What about student number 1, who turned up to class (\\(X_{1,1} = 1\\)) and read the textbook (\\(X_{2,1} = 1\\))? If we plug these numbers into the regression, we get \\[\n\\begin{array}\n\\hat{Y}_{1} &=& b_1 X_{1,1} + b_2 X_{2,1} + b_0 \\\\\n&=& (b_1 \\times 1) + ( b_2 \\times 1) + b_0 \\\\\n&=& b_1 + b_2 + b_0\n\\end{array}\n\\] So if we assume that attending class helps you get a good grade (i.e., \\(b_1 > 0\\)) and if we assume that reading the textbook also helps you get a good grade (i.e., \\(b_2 >0\\)), then our expectation is that student 1 will get a grade that that is higher than student 6 and student 7. \nAnd at this point, you won’t be at all suprised to learn that the regression model predicts that student 3, who read the book but didn’t attend lectures, will obtain a grade of \\(b_2 + b_0\\). I won’t bore you with yet another regression formula. Instead, what I’ll do is show you the following table of expected grades:\n\n \"attended - no\",\"$b_0$\",\"$b_0 + b_2$\", \n \"attended - yes\", \"$b_0 + b_1$\", \"$b_0 + b_1 + b_2$\"),\n col.names = c(\"\",\"read textbook? no\", \"read textbook? yes\"))\n```\n\nread textbook? no | read textbook? yes |\n | --- | --- |\nattended - no | \\(b_0\\) | \\(b_0 + b_2\\) |\nattended - yes | \\(b_0 + b_1\\) | \\(b_0 + b_1 + b_2\\) |\n\nAs you can see, the intercept term \\(b_0\\) acts like a kind of “baseline” grade that you would expect from those students who don’t take the time to attend class or read the textbook. Similarly, \\(b_1\\) represents the boost that you’re expected to get if you come to class, and \\(b_2\\) represents the boost that comes from reading the textbook. In fact, if this were an ANOVA you might very well want to characterise \\(b_1\\) as the main effect of attendance, and \\(b_2\\) as the main effect of reading! In fact, for a simple \\(2 \\times 2\\) ANOVA that’s exactly how it plays out.\n Okay, now that we’re really starting to see why ANOVA and regression are basically the same thing, let’s actually run our regression using the `rtfm.1` data and the `lm()` function to convince ourselves that this is really true. Running the regression in the usual way gives us the following output:241 \n\n```\nregression.model <- lm( grade ~ attend + reading, data = rtfm.1 )\n summary( regression.model )\n```\n\n There’s a few interesting things to note here. Firstly, notice that the intercept term is 43.5, which is close to the “group” mean of 42.5 observed for those two students who didn’t read the text or attend class. Moreover, it’s identical to the predicted group mean that we pulled out of our ANOVA using the `Effects()` function! Secondly, notice that we have the regression coefficient of \\(b_1 = 18.0\\) for the attendance variable, suggesting that those students that attended class scored 18% higher than those who didn’t. So our expectation would be that those students who turned up to class but didn’t read the textbook would obtain a grade of \\(b_0 + b_1\\), which is equal to \\(43.5 + 18.0 = 61.5\\). Again, this is similar to the observed group mean of 62.5, and identical to the expected group mean that we pulled from our ANOVA. You can verify for yourself that the same thing happens when we look at the students that read the textbook. Actually, we can push a little further in establishing the equivalence of our ANOVA and our regression. Look at the \\(p\\)-values associated with the `attend` variable and the `reading` variable in the regression output. They’re identical to the ones we encountered earlier when running the ANOVA. This might seem a little surprising, since the test used when running our regression model calculates a \\(t\\)-statistic and the ANOVA calculates an \\(F\\)-statistic. However, if you can remember all the way back to Chapter 9, I mentioned that there’s a relationship between the \\(t\\)-distribution and the \\(F\\)-distribution: if you have some quantity that is distributed according to a \\(t\\)-distribution with \\(k\\) degrees of freedom and you square it, then this new squared quantity follows an \\(F\\)-distribution whose degrees of freedom are 1 and \\(k\\). We can check this with respect to the \\(t\\) statistics in our regression model. For the `attend` variable we get a \\(t\\) value of 4.648. If we square this number we end up with 21.604, which is identical to the corresponding \\(F\\) statistic in our ANOVA. Finally, one last thing you should know. Because R understands the fact that ANOVA and regression are both examples of linear models, it lets you extract the classic ANOVA table from your regression model using the `anova()` function. All you have to do is this: \n\n```\nanova( regression.model )\n```\n\n```\n## Analysis of Variance Table\n## \n## Response: grade\n## Df Sum Sq Mean Sq F value Pr(>F) \n## attend 1 648 648 21.600 0.0055943 ** \n## reading 1 1568 1568 52.267 0.0007899 ***\n## Residuals 5 150 30 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n### 16.6.3 Changing the baseline category\n\n At this point, you’re probably convinced that the ANOVA and the regression are actually identical to each other. So there’s one last thing I should show you. What happens if I use the data from `rtfm.2` to run the regression? In `rtfm.2` , we coded the `attend` and `reading` variables as factors rather than as numeric variables. Does this matter? It turns out that it doesn’t. The only differences are superficial: \n\n```\nregression.model.2 <- lm( grade ~ attend + reading, data = rtfm.2 )\n summary( regression.model.2 )\n```\n\n The only thing that is different is that R labels the two variables differently: the output now refers to `attendyes` and `readingyes` . You can probably guess what this means. When R refers to `readingyes` it’s trying to indicate that it is assuming that “yes = 1” and “no = 0”. This is important. Suppose we wanted to say that “yes = 0” and “no = 1”. We could still run this as a regression model, but now all of our coefficients will go in the opposite direction, because the effect of `readingno` would be referring to the consequences of not reading the textbook. To show you how this works, we can use the `relevel()` function in R to change which level of the `reading` variable is set to “0”. Here’s how it works. First, let’s get R to print out the `reading` factor as it currently stands: `rtfm.2$reading` \n\n Notice that order in which R prints out the levels is “no” and then “yes”. Now let’s apply the `relevel()` function: \n\n```\nrelevel( x = rtfm.2$reading, ref = \"yes\" )\n```\n\nR now lists “yes” before “no”. This means that R will now treat “yes” as the “reference” level (sometimes called the baseline level) when you include it in an ANOVA. So let’s now create a new data frame with our factors recoded…\n\n```\nrtfm.3 <- rtfm.2 # copy the old data frame\n rtfm.3$reading <- relevel( rtfm.2$reading, ref=\"yes\" ) # re-level the reading factor\n rtfm.3$attend <- relevel( rtfm.2$attend, ref=\"yes\" ) # re-level the attend factor\n```\n\nFinally, let’s re-run our regression, this time using the re-coded data:\n\n```\nregression.model.3 <- lm( grade ~ attend + reading, data = rtfm.3 )\n summary( regression.model.3 )\n```\n\n As you can see, there are now a few changes. Most obviously, the `attendno` and `readingno` effects are both negative, though they’re the same magnitude as before: if you don’t read the textbook, for instance, you should expect your grade to drop by 28% relative to someone who did. The \\(t\\)-statistics have reversed sign too. The \\(p\\)-values remain the same, of course. The intercept has changed too. In our original regression, the baseline corresponded to a student who didn’t attend class and didn’t read the textbook, so we got a value of 43.5 as the expected baseline grade. However, now that we’ve recoded our variables, the baseline corresponds to a student who has read the textbook and did attend class, and for that student we would expect a grade of 89.5. \n\n### 16.6.4 How to encode non binary factors as contrasts\n\n At this point, I’ve shown you how we can view a \\(2\\times 2\\) ANOVA into a linear model. And it’s pretty easy to see how this generalises to a \\(2 \\times 2 \\times 2\\) ANOVA or a \\(2 \\times 2 \\times 2 \\times 2\\) ANOVA… it’s the same thing, really: you just add a new binary variable for each of your factors. Where it begins to get trickier is when we consider factors that have more than two levels. Consider, for instance, the \\(3 \\times 2\\) ANOVA that we ran earlier in this chapter using the `clin.trial` data. How can we convert the three-level `drug` factor into a numerical form that is appropriate for a regression? The answer to this question is pretty simple, actually. All we have to do is realise that a three-level factor can be redescribed as two binary variables. Suppose, for instance, I were to create a new binary variable called `druganxifree` . Whenever the `drug` variable is equal to `\"anxifree\"` we set `druganxifree = 1` . Otherwise, we set `druganxifree = 0` . This variable sets up a contrast, in this case between anxifree and the other two drugs. By itself, of course, the `druganxifree` contrast isn’t enough to fully capture all of the information in our `drug` variable. We need a second contrast, one that allows us to distinguish between joyzepam and the placebo. To do this, we can create a second binary contrast, called `drugjoyzepam` , which equals 1 if the drug is joyzepam, and 0 if it is not. Taken together, these two contrasts allows us to perfectly discriminate between all three possible drugs. The table below illustrates this: \n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2, ~V3,\n \"`placebo`\", \"0\", \"0\",\n \"`anxifree`\", \"1\", \"0\",\n \"`joyzepam`\", \"0\", \"1\"\n ), col.names = c( \"`drug`\", \"`druganxifree`\", \"`drugjoyzepam`\"))\n```\n\ndrugdruganxifreedrugjoyzepamplacebo00anxifree10joyzepam01\n If the drug administered to a patient is a placebo, then both of the two contrast variables will equal 0. If the drug is Anxifree, then the `druganxifree` variable will equal 1, and `drugjoyzepam` will be 0. The reverse is true for Joyzepam: `drugjoyzepam` is 1, and `druganxifree` is 0. Creating contrast variables manually is not too difficult to do using basic R commands. For example, here’s how we would create the `druganxifree` variable: \n\n```\ndruganxifree <- as.numeric( clin.trial$drug == \"anxifree\" )\n druganxifree\n```\n\n The\n\n```\nclin.trial$drug == \"anxifree\"\n```\n\npart of the command returns a logical vector that has a value of `TRUE` if the drug is Anxifree, and a value of `FALSE` if the drug is Joyzepam or the placebo. The `as.numeric()` function just converts `TRUE` to 1 and `FALSE` to 0. Obviously, this command creates the `druganxifree` variable inside the workspace. If you wanted to add it to the `clin.trial` data frame, you’d use a commmand like this instead: \n\n```\nclin.trial$druganxifree <- as.numeric( clin.trial$drug == \"anxifree\" )\n```\n\n You could then repeat this for the other contrasts that you wanted to use. However, it’s kind of tedious to do this over and over again for every single contrast that you want to create. To make it a little easier, the `lsr` package contains a simple function called `expandFactors()` that will convert every factor in a data frame into a set of contrast variables.242 We can use it to create a new data frame, `clin.trial.2` that contains the same data as `clin.trial` , but with the two factors represented in terms of the contrast variables: \n\n```\nclin.trial.2 <- expandFactors( clin.trial )\n```\n\n `clin.trial.2` \n\n It’s not as pretty as the original `clin.trial` data, but it’s definitely saying the same thing. We have now recoded our three-level factor in terms of two binary variables, and we’ve already seen that ANOVA and regression behave the same way for binary variables. However, there are some additional complexities that arise in this case, which we’ll discuss in the next section. \n\n### 16.6.5 The equivalence between ANOVA and regression for non-binary factors\n\n Now we have two different versions of the same data set: our original data frame `clin.trial` in which the `drug` variable is expressed as a single three-level factor, and the expanded data set `clin.trial.2` in which it is expanded into two binary contrasts. Once again, the thing that we want to demonstrate is that our original \\(3 \\times 2\\) factorial ANOVA is equivalent to a regression model applied to the contrast variables. Let’s start by re-running the ANOVA: \n\n```\ndrug.anova <- aov( mood.gain ~ drug + therapy, clin.trial )\n summary( drug.anova )\n```\n\n```\n## Df Sum Sq Mean Sq F value Pr(>F) \n## drug 2 3.453 1.7267 26.149 1.87e-05 ***\n## therapy 1 0.467 0.4672 7.076 0.0187 * \n## Residuals 14 0.924 0.0660 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n Obviously, there’s no surprises here. That’s the exact same ANOVA that we ran earlier, except for the fact that I’ve arbitrarily decided to rename the output variable as `drug.anova` for some stupid reason.243 Next, let’s run a regression, using `druganxifree` , `drugjoyzepam` and `therapyCBT` as the predictors. Here’s what we get: \n\n```\ndrug.regression <- lm( mood.gain ~ druganxifree + drugjoyzepam + therapyCBT, clin.trial.2 )\n summary( drug.regression )\n```\n\n Hm. This isn’t the same output that we got last time. Not surprisingly, the regression output prints out the results for each of the three predictors separately, just like it did every other time we used `lm()` . On the one hand, we can see that the \\(p\\)-value for the `therapyCBT` variable is exactly the same as the one for the `therapy` factor in our original ANOVA, so we can be reassured that the regression model is doing the same thing as the ANOVA did. On the other hand, this regression model is testing the `druganxifree` contrast and the `drugjoyzepam` contrast separately, as if they were two completely unrelated variables. It’s not surprising of course, because the poor `lm()` function has no way of knowing that `drugjoyzepam` and `druganxifree` are actually the two different contrasts that we used to encode our three-level `drug` factor. As far as it knows, `drugjoyzepam` and `druganxifree` are no more related to one another than `drugjoyzepam` and `therapyCBT` . However, you and I know better. At this stage we’re not at all interested in determining whether these two contrasts are individually significant. We just want to know if there’s an “overall” effect of drug. That is, what we want R to do is to run some kind of “omnibus” test, one in which the two “drug-related” contrasts are lumped together for the purpose of the test. Sound familiar? This is exactly the situation that we discussed in Section 16.5, and it is precisely this situation that the \\(F\\)-test is built to handle. All we need to do is specify our null model, which in this case would include the `therapyCBT` predictor, and omit both of the drug-related variables, and then run it through the `anova()` function: \n\n```\nnodrug.regression <- lm( mood.gain ~ therapyCBT, clin.trial.2 )\n anova( nodrug.regression, drug.regression )\n```\n\n```\n## Analysis of Variance Table\n## \n## Model 1: mood.gain ~ therapyCBT\n## Model 2: mood.gain ~ druganxifree + drugjoyzepam + therapyCBT\n## Res.Df RSS Df Sum of Sq F Pr(>F) \n## 1 16 4.3778 \n## 2 14 0.9244 2 3.4533 26.149 1.872e-05 ***\n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n Ah, that’s better. Our \\(F\\)-statistic is 26.1, the degrees of freedom are 2 and 14, and the \\(p\\)-value is \\(0.000019\\). The numbers are identical to the ones we obtained for the main effect of `drug` in our original ANOVA. Once again, we see that ANOVA and regression are essentially the same: they are both linear models, and the underlying statistical machinery for ANOVA is identical to the machinery used in regression. The importance of this fact should not be understated. Throughout the rest of this chapter we’re going to rely heavily on this idea. \n\n### 16.6.6 Degrees of freedom as parameter counting!\n\n At long last, I can finally give a definition of degrees of freedom that I am happy with. Degrees of freedom are defined in terms of the number of parameters that have to be estimated in a model. For a regression model or an ANOVA, the number of parameters corresponds to the number of regression coefficients (i.e. \\(b\\)-values), including the intercept. Keeping in mind that any \\(F\\)-test is always a comparison between two models, the first \\(df\\) is the difference in the number of parameters. For example, model comparison above, the null model (\n\n```\nmood.gain ~ therapyCBT\n```\n\n) has two parameters: there’s one regression coefficient for the `therapyCBT` variable, and a second one for the intercept. The alternative model ( \n\n```\nmood.gain ~ druganxifree + drugjoyzepam + therapyCBT\n```\n\n) has four parameters: one regression coefficient for each of the three contrasts, and one more for the intercept. So the degrees of freedom associated with the difference between these two models is \\(df_1 = 4-2 = 2\\). What about the case when there doesn’t seem to be a null model? For instance, you might be thinking of the \\(F\\)-test that appears at the very bottom of the regression output. I originally described that as a test of the regression model as a whole. However, that is still a comparison between two models. The null model is the trivial model that only includes an intercept, which is written as `outcome ~ 1` in R, and the alternative model is the full regression model. The null model in this case contains 1 regression coefficient, for the intercept term. The alternative model contains \\(K+1\\) regression coefficients, one for each of the \\(K\\) predictor variables and one more for the intercept. So the \\(df\\) value that you see in this \\(F\\) test is equal to \\(df_1 = K+1-1 = K\\). \nWhat about the second \\(df\\) value that appears in the \\(F\\)-test? This always refers to the degrees of freedom associated with the residuals. It is possible to think of this in terms of parameters too, but in a slightly counterintuitive way. Think of it like this: suppose that the total number of observations across the study as a whole is \\(N\\). If you wanted to perfectly describe each of these \\(N\\) values, you need to do so using, well… \\(N\\) numbers. When you build a regression model, what you’re really doing is specifying some of the numbers need to perfectly describe the data. If your model has \\(K\\) predictors and an intercept, then you’ve specified \\(K+1\\) numbers. So, without bothering to figure out exactly how this would be done, how many more numbers do you think are going to be needed to transform a \\(K+1\\) parameter regression model into a perfect redescription of the raw data? If you found yourself thinking that \\((K+1) + (N-K-1) = N\\), and so the answer would have to be \\(N-K-1\\), well done! That’s exactly right: in principle you can imagine an absurdly complicated regression model that includes a parameter for every single data point, and it would of course provide a perfect description of the data. This model would contain \\(N\\) parameters in total, but we’re interested in the difference between the number of parameters required to describe this full model (i.e. \\(N\\)) and the number of parameters used by the simpler regression model that you’re actually interested in (i.e., \\(K+1\\)), and so the second degrees of freedom in the \\(F\\) test is \\(df_2 = N-K-1\\), where \\(K\\) is the number of predictors (in a regression model) or the number of contrasts (in an ANOVA). In the example I gave above, there are \\(N=18\\) observations in the data set, and \\(K+1 = 4\\) regression coefficients associated with the ANOVA model, so the degrees of freedom for the residuals is \\(df_2 = 18-4 = 14\\).\n\n### 16.6.7 A postscript\n\n There’s one last thing I want to mention in this section. In the previous example, I used the `aov()` function to run an ANOVA using the `clin.trial` data which codes `drug` variable as a single factor. I also used the `lm()` function to run a regression using the `clin.trial` data in which we have two separate contrasts describing the drug. However, it’s also possible to use the `lm()` function on the the original data. That is, you could use a command like this: \n\n```\ndrug.lm <- lm( mood.gain ~ drug + therapy, clin.trial )\n```\n\n The fact that `drug` is a three-level factor does not matter. As long as the `drug` variable has been declared to be a factor, R will automatically translate it into two binary contrast variables, and will perform the appropriate analysis. After all, as I’ve been saying throughout this section, ANOVA and regression are both linear models, and `lm()` is the function that handles linear models. In fact, the `aov()` function doesn’t actually do very much of the work when you run an ANOVA using it: internally, R just passes all the hard work straight to `lm()` . However, I want to emphasise again that it is critical that your factor variables are declared as such. If `drug` were declared to be a numeric variable, then R would be happy to treat it as one. After all, it might be that `drug` refers to the number of drugs that one has taken in the past, or something that is genuinely numeric. R won’t second guess you here. It assumes your factors are factors and your numbers are numbers. Don’t make the mistake of encoding your factors as numbers, or R will run the wrong analysis. This is not a flaw in R: it is your responsibility as the analyst to make sure you’re specifying the right model for your data. Software really can’t be trusted with this sort of thing. Okay, warnings aside, it’s actually kind of neat to run your ANOVA using the `lm()` function in the way I did above. Because you’ve called the `lm()` function, the `summary()` that R pulls out is formatted like a regression. To save space I won’t show you the output here, but you can easily verify this by typing `summary( drug.lm )` \n\n However, because the `drug` and `therapy` variables were both factors, the `anova()` function actually knows which contrasts to group together for the purposes of running the \\(F\\)-tests, so you can extract the classic ANOVA table. Again, I won’t reproduce the output here since it’s identical to the ANOVA table I showed at the start of the section, but it’s worth trying the following command `anova( drug.lm )` \n\n```\n## Analysis of Variance Table\n## \n## Response: mood.gain\n## Df Sum Sq Mean Sq F value Pr(>F) \n## drug 2 3.4533 1.72667 26.1490 1.872e-05 ***\n## therapy 1 0.4672 0.46722 7.0757 0.01866 * \n## Residuals 14 0.9244 0.06603 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n just to see for yourself. However, this behaviour of the `anova()` function only occurs when the predictor variables are factors. If we try a command like \n\n```\nanova( drug.regression )\n```\n\n, the output will continue to treate `druganxifree` and `drugjoyzepam` as if they were two distinct binary factors. This is because in the `drug.regression` model we included all the contrasts as “raw” variables, so R had no idea which ones belonged together. However, when we ran the `drug.lm` model, we gave R the original factor variables, so it does know which contrasts go together. The behaviour of the `anova()` function reflects that. \n\n## 16.7 Different ways to specify contrasts\n\nIn the previous section, I showed you a method for converting a factor into a collection of contrasts. In the method I showed you, we specify a set of binary variables, in which define a table like this one:\n\n \"\\\"placebo\\\"\", \"0\", \"0\",\n \"\\\"anxifree\\\"\", \"1\", \"0\",\n \"\\\"joyzepam\\\"\", \"0\", \"1\"\n ), col.names = c(\"drug\", \"druganxifree\", \"drugjoyzepam\"))\n```\n\ndrug | druganxifree | drugjoyzepam |\n | --- | --- | --- |\n“placebo” | 0 | 0 |\n“anxifree” | 1 | 0 |\n“joyzepam” | 0 | 1 |\n\nEach row in the table corresponds to one of the factor levels, and each column corresponds to one of the contrasts. This table, which always has one more row than columns, has a special name: it is called a contrast matrix. However, there are lots of different ways to specify a contrast matrix. In this section I discuss a few of the standard contrast matrices that statisticians use, and how you can use them in R. If you’re planning to read the section on unbalanced ANOVA later on (Section 16.10) it’s worth reading this section carefully. If not, you can get away with skimming it, because the choice of contrasts doesn’t matter much for balanced designs.\n\n### 16.7.1 Treatment contrasts\n\n In the particular kind of contrasts that I’ve described above, one level of the factor is special, and acts as a kind of “baseline” category (i.e., `placebo` in our example), against which the other two are defined. The name for these kinds of contrast is treatment contrasts. The name reflects the fact that these contrasts are quite natural and sensible when one of the categories in your factor really is special because it actually does represent a baseline. That makes sense in our clinical trial example: the `placebo` condition corresponds to the situation where you don’t give people any real drugs, and so it’s special. The other two conditions are defined in relation to the placebo: in one case you replace the placebo with Anxifree, and in the other case your replace it with Joyzepam. R comes with a variety of functions that can generate different kinds of contrast matrices. For example, the table shown above is a matrix of treatment contrasts for a factor that has 3 levels. But suppose I want a matrix of treatment contrasts for a factor with 5 levels? The `contr.treatment()` function will do this: \n\n```\ncontr.treatment( n=5 )\n```\n\n Notice that, by default, the first level of the factor is always treated as the baseline category (i.e., it’s the one that has all zeros, and doesn’t have an explicit contrast associated with it). In Section 16.6.3 I mentioned that you can use the `relevel()` function to change which category is the first level of the factor.244 There’s also a special function in R called `contr.SAS()` that generates a treatment contrast matrix in which the last category is treated as the baseline: `contr.SAS( n=5 )` \n\n However, you can actually select any category you like as the baseline within the `contr.treatment()` function, by specifying the `base` argument in that function. See the help documentation for more details. \n\n### 16.7.2 Helmert contrasts\n\n Treatment contrasts are useful for a lot of situations, and they’re the default in R. However, they make most sense in the situation when there really is a baseline category, and you want to assess all the other groups in relation to that one. In other situations, however, no such baseline category exists, and it may make more sense to compare each group to the mean of the other groups. This is where Helmert contrasts, generated by the `contr.helmert()` function, can be useful. The idea behind Helmert contrasts is to compare each group to the mean of the “previous” ones. That is, the first contrast represents the difference between group 2 and group 1, the second contrast represents the difference between group 3 and the mean of groups 1 and 2, and so on. This translates to a contrast matrix that looks like this: `contr.helmert( n=5 )` \n\nOne useful thing about Helmert contrasts is that every contrast sums to zero (i.e., all the columns sum to zero). This has the consequence that, when we interpret the ANOVA as a regression, the intercept term corresponds to the grand mean \\(\\mu_{..})\\) if we are using Helmert contrasts. Compare this to treatment contrasts, in which the intercept term corresponds to the group mean for the baseline category. This property can be very useful in some situations. It doesn’t matter very much if you have a balanced design, which we’ve been assuming so far, but it will turn out to be important later when we consider unbalanced designs in Section 16.10. In fact, the main reason why I’ve even bothered to include this section on specifying is that contrasts become important if you want to understand unbalanced ANOVA.\n\n### 16.7.3 Sum to zero contrasts\n\nThe third option that I should briefly mention are “sum to zero” contrasts, which are used to construct pairwise comparisons between groups. Specifically, each contrast encodes the difference between one of the groups and a baseline category, which in this case corresponds to the last group:\n `contr.sum( n=5 )` \n\nMuch like Helmert contrasts, we see that each column sums to zero, which means that the intercept term corresponds to the grand mean when ANOVA is treated as a regression model. When interpreting these contrasts, the thing to recognise is that each of these contrasts is a pairwise comparison between group 5 and one of the other four groups. Specifically, contrast 1 corresponds to a “group 1 minus group 5” comparison, contrast 2 corresponds to a “group 2 minus group 5” comparison, and so on.\n\n### 16.7.4 Viewing and setting the default contrasts in R\n\n Every factor variable in R is associated with a contrast matrix. It has to be, otherwise R wouldn’t be able to run ANOVAs properly! If you don’t specify one explictly, or R will implicitly specify one for you. Here’s what I mean. When I created the `clin.trial` data, I didn’t specify any contrast matrix for either of the factors. You can see this by using the `attr()` function to print out the “contrasts” attribute of the factors. For example: \n\n `## NULL` The `NULL` output here means that R is telling you that the `drug` factor doesn’t have any attribute called “contrasts” for which it has any data. There is no contrast matrix stored anywhere explicitly for this factor. However, if we now ask R to tell us what contrasts are set up for this factor, it give us this: \n\n These are the same treatment contrast that we set up manually in Section 16.6. How did R know to set up treatment contrasts, even though I never actually told it anything about what contrasts I wanted? The answer is that R has a hidden list of default “options” that it looks up to resolve situations like this. You can print out all of the options by typing `options()` at the command prompt, but it’s not a very enlightening read. There are a lot of options, and we’re only interested in contrasts right now. Instead of printing out all of the options, we can ask for just one, like this: \n\n```\noptions( \"contrasts\" )\n```\n\n```\n## $contrasts\n## [1] \"contr.treatment\" \"contr.poly\"\n```\n\n What this is telling us is that the default contrasts for unordered factors (i.e., nominal scale variables) are treatment contrasts, and the default for ordered factors (i.e., interval scale variables) are “polynomial” contrasts. I don’t discuss ordered factors much in this book, and so I won’t go into what polynomial contrasts are all about. The key thing is that the `options()` function also allows you to reset these defaults (though only for the current session: they’ll revert to the original settings once you close R). Here’s the command: \n\n```\noptions(contrasts = c(\"contr.helmert\", \"contr.poly\"))\n```\n\nOnce we’ve done this, we can inspect the contrast settings again:\n `options(\"contrasts\")` \n\n```\n## $contrasts\n## [1] \"contr.helmert\" \"contr.poly\"\n```\n\n Now we see that the default contrasts for unordered factors have changed. So if I now ask R to tell me what contrasts are associated with the `drug` factor, it gives a different answer because I changed the default: \n\nThose are Helmert contrasts. In general, if you’re changing the default settings for something in R, it’s a good idea to reset them to their original values once you’re done. So let’s do that:\n\n```\noptions(contrasts = c(\"contr.treatment\", \"contr.poly\"))\n```\n\n### 16.7.5 Setting the contrasts for a single factor\n\n In the previous section, I showed you how to alter the default contrasts. However, suppose that all you really want to do is change the contrasts associated with a single factor, and leave the defaults as they are. To do this, what you need to do is specifically assign the contrast matrix as an “attribute’ of the factor. This is easy to do via the `contrasts()` function. For instance, suppose I wanted to use sum to zero contrasts for the `drug` factor, but keep the default treatment contrasts for everything else. I could do that like so: \n\n```\ncontrasts( clin.trial$drug ) <- contr.sum(3)\n```\n\nAnd if I now inspect the contrasts, I get the following\n\n```\ncontrasts( clin.trial$drug)\n```\n\n However, the contrasts for everything else will still be the defaults. You can check that we have actually made a specific change to the factor itself by checking to see if it now has an attribute, using the command\n\n. This will print out the same output shown above, because the contrast has in fact been attached to the `drug` factor, and does not rely on the defaults. If you want to wipe the attribute and revert the defaults, use a command like this: \n\n```\ncontrasts( clin.trial$drug ) <- NULL\n```\n\n### 16.7.6 Setting the contrasts for a single analysis\n\n One last way of changing contrasts. You might find yourself wanting to change the contrasts only for one specific analysis. That’s allowed too, because the `aov()` and `lm()` functions have a `contrasts` argument that you can use. To change contrasts for one specific analysis, we first set up a list variable that names245 the contrast types that you want to use for each of the factors: \n\n```\nmy.contrasts <- list( drug = contr.helmert, therapy = contr.helmert )\n```\n\n Next, fit the ANOVA model in the usual way, but this time we’ll specify the `contrasts` argument: \n\n```\nmod <- aov( mood.gain ~ drug*therapy, clin.trial, contrasts = my.contrasts )\n```\n\n If you try a command like `summary(aov)` you won’t see any difference in the output because the choice of contrasts does not affect the outcome when you have a balanced design (this won’t always be true later on). However, if you want to check that it has actually worked, you can inspect the value of `mod$contrasts` : `mod$contrasts` \n\n```\n## $drug\n## [,1] [,2]\n## placebo -1 -1\n## anxifree 1 -1\n## joyzepam 0 2\n## \n## $therapy\n## [,1]\n## no.therapy -1\n## CBT 1\n```\n\n As you can see, for the purposes of this one particular ANOVA, R has used Helmert contrasts for both variables. If I had omitted the part of the command that specified the `contrasts` argument, you’d be looking at treatment contrasts here because it would have reverted to whatever values the `contrasts()` function prints out for each of the factors. \n\n## 16.8 Post hoc tests\n\nTime to switch to a different topic. Let’s suppose you’ve done your ANOVA, and it turns out that you obtained some significant effects. Because of the fact that the \\(F\\)-tests are “omnibus” tests that only really test the null hypothesis that there are no differences among groups, obtaining a significant effect doesn’t tell you which groups are different to which other ones. We discussed this issue back in Chapter 14, and in that chapter our solution was to run \\(t\\)-tests for all possible pairs of groups, making corrections for multiple comparisons (e.g., Bonferroni, Holm) to control the Type I error rate across all comparisons. The methods that we used back in Chapter 14 have the advantage of being relatively simple, and being the kind of tools that you can use in a lot of different situations where you’re testing multiple hypotheses, but they’re not necessarily the best choices if you’re interested in doing efficient post hoc testing in an ANOVA context. There are actually quite a lot of different methods for performing multiple comparisons in the statistics literature (Hsu 1996), and it would be beyond the scope of an introductory text like this one to discuss all of them in any detail.\n That being said, there’s one tool that I do want to draw your attention to, namely Tukey’s “Honestly Significant Difference”, or Tukey’s HSD for short. For once, I’ll spare you the formulas, and just stick to the qualitative ideas. The basic idea in Tukey’s HSD is to examine all relevant pairwise comparisons between groups, and it’s only really appropriate to use Tukey’s HSD if it is pairwise differences that you’re interested in.246 For instance, in `model.2` , where we specified a main effect for drug and a main effect of therapy, we would be interested in the following four comparisons: \n\n* The difference in mood gain for people given Anxifree versus people given the placebo.\n * The difference in mood gain for people given Joyzepam versus people given the placebo.\n * The difference in mood gain for people given Anxifree versus people given Joyzepam.\n * The difference in mood gain for people treated with CBT and people given no therapy.\n\nFor any one of these comparisons, we’re interested in the true difference between (population) group means. Tukey’s HSD constructs simultaneous confidence intervals for all four of these comparisons. What we mean by 95% “simultaneous” confidence interval is that there is a 95% probability that all of these confidence intervals contain the relevant true value. Moreover, we can use these confidence intervals to calculate an adjusted \\(p\\) value for any specific comparison.\n The `TukeyHSD()` function in R is pretty easy to use: you simply input the model that you want to run the post hoc tests for. For example, if we were looking to run post hoc tests for `model.2` , here’s the command we would use: `TukeyHSD( model.2 )` \n\n```\n## Tukey multiple comparisons of means\n## 95% family-wise confidence level\n## \n## Fit: aov(formula = mood.gain ~ drug + therapy, data = clin.trial)\n## \n## $drug\n## diff lwr upr p adj\n## anxifree-placebo 0.2666667 -0.1216321 0.6549655 0.2062942\n## joyzepam-placebo 1.0333333 0.6450345 1.4216321 0.0000186\n## joyzepam-anxifree 0.7666667 0.3783679 1.1549655 0.0003934\n## \n## $therapy\n## diff lwr upr p adj\n## CBT-no.therapy 0.3222222 0.0624132 0.5820312 0.0186602\n```\n\nThe output here is (I hope) pretty straightforward. The first comparison, for example, is the Anxifree versus placebo difference, and the first part of the output indicates that the observed difference in group means is \\(.27\\). The next two numbers indicate that the 95% (simultaneous) confidence interval for this comparison runs from \\(-.12\\) to \\(.65\\). Because the confidence interval for the difference includes 0, we cannot reject the null hypothesis that the two group means are identical, and so we’re not all that surprised to see that the adjusted \\(p\\)-value is \\(.21\\). In contrast, if you look at the next line, we see that the observed difference between Joyzepam and the placebo is 1.03, and the 95% confidence interval runs from \\(.64\\) to \\(1.42\\). Because the interval excludes 0, we see that the result is significant \\((p<.001)\\).\n So far, so good. What about the situation where your model includes interaction terms? For instance, in `model.3` we allowed for the possibility that there is an interaction between drug and therapy. If that’s the case, the number of pairwise comparisons that we need to consider starts to increase. As before, we need to consider the three comparisons that are relevant to the main effect of `drug` and the one comparison that is relevant to the main effect of `therapy` . But, if we want to consider the possibility of a significant interaction (and try to find the group differences that underpin that significant interaction), we need to include comparisons such as the following: \n\n* The difference in mood gain for people given Anxifree and treated with CBT, versus people given the placebo and treated with CBT\n * The difference in mood gain for people given Anxifree and given no therapy, versus people given the placebo and given no therapy.\n * etc\n There are quite a lot of these comparisons that you need to consider. So, when we run the `TukeyHSD()` command for `model.3` we see that it has made a lot of pairwise comparisons (19 in total). Here’s the output: `TukeyHSD( model.3 )` \n\n```\n## Tukey multiple comparisons of means\n## 95% family-wise confidence level\n## \n## Fit: aov(formula = mood.gain ~ drug * therapy, data = clin.trial)\n## \n## $drug\n## diff lwr upr p adj\n## anxifree-placebo 0.2666667 -0.09273475 0.6260681 0.1597148\n## joyzepam-placebo 1.0333333 0.67393191 1.3927348 0.0000160\n## joyzepam-anxifree 0.7666667 0.40726525 1.1260681 0.0002740\n## \n## $therapy\n## diff lwr upr p adj\n## CBT-no.therapy 0.3222222 0.08256504 0.5618794 0.012617\n## \n## $`drug:therapy`\n## diff lwr\n## anxifree:no.therapy-placebo:no.therapy 0.10000000 -0.539927728\n## joyzepam:no.therapy-placebo:no.therapy 1.16666667 0.526738939\n## placebo:CBT-placebo:no.therapy 0.30000000 -0.339927728\n## anxifree:CBT-placebo:no.therapy 0.73333333 0.093405606\n## joyzepam:CBT-placebo:no.therapy 1.20000000 0.560072272\n## joyzepam:no.therapy-anxifree:no.therapy 1.06666667 0.426738939\n## placebo:CBT-anxifree:no.therapy 0.20000000 -0.439927728\n## anxifree:CBT-anxifree:no.therapy 0.63333333 -0.006594394\n## joyzepam:CBT-anxifree:no.therapy 1.10000000 0.460072272\n## placebo:CBT-joyzepam:no.therapy -0.86666667 -1.506594394\n## anxifree:CBT-joyzepam:no.therapy -0.43333333 -1.073261061\n## joyzepam:CBT-joyzepam:no.therapy 0.03333333 -0.606594394\n## anxifree:CBT-placebo:CBT 0.43333333 -0.206594394\n## joyzepam:CBT-placebo:CBT 0.90000000 0.260072272\n## joyzepam:CBT-anxifree:CBT 0.46666667 -0.173261061\n## upr p adj\n## anxifree:no.therapy-placebo:no.therapy 0.7399277 0.9940083\n## joyzepam:no.therapy-placebo:no.therapy 1.8065944 0.0005667\n## placebo:CBT-placebo:no.therapy 0.9399277 0.6280049\n## anxifree:CBT-placebo:no.therapy 1.3732611 0.0218746\n## joyzepam:CBT-placebo:no.therapy 1.8399277 0.0004380\n## joyzepam:no.therapy-anxifree:no.therapy 1.7065944 0.0012553\n## placebo:CBT-anxifree:no.therapy 0.8399277 0.8917157\n## anxifree:CBT-anxifree:no.therapy 1.2732611 0.0529812\n## joyzepam:CBT-anxifree:no.therapy 1.7399277 0.0009595\n## placebo:CBT-joyzepam:no.therapy -0.2267389 0.0067639\n## anxifree:CBT-joyzepam:no.therapy 0.2065944 0.2750590\n## joyzepam:CBT-joyzepam:no.therapy 0.6732611 0.9999703\n## anxifree:CBT-placebo:CBT 1.0732611 0.2750590\n## joyzepam:CBT-placebo:CBT 1.5399277 0.0050693\n## joyzepam:CBT-anxifree:CBT 1.1065944 0.2139229\n```\n\nIt looks pretty similar to before, but with a lot more comparisons made.\n\n## 16.9 The method of planned comparisons\n\nOkay, I have a confession to make. I haven’t had time to write this section, but I think the method of planned comparisons is important enough to deserve a quick discussion. In our discussions of multiple comparisons, in the previous section and back in Chapter 14, I’ve been assuming that the tests you want to run are genuinely post hoc. For instance, in our drugs example above, maybe you thought that the drugs would all have different effects on mood (i.e., you hypothesised a main effect of drug), but you didn’t have any specific hypothesis about how they would be different, nor did you have any real idea about which pairwise comparisons would be worth looking at. If that is the case, then you really have to resort to something like Tukey’s HSD to do your pairwise comparisons.\n\nThe situation is rather different, however, if you genuinely did have real, specific hypotheses about which comparisons are of interest, and you never ever have any intention to look at any other comparisons besides the ones that you specified ahead of time. When this is true, and if you honestly and rigourously stick to your noble intentions to not run any other comparisons (even when the data look like they’re showing you deliciously significant effects for stuff you didn’t have a hypothesis test for), then it doesn’t really make a lot of sense to run something like Tukey’s HSD, because it makes corrections for a whole bunch of comparisons that you never cared about and never had any intention of looking at. Under those circumstances, you can safely run a (limited) number of hypothesis tests without making an adjustment for multiple testing. This situation is known as the method of planned comparisons, and it is sometimes used in clinical trials. In a later version of this book, I would like to talk a lot more about planned comparisons.\n\n## 16.10 Factorial ANOVA 3: unbalanced designs\n\nFactorial ANOVA is a very handy thing to know about. It’s been one of the standard tools used to analyse experimental data for many decades, and you’ll find that you can’t read more than two or three papers in psychology without running into an ANOVA in there somewhere. However, there’s one huge difference between the ANOVAs that you’ll see in a lot of real scientific articles and the ANOVA that I’ve just described: in real life, we’re rarely lucky enough to have perfectly balanced designs. For one reason or another, it’s typical to end up with more observations in some cells than in others. Or, to put it another way, we have an unbalanced design.\n\nUnbalanced designs need to be treated with a lot more care than balanced designs, and the statistical theory that underpins them is a lot messier. It might be a consequence of this messiness, or it might be a shortage of time, but my experience has been that undergraduate research methods classes in psychology have a nasty tendency to ignore this issue completely. A lot of stats textbooks tend to gloss over it too. The net result of this, I think, is that a lot of active researchers in the field don’t actually know that there’s several different “types” of unbalanced ANOVAs, and they produce quite different answers. In fact, reading the psychological literature, I’m kind of amazed at the fact that most people who report the results of an unbalanced factorial ANOVA don’t actually give you enough details to reproduce the analysis: I secretly suspect that most people don’t even realise that their statistical software package is making a whole lot of substantive data analysis decisions on their behalf. It’s actually a little terrifying, when you think about it. So, if you want to avoid handing control of your data analysis to stupid software, read on…\n\n### 16.10.1 The coffee data\n\n As usual, it will help us to work with some data. The `coffee.Rdata` file contains a hypothetical data set (the `coffee` data frame) that produces an unbalanced \\(3 \\times 2\\) ANOVA. Suppose we were interested in finding out whether or not the tendency of people to `babble` when they have too much coffee is purely an effect of the coffee itself, or whether there’s some effect of the `milk` and `sugar` that people add to the coffee. Suppose we took 18 people, and gave them some coffee to drink. The amount of coffee / caffeine was held constant, and we varied whether or not milk was added: so `milk` is a binary factor with two levels, `\"yes\"` and `\"no\"` . We also varied the kind of sugar involved. The coffee might contain `\"real\"` sugar, or it might contain `\"fake\"` sugar (i.e., artificial sweetener), or it might contain `\"none\"` at all, so the `sugar` variable is a three level factor. Our outcome variable is a continuous variable that presumably refers to some psychologically sensible measure of the extent to which someone is “babbling”. The details don’t really matter for our purpose. To get a sense of what the data look like, we use the `some()` function in the `car` package. The `some()` function randomly picks a few of the observations in the data frame to print out, which is often very handy: \n\n```\nload(file.path(projecthome, \"data\",\"coffee.Rdata\"))\n some( coffee )\n```\n\n```\n## milk sugar babble\n## 1 yes real 4.6\n## 2 no fake 4.4\n## 3 no fake 3.9\n## 6 no real 5.5\n## 7 yes none 3.9\n## 8 yes none 3.5\n## 9 yes none 3.7\n## 10 no fake 5.6\n## 17 no none 5.3\n## 18 yes fake 5.7\n```\n\n If we use the `aggregate()` function to quickly produce a table of means, we get a strong impression that there are differences between the groups: \n\n```\naggregate( babble ~ milk + sugar, coffee, mean )\n```\n\n```\n## milk sugar babble\n## 1 yes none 3.700\n## 2 no none 5.550\n## 3 yes fake 5.800\n## 4 no fake 4.650\n## 5 yes real 5.100\n## 6 no real 5.875\n```\n\n This is especially true when we compare these means to the standard deviations for the `babble` variable, which you can calculate using `aggregate()` in much the same way. Across groups, this standard deviation varies from .14 to .71, which is fairly small relative to the differences in group means.247 So far, it’s looking like a straightforward factorial ANOVA, just like we did earlier. The problem arises when we check to see how many observations we have in each group: \n\n```\nxtabs( ~ milk + sugar, coffee )\n```\n\n```\n## sugar\n## milk none fake real\n## yes 3 2 3\n## no 2 4 4\n```\n\nThis violates one of our original assumptions, namely that the number of people in each group is the same. We haven’t really discussed how to handle this situation.\n\n### 16.10.2 “Standard ANOVA” does not exist for unbalanced designs\n\nUnbalanced designs lead us to the somewhat unsettling discovery that there isn’t really any one thing that we might refer to as a standard ANOVA. In fact, it turns out that there are three fundamentally different ways248 in which you might want to run an ANOVA in an unbalanced design. If you have a balanced design, all three versions produce identical results, with the sums of squares, \\(F\\)-values etc all conforming to the formulas that I gave at the start of the chapter. However, when your design is unbalanced they don’t give the same answers. Furthermore, they are not all equally appropriate to every situation: some methods will be more appropriate to your situation than others. Given all this, it’s important to understand what the different types of ANOVA are and how they differ from one another.\n\nThe first kind of ANOVA is conventionally referred to as Type I sum of squares. I’m sure you can guess what they other two are called. The “sum of squares” part of the name was introduced by the SAS statistical software package, and has become standard nomenclature, but it’s a bit misleading in some ways. I think the logic for referring to them as different types of sum of squares is that, when you look at the ANOVA tables that they produce, the key difference in the numbers is the SS values. The degrees of freedom don’t change, the MS values are still defined as SS divided by df, etc. However, what the terminology gets wrong is that it hides the reason why the SS values are different from one another. To that end, it’s a lot more helpful to think of the three different kinds of ANOVA as three different hypothesis testing strategies. These different strategies lead to different SS values, to be sure, but it’s the strategy that is the important thing here, not the SS values themselves. Recall from Section 16.5 and 16.6 that any particular \\(F\\)-test is best thought of as a comparison between two linear models. So when you’re looking at an ANOVA table, it helps to remember that each of those \\(F\\)-tests corresponds to a pair of models that are being compared. Of course, this leads naturally to the question of which pair of models is being compared. This is the fundamental difference between ANOVA Types I, II and III: each one corresponds to a different way of choosing the model pairs for the tests.\n\n### 16.10.3 Type I sum of squares\n\n The Type I method is sometimes referred to as the “sequential” sum of squares, because it involves a process of adding terms to the model one at a time. Consider the coffee data, for instance. Suppose we want to run the full \\(3 \\times 2\\) factorial ANOVA, including interaction terms. The full model, as we’ve discussed earlier, is expressed by the R formula\n\n, though we often shorten it by using the `sugar * milk` notation. The Type I strategy builds this model up sequentially, starting from the simplest possible model and gradually adding terms. The simplest possible model for the data would be one in which neither milk nor sugar is assumed to have any effect on babbling. The only term that would be included in such a model is the intercept, and in R formula notation we would write it as `babble ~ 1` . This is our initial null hypothesis. The next simplest model for the data would be one in which only one of the two main effects is included. In the coffee data, there are two different possible choices here, because we could choose to add milk first or to add sugar first (pardon the pun). The order actually turns out to matter, as we’ll see later, but for now let’s just make a choice arbitrarily, and pick sugar. So the second model in our sequence of models is `babble ~ sugar` , and it forms the alternative hypothesis for our first test. We now have our first hypothesis test: \n\n```\nknitr::kable(tibble::tribble(\n ~V1, ~V2,\n \"Null model:\", \"`babble ~ 1`\",\n \"Alternative model:\", \"`babble ~ sugar`\"\n ), col.names= c(\"\", \"\"))\n```\n\nNull model:babble ~ 1Alternative model:babble ~ sugar\n This comparison forms our hypothesis test of the main effect of `sugar` . The next step in our model building exercise it to add the other main effect term, so the next model in our sequence is \n\n. The second hypothesis test is then formed by comparing the following pair of models: \n\nNull model:babble ~ sugarAlternative model:babble ~ sugar + milk\n This comparison forms our hypothesis test of the main effect of `milk` . In one sense, this approach is very elegant: the alternative hypothesis from the first test forms the null hypothesis for the second one. It is in this sense that the Type I method is strictly sequential. Every test builds directly on the results of the last one. However, in another sense it’s very inelegant, because there’s a strong asymmetry between the two tests. The test of the main effect of `sugar` (the first test) completely ignores `milk` , whereas the test of the main effect of `milk` (the second test) does take `sugar` into account. In any case, the fourth model in our sequence is now the full model, \n\n, and the corresponding hypothesis test is \n\nNull model:babble ~ sugar + milkAlternative model:babble ~ sugar + milk + sugar:milk\n Type I sum of squares is the default hypothesis testing method used by the `anova()` function, so it’s easy to produce the results from a Type I analysis. We just type in the same commands that we always did. Since we’ve now reached the point that we don’t need to hide the fact that ANOVA and regression are both linear models, I’ll use the `lm()` function to run the analyses: \n\n```\n## Analysis of Variance Table\n## \n## Response: babble\n## Df Sum Sq Mean Sq F value Pr(>F) \n## sugar 2 3.5575 1.77876 6.7495 0.010863 * \n## milk 1 0.9561 0.95611 3.6279 0.081061 . \n## sugar:milk 2 5.9439 2.97193 11.2769 0.001754 **\n## Residuals 12 3.1625 0.26354 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nLeaving aside for one moment the question of how this result should be interpreted, let’s take note of the fact that our three \\(p\\)-values are \\(.0109\\), \\(.0811\\) and \\(.0018\\) respectively. Next, let’s see if we can replicate the analysis using tools that we’re a little more familiar with. First, let’s fit all four models:\n\n```\nmod.1 <- lm( babble ~ 1, coffee )\n mod.2 <- lm( babble ~ sugar, coffee )\n mod.3 <- lm( babble ~ sugar + milk, coffee )\n mod.4 <- lm( babble ~ sugar + milk + sugar:milk, coffee )\n```\n\n To run the first hypothesis test comparing `mod.1` to `mod.2` we can use the command `anova(mod.1, mod.2)` in much the same way that we did in Section 16.5. Similarly, we can use the commands `anova(mod.2, mod.3)` and `anova(mod.3, mod.4)` and to run the second and third hypothesis tests. However, rather than run each of those commands separately, we can enter the full sequence of models like this: \n\n```\nanova( mod.1, mod.2, mod.3, mod.4 )\n```\n\n```\n## Analysis of Variance Table\n## \n## Model 1: babble ~ 1\n## Model 2: babble ~ sugar\n## Model 3: babble ~ sugar + milk\n## Model 4: babble ~ sugar + milk + sugar:milk\n## Res.Df RSS Df Sum of Sq F Pr(>F) \n## 1 17 13.6200 \n## 2 15 10.0625 2 3.5575 6.7495 0.010863 * \n## 3 14 9.1064 1 0.9561 3.6279 0.081061 . \n## 4 12 3.1625 2 5.9439 11.2769 0.001754 **\n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nThis output is rather more verbose than the last one, but it’s telling essentially the same story.249\n\nThe big problem with using Type I sum of squares is the fact that it really does depend on the order in which you enter the variables. Yet, in many situations the researcher has no reason to prefer one ordering over another. This is presumably the case for our milk and sugar problem. Should we add milk first, or sugar first? It feels exactly as arbitrary as a data analysis question as it does as a coffee-making question. There may in fact be some people with firm opinions about ordering, but it’s hard to imagine a principled answer to the question. Yet, look what happens when we change the ordering:\n\n```\n## Analysis of Variance Table\n## \n## Response: babble\n## Df Sum Sq Mean Sq F value Pr(>F) \n## milk 1 1.4440 1.44400 5.4792 0.037333 * \n## sugar 2 3.0696 1.53482 5.8238 0.017075 * \n## milk:sugar 2 5.9439 2.97193 11.2769 0.001754 **\n## Residuals 12 3.1625 0.26354 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n The \\(p\\)-values for both main effect terms have changed, and fairly dramatically. Among other things, the effect of `milk` has become significant (though one should avoid drawing any strong conclusions about this, as I’ve mentioned previously). Which of these two ANOVAs should one report? It’s not immediately obvious. When you look at the hypothesis tests that are used to define the “first” main effect and the “second” one, it’s clear that they’re qualitatively different from one another. In our initial example, we saw that the test for the main effect of `sugar` completely ignores `milk` , whereas the test of the main effect of `milk` does take `sugar` into account. As such, the Type I testing strategy really does treat the first main effect as if it had a kind of theoretical primacy over the second one. In my experience there is very rarely if ever any theoretically primacy of this kind that would justify treating any two main effects asymmetrically. The consequence of all this is that Type I tests are very rarely of much interest, and so we should move on to discuss Type II tests and Type III tests. However, for the sake of completeness – on the off chance that you ever find yourself needing to run Type I tests – I’ll comment briefly on how R determines the ordering of terms in a Type I test. The key principle in Type I sum of squares is that the hypothesis testing be sequential, with terms added one at a time. However, it does also imply that main effects be added first (e.g., factors `A` , `B` , `C` etc), followed by first order interaction terms (e.g., terms like `A:B` and `B:C` ), then second order interactions (e.g., `A:B:C` ) and so on. Within each “block” you can specify whatever order you like. So, for instance, if we specified our model using a command like this, \n\n and then used `anova(mod)` to produce sequential hypothesis tests, what we’d see is that the main effect terms would be entered `A` then `B` and then `C` , but then the interactions would be entered in the order `B:C` first, then `A:B` and then finally `A:C` . Reordering the terms within each group will change the ordering, as we saw earlier. However, changing the order of terms across blocks has no effect. For instance, if we tried to move the interaction term `B:C` to the front, like this, \n\n it would have no effect. R would still enter the terms in the same order as last time. If for some reason you really, really need an interaction term to be entered first, then you have to do it the long way, creating each model manually using a separate `lm()` command and then using a command like \n\n```\nanova(mod.1, mod.2, mod.3, mod.4)\n```\n\nto force R to enter them in the order that you want. \n\n### 16.10.4 Type III sum of squares\n\n Having just finished talking about Type I tests, you might think that the natural thing to do next would be to talk about Type II tests. However, I think it’s actually a bit more natural to discuss Type III tests (which are simple) before talking about Type II tests (which are trickier). The basic idea behind Type III tests is extremely simple: regardless of which term you’re trying to evaluate, run the \\(F\\)-test in which the alternative hypothesis corresponds to the full ANOVA model as specified by the user, and the null model just deletes that one term that you’re testing. For instance, in the coffee example, in which our full model was\n\n, the test for a main effect of `sugar` would correspond to a comparison between the following two models: \n\nNull model:babble ~ milk + sugar:milkAlternative model:babble ~ sugar + milk + sugar:milk\n Similarly the main effect of `milk` is evaluated by testing the full model against a null model that removes the `milk` term, like so: \n\nNull model:babble ~ sugar + sugar:milkAlternative model:babble ~ sugar + milk + sugar:milk\n Finally, the interaction term `sugar:milk` is evaluated in exactly the same way. Once again, we test the full model against a null model that removes the `sugar:milk` interaction term, like so: \n\nNull model:babble ~ sugar + milkAlternative model:babble ~ sugar + milk + sugar:milk\n The basic idea generalises to higher order ANOVAs. For instance, suppose that we were trying to run an ANOVA with three factors, `A` , `B` and `C` , and we wanted to consider all possible main effects and all possible interactions, including the three way interaction `A:B:C` . The table below shows you what the Type III tests look like for this situation: \n\nTerm being tested isNull model is outcome ~ ...Alternative model is outcome ~ ...AB + C + A:B + A:C + B:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CBA + C + A:B + A:C + B:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CCA + B + A:B + A:C + B:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CA:BA + B + C + A:C + B:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CA:CA + B + C + A:B + B:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CB:CA + B + C + A:B + A:C + A:B:CA + B + C + A:B + A:C + B:C + A:B:CA:B:CA + B + C + A:B + A:C + B:CA + B + C + A:B + A:C + B:C + A:B:C\n As ugly as that table looks, it’s pretty simple. In all cases, the alternative hypothesis corresponds to the full model, which contains three main effect terms (e.g. `A` ), three first order interactions (e.g. `A:B` ) and one second order interaction (i.e., `A:B:C` ). The null model always contains 6 of thes 7 terms: and the missing one is the one whose significance we’re trying to test. At first pass, Type III tests seem like a nice idea. Firstly, we’ve removed the asymmetry that caused us to have problems when running Type I tests. And because we’re now treating all terms the same way, the results of the hypothesis tests do not depend on the order in which we specify them. This is definitely a good thing. However, there is a big problem when interpreting the results of the tests, especially for main effect terms. Consider the coffee data. Suppose it turns out that the main effect of `milk` is not significant according to the Type III tests. What this is telling us is that \n\nis a better model for the data than the full model. But what does that even mean? If the interaction term `sugar:milk` was also non-significant, we’d be tempted to conclude that the data are telling us that the only thing that matters is `sugar` . But suppose we have a significant interaction term, but a non-significant main effect of `milk` . In this case, are we to assume that there really is an “effect of sugar”, an “interaction between milk and sugar”, but no “effect of milk”? That seems crazy. The right answer simply must be that it’s meaningless250 to talk about the main effect if the interaction is significant. In general, this seems to be what most statisticians advise us to do, and I think that’s the right advice. But if it really is meaningless to talk about non-significant main effects in the presence of a significant interaction, then it’s not at all obvious why Type III tests should allow the null hypothesis to rely on a model that includes the interaction but omits one of the main effects that make it up. When characterised in this fashion, the null hypotheses really don’t make much sense at all. Later on, we’ll see that Type III tests can be redeemed in some contexts, but I’d better show you how to actually compute a Type III ANOVA first. The `anova()` function in R does not directly support Type II tests or Type III tests. Technically, you can do it by creating the various models that form the null and alternative hypotheses for each test, and then using `anova()` to compare the models to one another. I outlined the gist of how that would be done when talking about Type I tests, but speaking from first hand experience251 I can tell you that it’s very tedious. In practice, the `anova()` function is only used to produce Type I tests or to compare specific models of particular interest (see Section 16.5). If you want Type II or Type III tests you need to use the `Anova()` function in the `car` package. It’s pretty easy to use, since there’s a `type` argument that you specify. So, to return to our coffee example, our Type III tests are run as follows: \n\n```\nmod <- lm( babble ~ sugar * milk, coffee )\n Anova( mod, type=3 )\n```\n\n```\n## Anova Table (Type III tests)\n## \n## Response: babble\n## Sum Sq Df F value Pr(>F) \n## (Intercept) 41.070 1 155.839 3.11e-08 ***\n## sugar 5.880 2 11.156 0.001830 ** \n## milk 4.107 1 15.584 0.001936 ** \n## sugar:milk 5.944 2 11.277 0.001754 ** \n## Residuals 3.162 12 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n As you can see, I got lazy this time and used `sugar * milk` as a shorthand way of referring to \n\n```\nsugar + milk + sugar:milk\n```\n\n. The important point here is that this is just a regular ANOVA table, and we can see that our Type III tests are significant for all terms, even the intercept. \nExcept, as usual, it’s not that simple. One of the perverse features of the Type III testing strategy is that the results turn out to depend on the contrasts that you use to encode your factors (see Section 16.7 if you’ve forgotten what the different types of contrasts are). The results that I presented in the ANOVA table above are based on the R default, which is treatment contrasts; and as we’ll see later, this is usually a very poor choice if you want to run Type III tests. So let’s see what happens if switch to Helmert contrasts:\n\n```\nmy.contrasts <- list( milk = \"contr.Helmert\", sugar = \"contr.Helmert\" )\n mod.H <- lm( babble ~ sugar * milk, coffee, contrasts = my.contrasts )\n Anova( mod.H, type=3 )\n```\n\n Oh, that’s not good at all. In the case of `milk` in particular, the \\(p\\)-value has changed from .002 to .07. This is a pretty substantial difference, and hopefully it gives you a sense of how important it is that you take care when using Type III tests. Okay, so if the \\(p\\)-values that come out of Type III analyses are so sensitive to the choice of contrasts, does that mean that Type III tests are essentially arbitrary and not to be trusted? To some extent that’s true, and when we turn to a discussion of Type II tests we’ll see that Type II analyses avoid this arbitrariness entirely, but I think that’s too strong a conclusion. Firstly, it’s important to recognise that some choices of contrasts will always produce the same answers. Of particular importance is the fact that if the columns of our contrast matrix are all constrained to sum to zero, then the Type III analysis will always give the same answers. This means that you’ll get the same answers if you use `contr.Helmert` or `contr.sum` or `contr.poly` , but different answers for `contr.treatment` or `contr.SAS` . \n\n```\nrandom.contrasts <- matrix( rnorm(6), 3, 2 ) # create a random matrix\n random.contrasts[, 1] <- random.contrasts[, 1] - mean( random.contrasts[, 1] ) # contrast 1 sums to 0\n random.contrasts[, 2] <- random.contrasts[, 2] - mean( random.contrasts[, 2] ) # contrast 2 sums to 0\n random.contrasts # print it to check that we really have an arbitrary contrast matrix...\n```\n\n```\n## [,1] [,2]\n## [1,] 0.2421359 -0.8168047\n## [2,] 0.5476197 0.9503084\n## [3,] -0.7897556 -0.1335037\n```\n\n```\ncontrasts( coffee$sugar ) <- random.contrasts # random contrasts for sugar\n contrasts( coffee$milk ) <- contr.Helmert(2) # Helmert contrasts for the milk factor \n mod.R <- lm( babble ~ sugar * milk, coffee ) # R will use the contrasts that we assigned\n Anova( mod.R, type = 3 )\n```\n\nYep, same answers.\n\n### 16.10.5 Type II sum of squares\n\nOkay, so we’ve seen Type I and III tests now, and both are pretty straightforward: Type I tests are performed by gradually adding terms one at a time, whereas Type III tests are performed by taking the full model and looking to see what happens when you remove each term. However, both have some serious flaws: Type I tests are dependent on the order in which you enter the terms, and Type III tests are dependent on how you code up your contrasts. Because of these flaws, neither one is easy to interpret. Type II tests are a little harder to describe, but they avoid both of these problems, and as a result they are a little easier to interpret.\n Type II tests are broadly similar to Type III tests: start with a “full” model, and test a particular term by deleting it from that model. However, Type II tests are based on the marginality principle which states that you should not omit a lower order term from your model if there are any higher order ones that depend on it. So, for instance, if your model contains the interaction `A:B` (a 2nd order term), then it really ought to contain the main effects `A` and `B` (1st order terms). Similarly, if it contains a three way interaction term `A:B:C` , then the model must also include the main effects `A` , `B` and `C` as well as the simpler interactions `A:B` , `A:C` and `B:C` . Type III tests routinely violate the marginality principle. For instance, consider the test of the main effect of `A` in the context of a three-way ANOVA that includes all possible interaction terms. According to Type III tests, our null and alternative models are: \n\nNull model:outcome ~ B + C + A:B + A:C + B:C + A:B:CAlternative model:outcome ~ A + B + C + A:B + A:C + B:C + A:B:C\n Notice that the null hypothesis omits `A` , but includes `A:B` , `A:C` and `A:B:C` as part of the model. This, according to the Type II tests, is not a good choice of null hypothesis. What we should do instead, if we want to test the null hypothesis that `A` is not relevant to our `outcome` , is to specify the null hypothesis that is the most complicated model that does not rely on `A` in any form, even as an interaction. The alternative hypothesis corresponds to this null model plus a main effect term of `A` . This is a lot closer to what most people would intuitively think of as a “main effect of `A` ”, and it yields the following as our Type II test of the main effect of `A` .252 \n\nNull model:outcome ~ B + C + B:CAlternative model:outcome ~ A + B + C + B:C\n\nAnyway, just to give you a sense of how the Type II tests play out, here’s the full table of tests that would be applied in a three-way factorial ANOVA:\n\nTerm being tested isNull model is outcome ~ ...Alternative model is outcome ~ ...AB + C + B:CA + B + C + B:CBA + C + A:CA + B + C + A:CCA + B + A:BA + B + C + A:BA:BA + B + C + A:C + B:CA + B + C + A:B + A:C + B:CA:CA + B + C + A:B + B:CA + B + C + A:B + A:C + B:CB:CA + B + C + A:B + A:CA + B + C + A:B + A:C + B:CA:B:CA + B + C + A:B + A:C + B:CA + B + C + A:B + A:C + B:C + A:B:C\n In the context of the two way ANOVA that we’ve been using in the coffee data, the hypothesis tests are even simpler. The main effect of `sugar` corresponds to an \\(F\\)-test comparing these two models: \n\nNull model:babble ~ milkAlternative model:babble ~ sugar + milk\n The test for the main effect of `milk` is \n\nNull model:babble ~ sugarAlternative model:babble ~ sugar + milk\n Finally, the test for the interaction `sugar:milk` is: \n\nNull model:babble ~ sugar + milkAlternative model:babble ~ sugar + milk + sugar:milk\n Running the tests are again straightforward. We use the `Anova()` function, specifying `type=2` : \n\n```\nmod <- lm( babble ~ sugar*milk, coffee )\n Anova( mod, type = 2 )\n```\n\n```\n## Anova Table (Type II tests)\n## \n## Response: babble\n## Sum Sq Df F value Pr(>F) \n## sugar 3.0696 2 5.8238 0.017075 * \n## milk 0.9561 1 3.6279 0.081061 . \n## sugar:milk 5.9439 2 11.2769 0.001754 **\n## Residuals 3.1625 12 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\nType II tests have some clear advantages over Type I and Type III tests. They don’t depend on the order in which you specify factors (unlike Type I), and they don’t depend on the contrasts that you use to specify your factors (unlike Type III). And although opinions may differ on this last point, and it will definitely depend on what you’re trying to do with your data, I do think that the hypothesis tests that they specify are more likely to correspond to something that you actually care about. As a consequence, I find that it’s usually easier to interpret the results of a Type II test than the results of a Type I or Type III test. For this reason, my tentative advice is that, if you can’t think of any obvious model comparisons that directly map onto your research questions but you still want to run an ANOVA in an unbalanced design, Type II tests are probably a better choice than Type I or Type III.253\n\n### 16.10.6 Effect sizes (and non-additive sums of squares)\n\n The `etaSquared()` function in the `lsr` package computes \\(\\eta^2\\) and partial \\(\\eta^2\\) values for unbalanced designs and for different Types of tests. It’s pretty straightforward. All you have to do is indicate which `type` of tests you’re doing, \n\n```\netaSquared( mod, type=2 )\n```\n\n```\n## eta.sq eta.sq.part\n## sugar 0.22537682 0.4925493\n## milk 0.07019886 0.2321436\n## sugar:milk 0.43640732 0.6527155\n```\n\n and out pops the \\(\\eta^2\\) and partial \\(\\eta^2\\) values, as requested. However, when you’ve got an unbalanced design, there’s a bit of extra complexity involved. To see why, let’s expand the output from the `etaSquared()` function so that it displays the full ANOVA table: \n\n```\nes <- etaSquared( mod, type=2, anova=TRUE )\n es\n```\n\n```\n## eta.sq eta.sq.part SS df MS F\n## sugar 0.22537682 0.4925493 3.0696323 2 1.5348161 5.823808\n## milk 0.07019886 0.2321436 0.9561085 1 0.9561085 3.627921\n## sugar:milk 0.43640732 0.6527155 5.9438677 2 2.9719339 11.276903\n## Residuals 0.23219530 NA 3.1625000 12 0.2635417 NA\n## p\n## sugar 0.017075099\n## milk 0.081060698\n## sugar:milk 0.001754333\n## Residuals NA\n```\n\nOkay, if you remember back to our very early discussions of ANOVA, one of the key ideas behind the sums of squares calculations is that if we add up all the SS terms associated with the effects in the model, and add that to the residual SS, they’re supposed to add up to the total sum of squares. And, on top of that, the whole idea behind \\(\\eta^2\\) is that – because you’re dividing one of the SS terms by the total SS value – is that an \\(\\eta^2\\) value can be interpreted as the proportion of variance accounted for by a particular term.\n\nNow take a look at the output above. Because I’ve included the \\(\\eta^2\\) value associated with the residuals (i.e., proportion of variance in the outcome attributed to the residuals, rather than to one of the effects), you’d expect all the \\(\\eta^2\\) values to sum to 1. Because, the whole idea here was that the variance in the outcome variable can be divided up into the variability attributable to the model, and the variability in the residuals. Right? Right? And yet when we add up the \\(\\eta^2\\) values for our model…\n `sum( es[,\"eta.sq\"] )` `## [1] 0.9641783` \n… we discover that for Type II and Type III tests they generally don’t sum to 1. Some of the variability has gone “missing”. It’s not being attributed to the model, and it’s not being attributed to the residuals either. What’s going on here?\n\nBefore giving you the answer, I want to push this idea a little further. From a mathematical perspective, it’s easy enough to see that the missing variance is a consequence of the fact that in Types II and III, the individual SS values are not obliged to the total sum of squares, and will only do so if you have balanced data. I’ll explain why this happens and what it means in a second, but first let’s verify that this is the case using the ANOVA table. First, we can calculate the total sum of squares directly from the raw data:\n\n```\nss.tot <- sum( (coffee$babble - mean(coffee$babble))^2 )\n ss.tot\n```\n\n `## [1] 13.62` \nNext, we can read off all the SS values from one of our Type I ANOVA tables, and add them up. As you can see, this gives us the same answer, just like it’s supposed to:\n\n```\ntype.I.sum <- 3.5575 + 0.9561 + 5.9439 + 3.1625\n type.I.sum\n```\n\n `## [1] 13.62` \nHowever, when we do the same thing for the Type II ANOVA table, it turns out that the SS values in the table add up to slightly less than the total SS value:\n\n```\ntype.II.sum <- 0.9561 + 3.0696 + 5.9439 + 3.1625\n type.II.sum\n```\n\n `## [1] 13.1321` \nSo, once again, we can see that there’s a little bit of variance that has “disappeared” somewhere.\n\nOkay, time to explain what’s happened. The reason why this happens is that, when you have unbalanced designs, your factors become correlated with one another, and it becomes difficult to tell the difference between the effect of Factor A and the effect of Factor B. In the extreme case, suppose that we’d run a \\(2 \\times 2\\) design in which the number of participants in each group had been as follows:\n\nsugar | no sugar |\n | --- | --- |\nmilk | 100 | 0 |\nno milk | 0 | 100 |\n\nHere we have a spectacularly unbalanced design: 100 people have milk and sugar, 100 people have no milk and no sugar, and that’s all. There are 0 people with milk and no sugar, and 0 people with sugar but no milk. Now suppose that, when we collected the data, it turned out there is a large (and statistically significant) difference between the “milk and sugar” group and the “no-milk and no-sugar” group. Is this a main effect of sugar? A main effect of milk? Or an interaction? It’s impossible to tell, because the presence of sugar has a perfect association with the presence of milk. Now suppose the design had been a little more balanced:\n\nsugar | no sugar |\n | --- | --- |\nmilk | 100 | 5 |\nno milk | 5 | 100 |\n\nThis time around, it’s technically possible to distinguish between the effect of milk and the effect of sugar, because we have a few people that have one but not the other. However, it will still be pretty difficult to do so, because the association between sugar and milk is still extremely strong, and there are so few observations in two of the groups. Again, we’re very likely to be in the situation where we know that the predictor variables (milk and sugar) are related to the outcome (babbling), but we don’t know if the nature of that relationship is a main effect of one predictor, or the other predictor or the interaction.\n\nThis uncertainty is the reason for the missing variance. The “missing” variance corresponds to variation in the outcome variable that is clearly attributable to the predictors, but we don’t know which of the effects in the model is responsible. When you calculate Type I sum of squares, no variance ever goes missing: the sequentiual nature of Type I sum of squares means that the ANOVA automatically attributes this variance to whichever effects are entered first. However, the Type II and Type III tests are more conservative. Variance that cannot be clearly attributed to a specific effect doesn’t get attributed to any of them, and it goes missing.\n\n## 16.11 Summary\n\n* Factorial ANOVA with balanced designs, without interactions (Section 16.1) and with interactions included (Section 16.2)\n * Effect size, estimated means, and confidence intervals in a factorial ANOVA (Section 16.3)\n * Understanding the linear model underling ANOVA (Sections 16.5, 16.6 and 16.7)\n * Post hoc testing using Tukey’s HSD (Section 16.8), and a brief commentary on planned comparisons (Section 16.9)\n * Factorial ANOVA with unbalanced designs (Section 16.10)\n\nThe R command we need is\n\n```\nxtabs(~ drug+gender, clin.trial)\n```\n\nThe nice thing about the subscript notation is that generalises nicely: if our experiment had involved a third factor, then we could just add a third subscript. In principle, the notation extends to as many factors as you might care to include, but in this book we’ll rarely consider analyses involving more than two factors, and never more than three.↩\n * \n\nTechnically, marginalising isn’t quite identical to a regular mean: it’s a weighted average, where you take into account the frequency of the different events that you’re averaging over. However, in a balanced design, all of our cell frequencies are equal by definition, so the two are equivalent. We’ll discuss unbalanced designs later, and when we do so you’ll see that all of our calculations become a real headache. But let’s ignore this for now.↩\n * \n\nEnglish translation: “least tedious”.↩\n * \n\nThis chapter seems to be setting a new record for the number of different things that the letter R can stand for: so far we have R referring to the software package, the number of rows in our table of means, the residuals in the model, and now the correlation coefficient in a regression. Sorry: we clearly don’t have enough letters in the alphabet. However, I’ve tried pretty hard to be clear on which thing R is referring to in each case.↩\n * \n\nImplausibly large, I would think: the artificiality of this data set is really starting to show!↩\n * \n\nIn fact, there’s a function\n `Effect()` within the `effects` package that has slightly different arguments, but computes the same things, and won’t give you this warning message.↩ * \n\nDue to the way that the\n `leveneTest()` function is implemented, however, if you use a formula like \n\n```\nmood.gain \\~\\ drug + therapy + drug:therapy\n```\n\n, or input an ANOVA object based on a formula like this, you actually get the error message. That shouldn’t happen, because this actually is a fully crossed model. However, there’s a quirky shortcut in the way that the `leveneTest()` function checks whether your model is fully crossed that means that it doesn’t recognise this as a fully crossed model. Essentially what the function is doing is checking that you used `*` (which ensures that the model is fully crossed), and not `+` or `:` in your model formula. So if you’ve manually typed out all of the relevant terms for a fully crossed model, the `leveneTest()` function doesn’t detect it. I think this is a bug.↩ * \n\nThere could be all sorts of reasons for doing this, I would imagine.↩\n * \n\nThis is cheating in some respects: because ANOVA and regression are provably the same thing, R is lazy: if you read the help documentation closely, you’ll notice that the\n `aov()` function is actually just the `lm()` function in disguise! But we shan’t let such things get in the way of our story, shall we?↩ * \n\nIn the example given above, I’ve typed\n\n```\nsummary( regression.model )\n```\n\nto get the hypothesis tests. However, the `summary()` function does produce a lot of output, which is why I’ve used the `BLAH BLAH BLAH` text to hide the unnecessary parts of the output. But in fact, you can use the `coef()` function to do the same job. If you the command \n\n```\ncoef( summary( regression.model ))\n```\n\nyou’ll get exactly the same output that I’ve shown above (minus the `BLAH BLAH BLAH` ). Compare and contrast this to the output of \n\n```\ncoef( regression.model )\n```\n\nAdvanced users may want to look into the\n `model.matrix()` function, which produces similar output. Alternatively, you can use a command like \n\n```\ncontr.treatment(3)[clin.trial\\$drug,]\n```\n\n. I’ll talk about the `contr.treatment()` function later.↩ * \n\nFuture versions of this book will try to be a bit more consistent with the naming scheme for variables. One of the many problems with having to write a lengthy text very quickly to meet a teaching deadline is that you lose some internal consistency.↩\n * \n\nThe\n `lsr` package contains a more general function called `permuteLevels()` that can shuffle them in any way you like.↩ * \n\nTechnically, this list actually stores the functions themselves. R allows lists to contain functions, which is really neat for advanced purposes, but not something that matters for this book.↩\n * \n\nIf, for instance, you actually would find yourself interested to know if Group A is significantly different from the mean of Group B and Group C, then you need to use a different tool (e.g., Scheffe’s method, which is more conservative, and beyond the scope this book). However, in most cases you probably are interested in pairwise group differences, so Tukey’s HSD is a pretty useful thing to know about.↩\n * \n\nThis discrepancy in standard deviations might (and should) make you wonder if we have a violation of the homogeneity of variance assumption. I’ll leave it as an exercise for the reader to check this using the\n `leveneTest()` function.↩ * \n\nActually, this is a bit of a lie. ANOVAs can vary in other ways besides the ones I’ve discussed in this book. For instance, I’ve completely ignored the difference between fixed-effect models, in which the levels of a factor are “fixed” by the experimenter or the world, and random-effect models, in which the levels are random samples from a larger population of possible levels (this book only covers fixed-effect models). Don’t make the mistake of thinking that this book – or any other one – will tell you “everything you need to know” about statistics, any more than a single book could possibly tell you everything you need to know about psychology, physics or philosophy. Life is too complicated for that to ever be true. This isn’t a cause for despair, though. Most researchers get by with a basic working knowledge of ANOVA that doesn’t go any further than this book does. I just want you to keep in mind that this book is only the beginning of a very long story, not the whole story.↩\n * \n\nThe one thing that might seem a little opaque to some people is why the residual degrees of freedom in this output look different from one another (i.e., ranging from 12 to 17) whereas in the original one the residual degrees of freedom is fixed at 12. It’s actually the case that R uses a residual df of 12 in all cases (that’s why the \\(p\\) values are the same in the two outputs, and it’s enough to verify that\n\n```\npf(6.7495, 2,12, lower.tail=FALSE))\n```\n\ngives the correct answer of \\(p=.010863\\), for instance, whereas \n\n```\npf(6.7495, 2,15, lower.tail=FALSE))\n```\n\nwould have given a \\(p\\)-value of about \\(.00812\\). It’s the residual degrees of freedom in the full model (i.e., the last one) that matters here.↩ * \n\nOr, at the very least, rarely of interest.↩\n * \n\nYes, I’m actually a big enough nerd that I’ve written my own functions implementing Type II tests and Type III tests. I only did it to convince myself that I knew how the different Types of test worked, but it did turn out to be a handy exercise: the\n `etaSquared()` function in the `lsr` package relies on it. There’s actually even an argument in the `etaSquared()` function called `anova` . By default, `anova=FALSE` and the function just prints out the effect sizes. However, if you set `anova=TRUE` it will spit out the full ANOVA table as well. This works for Types I, II and III. Just set the `types` argument to select which type of test you want.↩ * \n\nNote, of course, that this does depend on the model that the user specified. If original ANOVA model doesn’t contain an interaction term for\n `B:C` , then obviously it won’t appear in either the null or the alternative. But that’s true for Types I, II and III. They never include any terms that you didn’t include, but they make different choices about how to construct tests for the ones that you did include.↩ * \n\nI find it amusing to note that the default in R is Type I and the default in SPSS is Type III (with Helmert contrasts). Neither of these appeals to me all that much. Relatedly, I find it depressing that almost nobody in the psychological literature ever bothers to report which Type of tests they ran, much less the order of variables (for Type I) or the contrasts used (for Type III). Often they don’t report what software they used either. The only way I can ever make any sense of what people typically report is to try to guess from auxiliary cues which software they were using, and to assume that they never changed the default settings. Please don’t do this… now that you know about these issues, make sure you indicate what software you used, and if you’re reporting ANOVA results for unbalanced data, then specify what Type of tests you ran, specify order information if you’ve done Type I tests and specify contrasts if you’ve done Type III tests. Or, even better, do hypotheses tests that correspond to things you really care about, and then report those!↩\n\n# Chapter 17 Bayesian statistics\n\nIn our reasonings concerning matter of fact, there are all imaginable degrees of assurance, from the highest certainty to the lowest species of moral evidence. A wise man, therefore, proportions his belief to the evidence. – .\n\nThe ideas I’ve presented to you in this book describe inferential statistics from the frequentist perspective. I’m not alone in doing this. In fact, almost every textbook given to undergraduate psychology students presents the opinions of the frequentist statistician as the theory of inferential statistics, the one true way to do things. I have taught this way for practical reasons. The frequentist view of statistics dominated the academic field of statistics for most of the 20th century, and this dominance is even more extreme among applied scientists. It was and is current practice among psychologists to use frequentist methods. Because frequentist methods are ubiquitous in scientific papers, every student of statistics needs to understand those methods, otherwise they will be unable to make sense of what those papers are saying! Unfortunately – in my opinion at least – the current practice in psychology is often misguided, and the reliance on frequentist methods is partly to blame. In this chapter I explain why I think this, and provide an introduction to Bayesian statistics, an approach that I think is generally superior to the orthodox approach.\n\nThis chapter comes in two parts. In Sections 17.1 through 17.3 I talk about what Bayesian statistics are all about, covering the basic mathematical rules for how it works as well as an explanation for why I think the Bayesian approach is so useful. Afterwards, I provide a brief overview of how you can do Bayesian versions of chi-square tests (Section 17.6), \\(t\\)-tests (Section 17.7), regression (Section 17.8) and ANOVA (Section 17.9).\n\n## 17.1 Probabilistic reasoning by rational agents\n\nFrom a Bayesian perspective, statistical inference is all about belief revision. I start out with a set of candidate hypotheses \\(h\\) about the world. I don’t know which of these hypotheses is true, but do I have some beliefs about which hypotheses are plausible and which are not. When I observe the data \\(d\\), I have to revise those beliefs. If the data are consistent with a hypothesis, my belief in that hypothesis is strengthened. If the data inconsistent with the hypothesis, my belief in that hypothesis is weakened. That’s it! At the end of this section I’ll give a precise description of how Bayesian reasoning works, but first I want to work through a simple example in order to introduce the key ideas. Consider the following reasoning problem:\n I’m carrying an umbrella. Do you think it will rain?\n\nIn this problem, I have presented you with a single piece of data (\\(d =\\) I’m carrying the umbrella), and I’m asking you to tell me your beliefs about whether it’s raining. You have two possible hypotheses, \\(h\\): either it rains today or it does not. How should you solve this problem?\n\n### 17.1.1 Priors: what you believed before\n\nThe first thing you need to do ignore what I told you about the umbrella, and write down your pre-existing beliefs about rain. This is important: if you want to be honest about how your beliefs have been revised in the light of new evidence, then you must say something about what you believed before those data appeared! So, what might you believe about whether it will rain today? You probably know that I live in Australia, and that much of Australia is hot and dry. And in fact you’re right: the city of Adelaide where I live has a Mediterranean climate, very similar to southern California, southern Europe or northern Africa. I’m writing this in January, and so you can assume it’s the middle of summer. In fact, you might have decided to take a quick look on Wikipedia255 and discovered that Adelaide gets an average of 4.4 days of rain across the 31 days of January. Without knowing anything else, you might conclude that the probability of January rain in Adelaide is about 15%, and the probability of a dry day is 85%. If this is really what you believe about Adelaide rainfall (and now that I’ve told it to you, I’m betting that this really is what you believe) then what I have written here is your prior distribution, written \\(P(h)\\):\n\nHypothesis | Degree of Belief |\n | --- | --- |\nRainy day | 0.15 |\nDry day | 0.85 |\n\n### 17.1.2 Likelihoods: theories about the data\n\nTo solve the reasoning problem, you need a theory about my behaviour. When does Dan carry an umbrella? You might guess that I’m not a complete idiot,256 and I try to carry umbrellas only on rainy days. On the other hand, you also know that I have young kids, and you wouldn’t be all that surprised to know that I’m pretty forgetful about this sort of thing. Let’s suppose that on rainy days I remember my umbrella about 30% of the time (I really am awful at this). But let’s say that on dry days I’m only about 5% likely to be carrying an umbrella. So you might write out a little table like this:\n\nHypothesis | Umbrella | No umbrella |\n | --- | --- | --- |\nRainy day | 0.30 | 0.70 |\nDry day | 0.05 | 0.95 |\n\nIt’s important to remember that each cell in this table describes your beliefs about what data \\(d\\) will be observed, given the truth of a particular hypothesis \\(h\\). This “conditional probability” is written \\(P(d|h)\\), which you can read as “the probability of \\(d\\) given \\(h\\)”. In Bayesian statistics, this is referred to as likelihood of data \\(d\\) given hypothesis \\(h\\).257\n\n### 17.1.3 The joint probability of data and hypothesis\n\nAt this point, all the elements are in place. Having written down the priors and the likelihood, you have all the information you need to do Bayesian reasoning. The question now becomes, how do we use this information? As it turns out, there’s a very simple equation that we can use here, but it’s important that you understand why we use it, so I’m going to try to build it up from more basic ideas.\n\nLet’s start out with one of the rules of probability theory. I listed it way back in Table 9.1, but I didn’t make a big deal out of it at the time and you probably ignored it. The rule in question is the one that talks about the probability that two things are true. In our example, you might want to calculate the probability that today is rainy (i.e., hypothesis \\(h\\) is true) and I’m carrying an umbrella (i.e., data \\(d\\) is observed). The joint probability of the hypothesis and the data is written \\(P(d,h)\\), and you can calculate it by multiplying the prior \\(P(h)\\) by the likelihood \\(P(d|h)\\). Mathematically, we say that: \\[ P(d,h) = P(d|h) P(h) \\]\n\nSo, what is the probability that today is a rainy day and I remember to carry an umbrella? As we discussed earlier, the prior tells us that the probability of a rainy day is 15%, and the likelihood tells us that the probability of me remembering my umbrella on a rainy day is 30%. So the probability that both of these things are true is calculated by multiplying the two:\n\n\\[ \\begin{array} P(\\mbox{rainy}, \\mbox{umbrella}) & = & P(\\mbox{umbrella} | \\mbox{rainy}) \\times P(\\mbox{rainy}) \\\\ & = & 0.30 \\times 0.15 \\\\ & = & 0.045 \\end{array} \\]\n\nIn other words, before being told anything about what actually happened, you think that there is a 4.5% probability that today will be a rainy day and that I will remember an umbrella. However, there are of course four possible things that could happen, right? So let’s repeat the exercise for all four. If we do that, we end up with the following table:\n\nUmbrella | No-umbrella |\n | --- | --- |\nRainy | 0.045 | 0.105 |\nDry | 0.0425 | 0.8075 |\n\nThis table captures all the information about which of the four possibilities are likely. To really get the full picture, though, it helps to add the row totals and column totals. That gives us this table:\n\nUmbrella | No-umbrella | Total |\n | --- | --- | --- |\nRainy | 0.0450 | 0.1050 | 0.15 |\nDry | 0.0425 | 0.8075 | 0.85 |\nTotal | 0.0875 | 0.9125 | 1 |\n\nThis is a very useful table, so it’s worth taking a moment to think about what all these numbers are telling us. First, notice that the row sums aren’t telling us anything new at all. For example, the first row tells us that if we ignore all this umbrella business, the chance that today will be a rainy day is 15%. That’s not surprising, of course: that’s our prior. The important thing isn’t the number itself: rather, the important thing is that it gives us some confidence that our calculations are sensible! Now take a look at the column sums, and notice that they tell us something that we haven’t explicitly stated yet. In the same way that the row sums tell us the probability of rain, the column sums tell us the probability of me carrying an umbrella. Specifically, the first column tells us that on average (i.e., ignoring whether it’s a rainy day or not), the probability of me carrying an umbrella is 8.75%. Finally, notice that when we sum across all four logically-possible events, everything adds up to 1. In other words, what we have written down is a proper probability distribution defined over all possible combinations of data and hypothesis.\n\nNow, because this table is so useful, I want to make sure you understand what all the elements correspond to, and how they written:\n\nUmbrella | No-umbrella |\n | --- | --- |\nRainy | \\(P\\)(Umbrella, Rainy) | \\(P\\)(No-umbrella, Rainy) | \\(P\\)(Rainy) |\nDry | \\(P\\)(Umbrella, Dry) | \\(P\\)(No-umbrella, Dry) | \\(P\\)(Dry) |\n\\(P\\)(Umbrella) | \\(P\\)(No-umbrella) |\n\nFinally, let’s use “proper” statistical notation. In the rainy day problem, the data corresponds to the observation that I do or do not have an umbrella. So we’ll let \\(d_1\\) refer to the possibility that you observe me carrying an umbrella, and \\(d_2\\) refers to you observing me not carrying one. Similarly, \\(h_1\\) is your hypothesis that today is rainy, and \\(h_2\\) is the hypothesis that it is not. Using this notation, the table looks like this:\n\n\\(d_1\\) | \\(d_2\\) |\n | --- | --- |\n\\(h_1\\) | \\(P(h_1, d_1)\\) | \\(P(h_1, d_2)\\) | \\(P(h_1)\\) |\n\\(h_2\\) | \\(P(h_2, d_1)\\) | \\(P(h_2, d_2)\\) | \\(P(h_2)\\) |\n\\(P(d_1)\\) | \\(P(d_2)\\) |\n\n### 17.1.4 Updating beliefs using Bayes’ rule\n\nThe table we laid out in the last section is a very powerful tool for solving the rainy day problem, because it considers all four logical possibilities and states exactly how confident you are in each of them before being given any data. It’s now time to consider what happens to our beliefs when we are actually given the data. In the rainy day problem, you are told that I really am carrying an umbrella. This is something of a surprising event: according to our table, the probability of me carrying an umbrella is only 8.75%. But that makes sense, right? A guy carrying an umbrella on a summer day in a hot dry city is pretty unusual, and so you really weren’t expecting that. Nevertheless, the problem tells you that it is true. No matter how unlikely you thought it was, you must now adjust your beliefs to accommodate the fact that you now know that I have an umbrella.258 To reflect this new knowledge, our revised table must have the following numbers:\n\nIn other words, the facts have eliminated any possibility of “no umbrella”, so we have to put zeros into any cell in the table that implies that I’m not carrying an umbrella. Also, you know for a fact that I am carrying an umbrella, so the column sum on the left must be 1 to correctly describe the fact that \\(P(\\mbox{umbrella})=1\\).\n\nWhat two numbers should we put in the empty cells? Again, let’s not worry about the maths, and instead think about our intuitions. When we wrote out our table the first time, it turned out that those two cells had almost identical numbers, right? We worked out that the joint probability of “rain and umbrella” was 4.5%, and the joint probability of “dry and umbrella” was 4.25%. In other words, before I told you that I am in fact carrying an umbrella, you’d have said that these two events were almost identical in probability, yes? But notice that both of these possibilities are consistent with the fact that I actually am carrying an umbrella. From the perspective of these two possibilities, very little has changed. I hope you’d agree that it’s still true that these two possibilities are equally plausible. So what we expect to see in our final table is some numbers that preserve the fact that “rain and umbrella” is slightly more plausible than “dry and umbrella”, while still ensuring that numbers in the table add up. Something like this, perhaps?\n\nWhat this table is telling you is that, after being told that I’m carrying an umbrella, you believe that there’s a 51.4% chance that today will be a rainy day, and a 48.6% chance that it won’t. That’s the answer to our problem! The posterior probability of rain \\(P(h|d)\\) given that I am carrying an umbrella is 51.4%\n\nHow did I calculate these numbers? You can probably guess. To work out that there was a 0.514 probability of “rain”, all I did was take the 0.045 probability of “rain and umbrella” and divide it by the 0.0875 chance of “umbrella”. This produces a table that satisfies our need to have everything sum to 1, and our need not to interfere with the relative plausibility of the two events that are actually consistent with the data. To say the same thing using fancy statistical jargon, what I’ve done here is divide the joint probability of the hypothesis and the data \\(P(d,h)\\) by the marginal probability of the data \\(P(d)\\), and this is what gives us the posterior probability of the hypothesis given that we know the data have been observed. To write this as an equation:259 \\[ P(h | d) = \\frac{P(d,h)}{P(d)} \\]\n\nHowever, remember what I said at the start of the last section, namely that the joint probability \\(P(d,h)\\) is calculated by multiplying the prior \\(P(h)\\) by the likelihood \\(P(d|h)\\). In real life, the things we actually know how to write down are the priors and the likelihood, so let’s substitute those back into the equation. This gives us the following formula for the posterior probability:\n\n\\[ P(h | d) = \\frac{P(d|h) P(h)}{P(d)} \\]\n\nAnd this formula, folks, is known as Bayes’ rule. It describes how a learner starts out with prior beliefs about the plausibility of different hypotheses, and tells you how those beliefs should be revised in the face of data. In the Bayesian paradigm, all statistical inference flows from this one simple rule.\n\n## 17.2 Bayesian hypothesis tests\n\nIn Chapter 11 I described the orthodox approach to hypothesis testing. It took an entire chapter to describe, because null hypothesis testing is a very elaborate contraption that people find very hard to make sense of. In contrast, the Bayesian approach to hypothesis testing is incredibly simple. Let’s pick a setting that is closely analogous to the orthodox scenario. There are two hypotheses that we want to compare, a null hypothesis \\(h_0\\) and an alternative hypothesis \\(h_1\\). Prior to running the experiment we have some beliefs \\(P(h)\\) about which hypotheses are true. We run an experiment and obtain data \\(d\\). Unlike frequentist statistics Bayesian statistics does allow to talk about the probability that the null hypothesis is true. Better yet, it allows us to calculate the posterior probability of the null hypothesis, using Bayes’ rule:\n\n\\[ P(h_0 | d) = \\frac{P(d|h_0) P(h_0)}{P(d)} \\]\n\nThis formula tells us exactly how much belief we should have in the null hypothesis after having observed the data \\(d\\). Similarly, we can work out how much belief to place in the alternative hypothesis using essentially the same equation. All we do is change the subscript:\n\n\\[ P(h_1 | d) = \\frac{P(d|h_1) P(h_1)}{P(d)} \\]\n\nIt’s all so simple that I feel like an idiot even bothering to write these equations down, since all I’m doing is copying Bayes rule from the previous section.260\n\n### 17.2.1 The Bayes factor\n\nIn practice, most Bayesian data analysts tend not to talk in terms of the raw posterior probabilities \\(P(h_0|d)\\) and \\(P(h_1|d)\\). Instead, we tend to talk in terms of the posterior odds ratio. Think of it like betting. Suppose, for instance, the posterior probability of the null hypothesis is 25%, and the posterior probability of the alternative is 75%. The alternative hypothesis is three times as probable as the null, so we say that the odds are 3:1 in favour of the alternative. Mathematically, all we have to do to calculate the posterior odds is divide one posterior probability by the other:\n\n\\[ \\frac{P(h_1 | d)}{P(h_0 | d)} = \\frac{0.75}{0.25} = 3 \\]\n\nOr, to write the same thing in terms of the equations above: \\[ \\frac{P(h_1 | d)}{P(h_0 | d)} = \\frac{P(d|h_1)}{P(d|h_0)} \\times \\frac{P(h_1)}{P(h_0)} \\]\n\nActually, this equation is worth expanding on. There are three different terms here that you should know. On the left hand side, we have the posterior odds, which tells you what you believe about the relative plausibilty of the null hypothesis and the alternative hypothesis after seeing the data. On the right hand side, we have the prior odds, which indicates what you thought before seeing the data. In the middle, we have the Bayes factor, which describes the amount of evidence provided by the data: \\[ \\begin{array}{ccccc}\\displaystyle \\frac{P(h_1 | d)}{P(h_0 | d)} &=& \\displaystyle\\frac{P(d|h_1)}{P(d|h_0)} &\\times& \\displaystyle\\frac{P(h_1)}{P(h_0)} \\\\[6pt] \\\\[-2pt] \\uparrow && \\uparrow && \\uparrow \\\\[6pt] \\mbox{Posterior odds} && \\mbox{Bayes factor} && \\mbox{Prior odds} \\end{array} \\] The Bayes factor (sometimes abbreviated as BF) has a special place in the Bayesian hypothesis testing, because it serves a similar role to the \\(p\\)-value in orthodox hypothesis testing: it quantifies the strength of evidence provided by the data, and as such it is the Bayes factor that people tend to report when running a Bayesian hypothesis test. The reason for reporting Bayes factors rather than posterior odds is that different researchers will have different priors. Some people might have a strong bias to believe the null hypothesis is true, others might have a strong bias to believe it is false. Because of this, the polite thing for an applied researcher to do is report the Bayes factor. That way, anyone reading the paper can multiply the Bayes factor by their own personal prior odds, and they can work out for themselves what the posterior odds would be. In any case, by convention we like to pretend that we give equal consideration to both the null hypothesis and the alternative, in which case the prior odds equals 1, and the posterior odds becomes the same as the Bayes factor.\n\n### 17.2.2 Interpreting Bayes factors\n\nOne of the really nice things about the Bayes factor is the numbers are inherently meaningful. If you run an experiment and you compute a Bayes factor of 4, it means that the evidence provided by your data corresponds to betting odds of 4:1 in favour of the alternative. However, there have been some attempts to quantify the standards of evidence that would be considered meaningful in a scientific context. The two most widely used are from Jeffreys (1961) and Kass and Raftery (1995). Of the two, I tend to prefer the Kass and Raftery (1995) table because it’s a bit more conservative. So here it is:\n\nBayes factor | Interpretation |\n | --- | --- |\n1 - 3 | Negligible evidence |\n3 - 20 | Positive evidence |\n20 - 150 | Strong evidence |\n$>$150 | Very strong evidence |\n\nAnd to be perfectly honest, I think that even the Kass and Raftery standards are being a bit charitable. If it were up to me, I’d have called the “positive evidence” category “weak evidence”. To me, anything in the range 3:1 to 20:1 is “weak” or “modest” evidence at best. But there are no hard and fast rules here: what counts as strong or weak evidence depends entirely on how conservative you are, and upon the standards that your community insists upon before it is willing to label a finding as “true”.\n\nIn any case, note that all the numbers listed above make sense if the Bayes factor is greater than 1 (i.e., the evidence favours the alternative hypothesis). However, one big practical advantage of the Bayesian approach relative to the orthodox approach is that it also allows you to quantify evidence for the null. When that happens, the Bayes factor will be less than 1. You can choose to report a Bayes factor less than 1, but to be honest I find it confusing. For example, suppose that the likelihood of the data under the null hypothesis \\(P(d|h_0)\\) is equal to 0.2, and the corresponding likelihood \\(P(d|h_0)\\) under the alternative hypothesis is 0.1. Using the equations given above, Bayes factor here would be:\n\n\\[ \\mbox{BF} = \\frac{P(d|h_1)}{P(d|h_0)} = \\frac{0.1}{0.2} = 0.5 \\]\n\nRead literally, this result tells is that the evidence in favour of the alternative is 0.5 to 1. I find this hard to understand. To me, it makes a lot more sense to turn the equation “upside down”, and report the amount op evidence in favour of the null. In other words, what we calculate is this:\n\n\\[ \\mbox{BF}^\\prime = \\frac{P(d|h_0)}{P(d|h_1)} = \\frac{0.2}{0.1} = 2 \\]\n\nAnd what we would report is a Bayes factor of 2:1 in favour of the null. Much easier to understand, and you can interpret this using the table above.\n\n## 17.3 Why be a Bayesian?\n\nUp to this point I’ve focused exclusively on the logic underpinning Bayesian statistics. We’ve talked about the idea of “probability as a degree of belief”, and what it implies about how a rational agent should reason about the world. The question that you have to answer for yourself is this: how do you want to do your statistics? Do you want to be an orthodox statistician, relying on sampling distributions and \\(p\\)-values to guide your decisions? Or do you want to be a Bayesian, relying on Bayes factors and the rules for rational belief revision? And to be perfectly honest, I can’t answer this question for you. Ultimately it depends on what you think is right. It’s your call, and your call alone. That being said, I can talk a little about why I prefer the Bayesian approach.\n\n### 17.3.1 Statistics that mean what you think they mean\n\nYou keep using that word. I do not think it means what you think it means\n – , 261 \nTo me, one of the biggest advantages to the Bayesian approach is that it answers the right questions. Within the Bayesian framework, it is perfectly sensible and allowable to refer to “the probability that a hypothesis is true”. You can even try to calculate this probability. Ultimately, isn’t that what you want your statistical tests to tell you? To an actual human being, this would seem to be the whole point of doing statistics: to determine what is true and what isn’t. Any time that you aren’t exactly sure about what the truth is, you should use the language of probability theory to say things like “there is an 80% chance that Theory A is true, but a 20% chance that Theory B is true instead”.\n\nThis seems so obvious to a human, yet it is explicitly forbidden within the orthodox framework. To a frequentist, such statements are a nonsense because “the theory is true” is not a repeatable event. A theory is true or it is not, and no probabilistic statements are allowed, no matter how much you might want to make them. There’s a reason why, back in Section 11.5, I repeatedly warned you not to interpret the \\(p\\)-value as the probability of that the null hypothesis is true. There’s a reason why almost every textbook on statstics is forced to repeat that warning. It’s because people desperately want that to be the correct interpretation. Frequentist dogma notwithstanding, a lifetime of experience of teaching undergraduates and of doing data analysis on a daily basis suggests to me that most actual humans thing that “the probability that the hypothesis is true” is not only meaningful, it’s the thing we care most about. It’s such an appealing idea that even trained statisticians fall prey to the mistake of trying to interpret a \\(p\\)-value this way. For example, here is a quote from an official Newspoll report in 2013, explaining how to interpret their (frequentist) data analysis:262\n Throughout the report, where relevant, statistically significant changes have been noted. All significance tests have been based on the 95 percent level of confidence. This means that if a change is noted as being statistically significant, there is a 95 percent probability that a real change has occurred, and is not simply due to chance variation. (emphasis added)\n\nNope! That’s not what \\(p<.05\\) means. That’s not what 95% confidence means to a frequentist statistician. The bolded section is just plain wrong. Orthodox methods cannot tell you that “there is a 95% chance that a real change has occurred”, because this is not the kind of event to which frequentist probabilities may be assigned. To an ideological frequentist, this sentence should be meaningless. Even if you’re a more pragmatic frequentist, it’s still the wrong definition of a \\(p\\)-value. It is simply not an allowed or correct thing to say if you want to rely on orthodox statistical tools.\n\nOn the other hand, let’s suppose you are a Bayesian. Although the bolded passage is the wrong definition of a \\(p\\)-value, it’s pretty much exactly what a Bayesian means when they say that the posterior probability of the alternative hypothesis is greater than 95%. And here’s the thing. If the Bayesian posterior is actually thing you want to report, why are you even trying to use orthodox methods? If you want to make Bayesian claims, all you have to do is be a Bayesian and use Bayesian tools.\n\nSpeaking for myself, I found this to be a the most liberating thing about switching to the Bayesian view. Once you’ve made the jump, you no longer have to wrap your head around counterinuitive definitions of \\(p\\)-values. You don’t have to bother remembering why you can’t say that you’re 95% confident that the true mean lies within some interval. All you have to do is be honest about what you believed before you ran the study, and then report what you learned from doing it. Sounds nice, doesn’t it? To me, this is the big promise of the Bayesian approach: you do the analysis you really want to do, and express what you really believe the data are telling you.\n\n## 17.4 Evidentiary standards you can believe\n\nIf [\\(p\\)] is below .02 it is strongly indicated that the [null] hypothesis fails to account for the whole of the facts. We shall not often be astray if we draw a conventional line at .05 and consider that [smaller values of \\(p\\)] indicate a real discrepancy.\n – (1925) \nConsider the quote above by , one of the founders of what has become the orthodox approach to statistics. If anyone has ever been entitled to express an opinion about the intended function of \\(p\\)-values, it’s Fisher. In this passage, taken from his classic guide Statistical Methods for Research Workers, he’s pretty clear about what it means to reject a null hypothesis at \\(p<.05\\). In his opinion, if we take \\(p<.05\\) to mean there is “a real effect”, then “we shall not often be astray”. This view is hardly unusual: in my experience, most practitioners express views very similar to Fisher’s. In essence, the \\(p<.05\\) convention is assumed to represent a fairly stringent evidentiary standard.\n\nWell, how true is that? One way to approach this question is to try to convert \\(p\\)-values to Bayes factors, and see how the two compare. It’s not an easy thing to do because a \\(p\\)-value is a fundamentally different kind of calculation to a Bayes factor, and they don’t measure the same thing. However, there have been some attempts to work out the relationship between the two, and it’s somewhat surprising. For example, Johnson (2013) presents a pretty compelling case that (for \\(t\\)-tests at least) the \\(p<.05\\) threshold corresponds roughly to a Bayes factor of somewhere between 3:1 and 5:1 in favour of the alternative. If that’s right, then Fisher’s claim is a bit of a stretch. Let’s suppose that the null hypothesis is true about half the time (i.e., the prior probability of \\(H_0\\) is 0.5), and we use those numbers to work out the posterior probability of the null hypothesis given that it has been rejected at \\(p<.05\\). Using the data from Johnson (2013), we see that if you reject the null at \\(p<.05\\), you’ll be correct about 80% of the time. I don’t know about you, but in my opinion an evidentiary standard that ensures you’ll be wrong on 20% of your decisions isn’t good enough. The fact remains that, quite contrary to Fisher’s claim, if you reject at \\(p<.05\\) you shall quite often go astray. It’s not a very stringent evidentiary threshold at all.\n\n## 17.5 The \\(p\\)-value is a lie.\n\nThe cake is a lie.\n The cake is a lie. The cake is a lie. The cake is a lie. – Portal263 \nOkay, at this point you might be thinking that the real problem is not with orthodox statistics, just the \\(p<.05\\) standard. In one sense, that’s true. The recommendation that Johnson (2013) gives is not that “everyone must be a Bayesian now”. Instead, the suggestion is that it would be wiser to shift the conventional standard to something like a \\(p<.01\\) level. That’s not an unreasonable view to take, but in my view the problem is a little more severe than that. In my opinion, there’s a fairly big problem built into the way most (but not all) orthodox hypothesis tests are constructed. They are grossly naive about how humans actually do research, and because of this most \\(p\\)-values are wrong.\n\nSounds like an absurd claim, right? Well, consider the following scenario. You’ve come up with a really exciting research hypothesis and you design a study to test it. You’re very diligent, so you run a power analysis to work out what your sample size should be, and you run the study. You run your hypothesis test and out pops a \\(p\\)-value of 0.072. Really bloody annoying, right?\n\nWhat should you do? Here are some possibilities:\n\n* You conclude that there is no effect, and try to publish it as a null result\n * You guess that there might be an effect, and try to publish it as a “borderline significant” result\n * You give up and try a new study\n * You collect some more data to see if the \\(p\\) value goes up or (preferably!) drops below the “magic” criterion of \\(p<.05\\)\n\nWhich would you choose? Before reading any further, I urge you to take some time to think about it. Be honest with yourself. But don’t stress about it too much, because you’re screwed no matter what you choose. Based on my own experiences as an author, reviewer and editor, as well as stories I’ve heard from others, here’s what will happen in each case:\n\nLet’s start with option 1. If you try to publish it as a null result, the paper will struggle to be published. Some reviewers will think that \\(p=.072\\) is not really a null result. They’ll argue it’s borderline significant. Other reviewers will agree it’s a null result, but will claim that even though some null results are publishable, yours isn’t. One or two reviewers might even be on your side, but you’ll be fighting an uphill battle to get it through.\n * \n\nOkay, let’s think about option number 2. Suppose you try to publish it as a borderline significant result. Some reviewers will claim that it’s a null result and should not be published. Others will claim that the evidence is ambiguous, and that you should collect more data until you get a clear significant result. Again, the publication process does not favour you.\n * \n\nGiven the difficulties in publishing an “ambiguous” result like \\(p=.072\\), option number 3 might seem tempting: give up and do something else. But that’s a recipe for career suicide. If you give up and try a new project else every time you find yourself faced with ambiguity, your work will never be published. And if you’re in academia without a publication record you can lose your job. So that option is out.\n * \n\nIt looks like you’re stuck with option 4. You don’t have conclusive results, so you decide to collect some more data and re-run the analysis. Seems sensible, but unfortunately for you, if you do this all of your \\(p\\)-values are now incorrect. All of them. Not just the \\(p\\)-values that you calculated for this study. All of them. All the \\(p\\)-values you calculated in the past and all the \\(p\\)-values you will calculate in the future. Fortunately, no-one will notice. You’ll get published, and you’ll have lied.\n\nWait, what? How can that last part be true? I mean, it sounds like a perfectly reasonable strategy doesn’t it? You collected some data, the results weren’t conclusive, so now what you want to do is collect more data until the the results are conclusive. What’s wrong with that?\n\nHonestly, there’s nothing wrong with it. It’s a reasonable, sensible and rational thing to do. In real life, this is exactly what every researcher does. Unfortunately, the theory of null hypothesis testing as I described it in Chapter 11 forbids you from doing this.264 The reason is that the theory assumes that the experiment is finished and all the data are in. And because it assumes the experiment is over, it only considers two possible decisions. If you’re using the conventional \\(p<.05\\) threshold, those decisions are:\n\nOutcome | Action |\n | --- | --- |\n\\(p\\) less than .05 | Reject the null |\n\\(p\\) greater than .05 | Retain the null |\n\nWhat you’re doing is adding a third possible action to the decision making problem. Specifically, what you’re doing is using the \\(p\\)-value itself as a reason to justify continuing the experiment. And as a consequence you’ve transformed the decision-making procedure into one that looks more like this:\n\nOutcome | Action |\n | --- | --- |\n\\(p\\) less than .05 | Stop the experiment and reject the null |\n\\(p\\) between .05 and .1 | Continue the experiment |\n\\(p\\) greater than .1 | Stop the experiment and retain the null |\n\nThe “basic” theory of null hypothesis testing isn’t built to handle this sort of thing, not in the form I described back in Chapter 11. If you’re the kind of person who would choose to “collect more data” in real life, it implies that you are not making decisions in accordance with the rules of null hypothesis testing. Even if you happen to arrive at the same decision as the hypothesis test, you aren’t following the decision process it implies, and it’s this failure to follow the process that is causing the problem.265 Your \\(p\\)-values are a lie.\n\nWorse yet, they’re a lie in a dangerous way, because they’re all too small. To give you a sense of just how bad it can be, consider the following (worst case) scenario. Imagine you’re a really super-enthusiastic researcher on a tight budget who didn’t pay any attention to my warnings above. You design a study comparing two groups. You desperately want to see a significant result at the \\(p<.05\\) level, but you really don’t want to collect any more data than you have to (because it’s expensive). In order to cut costs, you start collecting data, but every time a new observation arrives you run a \\(t\\)-test on your data. If the \\(t\\)-tests says \\(p<.05\\) then you stop the experiment and report a significant result. If not, you keep collecting data. You keep doing this until you reach your pre-defined spending limit for this experiment. Let’s say that limit kicks in at \\(N=1000\\) observations. As it turns out, the truth of the matter is that there is no real effect to be found: the null hypothesis is true. So, what’s the chance that you’ll make it to the end of the experiment and (correctly) conclude that there is no effect? In an ideal world, the answer here should be 95%. After all, the whole point of the \\(p<.05\\) criterion is to control the Type I error rate at 5%, so what we’d hope is that there’s only a 5% chance of falsely rejecting the null hypothesis in this situation. However, there’s no guarantee that will be true. You’re breaking the rules: you’re running tests repeatedly, “peeking” at your data to see if you’ve gotten a significant result, and all bets are off.\n\nSo how bad is it? The answer is shown as the solid black line in Figure 17.1, and it’s astoundingly bad. If you peek at your data after every single observation, there is a 49% chance that you will make a Type I error. That’s, um, quite a bit bigger than the 5% that it’s supposed to be. By way of comparison, imagine that you had used the following strategy. Start collecting data. Every single time an observation arrives, run a Bayesian \\(t\\)-test (Section 17.7 and look at the Bayes factor. I’ll assume that Johnson (2013) is right, and I’ll treat a Bayes factor of 3:1 as roughly equivalent to a \\(p\\)-value of .05.266 This time around, our trigger happy researcher uses the following procedure: if the Bayes factor is 3:1 or more in favour of the null, stop the experiment and retain the null. If it is 3:1 or more in favour of the alternative, stop the experiment and reject the null. Otherwise continue testing. Now, just like last time, let’s assume that the null hypothesis is true. What happens? As it happens, I ran the simulations for this scenario too, and the results are shown as the dashed line in Figure 17.1. It turns out that the Type I error rate is much much lower than the 49% rate that we were getting by using the orthodox \\(t\\)-test.\n\nIn some ways, this is remarkable. The entire point of orthodox null hypothesis testing is to control the Type I error rate. Bayesian methods aren’t actually designed to do this at all. Yet, as it turns out, when faced with a “trigger happy” researcher who keeps running hypothesis tests as the data come in, the Bayesian approach is much more effective. Even the 3:1 standard, which most Bayesians would consider unacceptably lax, is much safer than the \\(p<.05\\) rule.\n\n### 17.5.1 Is it really this bad?\n\nThe example I gave in the previous section is a pretty extreme situation. In real life, people don’t run hypothesis tests every time a new observation arrives. So it’s not fair to say that the \\(p<.05\\) threshold “really” corresponds to a 49% Type I error rate (i.e., \\(p=.49\\)). But the fact remains that if you want your \\(p\\)-values to be honest, then you either have to switch to a completely different way of doing hypothesis tests, or you must enforce a strict rule: no peeking. You are not allowed to use the data to decide when to terminate the experiment. You are not allowed to look at a “borderline” \\(p\\)-value and decide to collect more data. You aren’t even allowed to change your data analyis strategy after looking at data. You are strictly required to follow these rules, otherwise the \\(p\\)-values you calculate will be nonsense.\n\nAnd yes, these rules are surprisingly strict. As a class exercise a couple of years back, I asked students to think about this scenario. Suppose you started running your study with the intention of collecting \\(N=80\\) people. When the study starts out you follow the rules, refusing to look at the data or run any tests. But when you reach \\(N=50\\) your willpower gives in… and you take a peek. Guess what? You’ve got a significant result! Now, sure, you know you said that you’d keep running the study out to a sample size of \\(N=80\\), but it seems sort of pointless now, right? The result is significant with a sample size of \\(N=50\\), so wouldn’t it be wasteful and inefficient to keep collecting data? Aren’t you tempted to stop? Just a little? Well, keep in mind that if you do, your Type I error rate at \\(p<.05\\) just ballooned out to 8%. When you report \\(p<.05\\) in your paper, what you’re really saying is \\(p<.08\\). That’s how bad the consequences of “just one peek” can be.\n\nNow consider this … the scientific literature is filled with \\(t\\)-tests, ANOVAs, regressions and chi-square tests. When I wrote this book I didn’t pick these tests arbitrarily. The reason why these four tools appear in most introductory statistics texts is that these are the bread and butter tools of science. None of these tools include a correction to deal with “data peeking”: they all assume that you’re not doing it. But how realistic is that assumption? In real life, how many people do you think have “peeked” at their data before the experiment was finished and adapted their subsequent behaviour after seeing what the data looked like? Except when the sampling procedure is fixed by an external constraint, I’m guessing the answer is “most people have done it”. If that has happened, you can infer that the reported \\(p\\)-values are wrong. Worse yet, because we don’t know what decision process they actually followed, we have no way to know what the \\(p\\)-values should have been. You can’t compute a \\(p\\)-value when you don’t know the decision making procedure that the researcher used. And so the reported \\(p\\)-value remains a lie.\n\nGiven all of the above, what is the take home message? It’s not that Bayesian methods are foolproof. If a researcher is determined to cheat, they can always do so. Bayes’ rule cannot stop people from lying, nor can it stop them from rigging an experiment. That’s not my point here. My point is the same one I made at the very beginning of the book in Section 1.1: the reason why we run statistical tests is to protect us from ourselves. And the reason why “data peeking” is such a concern is that it’s so tempting, even for honest researchers. A theory for statistical inference has to acknowledge this. Yes, you might try to defend \\(p\\)-values by saying that it’s the fault of the researcher for not using them properly. But to my mind that misses the point. A theory of statistical inference that is so completely naive about humans that it doesn’t even consider the possibility that the researcher might look at their own data isn’t a theory worth having. In essence, my point is this:\n Good laws have their origins in bad morals.\n – 267 \nGood rules for statistical testing have to acknowledge human frailty. None of us are without sin. None of us are beyond temptation. A good system for statistical inference should still work even when it is used by actual humans. Orthodox null hypothesis testing does not.268\n\n## 17.6 Bayesian analysis of contingency tables\n\nTime to change gears. Up to this point I’ve been talking about what Bayesian inference is and why you might consider using it. I now want to briefly describe how to do Bayesian versions of various statistical tests. The discussions in the next few sections are not as detailed as I’d like, but I hope they’re enough to help you get started. So let’s begin.\n The first kind of statistical inference problem I discussed in this book appeared in Chapter 12, in which we discussed categorical data analysis problems. In that chapter I talked about several different statistical problems that you might be interested in, but the one that appears most often in real life is the analysis of contingency tables. In this kind of data analysis situation, we have a cross-tabulation of one variable against another one, and the goal is to find out if there is some association between these variables. The data set I used to illustrate this problem is found in the `chapek9.Rdata` file, and it contains a single data frame `chapek9` \n\n```\nload(file.path(projecthome, \"data\",\"chapek9.Rdata\"))\nhead(chapek9)\n```\n\n In this data set, we supposedly sampled 180 beings and measured two things. First, we checked whether they were humans or robots, as captured by the `species` variable. Second, we asked them to nominate whether they most preferred flowers, puppies, or data. When we produce the cross-tabulation, we get this as the results: \n\n```\ncrosstab <- xtabs( ~ species + choice, chapek9 )\ncrosstab\n```\n\nSurprisingly, the humans seemed to show a much stronger preference for data than the robots did. At the time we speculated that this might have been because the questioner was a large robot carrying a gun, and the humans might have been scared.\n\n### 17.6.1 The orthodox text\n\n Just to refresh your memory, here’s how we analysed these data back in Chapter@refch:chisquare. Because we want to determine if there is some association between `species` and `choice` , we used the `associationTest()` function in the `lsr` package to run a chi-square test of association. The results looked like this: \n\n```\nlibrary(lsr)\nassociationTest( ~species + choice, chapek9 )\n```\n\nBecause we found a small \\(p\\) value (in this case \\(p<.01\\)), we concluded that the data are inconsistent with the null hypothesis of no association, and we rejected it.\n\n### 17.6.2 The Bayesian test\n\n How do we run an equivalent test as a Bayesian? Well, like every other bloody thing in statistics, there’s a lot of different ways you could do it. However, for the sake of everyone’s sanity, throughout this chapter I’ve decided to rely on one R package to do the work. Specifically, I’m going to use the `BayesFactor` package written by and , which as of this writing is in version 0.9.10. For the analysis of contingency tables, the `BayesFactor` package contains a function called `contingencyTableBF()` . The data that you need to give to this function is the contingency table itself (i.e., the `crosstab` variable above), so you might be expecting to use a command like this: \n\n```\nlibrary( BayesFactor ) # ...because we have to load the package\ncontingencyTableBF( crosstab ) # ...because that makes sense, right?\n```\n\n However, if you try this you’ll get an error message. This is because the `contingencyTestBF()` function needs one other piece of information from you: it needs to know what sampling plan you used to run your experiment. You can specify the sampling plan using the `sampleType` argument. So I should probably tell you what your options are! The `contingencyTableBF()` function distinguishes between four different types of experiment: \n\n* Fixed sample size. Suppose that in our\n `chapek9` example, our experiment was designed like this: we deliberately set out to test 180 people, but we didn’t try to control the number of humans or robots, nor did we try to control the choices they made. In this design, the total number of observations \\(N\\) is fixed, but everything else is random. This is referred to as “joint multinomial” sampling, and if that’s what you did you should specify \n\n. In the case of the `chapek9` data, that’s actually what I had in mind when I invented the data set. * Fixed row (or column) totals. A different kind of design might work like this. We decide ahead of time that we want 180 people, but we try to be a little more systematic about it. Specifically, the experimenter constrains it so that we get a predetermined number of humans and robots (e.g., 90 of each). In this design, either the row totals or the column totals are fixed, but not both. This is referred to as “independent multinomial” sampling, and if that’s what you did you should specify\n\n. * Both row and column totals fixed. Another logical possibility is that you designed the experiment so that both the row totals and the column totals are fixed. This doesn’t make any sense at all in the\n `chapek9` example, but there are other deisgns that can work this way. Suppose that I show you a collection of 20 toys, and then given them 10 stickers that say `boy` and another 10 that say `girl` . I then give them 10 `blue` stickers and 10 `pink` stickers. I then ask you to put the stickers on the 20 toys such that every toy has a colour and every toy has a gender. No matter how you assign the stickers, the total number of pink and blue toys will be 10, as will the number of boys and girls. In this design both the rows and columns of the contingency table are fixed. This is referred to as “hypergeometric” sampling, and if that’s what you’ve done you should specify \n\n```\nsampleType = \"hypergeom\"\n```\n\n. * Nothing is fixed. Finally, it might be the case that nothing is fixed. Not the row columns, not the column totals, and not the total sample size either. For instance, in the\n `chapek9` scenario, suppose what I’d done is run the study for a fixed length of time. By chance, it turned out that I got 180 people to turn up to study, but it could easily have been something else. This is referred to as “Poisson” sampling, and if that’s what you’ve done you should specify `sampleType=\"poisson\"` . Okay, so now we have enough knowledge to actually run a test. For the `chapek9` data, I implied that we designed the study such that the total sample size \\(N\\) was fixed, so we should set \n\n. The command that we need is, \n\n```\nlibrary( BayesFactor )\ncontingencyTableBF( crosstab, sampleType = \"jointMulti\" )\n```\n\nAs with most R commands, the output initially looks suspiciously similar to utter gibberish. Fortunately, it’s actually pretty simple once you get past the initial impression. Firstly, note that the stuff at the top and bottom are irrelevant fluff. You already know that you’re doing a Bayes factor analysis. You already know that you’re analysing a contingency table, and you already know that you specified a joint multinomial sampling plan. So let’s strip that out and take a look at what’s left over:\n\n```\n[1] Non-indep. (a=1) : 15.92684 @plusorminus0%\n\nAgainst denominator:\n Null, independence, a = 1\n```\n\n Let’s also ignore those two `a=1` bits, since they’re technical details that you don’t need to know about at this stage.269 The rest of the output is actually pretty straightforward. At the bottom, the output defines the null hypothesis for you: in this case, the null hypothesis is that there is no relationship between `species` and `choice` . Or, to put it another way, the null hypothesis is that these two variables are independent. Now if you look at the line above it, you might (correctly) guess that the `Non-indep.` part refers to the alternative hypothesis. In this case, the alternative is that there is a relationship between `species` and `choice` : that is, they are not independent. So the only thing left in the output is the bit that reads \n\n```\n15.92684 @plusorminus0%\n```\n\nThe 15.9 part is the Bayes factor, and it’s telling you that the odds for the alternative hypothesis against the null are about 16:1. The \\(\\pm0\\%\\) part is not very interesting: essentially, all it’s telling you is that R has calculated an exact Bayes factor, so the uncertainty about the Bayes factor is 0%.270 In any case, the data are telling us that we have moderate evidence for the alternative hypothesis.\n\n### 17.6.3 Writing up the results\n\nWhen writing up the results, my experience has been that there aren’t quite so many “rules” for how you “should” report Bayesian hypothesis tests. That might change in the future if Bayesian methods become standard and some task force starts writing up style guides, but in the meantime I would suggest using some common sense. For example, I would avoid writing this:\n A Bayesian test of association found a significant result (BF=15.92)\n\nTo my mind, this write up is unclear. Even assuming that you’ve already reported the relevant descriptive statistics, there are a number of things I am unhappy with. First, the concept of “statistical significance” is pretty closely tied with \\(p\\)-values, so it reads slightly strangely. Second, the “BF=15.92” part will only make sense to people who already understand Bayesian methods, and not everyone does. Third, it is somewhat unclear exactly which test was run and what software was used to do so.\n\nOn the other hand, unless precision is extremely important, I think that this is taking things a step too far:\n We ran a Bayesian test of association using version 0.9.10-1 of the BayesFactor package using default priors and a joint multinomial sampling plan. The resulting Bayes factor of 15.92 to 1 in favour of the alternative hypothesis indicates that there is moderately strong evidence for the non-independence of species and choice.\n\nEverything about that passage is correct, of course. Rouder (2015) built their Bayesian tests of association using the paper by (1974), the specific test we used assumes that the experiment relied on a joint multinomial sampling plan, and indeed the Bayes factor of 15.92 is moderately strong evidence. It’s just far too wordy.\n\nIn most situations you just don’t need that much information. My preference is usually to go for something a little briefer. First, if you’re reporting multiple Bayes factor analyses in your write up, then somewhere you only need to cite the software once, at the beginning of the results section. So you might have one sentence like this:\n All analyses were conducted using the BayesFactor package in R , and unless otherwise stated default parameter values were used\n\nNotice that I don’t bother including the version number? That’s because the citation itself includes that information (go check my reference list if you don’t believe me). There’s no need to clutter up your results with redundant information that almost no-one will actually need. When you get to the actual test you can get away with this:\n A test of association produced a Bayes factor of 16:1 in favour of a relationship between species and choice.\n Short and sweet. I’ve rounded 15.92 to 16, because there’s not really any important difference between 15.92:1 and 16:1. I spelled out “Bayes factor” rather than truncating it to “BF” because not everyone knows the abbreviation. I indicated exactly what the effect is (i.e., “a relationship between species and choice”) and how strong the evidence was. I didn’t bother indicating whether this was “moderate” evidence or “strong” evidence, because the odds themselves tell you! There’s nothing stopping you from including that information, and I’ve done so myself on occasions, but you don’t strictly need it. Similarly, I didn’t bother to indicate that I ran the “joint multinomial” sampling plan, because I’m assuming that the method section of my write up would make clear how the experiment was designed. (I might change my mind about that if the method section was ambiguous.) Neither did I bother indicating that this was a Bayesian test of association: if your reader can’t work that out from the fact that you’re reporting a Bayes factor and the fact that you’re citing the `BayesFactor` package for all your analyses, then there’s no chance they’ll understand anything you’ve written. Besides, if you keep writing the word “Bayes” over and over again it starts to look stupid. Bayes Bayes Bayes Bayes Bayes. See? \n\n### 17.6.4 Other sampling plans\n\n Up to this point all I’ve shown you is how to use the `contingencyTableBF()` function for the joint multinomial sampling plan (i.e., when the total sample size \\(N\\) is fixed, but nothing else is). For the Poisson sampling plan (i.e., nothing fixed), the command you need is identical except for the `sampleType` argument: \n\n```\ncontingencyTableBF(crosstab, sampleType = \"poisson\" )\n```\n\n```\n## Bayes factor analysis\n## --------------\n## [1] Non-indep. (a=1) : 28.20757 ±0%\n## \n## Against denominator:\n## Null, independence, a = 1 \n## ---\n## Bayes factor type: BFcontingencyTable, poisson\n```\n\nNotice that the Bayes factor of 28:1 here is not the identical to the Bayes factor of 16:1 that we obtained from the last test. The sampling plan actually does matter.\n What about the design in which the row columns (or column totals) are fixed? As I mentioned earlier, this corresponds to the “independent multinomial” sampling plan. Again, you need to specify the `sampleType` argument, but this time you need to specify whether you fixed the rows or the columns. For example, suppose I deliberately sampled 87 humans and 93 robots, then I would need to indicate that the `fixedMargin` of the contingency table is the `\"rows\"` . So the command I would use is: \n\n```\ncontingencyTableBF(crosstab, sampleType = \"indepMulti\", fixedMargin=\"rows\")\n```\n\nAgain, the Bayes factor is different, with the evidence for the alternative dropping to a mere 9:1. As you might expect, the answers would be diffrent again if it were the columns of the contingency table that the experimental design fixed.\n\nFinally, if we turn to hypergeometric sampling in which everything is fixed, we get…\n\n```\ncontingencyTableBF(crosstab, sampleType = \"hypergeom\")\n#Error in contingencyHypergeometric(as.matrix(data2), a) : \n# hypergeometric contingency tables restricted to 2 x 2 tables; see help for contingencyTableBF()\n```\n\n… an error message. Okay, some quick reading through the help files hints that support for larger contingency tables is coming, but it’s not been implemented yet. In the meantime, let’s imagine we have data from the “toy labelling” experiment I described earlier in this section. Specifically, let’s say our data look like this:\n\n```\ntoys <- data.frame(stringsAsFactors=FALSE,\n gender = c(\"girl\", \"boy\"),\n pink = c(8, 2),\n blue = c(2, 8)\n )\n```\n\nThe Bayesian test with hypergeometric sampling gives us this:\n\n```\ncontingencyTableBF(toys, sampleType = \"hypergeom\")\n\n#Bayes factor analysis\n#--------------\n#[1] Non-indep. (a=1) : 8.294321 @plusorminus0%\n#\n#Against denominator:\n# Null, independence, a = 1 \n#---\n#Bayes factor type: BFcontingencyTable, hypergeometric\n```\n\nThe Bayes factor of 8:1 provides modest evidence that the labels were being assigned in a way that correlates gender with colour, but it’s not conclusive.\n\n## 17.7 Bayesian \\(t\\)-tests\n\n The second type of statistical inference problem discussed in this book is the comparison between two means, discussed in some detail in the chapter on \\(t\\)-tests (Chapter 13. If you can remember back that far, you’ll recall that there are several versions of the \\(t\\)-test. The `BayesFactor` package contains a function called `ttestBF()` that is flexible enough to run several different versions of the \\(t\\)-test. I’ll talk a little about Bayesian versions of the independent samples \\(t\\)-tests and the paired samples \\(t\\)-test in this section. \n\n### 17.7.1 Independent samples \\(t\\)-test\n\nThe most common type of \\(t\\)-test is the independent samples \\(t\\)-test, and it arises when you have data that look something like this:\n\n```\nload(file.path(projecthome, \"data\",\"harpo.Rdata\"))\nhead(harpo)\n```\n\n In this data set, we have two groups of students, those who received lessons from Anastasia and those who took their classes with Bernadette. The question we want to answer is whether there’s any difference in the grades received by these two groups of student. Back in Chapter@refch:ttest I suggested you could analyse this kind of data using the\n\nfunction in the `lsr` package. For example, if you want to run a Student’s \\(t\\)-test, you’d use a command like this: \n\n```\nindependentSamplesTTest(\n formula = grade ~ tutor, \n data = harpo, \n var.equal = TRUE \n )\n```\n\n Like most of the functions that I wrote for this book, the\n\nis very wordy. It prints out a bunch of descriptive statistics and a reminder of what the null and alternative hypotheses are, before finally getting to the test results. I wrote it that way deliberately, in order to help make things a little clearer for people who are new to statistics. \nAgain, we obtain a \\(p\\)-value less than 0.05, so we reject the null hypothesis.\n What does the Bayesian version of the \\(t\\)-test look like? Using the `ttestBF()` function, we can obtain a Bayesian analog of Student’s independent samples \\(t\\)-test using the following command: \n\n```\n## Bayes factor analysis\n## --------------\n## [1] Alt., r=0.707 : 1.754927 ±0%\n## \n## Against denominator:\n## Null, mu1-mu2 = 0 \n## ---\n## Bayes factor type: BFindepSample, JZS\n```\n\n Notice that format of this command is pretty standard. As usual we have a `formula` argument in which we specify the outcome variable on the left hand side and the grouping variable on the right. The `data` argument is used to specify the data frame containing the variables. However, notice that there’s no analog of the `var.equal` argument. This is because the `BayesFactor` package does not include an analog of the Welch test, only the Student test.271 In any case, when you run this command you get this as the output: So what does all this mean? Just as we saw with the `contingencyTableBF()` function, the output is pretty dense. But, just like last time, there’s not a lot of information here that you actually need to process. Firstly, let’s examine the bottom line. The `BFindepSample` part just tells you that you ran an independent samples \\(t\\)-test, and the `JZS` part is technical information that is a little beyond the scope of this book.272 Clearly, there’s nothing to worry about in that part. In the line above, the text `Null, mu1-mu2 = 0` is just telling you that the null hypothesis is that there are no differences between means. But you already knew that. So the only part that really matters is this line here: \n\n```\n[1] Alt., r=0.707 : 1.754927 @plusorminus0%\n```\n\n Ignore the `r=0.707` part: it refers to a technical detail that we won’t worry about in this chapter.273 Instead, you should focus on the part that reads `1.754927` . This is the Bayes factor: the evidence provided by these data are about 1.8:1 in favour of the alternative. \nBefore moving on, it’s worth highlighting the difference between the orthodox test results and the Bayesian one. According to the orthodox test, we obtained a significant result, though only barely. Nevertheless, many people would happily accept \\(p=.043\\) as reasonably strong evidence for an effect. In contrast, notice that the Bayesian test doesn’t even reach 2:1 odds in favour of an effect, and would be considered very weak evidence at best. In my experience that’s a pretty typical outcome. Bayesian methods usually require more evidence before rejecting the null.\n\n### 17.7.2 Paired samples \\(t\\)-test\n\n Back in Section 13.5 I discussed the `chico` data frame in which students grades were measured on two tests, and we were interested in finding out whether grades went up from test 1 to test 2. Because every student did both tests, the tool we used to analyse the data was a paired samples \\(t\\)-test. To remind you of what the data look like, here’s the first few cases: \n\n```\nload(file.path(projecthome, \"data\",\"chico.Rdata\"))\nhead(chico)\n```\n\n We originally analysed the data using the `pairedSamplesTTest()` function in the `lsr` package, but this time we’ll use the `ttestBF()` function from the `BayesFactor` package to do the same thing. The easiest way to do it with this data set is to use the `x` argument to specify one variable and the `y` argument to specify the other. All we need to do then is specify `paired=TRUE` to tell R that this is a paired samples test. So here’s our command: \n\n```\nttestBF(\n x = chico$grade_test1,\n y = chico$grade_test2,\n paired = TRUE\n )\n```\n\n```\n## Bayes factor analysis\n## --------------\n## [1] Alt., r=0.707 : 5992.05 ±0%\n## \n## Against denominator:\n## Null, mu = 0 \n## ---\n## Bayes factor type: BFoneSample, JZS\n```\n\nAt this point, I hope you can read this output without any difficulty. The data provide evidence of about 6000:1 in favour of the alternative. We could probably reject the null with some confidence!\n\n## 17.8 Bayesian regression\n\nOkay, so now we’ve seen Bayesian equivalents to orthodox chi-square tests and \\(t\\)-tests. What’s next? If I were to follow the same progression that I used when developing the orthodox tests you’d expect to see ANOVA next, but I think it’s a little clearer if we start with regression.\n\n### 17.8.1 A quick refresher\n\n In Chapter 15 I used the `parenthood` data to illustrate the basic ideas behind regression. To remind you of what that data set looks like, here’s the first six observations: \n\n```\nload(file.path(projecthome, \"data\",\"parenthood.Rdata\"))\nhead(parenthood)\n```\n\n```\n## dan.sleep baby.sleep dan.grump day\n## 1 7.59 10.18 56 1\n## 2 7.91 11.66 60 2\n## 3 5.14 7.92 82 3\n## 4 7.71 9.61 55 4\n## 5 6.68 9.75 67 5\n## 6 5.99 5.04 72 6\n```\n\n Back in Chapter 15 I proposed a theory in which my grumpiness ( `dan.grump` ) on any given day is related to the amount of sleep I got the night before ( `dan.sleep` ), and possibly to the amount of sleep our baby got ( `baby.sleep` ), though probably not to the `day` on which we took the measurement. We tested this using a regression model. In order to estimate the regression model we used the `lm()` function, like so: \n\n The hypothesis tests for each of the terms in the regression model were extracted using the `summary()` function as shown below: `summary(model)` \n\n```\n## \n## Call:\n## lm(formula = dan.grump ~ dan.sleep + day + baby.sleep, data = parenthood)\n## \n## Residuals:\n## Min 1Q Median 3Q Max \n## -10.906 -2.284 -0.295 2.652 11.880 \n## \n## Coefficients:\n## Estimate Std. Error t value Pr(>|t|) \n## (Intercept) 126.278707 3.242492 38.945 <2e-16 ***\n## dan.sleep -8.969319 0.560007 -16.016 <2e-16 ***\n## day -0.004403 0.015262 -0.288 0.774 \n## baby.sleep 0.015747 0.272955 0.058 0.954 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n## \n## Residual standard error: 4.375 on 96 degrees of freedom\n## Multiple R-squared: 0.8163, Adjusted R-squared: 0.8105 \n## F-statistic: 142.2 on 3 and 96 DF, p-value: < 2.2e-16\n```\n\n When interpreting the results, each row in this table corresponds to one of the possible predictors. The `(Intercept)` term isn’t usually interesting, though it is highly significant. The important thing for our purposes is the fact that `dan.sleep` is significant at \\(p<.001\\) and neither of the other variables are. \n\n### 17.8.2 The Bayesian version\n\n Okay, so how do we do the same thing using the `BayesFactor` package? The easiest way is to use the `regressionBF()` function instead of `lm()` . As before, we use `formula` to indicate what the full regression model looks like, and the `data` argument to specify the data frame. So the command is: \n\n So that’s pretty straightforward: it’s exactly what we’ve been doing throughout the book. The output, however, is a little different from what you get from `lm()` . The format of this is pretty familiar. At the bottom we have some techical rubbish, and at the top we have some information about the Bayes factors. What’s new is the fact that we seem to have lots of Bayes factors here. What’s all this about? The trick to understanding this output is to recognise that if we’re interested in working out which of the 3 predictor variables are related to `dan.grump` , there are actually 8 possible regression models that could be considered. One possibility is the intercept only model, in which none of the three variables have an effect. At the other end of the spectrum is the full model in which all three variables matter. So what `regressionBF()` does is treat the intercept only model as the null hypothesis, and print out the Bayes factors for all other models when compared against that null. For example, if we look at line 4 in the table, we see that the evidence is about \\(10^{33}\\) to 1 in favour of the claim that a model that includes both `dan.sleep` and `day` is better than the intercept only model. Or if we look at line 1, we can see that the odds are about \\(1.6 \\times 10^{34}\\) that a model containing the `dan.sleep` variable (but no others) is better than the intercept only model. \n\n### 17.8.3 Finding the best model\n\n In practice, this isn’t super helpful. In most situations the intercept only model is one that you don’t really care about at all. What I find helpful is to start out by working out which model is the best one, and then seeing how well all the alternatives compare to it. Here’s how you do that. In this case, it’s easy enough to see that the best model is actually the one that contains `dan.sleep` only (line 1), because it has the largest Bayes factor. However, if you’ve got a lot of possible models in the output, it’s handy to know that you can use the `head()` function to pick out the best few models. First, we have to go back and save the Bayes factor information to a variable: \n\n Let’s say I want to see the best three models. To do this, I use the `head()` function specifying `n=3` , and here’s what I get as the result: `head( models, n = 3)` \n\n This is telling us that the model in line 1 (i.e.,\n\n) is the best one. That’s almost what I’m looking for, but it’s still comparing all the models against the intercept only model. That seems silly. What I’d like to know is how big the difference is between the best model and the other good models. For that, there’s this trick: \n\n```\nhead( models/max(models), n = 3)\n```\n\n```\n## Bayes factor analysis\n## --------------\n## [1] dan.sleep : 1 ±0%\n## [2] dan.sleep + day : 0.0626532 ±0.01%\n## [3] dan.sleep + baby.sleep : 0.0602154 ±0.01%\n## \n## Against denominator:\n## dan.grump ~ dan.sleep \n## ---\n## Bayes factor type: BFlinearModel, JZS\n```\n\n Notice the bit at the bottom showing that the “denominator” has changed. What that means is that the Bayes factors are now comparing each of those 3 models listed against the\n\nmodel. Obviously, the Bayes factor in the first line is exactly 1, since that’s just comparing the best model to itself. More to the point, the other two Bayes factors are both less than 1, indicating that they’re all worse than that model. The Bayes factors of 0.06 to 1 imply that the odds for the best model over the second best model are about 16:1. You can work this out by simple arithmetic (i.e., \\(0.06 / 1 \\approx 16\\)), but the other way to do it is to directly compare the models. To see what I mean, here’s the original output: `models` \n\nThe best model corresponds to row 1 in this table, and the second best model corresponds to row 4. All you have to do to compare these two models is this:\n\n```\nmodels[1] / models[4]\n```\n\n```\n## Bayes factor analysis\n## --------------\n## [1] dan.sleep : 15.96088 ±0.01%\n## \n## Against denominator:\n## dan.grump ~ dan.sleep + day \n## ---\n## Bayes factor type: BFlinearModel, JZS\n```\n\n And there you have it. You’ve found the regression model with the highest Bayes factor (i.e.,\n\n), and you know that the evidence for that model over the next best alternative (i.e., \n\n) is about 16:1. \n\n### 17.8.4 Extracting Bayes factors for all included terms\n\n Okay, let’s say you’ve settled on a specific regression model. What Bayes factors should you report? In this example, I’m going to pretend that you decided that\n\nis the model you think is best. Sometimes it’s sensible to do this, even when it’s not the one with the highest Bayes factor. Usually this happens because you have a substantive theoretical reason to prefer one model over the other. However, in this case I’m doing it because I want to use a model with more than one predictor as my example! Having figured out which model you prefer, it can be really useful to call the `regressionBF()` function and specifying `whichModels=\"top\"` . You use your “preferred” model as the `formula` argument, and then the output will show you the Bayes factors that result when you try to drop predictors from this model: \n\n```\nregressionBF( \n formula = dan.grump ~ dan.sleep + baby.sleep,\n data = parenthood,\n whichModels = \"top\"\n)\n```\n\n```\n## \n | \n | | 0%\n | \n |====================== | 33%\n | \n |=========================================== | 67%\n | \n |=================================================================| 100%\n```\n\n```\n## Bayes factor top-down analysis\n## --------------\n## When effect is omitted from dan.sleep + baby.sleep , BF is...\n## [1] Omit baby.sleep : 16.60705 ±0.01%\n## [2] Omit dan.sleep : 1.025403e-26 ±0.01%\n## \n## Against denominator:\n## dan.grump ~ dan.sleep + baby.sleep \n## ---\n## Bayes factor type: BFlinearModel, JZS\n```\n\n Okay, so now you can see the results a bit more clearly. The Bayes factor when you try to drop the `dan.sleep` predictor is about \\(10^{-26}\\), which is very strong evidence that you shouldn’t drop it. On the other hand, the Bayes factor actually goes up to 17 if you drop `baby.sleep` , so you’d usually say that’s pretty strong evidence for dropping that one. \n\n## 17.9 Bayesian ANOVA\n\n As you can tell, the `BayesFactor` package is pretty flexible, and it can do Bayesian versions of pretty much everything in this book. In fact, it can do a few other neat things that I haven’t covered in the book at all. However, I have to stop somewhere, and so there’s only one other topic I want to cover: Bayesian ANOVA. \n\n### 17.9.1 A quick refresher\n\n As with the other examples, I think it’s useful to start with a reminder of how I discussed ANOVA earlier in the book. First, let’s remind ourselves of what the data were. The example I used originally is the `clin.trial` data frame, which looks like this \n\n```\nload(file.path(projecthome, \"data\",\"clinicaltrial.Rdata\"))\nhead(clin.trial)\n```\n\n```\n## drug therapy mood.gain\n## 1 placebo no.therapy 0.5\n## 2 placebo no.therapy 0.3\n## 3 placebo no.therapy 0.1\n## 4 anxifree no.therapy 0.6\n## 5 anxifree no.therapy 0.4\n## 6 anxifree no.therapy 0.2\n```\n\n To run our orthodox analysis in earlier chapters we used the `aov()` function to do all the heavy lifting. In Chapter 16 I recommended using the `Anova()` function from the `car` package to produce the ANOVA table, because it uses Type II tests by default. If you’ve forgotten what “Type II tests” are, it might be a good idea to re-read Section 16.10, because it will become relevant again in a moment. In any case, here’s what our analysis looked like: \n\n```\nlibrary(car)\nmodel <- aov( mood.gain ~ drug * therapy, data = clin.trial )\nAnova(model)\n```\n\n```\n## Anova Table (Type II tests)\n## \n## Response: mood.gain\n## Sum Sq Df F value Pr(>F) \n## drug 3.4533 2 31.7143 1.621e-05 ***\n## therapy 0.4672 1 8.5816 0.01262 * \n## drug:therapy 0.2711 2 2.4898 0.12460 \n## Residuals 0.6533 12 \n## ---\n## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1\n```\n\n That’s pretty clearly showing us evidence for a main effect of `drug` at \\(p<.001\\), an effect of `therapy` at \\(p<.05\\) and no interaction. \n\n### 17.9.2 The Bayesian version\n\n How do we do the same thing using Bayesian methods? The `BayesFactor` package contains a function called `anovaBF()` that does this for you. It uses a pretty standard `formula` and `data` structure, so the command should look really familiar. Just like we did with regression, it will be useful to save the output to a variable: \n\n```\nmodels <- anovaBF( \n formula = mood.gain ~ drug * therapy,\n data = clin.trial\n)\n```\n\n```\n## \n | \n | | 0%\n | \n |================ | 25%\n | \n |================================ | 50%\n | \n |================================================= | 75%\n | \n |=================================================================| 100%\n```\n\nThe output is quite different to the traditional ANOVA, but it’s not too bad once you understand what you’re looking for. Let’s take a look:\n `models` This looks very similar to the output we obtained from the `regressionBF()` function, and with good reason. Remember what I said back in Section 16.6: under the hood, ANOVA is no different to regression, and both are just different examples of a linear model. Becasue of this, the `anovaBF()` reports the output in much the same way. For instance, if we want to identify the best model we could use the same commands that we used in the last section. One variant that I find quite useful is this: `models/max(models)` \n\n```\n## Bayes factor analysis\n## --------------\n## [1] drug : 0.3436062 ±1.34%\n## [2] therapy : 0.001022285 ±1.34%\n## [3] drug + therapy : 1 ±0%\n## [4] drug + therapy + drug:therapy : 0.954423 ±1.64%\n## \n## Against denominator:\n## mood.gain ~ drug + therapy \n## ---\n## Bayes factor type: BFlinearModel, JZS\n```\n\n By “dividing” the `models` output by the best model (i.e., `max(models)` ), what R is doing is using the best model (which in this case is `drugs + therapy` ) as the denominator, which gives you a pretty good sense of how close the competitors are. For instance, the model that contains the interaction term is almost as good as the model without the interaction, since the Bayes factor is 0.98. In other words, the data do not clearly indicate whether there is or is not an interaction. \n\n### 17.9.3 Constructing Bayesian Type II tests\n\nOkay, that’s all well and good, you might be thinking, but what do I report as the alternative to the \\(p\\)-value? In the classical ANOVA table, you get a single \\(p\\)-value for every predictor in the model, so you can talk about the significance of each effect. What’s the Bayesian analog of this?\n\nIt’s a good question, but the answer is tricky. Remember what I said in Section 16.10 about ANOVA being complicated. Even in the classical version of ANOVA there are several different “things” that ANOVA might correspond to. Specifically, I discussed how you get different \\(p\\)-values depending on whether you use Type I tests, Type II tests or Type III tests. To work out which Bayes factor is analogous to “the” \\(p\\)-value in a classical ANOVA, you need to work out which version of ANOVA you want an analog for. For the purposes of this section, I’ll assume you want Type II tests, because those are the ones I think are most sensible in general. As I discussed back in Section 16.10, Type II tests for a two-way ANOVA are reasonably straightforward, but if you have forgotten that section it wouldn’t be a bad idea to read it again before continuing.\n Assuming you’ve had a refresher on Type II tests, let’s have a look at how to pull them from the Bayes factor table. Suppose we want to test the main effect of `drug` . The null hypothesis for this test corresponds to a model that includes an effect of `therapy` , but no effect of `drug` . The alternative hypothesis is the model that includes both. In other words, what we want is the Bayes factor corresponding to this comparison: \n\nAs it happens, we can read the answer to this straight off the table because it corresponds to a comparison between the model in line 2 of the table and the model in line 3: the Bayes factor in this case represents evidence for the null of 0.001 to 1. Or, more helpfully, the odds are about 1000 to 1 against the null.\n The main effect of `therapy` can be calculated in much the same way. In this case, the null model is the one that contains only an effect of drug, and the alternative is the model that contains both. So the relevant comparison is between lines 2 and 1 in the table. The odds in favour of the null here are only 0.35 to 1. Again, I find it useful to frame things the other way around, so I’d refer to this as evidence of about 3 to 1 in favour of an effect of `therapy` . \nFinally, in order to test an interaction effect, the null model here is one that contains both main effects but no interaction. The alternative model adds the interaction. That is:\n\nIf we look those two models up in the table, we see that this comparison is between the models on lines 3 and 4 of the table. The odds of 0.98 to 1 imply that these two models are fairly evenly matched.\n You might be thinking that this is all pretty laborious, and I’ll concede that’s true. At some stage I might consider adding a function to the `lsr` package that would automate this process and construct something like a “Bayesian Type II ANOVA table” from the output of the `anovaBF()` function. However, I haven’t had time to do this yet, nor have I made up my mind about whether it’s really a good idea to do this. In the meantime, I thought I should show you the trick for how I do this in practice. The command that I use when I want to grab the right Bayes factors for a Type II ANOVA is this one: `max(models)/models` \n\n```\n## denominator\n## numerator drug therapy drug + therapy\n## drug + therapy 2.910308 978.2007 1\n## denominator\n## numerator drug + therapy + drug:therapy\n## drug + therapy 1.047753\n```\n\n The output isn’t quite so pretty as the last one, but the nice thing is that you can read off everything you need. The best model is `drug + therapy` , so all the other models are being compared to that. What’s the Bayes factor for the main effect of `drug` ? The relevant null hypothesis is the one that contains only `therapy` , and the Bayes factor in question is 954:1. The main effect of `therapy` is weaker, and the evidence here is only 2.8:1. Finally, the evidence against an interaction is very weak, at 1.01:1. Reading the results off this table is sort of counterintuitive, because you have to read off the answers from the “wrong” part of the table. For instance, the evidence for an effect of `drug` can be read from the column labelled `therapy` , which is pretty damned weird. To be fair to the authors of the package, I don’t think they ever intended for the `anovaBF()` function to be used this way. My understanding274 is that their view is simply that you should find the best model and report that model: there’s no inherent reason why a Bayesian ANOVA should try to follow the exact same design as an orthodox ANOVA.275 \nIn any case, if you know what you’re looking for, you can look at this table and then report the results of the Bayesian analysis in a way that is pretty closely analogous to how you’d report a regular Type II ANOVA. As I mentioned earlier, there’s still no convention on how to do that, but I usually go for something like this:\n A Bayesian Type II ANOVA found evidence for main effects of drug (Bayes factor: 954:1) and therapy (Bayes factor: 3:1), but no clear evidence for or against an interaction (Bayes factor: 1:1).\n\n## 17.10 Summary\n\nThe first half of this chapter was focused primarily on the theoretical underpinnings of Bayesian statistics. I introduced the mathematics for how Bayesian inference works (Section 17.1), and gave a very basic overview of how Bayesian hypothesis testing is typically done (Section 17.2). Finally, I devoted some space to talking about why I think Bayesian methods are worth using (Section 17.3.\n The second half of the chapter was a lot more practical, and focused on tools provided by the `BayesFactor` package. Specifically, I talked about using the `contingencyTableBF()` function to do Bayesian analogs of chi-square tests (Section 17.6, the `ttestBF()` function to do Bayesian \\(t\\)-tests, (Section 17.7), the `regressionBF()` function to do Bayesian regressions, and finally the `anovaBF()` function for Bayesian ANOVA. \nIf you’re interested in learning more about the Bayesian approach, there are many good books you could look into. ’s book Doing Bayesian Data Analysis is a pretty good place to start (Kruschke 2011), and is a nice mix of theory and practice. His approach is a little different to the “Bayes factor” approach that I’ve discussed here, so you won’t be covering the same ground. If you’re a cognitive psychologist, you might want to check out and ’ book Bayesian Cognitive Modeling (Lee and Wagenmakers 2014). I picked these two because I think they’re especially useful for people in my discipline, but there’s a lot of good books out there, so look around!\n\n. 1925. Statistical Methods for Research Workers. Edinburgh, UK: Oliver; Boyd.\n\nIt’s a leap of faith, I know, but let’s run with it okay?↩\n * \n\nUm. I hate to bring this up, but some statisticians would object to me using the word “likelihood” here. The problem is that the word “likelihood” has a very specific meaning in frequentist statistics, and it’s not quite the same as what it means in Bayesian statistics. As far as I can tell, Bayesians didn’t originally have any agreed upon name for the likelihood, and so it became common practice for people to use the frequentist terminology. This wouldn’t have been a problem, except for the fact that the way that Bayesians use the word turns out to be quite different to the way frequentists do. This isn’t the place for yet another lengthy history lesson, but to put it crudely: when a Bayesian says “a likelihood function” they’re usually referring one of the rows of the table. When a frequentist says the same thing, they’re referring to the same table, but to them “a likelihood function” almost always refers to one of the columns. This distinction matters in some contexts, but it’s not important for our purposes.↩\n * \n\nIf we were being a bit more sophisticated, we could extend the example to accommodate the possibility that I’m lying about the umbrella. But let’s keep things simple, shall we?↩\n * \n\nYou might notice that this equation is actually a restatement of the same basic rule I listed at the start of the last section. If you multiply both sides of the equation by \\(P(d)\\), then you get \\(P(d) P(h| d) = P(d,h)\\), which is the rule for how joint probabilities are calculated. So I’m not actually introducing any “new” rules here, I’m just using the same rule in a different way.↩\n * \n\nObviously, this is a highly simplified story. All the complexity of real life Bayesian hypothesis testing comes down to how you calculate the likelihood \\(P(d|h)\\) when the hypothesis \\(h\\) is a complex and vague thing. I’m not going to talk about those complexities in this book, but I do want to highlight that although this simple story is true as far as it goes, real life is messier than I’m able to cover in an introductory stats textbook.↩\n * \n\nhttp://www.imdb.com/title/tt0093779/quotes. I should note in passing that I’m not the first person to use this quote to complain about frequentist methods. and colleagues had the idea first. I’m shamelessly stealing it because it’s such an awesome pull quote to use in this context and I refuse to miss any opportunity to quote The Princess Bride.↩\n * \n\nhttp://about.abc.net.au/reports-publications/appreciation-survey-summary-report-2013/↩\n * \n\nIn the interests of being completely honest, I should acknowledge that not all orthodox statistical tests that rely on this silly assumption. There are a number of sequential analysis tools that are sometimes used in clinical trials and the like. These methods are built on the assumption that data are analysed as they arrive, and these tests aren’t horribly broken in the way I’m complaining about here. However, sequential analysis methods are constructed in a very different fashion to the “standard” version of null hypothesis testing. They don’t make it into any introductory textbooks, and they’re not very widely used in the psychological literature. The concern I’m raising here is valid for every single orthodox test I’ve presented so far, and for almost every test I’ve seen reported in the papers I read.↩\n * \n\nA related problem: http://xkcd.com/1478/↩\n * \n\nSome readers might wonder why I picked 3:1 rather than 5:1, given that Johnson (2013) suggests that \\(p=.05\\) lies somewhere in that range. I did so in order to be charitable to the \\(p\\)-value. If I’d chosen a 5:1 Bayes factor instead, the results would look even better for the Bayesian approach.↩\n * \n\nOkay, I just know that some knowledgeable frequentists will read this and start complaining about this section. Look, I’m not dumb. I absolutely know that if you adopt a sequential analysis perspective you can avoid these errors within the orthodox framework. I also know that you can explictly design studies with interim analyses in mind. So yes, in one sense I’m attacking a “straw man” version of orthodox methods. However, the straw man that I’m attacking is the one that is used by almost every single practitioner. If it ever reaches the point where sequential methods become the norm among experimental psychologists and I’m no longer forced to read 20 extremely dubious ANOVAs a day, I promise I’ll rewrite this section and dial down the vitriol. But until that day arrives, I stand by my claim that default Bayes factor methods are much more robust in the face of data analysis practices as they exist in the real world. Default orthodox methods suck, and we all know it.↩\n * \n\nIf you’re desperate to know, you can find all the gory details in Gunel and Dickey (1974). However, that’s a pretty technical paper. The help documentation to the\n `contingencyTableBF()` gives this explanation: “the argument `priorConcentration` indexes the expected deviation from the null hypothesis under the alternative, and corresponds to Gunel and Dickey’s (1974) \\(a\\) parameter.” As I write this I’m about halfway through the Gunel and Dickey paper, and I agree that setting \\(a=1\\) is a pretty sensible default choice, since it corresponds to an assumption that you have very little a priori knowledge about the contingency table.↩ * \n\nIn some of the later examples, you’ll see that this number is not always 0%. This is because the\n `BayesFactor` package often has to run some simulations to compute approximate Bayes factors. So the answers you get won’t always be identical when you run the command a second time. That’s why the output of these functions tells you what the margin for error is.↩ * \n\nApparently this omission is deliberate. I have this vague recollection that I spoke to about this once, and his opinion was that when homogeneity of variance is violated the results of a \\(t\\)-test are uninterpretable. I can see the argument for this, but I’ve never really held a strong opinion myself. (Jeff, if you never said that, I’m sorry)↩\n * \n\nJust in case you’re interested: the “JZS” part of the output relates to how the Bayesian test expresses the prior uncertainty about the variance \\(\\sigma^2\\), and it’s short for the names of three people: “”. See Rouder et al. (2009) for details.↩\n * \n\nAgain, in case you care … the null hypothesis here specifies an effect size of 0, since the two means are identical. The alternative hypothesis states that there is an effect, but it doesn’t specify exactly how big the effect will be. The \\(r\\) value here relates to how big the effect is expected to be according to the alternative. You can type\n `?ttestBF` to get more details.↩ * \n\nAgain, guys, sorry if I’ve misread you.↩\n * \n\nI don’t even disagree with them: it’s not at all obvious why a Bayesian ANOVA should reproduce (say) the same set of model comparisons that the Type II testing strategy uses. It’s precisely because of the fact that I haven’t really come to any strong conclusions that I haven’t added anything to the\n `lsr` package to make Bayesian Type II tests easier to produce.↩\n\n# Epilogue\n\n“Begin at the beginning”, the King said, very gravely, “and go on till you come to the end: then stop”\n – \nIt feels somewhat strange to be writing this chapter, and more than a little inappropriate. An epilogue is what you write when a book is finished, and this book really isn’t finished. There are a lot of things still missing from this book. It doesn’t have an index yet. A lot of references are missing. There are no “do it yourself exercises”. And in general, I feel that there a lot of things that are wrong with the presentation, organisation and content of this book. Given all that, I don’t want to try to write a “proper” epilogue. I haven’t finished writing the substantive content yet, so it doesn’t make sense to try to bring it all together. But this version of the book is going to go online for students to use, and you will be able to purchase a hard copy too, so I want to give it at least a veneer of closure. So let’s give it a go, shall we?\n\n## The undiscovered statistics\n\nFirst, I’m going to talk a bit about some of the content that I wish I’d had the chance to cram into this version of the book, just so that you can get a sense of what other ideas are out there in the world of statistics. I think this would be important even if this book were getting close to a final product: one thing that students often fail to realise is that their introductory statistics classes are just that: an introduction. If you want to go out into the wider world and do real data analysis, you have to learn a whole lot of new tools that extend the content of your undergraduate lectures in all sorts of different ways. Don’t assume that something can’t be done just because it wasn’t covered in undergrad. Don’t assume that something is the right thing to do just because it {} covered in an undergrad class. To stop you from falling victim to that trap, I think it’s useful to give a bit of an overview of some of the other ideas out there.\n\n### Omissions within the topics covered\n\nEven within the topics that I have covered in the book, there are a lot of omissions that I’d like to redress in future version of the book. Just sticking to things that are purely about statistics (rather than things associated with R), the following is a representative but not exhaustive list of topics that I’d like to expand on in later versions:\n\n* Other types of correlations In Chapter 5 I talked about two types of correlation: Pearson and Spearman. Both of these methods of assessing correlation are applicable to the case where you have two continuous variables and want to assess the relationship between them. What about the case where your variables are both nominal scale? Or when one is nominal scale and the other is continuous? There are actually methods for computing correlations in such cases (e.g., polychoric correlation), but I just haven’t had time to write about them yet.\n * More detail on effect sizes In general, I think the treatment of effect sizes throughout the book is a little more cursory than it should be. In almost every instance, I’ve tended just to pick one measure of effect size (usually the most popular one) and describe that. However, for almost all tests and models there are multiple ways of thinking about effect size, and I’d like to go into more detail in the future.\n * Dealing with violated assumptions In a number of places in the book I’ve talked about some things you can do when you find that the assumptions of your test (or model) are violated, but I think that I ought to say more about this. In particular, I think it would have been nice to talk in a lot more detail about how you can tranform variables to fix problems. I talked a bit about this in Sections 7.2, 7.3 and 15.9.4, but the discussion isn’t detailed enough I think.\n * Interaction terms for regression In Chapter 16 I talked about the fact that you can have interaction terms in an ANOVA, and I also pointed out that ANOVA can be interpreted as a kind of linear regression model. Yet, when talking about regression in Chapter 15 I made no mention of interactions at all. However, there’s nothing stopping you from including interaction terms in a regression model. It’s just a little more complicated to figure out what an “interaction” actually means when you’re talking about the interaction between two continuous predictors, and it can be done in more than one way. Even so, I would have liked to talk a little about this.\n * Method of planned comparison As I mentioned this in Chapter 16, it’s not always appropriate to be using post hoc correction like Tukey’s HSD when doing an ANOVA, especially when you had a very clear (and limited) set of comparisons that you cared about ahead of time. I would like to talk more about this in a future version of book.\n * Multiple comparison methods Even within the context of talking about post hoc tests and multiple comparisons, I would have liked to talk about the methods in more detail, and talk about what other methods exist besides the few options I mentioned.\n\n## Statistical models missing from the book\n\nStatistics is a huge field. The core tools that I’ve described in this book (chi-square tests, \\(t\\)-tests, ANOVA and regression) are basic tools that are widely used in everyday data analysis, and they form the core of most introductory stats books. However, there are a lot of other tools out there. There are so very many data analysis situations that these tools don’t cover, and in future versions of this book I want to talk about them. To give you a sense of just how much more there is, and how much more work I want to do to finish this thing, the following is a list of statistical modelling tools that I would have liked to talk about. Some of these will definitely make it into future versions of the book.\n\nAnalysis of covariance In Chapter 16 I spent a bit of time discussing the connection between ANOVA and regression, pointing out that any ANOVA model can be recast as a kind of regression model. More generally, both are examples of linear models, and it’s quite possible to consider linear models that are more general than either. The classic example of this is “analysis of covariance” (ANCOVA), and it refers to the situation where some of your predictors are continuous (like in a regression model) and others are categorical (like in an ANOVA).\n * \n\nNonlinear regression When discussing regression in Chapter 15, we saw that regression assume that the relationship between predictors and outcomes is linear. One the other hand, when we talked about the simpler problem of correlation in Chapter 5, we saw that there exist tools (e.g., Spearman correlations) that are able to assess non-linear relationships between variables. There are a number of tools in statistics that can be used to do non-linear regression. For instance, some non-linear regression models assume that the relationship between predictors and outcomes is monotonic (e.g., isotonic regression), while others assume that it is smooth but not necessarily monotonic (e.g., Lowess regression), while others assume that the relationship is of a known form that happens to be nonlinear (e.g., polynomial regression).\n * \n\nLogistic regression Yet another variation on regression occurs when the outcome variable is binary valued, but the predictors are continuous. For instance, suppose you’re investigating social media, and you want to know if it’s possible to predict whether or not someone is on Twitter as a function of their income, their age, and a range of other variables. This is basically a regression model, but you can’t use regular linear regression because the outcome variable is binary (you’re either on Twitter or you’re not): because the outcome variable is binary, there’s no way that the residuals could possibly be normally distributed. There are a number of tools that statisticians can apply to this situation, the most prominent of which is logistic regression.\n * \n\nThe General Linear Model (GLM) The GLM is actually a family of models that includes logistic regression, linear regression, (some) nonlinear regression, ANOVA and many others. The basic idea in the GLM is essentially the same idea that underpins linear models, but it allows for the idea that your data might not be normally distributed, and allows for nonlinear relationships between predictors and outcomes. There are a lot of very handy analyses that you can run that fall within the GLM, so it’s a very useful thing to know about.\n * \n\nSurvival analysis In Chapter 2 I talked about “differential attrition”, the tendency for people to leave the study in a non-random fashion. Back then, I was talking about it as a potential methodological concern, but there are a lot of situations in which differential attrition is actually the thing you’re interested in. Suppose, for instance, you’re interested in finding out how long people play different kinds of computer games in a single session. Do people tend to play RTS (real time strategy) games for longer stretches than FPS (first person shooter) games? You might design your study like this. People come into the lab, and they can play for as long or as little as they like. Once they’re finished, you record the time they spent playing. However, due to ethical restrictions, let’s suppose that you cannot allow them to keep playing longer than two hours. A lot of people will stop playing before the two hour limit, so you know exactly how long they played. But some people will run into the two hour limit, and so you don’t know how long they would have kept playing if you’d been able to continue the study. As a consequence, your data are systematically censored: you’re missing all of the very long times. How do you analyse this data sensibly? This is the problem that survival analysis solves. It is specifically designed to handle this situation, where you’re systematically missing one “side” of the data because the study ended. It’s very widely used in health research, and in that context it is often literally used to analyse survival. For instance, you may be tracking people with a particular type of cancer, some who have received treatment A and others who have received treatment B, but you only have funding to track them for 5 years. At the end of the study period some people are alive, others are not. In this context, survival analysis is useful for determining which treatment is more effective, and telling you about the risk of death that people face over time.\n * \n\nRepeated measures ANOVA When talking about reshaping data in Chapter 7, I introduced some data sets in which each participant was measured in multiple conditions (e.g., in the drugs data set, the working memory capacity (WMC) of each person was measured under the influence of alcohol and caffeine). It is quite common to design studies that have this kind of repeated measures structure. A regular ANOVA doesn’t make sense for these studies, because the repeated measurements mean that independence is violated (i.e., observations from the same participant are more closely related to one another than to observations from other participants. Repeated measures ANOVA is a tool that can be applied to data that have this structure. The basic idea behind RM-ANOVA is to take into account the fact that participants can have different overall levels of performance. For instance, Amy might have a WMC of 7 normally, which falls to 5 under the influence of caffeine, whereas Borat might have a WMC of 6 normally, which falls to 4 under the influence of caffeine. Because this is a repeated measures design, we recognise that – although Amy has a higher WMC than Borat – the effect of caffeine is identical for these two people. In other words, a repeated measures design means that we can attribute some of the variation in our WMC measurement to individual differences (i.e., some of it is just that Amy has higher WMC than Borat), which allows us to draw stronger conclusions about the effect of caffeine.\n * \n\nMixed models Repeated measures ANOVA is used in situations where you have observations clustered within experimental units. In the example I gave above, we have multiple WMC measures for each participant (i.e., one for each condition). However, there are a lot of other ways in which you can end up with multiple observations per participant, and for most of those situations the repeated measures ANOVA framework is insufficient. A good example of this is when you track individual people across multiple time points. Let’s say you’re tracking happiness over time, for two people. Aaron’s happiness starts at 10, then drops to 8, and then to 6. Belinda’s happiness starts at 6, then rises to 8 and then to 10. Both of these two people have the same “overall” level of happiness (the average across the three time points is 8), so a repeated measures ANOVA analysis would treat Aaron and Belinda the same way. But that’s clearly wrong. Aaron’s happiness is decreasing, whereas Belinda’s is increasing. If you want to optimally analyse data from an experiment where people can change over time, then you need a more powerful tool than repeated measures ANOVA. The tools that people use to solve this problem are called “mixed” models, because they are designed to learn about individual experimental units (e.g. happiness of individual people over time) as well as overall effects (e.g. the effect of money on happiness over time). Repeated measures ANOVA is perhaps the simplest example of a mixed model, but there’s a lot you can do with mixed models that you can’t do with repeated measures ANOVA.\n * \n\nReliability analysis Back in Chapter 2 I talked about reliability as one of the desirable characteristics of a measurement. One of the different types of reliability I mentioned was inter-item reliability. For example, when designing a survey used to measure some aspect to someone’s personality (e.g., extraversion), one generally attempts to include several different questions that all ask the same basic question in lots of different ways. When you do this, you tend to expect that all of these questions will tend to be correlated with one another, because they’re all measuring the same latent construct. There are a number of tools (e.g., Cronbach’s \\(\\alpha\\)) that you can use to check whether this is actually true for your study.\n * \n\nFactor analysis One big shortcoming with reliability measures like Cronbach’s \\(\\alpha\\) is that they assume that your observed variables are all measuring a single latent construct. But that’s not true in general. If you look at most personality questionnaires, or IQ tests, or almost anything where you’re taking lots of measurements, it’s probably the case that you’re actually measuring several things at once. For example, all the different tests used when measuring IQ do tend to correlate with one another, but the pattern of correlations that you see across tests suggests that there are multiple different “things” going on in the data. Factor analysis (and related tools like principal components analysis and independent components analsysis) is a tool that you can use to help you figure out what these things are. Broadly speaking, what you do with these tools is take a big correlation matrix that describes all pairwise correlations between your variables, and attempt to express this pattern of correlations using only a small number of latent variables. Factor analysis is a very useful tool – it’s a great way of trying to see how your variables are related to one another – but it can be tricky to use well. A lot of people make the mistake of thinking that when factor analysis uncovers a latent variable (e.g., extraversion pops out as a latent variable when you factor analyse most personality questionnaires), it must actually correspond to a real “thing”. That’s not necessarily true. Even so, factor analysis is a very useful thing to know about (especially for psychologists), and I do want to talk about it in a later version of the book.\n * \n\nMultidimensional scaling Factor analysis is an example of an “unsupervised learning” model. What this means is that, unlike most of the “supervised learning” tools I’ve mentioned, you can’t divide up your variables in to predictors and outcomes. Regression is supervised learning; factor analysis is unsupervised learning. It’s not the only type of unsupervised learning model however. For example, in factor analysis one is concerned with the analysis of correlations between variables. However, there are many situations where you’re actually interested in analysing similarities or dissimilarities between objects, items or people. There are a number of tools that you can use in this situation, the best known of which is multidimensional scaling (MDS). In MDS, the idea is to find a “geometric” representation of your items. Each item is “plotted” as a point in some space, and the distance between two points is a measure of how dissimilar those items are.\n * \n\nClustering Another example of an unsupervised learning model is clustering (also referred to as classification), in which you want to organise all of your items into meaningful groups, such that similar items are assigned to the same groups. A lot of clustering is unsupervised, meaning that you don’t know anything about what the groups are, you just have to guess. There are other “supervised clustering” situations where you need to predict group memberships on the basis of other variables, and those group memberships are actually observables: logistic regression is a good example of a tool that works this way. However, when you don’t actually know the group memberships, you have to use different tools (e.g., \\(k\\)-means clustering). There’s even situations where you want to do something called “semi-supervised clustering”, in which you know the group memberships for some items but not others. As you can probably guess, clustering is a pretty big topic, and a pretty useful thing to know about.\n * \n\nCausal models One thing that I haven’t talked about much in this book is how you can use statistical modeling to learn about the causal relationships between variables. For instance, consider the following three variables which might be of interest when thinking about how someone died in a firing squad. We might want to measure whether or not an execution order was given (variable A), whether or not a marksman fired their gun (variable B), and whether or not the person got hit with a bullet (variable C). These three variables are all correlated with one another (e.g., there is a correlation between guns being fired and people getting hit with bullets), but we actually want to make stronger statements about them than merely talking about correlations. We want to talk about causation. We want to be able to say that the execution order (A) causes the marksman to fire (B) which causes someone to get shot (C). We can express this by a directed arrow notation: we write it as \\(A \\rightarrow B \\rightarrow C\\). This “causal chain” is a fundamentally different explanation for events than one in which the marksman fires first, which causes the shooting \\(B \\rightarrow C\\), and then causes the executioner to “retroactively” issue the execution order, \\(B \\rightarrow A\\). This “common effect” model says that A and C are both caused by B. You can see why these are different. In the first causal model, if we had managed to stop the executioner from issuing the order (intervening to change A), then no shooting would have happened. In the second model, the shooting would have happened any way because the marksman was not following the execution order. There is a big literature in statistics on trying to understand the causal relationships between variables, and a number of different tools exist to help you test different causal stories about your data. The most widely used of these tools (in psychology at least) is structural equations modelling (SEM), and at some point I’d like to extend the book to talk about it.\n\nOf course, even this listing is incomplete. I haven’t mentioned time series analysis, item response theory, market basket analysis, classification and regression trees, or any of a huge range of other topics. However, the list that I’ve given above is essentially my wish list for this book. Sure, it would double the length of the book, but it would mean that the scope has become broad enough to cover most things that applied researchers in psychology would need to use.\n\n## Other ways of doing inference\n\nA different sense in which this book is incomplete is that it focuses pretty heavily on a very narrow and old-fashioned view of how inferential statistics should be done. In Chapter 10 I talked a little bit about the idea of unbiased estimators, sampling distributions and so on. In Chapter 11 I talked about the theory of null hypothesis significance testing and \\(p\\)-values. These ideas have been around since the early 20th century, and the tools that I’ve talked about in the book rely very heavily on the theoretical ideas from that time. I’ve felt obligated to stick to those topics because the vast majority of data analysis in science is also reliant on those ideas. However, the theory of statistics is not restricted to those topics, and – while everyone should know about them because of their practical importance – in many respects those ideas do not represent best practice for contemporary data analysis. One of the things that I’m especially happy with is that I’ve been able to go a little beyond this. Chapter 17 now presents the Bayesian perspective in a reasonable amount of detail, but the book overall is still pretty heavily weighted towards the frequentist orthodoxy. Additionally, there are a number of other approaches to inference that are worth mentioning:\n\nBootstrapping Throughout the book, whenever I’ve introduced a hypothesis test, I’ve had a strong tendency just to make assertions like “the sampling distribution for BLAH is a \\(t\\)-distribution” or something like that. In some cases, I’ve actually attempted to justify this assertion. For example, when talking about \\(\\chi^2\\) tests in Chapter 12, I made reference to the known relationship between normal distributions and \\(\\chi^2\\) distributions (see Chapter 9) to explain how we end up assuming that the sampling distribution of the goodness of fit statistic is \\(\\chi^2\\). However, it’s also the case that a lot of these sampling distributions are, well, wrong. The \\(\\chi^2\\) test is a good example: it is based on an assumption about the distribution of your data, an assumption which is known to be wrong for small sample sizes! Back in the early 20th century, there wasn’t much you could do about this situation: statisticians had developed mathematical results that said that “under assumptions BLAH about the data, the sampling distribution is approximately BLAH”, and that was about the best you could do. A lot of times they didn’t even have that: there are lots of data analysis situations for which no-one has found a mathematical solution for the sampling distributions that you need. And so up until the late 20th century, the corresponding tests didn’t exist or didn’t work. However, computers have changed all that now. There are lots of fancy tricks, and some not-so-fancy, that you can use to get around it. The simplest of these is bootstrapping, and in it’s simplest form it’s incredibly simple. Here it is: simulate the results of your experiment lots and lots of time, under the twin assumptions that (a) the null hypothesis is true and (b) the unknown population distribution actually looks pretty similar to your raw data. In other words, instead of assuming that the data are (for instance) normally distributed, just assume that the population looks the same as your sample, and then use computers to simulate the sampling distribution for your test statistic if that assumption holds. Despite relying on a somewhat dubious assumption (i.e., the population distribution is the same as the sample!) bootstrapping is quick and easy method that works remarkably well in practice for lots of data analysis problems.\n * \n\nCross validation One question that pops up in my stats classes every now and then, usually by a student trying to be provocative, is “Why do we care about inferential statistics at all? Why not just describe your sample?” The answer to the question is usually something like this: “Because our true interest as scientists is not the specific sample that we have observed in the past, we want to make predictions about data we might observe in the future”. A lot of the issues in statistical inference arise because of the fact that we always expect the future to be similar to but a bit different from the past. Or, more generally, new data won’t be quite the same as old data. What we do, in a lot of situations, is try to derive mathematical rules that help us to draw the inferences that are most likely to be correct for new data, rather than to pick the statements that best describe old data. For instance, given two models A and B, and a data set X you collected today, try to pick the model that will best describe a new data set Y that you’re going to collect tomorrow. Sometimes it’s convenient to simulate the process, and that’s what cross-validation does. What you do is divide your data set into two subsets, X1 and X2. Use the subset X1 to train the model (e.g., estimate regression coefficients, let’s say), but then assess the model performance on the other one X2. This gives you a measure of how well the model generalises from an old data set to a new one, and is often a better measure of how good your model is than if you just fit it to the full data set X.\n * \n\nRobust statistics Life is messy, and nothing really works the way it’s supposed to. This is just as true for statistics as it is for anything else, and when trying to analyse data we’re often stuck with all sorts of problems in which the data are just messier than they’re supposed to be. Variables that are supposed to be normally distributed are not actually normally distributed, relationships that are supposed to be linear are not actually linear, and some of the observations in your data set are almost certainly junk (i.e., not measuring what they’re supposed to). All of this messiness is ignored in most of the statistical theory I developed in this book. However, ignoring a problem doesn’t always solve it. Sometimes, it’s actually okay to ignore the mess, because some types of statistical tools are “robust”: if the data don’t satisfy your theoretical assumptions, they still work pretty well. Other types of statistical tools are not robust: even minor deviations from the theoretical assumptions cause them to break. Robust statistics is a branch of stats concerned with this question, and they talk about things like the “breakdown point” of a statistic: that is, how messy does your data have to be before the statistic cannot be trusted? I touched on this in places. The mean is not a robust estimator of the central tendency of a variable; the median is. For instance, suppose I told you that the ages of my five best friends are 34, 39, 31, 43 and 4003 years. How old do you think they are on average? That is, what is the true population mean here? If you use the sample mean as your estimator of the population mean, you get an answer of 830 years. If you use the sample median as the estimator of the population mean, you get an answer of 39 years. Notice that, even though you’re “technically” doing the wrong thing in the second case (using the median to estimate the mean!) you’re actually getting a better answer. The problem here is that one of the observations is clearly, obviously a lie. I don’t have a friend aged 4003 years. It’s probably a typo: I probably meant to type 43. But what if I had typed 53 instead of 43, or 34 instead of 43? Could you be sure if this was a typo? Sometimes the errors in the data are subtle, so you can’t detect them just by eyeballing the sample, but they’re still errors that contaminate your data, and they still affect your conclusions. Robust statistics is a concerned with how you can make safe inferences even when faced with contamination that you don’t know about. It’s pretty cool stuff.\n\n### Miscellaneous topics\n\nMissing data Suppose you’re doing a survey, and you’re interested in exercise and weight. You send data to four people. Adam says he exercises a lot and is not overweight. Briony says she exercises a lot and is not overweight. Carol says she does not exercise and is overweight. Dan says he does not exercise and refuses to answer the question about his weight. Elaine does not return the survey. You now have a missing data problem. There is one entire survey missing, and one question missing from another one, What do you do about it? I’ve only barely touched on this question in this book, in Section5.8, and in that section all I did was tell you about some R commands you can use to ignore the missing data. But ignoring missing data is not, in general, a safe thing to do. Let’s think about Dan’s survey here. Firstly, notice that, on the basis of my other responses, I appear to be more similar to Carol (neither of us exercise) than to Adam or Briony. So if you were forced to guess my weight, you’d guess that I’m closer to her than to them. Maybe you’d make some correction for the fact that Adam and I are males and Briony and Carol are females. The statistical name for this kind of guessing is “imputation”. Doing imputation safely is hard, but important, especially when the missing data are missing in a systematic way. Because of the fact that people who are overweight are often pressured to feel poorly about their weight (often thanks to public health campaigns), we actually have reason to suspect that the people who are not responding are more likely to be overweight than the people who do respond. Imputing a weight to Dan means that the number of overweight people in the sample will probably rise from 1 out of 3 (if we ignore Dan), to 2 out of 4 (if we impute Dan’s weight). Clearly this matters. But doing it sensibly is more complicated than it sounds. Earlier, I suggested you should treat me like Carol, since we gave the same answer to the exercise question. But that’s not quite right: there is a systematic difference between us. She answered the question, and I didn’t. Given the social pressures faced by overweight people, isn’t it likely that I’m more overweight than Carol? And of course this is still ignoring the fact that it’s not sensible to impute a single weight to me, as if you actually knew my weight. Instead, what you need to do it is impute a range of plausible guesses (referred to as multiple imputation), in order to capture the fact that you’re more uncertain about my weight than you are about Carol’s. And let’s not get started on the problem posed by the fact that Elaine didn’t send in the survey. As you can probably guess, dealing with missing data is an increasingly important topic. In fact, I’ve been told that a lot of journals in some fields will not accept studies that have missing data unless some kind of sensible multiple imputation scheme is followed.\n * \n\nPower analysis In Chapter 11 I discussed the concept of power (i.e., how likely are you to be able to detect an effect if it actually exists), and referred to power analysis, a collection of tools that are useful for assessing how much power your study has. Power analysis can be useful for planning a study (e.g., figuring out how large a sample you’re likely to need), but it also serves a useful role in analysing data that you already collected. For instance, suppose you get a significant result, and you have an estimate of your effect size. You can use this information to estimate how much power your study actually had. This is kind of useful, especially if your effect size is not large. For instance, suppose you reject the null hypothesis \\(p<.05\\), but you use power analysis to figure out that your estimated power was only \\(.08\\). The significant result means that, if the null hypothesis was in fact true, there was a 5% chance of getting data like this. But the low power means that, even if the null hypothesis is false, the effect size was really as small as it looks, there was only an 8% chance of getting data like the one you did. This suggests that you need to be pretty cautious, because luck seems to have played a big part in your results, one way or the other!\n * \n\nData analysis using theory-inspired models In a few places in this book I’ve mentioned response time (RT) data, where you record how long it takes someone to do something (e.g., make a simple decision). I’ve mentioned that RT data are almost invariably non-normal, and positively skewed. Additionally, there’s a thing known as the speed-accuracy tradeoff: if you try to make decisions too quickly (low RT), you’re likely to make poorer decisions (lower accuracy). So if you measure both the accuracy of a participant’s decisions and their RT, you’ll probably find that speed and accuracy are related. There’s more to the story than this, of course, because some people make better decisions than others regardless of how fast they’re going. Moreover, speed depends on both cognitive processes (i.e., time spend thinking) but also physiological ones (e.g., how fast can you move your muscles). It’s starting to sound like analysing this data will be a complicated process. And indeed it is, but one of the things that you find when you dig into the psychological literature is that there already exist mathematical models (called “sequential sampling models”) that describe how people make simple decisions, and these models take into account a lot of the factors I mentioned above. You won’t find any of these theoretically-inspired models in a standard statistics textbook. Standard stats textbooks describe standard tools, tools that could meaningfully be applied in lots of different disciplines, not just psychology. ANOVA is an example of a standard tool: it is just as applicable to psychology as to pharmacology. Sequential sampling models are not: they are psychology-specific, more or less. This doesn’t make them less powerful tools: in fact, if you’re analysing data where people have to make choices quickly, you should really be using sequential sampling models to analyse the data. Using ANOVA or regression or whatever won’t work as well, because the theoretical assumptions that underpin them are not well-matched to your data. In contrast, sequential sampling models were explicitly designed to analyse this specific type of data, and their theoretical assumptions are extremely well-matched to the data. Obviously, it’s impossible to cover this sort of thing properly, because there are thousands of context-specific models in every field of science. Even so, one thing that I’d like to do in later versions of the book is to give some case studies that are of particular relevance to psychologists, just to give a sense for how psychological theory can be used to do better statistical analysis of psychological data. So, in later versions of the book I’ll probably talk about how to analyse response time data, among other things.\n\n## Learning the basics, and learning them in R\n\nOkay, that was… long. And even that listing is massively incomplete. There really are a lot of big ideas in statistics that I haven’t covered in this book. It can seem pretty depressing to finish a 600-page textbook only to be told that this only the beginning, especially when you start to suspect that half of the stuff you’ve been taught is wrong. For instance, there are a lot of people in the field who would strongly argue against the use of the classical ANOVA model, yet I’ve devote two whole chapters to it! Standard ANOVA can be attacked from a Bayesian perspective, or from a robust statistics perspective, or even from a “it’s just plain wrong” perspective (people very frequently use ANOVA when they should actually be using mixed models). So why learn it at all?\n\nAs I see it, there are two key arguments. Firstly, there’s the pure pragmatism argument. Rightly or wrongly, ANOVA is widely used. If you want to understand the scientific literature, you need to understand ANOVA. And secondly, there’s the “incremental knowledge” argument. In the same way that it was handy to have seen one-way ANOVA before trying to learn factorial ANOVA, understanding ANOVA is helpful for understanding more advanced tools, because a lot of those tools extend on or modify the basic ANOVA setup in some way. For instance, although mixed models are way more useful than ANOVA and regression, I’ve never heard of anyone learning how mixed models work without first having worked through ANOVA and regression. You have to learn to crawl before you can climb a mountain.\n Actually, I want to push this point a bit further. One thing that I’ve done a lot of in this book is talk about fundamentals. I spent a lot of time on probability theory. I talked about the theory of estimation and hypothesis tests in more detail than I needed to. When talking about R, I spent a lot of time talking about how the language works, and talking about things like writing your own scripts, functions and programs. I didn’t just teach you how to draw a histogram using `hist()` , I tried to give a basic overview of how the graphics system works. Why did I do all this? Looking back, you might ask whether I really needed to spend all that time talking about what a probability distribution is, or why there was even a section on probability density. If the goal of the book was to teach you how to run a \\(t\\)-test or an ANOVA, was all that really necessary? Or, come to think of it, why bother with R at all? There are lots of free alternatives out there: PSPP, for instance, is an SPSS-like clone that is totally free, has simple “point and click” menus, and can (I think) do every single analysis that I’ve talked about in this book. And you can learn PSPP in about 5 minutes. Was this all just a huge waste of everyone’s time??? The answer, I hope you’ll agree, is no. The goal of an introductory stats is not to teach ANOVA. It’s not to teach \\(t\\)-tests, or regressions, or histograms, or \\(p\\)-values. The goal is to start you on the path towards becoming a skilled data analyst. And in order for you to become a skilled data analyst, you need to be able to do more than ANOVA, more than \\(t\\)-tests, regressions and histograms. You need to be able to think properly about data. You need to be able to learn the more advanced statistical models that I talked about in the last section, and to understand the theory upon which they are based. And you need to have access to software that will let you use those advanced tools. And this is where – in my opinion at least – all that extra time I’ve spent on the fundamentals pays off. If you understand the graphics system in R, then you can draw the plots that you want, not just the canned plots that someone else has built into R for you. If you understand probability theory, you’ll find it much easier to switch from frequentist analyses to Bayesian ones. If you understand the core mechanics of R, you’ll find it much easier to generalise from linear regressions using `lm()` to using generalised linear models with `glm()` or linear mixed effects models using `lme()` and `lmer()` . You’ll even find that a basic knowledge of R will go a long way towards teaching you how to use other statistical programming languages that are based on it. Bayesians frequently rely on tools like WinBUGS and JAGS, which have a number of similarities to R, and can in fact be called from within R. In fact, because R is the “lingua franca of statistics”, what you’ll find is that most ideas in the statistics literature has been implemented somewhere as a package that you can download from CRAN. The same cannot be said for PSPP, or even SPSS. \nIn short, I think that the big payoff for learning statistics this way is extensibility. For a book that only covers the very basics of data analysis, this book has a massive overhead in terms of learning R, probability theory and so on. There’s a whole lot of other things that it pushes you to learn besides the specific analyses that the book covers. So if your goal had been to learn how to run an ANOVA in the minimum possible time, well, this book wasn’t a good choice. But as I say, I don’t think that is your goal. I think you want to learn how to do data analysis. And if that really is your goal, you want to make sure that the skills you learn in your introductory stats class are naturally and cleanly extensible to the more complicated models that you need in real world data analysis. You want to make sure that you learn to use the same tools that real data analysts use, so that you can learn to do what they do. And so yeah, okay, you’re a beginner right now (or you were when you started this book), but that doesn’t mean you should be given a dumbed-down story, a story in which I don’t tell you about probability density, or a story where I don’t tell you about the nightmare that is factorial ANOVA with unbalanced designs. And it doesn’t mean that you should be given baby toys instead of proper data analysis tools. Beginners aren’t dumb; they just lack knowledge. What you need is not to have the complexities of real world data analysis hidden from from you. What you need are the skills and tools that will let you handle those complexities when they inevitably ambush you in the real world.\n\nAnd what I hope is that this book – or the finished book that this will one day turn into – is able to help you with that.\n\n# References\n\n———. 2002. Categorical Data Analysis. 2nd ed. Hoboken, NJ: Wiley.\n\n———. 1922b. “On the Mathematical Foundation of Theoretical Statistics.” Philosophical Transactions of the Royal Society A 222: 309–68.\n\n———. 1925. Statistical Methods for Research Workers. Edinburgh, UK: Oliver; Boyd.\n\n., , and . 2005. Introduction to Mathematical Statistics. 6th ed. Upper Saddle River, NJ: Pearson.\n\n., , , , and . 2009. “Bayesian T-Tests for Accepting and Rejecting the Null Hypothesis.” Psychonomic Bulletin & Review 16: 225–37.\n\n. 1980. “A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity.” Econometrika 48: 817–38."}}},{"rowIdx":99,"cells":{"project":{"kind":"string","value":"block_keys"},"source":{"kind":"string","value":"hex"},"language":{"kind":"string","value":"Erlang"},"content":{"kind":"string","value":"Toggle Theme\n\nBlockKeys v0.1.10\n API Reference\n===\n\n Modules\n---\n\n[BlockKeys](BlockKeys.html)\n\nGenerates or restores a wallet from mnemonic phrases\n\n[BlockKeys.Base58](BlockKeys.Base58.html)\n\nDocumentation for Base58\n\n[BlockKeys.Base58.Check](BlockKeys.Base58.Check.html)\n\n[BlockKeys.Base58.Encoder](BlockKeys.Base58.Encoder.html)\n\n[BlockKeys.Bitcoin](BlockKeys.Bitcoin.html)\n\nHelper module to derive and convert to a Bitcoin Address\n\n[BlockKeys.Bitcoin.Address](BlockKeys.Bitcoin.Address.html)\n\nConverts a public extended key into a Bitcoin Address\n\n[BlockKeys.CKD](BlockKeys.CKD.html)\n\nThis module derives children keys given an extended public or private key and a path\n\n[BlockKeys.Crypto](BlockKeys.Crypto.html)\n\nThis module is a wrapper around cryptographic functions\n\n[BlockKeys.Encoding](BlockKeys.Encoding.html)\n\nThis module contains Base58check encoding and decoding functions for extended keys\n\n[BlockKeys.Ethereum](BlockKeys.Ethereum.html)\n\nHelper module to derive and convert to an Ethereum Address\n\n[BlockKeys.Ethereum.Address](BlockKeys.Ethereum.Address.html)\n\nConverts a public extended key into an Ethereum Address\n\n[BlockKeys.Mnemonic](BlockKeys.Mnemonic.html)\n\nBIP32 implementation responsible for generating mnemonic phrases, seeds and public / private address trees\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys\n\n \n===\n\nGenerates or restores a wallet from mnemonic phrases\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[from_mnemonic(phrase, network \\\\ :mainnet)](#from_mnemonic/2)\n\n[generate(network \\\\ :mainnet)](#generate/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Base58\n\n \n===\n\nDocumentation for Base58.\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[decode(data)](#decode/1)\n\n[decode_check(data, length \\\\ 25)](#decode_check/2)\n\n[encode(data, hash \\\\ \"\")](#encode/2)\n\n[encode_check(data, prefix)](#encode_check/2)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Base58.Check\n\n \n===\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[decode_check(data, length \\\\ 25)](#decode_check/2)\n\n[encode_check(data, prefix)](#encode_check/2)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Base58.Encoder\n\n \n===\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[alphabet()](#alphabet/0)\n\n[decode(data)](#decode/1)\n\n[encode(data, hash \\\\ \"\")](#encode/2)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Bitcoin\n\n \n===\n\nHelper module to derive and convert to a Bitcoin Address\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[address(key, path)](#address/2)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Bitcoin.Address\n\n \n===\n\nConverts a public extended key into a Bitcoin Address\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[from_xpub(xpub)](#from_xpub/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.CKD\n\n \n===\n\nThis module derives children keys given an extended public or private key and a path\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[child_key(error, index)](#child_key/2)\n\n[child_key_private(key, child_index)](#child_key_private/2)\n\n[child_key_public(key, child_index)](#child_key_public/2)\n\n[derive(extended_key, path)](#derive/2)\n\nReturns a Base58 encode check child extended key given an extended key and a path\n\n[master_keys(encoded_seed)](#master_keys/1)\n\n[master_private_key(arg, network \\\\ :mainnet)](#master_private_key/2)\n\n[master_public_key(key)](#master_public_key/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nReturns a Base58 encode check child extended key given an extended key and a path\n\n### \n\n Examples\n```\n iex> BlockKeys.derive(\"\", \"m/44'/0'/0'\")\n \"\"\n```\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Crypto\n\n \n===\n\nThis module is a wrapper around cryptographic functions\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[ec_point_addition(parent_key, child_key)](#ec_point_addition/2)\n\n[hash160(data)](#hash160/1)\n\n[public_key(private_key)](#public_key/1)\n\n[public_key_decompress(public_key)](#public_key_decompress/1)\n\n[ripemd160(data)](#ripemd160/1)\n\n[sha256(data)](#sha256/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Encoding\n\n \n===\n\nThis module contains Base58check encoding and decoding functions for extended keys\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[base58_encode(bytes, version_prefix \\\\ \"\")](#base58_encode/2)\n\n[decode_extended_key(key)](#decode_extended_key/1)\n\n[encode_extended_key(version_number, depth, fingerprint, index, chain_code, key)](#encode_extended_key/6)\n\n[encode_private(map)](#encode_private/1)\n\n[encode_public(payload)](#encode_public/1)\n\n[private_version_number(atom)](#private_version_number/1)\n\n[public_version_number(atom)](#public_version_number/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Ethereum\n\n \n===\n\nHelper module to derive and convert to an Ethereum Address\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[address(key, path)](#address/2)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Ethereum.Address\n\n \n===\n\nConverts a public extended key into an Ethereum Address\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[from_xpub(xpub)](#from_xpub/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nToggle Theme\n\nBlockKeys v0.1.10 BlockKeys.Mnemonic\n\n \n===\n\nBIP32 implementation responsible for generating mnemonic phrases, seeds and public / private address trees.\n\n[Link to this section](#summary)\n Summary\n===\n\n[Functions](#functions)\n---\n\n[entropy_from_phrase(phrase)](#entropy_from_phrase/1)\n\nTakes a string of word phrases and converts them back to 256bit entropy\n\n[generate_phrase(entropy \\\\ :crypto.strong_rand_bytes(32))](#generate_phrase/1)\n\nGenerates the 24 random manmonic words\n\n[generate_seed(mnemonic, password \\\\ \"\")](#generate_seed/2)\n\nGiven a binary of entropy it will generate teh hex encoded seed\n\n[salt(password)](#salt/1)\n\n[Link to this section](#functions)\nFunctions\n===\n\nTakes a string of word phrases and converts them back to 256bit entropy\n\n Examples\n---\n```\niex> BlockKeys.Mnemonic.entropy_from_phrase(\"safe result wire cattle sauce luggage couple legend pause rather employ pear trigger live daring unlock music lyrics smoke mistake endorse kite obey siren\"\")\nbe16fbf0922bf9098c4bfca1764923d10e89054de77091f0af3346f49cf665fe \n```\n\nGenerates the 24 random manmonic words.\n\nCan optionally accept entropy string to used to generate a mnemonic.\n\n Examples\n---\n```\niex> BlockKeys.Bip32Mnemonic.generate_phrase()\n\"baby shadow city tower diamond magnet avocado champion crash ...\"\niex> BlockKeys.Mnemonic.generate_phrase(\"1234\")\n\"couple muscle snack\"\n```\nNOTE: For now the seed can be only generated from 32 bytes entropy\n\nGiven a binary of entropy it will generate teh hex encoded seed\n\n Examples\n---\n```\niex> BlockKeys.Mnemonic.generate_seed(\"weather neither click twin monster night bridge door immense tornado crack model canal answer harbor weasel winter fan universe burden price quote tail ride\"\n\"af7f48a70d0ecedc77df984117e336e12f0f0e681a4c95b25f4f17516d7dc4cca456e3a400bd1c6a5a604af67eb58dc6e0eb46fd520ad99ef27855d119dca517\"\n```\n"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":0,"numItemsPerPage":100,"numTotalItems":59733,"offset":0,"length":100}},"jwt":"eyJhbGciOiJFZERTQSJ9.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc2MTkwMzMzOCwic3ViIjoiL2RhdGFzZXRzL1NpdmlsVGFyYW0vc3RhcmNvZGVyMi1kb2N1bWVudGF0aW9uIiwiZXhwIjoxNzYxOTA2OTM4LCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.amXJ5mULTuNaU2voAeshErt-kkTk0ZqX6Djix8vmmVUPA9bYQN4Yvhfmng39LL1ckIdniSzNHxz_LYy3NYRIDA","displayUrls":true,"sqlConsoleEnabled":true},"discussionsStats":{"closed":0,"open":2,"total":2},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[{"views":[{"key":"default/train","displayName":"train","viewName":"train"}],"sql":"-- The SQL console is powered by DuckDB WASM and runs entirely in the browser.\n-- Get started by typing a query or selecting a view from the options below.\nSELECT *\nFROM train\nWHERE source = 'free_programming_book'\n AND lower(content) LIKE '%python%'\n AND LOWER(content) LIKE '%algorithms%';","title":"Python Algorithms in Free Books","createdAt":"2025-04-03T17:17:56.697Z","slug":"RUf2lH9","private":false,"justification":"Retrieves entries related to both Python and algorithms from a specific source, providing a filtered dataset but lacking broader analytical depth.","viewNames":["train"]},{"views":[{"key":"default/train","displayName":"train","viewName":"train"}],"sql":"-- The SQL console is powered by DuckDB WASM and runs entirely in the browser.\n-- Get started by typing a query or selecting a view from the options below.\nSELECT *\nFROM train\nWHERE source = 'free_programming_book'\n AND language = 'Python'\nLIMIT 100;","title":"Python Code from Free Books","createdAt":"2025-04-03T11:27:46.424Z","slug":"WGn_NVw","private":false,"justification":"This query retrieves a limited number of records related to Python programming books, providing basic filtering but minimal analytical insight.","viewNames":["train"]}],"user":[]}}">