R for

Statistics

Pierre-Andr Cornillon Julie Josse

Arnaud Guyader Maela Kloareg
Franois Husson Eric Matzner-Lber
Nicolas Jgou Laurent Rouvire
 R for

Statistics
This page intentionally left blank
 R for

Statistics

Pierre-Andr Cornillon Julie Josse

Arnaud Guyader Maela Kloareg
Franois Husson Eric Matzner-Lber
Nicolas Jgou Laurent Rouvire
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
2012 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20120113

International Standard Book Number-13: 978-1-4398-8146-0 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a pho-
tocopy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

Preface xiii

I An Overview of R 1
1 Main Concepts 3
1.1 Installing R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Work Session . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Session in Linux . . . . . . . . . . . . . . . . . . . . . 4
1.2.2 Session in Windows . . . . . . . . . . . . . . . . . . . 4
1.2.3 Session on a Mac . . . . . . . . . . . . . . . . . . . . . 5
1.3 Help . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Online Assistance . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Help on CRAN . . . . . . . . . . . . . . . . . . . . . . 6
1.4 R Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 Creating, Displaying and Deleting Objects . . . . . . . 6
1.4.2 Type of Object . . . . . . . . . . . . . . . . . . . . . . 7
1.4.3 The Missing Value . . . . . . . . . . . . . . . . . . . . 8
1.4.4 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4.4.1 Numeric Vectors . . . . . . . . . . . . . . . . 9
1.4.4.2 Character Vectors . . . . . . . . . . . . . . . 10
1.4.4.3 Logical Vectors . . . . . . . . . . . . . . . . . 11
1.4.4.4 Selecting Part of a Vector . . . . . . . . . . . 12
1.4.4.5 Selection in Practice . . . . . . . . . . . . . . 13
1.4.5 Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4.5.1 Selecting Elements or Part of a Matrix . . . 14
1.4.5.2 Calculating with the Matrices . . . . . . . . 16
1.4.5.3 Row and Column Operations . . . . . . . . . 17
1.4.6 Factors . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.7 Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.7.1 Creating a List . . . . . . . . . . . . . . . . . 20
1.4.7.2 Extraction . . . . . . . . . . . . . . . . . . . 20
1.4.7.3 A Few List Functions . . . . . . . . . . . . . 21
1.4.7.4 Dimnames List . . . . . . . . . . . . . . . . . 21
1.4.8 Data-Frames . . . . . . . . . . . . . . . . . . . . . . . 22
1.5 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.5.1 Arguments of a Function . . . . . . . . . . . . . . . . 23

v
vi R for Statistics

1.5.2 Output . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.6 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.6.1 What Is a Package ? . . . . . . . . . . . . . . . . . . . 24
1.6.2 Installing a Package . . . . . . . . . . . . . . . . . . . 24
1.6.3 Updating Packages . . . . . . . . . . . . . . . . . . . . 25
1.6.4 Using Packages . . . . . . . . . . . . . . . . . . . . . . 25
1.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2 Preparing Data 29
2.1 Reading Data from File . . . . . . . . . . . . . . . . . . . . . 29
2.2 Exporting Results . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Manipulating Variables . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Changing Type . . . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Dividing into Classes . . . . . . . . . . . . . . . . . . . 35
2.3.3 Working on Factor Levels . . . . . . . . . . . . . . . . 36
2.4 Manipulating Individuals . . . . . . . . . . . . . . . . . . . . 39
2.4.1 Identifying Missing Data . . . . . . . . . . . . . . . . . 39
2.4.2 Finding Outliers . . . . . . . . . . . . . . . . . . . . . 42
2.5 Concatenating Data Tables . . . . . . . . . . . . . . . . . . . 43
2.6 Cross-Tabulation . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3 R Graphics 51
3.1 Conventional Graphical Functions . . . . . . . . . . . . . . . 51
3.1.1 The Plot Function . . . . . . . . . . . . . . . . . . . . 52
3.1.2 Representing a Distribution . . . . . . . . . . . . . . . 58
3.1.3 Adding to Graphs . . . . . . . . . . . . . . . . . . . . 60
3.1.4 Graphs with Multiple Dimensions . . . . . . . . . . . 64
3.1.5 Exporting Graphs . . . . . . . . . . . . . . . . . . . . 66
3.1.6 Multiple Graphs . . . . . . . . . . . . . . . . . . . . . 67
3.1.7 Multiple Windows . . . . . . . . . . . . . . . . . . . . 69
3.1.8 Improving and Personalising Graphs . . . . . . . . . . 70
3.2 Graphical Functions with lattice . . . . . . . . . . . . . . . . 73
3.2.1 Characteristics of a Lattice Graph . . . . . . . . . . 75
3.2.2 Formulae and Groups . . . . . . . . . . . . . . . . . . 76
3.2.3 Customising Graphs . . . . . . . . . . . . . . . . . . . 79
3.2.3.1 Panel Function . . . . . . . . . . . . . . . . . 79
3.2.3.2 Controlling Size . . . . . . . . . . . . . . . . 80
3.2.3.3 Panel Layout . . . . . . . . . . . . . . . . . . 81
3.2.4 Exportation . . . . . . . . . . . . . . . . . . . . . . . . 81
3.2.5 Other Packages . . . . . . . . . . . . . . . . . . . . . . 82
3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Contents vii

4 Making Programs with R 87

4.1 Control Flows . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.1.1 Grouped Expressions . . . . . . . . . . . . . . . . . . . 87
4.1.2 Loops (for or while) . . . . . . . . . . . . . . . . . . . 87
4.1.3 Conditions (if, else) . . . . . . . . . . . . . . . . . . . 89
4.2 Predefined Functions . . . . . . . . . . . . . . . . . . . . . . 90
4.3 Creating a Function . . . . . . . . . . . . . . . . . . . . . . . 97
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

II Statistical Methods 101

Introduction to the Statistical Methods Part 103

5 A Quick Start with R 105

5.1 Installing R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Opening and Closing R . . . . . . . . . . . . . . . . . . . . . 105
5.3 The Command Prompt . . . . . . . . . . . . . . . . . . . . . 105
5.4 Attribution, Objects, and Function . . . . . . . . . . . . . . 106
5.5 Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.6 Other . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.7 Rcmdr Package . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.8 Importing (or Inputting) Data . . . . . . . . . . . . . . . . . 107
5.9 Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.10 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . 108

6 Hypothesis Test 109

6.1 Confidence Intervals for a Mean . . . . . . . . . . . . . . . . 109
6.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.1.4 Processing the Example . . . . . . . . . . . . . . . . . 110
6.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 112
6.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 112
6.2 Chi-Square Test of Independence . . . . . . . . . . . . . . . . 113
6.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2.4 Processing the Example . . . . . . . . . . . . . . . . . 114
6.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 116
6.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 117
6.3 Comparison of Two Means . . . . . . . . . . . . . . . . . . . 118
6.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.3.4 Processing the Example . . . . . . . . . . . . . . . . . 119
6.3.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 123
viii R for Statistics

6.3.6 Taking Things Further . . . . . . . . . . . . . . . . . . 124

6.4 Testing Conformity of a Proportion . . . . . . . . . . . . . . 125
6.4.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4.3 Step . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4.4 Processing the Example . . . . . . . . . . . . . . . . . 125
6.4.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 126
6.4.6 Taking Things Further . . . . . . . . . . . . . . . . . . 126
6.5 Comparing Several Proportions . . . . . . . . . . . . . . . . 127
6.5.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5.3 Step . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5.4 Processing the Example . . . . . . . . . . . . . . . . . 127
6.5.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 128
6.5.6 Taking Things Further . . . . . . . . . . . . . . . . . . 128
6.6 The Power of a Test . . . . . . . . . . . . . . . . . . . . . . . 129
6.6.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.6.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.6.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.6.4 Processing the Example . . . . . . . . . . . . . . . . . 130
6.6.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 130
6.6.6 Taking Things Further . . . . . . . . . . . . . . . . . . 131

7 Regression 133
7.1 Simple Linear Regression . . . . . . . . . . . . . . . . . . . . 133
7.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.1.4 Processing the Example . . . . . . . . . . . . . . . . . 134
7.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 138
7.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 139
7.2 Multiple Linear Regression . . . . . . . . . . . . . . . . . . . 140
7.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.2.4 Processing the Example . . . . . . . . . . . . . . . . . 141
7.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 145
7.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 146
7.3 Partial Least Squares (PLS) Regression . . . . . . . . . . . . 147
7.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.3.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.3.4 Processing the Example . . . . . . . . . . . . . . . . . 148
7.3.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 153
7.3.6 Taking Things Further . . . . . . . . . . . . . . . . . . 153
Contents ix

8 Analysis of Variance and Covariance 157

8.1 One-Way Analysis of Variance . . . . . . . . . . . . . . . . . 157
8.1.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . 157
8.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.1.4 Processing the Example . . . . . . . . . . . . . . . . . 159
8.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 164
8.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 165
8.2 Multi-Way Analysis of Variance with Interaction . . . . . . . 166
8.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.2.4 Processing the Example . . . . . . . . . . . . . . . . . 168
8.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 171
8.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 172
8.3 Analysis of Covariance . . . . . . . . . . . . . . . . . . . . . 173
8.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.3.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.3.4 Processing the Example . . . . . . . . . . . . . . . . . 175
8.3.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 179
8.3.6 Taking Things Further . . . . . . . . . . . . . . . . . . 179

9 Classification 181
9.1 Linear Discriminant Analysis . . . . . . . . . . . . . . . . . . 181
9.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 182
9.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
9.1.4 Processing the Example . . . . . . . . . . . . . . . . . 183
9.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 188
9.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 188
9.2 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . 190
9.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 190
9.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.2.4 Processing the Example . . . . . . . . . . . . . . . . . 191
9.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 197
9.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 198
9.3 Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
9.3.4 Processing the Example . . . . . . . . . . . . . . . . . 199
9.3.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 208
9.3.6 Taking Things Further . . . . . . . . . . . . . . . . . . 208
x R for Statistics

10 Exploratory Multivariate Analysis 209

10.1 Principal Component Analysis . . . . . . . . . . . . . . . . . 209
10.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 209
10.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
10.1.4 Processing the Example . . . . . . . . . . . . . . . . . 210
10.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 218
10.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 219
10.2 Correspondence Analysis . . . . . . . . . . . . . . . . . . . . 222
10.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 222
10.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 222
10.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
10.2.4 Processing the Example . . . . . . . . . . . . . . . . . 223
10.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 227
10.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 228
10.3 Multiple Correspondence Analysis . . . . . . . . . . . . . . . 230
10.3.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 230
10.3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 230
10.3.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
10.3.4 Processing the Example . . . . . . . . . . . . . . . . . 231
10.3.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 239
10.3.6 Taking Things Further . . . . . . . . . . . . . . . . . . 239

11 Clustering 241
11.1 Ascending Hierarchical Clustering . . . . . . . . . . . . . . . 241
11.1.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 241
11.1.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 241
11.1.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
11.1.4 Processing the Example . . . . . . . . . . . . . . . . . 242
11.1.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 246
11.1.6 Taking Things Further . . . . . . . . . . . . . . . . . . 247
11.2 The k-Means Method . . . . . . . . . . . . . . . . . . . . . . 252
11.2.1 Objective . . . . . . . . . . . . . . . . . . . . . . . . . 252
11.2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . 252
11.2.3 Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
11.2.4 Processing the Example . . . . . . . . . . . . . . . . . 252
11.2.5 Rcmdr Corner . . . . . . . . . . . . . . . . . . . . . . . 255
11.2.6 Taking Things Further . . . . . . . . . . . . . . . . . . 256

Appendix 257
A.1 The Most Useful Functions . . . . . . . . . . . . . . . . . . . 257
A.1.1 Generic Functions . . . . . . . . . . . . . . . . . . . . 257
A.1.2 Numerical Functions . . . . . . . . . . . . . . . . . . . 257
A.1.3 Data Handling Functions . . . . . . . . . . . . . . . . 258
A.1.4 Probability Distributions . . . . . . . . . . . . . . . . 260
Contents xi

A.1.5 Basic Statistical Functions . . . . . . . . . . . . . . . . 261

A.1.6 Advanced Statistical Functions . . . . . . . . . . . . . 262
A.1.7 Graphical Functions . . . . . . . . . . . . . . . . . . . 263
A.1.8 Import and Export Functions . . . . . . . . . . . . . . 264
A.1.9 Text Management . . . . . . . . . . . . . . . . . . . . 265
A.1.10 Some Other Useful Functions . . . . . . . . . . . . . . 265
A.2 Writing a Formula for the Models . . . . . . . . . . . . . . . 266
A.3 The Rcmdr Package . . . . . . . . . . . . . . . . . . . . . . . 267
A.4 The FactoMineR Package . . . . . . . . . . . . . . . . . . . . 269
A.4.1 The FactoMineR Package . . . . . . . . . . . . . . . . 269
A.4.2 The Drop-Down Menu . . . . . . . . . . . . . . . . . . 270
A.5 Answers to the Exercises . . . . . . . . . . . . . . . . . . . . 270
A.5.1 Exercises: Chapter 1 . . . . . . . . . . . . . . . . . . . 270
A.5.2 Exercises: Chapter 2 . . . . . . . . . . . . . . . . . . . 278
A.5.3 Exercises: Chapter 3 . . . . . . . . . . . . . . . . . . . 287
A.5.4 Exercises: Chapter 4 . . . . . . . . . . . . . . . . . . . 292

Bibliography 295

Index of the Functions 297

Index 301
This page intentionally left blank
Preface

This book is the English adaptation of the second edition of the book Statis-
tiques avec R which was published in 2008 and was a great success in the
French-speaking world. In this version, a number of worked examples have
been supplemented and new examples have been added. We hope that readers
will enjoy using this book for reference when working with R.

This book is aimed at statisticians in the widest sense, that is to say, all
those working with datasets: science students, biologists, economists, etc. All
statistical studies depend on vast quantities of information, and computerised
tools are therefore becoming more and more essential. There are currently
a wide variety of software packages which meet these requirements. Here we
have opted for R, which has the triple advantage of being free, comprehen-
sive, and its use is booming. However, no prior experience of the software
is required. This work aims to be accessible and useful for both novices and
experts alike.

This book is organised into two main parts: Part I focuses on the R soft-
ware and the way it works, and Part II on the implementation of traditional
statistical methods with R. In order to render them as independent as pos-
sible, a brief chapter offers extra help getting started (Chapter 5, A Quick
Start with R) and acts as a transition: It will help those readers who are
more interested in statistics than in software to be operational more quickly.

In the first chapter we present the basic elements of the software. The
second chapter deals with data processing (reading data from file, merging
factor levels, etc.), that is to say, common operations in statistics. As any
statistical report depends on clear, concise visualisation of results, in Chap-
ter 3 we detail the variety of possibilities available in this domain with R.
We first present the construction of simple graphs with the various available
options, and then detail the use of more complex graphs. Some programming
basics are then presented in Chapter 4: We explain how to construct your own
functions but also present some of the useful pre-defined functions to conduct
repetitive analyses automatically. Focusing on the R software itself, this first
main part enables the reader to understand the commands used in subsequent
statistical studies.

The second part of the book offers solutions for dealing with a wide

xiii
xiv R for Statistics

range of traditional statistical data processing techniques: hypothesis tests

(Chapter 6), regression methods (Chapter 7), analyses of variance and co-
variance (Chapter 8), classification methods (Chapter 9), exploratory multi-
variate analysis methods (Chapter 10), and clustering methods (Chapter 11).
Each technique is preceded by a concrete example and is dealt with indepen-
dently in its own specific worked example. Following a short presentation
of the method, the R command lines are explicitly detailed and the results
commented. Readers are able to download the datasets from the address
http://www.agrocampus-ouest.fr/math/RforStat, where they can easily
find all the results described in this work. R is a command line interface, but
another easy-to-use graphical interface is available: At the end of each worked
example there is a section called Rcmdr Corner which can be used to process
the example using this second interface. This section also helps readers get to
know the R commands better as, with each click on the drop-down menu, the
lines of code corresponding to the chosen action are generated.

To draw this preface to a close, we would like to thank Jerome Pages and
Dominique Dehay, who enabled us to write the French version under the best
possible conditions. We would also like to thank Rebecca Clayton for the
English translation and Agrocampus Ouest for financing this translation.
 Part I

An Overview of R

1
This page intentionally left blank
1
Main Concepts

This chapter provides a general presentation of R software. We first describe

how to install it, the differences in use depending on the operating system,
and the different help available. We then go on to list the various elements
used in R (vectors, matrices, etc.), as well as a few comments about their
functions. We conclude by presenting the packages, or libraries, of external
programs which will be useful throughout this book.

1.1 Installing R
R is an open source software environment for statistics freely distributed by
CRAN (Comprehensive R Archive Network) at the following address http:
//cran.r-project.org/. The installation of R varies according to the oper-
ating system (Windows, Mac OS X, or Linux) but the functions are exactly
the same and most of the programs are portable from one system to another.
Installing R is very simple, just follow the instructions.

1.2 Work Session

To use R, first open a work session. R then opens a window with the following
notice:
R version 2.14.1 (2011-12-22)
Copyright (C) 2011 The R Foundation for Statistical Computing
ISBN 3-900051-07-0
Platform: i386-pc-mingw32/i386 (32-bit)

R is free software and comes with ABSOLUTELY NO WARRANTY.

You are welcome to redistribute it under certain conditions.
Type license() or licence() for distribution details.

Natural language support but running in an English locale

4 R for Statistics

R is a collaborative project with many contributors.

Type contributors() for more information and
citation() on how to cite R or R packages in publications.

Type demo() for some demos, help() for on-line help, or

help.start() for an HTML browser interface to help.
Type q() to quit R.

>
R then waits for an instruction, as indicated by the symbol > at the begin-
ning of the line. Each instruction must be validated by Enter to be run. If
the instruction is correct, R again waits for another instruction, as indicated
by >. If the instruction is incomplete, R yields the symbol +. You must
then complete the instruction or take control again by typing Ctrl + c or
Esc. If the instruction is incorrect, an error message will appear.

For each project, we recommend you create a file in which an image of the
session concerning this project will be saved. We also recommend that users
write the commands in a text file in order to be able to use them again when
needed.
By default, R saves all the objects created (variables, results tables, etc.).
At the end of the session, these objects can be saved in an image of the session
.RData, using the command save.image() or when closing a session. To close
a session in R, use the quit function:
> q()
Save workspace image? [y/n/c]:
The backup file can be found in the project file (we will see later how
to change the destination). Saved objects are thus available for future use
using the load function. We will now look at the differences between a session
opened in Linux, in Windows, and on a Mac.

1.2.1 Session in Linux

To open an R session, type the command R in a command window and indicate
the path to the working directory using the command setwd(). We can also
open an R session directly in the directory in which we wish to work. The
destination can be checked using getwd().

1.2.2 Session in Windows

To begin an R session, click on the R icon on the desktop to open a window.
The data and the commands will be saved in the same place as R was opened,
that is to say, in the same place as R is installed.
Main Concepts 5

The best way to change the working directory is to use the setwd() func-
tion or by going to File then Change dir....

To close a session in R, go to File, select Exit (or equivalently type q() in

the command window) and answer yes when asked Save workspace image?
An R icon will then be created in the working directory. When you click
on this icon, a new session will open directly in the given directory and the
objects created in the previous session will be available.

1.2.3 Session on a Mac

There are two ways of opening an R session. The first is to open the working
directory and type the R command. The second option is to click on the R
icon in the Applications menu and an R window will open. The working data
and the commands used must therefore be saved in the same place as R was
opened. It is also possible to choose the directory in which we wish to work
using the option Change Working Directory of the Misc icon.

To close a session in R, go to R, then Quit R and choose whether or not to

save the session. A .RData file is created in the working directory. However,
it is not possible to view hidden files with Finder and it is therefore impossible
to click on this file.

1.3 Help
1.3.1 Online Assistance
For help with a given function, for example the mean function, simply use
> help(mean)

or
> ?mean
Help is displayed directly in the interface. It is also possible to display
help in HTML format in a web browser. For this option, type

> help.start()
HTML help offers a search function as well as functions grouped under differ-
ent headings. This is probably the most appropriate help option for a novice
user.
6 R for Statistics

Within the help section for each function you will find the list of arguments
to be detailed, along with their default value where appropriate, the list of
results returned by the function, and the name of the programmer. At the
end of the help section, one or more examples of how the function is used are
presented and can be copied and pasted into R in order to better understand
its use.

1.3.2 Help on CRAN

The help provided by R is not always sufficient. Indeed, sometimes for a
specific question, a user will not find an appropriate solution. For this type of
question, there is a website where users can search for solutions in the archives
of R users. This help can be accessed from an R command:
> RSiteSearch("mean weighted")
or from the website http://finzi.psych.upenn.edu/search.html. A FAQ
section is also available at the address http://cran.r-project.org/doc/
FAQ/R-FAQ.html. Usually, the question has already been asked by other users
and the answer to your problem can be found here. Finally, you can also
search directly on the Internet using a search engine such as Google, with
the keywords R or CRAN or use the website http://www.rseek.org/ which is
dedicated to R search on the Internet.

1.4 R Objects
R uses functions or operators which act on objects (vectors, matrices, etc.).
All the functions in this book are written in bold. This section deals with
presenting and manipulating different R objects.

1.4.1 Creating, Displaying and Deleting Objects

An object can be created by being assigned to one of the three operators <-,
->, = and by naming the object:
> b<-41.3 # set up object b consisting of one number: 41.3
> x<-b # b is assigned to x
> x=b # b is assigned to x
> b->x # b is assigned to x
For the rest of this book we use the assigning operator <-. The symbol #
indicates that all the rest of the line will not be interpreted by R; this is so
that comments can be added to a program.
Existing R data can also be retrieved using the following function data:
Main Concepts 7

> data(iris)

This command is used to load the iris dataset, which is now available. Data
can also be read directly from a file using the function read.table (see Sec-
tion 2.1), for example,
> ozone <- read.table("ozone.txt",header=TRUE)
If an object does not exist, the allocation creates it. If not, the allocation
replaces the previous value without warning. We display the value of an
object x via the command
> print(x)
or, more simply

> x
By default, R saves all the objects created in the session. We therefore
recommend that you delete these objects regularly. To find the objects from
a session, use the function objects() or ls(). To delete the object x, type

> rm(x)
and to delete multiple objects
> rm(object1,object2,object3)
Finally, to delete a list of objects for which a part of the name is common, for
example the letter a, use
> rm(list=ls(pattern=".*a.*"))

1.4.2 Type of Object

Before looking at the different objects in R, it is important to know the main
modes or types. These types are
Null (empty object): NULL

Logical (Boolean): TRUE, FALSE or T, F

Numeric (real number): 1, 2.3222, pi, 1e-10
Complex (complex number): 2+0i, 2i
Character (chain of characters): hi, "K"

In order to know the mode of an object x of R, simply run the command

> mode(x)
8 R for Statistics

It is also possible to test which mode a particular object x belongs to. The
results are Booleans with values that are either TRUE or FALSE:
> is.null(x)
> is.logical(x)
> is.numeric(x)
> is.complex(x)
> is.character(x)
It is possible to convert the object x of a given mode explicitly using the
following commands:
> as.logical(x)
> as.numeric(x)
> as.complex(x)
> as.character(x)
However, we must be careful with regard to the meaning of these conver-
sions. R always yields a result for every conversion instruction even if it is
meaningless. For example, from the following conversion table:
From To Function Conversions
Logical Numeric as.numeric FALSE 0
TRUE 1
Logical Character as.character FALSE "FALSE"
TRUE "TRUE"
Character Numeric as.numeric "1", "2", ... 1, 2, ...
"A", ... NA
Character Logical as.logical "FALSE", "F" FALSE
"TRUE", "T" TRUE
Other characters NA
Numeric Logical as.logical 0 FALSE
Other numerics TRUE
Numeric Character as.character 1, 2, ... "1", "2", ...

Each object has two intrinsic attributes: its mode mode() and its length
length(). There are also other specific attributes which vary according to the
type of object: dim, dimnames, class, names. It is possible to request a list of
these specific attributes by running the command
> attributes(object)

1.4.3 The Missing Value

For a number of reasons, certain elements of data may not be collected during
an experiment or study. These elements are known as missing data. They
are therefore not available for the user, and R denotes them NA for Not
Available. This is not a true mode and it has its own calculation rules. For
example,
Main Concepts 9

> x <- NA
> print(x+1)
> NA
In order to know where to find the missing values for an object x, we must
ask the question:
> is.na(x)

This yields a Boolean of the same length as x. The question is also asked ele-
ment by element. In the case of a vector, this yields a logical vector the same
length as x with TRUE if the corresponding element in x is NA, and FALSE if not.

Other special values should be noted here: Inf for infinity and NaN for
Not a Number, values resulting from calculation problems; for example
exp(1e10) or log(-2), respectively.

1.4.4 Vectors
Vectors are atomic objects, that is to say, they are of a unique type (null, log-
ical, numeric, etc.), made up of a set of values called components, coordinates
or elements. The attribute length, obtained using the function length, gives
the number of elements in the vector.

1.4.4.1 Numeric Vectors

There are a number of different ways of constructing a vector, some of which
are detailed below:

Construction by the collect function c:

> x <- c(5.6,-2,78,42.3) # numeric vector with 4 entries

> x
[1] 5.6 -2.0 78.0 42.3
> x <- c(x,3,c(12,8)) # vector with 7 entries
> x
[1] 5.6 -2.0 78.0 42.3 3.0 12.0 8.0
> x <- 2 # vector of length 1
> x
[1] 2

Construction by sequence operator ::

> 1:6
[1] 1 2 3 4 5 6

Construction by the seq function:

10 R for Statistics

> seq(1,6,by=0.5)
[1] 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0
> seq(1,6,length=5)
[1] 1.00 2.25 3.50 4.75 6.00

Construction by the rep function:

> rep(1,4)
[1] 1 1 1 1
> rep(c(1,2),each=3)
[1] 1 1 1 2 2 2

Construction by the scan function. R will then ask you to enter the el-
ements one at a time. Using the argument n=3, we specify that three
elements will be collected. If we do not specify n, the collection is stopped
by putting in an empty value.

> marks <- scan(n=3)

1: 12
2: 18
3: 15
Read 3 items

1.4.4.2 Character Vectors

It is possible to create character vectors in the same way with the functions c
or rep. For example,
> x <- c("A","BB","C1")
> x
[1] "A" "BB" "C1"
> x <- rep(A,5)
> x
[1] "A" "A" "A" "A" "A"
Even if R interprets " and in the same way, we will use " from here on in.
It is also possible to create character vectors using the vector letters which
contains the letters of the alphabet. The function format is used to organise
numerical data in chains of characters of the same length (see also the toString
function).

Another way of doing this is to manipulate different R objects and to

concatenate them or to extract part of them. To concatenate the objects, use
the paste function:
> paste("X",1:5,sep="-")
[1] "X-1" "X-2" "X-3" "X-4" "X-5"
Main Concepts 11

> paste(c("X","Y"),1:5,"txt",sep=".")
[1] "X.1.txt" "Y.2.txt" "X.3.txt" "Y.4.txt" "X.5.txt"
> paste(c("X","Y"),1:5,sep=".",collapse="+")
[1] "X.1+Y.2+X.3+Y.4+X.5"
The argument collapse groups together all the elements in a vector with a
length of 1. For an extraction, use the function substr:
> substr("freerider",5,9)
[1] "rider"
This extracts the characters ranked from 5 to 9 from freerider which yields
"rider".

1.4.4.3 Logical Vectors

Boolean vectors are usually generated with logical operators: >, >=, <,
<=, ==, !=, etc. They can also be generated using the functions seq,
rep and c. They can be used to make complex selections (see Section 1.4.4.4,
p. 12) or conditions operations (see Section 4.1, p. 87). Let us examine the
following example:
> 1>0
[1] TRUE
This command yields a logical vector with a length of 1 which is TRUE, as 1 is
greater than 0. The command
> x>13
yields a logical vector of the same length as x. Its elements have the value
TRUE when the corresponding element meets the given condition (here strictly
higher than 13) and the value FALSE if it does not. During the calculations, the
logicals are changed into numerics, following the convention that the FALSE
become 0 and the TRUE become 1. Let us examine an example of this. An
object test is created, which is the vector of the logicals (FALSE,FALSE,TRUE).
We then calculate the following product:
> x <- c(-1,0,2)
> test <- x>1
> (1+x^2)*(x>1)
[1] 0 0 5
The functions all and any can also be used. The function all yields TRUE if
all the elements meet the condition and if not, it yields FALSE. The function
any yields TRUE if one of the elements meets the condition, and if not, FALSE:
> all(x>1)
[1] FALSE
> any(x>1)
[1] TRUE
12 R for Statistics

1.4.4.4 Selecting Part of a Vector

Selections are made using the selection operator [ ] and a selection vector:
> x[indexvector]
The selection vector can be a vector of positive integers, negative integers or
logicals. The most natural selection is that of positive integer vectors. The
integers are indices of the elements to be selected and must be between 1 and
length(x). The length of the index vector can vary:
> v <- 1:100
> v[6] # sixth element of v
> v[6:8] # 6th to 8th elements
> v[c(6,6,1:2)] # elements 6, 6, 1, and 2
> v[10:1] # elements 10, 9 ..., and 1
Another method is to remove the elements of the vector that you do not
wish to retain: This is selection by negative integer vectors. The index vector
indicates the indices of the elements to be excluded from the results:
> v[-(1:5)] # v without its five first elements
> v[-c(1,5)] # v without its first and its fifth elements
Another method is to select elements of the vector according to their value,
or other elements originating from other R objects. This leads to selection by
logical vectors. Using this selection method it is possible to extract specific
elements which we already know how to characterise with a circumlocution,
or by logical condition: the element of sex M or the element with a value
of less than 5 and/or more than or equal to 12:
> v <- 1:15
> print(v)
[1] 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
> v[(v<5)]
[1] 1 2 3 4
> v[(v<5)&(v>=12)] # & means "and"
integer(0)
As no elements are both less than 5 and greater than 12, the function yields
an empty set with integer(0).
> v[(v<5)|(v>=12)] # | means "or"
[1] 1 2 3 4 12 13 14 15
It is also possible to select the values of a vector from the values of another
vector of the same length:
> T <- c(23, 28, 24, 32)
> O3 <- c(80, 102, 87, 124)
> O3[T>25]
[1] 102 124
Main Concepts 13

1.4.4.5 Selection in Practice

We can substitute a selected subset with new values:
> x[is.na(x)] <- 0 # missing values (NA) are replaced by 0
> y[y<0]<- -y[y<0] #negative elts are replaced by their opposite
This action means taking the absolute value:
> y <- abs(y)
It is common to select the maximum or the minimum of a given function.
Below are three possibilities for searching for the coordinate of the vectors
smallest element x:
> which.min(x)
> which(x==min(x))
> (1:length(x))[x==min(x)]
The logical operator == is used to test equality and yields a Boolean.

1.4.5 Matrices
Matrices are atomic objects, that is to say, they are of the same mode or
type for all values. Each value of the matrix can be located by its row and
column numbers. The two intrinsic attributes of an R object are its length,
which here corresponds to the total number of elements in the matrix, and
its in the object mode, which here corresponds to the mode of the elements
in this matrix. The matrices also include the dimension attribute dim, which
yields the number of rows and the number of columns. They can also possess
an optional attribute dimnames (see Section 1.4.7.4, p. 21). Hereafter are the
main ways of creating a matrix. The most widely used is the matrix function
which takes as its arguments the vector of elements and the number of rows
or columns in the matrix:
> m <- matrix(c(1,17,12,3,6,0),ncol=2)
> m
[,1] [,2]
[1,] 1 3
[2,] 17 6
[3,] 12 0
> m <- matrix(1:8,nrow=2)
> m
[,1] [,2] [,3] [,4]
[1,] 1 3 5 7
[2,] 2 4 6 8
By default, R ranks the values in a matrix by column. To rank the elements
by row, use the argument byrow:
14 R for Statistics

> m <- matrix(1:8,nrow=2,byrow=TRUE)

> m
[,1] [,2] [,3] [,4]
[1,] 1 2 3 4
[2,] 5 6 7 8
If the length of the vector is different from the number of elements in the
matrix, R fills the whole matrix. If the vector is too large, it takes the first
elements; and if the vector is too small, R repeats it:
> m <- matrix(1:4,nrow=3,ncol=3)
> m
[,1] [,2] [,3]
[1,] 1 4 3
[2,] 2 1 4
[3,] 3 2 1
It is possible to fill a matrix with one single element without having to create
the element vector:
> one <- matrix(1,nrow=2,ncol=4)
> one
[,1] [,2] [,3] [,4]
[1,] 1 1 1 1
[2,] 1 1 1 1
Vectors are not considered by R to be matrices. However, it is possible to
change a vector into a one-column matrix using the function as.matrix:
> x <- seq(1,10,by=2)
> x
[1] 1 3 5 7 9
> as.matrix(x)
[,1]
[1,] 1
[2,] 3
[3,] 5
[4,] 7
[5,] 9

1.4.5.1 Selecting Elements or Part of a Matrix

The position of an element within a matrix is generally indicated by its row
and column numbers. Thus, to select the element (i, j) from the matrix m, we
write
> m[i,j]
Main Concepts 15

It is unusual to only need to select one element in a matrix. Generally, multiple

rows and/or columns are selected. Let us examine the different cases we might
encounter in practice:
Selection by positive integers:

> m[i,]

N.B. This yields row i in the form of a vector.

> m[i,,drop=F]

gives row i in the form of a one-row matrix and not of a vector, and thus
the row name can be retained.

> m[,c(2,2,1)]

gives the second, the second and the first columns: This is a three-column
matrix.
Selection by negative integers:

> m[-1,] # matrix m without its first row

> m[1:2,-1] # two first rows of m without its first column

Selection by logicals:

> m <- matrix(1:8,ncol=4,byrow=T)

> m
[,1] [,2] [,3] [,4]
[1,] 1 2 3 4
[2,] 5 6 7 8

The following instruction yields only the columns of m for which the value
on the first line is strictly greater than 2:

> m[,m[1,]>2]
[,1] [,2]
[1,] 3 4
[2,] 7 8

This is of course a matrix, whereas the following instruction yields a vector

containing the values of m which are greater than 2:

> m[m>2]
[1] 5 6 3 7 4 8
16 R for Statistics

The following instruction replaces the values of m which are greater than
2 with NA:

> m[m>2] <- NA

> m
[,1] [,2] [,3] [,4]
[1,] 1 2 NA NA
[2,] NA NA NA NA

1.4.5.2 Calculating with the Matrices

> m <- matrix(1:4,ncol=2)
> m
[,1] [,2]
[1,] 1 3
[2,] 2 4
> n <- matrix(3:6,ncol=2,byrow=T)
> n
[,1] [,2]
[1,] 3 4
[2,] 5 6
> m+n
[,1] [,2]
[1,] 4 7
[2,] 7 10
> m*n # product element by element
[,1] [,2]
[1,] 3 12
[2,] 10 24
> sin(m) # sinus element by element
> exp(m) # exponential element by element
> m^4 # power 4 element by element

The following table gives the most useful functions for linear algebra.

Function Description
X%*%Y Product of matrices
t(X) Transposition of a matrix
diag(5) Identity matrix of order 5
diag(vec) Diagonal matrix with the values of vector vec in the
diagonal
crossprod(X,Y) Cross product (t(X)%*%Y)
det(X) Determinant of matrix X
svd(X) Singular value decomposition
eigen(X) Matrix diagonalisation
solve(X) Matrix inversion
Main Concepts 17

Function Description
solve(A,b) Solving linear systems
chol(Y) Cholesky decomposition
qr(Y) QR decomposition

Below is an example of how the functions eigen and solve are used:
> A <- matrix(1:4,ncol=2)
> B <- matrix(c(5,7,6,8),ncol=2)
> D <- A%*%t(B)
> D
[,1] [,2]
[1,] 23 31
[2,] 34 46
> eig <- eigen(D)
> eig
$values
[1] 68.9419802 0.0580198

$vectors
[,1] [,2]
[1,] -0.5593385 -0.8038173
[2,] -0.8289393 0.5948762
The eigen function yields the eigenvalues in the object values and the eigen-
vectors in the object vectors. To extract the first eigenvector, simply write

> eig$vectors[,1]
For example, to solve the following system of equations:

23 x + 31 y = 1
34 x + 46 y = 2

which is the same as writing Dz=V, with V=c(1,2), we use the solve function
> V <- c(1,2)
> solve(D,V)
[1] -4 3
The solution is therefore x = 4 and y = 3.

1.4.5.3 Row and Column Operations

In this subsection we present a number of useful functions:
Dimensions: dim(X), nrow(X) and ncol(X), respectively, give the dimen-
sion, the number of rows and the number of columns of X:
18 R for Statistics

> X <- matrix(1:6,ncol=3)

> X
[,1] [,2] [,3]
[1,] 1 3 5
[2,] 2 4 6
> ncol(X)
[1] 3
> nrow(X)
[1] 2
> dim(X)
[1] 2 3

These functions yield NULL if X is a vector.

Concatenation: by column using the function cbind, by row using the
function rbind (for further information, see Section 2.5, p. 43):

> cbind(c(1,2),c(3,4))
[,1] [,2]
[1,] 1 3
[2,] 2 4

The apply function is used to apply a function f to the rows (MARGIN=1)

or the columns (MARGIN=2) of the matrix (for further information, see
Section 4.2, p. 90). For example,

> apply(X,MARGIN=2,sum) # sum by column

[1] 3 7 11
> apply(X,1,mean) # mean by row
[1] 3 4

sweep(X,2,stdev,FUN="/") is used to divide each column (as MARGIN=2)

of X by an element of the vector stdev (for further information, see Sec-
tion 4.2, p. 93). This example is used to standardise the variables in a
matrix.
Remark
Data cubes or higher dimensions tables can be constructed using the function
array. These tables are manipulated in a similar way as the matrices.

1.4.6 Factors
Factors are vectors which enable the user to manipulate qualitative data.
Length is determined by the length function, mode by mode and the categories
of the factor by levels. They form a class of objects and are treated differently
depending on the function, such as the plot function for graphs. Factors can be
non-ordinal (male, female) or ordinal (level of skiing ability). Three functions
can be used to create factors:
Main Concepts 19

The factor function

> sex <- factor(c("M","M","F","M","F","M","M","M"))

> sex
[1] M M F M F M M M
Levels: F M

We can also name each level as the factor is constructed:

> sex <- factor(c(2,2,1,2,1,2,1),labels=c("woman","man"))

> sex
[1] man man woman man woman man man man
Levels: woman man

The ordered function

> ability <- ordered(c("beginner","beginner","champion",

"champion","intermediate","intermediate","intermediate",
"champion"),levels=c("beginner","intermediate","champion"))
> ability
[1] beginner beginner champion champion
[5] intermediate intermediate intermediate champion
Levels: beginner < intermediate < champion

The as.factor function

> somersault <- c(1:5,5:1)

> somersault
[1] 1 2 3 4 5 5 4 3 2 1
> somersault.f <- as.factor(somersault)
> somersault.f
[1] 1 2 3 4 5 5 4 3 2 1
Levels: 1 2 3 4 5

To know the levels, the number of levels and the number of factors per
level of somersault.f, we use

> levels(somersault.f)
[1] "1" "2" "3" "4" "5"
> nlevels(somersault.f)
[1] 5
> table(somersault.f)
somersault.f
1 2 3 4 5
2 2 2 2 2
20 R for Statistics

The table function can also be used to construct cross tabulations (Sec-
tion 2.6, p. 46).
To convert the factors into numerics, we use the following instructions:
> x <- factor(c(10,11,13))
> as.numeric(x)
[1] 1 2 3
> as.numeric(as.character(x))
[1] 10 11 13
Details of this conversion are given in Section 2.3.1 (p. 35).

1.4.7 Lists
A list is a heterogeneous object. It is a set of ranked objects which do not
always have the same mode or length. The objects are referred to as compo-
nents of the list. These components can be named. Lists have the same two
attributes as vectors (length and mode), as well as the additional attribute
names. Lists are important objects as all the functions which yield multiple
objects do so in the form of a list (see Section 1.5).

1.4.7.1 Creating a List

> myvector <- seq(2,10,by=3)
> mymatrix <- matrix(1:8,ncol=2)
> myfactor <- factor(c("M","M","F","M","F","M","M","M"))
> myordered <- ordered(c("beginner","beginner","champion",
"champion","intermediate","intermediate","intermediate",
"champion"),levels=c("beginner","intermediate","champion"))
> mylist <- list(myvector,mymatrix,myfactor,myordered)
> length(mylist)
[1] 4
> mode(mylist)
[1] "list"
We name the components of the list:
> names(mylist)
NULL
> names(mylist) <- c("vec","mat","sex","skilevel")
> names(mylist)
[1] "vec" "mat" "sex" "skilevel"

1.4.7.2 Extraction
To extract a component from the list, we can simply indicate the position of
the element that we want to extract. [[ ]] are used to identify the element
in the list:
Main Concepts 21

> mylist[[3]]
[1] M M F M F M M M
Levels: F M
> mylist[[1]]
[1] 2 5 8
We can also use the name of the element, if it has one, which can be written
in two ways:
> mylist$sex
[1] M M F M F M M M
Levels: F M

> mylist[["sex"]]
[1] M M F M F M M M
Levels: F M
It is possible to extract multiple elements from the same list, thus creating a
sub-list. N.B. Here we use [ ] and not [[ ]]:
> mylist[c(1,3)]
$vec
[1] 2 5 8

$sex
[1] M M F M F M M M
Levels: F M

1.4.7.3 A Few List Functions

lapply successively applies a function (such as the mean, the variance, etc.)
to each of the components of a list.
unlist(mylist) creates a single vector containing all the elements of the
list. As all the elements of a vector are always of the same mode, automatic
conversions in R must be used with caution.
c(list1,list2) concatenates two lists.

1.4.7.4 Dimnames List

The dimnames list is an optional attribute of a matrix which contains two
components: row names and column names. For example,
> X <- matrix(1:12,nrow=4,ncol=3)
> rowname <- c("row1","row2","row3","row4")
> colname <- c("col1","col2","col3")
> dimnames(X) <- list(rowname,colname)
> X
22 R for Statistics

col1 col2 col3

row1 1 5 9
row2 2 6 10
row3 3 7 11
row4 4 8 12
It is possible to select the rows and/or the columns using dimnames:
> X[c("row4","row1"),c("col3","col2")]
col3 col2
row4 12 8
row1 9 5
It is also possible to modify (or delete) the row and/or column names using
> dimnames(X) <- list(NULL,dimnames(X)[[2]])
> X
col1 col2 col3
[1,] 1 5 9
[2,] 2 6 10
[3,] 3 7 11
[4,] 4 8 12

1.4.8 Data-Frames
Data-frames are special lists with components of the same length but with po-
tentially different modes. The data tables generally used in statistics are often
referred to as data-frames. Indeed, a data table is made up of quantitative
and/or qualitative variables taken from the same individuals.
The main ways of creating a data-frame are to use the data.frame function
to concatenate vectors of the same size but which may have different modes; or
the read.table function (see Section 2.1) to read a data table; the as.data.frame
function for explicit conversion. For example, if we concatenate the numeric
vector vec1 and the alphabetic vector vec2:
> vec1 <- 1:5
> vec2 <- c("a","b","c","c","b")
> df <- data.frame(name.var1 = vec1, name.var2 = vec2)
> df
name.var1 name.var2
1 1 a
2 2 b
3 3 c
4 4 c
5 5 b
In order to extract components from the data-frame, we can use the same
methods as those presented for lists or matrices.
Main Concepts 23
Remarks
To convert a matrix to a data-frame, use the as.data.frame function.

To convert a data-frame to a matrix, use the data.matrix function.

1.5 Functions
Functions are R objects. A great number of functions are predefined in R, for
example, mean which calculates the mean, min which calculates the minimum,
etc. It is also possible to create your own functions (see Section 4.3). Functions
admit arguments as input and yield results as output.

1.5.1 Arguments of a Function

The arguments of a function can be compulsory or optional. When they are
optional, they nonetheless have a default value.
Let us take the example of the function rnorm which generates random
numbers according to the normal distribution. This function admits three
arguments: n the number of values, mean the mean, and sd the standard
deviation of the distribution. These arguments are by default set at 0 and 1.
In order to be able to observe the same results from one simulation to another,
you must set the seed of the random number generator using the function
set.seed to a given value, here 13.

> set.seed(13) # specify generator seeds

> rnorm(n=3)
[1] 0.5543269 -0.2802719 1.7751634
These three numbers were thus chosen according to the standard normal dis-
tribution N (0, 1). If we want to select four numbers according to a normal
distribution with mean 5 and standard deviation 0.5, we use
> set.seed(13) # specify generator seeds
> rnorm(n=4,mean=5,sd=0.5)
[1] 5.277163 4.859864 5.887582 5.093660

It is not crucial to specify the name of the arguments, as long as the order
is respected. This order is defined when the function is created. The args
function yields the arguments of a function:
> args(rnorm)
function (n, mean = 0, sd = 1)
NULL
24 R for Statistics

It is also possible to consult the help section for the function.

Some functions admit optional unnamed arguments. These are represented
by ..., as for the plot function, for example,
> args(plot)
function (x, y, ...)
NULL
This function draws y according to x, with ... used to control the appearance
of the graph (see Section 3.1.1, p. 52).

1.5.2 Output
Most of the functions yield multiple results and the entirety of these results is
contained within a list. To visualise all these outputs, you first need to know
all the elements. This information can be accessed using the names function.
Use [[j]] to extract element j from the list, or indeed $ if the elements in
the list are named (see Section 1.4.7, p. 20).

1.6 Packages
1.6.1 What Is a Package ?
A package or library of external programs is simply a set of R programs which
supplements and enhances the functions of R. Packages are generally reserved
for specific methods or fields of application.
There are more than 3000 packages (as of June 2011). A certain number
of these are considered essential (MASS, rpart, etc.) and are provided with
R. Others are more recent statistical advances and can be downloaded freely
from the CRAN network. Furthermore, many more packages are available
within projects, such as BioConductor which contains packages dedicated to
processing genomic data.

1.6.2 Installing a Package

In this section we download a package and install it. Packages need only be
installed once. Once they are installed, simply call the package using the
library function each time you open an R session:
> library(package.name)
Packages are available on the CRAN network (http://cran.r-project.
org/). A number of mirror sites, or exact copies of the CRAN site, are
available. To install a package from CRAN, simply run the install.packages
function:
Main Concepts 25

> install.packages(dependencies=TRUE)
Then choose the nearest mirror to you and select the package to install, such
as the FactoMineR package. You then simply need to load it in order to use
it in an R session:
> library(FactoMineR)
It is also possible to download a package using one of the drop-down menus
in Windows (Packages Install package(s)...) or Mac OS X (Packages
& Data Package Installer). It is also possible to configure the use of a
proxy (help(download.file)) and the destinations where the packages will
be installed. All the packages installed on a given computer can be listed
using the command
> library()

1.6.3 Updating Packages

Some packages constantly evolve, with new versions regularly available. It is
therefore important to update them from time to time. To do this, they must
not be in use. The best solution is to update packages when you open an R
session by using
> update.packages()
to choose a mirror site near you and answer any questions.
In Windows and Mac OS X, there is a drop-down menu for updating
packages: simply go to Packages Update packages ... and choose a
mirror site near you. A list of available package updates is then displayed,
with all packages selected by default.
N.B. Package updates depend on the version of R used. Some packages are
not compatible with past versions, or have not been compiled for these. We
therefore also recommend that you regularly update your version of R before
updating packages.

1.6.4 Using Packages

Once they are installed, packages must be called by the user before they
can be used. Suppose we want to generate three vectors of R2 according to
a multivariate

normal distribution with mean (0, 1) and covariance matrix
1 0.5 . This is possible thanks to the mvrnorm function from the MASS
0.5 1
package. In order to use it, first call the MASS package and then use the
function mvrnorm:
> library(MASS)
> set.seed(45) # specify generator seeds
> mvrnorm(3,mu=c(0,1),Sigma=matrix(c(1,0.5,0.5,1),2,2))
26 R for Statistics

[,1] [,2]
[1,] -0.07788253 1.6681649
[2,] -1.05816423 0.8399431
[3,] -0.49608637 0.8387077

1.7 Exercises
Exercise 1.1 (Creating Vectors)
1. Create the following vectors using the rep function:

vec1 = 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
vec2 = 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5
vec3 = 1 1 2 2 2 3 3 3 3 4 4 4 4 4

2. Create the following vector using the paste function:

vec4 = "A0)" "A1)" "A2)" "A3)" "A4)" "A5)" "A6)" "A7)"

"A8)" "A9)" "A10)"

3. letters is a vector containing the 26 letters of the alphabet. Find the

index of the letter q and create the vector containing a1,b2, until q
with the index of q.
Exercise 1.2 (Working with NA)
1. Fix the seed generator (set.seed) at 007 and create a vector vec1 of 100
numbers uniformly generated between 0 and 7 (runif ). Calculate the mean
and variance of vec1.
2. Create a vector vec2 equal to vec1 and select ten values at random (sample)
and replace them by NA (missing values). Find the positions of the NA using
the is.na function.
3. Calculate the mean and variance of the vector vec2.
4. Create a vector vec3 from vec2 by deleting the NA values and calculate
the mean and variance.
5. Create a vector vec4 from vec2 where the NA values are replaced by the
mean of the values of vec2. What happens to the mean and the variance of
vec4?
6. Create a vector vec5 from vec2 where the NA values are replaced by values
drawn from a Gaussian distribution with the mean and variance of vec2.
7. Same question as the previous one using values drawn from a Uniform
distribution between the minimum and the maximum of the sample.
Main Concepts 27

8. Create a vector vec7 from vec2 where the NA values are replaced by values
selected at random from vec2.
Exercise 1.3 (Creating, Manipulating and Inverting a Matrix)
1. Create the following matrix mat (with the column and row names):

column 1 column 2 column 3 column 4

row-1 1 5 5 0
row-2 0 5 6 1
row-3 3 0 3 3
row-4 4 4 4 2

2. Create a vector containing the diagonal elements of the matrix mat.

3. Create a matrix containing the first 2 rows of mat.
4. Create a matrix containing the last 2 columns of mat.

5. Calculate the determinant and then invert the matrix using the appropriate
functions.
Exercise 1.4 (Selecting and Sorting in a Data-Frame)
1. From the iris dataset available in R (use data(iris) to load it and then
iris[1:5,] to visualise the first five rows), create a sub-dataset comprising
only the data for the category versicolor of the variable species (call this
new dataset iris2).

2. Sort the data in iris2 in descending order according to the variable

Sepal.Length (you can use the order function).

Exercise 1.5 (Using the apply Function)

1. Calculate the benchmark statistics (mean, min, max, etc.) for the three
variables of the ethanol dataset (available in the lattice package).
2. Calculate the quartiles for each of the three variables. To do this, use the
apply function with the quantile function.

3. Again using the apply function, calculate all the deciles for each of the
three variables using the argument probs of the quantile function.

Exercise 1.6 (Selection in a Matrix with the apply Function)

1. From the matrix mat created in Exercise 1.3, create a matrix containing
the columns of mat having all values smaller than 6 (use apply and all).
2. In the same way, create a matrix containing the rows of mat which do not
contain 0.
Exercise 1.7 (Using the lapply Function)
1. Load the Aids2 dataset of package MASS and summarise the Aids2 dataset.
28 R for Statistics

2. Using the is.numeric and the lapply functions, create a Boolean vector of
length the number of columns of the data-frame Aids2 which takes the value
TRUE when the corresponding column of Aids2 is not numeric and FALSE when
it is numeric.
3. Select the qualitative variables of the Aids2 data-frame and assign the
result in a data-frame: Aids2.qual.

4. Give the levels of each qualitative variables of the Aids2.qual data-frame.

Exercise 1.8 (Levels of the Qualitative Variables in a Subset)

1. Load the Aids2 dataset of package MASS.

2. Select the rows of Aids2 where the sex is Male (M) and the state is not
Other. Put the result in a new data-frame: res.
3. Summarise the res dataset and check that a) the sex variable has no
female and b) the female level is still present.

4. Print the attributes of the sex variable (function attributes). Check that
a) there is a level attribute with F and M, and b) the class of the object is
factor.
5. Using as.character, transform the sex variable (of res) into an object
named sexc of type character. Check that sexc has no attributes.

6. Using as.factor, transform sexc into an object named sexf of type factor.
Check that the resulting object sexf has a) a level attribute, b) a class factor
and c) no more F level.
Conclusion: Transforming a factor into a vector of characters and transforming
back into a factor resets the levels.

7. Using Exercise 1.7, create a Boolean vector of length the number of columns
of res which takes the value TRUE for categorical variables and FALSE for
numeric variables.
8. Transform all factors of res into characters using lapply.

9. Transform back into factors using lapply.

2
Preparing Data

In statistics, data is the starting point for any analysis. It is therefore es-
sential to know how to master operations such as importing and exporting
data, changing types, identifying individuals with missing values, concatenat-
ing factor levels, etc. These different notions are therefore presented in this
chapter, which aims to be both concise and satisfactory in practice.

2.1 Reading Data from File

The data is first collected, and may be pre-processed using software, be it a
spreadsheet or statistical software. Each piece of software has its own storage
format; the simplest option is to exchange data in a format common to all
software, that is to say, in text format (extension .txt or .csv).
Data is generally contained within a file in which the individuals are pre-
sented in rows and the variables in columns. Text format contains data sepa-
rated by a column separator. The table below thus groups together measure-
ments of four variables (height, weight, shoe size, and sex) for three individ-
uals, and the column separator used here is ,. The first column features
the identifiers of the individuals, which here are their first names: all the first
names must therefore be different. This identifier is generally a string of char-
acters or the number of the individual. In some cases, individuals do not have
identifiers and/or the variables do not have names.
surname,height,weight,size,sex
tony,184,80,9.5,M
james,175.5,78,8.5,M
john,158,72,8,M
Let us suppose that this data is contained within a mydata.csv file within a
DATA file directory located in the current working directory (see Section 1.2.1
or 1.2.2 or 1.2.3). If inputting the data manually, make sure that you start a
new line at the end of the last row; this operation is automatic in spreadsheets.
The data is imported into the object tab as follows:
> tab <- read.table("DATA/mydata.csv",sep=",",header=TRUE,
dec=".",row.names=1)
30 R for Statistics

The file name is between speech marks (" or but not the same word-
processing speech marks found in Microsoft R Word). The sep option in-
dicates that the character which separates the columns is here ",". For a
space, use " ", and for a tabulation, "\t". The argument header indicates
whether or not the first row contains the names of the variables. The argu-
ment dec specifies that the decimal point is a "." (here for the number 175.5).
Finally, the argument row.names indicates that column 1 is not a variable but
the row identifier: here the individuals names.

N.B. Windows users are accustomed to specifying paths using \. With

R, even in Windows, we use /. Therefore, if the dataset is situated in the
drive C: in the Temp directory, the reading becomes
> tab <- read.table("C:/Temp/mydata.csv",sep=",",header=TRUE,
dec=".",row.names=1)
The path can also be a URL, as in the worked examples in the second part
of this book. Thus, to import the dataset, it is possible to use the following
command:
> decath <- read.table("http://www.agrocampus-ouest.fr/math/
RforStat/decathlon.csv",sep=",",dec=".",header=TRUE,
row.names=1)
Sometimes, a special character can be used to indicate that a value is
missing. In R, this character is "NA" (Not Available). It is important to
specify this character for missing data. Let us consider the following file
mydata2.csv:
height weight size sex
184 80 9.5 "M"
175.5 78 8.5 "M"
. 72 8 "M"
178 . 7 "F"
It is imported by modifying the default value of na.strings, which is "NA",
using
> tab <- read.table("DATA/mydata2.csv",sep=" ",header=TRUE,
na.strings = ".")
The quote option (not used in our example) can be used to control any char-
acters which may surround a string of characters. By default, this is a single
or double speech mark. For other options used to control the reading more
precisely, refer to the help of the read.table function.
It must be noted that the result of the reading is always a data-frame type
object. The types of variables are inferred by the reading function, which
can be wrong! It is therefore necessary to check each variable type using the
summary function. In the tab table from the previous inputting, we obtain
Preparing Data 31

> summary(tab)
height weight size sex
Min. :175.5 Min. :72.00 Min. :7.00 F:1
1st Qu.:176.8 1st Qu.:75.00 1st Qu.:7.75 M:3
Median :178.0 Median :78.00 Median :8.25
Mean :179.2 Mean :76.67 Mean :8.25
3rd Qu.:181.0 3rd Qu.:79.00 3rd Qu.:8.75
Max. :184.0 Max. :80.00 Max. :9.50
NAs : 1.0 NAs : 1.00
Here, the variable types are correct: height, weight and shoe size are quan-
titative and sex is qualitative with two categories (F and M). If this is not
the case, it may be necessary to change the variable type (see Section 2.3.1).
Sometimes, if the reading does not function correctly, the error may stem from
the text file itself. Below is a list of some of the most common problems:
Column separator incorrectly specified
Decimal point incorrectly specified (the variables may be mistakenly con-
sidered qualitative)

Tabulation replacing a space

Row or column name includes an apostrophe
Problem of open speech marks
In the first case, it is often difficult to find the mistake. One unsophisticated
solution is offered in Exercise 2.1.
Remarks
It is possible to import data using the Rcmdr package. We here advise you
to turn to the appendix on Rcmdr (see Appendix A.3, p. 267).
It is possible to import directly from databases such as access or mysql
using the RMySQL package.

It must also be noted that certain (older) spreadsheets impose a limit of

256 columns. If the data has more than 256 variables, as in spectrometry
or genomics, users of this type of spreadsheet should therefore present their
datasets in inverted format: variablesindividuals. In such cases, you must
first transpose the table once you have imported it. This poses no further
problems as long as all the variables are numerics. If even just one of the
variables is qualitative, transposition will mean that there is a problem with
variable type. Let us examine the following brief example:

sex M F M
age 18 17 22
32 R for Statistics

This file was imported in the traditional manner, remembering that the first
column contains the names of the variables:

> tab <- read.table("toyex.txt",sep=" ",header=F,row.names=1)

> summary(tab)
V2 V3 V4
18:1 17:1 22:1
M :1 F :1 M :1

We transpose the table (the rows become the columns and vice versa):
> tab2 <- t(tab)
> summary(tab2)
sex age
F:1 17:1
M:2 18:1
22:1
The sex variable poses no problem, but the age variable is summarised as a
three-level factor. In fact, the result is no longer a data-frame but a character
matrix.

> is.data.frame(tab2)
[1] FALSE
> is.matrix(tab2)
[1] TRUE

We must therefore convert our matrix to a data-frame (object which can

contain both qualitative and quantitative variables) and then transform the
second column into numerics. To do this, type the functions as.character and
as.numeric:
> tab2 <- data.frame(tab2)
> tab2[,"age"] <- as.numeric(as.character(tab2[,"age"]))
> tab2
sex age
V2 M 18
V3 F 17
V4 M 22

A detailed explanation of the second row is given in Section 2.3.1, p. 35. Of

course, if there are multiple numeric variables, they must all be checked, per-
haps with the help of a loop (see Section 4.1.2, p. 87). To attribute a row
heading other than V2, V3, V4, we use the rownames function. We can, for
example, attribute the row names NULL (no names for the rows):

> rownames(tab2) <- NULL

> tab2
Preparing Data 33

sex age
1 M 18
2 F 17
3 M 22

2.2 Exporting Results

Once the analyses have been conducted and the results obtained, it is often
important to send the results to other people or export them to use with other
software. For this purpose, a text file is again the most appropriate format.
Results are mostly presented in table format and we will therefore export a
table. Exporting a table is very simple:
> write.table(tab,"myfile.csv",sep=",",row.names=FALSE)
The object tab is exported to the file myfile.csv in the current working
directory (see Section 1.2.1 or 1.2.2 or 1.2.3). We can of course specify a
particular path. The exportation above can be used to control the column
separator fixed at "," and the file name myfile.csv. It is also possible to
control other options such as
Whether or not the results contain the column names (col.names); by
default col.names= TRUE
Whether or not the results contain the row names (row.names); by default
row.names= TRUE
Whether or not character strings are defined by speech marks; by default
quote= TRUE
The decimal point (dec) which is "." by default and the character string
for missing values (na) which is "NA" by default.
Therefore, to export without speech marks, without the row and column
names, and with a tabulation separator, we use
> write.table(tab,"myfile.txt",quote=FALSE,row.names=FALSE,
col.names=FALSE,sep="\t")
This type of file which mixes tabulation (separator) and spaces (missing value)
is not recommended as it can be difficult to find the origin of the problem,
should one occur. Judging from experience, the separator ";" seems the most
appropriate choice.
Remarks
The write.csv function is a possible alternative to the write.table function.
34 R for Statistics

The write.infile function from the FactoMineR package can be used to write
all the objects of a list in the same file, without having to specify each of
the objects in the list.

2.3 Manipulating Variables

2.3.1 Changing Type
It is often necessary to change the variable type. For example, following
inputting, a qualitative variable whose categories are coded using figures is
understood by R to be a quantitative variable. In such cases, they must be
changed from quantitative (type: numeric) to qualitative (type: factor).
Converting quantitative variables to qualitative variables is easy: simply
create a factor using the factor function. Let us take the example of a variable
X made up of numerics:
> X <- c(rep(10,3),rep(12,2),rep(13,4))
> X
[1] 10 10 10 12 12 13 13 13 13

There are classical methods for knowing whether a vector-type object is a

quantitative or a qualitative variable without displaying the entire vector.
The first way of doing this is to ask R about the type:
> is.factor(X)
[1] FALSE
> is.numeric(X)
[1] TRUE
The second is to summarise the variable (summary). When it is a quantita-
tive variable, the minimum, the maximum, the quartiles and the mean are
displayed. However, for a factor, the number of observations for the first six
levels of the qualitative variable is given:
> summary(X)
Min. 1st Qu. Median Mean 3rd Qu. Max.
10.00 10.00 12.00 11.78 13.00 13.00
This, of course, is a quantitative variable. This is the most practical in-
struction for data tables (data-frames) which have been imported. Indeed,
the summary is offered variable by variable and enables the user to quickly
identify any errors of variable type.
Converting to a factor is just as simple; use the factor function:
> Xfac <- factor(X)
Preparing Data 35

> Xfac
[1] 10 10 10 12 12 13 13 13 13
Levels: 10 12 13
> summary(Xfac)
10 12 13
3 2 4
Displaying a factor enables the user to distinguish it from a numeric, thanks
to the presence of different levels at the end of the display. This is also the
case for the summary yielded by summary.
Converting a factor to a numeric is a two-step process. We first convert
the factor to a character vector and then convert this vector to a numeric.
If we convert the factor directly to a numeric, the levels will be re-coded in
order (the first level will be 1, the second 2, etc.):

## coerce to factor + label change

> as.numeric(Xfac)
[1] 1 1 1 2 2 3 3 3 3
## coerce to factor without label change: 2 steps
> foo <- as.character(Xfac)
> foo
[1] "10" "10" "10" "12" "12" "13" "13" "13" "13"
> as.numeric(foo)
[1] 10 10 10 12 12 13 13 13 13

2.3.2 Dividing into Classes

Changing a quantitative variable to a qualitative variable, or dividing it into
classes, is common practice in statistics. Thus, for example, to construct the
axes of a Multiple Correspondence Analysis with quantitative variables (see
Worked Example 10.3, p. 230), they must be converted to qualitative variables.
Continuous variables such as X can be divided in one of two ways:
Division according to thresholds defined by the user
Automatic division into similar-sized classes

Let us start with the first option. We assign X a vector of fifteen random
numbers following a standard normal distribution:
> set.seed(654) ## set generator seeds
> X <- rnorm(15)
> X
[1] -0.76031762 -0.38970450 1.68962523 -0.09423560 0.095301
[7] 1.06576755 0.93984563 0.74121222 -0.43531214 -0.107260
[13] -0.98260589 -0.82037099 -0.87143256
36 R for Statistics

Let us divide this variable into three levels: between the minimum and 0.2,
between 0.2 and 0.2, and finally between 0.2 and the maximum. The cut
function can be used to conduct this step automatically. The classes are of the
format (ai ; ai+1 ]. In order to include the minimum of X in the first class, you
must use the argument include.lowest=TRUE. The first class will therefore
be of the form [a1 , a2 ]:
> Xfac <- cut(X,breaks=c(min(X),-0.2,0.2,max(X)),
include.lowest=TRUE)
> Xfac
[1] [-0.983,-0.2] [-0.983,-0.2] (0.2,1.69] (-0.2,0.2]
[5] (-0.2,0.2] (0.2,1.69] (0.2,1.69] (0.2,1.69]
[9] (0.2,1.69] [-0.983,-0.2] (-0.2,0.2] [-0.983,-0.2]
[13] [-0.983,-0.2] [-0.983,-0.2] [-0.983,-0.2]
Levels: [-0.983,-0.2] (-0.2,0.2] (0.2,1.69]
> table(Xfac)
Xfac
[-0.983,-0.2] (-0.2,0.2] (0.2,1.69]
7 3 5
The result of the division is a factor. When we list observations by category
(with the table function), we obtain unequal class sizes. If we require similar
class sizes for each of the three categories, we use the quantile function. This
function is used to calculate empirical quantiles and thus gives the required
thresholds. The thresholds are as follows:
The minimum, which is the empirical quantile associated with the proba-
bility 0: Pn (X < min(X)) = 0
The first threshold, which is the empirical quantile associated with the
probability 1/3: Pn (X threshold1 ) = 1/3
The second threshold, which is the empirical quantile associated with the
probability 2/3: Pn (X threshold2 ) = 2/3
The maximum, which is the empirical quantile associated with the prob-
ability 1: Pn (X max(X)) = 1
> mybreaks <- quantile(X,probs=seq(0,1,length=4))
> Xfac <- cut(X,breaks=mybreaks,include.lowest=TRUE)
> table(Xfac)
Xfac
[-0.983,-0.544] (-0.544,0.311] (0.311,1.69]
5 5 5

2.3.3 Working on Factor Levels

This section helps to resolve some of the classical problems when it comes to
organising data:
Preparing Data 37

How can multiple factor levels be merged if, for example, they have insuf-
ficient samples?

How can the levels of a factor be renamed?

The answer to these two questions is the same in R. Let us use the variable
Xfac from the previous section:
> Xfac
[1] [-0.983,-0.544] (-0.544,0.311] (0.311,1.69]
[4] (-0.544,0.311] (-0.544,0.311] (0.311,1.69]
[7] (0.311,1.69] (0.311,1.69] (0.311,1.69]
[10] (-0.544,0.311] (-0.544,0.311] [-0.983,-0.544]
[13] [-0.983,-0.544] [-0.983,-0.544] [-0.983,-0.544]
Levels: [-0.983,-0.544] (-0.544,0.311] (0.311,1.69]

We want to change the titles of the levels. To do this, simply assign a new
value to the levels of Xfac called using the levels function:
> levels(Xfac) <- c("lev1","lev2","lev3")
> Xfac
[1] lev1 lev2 lev3 lev2 lev2 lev3 lev3 lev3 lev3
[10] lev2 lev2 lev1 lev1 lev1 lev1
Levels: lev1 lev2 lev3
To merge two levels such as levels 1 and 3, simply rename the two levels with
the same name (i.e. repeat it):

> levels(Xfac) <- c("lev1+3","lev2","lev1+3")

> Xfac
[1] lev1+3 lev2 lev1+3 lev2 lev2 lev1+3 lev1+3 lev1+3
[9] lev1+3 lev2 lev2 lev1+3 lev1+3 lev1+3 lev1+3
Levels: lev1+3 lev2

The order in which the levels of a factor appear is not insignificant. For
ordinal factors, simply specify the order of the levels when the factor is created
(Section 1.4.6, p. 18). In theory this order is not important for factors whose
categories are not ordinal. However, in certain methods such as analyses of
variance (see Worked Example 8.1, p. 157), the first level is used as a base level,
and the others are all compared to it. In such cases, specifying the first factor
is thus essential. Les us consider, for example, a qualitative variable with
levels 1, 2 or 3, depending on whether the individual received the classic,
new or placebo treatment. The data is as follows:
> X <- c(1,1,2,2,2,3)

and to convert it to a factor:

38 R for Statistics

> Xfac <- factor(X,label=c("classic","new","placebo"))

> Xfac
[1] classic classic new new new placebo
Levels: classic new placebo
If we want placebo to be the base level, then we use the relevel function:
> Xfac2 <- relevel(Xfac,ref="placebo")
> Xfac2
[1] classic classic new new new placebo
Levels: placebo classic new
In order to control the order of the levels, simply recreate a factor by modifying
an existing one, specifying the order in which the levels should be displayed.
Thus, to choose the order of levels placebo then new and finally classic, we
use
> Xfac3 <- factor(Xfac,levels=c("placebo","new","classic"))
> Xfac3
[1] classic classic new new new placebo
Levels: placebo new classic
Sometimes we have to delete atypical individuals1 or work with a subset
of individuals. This may mean that there could be qualitative variables for
which a level has no individuals. Let us construct an example to illustrate
this problem:
> myfactor <- factor(c(rep("A",3),"B",rep("C",4)))
> myfactor
[1] A A A B C C C C
Levels: A B C
If we need to remove the fourth individual, the factor is as follows:
> myfactor <- myfactor[-4]
> myfactor
[1] A A A C C C C
Levels: A B C
> table(myfactor)
myfactor
A B C
3 0 4
The category B thus no longer has any individuals. We must therefore delete
this category. In order to do so, it is possible to reassign this category to
another, such as category A. Another method is to convert the factor into a
1 The last mass trials were a great success. There are going to be fewer but better

Russians. Greta Garbo in Ninotchka (1939), by Ernst Lubitsch.

Preparing Data 39

character vector and then to convert it back to a factor. The first conversion
removes the notion of levels, and the second returns it, taking into account
the data vector. As category B no longer exists, it disappears:
> myfactor <- as.character(myfactor)
> myfactor <- factor(myfactor)
> myfactor
[1] A A A C C C C
Levels: A C
Sometimes, when a categorys sample size is too small, we break down the
individuals into other categories at random. This process, known as ventila-
tion (random allocation), is examined in Exercise 2.6.

2.4 Manipulating Individuals

2.4.1 Identifying Missing Data
In R, missing data is represented by NA (Not Available). In order to identify
the individuals for which they are missing, simply use the is.na function, which
yields TRUE if the value is NA and FALSE if it is not. Let us construct a variable
measured from ten individuals by conducting a random draw according to
normal distribution N (0, 1). Then, we assign missing data, for individuals 3,
4 and 6 for example:
> set.seed(23)
> variable <- rnorm(10,mean=0,sd=1)
> variable[c(3,4,6)] <- NA
Now let us imagine that the statistician needs to know how to identify those
individuals with missing values. His first task will be to discover the identifiers,
or numbers for these individuals. In order to find them, simply use the is.na
function:
> miss <- is.na(variable)
> miss
[1] FALSE FALSE TRUE TRUE FALSE TRUE FALSE FALSE FALSE FALSE
Elements 3, 4 and 6 yield the value TRUE. They therefore carry the value NA
in the variable vector. The which function automatically yields the indices
of TRUE:
> which(miss)
[1] 3 4 6
40 R for Statistics

If we want to eliminate these individuals, we must conserve all those which

are not missing, that is to say, all the FALSE of the miss vector. To do this,
we conserve all the TRUE for negations of miss, that is to say !miss:
> variable2 <- variable[!miss]
> variable2
[1] 0.19321233 -0.43468211 0.99660511 -0.27808628 1.01920549
[6] 0.04543718 1.57577959
Another way to do this is to eliminate the coordinates using selection by
negative integers:
> variable3 <- variable[-which(miss)]
> variable3
[1] 0.19321233 -0.43468211 0.99660511 -0.27808628 1.01920549
[6] 0.04543718 1.57577959
Let us extend this to a data table which admits multiple variables. We
create a second variable and construct the table don:
> myfac <- factor(c(rep("M",3),NA,NA,rep("F",5)))
> tab <- cbind.data.frame(variable,myfac)
> tab
variable myfac
1 0.19321233 M
2 -0.43468211 M
3 NA M
4 NA <NA>
5 0.99660511 <NA>
6 NA F
7 -0.27808628 F
8 1.01920549 F
9 0.04543718 F
10 1.57577959 F
The statistical summary of the dataset is
> summary(tab)
variable myfac
Min. :-0.4347 F :5
1st Qu.:-0.1163 M :3
Median : 0.1932 NAs:2
Mean : 0.4454
3rd Qu.: 1.0079
Max. : 1.5758
NAs : 3.0000
The total of missing individuals is specified variable by variable. If we want
to eliminate those individuals with at least one missing value, we must first
identify them:
Preparing Data 41

> miss <- is.na(tab)

> miss
variable varqual
[1,] FALSE FALSE
[2,] FALSE FALSE
[3,] TRUE FALSE
[4,] TRUE TRUE
[5,] FALSE TRUE
[6,] TRUE FALSE
[7,] FALSE FALSE
[8,] FALSE FALSE
[9,] FALSE FALSE
[10,] FALSE FALSE

Next, the rows of tab which correspond to the rows of miss with at least one
TRUE must be eliminated: here the rows 3, 4, 5 and 6. In order to do this,
for each row we need to know whether there is at least one TRUE. The any
function applied to a vector yields TRUE if there is at least one TRUE in the
vector. To apply this function to every row in the table, we use apply on the
rows (MARGIN=1):

> deletion <- apply(miss,MARGIN=1,FUN=any)

> deletion
[1] FALSE FALSE TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE
The last vector therefore contains TRUE in each row which contains at least
one NA. The rows to be conserved are therefore those which correspond to all
of the TRUE of the negation of deletion. The refined data is therefore
> tab2 <- tab[!deletion,]
> tab2
variable myfac
1 0.19321233 M
2 -0.43468211 M
7 -0.27808628 F
8 1.01920549 F
9 0.04543718 F
10 1.57577959 F

Finally, if we want to know the row column pairs of individuals with a

missing value, we can use the option arr.ind (for array indices) of the which
function:
> which(is.na(tab),arr.ind=TRUE)
row col
[1,] 3 1
[2,] 4 1
42 R for Statistics

[3,] 6 1
[4,] 4 2
[5,] 5 2
The first piece of missing data is in row 3 and column 1 ..., the fifth piece of
missing data is in row 5 and column 2.

2.4.2 Finding Outliers

We use data offered by R to illustrate this problem. The dataset is called
kyphosis and is part of the rpart package (see also Section 1.6). Load the
package and then the dataset to the session in order to be able to use them:
> library(rpart)
> data(kyphosis)
We use the Number variable from this dataset. Let us draw a boxplot (Fig-
ure 2.1) for this variable:
> boxplot(kyphosis[,"Number"])
10
8
6
4
2

Figure 2.1
Boxplot for a variable with outliers (denoted by arrows).

The two individuals above the higher whisker (see Figure 2.1) are often consid-
ered to be outliers. How can we determine their number (or other identifier)?
One solution is to use the result of the boxplot function. The component out
from the ouput list gives us the outliers values:
> results <- boxplot(kyphosis[,"Number"])
> outliers <- results$out
> outliers
[1] 9 10
The outlier values for the variable Number are therefore 9 and 10. We still need
to know the identifier for the individuals in question. In order to do this, we
search for individuals whose value corresponds to one of the values contained
Preparing Data 43

within outliers. The operator %in% yields TRUE or FALSE, depending on

whether or not the value is within the set (here outliers). Then, to know
the number of observations which are found within this set, we use which:
> which(kyphosis[,"Number"]%in%outliers)
[1] 43 53
Individuals 43 and 53 are therefore the outliers.
Another way of identifying these outliers is to use the identify function.
By clicking on a point on a graph, the function yields the row number cor-
responding to this individual. The identify function takes the abscissa and
ordinate values from the scatterplot as the arguments:
> identify(rep(1,length(kyphosis[,"Number"])),kyphosis[,"Number"])
Here we can see that only one variable is represented on the ordinate. The
abscissa is simply 1 as we only have one variable for all the observations. The
number of observations is here length(kyphosis[,"Number"]).
We can therefore identify the outliers with a single click. When you have
finished, simply click elsewhere in the graph window with the right button of
the mouse, or type Esc in the command window.

2.5 Concatenating Data Tables

Let us imagine that we have two data tables that we want to merge. This
can be done in one of two ways: putting the tables one on top of the other
(juxtaposition of rows, see Figure 2.2) or putting them one next to the other
(juxtaposition of columns, see Figure 2.3).

Y
X Y Z

Figure 2.3
Figure 2.2 Column concatenation: cbind(X,Y).
Row concatenation: rbind(X,Y).

In order to concatenate by row for the tables X and Y, we use

> Z <- rbind(X,Y)
44 R for Statistics

X and Y must therefore have the same number of columns. However, the logic
is a little different, depending on whether we concatenate two matrices or two
data-frames. This question is illustrated in the example below.
Let us first construct a matrix with its row identifiers and variable names
X1 and X2:
> X <- matrix(c(1,2,3,4),2,2)
> rownames(X) <- paste("row",1:2,sep="")
> colnames(X) <- paste("X",1:2,sep="")
> X
X1 X2
row1 1 3
row2 2 4

And the same for the Y matrix:

> Y <- matrix(11:16,3,2)
> colnames(Y) <- paste("Y",1:2,sep="")
> Y
Y1 Y2
[1,] 11 14
[2,] 12 15
[3,] 13 16
We concatenate X and Y in one single matrix:
> Z <- rbind(X,Y)
> Z
X1 X2
row1 1 3
row2 2 4
11 14
12 15
13 16
In this example, R concatenates the matrices conserving only the variable
names of the first matrix.
If we use two data-frames with identical structures, an error message ap-
pears when the variable names are different:

> Z <- rbind(data.frame(X),data.frame(Y))

Error in match.names(clabs, names(xi)) :
names do not match previous names
We must therefore change the variable names for one of the data-frames:

> Xd <- data.frame(X)

> Yd <- data.frame(Y)
Preparing Data 45

> colnames(Yd) <- c("X2","X1")

> rbind(Xd,Yd)
X1 X2
row1 1 3
row2 2 4
3 14 11
4 15 12
5 16 13

We can see (on the third line) that the order of the variables is inconsistent.
In this case, the variables are re-classified prior to concatenation.
Concatenation by column is done in the same way as for the matrices or
data-frames. The cbind function does not check the row names. The names
from the first table are retained:

> X <- matrix(c(1,2,3,4),2,2)

> rownames(X) <- paste("row",1:2,sep="")
> Y <- matrix(11:16,2,3)
> cbind(data.frame(X),Y)
X1 X2 X1 X2 X3
row1 1 3 11 13 15
row2 2 4 12 14 16
This type of concatenation that is to say, by row or by column is not
the only possibility. Two tables can be merged according to a key, merge (see
Figure 2.4).

Key Key Key

X Y

X Y Z

Figure 2.4
Merging by key: merge(X,Y,by="key").

Let us construct two data tables, one with more rows than the other. The
first table merges (column concatenation) a continuous variable (age) and two
qualitative variables in a data-frame. The cbind.data.frame function can be
used to specify that the result must be a data-frame:
46 R for Statistics

> age <- c(7,38,32)

> surname <- c("arnaud","nicolas","laurent")
> town <- factor(c("rennes","rennes","marseille"))
> indiv <- cbind.data.frame(age,surname,town)
> indiv
age surname town
1 7 arnaud rennes
2 38 nicolas rennes
3 32 laurent marseille
A second table groups together the characteristics of the towns:
> population <- c(200,500,800)
> towns <- cbind.data.frame(c("rennes","lyon","marseille"),
population)
> names(towns) <- c("town","pop")
> towns
town pop
1 rennes 200
2 lyon 500
3 marseille 800
If we want to merge these two tables into one, where the characteristics of the
towns will be repeated in each row, we merge by the key town:
> merge(indiv,towns,by="town")
town age surname pop
1 marseille 32 laurent 800
2 rennes 7 arnaud 200
3 rennes 38 nicolas 200
The key town is therefore the first column of the dataset.

2.6 Cross-Tabulation
When two qualitative variables are analysed, we can present the data in two
ways: a classical table where an individual (a row) is described by the variables
(two variables), or a cross-tabulation (also known as a contingency table),
which gives the sample size for each confrontation of categories. Of course,
this can be generalised to more than two qualitative variables.
In order to get a clear idea of both types of table, let us first construct
two qualitative variables (wool and tension) collected from ten individuals.
The variable wool corresponds to the following three types of wool: Angora,
Merinos and Texel. The variable Tension indicates the value Low or High for
the wools tension resistance.
Preparing Data 47

> tension <- factor(c(rep("Low",5),rep("High",5)))

> tension
[1] Low Low Low Low Low High High High High High
Levels: High Low
> wool <- factor(c(rep("Mer",3),rep("Ang",3),rep("Tex",4)))
> wool
[1] Mer Mer Mer Ang Ang Ang Tex Tex Tex Tex
Levels: Ang Mer Tex

Let us merge these two variables in the table don:

> tab <- cbind.data.frame(tension,wool)
> tab
tension wool
1 Low Mer
2 Low Mer
3 Low Mer
4 Low Ang
5 Low Ang
6 High Ang
7 High Tex
8 High Tex
9 High Tex
10 High Tex
The cross-tabulation is therefore obtained directly on the objects using

> table(tab$tension,tab$wool)
Ang Mer Tex
High 1 0 4
Low 2 3 0
Another method is to use the column names from the data-frame and to
conduct the confrontation using a formula:
> crosstab <- xtabs(~tension+wool,data=tab)
> crosstab
wool
tension Ang Mer Tex
High 1 0 4
Low 2 3 0
Many functions in R are based on the hypothesis of an individuals
variables table, but the opposite operation must also be made possible for
situations in which the data are contained within a cross-tabulation. However,
an individuals variables table is not constructed immediately:
> tabframe <- as.data.frame(crosstab)
48 R for Statistics

> tabframe
tension wool Freq
1 High Ang 1
2 Low Ang 2
3 High Mer 0
4 Low Mer 3
5 High Tex 4
6 Low Tex 0

We obtain the frequencies for each combination rather than a row for each
individual. To reconstruct the initial table, see Exercise 2.8.

2.7 Exercises
Exercise 2.1 (Robust Reading of Data)
In some cases, reading data fails and it can be difficult to understand why. In
such cases, here is a more robust procedure which can be used to identify the
problem.
1. Read the file mydata.csv in a character vector myvector using the scan
function (in a complex reading, this step can also be used to check the column
separator).
2. Change the decimal separator from "," to "." (function gsub) in myvector.
It must be noted that this step can be conducted more simply using a spread-
sheet or word processor.
3. Construct a character matrix mymatrix of four rows and five columns which
contains the data and the names of the individuals and variables.

4. Gather the names of individuals and variables in an object namerow and

namecol. Allocate the remaining data within a three-row and four-column
matrix (mymatrix). For a complex or delicate inputting, display the data
column by column in order to make sure that the columns are not mixed
up.
5. Convert the matrix mymatrix to a data-frame mydata and check that the
variable type is correct. If not, convert the factors into numerics. See Sec-
tion 2.3.1, p. 35.

Exercise 2.2 (Reading Data from File)

Read the files test1.csv, test1.prn, test2.csv and test3.csv.

Exercise 2.3 (Reading Data from File with Date Format)

1. Read the file test4.csv (Hint: Use the argument skip to delete the first 2
Preparing Data 49

lines). The data presents a sample of friends with the age, gender (coded 0
for male and 1 for female) and the first date of skiing (year-month-day).

2. Read the file again with the option colClasses to specify that the gender
should be considered as a factor and the first time skiing as a date.

Exercise 2.4 (Reading Data from File and Merging)

1. Read the files state1.csv, state2.csv and state3.csv.
2. Merge the three data-frames into one using the common columns to repeat
the rows adequately.

Exercise 2.5 (Merging and Selection)

1. Read the files fusion1.xls and fusion2.xls.
2. Retain only the variables yhat1, yhat3, Rhamnos, and Arabinos. With
these four variables, create one single data-frame.

3. Create the variable yres1, which is the difference between yhat1 and
Rhamnos, and create the variable yres2, which is the difference between yhat3
and Arabinos. Add these variables to the data-frame.
Exercise 2.6 (Ventilation (allocation at random))
In the presence of the following qualitative variable,
> Xfac <- factor(c(rep("A",60),rep("B",20),rep("C",17),
rep("D",3)))

1. Calculate the frequencies of each category (or factor level).

2. Display the names of the categories with sample sizes smaller than 5% of
the total sample size.
3. Calculate the frequencies of each category without the category or cate-
gories of the previous question. The result will be placed in the freq vector.
4. Select the individuals which take the category or categories of the Exer-
cise 2.6.2. Allocate them a value selected from the remaining categories, ac-
cording to a draw for which the probabilities are calculated in Exercise 2.6.3
(use the sample function). This process is known as ventilation2 .

Exercise 2.7 (Ventilation by Ordinal Factors)

In the presence of the following qualitative variable with ordinal categories:
2 Have you ever seen anything like that? Totally calm, singing away and then bang!

Right in the face! This man is crazy! And I treat madmen every day; Ill write him a
prescription, and a strong one! Ill show him whos boss. Well find bits of him all around
Paris like pieces of a jigsaw puzzle... When it goes too far, I dont correct people any more,
I destroy them... Spread them out, ventilate them.... Bernard Blier in Crooks in Clover,
by Georges Lautner (1963), script by Michel Audiard.
50 R for Statistics

> Xfac <- factor(c(rep("0-10",1),rep("11-20",3),rep("21-30",5),

rep("31-40",20),rep("41-50",2),rep("51-60",2),
rep("61-70",1),rep("71-80",31),rep("+ 80",20)))
1. Calculate the frequencies of each category (or factor level).
2. Display the names of the categories with sample sizes smaller than 5% of
the total sample size.
3. Start with the weakest (according to the order of the categories) of the
categories in Exercise 2.7.2. Merge this category with the category immedi-
ately above. If the sample size of these two merged categories is greater
than 5% of the total, go on to the next category of Exercise 2.7.2 (if there
is one). If not, merge again with the category immediately above (out of
all the categories). Proceed in the same way until all the categories from
Exercise 2.7.2 have been used.
Exercise 2.8 (Cross-Tabulation Data Table)
1. Create the contingency table confronting the qualitative variables types of
wool (wool) and tension (tension):

Ang Mer Tex

High 1 0 4
Low 2 3 0

2. From this table, create a character matrix tabmat with three columns and
as many rows as confrontations of categories (or cells in the contingency table).
This character matrix will be filled row by row with tension (rows from the
previous table), wool type (columns from the previous table), and the sample
size for the confrontation of categories. In order to do so, we use the functions
matrix and rep.
3. Convert the character matrix from the previous question to a data-frame;
name it tabframe, for example, and check the variable types. Using this data-
frame, and assign the total number of individuals to n (sum) and the number
of qualitative variables to nbfac (ncol).
4. Create a counter iter set at 1 and a character matrix tabcomplete, for
example the character "", the same size as the final dataset.
5. Make a loop around the number of rows of tabmat (see Section 4.1.2,
p. 87). Each row i corresponds to a confrontation of categories (or strata). If
the number of individuals with this confrontation of categories i is not null,
then repeat in as many rows as necessary in tabcomplete, the confrontation
of category i (by distributing the categories in the corresponding columns).
It is possible to use a loop, the iter counter, the tabmat matrix and the
tabframe data-frame.
The result tabcomplete will be the same as the initial table.
3
R Graphics

Graphs are often the starting point for statistical analysis. One of the main
advantages of R is how easy it is for the user to create many different kinds of
graphs. We begin this chapter by studying conventional graphs, followed by
an examination of some more complex representations. This final part, using
the lattice package, can be omitted if using R for the first time.

3.1 Conventional Graphical Functions

To begin with, it may be interesting to examine a few example of graphical
representations which can be constructed with R. We use the demo function:
> demo(graphics)
It must be noted that there are other demonstrations available, a list of which
can be found using demo().
We give examples of representations for quantitative and qualitative vari-
ables. We use the data file ozone.txt, imported using
> ozone <- read.table("ozone.txt",header=T)

We obtain meteorological variables and an ozone pollution variable collected

in Rennes (France) during the summer of 2001. The variables available are
maxO3 (maximum daily ozone), T12 (temperature at midday), wind (direc-
tion), rain and Wx12 (projection of the wind speed vector on the east-west
axis at midday):

> ozone <- ozone[,c("T12","maxO3","wind","rain","Wx12")]

> summary(ozone)
T12 maxO3 wind rain
Min. :14.00 Min. : 42.00 East :10 Dry :69
1st Qu.:18.60 1st Qu.: 70.75 North:31 Rainy:43
Median :20.55 Median : 81.50 South:21
Mean :21.53 Mean : 90.30 West :50
3rd Qu.:23.55 3rd Qu.:106.00
Max. :33.50 Max. :166.00
52 R for Statistics

Wx12
Min. :-7.878
1st Qu.:-3.565
Median :-1.879
Mean :-1.611
3rd Qu.: 0.000
Max. : 6.578
The variables T12, maxO3 and Wx12 are continuous quantitative variables (nu-
merics), whereas wind and rain are factors.

3.1.1 The Plot Function

The plot function is a generic function (see Appendix A.1, p. 257) of R used
to represent all kinds of data. Classical use of the plot function consists of
representing a scatterplot for a variable y according to another variable x.
For example, to represent the graph of the function x 7 sin(2x) on [0, 1], at
regular steps we use the following commands:
> gridx <- seq(0,1,length=50)
> fx <- sin(2*pi*gridx)
> plot(x=gridx,y=fx)
These commands open a graph window (known as a device).
1.0
0.5
0.0
fx

0.5
1.0

0.0 0.2 0.4 0.6 0.8 1.0

gridx

Figure 3.1
Scatterplot for the function x 7 sin(2x).

In statistics it is more common to use the plot function with formulae of

the type y~x. This is the format that we will use. For example, the graph in
Figure 3.1 can also be obtained using
> plot(fx~gridx)
R Graphics 53

We continue to use the ozone example and represent different pairs of

variables. Let us start by representing two quantitative variables: maximum
ozone (maxO3) according to temperature (T12) (Figure 3.2):
> plot(ozone[,"max03"]~ozone[,"T12"])
160
140
ozone[, "maxO3"]

120
100
80
60
40

15 20 25 30

ozone[, "T12"]

Figure 3.2
Scatterplot (T12, max03).

As the two variables are contained and named within the same table, a simpler
syntax can be used, which automatically inserts the variables as labels for the
axes (Figure 3.3):
> plot(maxO3~T12,data=ozone)
160
140
120
maxO3

100
80
60
40

15 20 25 30

T12

Figure 3.3
Scatterplot (T12, maxO3) with explicit axis labels.
54 R for Statistics

This graphical representation could also have been obtained using the com-
mand:

> plot(ozone[,"max03"]~ozone[,"T12"],xlab="T12",ylab="maxO3")

Similarly, to represent a quantitative variable (maxO3) according to a qual-

itative variable (wind), we write
> plot(maxO3~wind,data=ozone,xlab="Wind directions",
ylab="Maximum Ozone Concentration")
160
Maximum Ozone Concentration

140
120
100
80
60
40

East North South West

Wind directions

Figure 3.4
Boxplots of max03 according to wind.

In this case, the plot function yields a boxplot for each category of the variable
wind (Figure 3.4). In this graph we can see if the variable wind has an
effect on the ozone. Here, the easterly wind seems to be associated with high
concentrations of ozone. This graph can also be obtained using the boxplot
function:
> boxplot(maxO3~wind,data=ozone)

Given the nature of the variables, the plot function chose to use the boxplot
function as it considers it the most relevant.
We can also represent two qualitative variables using a bar chart. For
example,
> plot(rain~wind,data=ozone)
can be used to obtain, for each category of the explanatory factor (here wind),
the relative frequencies for each category of the response factor (here rain).
The width of the bar is proportional to the frequency of the category of the
R Graphics 55

explanatory factor (here wind). This graph can also be obtained using the
spineplot function.

1.0
Rainy

0.8
0.6
rain

Dry

0.4
0.2
0.0
East North South West

wind

Figure 3.5
Bar chart for rain according to wind.

In Figure 3.5, we find that the weather type rain is proportionally greater
when the wind blows from the west (which indeed seems to be the case in
Rennes). Furthermore, as this is the widest bar, we can also conclude that
this is the most common wind direction in Rennes.
Finally, it is possible to represent a qualitative variable according to a
quantitative variable, here wind according to T12:
> plot(wind~T12,data=ozone)
1.0
West

0.8
South

0.6
wind

North

0.4
0.2
East

0.0

14 18 20 22 24 28 32

T12

Figure 3.6
Bar chart for wind according to T12.
56 R for Statistics

We here obtain a bar chart (Figure 3.6). On the abscissa, the quantitative
variable is divided into classes in the same default manner as a histogram
(hist). In each of these classes, the relative frequencies of each category of
wind are calculated, giving the height of each bar. The colour corresponds to
the category. This graph can also be obtained automatically using
> spineplot(wind~T12,data=ozone)
This bar chart is difficult to interpret. It is possible to return to the first
formulation of the plot function to represent the scatterplot (T12,wind); see
Figure 3.7:
> plot(ozone[,"T12"],ozone[,"wind"],yaxt="n",
xlab="T12",ylab="wind",pch="+")
> axis(side=2,at=1:4,labels=levels(ozone[,"wind"]))
It must here be noted that the graph must be reworked: The argument
yaxt="n" deletes the scale from the y axis, which here takes the different
categories of the qualitative variable converted to numerics by default: 1 to
4. The axis function is therefore used to rename the scale.
West
South
wind

North
East

15 20 25 30

T12

Figure 3.7
Graph of the qualitative variable wind according to the quantitative variable
T12.

This type of graph is interesting when there are only a few measurements
for each category. If not, the data should be represented using a boxplot (see
Figure 3.4).
Even if, intrinsically, the plot function expects the two input arguments
x and y for the abscissa and ordinate, respectively, it is possible to specify
only one argument. For example, for one quantitative variable, R draws this
variable sequentially with the observation number on the abscissa axis (see
Figure 3.8).
R Graphics 57

> plot(ozone[,"maxO3"],xlab="num.",ylab="maxO3",cex=.5,pch=16)

160

140

120

maxO3

100

80

60

40

0 20 40 60 80 100

num.

Figure 3.8
Scatterplot for ozone peaks (maxO3) according to index.

This command yields an image where, by default, the abscissa axis carries the
ordinal continuation 1, , n known as index (where n is the number of ob-
servations). The size of the symbol can be controlled using the argument cex
which manages the increase (or decrease) factor for size (by default cex=1).
The argument pch is used to specify the shape of the points. This argument
accepts numerical values between 0 and 25 (see Figure 3.9) or a character
directly rewritten on the screen.

21 22 23 24 25

16 17 18 19 20

11 12 13 14 15

6 7 8 9 10

1 2 3 4 5

+ "+" x "x" "." # "#" 0

Figure 3.9
Symbol obtained for the value of the argument pch.

It is also possible to modify the type of line drawn using the argument
type: "p" to only draw the points (default option), "l" (line) to link them,
"b" or "o" to do both. The argument type can also be used to draw vertical
58 R for Statistics

bars ("h", high-density) or steps after the points ("s", step) or before ("S",
step). A graphical illustration of these arguments is given on Figure 3.10.

p l b h s S

Figure 3.10
Argument type for the values "p", "l", "b", "h", "s", "S".

The evolution of ozone maxima can be obtained using the argument type="l"
(Figure 3.11):
> plot(ozone[,"maxO3"],xlab="num.",ylab="maxO3",type="l")
160
140
120
maxO3

100
80
60
40

0 20 40 60 80 100

num.

Figure 3.11
Evolution of ozone peaks (maxO3) during the summer of 2001.

This is a very classical graph for representing temporal data.

In conclusion, the plot function adapts its behaviour to each type of data.

3.1.2 Representing a Distribution

To represent the distribution of a continuous variable (e.g. the vector of nu-
merics ozone[,"maxO3"]), there are a number of pre-programmed classical
R Graphics 59

solutions. The histogram for the variable maxO3 (Figure 3.12) estimates the
density if we specify the argument prob=TRUE of the hist function:

> hist(ozone[,"maxO3"],main="Histogram",prob=TRUE,xlab="Ozone")

Histogram

0.015
Density

0.010
0.005
0.000

40 60 80 100 120 140 160 180

Ozone

Figure 3.12
Histogram for ozone.

A kernel density estimate of the variable maxO3 can be obtained using the
density function. The graphical representation (Figure 3.13) is obtained as
follows:

> plot(density(ozone[,"maxO3"]),main="Kernel Density Estimate",

xlab="Ozone")

Kernel Density Estimate

0.015
0.010
Density

0.005
0.000

50 100 150 200

Ozone

Figure 3.13
Kernel density estimate for ozone.
60 R for Statistics

To represent the distribution of a qualitative variable, we use a bar chart,

obtained using the barplot function. We must count the number of occurrences
at each level using the table function. For the factor wind,
> barplot(table(ozone[,"wind"]))
R yields a bar chart (Figure 3.14) with, on the x-axis the values of the factor,
and on the y-axis the number of observations for each category.
50
40
30
20
10
0

East North South West

Figure 3.14
Bar chart for a qualitative variable.

It is also possible to use the plot function directly on the vector of the
categories:
> plot(ozone[,"wind"])
Another method, but which is not recommended, is circular representation,
or a pie chart, which can be obtained using the pie function.

3.1.3 Adding to Graphs

Once the graph is drawn, it is possible to add additional information: extra
lines (lines), points (points), texts (text), symbols (circles, squares, stars, etc.;
symbols), arrows (arrows), segments (segments) or polygons (polygon). We
will only present the addition of texts, lines and points.
We again refer to the graphical representation of ozone maxima maxO3
according to temperature T12 (Figure 3.15). To make the date appear we
need to add text (text) to the scatterplot obtained using plot.
Nonetheless, in order to avoid overlapping the symbols (the small circle
pch=1) and the names, we do not draw any symbols (type="n"). We simply
write the month and the day (i.e. the last four characters of the row names)
with a reduced font size (cex=.75):
R Graphics 61

> plot(maxO3~T12,data=ozone,type="n")
> text(ozone[,"T12"],ozone[,"maxO3"],substr(rownames(ozone),5,8),
cex=.75)

160 0824
0727 0729 0825
0730
0726
0725
0728
0620 0623 0625
140

0703
0823
0801
120

06220624
0802 0822 0814
0707 0604 0821
0828 0827
maxO3

0815 0812
0612 0626 0702
100

0829 0613
0611 0826
0928
0922
0921 0811
0706
0705
0923
0603 0605
07190614
06180720 0709
0601
0925
0617 0602 09290716 0724
0627 0731
0621
0723 0629
80

0607 0610 07040606

0901 0912
0803
0924 0830 0810
0710
0927
08130820 0902
09130615 0711
0708
0807
0722
0718 0916
0914
08090920
0616
0915 0904
0930 0701
0919 0715 0907
0714
0905 0819 0630
0721 0806
0908
0717
0804
60

0917 0906 083108080628

0712 0805
0903
0713
0918
40

15 20 25 30

T12

Figure 3.15
Text positioned via the points coordinates.

The subsequent graph (Figure 3.15) does not show the exact position of
the name. It is therefore recommended to plot crosses on the graph and then
to add text above these crosses. We thus draw these crosses on the coordinates
(plot function), and then add the text on top of that (pos=3) offset by 0.3
(offset):
> plot(maxO3~T12,data=ozone,type="p",pch=3,cex=.75)
> text(ozone[,"T12"],ozone[,"maxO3"],
substr(rownames(ozone),5,8),cex=.75,pos=3,offset=.3)

If in addition we wish to add a coloured symbol for the dates corresponding

to rainy days, we will select their coordinates and then draw a grey dot at
their coordinates (Figure 3.16):
> selection <- ozone[,"rain"]=="Rainy"
> points(ozone[selection,"T12"],ozone[selection,"maxO3"],
pch=21,bg="grey70",cex=.75)
62 R for Statistics
0824
0727 0729 0825

160
0730
0726
0725
0728
0620 0623 0625
0703

140
0823
0801
06220624

120
0802 0822 0814
0707 0604 0821
0828 0827
0815 0812
0612 0626

maxO3

0702
0829 0613 0611 0826
0928
100
0922
0921 0811
0706
0705 0605

0603
0923 0719 0614 0709
0618
0601 0720

0925
06170602 0929 0716

0724
0627 0731
0607 0610 0704 0606 0621
0723 0629
0803 0901

08100927
0710 0912
80

0924 0830
0913 0813
0820 0902

0920
0916
0914 0615
0904 0708
07110807
0722
0718 0809 0616
0915 0930

0907 0701
0919 0714
0715 0819 0630
0905
0908
0721 0806
0804
0717
0917
0906 0831 0808
0628
60

0712 0805
0903


0713
0918

40

15 20 25 30

T12

Figure 3.16
Scatterplot with the rainy days in grey.

From this graph we can see that it does not rain when the temperature exceeds
27 degrees. This limit can be marked using the abline function and adding
the command (Figure 3.17):
> abline(v=27,lty=2)

0824
0727 0729 0825
160

0730
0726
0725
0728
0620 0623 0625
0703
140

0823
0801
06220624
120

0802 0822 0814

0707 0604 0821
0828 0827
0815 0812
0612 0626

maxO3

0702
0829 0613 0611 0826
0928
100

0922
0921 0811
0706
0705 0605

0603
0923 0719

060106180614
0720 0709
0925
0617 0716
0929
0602 0606

0724
0627 0731
0607 0610 07040810 0621
0723 0629
0803 0901

0927
0710 0912
80

0924 0830
0913 0813
0820 0902

0916
0914
08090920
0616 0615 0711
0904
0930 0708
0807
0722
0701
0718 0915
0907
0714
0715 0819

0630
0919 0905
0908
0721 0806
0717
0804
0917
0906 0831 0808
0628
60

0712 0805
0903


0713
0918

40

15 20 25 30

T12

Figure 3.17
Use of the abline function.

We can also add a horizontal line using abline(h= ) or any type of line (by
specifying the intercept and the slope using abline(c(intercept,slope)).
R Graphics 63

The lty (line type) setting is used to control the type of line (1: continuous
line, 2: dashed, 3: dotted, etc.).
As its name indicates, the abline function adds a straight line onto the
existing graph. If we want to add two broken lines to compare, for example,
the evolution of maximum ozone levels over two different weeks, we use the
lines function. Let us compare the evolution of ozone levels during the first
two weeks:

> plot(ozone[1:7,"maxO3"],type="l")
> lines(ozone[8:14,"maxO3"],col="grey50") # add lines (2nd week)
110
ozone[1:7, "maxO3"]

100
95
90
85
80

1 2 3 4 5 6 7

Index

Figure 3.18
Adding a grey line using lines.

The graph (Figure 3.18) does not show the sixth and seventh observations for
this second week. Indeed, if we want to draw a line on the graph, we need to
know the minimum and maximum for each axis in order to be able to prepare
them and to prepare the scales. Here, the minimum for the second line is
less than that of the first. The graph is scaled using the plot function. When
called upon, this function is only aware of the information concerning the
plot instruction. Scaling is not automatic following a lines (or points) order.
It is therefore important, right from the plot order, to specify the minimum
and maximum for both weeks. Simply give the ylim argument the results
of the range function. The latter yields the minimum and maximum. This
argument tells the plot function between which ordinates the graph is situated
(Figure 3.19).
> rangey <- range(ozone[1:7,"maxO3"],ozone[8:14,"maxO3"])
> plot(ozone[1:7,"maxO3"],type="l",ylim=rangey,lty=1)
> lines(ozone[8:14,"maxO3"],col="grey50",lty=1)
64 R for Statistics

110
ozone[1:7, "maxO3"]

100
90
80
70

1 2 3 4 5 6 7

Index

Figure 3.19
Adding a grey line using lines and scaling the ordinates using the ylim argu-
ment.

Of course, if the abscissas are the same, you do not need to specify the mini-
mum and maximum on the abscissa.

3.1.4 Graphs with Multiple Dimensions

The most classical functions for representing graphs in 3D on a grid of points
are the functions persp (3D with perspective effect), contour (contour lines)
and image (contour lines in colour).
The classic Mexican hat, given in the help section
p of persp,

p is the graph
of the function (x, y) 7 z = f (x, y) = 10 sin x2 + y 2 / x2 + y 2 . To
implement this function, we have chosen a regular square grid with thirty
different values between 10 and 10 (on x and y):

> f <- function(x,y) 10 * sin(sqrt(x^2+y^2))/sqrt(x^2+y^2)

> x <- seq(-10, 10, length= 30)
> y <- x
We evaluate the f function at each point of the grid: for each pair (x[i],
y[j]), we calculate f(x[i],y[j]). In order to avoid creating a (double) loop,
we use the outer function which enables this type of evaluation:
> z <- outer(x, y, f)
We draw this function in 3D (Figure 3.20)

> persp(x, y, z, theta = 30, phi = 30, expand = 0.5)

R Graphics 65

y
x

Figure 3.20
Example of use of the persp function.

We could obtain level sets using contour(x, y, z) or image(x, y, z).

We can also use the rgl package to construct the response surface for the
previous graph:
> library(rgl)
> rgl.surface(x, y, z)
This package is recommended for all problems with 3D visualisation. For
example, it can be used to rotate graphs with the mouse and to use symbols
with lighting effects.
If we want to draw a scatterplot representing n individuals characterised
by three quantitative variables, it is not possible to use conventional graphs.
We must therefore use a package such as rgl or lattice. We start with the lattice
package which is one of the default packages provided with R.
We again use the ozone example, representing the individuals for the three
quantitative variables maxO3, T12 and Wx12. As we are looking to explain the
variable maxO3, it is only natural to represent it on the third axis (Figure 3.21):

> library(lattice)
> cloud(maxO3~T12+Wx12,type=c("p","h"),data=ozone)
Here, the lattice package enables the user to apply two arguments for type,
which is not possible in conventional graphs.
66 R for Statistics

maxO3

Wx12 T12

Figure 3.21
3D graph characterised by measurements of the quantitative variables maxO3,
T12 and Wx12.

If we want to use the rgl package, simply type:

> plot3d(ozone[,"T12"],ozone[,"Wx12"],ozone[,"maxO3"],radius=2,
xlab="T12",ylab="Wx12",zlab="maxO3",type="s")
It must be noted that the type of points type="s", specific to rgl, enables
spheres to be drawn with a fixed radius of 2, thanks to the argument radius.

3.1.5 Exporting Graphs

Graphs are used as a visual support and are often needed for use in presenta-
tions or other types of external documents. It is therefore important to be able
to export them in a format which is compatible with other software. There
are many different formats available, such as jpeg, png, pdf and postscript
(ps), or vectorial graphical formats such as emf (in Windows), svg or xfig.
A device is defined as either the graph window or a file into which the
graph is exported. The device changes with the export format (see Table
3.1).
In Windows, it is easiest to save the graph from the graph window. In order
to do so, go to File Save as and choose the desired format, preferably a
vector format (metafile emf). It is also possible to copy the graph and to
paste it into a document by right-clicking on the graph with the mouse and
then choosing Copy as metafile.
Otherwise, exporting a graph always follows the same procedure. For
example, if we want to export in pdf, the name of the device is (pdf) and is
activated by the function of the same name. To save the graph Figure 3.19 in
pdf format, write
R Graphics 67
TABLE 3.1
Non-exhaustive List of Devices
Device Description Graph Window
quartz default Mac OS X device Yes
X11 Unix/Linux default device Yes
jpeg jpeg file No
png png file No
bmp bmp file No
pdf pdf file No
ps ps file No
pictex picTeX macros No
svg svg vector format No
xfig xfig vector format No

> pdf("graphik.pdf")
> rangey <- range(ozone[1:7,"maxO3"],ozone[8:14,"maxO3"])
> plot(ozone[1:7,"maxO3"],type="l",ylim=rangey,lty=1)
> lines(ozone[8:14,"maxO3"],col="grey50",lty=1)
> dev.off()
The dev.off function closes the device and finalises the file graphik.pdf which
will be found in the current working directory (see Section 1.2.1 or 1.2.2
or 1.2.3). It is of course possible to specify the destination pathway using
pdf("mypath/graphik.pdf").
There are a number of different options when exporting graphs (paper size,
choice of fonts, etc.). These options all depend on the export format chosen
and can be consulted in the help section. Not all the devices are necessarily
available: they depend on the platform used, the compilation options and the
packages installed.
If we want to modify a graph (move a legend, text, increase the size of
a given element, etc.), vector graphics software is required. The graph must
therefore be exported (format svg, emf or xfig, for example) and then altered
using the appropriate software. It must be noted that inkscape and oodraw
(from openoffice) are freely available open source software which import graphs
in svg format. In Mac OS X and Unix/Linux, xfig is another alternative to
these two pieces of software.

3.1.6 Multiple Graphs

If we want to feature more than one graph in the same window, after open-
ing the device, simply state the number of graphs required. Traditionally,
we use the par function. The instruction par(mfrow=c(n,p)) organises np
graphs into n rows and p columns. To obtain two graphs on the same row
(Figure 3.22), write:
68 R for Statistics

> par(mfrow=c(1,2))
> plot(1:10,10:1,pch=0)
> plot(rep(1,4),type="l")

1.4
10
8

1.2
rep(1, 4)
6
10:1

1.0
4

0.8
2

2 4 6 8 10 0.6 1.0 1.5 2.0 2.5 3.0 3.5 4.0

1:10 Index

Figure 3.22
Graphs arranged in 1 row 2 columns using par(mfrow=c(1,2)).

To come back to one graph per window, simply kill the graph window or use
the command par(mfrow=c(1,1)).
Sometimes, we need to use graphs of different sizes. In such cases, we
use the layout function: this function divides the graph window into areas
before distributing them. As its argument, it therefore admits a nrow ncol
matrix, which creates nrow ncol number of individual tiles. The values of
the matrix therefore correspond to the numbers of the graphs featured in each
tile. For example, to set out three graphs on two rows, as in Figure 3.23, we
first create the following matrix:
 
1 1
2 3
> mat <- matrix(c(1,1,2,3), nrow=2, ncol=2, byrow = TRUE)

2 3

Figure 3.23
Positioning three graphs on two rows using layout.
R Graphics 69

Then we draw the three graphs (Figure 3.24):

> layout(mat)
> plot(1:10,10:1,pch=0)
> plot(rep(1,4),type="l")
> plot(c(2,3,-1,0),type="b")
10
8
6
10:1

4
2

2 4 6 8 10

1:10
1.4

3
1.2

2
c(2, 3, 1, 0)
rep(1, 4)

1.0

1
0.8

0
0.6

1.0 1.5 2.0 2.5 3.0 3.5 4.0 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Index Index

Figure 3.24
Three graphs on two rows using layout.

It is also possible to specify the height and width of the columns (see help for
layout and Exercise 3.8). It must be noted that the graphs might be (too) far
apart, but this can easily be resolved (see Exercise 3.8).

3.1.7 Multiple Windows

Sometimes, more than one graph window is required. To create a graph win-
dow, simply open the device which corresponds to the operating system: X11
in Unix/Linux or Windows and quartz in Mac OS X. Therefore, if we want
to draw up two graphs in two separate windows, in Unix/Linux or Windows
simply use
> plot(1:10,10:1,pch=0)
> X11()
> plot(rep(1,4),type="l")
70 R for Statistics

The first graph is thus drawn. If no graph window was open prior to the
creation of the first graph, it will be created automatically. A second window
is then opened, thus rendering the first inactive. The second graph is drawn
in this new window. Subsequent graphs will be drawn in the active window.

3.1.8 Improving and Personalising Graphs

In order to improve a graph, it is possible
To use colours:

> plot(x,y,type="p",pch=o,col="red")

These colours can be specified using a number (generally, 1=black, 2=red,

3=green, etc.), a name in English from the list given by the vector colors(),
or an RGB code with transparency (see rgb). The correspondence between
the number of the colour and the colour itself is ensured by the palette.
It is possible to change the background and line colours for a device using
the par function, which manages all the graphical parameters for a given
device:

> par(fg="blue",bg="#f2a1c2")

To modify the palette, ensuring a correspondence between the number and

the colour:

> palette(gray(seq(0,0.9,len=10))) # palette of grey

> plot(1:10,rep(1,10),type="p",pch=25,bg=1:10)
> palette("default")

To add a title:

> plot(x,y,type="p",pch=o,main="My Title")

To control the appearance of the axes or to delete them (and the corre-
sponding legend):

> plot(c(10,11),c(12,11),type="p",axes=FALSE,xlab="",ylab="")

to finally redefine them by hand:

> axis(1,at=c(10,11),label=c("Dim 1","Dim 2"))

To construct orthonormal axes (argument asp=1):

> plot(x,y,type="p",asp=1)
R Graphics 71

To add a legend:

> rangey <- range(ozone[1:7,"maxO3"],ozone[8:14,"maxO3"])

> plot(ozone[1:7,"maxO3"],type="l")
> lines(ozone[8:14,"maxO3"],ylim=rangey,col="grey50"))
> legend("topleft",legend=c("week 1","week 2"),
col=c("black","grey50"),lty=1)

To insert symbols or mathematical formulae (help(plotmath)), for exam-

ple for the legend of an axis:

> plot(1,1,xlab=expression(bar(x)==sum(frac(x[i], n), i==1,n)))

Some parameters can be modified directly in the graph command, and

others are accessible using the par function.

Argument Description
adj Controls the position of the text in relation to the left
edge of the text: 0 to the left, 0.5 centred, 1 to the
right; if two values are given c(x,y), the text will be
horizontally and vertically justified
asp Specifies the ratio between y and x: asp=1 constructs
orthonormal graphs
axes By default TRUE, the axes and the frame are shown
bg Specifies the background colour 1, 2, or a colour chosen
from colors()
bty Controls the frame line, permitted values: o,l, 7,
c, u or ] (with the shape of frame resembling that
of the character); bty="n" deletes the frame
cex Controls the size of the characters and symbols compared
to the default value which is 1
cex.axis Controls the size of the characters on the scales of the
axes
cex.lab Controls the size of the characters for the axis labels
cex.main Controls the size of the characters of the title
cex.sub Controls the size of the characters in the caption
col Specifies the colour of the graph: possible values 1, 2, or
a colour chosen from colors()
col.axis Specifies the colour of the axes
col.main Specifies the colour of the title
font Controls the style of the text (1: normal, 2: italics, 3:
bold, 4: bold italics)
font.axis Controls the style for the scales
font.lab Controls the style for the axis labels
font.main Controls the style of the title
font.sub Controls the style of the caption
72 R for Statistics

Argument Description
las Controls the distribution of annotations on the axes (0,
default value: parallel to the axes, 1: horizontal, 2: per-
pendicular to the axes, 3: vertical)
lty Controls the type of line drawn (1: continuous, 2:
dashed, 3: dotted, 4: alternately dotted and dashed,
5: long dashes, 6: alternately long and short dashes), or
write in words solid, dashed, dotted, dotdash,
longdash, twodash or blank (to write nothing)
lwd Controls the thickness of the lines
main Specifies the title of the graph, for example main="My
Title"
mfrow Vector c(nr,nc) which divides the graph window into
nr rows and nc columns; the graphs are then drawn in a
line
offset Specifies the position of the text with respect to the point
position (text function)
pch Integer (between 0 and 25) which controls the type of
symbol (or potentially any character in speech marks)
pos Specifies the position of the text; permitted values 1, 2,
3 and 4 (text function)
ps Integer controlling the size of the text and symbols in
points
sub Specifies the caption of the graph, for example sub="My
Caption"
tck, tcl Specifies the distances between graduations on the axes
type Specifies the type of graph drawn: permitted values
"n","p","l","b,"h","s","S"
xlim, ylim Specifies the limits of the axes, for example
xlim=c(0,10)
xlab, ylab Specifies the annotations on the axes, for example
xlab="Abscissa"

Below is a list of classical graphical functions.

Function Description
barplot(x) Draws a bar chart for the values of x
boxplot(x) Draws a boxplot for x
contour(x,y,z) Draws contour lines, see also
filled.contour(x,y,z)
coplot(x~y|f1) Draws the bivariate plot of (x, y) for each value
of f1 (or for each small interval for the values of
f1)
filled.contour(x, Draws contour lines, but the areas between the
y,z) contours are coloured, see also image(x,y,z)
R Graphics 73

Function Description
hist(x,prob=TRUE) Draws a histogram for x
image(x,y,z) Draws rectangles at the coordinates x, y coloured
according to z, see also contour(x,y,z)
interaction.plot(f1, Draws the graph for the averages of x according
f2,x,fun=mean) to the values of f1 (on the abscissa axis) and f2
(one broken line per category of f2)
matplot(x,y) Draws the bivariate plot for the first column of x
confronted with the first column of y, the second
column of x with the second column of y, etc.
pairs(x) If x is a matrix or data-frame, draws all the bi-
variate plots between the columns of x
persp(x,y,z) Draws a response surface in 3D, see demo(persp)
pie(x) Draws a pie chart
plot(objet) Draws a graph corresponding to the class of
objet
plot(x,y) Draws y against x
qqnorm(x) Draws the quantiles of x in relation to those ex-
pected from a normal distribution
qqplot(x,y) Draws the quantiles of y in relation to those of x
spineplot(f1,f2) Draws the bar chart corresponding to f1 and f2
stripplot(x) Draws the graph for the values of x on a line
sunflowerplot(x,y) Idem, but the overlapped points are drawn in the
shape of flowers with the number of petals corre-
sponding to the number of points
symbols(x,y,...) For the coordinates x and y, draws different sym-
bols (stars, circles, boxplots, etc.)

3.2 Graphical Functions with lattice

The lattice package is used to represent those data for which the variable
represented on the ordinate axis is based on multiple variables. To do this,
the lattice package relies on low-level graphical functions which are not those
we traditionally use, but instead are those offered in the grid package.
The grid package is more comprehensive and more accurate than older low-
level graphical functions used, for example, in plot. However, they are more
complex and we refer the interested reader to the book by Murrel (2005). Here
we present a brief overview of the possibilities offered by the lattice package.
We use the quakes dataset available in R. The dataset gives the locations
of 1000 seismic events of MB greater than 4.0. The events occurred near
Fiji since 1964. These include the Richter magnitude magn, longitude long,
latitude lat, and the depth depth. The spatial distribution of the earthquakes
74 R for Statistics

provides information about the location of the trench. The most natural way
of representing this is to feature the longitude on the abscissa, the latitude
on the ordinate, and to plot one point for each earthquake. As it is probable
that the location will affect the depth of the earthquake, we have chosen to
represent a symbol with a colour which darkens as the depth increases. To do
this, we divide the depth variable into six classes (see Section 2.3.2, p. 35):
> library(lattice)
> data(quakes)
> breakpts <- quantile(quakes$depth,probs=seq(0,1,length=7))
> class <- cut(quakes$depth,breaks=breakpts,include.lowest=TRUE)
> levels(class) <- c("grey80","grey65","grey50","grey35",
"grey10","black")
> plot(lat ~ long,data=quakes,col=as.character(class),pch=16)
Quakes seem therefore to be distributed according to two planes (see Fig-
ure 3.25). We can represent them according to six depth levels using the
10

15

20

lat

25

30

35

165 170 175 180 185

long

Figure 3.25
Latitude and longitude of earthquakes near Fiji.

functions xyplot and equal.count. The sample sizes for each level are equal
and the intervals overlap: 5% of the data are common to two consecutive
intervals (overlap=0.05).
> Depth <- equal.count(quakes$depth,number = 6,overlap=0.05)
> xyplot(lat ~ long | Depth, data = quakes,pch="+",
xlab = "Longitude",ylab = "Latitude")
The depth interval is given in the strip above each panel of the overall graph
(Figure 3.26).
R Graphics 75
165 170 175 180 185

Depth Depth Depth

10
+++
++ ++++ + +
++++ + +
+ +++++++
+ +++ + +++
++
+
++
15
++++++ ++ + + +++++
+++ + +++++++++++
++++++++
+++ ++ +
+ +
+++ + ++ ++
+ +++ +
++++ ++++++ +
+ +++
++
++
++++
++ +
+ + +++ ++ +++
+ +++++ 20
++++
+++ + +++
++++
++++ + ++
+ ++
+ ++
++++ ++ + +++
+++
+++++ ++++ +
++++
++
+
+
+
++
+
+ +++
+++
+++ + +
+++
+ + +
+++
+ + ++ ++
+++++
+++
+ ++
+++++
+ +++
++ +
+++
+++ +
+
25
++
+
+++++ 30
++
+ +
35
+
Latitude

Depth Depth Depth

10 + ++ ++++ ++++
+++
++ ++++
++ +++
+++++ +++
+++
+
+
++
+ + +++ +
+++
15 ++++
++++
++
+++ +++
++
++
+
+++++ +
+
++++++
+ + ++++++ ++++++
+
+++++ +++++
++ +
++++++++
+
+ ++ + ++ +++
+ +++ +++ ++ + + ++ +++ ++
+
++
++ +++ ++++
+++
+ +++ +++ ++ + ++++ +++
+++++
20 +++ +++ + + ++++++++
++
+
++
++
+++
+++ ++ + ++++
++ + + +++
+ +++++
+++++
++++ + +++ + ++
++++ + +++ +++
+++
++
+
+
+
+
+ ++
+++ +
+++ +++ +
+++
25 ++ +++ + +
++
++ + +++
+ +++
+++
+++ ++
++ +++
+++ ++
+
+++
+++++
+ ++++
++ + +
30 +++ ++ +++ +++
++
+
++++
+ + ++
+ +++
+ ++ ++ + + +
35 ++ + +
+ ++
+ ++ ++

165 170 175 180 185 165 170 175 180 185

Longitude

Figure 3.26
Latitude and longitude of earthquakes depending on depth.

3.2.1 Characteristics of a Lattice Graph

Below are some of the characteristics of graphs from the lattice package:
In general, each graph is composed of a series of panels (elementary graphs)
arranged into a rectangular table.
Panels are indexed from left to right and from bottom to top.

The abscissa and ordinate axes are identical from one panel to the next.
A panel is a graph of one variable (ordinate) in relation to another (ab-
scissa) conditioned by one (or multiple) other variable(s). The conditional
value appears in the strip for each panel.

Lattice type graphs are similar to normal graphs but they can be used
to divide groups of variables. Table 3.2 illustrates the correspondence between
conventional and lattice graphs.
Conditioning relies on a qualitative variable or on the way a continuous
variable is partitioned. In the latter case, the resulting object is referred to
as shingles. This object is created using the equal.count function and can
be represented graphically (Figure 3.27) using the generic plot function:
> plot(Depth)
76 R for Statistics
TABLE 3.2
Correspondence between Conventional and Lattice Graphs
Lattice Description Conventional
barchart Bar chart barplot
bwplot Boxplot boxplot
densityplot() Estimation of density plot(density())
histogram Histogram hist
qqmath Normal QQ plot qqnorm
qq Quantiles - quantiles (2 samples) qqplot
xyplot Scatterplot plot
levelplot Greyscale (3D) image
contourplot Contour lines (3D) contour
cloud Scatterplot (3D) none
wireframe Perspective (3D) persp
splom Matrix of scatterplot pairs
parallel Diagonal none

6
5
Panel

4
3
2
1

200 400 600

Range

Figure 3.27
Shingles of earthquake depth.

3.2.2 Formulae and Groups

Lattice graphs are based on a formula which can be used to specify the
abscissa and ordinate axes and the conditioning. For example, if within the
same data-frame, we have the variables yaxis, xaxis, factor, factor1 and
factor2, and if we have created the object shingles, then the following
formulae

yaxis~xaxis|factor
yaxis~xaxis|factor1*factor2
yaxis~xaxis|shingles*factor
can be used to plot the points for which the coordinates are contained within
xaxis and yaxis. This will condition, according to a) the categories of
R Graphics 77

factor, b) the confrontation of the categories of factor1, factor2 and c)

those of factor and the intervals defined in shingles.
Below we present two examples from the states.x77 table available in R.
We take the table states.x77 and add the names of the different states and
regions of the United States:
> data(state)
> states <- data.frame(state.x77,state.name=rownames(state.x77),
state.region = state.region)
We examine the representation of crimes according to population in the four
regions of the United States (simple conditioning, Figure 3.28), and then add
the variable income (multiple conditioning). To conduct this second condi-
tioning, income must be broken down; we here choose three classes:

> income <- equal.count(states[,"Income"],number=3)

We draw two graphs:
> xyplot(Murder ~ Population | state.region, data = states)

0 5000 10000 15000 20000

North Central West

15

10

5
Murder

Northeast South
15

10

0 5000 10000 15000 20000

Population

Figure 3.28
Crime rates and population size by region.

> xyplot(Murder ~ Population | state.region*income,

data = states)
78 R for Statistics
0 5000 15000 0 5000 15000

income income income income

Northeast South North Central West
15

10

income income income income

Northeast South North Central West
15
Murder

10

income income income income

Northeast South North Central West
15

10

0 5000 15000 0 5000 15000

Population

Figure 3.29
Crime rates and population size by region and class of income.

We thus obtain 4 3 panels in which we can visualise the relationship

between crime rate and population for each class of income region (Fig-
ure 3.29). This graph features numerous panels. It can often be easier to
read if one of the variables is represented using a specific colour or a symbol.
For each class of income we represent the crime rate with different symbols
/ colours for each region. On the computer screen, the different colours are
visible; however, if the graphs are exported in black and white, it is better to
use symbols. Within each panel (each class of income), there are four groups
(the regions).

> xyplot(Murder~Population|income, groups=state.region,

data=states,auto.key=TRUE,type="p")
A legend can be added using the argument auto.key, which is used to control
the appearance of the legends: visualisation of lines (instead of symbols),
position (top, right, bottom, or left), etc. (Figure 3.30).
R Graphics 79

income
15

10

Northeast
South
Murder

income income North Central

15 West

10

0 5000 10000 15000 20000

Population

Figure 3.30
Crime rates and population size by class of income using postscript theme
(see Section 3.2.4, p. 81).

3.2.3 Customising Graphs

3.2.3.1 Panel Function
The panel option is used to more accurately control the appearance of the
graph within each panel. Each lattice type graph function (bwplot, xyplot,
etc., see Table 3.2 for other examples) has its own panel function (panel.bwplot,
panel.xyplot, etc.). Other panel functions are also available, such as loess
smoothing (panel.loess), linear regression, etc. (the most useful panel func-
tions can be obtained with help(panel.functions)). Of course, it is possible
to create ones own panel functions, but the graph commands must be com-
patible with the lattice package: for example, to insert text, one must not use
text, but ltext (l for lattice). For further information, we recommend the book
by Murrel (2005). However, it is relatively simple to merge two panel func-
tions together. In the following example (see Figure 3.31), we shall smooth
the crime rate according to the population sizes (panel.loess) and draw the
points which correspond to the observations (panel.xyplot):
> mypanel <- function(...) {
+ panel.loess(...)
+ panel.xyplot(...)
+ }
80 R for Statistics

> xyplot(Murder ~ Population | income,data=states,panel=mypanel)

The mypanel function may seem strange, but the ... are indeed part of the
function, and are not arguments to be detailed.

income
15

10

5
Murder

income income
15

10

0 5000 10000 15000 20000

Population

Figure 3.31
Crime rates according to population size by class of income: smoothing and
data representation.

3.2.3.2 Controlling Size

Often, the characters are too large (or small), or the banner is too wide, and
it can therefore be useful to know how to modify the size of these elements.
The following command can be used to change the overall size of the symbols
and text:
> fontsize <-trellis.par.get("fontsize")
> fonsize$text <- 8
> fontsize$points <- 5
> trellis.par.set("fontsize",fonsize)
The fontsize option is the name of the list, the values of which are within
another list which controls the size of the points and the font. In order to
control the size of the points in relation to the default size (option cex),
another option is also available:
> plot.symbol <- trellis.par.get("plot.symbol")
R Graphics 81

> plot.symbol$cex <- 0.5

> trellis.par.get("plot.symbol",plot.symbol)
For a complete list of parameters, use trellis.par.get().
However, it is also possible to control symbol size, legend font size, and
graduations (x and y axes) independently, using the following options:
> cex <- 0.5 # symbol magnification
> xlab <- list("my title for x-axis",cex=0.7) # title for x axis
> ylab <- list(cex=0.7) # no title for y axis
> scales <- list(cex=0.5) # size of the graduations

3.2.3.3 Panel Layout

It is very simple to change the layout of the panels in a lattice graph. Simply
use the layout argument from the graph functions. Thus, if we want to draw
density estimators (densityplot, see Table 3.2) for the crime rate depending
on class of income on a single row, simply request a layout with three panels
on one row (Figure 3.32):
> densityplot(~Murder|income,data=states,layout=c(3,1))

5 0 5 10 15 20

income income income

0.10

0.08
Density

0.06

0.04

0.02

0.00

5 0 5 10 15 20 5 0 5 10 15 20

Murder

Figure 3.32
Crime rates according to population size by class of income: smoothing and
data representation.

3.2.4 Exportation
To export lattice graphs, we use the same procedure as for conventional
graphs (see Section 3.1.5, p. 66). It is sometimes necessary to export a graph
in black and white. In such cases, we recommend that you replace the colours
with different symbols. This tedious step is actually managed automatically
using the trellis.device function. This function is used to manage the themes.
82 R for Statistics

Themes control the appearance of the graph by suggesting a combination of

colours, symbols and different types of lines. One simple way of visualising
these themes is to use the show.settings function:
> show.settings(standard.theme("postscript")) # black and white
> show.settings(col.whitebg()) # default
> show.settings(standard.theme("quartz")) # quartz

To export a black and white lattice graph in pdf format, simply use
> trellis.device(device="pdf",theme = standard.theme("postscript"),
file="export.pdf")
> my lattice expression(s)
> dev.off()

3.2.5 Other Packages

Other packages are available which can be used to create more advanced
graphs. Here, we simply mention their existence. The rgl package is used to
represent 3D data and to rotate them using the mouse. The ggplot2 package
attempts to combine the positive elements of conventional and lattice graphs.
Finally, the iplots package is used to create interactive graphs.

3.3 Exercises
Exercise 3.1 (Draw a function)
1. Draw the sine function between 0 and 2 (use pi).
2. Add the following title (title): plot of sine function.

Exercise 3.2 (Comparison of Distributions)

1. Draw the pdf (probability distribution function) of the standard Gaussian
distribution between 4 and 4 (use dnorm).

2. On the same graph, draw Students t-distribution to 5 and 30 degrees of

freedom. Use the curve function and a different colour for each line.
3. Add a legend at the top left to differentiate between each distribution.

Exercise 3.3 (Plotting Points)

1. Import the ozone dataset and plot the scatterplot for maximum ozone
(maxO3) and temperature (T12).

2. Plot the scatterplot maxO3 according to T12, with lines connecting the
points.
R Graphics 83

3. Draw graph in Figure 3.33 using the order function.

160
140
120
maxO3

100
80
60
40

15 20 25 30

T12

Figure 3.33
Scatterplot maxO3 according to T12.

Exercise 3.4 (Law of Large Numbers)

1. Having set the seed of the random generator (set.seed), simulate a sample
(x1 , ..., x1000 ) from Bernoullis distribution with parameter p = 0.6.
Pl
2. Calculate the successive means Ml = Sl /l where Sl = i=1 Xi . Draw Ml
according to l, then add the horizontal line with equation y = 0.6.

Exercise 3.5 (Central Limit Theorem)

1. Let us denote X1 , X2 , ... XN i.i.d. random variables following Bernoullis
distribution with parameter p. Recall the distribution of SN = X1 + . . . + XN .
Specify the mean and standard deviation.
2. Set p = 0.5. For N = 10, using the rbinom function, simulate n = 1000
occurrences S1 , . . . , S1000 of a binomial distribution with parameters N and
p. Organise the quantities Si N p into a vector U10. Do the same with
N p(1p)
N = 30 and N = 1000 to obtain two new vectors U30 and U1000.
3. In one window (par(mfrow)) represent histograms for U10, U30 and U1000,
each time overlapping the density of the standard Gaussian distribution (ob-
tained using dnorm).

Exercise 3.6 (Draw Sunspots)

1. Read data from file sunspots.csv series which details, date by date, the
relative number of sunspots. Check the variable types after reading.

2. Create a qualitative variable thirty with a value of 1 for the first year
(1749), increasing by 1 every thirty years. In order to do so, round up the
result (floor) of the division by 30 (or the division integer).
84 R for Statistics

3. Enter the colour vector which contains the following colours: green, yel-
low, magenta, orange, cyan, grey, red, green and blue. Automatically check
that these colours are indeed contained within the colors() vector (instruc-
tions %in% and all).
4. Draw the time series as seen in Figure 3.34. Use the palette function, plot,
lines and a loop (see also unique).

+
250

+ +
+
++
+ + ++
200

++ +++ ++
+ + +
Number of sunspots

+ + + + +
++
+
++ + + ++ ++
++
+ ++ + + + + + ++
++
+ ++++
+
+ ++ ++ + ++ + + +
+ + ++ + ++ +
150

+ + + ++ ++ + + ++
+ +++ ++
++ + ++
++ ++ + +
+ + ++++
+ ++ + ++ +
++
+ + +
+ + + + + ++
+++ ++ +++ +
+
+ + +
+ ++++
++
++ + ++ + +++ + ++ + ++ + + ++ +
+
+ +++ +++ ++ ++ ++ +++ +
++ + ++ ++ ++++ ++
++ + +++++ + ++ + + + + ++ + + +++ ++
+
+
++ + ++
+ + ++++ + ++
++++
++ + +++ ++ + ++ + +++ ++ + +++ + + + ++ + + + + +
+ +++
100

+
++ +
+ + +++ + + + ++ + ++++ ++ + +
++ + + + +
++
+ +++ +++ + ++ + +
+ ++ + +
++ +
+
+
+
+
+
++ + +
++ ++ ++++ + + ++ + ++ +++
++
++ ++ ++ + + ++ +++ +++ +++ + + + + +++ +
+++ + ++
++ ++ + ++
+++ ++ + ++
+ + +
++
++ ++ +
+
++ +
+
+
+ +
++
+
+ ++ + +++ ++ + ++
++ + ++
+ ++++ +
+
+
+ + +++ ++ ++ ++ + +
++ +
+
+++ + ++++
+
+ + ++++++ +
++ + +
+ +
++++ + +
+
+ ++ + + ++ + + + + +
++
+ + + ++ +++
+ ++ + + + +++ +++ + + + + + + +
+++
+ ++ +++ + +++ ++ + + ++ +++++++++ + ++
+
+ ++
++
++++
+ ++
++ +++
+ +++ +++
++++ + ++++++ +++ +++ + +++ ++ + ++ +++
++ +++
++ +++ + +++ +++++++ ++
+
+ +
++ +++++++
++ +
++ ++
++++ ++ ++ +
+++ ++
++
+ ++
+ ++++++ ++ +
++ ++ ++ +
++++ +++++
+ +
++
++ +++
++ ++ ++++ +++++++ ++ +++ +++ + ++ +
+ ++
+ ++ ++ ++ ++ + + + +++ + ++ + ++ ++ ++ + + ++++
50

++ +
+ +
+ + + + + +
+ +
+++ + + +
+ +
+ + + +++ ++
+++ + +
+ +
+ +
++ + + ++ + ++ + + + + + +++ + +
+ + ++
+ + +
++ +
+ ++
++++
+++ +
+++ +++ + + + +
++ + +
+++++ + + ++
+ +++ ++++ + + +
+ ++ + +
+ + +++ + + +
++ + + + ++++++ +
++ ++++++ ++++ + +
+
++++ ++ + ++ ++ ++ +
+
+ +++
++ +
+++ ++ ++++ + ++ + +++ ++
++
+ + ++
++ ++
++ ++ ++ ++
+++ ++ ++ ++ ++++ +++++ ++++ ++ +++ ++ + ++ +
+
+ ++
++++
+ ++++ ++ ++ ++ + +++ ++
+
++++
++ ++ +++
+++
+
++
+ +
++++++ +++ +
+
+ ++ ++ +
+ + ++ +
+ ++
++ +
++ +
+++ ++ ++++
+ ++++ ++++ + ++
+ ++ ++ + +
++++ ++
+
++ ++ +++ ++ +++ + ++++
+++
+ +
+ + ++ +
+++ +
++
+ +++
+++++
+ ++ +++
++
+
+ +++++ +
++
++ ++
+ + +
++ +++
+++ ++ + +++
+++ ++ ++++++ + +
+ ++++ ++++ + +++
++++ ++++ ++ ++
++
+
+++ +
+
+
+
+
+
+++
+
+
+++
+
++ + + +++ + + +
+++ +
+
+++
++
++
+++++ +++++
++
+ +++
+++++ ++ +
++ + +
++ + + + +++
+
++ +
+ ++ ++ +++
++++ +
+
++
+
++ ++
+ +++ +
++
+ +++ ++ + +
++ +
+ ++ + + ++ ++
+++
+
0

0 500 1000 1500 2000 2500

Time

Figure 3.34
Sunspots according to observation number.

Exercise 3.7 (Draw a Density)

The objective of this exercise is to reproduce the following graph (Figure 3.35).
0.4
0.3
Density

A  5%
0.2
0.1
0.0

3 2 1 0 1 2 3

Figure 3.35
Density of a normal distribution.

1. Draw the density of the random variable X, X N (0, 1) (see dnorm).

2. Add the abscissa axis (see abline).
R Graphics 85

3. Identify the area below the curve to the right of q corresponding to a

probability of 5% and fill it blue (polygon); see Figure 3.35.

4. Add an arrow identifying the coloured area (arrows).

5. At the end of the arrow, write = 5% (text and expression).

Exercise 3.8 (Multiple Graphs)

1. Reproduce the graph in Figure 3.36.
10
8
6
4
2

2 4 6 8 10
1.4

1:10
1.2

c(2, 3, 1, 0)

2
1.0

1
0.8

0
0.6

1.0 1.5 2.0 2.5 3.0 3.5 4.0 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Figure 3.36
Three graphs on two rows.

2. Reproduce the same graph as above but with a graph at the bottom right
with a width of 1 for 4 compared with that at the bottom left (see the argu-
ments for layout).

Exercise 3.9 (Draw Points and Lines)

1. Import the ozone table.
2. Use the lattice package to draw the scatterplot maxO3 according to T12.
3. Use the lattice package to draw the scatterplot maxO3 according to T12
with a line connecting the points (type "p" and type "l").

4. Use the lattice package to draw the scatterplot maxO3 according to T12,
with lines connecting the points (type "a"). Compare this line to that in
Figure 3.33 (Exercise 3.3, p. 83). What changed using the type="a" option
specific to the lattice package?

Exercise 3.10 (Panel)

1. Construct a data-frame with the first three rows of maxO3 and T12 from
the ozone table.
86 R for Statistics

2. Concatenate the table according to the rows so that the table appears six
times.

3. Construct an ordinal factor for which the observations are 3 times "h", 3
times "b", 3 times "p", 3 times "l", 3 times "s" and 3 times "S". The order
of levels will be "p", "l", "b", "h", "s" and "S".
4. Add this ordinal factor to the data of question 2.

5. Reproduce graph in Figure 3.10 (p. 58) with xyplot and the appropriate
function panel (see subscripts in the help section of xyplot).
6. Reproduce the same graph but instead of using an ordinal factor, use a
non-ordinal (classical) factor and observe the difference.
4
Making Programs with R

Programming in R is based on the same principles as for other software for sci-
entific calculations. It therefore uses familiar programming structures (loops,
condition if else, etc.) as well as predefined functions specific to statistical
practices.

4.1 Control Flows

4.1.1 Grouped Expressions
A group of commands is similar to brackets in mathematics: Those commands
which are grouped together are executed together. In R, groups of commands
are contained within curly brackets:
> {
+ expr1
+ expr2
+ ...
+ }
Two successive commands are separated by a new line (Enter key). However,
it is possible to separate two commands and to keep them on the same line.
To do this, simply separate them with a semi-colon. Therefore the above
example could also be written as
> { expr1; expr2; ... }

4.1.2 Loops (for or while)

Conventional loops are available in R. Let us start with the for loop. We
want to display all the integers from 1 to 99. One solution is as follows:
> for (i in 1:99) print(i)

The i index takes as its values all the coordinates of the chosen vector. If we
want to display only odd integers, simply create a vector which starts at 1
88 R for Statistics

and goes up to 99, at intervals of two. The seq function is used to create just
such a vector. The loop becomes
> for (i in seq(1,99,by=2)) print(i)
This method can easily be generalised to any vector. Thus, if we choose a
character vector which represents the first three days of the week, we obtain
> vector <- c("Monday","Tuesday","Wednesday")
> for (i in vector) print(i)
[1] "Monday"
[1] "Tuesday"
[1] "Wednesday"
In general there are multiple orders to be executed for each iteration. To
do so, the commands must be grouped together. Generally speaking, the for
loop is expressed:
> for (i in vector) {
+ expr1
+ expr2
+ ...
+ }
Another possibility of a loop is the while condition. Its general syntax is
as follows:
> while (condition) {
+ expr1
+ expr2
+ ...
+ }
The orders expr1, expr2, etc. are executed as long as the condition is true,
and this is evaluated at the beginning of the loop. As soon as the condition
is found to be false, the loop is stopped. Thus,
> i <- 1
> while (i<3) {
+ print(i)
+ i <- i+1
+ }
[1] 1
[1] 2
is used to display i and to increase it by 1 when i is less than 3.
One final possibility for a loop is the order repeat. It is understood as:
repeat the orders indefinitely. To ensure the loop is stopped, we use the order
break. This order can be used for any loop. An example is given in the
following subsection.
Making Programs with R 89

4.1.3 Conditions (if, else)

This involves executing an order under condition: the order is executed if,
and only if, the condition is true. In its simplest form it is written as:
> if (condition) {
+ expr1
+ expr2
+ ...
+ }
For example, if we want to use a repeat loop to print i varying from 1 to
3 inclusive, we must come out of the loop prior to printing, and when i is
greater than 3,
> i <- 1
> repeat {
+ print(i)
+ i <- i+1
+ if (i>3) break
+ }
Here, if i is greater than 3, we only need to execute a (break) order. Therefore
the commands do not need to be grouped together (the curly brackets can be
omitted).
Another condition can be added after if: the else condition, which can
be used to distinguish between two cases. Either the condition is true, in
which case the order (or group of orders) after if is executed, or it is not
true and therefore the order (or group of orders) after else is executed. In
its most general form, the if, else condition is written as:

> if (condition) {
+ expr1
+ expr2
+ ...
+ } else {
+ expr3
+ expr4
+ ...
+ }
Be aware that the order else must be on the same line as the } closing
bracket of the if clause (as above).
90 R for Statistics

4.2 Predefined Functions

Certain functions in R are predefined in order to avoid having to use loops
which generally have lengthy calculation times. The most common of these
functions is undoubtedly apply, which is used to apply one function to all the
margins of a table. Let us consider the table X made up of twenty randomly
chosen integers between 1 and 20 (sample). The mean for each column is
calculated as follows:
> set.seed(1234)
> X <- matrix(sample(1:20,20),ncol=4)
> X
[,1] [,2] [,3] [,4]
[1,] 3 10 7 9
[2,] 12 1 5 17
[3,] 11 4 20 16
[4,] 18 8 15 19
[5,] 14 6 2 13
> apply(X,MARGIN=2,FUN=mean)
[1] 11.6 5.8 9.8 14.8
It is also possible to add additional arguments to the function which is applied
to each column. For example, if there is a missing piece of data in X, it can
be interesting to use the argument na.rm=TRUE of the mean function, which
only calculates the mean of the present data:
> X[1,1] <- NA
> apply(X,MARGIN=2,FUN=mean)
[1] NA 5.8 9.8 14.8
> apply(X,MARGIN=2,FUN=mean,na.rm=TRUE)
[1] 13.75 5.80 9.80 14.80
Due to the frequent use in statistics of means per column (or row), there is a
shortcut in the form of a function colMeans (or rowMeans):
> colMeans(X,na.rm=TRUE)
[1] 13.75 5.80 9.80 14.80
In the same way, column (or row) sums can also be executed directly using
colSums (or rowSums).
If the data table is three dimensional (data cube), it is possible to ex-
ecute the function by row (MARGIN=1), by column (MARGIN=2) or by depth
(MARGIN=3). It is also possible to execute the function for confronting rows and
columns (MARGIN=c(1,2)), or rows and depth (MARGIN=c(1,3)) or columns
and depth (MARGIN=c(2,3)). As an example, let us calculate the sum of a
table with three entries for each row/column pair:
Making Programs with R 91

> set.seed(1234)
> Y <- array(sample(24),dim=c(4,3,2))
> Y
, , 1
[,1] [,2] [,3]
[1,] 3 18 11
[2,] 15 13 8
[3,] 14 1 10
[4,] 22 4 23

, , 2
[,1] [,2] [,3]
[1,] 17 21 5
[2,] 16 2 20
[3,] 24 7 9
[4,] 19 6 12
> apply(Y,MARGIN=c(1,2),FUN=sum,na.rm=TRUE)
[,1] [,2] [,3]
[1,] 20 39 16
[2,] 31 15 28
[3,] 38 8 19
[4,] 41 10 35
Here we have used the R functions mean and sum, but we could just as easily
have used functions which we programmed in advance. For example,

> MyFunction <- function(x,y) {

+ z <- x+sqrt(y)
+ return(1/z)
+ }
> set.seed(1234)
> X <- matrix(sample(20),ncol=4)
> X
[,1] [,2] [,3] [,4]
[1,] 3 10 7 9
[2,] 12 1 5 17
[3,] 11 4 20 16
[4,] 18 8 15 19
[5,] 14 6 2 13
> apply(X,MARGIN=c(1,2),FUN=MyFunction,y=2)
[,1] [,2] [,3] [,4]
[1,] 0.22654092 0.08761007 0.11884652 0.09602261
[2,] 0.07454779 0.41421356 0.15590376 0.05430588
[3,] 0.08055283 0.18469903 0.04669796 0.05742436
[4,] 0.05150865 0.10622236 0.06092281 0.04898548
[5,] 0.06487519 0.13487607 0.29289322 0.06937597
92 R for Statistics

Many functions are based on the same principle as the apply function. For
example, the tapply function applies one same function, not to the margins
of the table but instead at each level of a factor or combination of factors:
> Z <- 1:5
> Z
[1] 1 2 3 4 5
> vec1 <- c(rep("A1",2),rep("A2",2),rep("A3",1))
> vec1
[1] "A1" "A1" "A2" "A2" "A3"
> vec2 <- c(rep("B1",3),rep("B2",2))
> vec2
[1] "B1" "B1" "B1" "B2" "B2"
> tapply(Z,vec1,sum)
A1 A2 A3
3 7 5
> tapply(Z,list(vec1,vec2),sum)
B1 B2
A1 3 NA
A2 3 4
A3 NA 5
lapply, or its equivalent, sapply, applies one same function to every element
in a list. The difference between these two functions is as follows: The lapply
function by default yields a list when the sapply function yields a matrix
or a vector. Let us therefore create a list containing two matrices and then
calculate the mean of each element of the list; here is each matrix:
> set.seed(545)
> mat1 <- matrix(sample(12),ncol=4)
> mat1
[,1] [,2] [,3] [,4]
[1,] 9 4 1 11
[2,] 10 2 3 6
[3,] 7 5 8 12
> mat2 <- matrix(sample(4),ncol=2)
> mat2
[,1] [,2]
[1,] 4 3
[2,] 2 1
> mylist <- list(matrix1=mat1,matrix2=mat2)
> lapply(mylist,mean)
$matrix1
[1] 6.5

$matrix2
[1] 2.5
Making Programs with R 93

It is even possible to calculate the sum by column of each element of the list
using the apply functions as FUN function of the lapply function:
> lapply(mylist,apply,2,sum,na.rm=T)
$matrix1
[1] 26 11 12 29

$matrix2
[1] 6 4
The aggregate function works on data-frames. It separates the data into
sub-groups defined by a vector, and calculates a statistic for all the variables
of the data-frame for each sub-group. Let us reexamine the data which we
generated earlier and create a data-frame with two variables Z and T:
> Z <- 1:5
> T <- 5:1
> vec1 <- c(rep("A1",2),rep("A2",2),rep("A3",1))
> vec2 <- c(rep("B1",3),rep("B2",2))
> df <- data.frame(Z,T,vec1,vec2)
> df
Z T vec1 vec2
1 1 5 A1 B1
2 2 4 A1 B1
3 3 3 A2 B1
4 4 2 A2 B2
5 5 1 A3 B2
> aggregate(df[,1:2],list(FactorA=vec1),sum)
FactorA Z T
1 A1 3 9
2 A2 7 5
3 A3 5 1
Sub-groups can also be defined by a vector generated by two factors:
> aggregate(df[,1:2],list(FactorA=vec1,FactorB=vec2),sum)
FactorA FactorB Z T
1 A1 B1 3 9
2 A2 B1 3 3
3 A2 B2 4 2
4 A3 B2 5 1
The sweep function is used to apply one single procedure to all the margins
in a table. For example, if we want to centre and standardise the columns of
a matrix X, we write
> set.seed(1234)
> X <- matrix(sample(12),ncol=3)
94 R for Statistics

> X
[,1] [,2] [,3]
[1,] 2 10 3
[2,] 7 5 8
[3,] 11 1 4
[4,] 6 12 9
> mean.X <- apply(X,2,mean)
> mean.X
[1] 6.5 7.0 6.0
> sd.X <- apply(X,2,sd)
> sd.X
[1] 3.696846 4.966555 2.943920
> Xc <- sweep(X,2,mean.X,FUN="-")
> Xc
[,1] [,2] [,3]
[1,] -4.5 3 -3
[2,] 0.5 -2 2
[3,] 4.5 -6 -2
[4,] -0.5 5 3
> Xcr <- sweep(Xc,2,sd.X,FUN="/")
> Xcr
[,1] [,2] [,3]
[1,] -1.2172540 0.6040404 -1.0190493
[2,] 0.1352504 -0.4026936 0.6793662
[3,] 1.2172540 -1.2080809 -0.6793662
[4,] -0.1352504 1.0067341 1.0190493
It must be noted that to centre and standardise table X, we can more
simply use the scale function. The by function, however, is used to apply one
function to a whole data-frame for the different levels of a factor or list of
factors. This function is thus similar to the tapply function, as it is used with
a data-frame rather than a vector. Let us generate some data:
> set.seed(1234)
> T <- rnorm(100)
> Z <- rnorm(100)+3*T+5
> vec1 <- c(rep("A1",25),rep("A2",25),rep("A3",50))
> don <- data.frame(Z,T)
We can thus obtain a summary of each variable for each level of the factor
vec1:
> by(don,list(FactorA=vec1),summary)
FactorA: A1
Z T
Min. :-2,540 Min. :-2.3457
1st Qu.: 2.380 1st Qu.:-0.7763
Making Programs with R 95

Median : 3.662 Median :-0.4907

Mean : 4.331 Mean :-0.2418
3rd Qu.: 5.737 3rd Qu.: 0.2774
Max. :12.221 Max. : 2.4158
------------------------------------
FactorA: A2
Z T
Min. :-1.856 Min. :-2.1800
1st Qu.: 1.520 1st Qu.:-1.1073
Median : 3.071 Median :-0.8554
Mean : 2.991 Mean :-0.6643
3rd Qu.: 4.050 3rd Qu.:-0.4659
Max. : 9.321 Max. : 1.4495
------------------------------------
FactorA: A3
Z T
Min. :-0.7953 Min. :-1.80603
1st Qu.: 3.2903 1st Qu.:-0.56045
Median : 5.1260 Median :-0.04396
Mean : 5.4811 Mean : 0.13953
3rd Qu.: 8.0739 3rd Qu.: 0.81208
Max. :12.6221 Max. : 2.54899
Be careful: If we request the sum, we obtain the sum for all the variables
for each level of the factor vec1 rather than the sum variable by variable for
each level of vec1:
> by(don,list(FactorA=vec1),sum)
FactorA: A1
[1] 102.2201
-------------
FactorA: A2
[1] 58.16823
-------------
FactorA: A3
[1] 281.0313
It is also possible to perform more complex calculations. For example, if
we want to calculate a regression for each level of the variable vec1, we can call
on by to repeat a function which we will define ourselves (Section 4.3), which
here is the regression of variable Z according to variable T from the dataset
x. This functions only argument is x but it also calculates the coefficients of
linear regression.
> myfunction <- function(x){
+ summary(lm(Z~T, data=x))$coef
+ }
96 R for Statistics

> by(don, vec1, myfunction)

vec1: A1
Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.066900 0.1732763 29.24174 1.073704e-19
T 3.045363 0.1850066 16.46084 3.214797e-14
------------------------------------------------------
vec1: A2
Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.090834 0.2419277 21.04279 1.592367e-16
T 3.160780 0.2340362 13.50551 2.018228e-12
------------------------------------------------------
vec1: A3
Estimate Std. Error t value Pr(>|t|)
(Intercept) 5.081208 0.1700007 29.88935 1.120636e-32
T 2.865977 0.1640595 17.46914 1.939013e-22
This type of calculation is also automated in a different way by the lmList
function available in the nlme package.
There are also other functions such as the replicate function, which is
used to repeat an expression n times (see Table 4.1). Usually, the expression
includes a random number generation, and evaluating this expression n times
does not yield the same result. This therefore means making a loop for (i
in 1:n). For example,
> set.seed(1234)
> replicate(n=8, mean(rnorm(100)))
[1] -0.224412908 0.165109001 0.129179613 -0.012143006
[5] 0.055836342 -0.002691679 0.091426757 0.078374289
The outer function is used to repeat a function for each occurrence of the
combination of two vectors. For example,
> Month <- c("Jan","Feb","Mar")
> Year <- 2008:2010
> outer(Month,Year,FUN="paste")
[,1] [,2] [,3]
[1,] "Jan 2008" "Jan 2009" "Jan 2010"
[2,] "Feb 2008" "Feb 2009" "Feb 2010"
[3,] "Mar 2008" "Mar 2009" "Mar 2010"
Again, it is possible here to add arguments to the chosen function. Here, we
separate the months and the years by a dash rather than a space, which is
the default character:
> outer(Month,Year,FUN="paste",sep="-")
[,1] [,2] [,3]
[1,] "Jan-2008" "Jan-2009" "Jan-2010"
[2,] "Feb-2008" "Feb-2009" "Feb-2010"
[3,] "Mar-2008" "Mar-2009" "Mar-2010"
Making Programs with R 97
TABLE 4.1
Summary of a Few Functions that Can Be Used to Avoid loops
Function Description
aggregate Applies a function to a subset of rows from a data-
frame (mean of all the variables by sex, etc.). Yields a
data-frame (usually)
apply Applies a function to a margin (row, column, etc.) of
a matrix or table with multiple entries (column sum,
row sum, etc.)
by, tapply Applies a function to a subset of rows of a data-frame
or vector (mean of all variables by sex, etc.). Generally
yields a list
mapply Repeats the application of a function taking the first
coordinates of each vector as its first argument, the sec-
ond coordinates of each vector as its second argument,
etc. (see also Vectorize)
outer Applies a (vectorised) function to each pair of two vec-
tors
replicate Evaluates the same expression n times
sapply, lapply Applies a function to each coordinate of a vector or
each component of a list. Yields a vector, a matrix or
a list
sweep Applies a function to a margin (row, column, etc.) of
a matrix or table with multiple entries with a STATS
argument which varies with each margin (divides each
column by its standard deviation, substracts the mean
from each column, etc.)

4.3 Creating a Function

Functions are used to carry out many different R commands. These orders
can depend on arguments provided as input, but this is not always the case.
Functions provide unique result objects (see Figure 4.1). These resulting
objects are specified within the functions by the return function. By default,
if the written function yields no results, the last result obtained prior to output
is given as the result.
98 R for Statistics

function

Argument 1 expression 1
Argument 2 expression 2 unique
... ... result

Figure 4.1
Diagram of an R function.

Let us begin with a simple example: the sum of the first n integers. The
number n is an integer which is the input argument; the result is simply the
sum, as requested:

> mysum <- function(n) {

result <- sum(1:n)
return(result)
}
The function is called upon using its name (mysum), followed by the input
arguments between brackets. Here, the function has an argument as an input.
It is therefore called upon by simply using
> mysum(3)
[1] 6
It is possible to attribute the result of a function to an R object using

> res <- mysum(3)

> res
[1] 6
This function can be improved as follows:

> mysum <- function(n) {

+ if (n<=0) stop("n must be a positive integer")
+ if (floor(n)!=n) warning(paste("rounds",n,"as",floor(n)))
+ result <- sum(1:floor(n))
+ return(result)
+ }
Of course, when n is not larger than one, the sum is not meaningful. The
function is thus stopped (stop) and displays an error message. On the other
hand, when n is larger than one but not a whole number, the function rounds
n down to its integer part (floor) and sends a message indicating that n has
been replaced (warnings).
The function is called in much the same way:
Making Programs with R 99

> mysum(4.325)
[1] 10
Warning message:
In mysum(4.325) : rounds 4.325 as 4
We now suggest a function with two input arguments: factor1 and factor2,
two qualitative variables. This function yields the contingency table as well
as the character vector for the levels of factor1 and factor2, for which the
combined sample size is zero. Here, more than one result will be given. As
one is a table and the other a character vector (or character matrix), these
two objects can neither be brought together in a matrix (as they are not of
the same type), nor can they be combined in a data-frame (as they are not
the same length). The only result will therefore be a list grouping together
these two results.
The function will therefore calculate the contingency table (table) and
then select the void cells. We then need to know the indices corresponding to
the void cells of the contingency table (which function, option arr.ind=TRUE)
and identify the names of the corresponding categories:
> myfun <- function(factor1,factor2) {
+ res1 <- table(factor1,factor2)
+ selection <- which(res1==0,arr.ind = TRUE)
+ res2 <- matrix("",nrow=nrow(selection),ncol=2)
+ res2[,1] <- levels(factor1)[selection[,1]]
+ res2[,2] <- levels(factor2)[selection[,2]]
+ return(list(tab=res1,level=res2))
+ }
If we call upon the function with the factors wool and tension:
> tension <- factor(c(rep("Low",5),rep("High",5)))
> wool <- factor(c(rep("Mer",3),rep("Ang",3),rep("Tex",4)))
This yields
> myfun(tension,wool)
$tab
factor2
factor1 Ang Mer Tex
High 1 0 4
Low 2 3 0

$level
[,1] [,2]
[1,] "High" "Mer"
[2,] "Low" "Tex"
To find out more about using a function, see Section 1.5.
100 R for Statistics

4.4 Exercises
Exercise 4.1 (Factorial)
1. Programme the factorial function: n! = n (n 1) 2 1 using the
prod function.
2. Programme the factorial function using a loop for.

Exercise 4.2 (Ventilation (allocation at random))

1. Repeat Exercise 2.6 (p. 49) and programme a ventilation function. As its
arguments, this function will take both a variable and the threshold from
which categories must be ventilated. Set the threshold to 5% by default.
2. Write a function so that all the qualitative variables in a table can be
ventilated (see previous question).

Exercise 4.3 (Ventilation on an Ordered Factor)

1. Repeat Exercise 2.7 (p. 49) and programme a ventilation function on an
ordered factor. As its arguments, this function will take both a variable and
the threshold from which categories must be ventilated. Set the threshold to
5% by default.

2. Write a function so that all the ordered qualitative variables in a table can
be ventilated (see previous question).
 Part II

Statistical Methods

101
This page intentionally left blank
Introduction to the Statistical
Methods Part

In this part we present most of the classical statistical methods in the form of
worked examples. Each method is illustrated using a concrete example.
The structure of the worked examples is always the same. We start with
a quick presentation of the method with theoretical reminders (section objec-
tive). Then follows a description of the dataset and the associated problem
(section example). We present the steps to solve this problem (section steps).
The example is processed step by step, and a brief commentary of the results
is given (section processing the example). In this part, the different instruc-
tions are given using command lines. However, wherever possible, we use the
R Commander interface available in the Rcmdr package (see Appendix A.3,
p. 267) which allows one to perform the various methods using easy-to-use
drop-down menus. In conclusion, a section called Taking Things Further
gives reference books or less-standard extensions to the method.
Finally, it must be noted that, as each worked example is designed to be
consulted independently of the others, there will certainly be some overlaps.
From a purely practical standpoint, each dataset analysed in each worked
example is available on the website
http://www.agrocampus-ouest.fr/math/RforStat
There are therefore two ways of importing data, both of which must be done
at the beginning of the study:
Either by saving the file in the working directory and using the following
command line

> mydata <- read.table("MyDataFile.csv", header=TRUE, sep=";")

Or by reading the dataset directly from the website:

> mydata <- read.table("http://www.agrocampus-ouest.fr/math/

RforStat/MyDataFile.csv", header=TRUE, sep=";")

We believe that learning R requires a solid knowledge base detailed in the

first four chapters. However, teaching experience has shown that impatient
users often try to learn to use the software on their own. For such users, here
follows a brief summary of the previous sections.
This page intentionally left blank
5
A Quick Start with R

5.1 Installing R
Simply go to the CRAN website at http://cran.r-project.org/ and in-
stall the version of R best suited to your computers operating system. For
comprehensive documentation on the installation process, visit http://cran.
r-project.org/doc/manuals.

5.2 Opening and Closing R

In Windows and Mac OS X, look for R in Program Files. In Linux, simply
type R in a command window. To close R, either use the menu or type the
command q(). When exiting the program, the software will ask if you want
to save the work session. This means that the next time that you open R, you
will be able to retrieve all the objects you have constructed. Most of the time
it is preferable to save the instructions rather than all the objects. To do so,
write and save the instructions in a text file, known as a script.

5.3 The Command Prompt

When R is launched, a window opens and waits for some instructions with a
prompt >. R can be used immediately as a calculator, for example,
> 2+3.2
[1] 5.2

The [1] indicates that the first (and only) coordinate of the resulting vector
is 5.2. When R is waiting for the next instruction, the command prompt
becomes +. For example, if you type
> 1 -
106 R for Statistics

R will wait for the second part of the subtraction and the command prompt
will be +. To substract 2 from 1, type
+ 2
[1] -1
Generally, either a ) or " has been forgotten. Simply suggest one (or more)
bracket(s) or speech mark(s) to end the command.

5.4 Attribution, Objects, and Function

R makes calculations using functions (which will be written in bold):
> sqrt(2)
[1] 1.414214
Results can be attributed to more or less complex objects using <- or =. Thus
we create vector x with integer values from 3 to 7:
> x <- 3:7
> x
[1] 3 4 5 6 7
The content can be displayed simply by typing the name of the object.

5.5 Selection
To construct the vector y with coordinates 2 and 4 of vector x, type
> y <- x[c(2,4)]
> y
[1] 4 6
Select columns 1 and 3 of a matrix m and rows 2 and 5 respectively using
> m <- matrix(1:15,ncol=3) #creates the matrix
> m[,c(1,3)] #selects columns 1 and 3
> m[c(2,5),] #selects rows 2 and 5
In combination, rows 4 and 2 of columns 2 and 3 are obtained as follows:
> m[c(4,2),c(2,3)]
[,1] [,2]
[1,] 9 14
[2,] 7 12
A Quick Start with R 107

We can also assign objects of different kinds to lists:

> mylist <- list(vector=x,mat=m)

The component mat from the list mylist is selected using
> mylist$mat

5.6 Other
All commands can be stopped using the shortcut Ctrl + c (or the STOP
icon). It is possible to reuse preceding commands using the up or down arrows.
Additionally, to obtain help for the mean function, type

> help(mean)

5.7 Rcmdr Package

The simplest way to get to grips with R is undoubtedly to use the Rcmdr pack-
age which combines R with an easy-to-use drop-down menu. The advantage
of this package is that is also provides the lines of code which correspond to
the analyses carried out: Users can therefore familiarise themselves with the
programming by seeing which functions are used. To use this package, install
it and then launch it using
> library(Rcmdr)

5.8 Importing (or Inputting) Data

Usually, the first thing to do is to input or import data. Data can be easily
imported from a spreadsheet (such as Excel). Generally speaking, data files
contain the names of the variables in the first row and sometimes identifiers
for the individuals in the first column. We recommend that this file be saved
in a format which is easy to export, such as txt or csv. The file myfile.csv
contains the values for two variables x and y, with measurements taken from
fifty individuals. Let us suppose that the file is located at the same destination
as R. The command by which the file can be imported and placed within an
object named dat is
108 R for Statistics

> dat <- read.table("myfile.csv", sep=";", header=TRUE, dec=".")

if for example the column separator is ;, the name of the variables is present
on the first row (header=TRUE) and the decimal separator is .. Then check
that the reading was successful by summarising the dataset
> summary(dat)

5.9 Graphs
To construct a histogram and then a boxplot for all the individuals of the
second variable, we write
> hist(dat[,2])
> boxplot(dat[,2])
If the second variable is called y, we can use the name of the variable rather
than selecting the column by writing dat[,"y"] instead of dat[,2]. If the
first variable is called x, y is represented in confrontation with x by writing
> plot(y~x,data=dat)
If the variable x is quantitative, the graph is a scatterplot. If x is qualitative,
the graph is a boxplot for each category of x.

5.10 Statistical Analysis

Different functions can be used to conduct various statistical analyses. The
easiest way to know how to use a function is to refer to its help section by
typing help(myfunction). The help section demonstrates how the function is
defined, all of its arguments and, finally, usage examples that can be directly
copied and pasted into R. If the outputs of the function are assigned to an
object named result, it is possible to list all the objects present in result
using names(result). For example, after the regression of y according to x,
> result <- lm(y~x,data=dat)
> names(result)
[1] "coefficients" "residuals" "effects" "rank"
[5] "fitted.values" "assign" "qr" "df.residual"
[9] "xlevels" "call" "terms" "model"
The result$coefficients command gives access to the object containing
the regression coefficients. For certain functions, the summary function can
be used to summarise the main results.
6
Hypothesis Test

6.1 Confidence Intervals for a Mean

6.1.1 Objective
The aim of this worked example is to construct a confidence interval for the
mean of a quantitative variable from a sample. The method presented here
can be extended to construct a confidence interval of a proportion (see Sec-
tion 6.1.6, Taking Things Further).
The confidence interval for the population mean for a variable X is
expressed as:
 

x t1/2 (n 1) , x + t1/2 (n 1)
n n
where x is the empirical mean of the sample, the estimated standard devia-
tion, n the sample size, and t1/2 (n 1) the quantile 1 /2 of the Students
t-distribution with (n 1) degrees of freedom. It is important to remember
that the procedure assumes that the estimator of the mean X follows a normal
distribution. This is true if X follows a normal distribution or if the sample
size is large enough (in practice, n > 30, thanks to the central limit theorem).

6.1.2 Example
We examine the weight of adult female octopuses. We have access to a sample
from 240 female octopuses fished off the coast of Mauritania. For the overall
population, we would like to obtain an estimation of the mean of the weight
and a confidence interval for this mean with a threshold of 95%.

6.1.3 Steps
1. Read the data.
2. Calculate the descriptive statistics (mean, standard deviation).
3. Construct a histogram.
4. (Optional) Test the normality of the data.
110 R for Statistics

5. Construct the confidence interval.

6.1.4 Processing the Example

1. Reading the data:
Import the data by using the read.table function. Then check that the data
has been correctly imported:

> octopus <- read.table("OctopusF.txt",header=T)

> dim(octopus)
[1] 240 1
> summary(octopus)

Weight
Min. : 40.0
1st Qu.: 300.0
Median : 545.0
Mean : 639.6
3rd Qu.: 800.0
Max. :2400.0

Check that the sample size is equal to n = 240. Then check that the vari-
able Weight is indeed quantitative and obtain the minimum, the mean, the
quartiles and the maximum for this variable.

2. Calculating the descriptive statistics:

Next, estimate the mean (x) and the standard deviation () within the sample:

> mean(octopus)
Weight
639.625
> sd(octopus)
Weight
445.8965

3. Constructing a histogram:
Prior to any statistical analysis, it can be useful to represent the data graph-
ically. As the data is quantitative, a histogram is used:

> hist(octopus[,"Weight"],main="",nclass=15,freq=FALSE,
ylab="Density",xlab="Weight")

We use the argument freq=FALSE, which is equivalent to prob=TRUE, to obtain

a representation of the density, rather than frequency. The argument nclass
Confidence Intervals for a Mean 111

is not compulsory but can be used to choose the number of intervals. The
histogram (Figure 6.1) gives an idea of the weight distribution.
0.0012
0.0008
Density

0.0004
0.0000

0 500 1000 1500 2000

Weight

Figure 6.1
Histogram of octopus weight.

This distribution is asymmetric and is far from a normal distribution.

4. (Optional) Testing the normality of the data:
In order to construct a mean confidence interval, it is assumed that its esti-
mator follows a normal distribution. For any large sample (over thirty indi-
viduals), this is a reasonable hypothesis. For small samples, the normality of
the sample should be tested, rather than that of the mean. In order to do so,
it is possible to have an account of probability plotting for informal assess-
ment of normality (using the qqnorm function) or to use the Shapiro-Wilk test
(shapiro.test). If normality is refused, we can construct a confidence interval
using, for example, a bootstrap procedure (use the function boot.ci from the
package boot). In this example, the sample size is 240 and these tests are
therefore unnecessary.
5. Constructing the confidence interval:
The confidence interval is calculated using the function t.test:

> t.test(octopus$Weight,conf.level=0.95)$conf.int
[1] 582.9252 696.3248
attr(,"conf.level")
[1] 0.95

Strictly speaking, the unknown mean of X, estimated to be 639.625 g, either

does or does not belong to the interval [583, 696]. It is therefore within the
interval with a probability of 0 or 1. However, the procedure as a whole
112 R for Statistics

guarantees that, if it is repeated infinitely with new samples of the same size
n = 240, then 95% of the confidence intervals will contain the true unknown
value of . In practice, it is faster to say that the true mean is between
583 g and 696 g at a confidence level of 95%.
It is also possible tocalculate the confidence interval by hand by calculating
xt239 (0.975)/ n, where t239 (0.975) is the quantile 97.5% of the Students
t-distribution with 239 degrees of freedom:
> mean(octopus$Weight)-qt(0.975, df=239)*
sd(octopus$Weight)/sqrt(240)
[1] 582.9252
> mean(octopus$Weight)+qt(0.975, df=239)*
sd(octopus$Weight)/sqrt(240)
[1] 696.3248

6.1.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
then select the file.
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
2. Calculating the descriptive statistics:
Statistics Summaries Numerical summaries...
3. Constructing a histogram:
Graphs Histogram... Then choose the number of classes and the scale
of the Densities axes.
4. (Optional) Testing the normality of the data: Statistics Summaries
Shapiro-Wilk test of normality...
5. Constructing the confidence interval:
We use a function which tests the equality of the mean with a predefined value
(which is 0 by default) and as output this function provides the confidence
interval: Statistics Means Single-sample t-test...

6.1.6 Taking Things Further

To construct a confidence interval for a proportion, we use the function prop.test
(see also the worked example on test for equality of two proportions in Sec-
tion 6.4, p. 125).

These classical methods are explained in detail in many books such as

Clarke and Cooke (2004).
Chi-Square Test of Independence 113

6.2 Chi-Square Test of Independence

6.2.1 Objective
The objective of this worked example is to test the independence between
two qualitative variables. First, we consider that the data is in a contingency
table (two-way table). In Section 6.2.6, Taking Things Further, we examine
a case in which the data is in a table of individuals variables.
To test the independence of two qualitative variables, we test the null
hypothesis H0 : the two variables are independent against the alternative
hypothesis H1 : the two variables are not independent. In order to do so,
we calculate the following test statistic:
I X J
X (nij Tij )2
2 = ,
i=1 j=1
Tij

where nij is the number of individuals (observed frequency) who take the
category i of the first variable and the category j of the second, and Tij
corresponds to the expected frequency under the null hypothesis, and I and
J are the number of categories for each
P
of the variables.P Thus, Tij = npi pj
nij n
with n the total sample size, pi = jn and pj = in ij . Under H0 , 2
follows a chi-square distribution with (I 1) (J 1) degrees of freedom.
If there is no independence, it is interesting to calculate the contribution of
pairs of categories to the chi-square statistic in order to observe associations
between categories.

6.2.2 Example
This is a historic example by Fisher. He studied the hair colour of boys and
girls in a Scottish county:

Fair Red Medium Dark Jet Black

Boys 592 119 849 504 36
Girls 544 97 677 451 14

We would like to know if hair colour is independent of sex with a type-one

error rate of 5%. If hair colour does depend on sex, we would like to study the
association between categories, for example to know which colours are more
common in girls than in boys.

6.2.3 Steps
1. Input the data.
114 R for Statistics

2. Visualise the data.

3. (Optional) Calculate row and column profiles.
4. Construct the chi-square test.
5. Calculate the contributions to the chi-square statistic.

6.2.4 Processing the Example

1. Inputting the data:
Input the data manually into a matrix. Assign row names (rownames) and
column names (colnames):
> colour<-matrix(c(592,544,119,97,849,677,504,451,36,14),ncol=5)
> rownames(colour)<-c("Boys","Girls")
> colnames(colour)<-c("Fair","Red","Medium","Dark","Jet Black")
2. Visualising the data:
The data can be represented as more common in girls than in boys according
to sex in one graph window (Figure 6.2). In order to do so, we use the par
function, (see Section 3.1.6, p. 67) and the attribute which defines the number
of graphs by row and column (mfrow=c(2,1)):
> par(mfrow=c(2,1))
> barplot(colour[1,],main="Boys")
> barplot(colour[2,],main="Girls")

Boys
200 400 600 800
Observed frequency

Fair Red Medium Dark Jet Black

Hair colour
Girls
600
Observed frequency

400
200
0

Fair Red Medium Dark Jet Black

Hair colour

Figure 6.2
Distributions of hair colour by sex.
Chi-Square Test of Independence 115

3. (Optional) Calculating row and column profiles:

To explore the data, it can be interesting to calculate the row and column
profiles. First calculate the joint frequencies, that is to say, the percentages
nij /n. The results can be presented by multiplying by 100 and then rounding
to the first decimal place using the round function:

> round(100 * colour / sum(colour), 1)

Fair Red Medium Dark Jet Black
Boys 15.2 3.1 21.9 13.0 0.9
Girls 14.0 2.5 17.4 11.6 0.4
P
The row profiles (nij / j nij ) are calculated using the prop.table function
and the argument margin = 1. Note that, to use this function, the dataset
must be a matrix. If this is not the case, it must be converted to a matrix
using the function as.matrix.

> round(100 * prop.table (colour, margin = 1), 1)

Fair Red Medium Dark Jet Black
Boys 28.2 5.7 40.4 24.0 1.7
Girls 30.5 5.4 38.0 25.3 0.8

This yields the distributions of hair colour according to sex which are repre-
sented in Figure 6.2. There is very little difference between these row profiles.
The link between the two variables is not immediately obvious; it can there-
fore be useful to conduct a test.

P
The column profiles are calculated in the same way (nij / i nij ) while speci-
fying that this time we are working on columns (margin = 2):

> round(100 * prop.table (colour, margin = 2), 1)

Fair Red Medium Dark Jet Black
Boys 52.1 55.1 55.6 52.8 72
Girls 47.9 44.9 44.4 47.2 28

With regards to the column profiles, the greatest differences are for those with
jet black hair: 72% of those with jet black hair are boys.
4. Constructing the chi-square test:
In order to conduct the test of independence between the variables sex and
hair colour, we calculate the 2obs value and determine the p-value associated
with the test:

> results <- chisq.test(colour)

> results
116 R for Statistics

Pearsons Chi-squared test

data: colour
X-squared = 10.4674, df = 4, p-value = 0.03325

The p-value indicates that such a big 2obs value would have 3.32% probability
of being observed in a sample of this size if hair colour were independent of
sex. Thus, at the 5% threshold, we reject the independence hypothesis and
conclude that, according to this dataset, hair colour depends on sex.
5. Calculating the contributions to the 2 -statistic:
This relationship between the two variables can be studied in greater detail
(n T )2
by calculating the contributions ijTij ij to the 2obs statistic. The square
roots of these contributions are in the object residuals. By dividing each
term by the total (i.e. the 2obs value contained in the stat object), we obtain
the contributions expressed as percentage:

> round(100 * results$residuals^2 / results$stat, 1)

Fair Red Medium Dark Jet Black
Boys 7.8 0.4 6.5 2.9 28.4
Girls 9.2 0.5 7.7 3.4 33.4

The combinations which contribute the most to the non-independence of the

two variables are those concerning Jet Black. In order to interpret the con-
tribution, we inspect the residuals, and more precisely their positivity or neg-
ativity:

> round(results$residuals, 3)
Fair Red Medium Dark Jet Black
Boys -0.903 0.202 0.825 -0.549 1.723
Girls 0.979 -0.219 -0.896 0.596 -1.870

It can be said that the number of boys with jet black hair is greater than
expected (implied: greater than if the independence hypothesis were true)
and there are fewer girls than expected.

6.2.5 Rcmdr Corner

In Rcmdr, it is impossible to construct a chi-square test of independence from a
contingency table previously input. If the data is in the form of a contingency
table, we recommend that you conduct the test using the command window
(see above).
Nevertheless, it is possible:
To input data directly into Rcmdr and then to analyse that data. In this
Chi-Square Test of Independence 117

way, we can obtain the row or column percentages, the contributions and
the test statistic:
Statistics Contingency tables
Enter and analyze two-way table...
To use the existing data in the form of a table of individuals vari-
ables (see Section 6.2.6, Taking Things Further). We then use the
menu Statistics Contingency tables Two-way table to select
two variables and obtain a contingency table. This procedure yields the
result of the test and the contributions to the 2 statistic.

6.2.6 Taking Things Further

The data may arrive from a table of individuals variables: with the children
in rows (i.e. the statistical individual), and the qualitative variables in the
columns, which here are colour and sex. The same procedure can therefore
be carried out as before using the function xtabs (see Section 2.6, p. 46).
Once the contingency table has been constructed, it is possible to analyse it
as previously described:
> cont.tab <- xtabs(~colour+sex, data=dataset)
> chisq.test(cont.tab)

The contingency tables can be visualised with a correspondence analy-

sis (CA) by using the CA function (Worked Example 10.2, p. 222) from the
FactoMineR package.
Theoretical explanations and examples of the chi-square test are available
in many books, such as Clarke and Cooke (2004).
118 R for Statistics

6.3 Comparison of Two Means

6.3.1 Objective
The aim of this worked example is to test the equality of the means of two sub-
populations (1 and 2 ) for a quantitative variable X. In formal terms, we test
the hypothesis H0 : 1 = 2 against the alternative hypothesis H1 : 1 6= 2
(or H1 : 1 > 2 or H1 : 1 < 2 ).
In order to carry out this test, we have to know if the variances of the vari-
able X (12 and 22 ) in each sub-population are equal or not. Consequently, we
must first test the hypothesis H0 : 12 = 22 against the alternative hypothesis
2
H1 : 12 6= 22 . The test statistic is F = 12 and under the null hypothesis this
2
quantity follows a Fishers distribution with, respectively, n1 1 and n2 1 de-
grees of freedom (where n1 and n2 are the sample size in each sub-population).

If we accept the equality of the two variances, we test the equality of

the two means using the following t-test. The  variance
 of the difference
2 (n 1)12 +(n2 1)22 1 1
(X1 X2 ) is equal to D = 1 n1 +n 2 2 n1 + n2 . The test statistic
is T = X1DX2 and, under the null hypothesis (H0 : 1 = 2 ), this quantity
follows a Students t distribution with n1 + n2 2 degrees of freedom.
If we reject the equality of the two variances, we test the equality of the
two means using Welchs t-test. The variance of the difference (X1 X2 ) is
2 2
2
equal to D = n11 + n22 . The test statistic is T = X1DX2 and, under the null
hypothesis (H0 : 1 = 2 ), this quantity follows a Students t distribution
 2 2  2 2
/n /n
with degrees of freedom, where 1 = n111 12 1 + n211 22 2 .
D D

6.3.2 Example
We want to compare the weights of male and female adult octopuses. We
have a dataset with fifteen male and thirteen female octopuses fished off the
coast of Mauritania. Table 6.1 features an extract from the dataset.
TABLE 6.1
Extract from the Octopus
Dataset (weight in grams)
Weight Sex
300 Female
700 Female
850 Female
.. ..
. .
5400 Male
Comparison of Two Means 119

We would like to test the equality of the unknown theoretical mean female
(1 ) and male (2 ) octopus weights, with a type-one error rate set at 5%.

6.3.3 Steps
1. Read the data.
2. Compare the two sub-populations graphically.
3. Calculate the descriptive statistics (mean, standard deviation, and quar-
tiles) for each sub-population.
4. (Optional) Test the normality of the data in each sub-population.
5. Test the equality of variances.
6. Test the equality of means.

6.3.4 Processing the Example

1. Reading the data:

> octopus <- read.table("Octopus.csv",header=T,sep=";")

Summary of the dataset:

> summary(octopus)
Weight Sex
Min. : 300 Female:13
1st Qu.:1480 Male :15
Median :1800
Mean :2099
3rd Qu.:2750
Max. :5400

Using this procedure, we obtain descriptive statistics and we can check that
the variable Weight is quantitative and that the variable Sex is qualita-
tive. Be aware that if, for example, the categories of the variable Sex were
coded 1 and 2, it would be necessary to transform this variable into a fac-
tor prior to any analysis as R would consider it to be quantitative (see Sec-
tion 1.4.2, p. 7). In order to do so, we would write octopus[,"Sex"] <-
factor(octopus[,"Sex"]).
2. Comparing the two sub-populations graphically:
Before any analysis, it may be interesting to visualise the data. Boxplots are
used to compare the distribution of weights in each category of the variable
Sex:
120 R for Statistics

> boxplot(Weight ~ Sex, ylab="Weight", xlab="Sex", data=octopus)

Figure 6.3 shows that the males are generally heavier than the females as both
weight medians and quartiles are greater for males.
5000
4000
3000
Weight

2000
1000

Female Male

Sex

Figure 6.3
Boxplots of octopus weights according to sex.

3. Calculating the descriptive statistics for each sub-population:

Now calculate the mean, standard deviation and quartiles for each sex using
the tapply function (the argument na.rm = TRUE is in fact useless here as
there is no missing data):

> tapply(octopus[,"Weight"], octopus[,"Sex"], mean, na.rm=TRUE)

Female Male
1405.385 2700.000
> tapply(octopus[,"Weight"], octopus[,"Sex"], sd, na.rm=TRUE)
Female Male
621.9943 1158.3547
> tapply(octopus[,"Weight"],octopus[,"Sex"],quantile,na.rm=TRUE)
$Female
0% 25% 50% 75% 100%
300 900 1500 1800 2400

$Male
0% 25% 50% 75% 100%
1150 1800 2700 3300 5400

4. (Optional) Testing the normality of the data in each sub-population:

In order to construct the comparison of means test, we assume that the mean
estimator in each sub-population follows a normal distribution. This is true
Comparison of Two Means 121

if the data is normally distributed or if the sample size is large enough (in
practice, greater than 30, thanks to the central limit theorem). Here there are
less than thirty individuals; therefore data normality must be tested for each
sub-population. In order to do so, the Shapiro-Wilk test is used. To test the
normality of the males alone, select the weight of the males by requiring that
the qualitative variable Sex carries the category Male. Rows are selected by
building the logic vector select.males. The components of this vector are
TRUE for males and are otherwise FALSE. We then conduct the Shapiro-Wilk
test on the individuals in this selection. Before conducting the Shapiro-Wilk
test, we draw the normal QQ-plot that gives an account of probability plotting
for informal assessment of normality (see Figure 6.4):

Figure 6.4
Normal QQ-plot for male octopus weights.

> select.males <- octopus[,"Sex"]=="Male"

> qqnorm(octopus[select.males,"Weight"])
> qqline(octopus[select.males,"Weight"],col="grey")
> shapiro.test(octopus[select.males,"Weight"])

Shapiro-Wilk normality test

data: octopus[select.males, "Weight"]

W = 0.935, p-value = 0.3238
122 R for Statistics

As the p-value associated with the test is greater than 5%, the normality of
the males weights is accepted. We will not provide the output for the females,
but the assumption of normality is also accepted.
When the assumption of normality is rejected, the test of equality of means can
be conducted using non-parametric tests such as that of Wilcoxon (wilcox.test)
or Kruskal-Wallis (kruskal.test).
5. Testing the equality of variances:
In order to compare the mean of the two male and female sub-populations,
there are two possible types of tests: one in which the unknown variances of the
two sub-populations are different and the other in which they are equal. We
must therefore test the equality of variances H0 : 12 = 22 against H1 : 12 6= 22
using an F -test:

> var.test(Weight ~ Sex, conf.level=.95,data=octopus)

F test to compare two variances

data: Weight by Sex

F = 0.2883, num df = 12, denom df = 14, p-value = 0.03713
alternative hypothesis: true ratio of variances is
not equal to 1
95 percent confidence interval:
0.0945296 0.9244467
sample estimates:
ratio of variances
0.2883299

The p-value associated with the test of comparison of variances is 0.037: H0

can therefore be rejected and we can thus consider that the variances are
significantly different. The variance ratio (1 /2 ) is 0.288 and the confidence
interval for this ratio (to 95%) is [0.09 ; 0.92]. In the variance ratio, the
numerator is the variance of the weights at the first level of the variable
Sex alphabetically (here, Female) and the denominator is the variance of the
weights at the second level of the variable Sex.
6. Testing the equality of means:
We here use the t.test function. As the variances are different, Welchs test is
used to compare the means. In order to do so, specify that the variances are
unequal using the argument var.equal=FALSE. If the variance were equal, we
would have used a Students t-test with the argument var.equal=TRUE. The
default test is bilateral (alternative=two.sided), however the alternative
hypothesis H1 may be that the males are lighter (alternative="less") or
heavier (alternative="greater"). We here consider the category of refer-
ence to be Female (first level of the variable Sex) and the other category is
tested according to this reference.
Comparison of Two Means 123

> t.test(Weight~Sex, alternative=two.sided, conf.level=.95,

var.equal=FALSE, data=octopus)

Welch Two Sample t-test

data: Weight by Sex

t = -3.7496, df = 22.021, p-value = 0.001107
alternative hypothesis: true difference in means
is not equal to 0
95 percent confidence interval:
-2010.624 -578.607
sample estimates:
mean in group Female mean in group Male
1405.385 2700.000

The p-value (0.001) associated with the test of unequal variances indicates
that the means are significantly different. The mean weight for males within
the population (estimated at 2700 g) is thus significantly different from that
of females (estimated at 1405 g).

6.3.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
It must then be specified that the column separator is ;.
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
2. Comparing the two sub-populations graphically:
Graphs Boxplot...
Then click on Plot by groups... to obtain one box for each category of the
qualitative variable (Identify outliers with mouse is used to identify the
extreme individuals, i.e. those beyond the boxplots).
3. Calculating the descriptive statistics for each sub-population:
Statistics Summaries Numerical summaries...
Then click on Summarize by groups...
4. (Optional) Testing the normality of the data in each sub-population:
Be aware that the test for normality is conducted for the overall data and
cannot be conducted directly on a subset. It is therefore necessary to define
the subsets.
Data Active data set Subset active data set...
124 R for Statistics

The subset is determined using Subset expression (type Sex == "Male")

and it is essential to rename the dataset in order not to lose the original (Name
for new data set).
We can then conduct the Shapiro-Wilk test:
Statistics Summaries Shapiro-Wilk test of normality...
We then come back to the original data in order to conduct the same analysis
for the females: We again select the initial dataset as active by clicking on
Data. We conduct the test of normality for the females (in the same way as
for the males). We then select the initial dataset.
In summary, all of this is somewhat restrictive and it is preferable to conduct
these tests via a command line.
5. Testing the equality of variances:
Statistics Variances Two-variances F-test...
Choose the qualitative variable (Groups (pick one)), the quantitative vari-
able (Response variable (pick one)) and the alternative test hypothesis
(by default, the alternative hypothesis is that the variances are different and
the test is Two-sided).

6. Testing the equality of means:

Statistics Means Independent samples t-test...
Choose the qualitative variable (Groups (pick one)), the quantitative vari-
able (Response variable (pick one)) and the alternative test hypothesis
(by default, the alternative hypothesis is that the means are different and the
test is Two-sided). Specify whether you consider the variances to be equal or
unequal (Assume equal variances?).

6.3.6 Taking Things Further

If we need to compare more than two means, a number of tests are available.
The choice of test depends whether or not the variances in each sub-population
are equal. If the variances are equal, refer to the worked example on one-way
analysis of variance (p. 157); if the variances are unequal, use the function
oneway.test (Welch, 1951). Bartletts test (bartlett.test) can be used to test
equality of variances.

Both the tests of comparisons of variance and of means are explained in

detail in most books such as Clarke and Cooke (2004).
Testing Conformity of a Proportion 125

6.4 Testing Conformity of a Proportion

6.4.1 Objective
The aim of this worked example is to test the equality of a proportion at a
given value 0 . If P is the estimator of and n the sample size, nP follows
a binomial distribution with parameters n and . This distribution allows
one to perform an exact test about the probability of success in a Bernoulli
experiment.
This test can also be used to test the equality of two proportions. Indeed,
testing 1 = 2 remains to test 1 = 0.5.

6.4.2 Example
We are interested in the intention to vote for a candidate A in the second
round of presidential elections. In a poll of 1040 voters, candidate A wins
52.4% of the vote. Can we consider to the 95% threshold that this candidate
will win the election? Implicitly, we assume that the sample is representative
of the population as a whole, drawn with replacement, and that the voters
will vote as predicted by the poll. However, simple random polls use samples
drawn without replacement. Nevertheless, the procedure explained below is a
reasonable approximation when the poll rate is low, that is, when the size of
the population is much greater than that of the sample.

6.4.3 Step
Test the equality of the proportion to 50% with a type-one error rate of 5%.

6.4.4 Processing the Example

As an alternative hypothesis to the test of the equality of a proportion of
50%, we choose the percentage of voting intentions to be greater than 50%
(alternative="greater"). The success rate (i.e. the number of votes for
candidate A) must be a whole number. We therefore round the decimal yielded
by the problem to the nearest whole number (using the round function with
0 decimal digits). We then use the binom.test function, but the prop.test
function yields exactly the same result:
> nbr.vot.A <- round(0.524 * 1040,0)
> binom.test(nbr.vot.A, n=1040, p=0.5, alternative="greater")
Exact binomial test

data: nbr.vot.A and 1040

nb of successes = 545, nb of trials = 1040, p-value = 0.06431
126 R for Statistics

alternative hypothesis: true probability of success is > 0.5

95 percent confidence interval:
0.4980579 1.0000000
sample estimates:
probability of success
0.5240385
As the p-value is greater than 5%, we accept the hypothesis that the pro-
portion is equal to 50%: we cannot therefore confirm from this poll that
candidate A will win the election.
The previous exact binomial test yields a confidence interval for the pro-
portion at the 95% level. However, since we have constructed a unilateral
test, this interval places all the error (i.e. 5%) on one side. More pre-
cisely, the interval provided is of the format [a, 1] where a is determined so as
Pr(a p 1) 0.95. To obtain a confidence interval of the format [a, b], with
approximately 2.5% on each side, a bilateral test is required (two.sided), and
this is the default option:
> binom.test(nbr.vot.A,n=1040)
Exact binomial test

data: nbr.vot.A and 1040

number of successes=545, number of trials=1040, p-value=0.1286
alternative hypothesis: true probability of success
is not equal to 0.5
95 percent confidence interval:
0.4931733 0.5547673
sample estimates:
probability of success
0.5240385
Strictly speaking, the unknown probability p estimated at 0.524 either does or
does not belong to the interval [0.493, 0.555]. It is therefore within the interval
with a probability of 0 or 1. However, the overall procedure guarantees that,
if it were to be repeated infinitely with new samples of the same size n = 1040
participants, 95% of the confidence intervals would contain the true unknown
value p. In practice, even if that is not the case, it is quicker to say that the
percentage of voting intentions is between 49.3% and 55.5%.

6.4.5 Rcmdr Corner

There is no test for the conformity of a proportion in Rcmdr.

6.4.6 Taking Things Further

The tests of conformity to a proportion are explained in detail in most books
such as Moore et al. (2007).
Comparing Several Proportions 127

6.5 Comparing Several Proportions

6.5.1 Objective
The aim of this worked example is to test the equality of several proportions.

6.5.2 Example
We shall again examine the example of hair colour for boys and girls in a
Scottish county (see Worked Example 6.2, p. 113):

Fair Red Medium Dark Jet Black

Boys 592 119 849 504 36
Girls 544 97 677 451 14

We would like to compare the proportions of boys for different groups, which
here is different hair colours. We will test whether or not these proportions
are equal in all the groups, with a type-one error set at 5%.

6.5.3 Step
Test the equality of proportions for boys (for different hair colours).

6.5.4 Processing the Example

To conduct the test of equality of proportions, we provide a list with the
number of boys for each hair colour, then a list with the total number of
individuals for each hair colour:
> prop.test(c(592,119,849,504,36),n=c(1136,216,1526,955,50))

5-sample test for equality of proportions without continuity

correction

data: c(592,119,849,504,36) out of c(1136,216,1526,955,50)

X-squared = 10.4674, df = 4, p-value = 0.03325
alternative hypothesis: two.sided
sample estimates:
prop 1 prop 2 prop 3 prop 4 prop 5
0.5211268 0.5509259 0.5563565 0.5277487 0.7200000
Significant differences between proportions can thus be confirmed (the p-
value is lower than 5%). We can see that the proportion of boys estimated for
the jet black group is much greater than for the other groups. Note that the
128 R for Statistics

test which is performed is a chi-square test of independence and the results are
exactly the same as the ones obtained in the worked example on chi-square
(p. 113).

6.5.5 Rcmdr Corner

There is no test of comparison of proportions in Rcmdr.

6.5.6 Taking Things Further

The prop.test function can also be applied in cases where there are two propor-
tions (corresponding to two variables with two categories) as Yates correction
for continuity is offered by default in this situation.

Tests for equality of proportions are explained in detail in books such as

Moore et al. (2007).
Power of a Test 129

6.6 The Power of a Test

6.6.1 Objective
The aim of this worked example is to calculate the power of the test of equality
of means of two sub-populations. The test of equality of means is used to
choose between the hypothesis H0 : 1 = 2 and the alternative hypothesis
H1 : 1 6= 2 (or H1 : 1 > 2 or H1 : 1 < 2 ) (see Worked Example 6.3).
The power of the test is the probability of rejecting the hypothesis H0 when
H1 is true. Power is equal to 1 , being the type-two error (that is to
say, the risk of mistakenly accepting H0 ). The power therefore corresponds
to the probability of detecting a difference in means, if indeed this difference
exists. The advantage of calculating the power of a test prior to conducting
an experiment lies mainly in the ability to optimise the number of trials (i.e.
statistical individuals) according to the aim of the experimenter. Indeed, the
power of the test is directly related to the number of individuals per group
(n), the amplitude of the difference that we want to detect (), the within
-group variability (), and the type-one error ().

6.6.2 Example
We now examine the example of an experiment in milk production. Re-
searchers at INRA (the French National Institute for Agricultural Research)
selected two genetically different types of dairy cow according to the volume
of milk produced. The aim is to detect a potential difference in the protein
levels in the milks produced by these two sub-populations.
During a previous study, the standard deviation of protein levels in the
milk from a herd of Normandy cows was found to be 1.7 g/kg of milk. As
an approximation we will therefore use the standard deviation = 1.7 and
use the classical = 5% threshold. The aim is to have = 80% chance
of detecting a difference in the means of the protein levels of = 1 g/kg of
milk from the two populations. To meet this objective, we will determine the
number of dairy cows required using the function power.t.test.

6.6.3 Steps
1. Calculate the number of individuals required to obtain a power of 80%.

2. Calculate the power of the test with twenty individuals per group.
3. Calculate the difference of means detectable at 80% with twenty individuals
per group.
130 R for Statistics

6.6.4 Processing the Example

1. Calculating the number of individuals required to obtain a power of 80%:

> power.t.test(delta=1, sd=1.7, sig.level=0.05, power=0.8)

Two-sample t test power calculation

n = 46.34674
delta = 1
sd = 1.7
sig.level = 0.05
power = 0.8
alternative = two.sided

NOTE: n is number in *each* group

Here we can see that a minimum of 47 individuals are required per population
in order to have more than an 80% chance of detecting a difference in means
in the protein levels of 1 g/kg of milk between the two populations.
2. Calculating the power of the test with twenty individuals per group:

> power.t.test(n=20, delta=1, sd=1.7, sig.level=0.05)$power

[1] 0.4416243

If we decide to use only twenty cows per population in the experiment, there
will be a 44% chance of detecting a difference in means in the protein levels
of 1 g/kg of milk between the two populations.
3. Calculating the difference detectable at 80% with twenty individuals per
group:

> power.t.test(n=20, sd=1.7, sig.level=0.05, power=0.8)$delta

[1] 1.545522

If we decide to use only twenty cows per population in the experiment, there
will then be an 80% chance of detecting a difference in means in the protein
levels of 1.55 g/kg of milk between the two populations.

6.6.5 Rcmdr Corner

It is not possible to calculate the power of a test with Rcmdr.
Power of a Test 131

6.6.6 Taking Things Further

If we want to compare more than two means, and if we suppose that the vari-
ances in each sub-population are equal, we can construct a one-way analysis of
variance model (see p. 157). In order to calculate the power of the test before-
hand, use function power.anova.test and then test the overall effect. In order
to do so, we must first have an idea of the within-group variance (or residual
mean square) which will be observed, as well as defining the between-group
variance (or mean square associated to the factor) that we wish to detect.
We examine the same example, but this time consider that there are three
genetic populations. We again assume that the residual standard deviation
of the protein level is 1.7 g/kg of milk and that we want to detect (with a
threshold 5%) whether there really is a difference between the three popula-
tions if the true means of protein levels are, for example, 28, 30 and 31 g/kg,
respectively.
> 1.7^2
[1] 2.89
> var(c(28, 30, 31))
[1] 2.333333
> power.anova.test(groups=3, between.var=2.3333,
within.var=2.89, power=.80)

Balanced one-way analysis of variance power calculation

groups = 3
n = 7.067653
between.var = 2.3333
within.var = 2.89
sig.level = 0.05
power = 0.8

NOTE: n is number in each group

The foreseen residual variance is 2.89 g2 /kg2 and the between-group vari-
ance that we want to detect as significant is 2.33 g2 /kg2 . We consider that
eight cows are therefore needed per group in order to have an 80% chance
of detecting the effect of the genetical population on the protein level of the
milk.
The power of a t-test for mean comparison is explained in books such as
Clarke and Cooke (2004). The power of an F -test in analysis of variance is
described in Sahai and Ageel (2000), for example.
This page intentionally left blank
7
Regression

7.1 Simple Linear Regression

7.1.1 Objective
Simple linear regression is a statistical method used to model the linear re-
lationship between two quantitative variables for explanatory or prediction
purposes. Here we have one explanatory variable (denoted X) and one re-
sponse variable (denoted Y ), connected by the following model:
Y = 0 + 1 X +
where is the variable for noise or measurement error rate. The parameters
0 and 1 are unknown. The aim is to estimate them from a sample of n pairs
(x1 , y1 ), . . . , (xn , yn ). The model is written in indexed form:
yi = 0 + 1 xi + i
The 0 coefficient corresponds to the intercept and 1 to the slope. We esti-
mate these parameters by minimising the least-square criterion
n
X
(0 , 1 ) = argmin (yi 0 1 xi )2
0 ,1 i=1

Once the parameters have been estimated (0 and 1 ), we obtain the

regression line:
f (x) = 0 + 1 x
from which predictions can be made. The adjusted or smoothed values are
defined by
yi = 0 + 1 xi
and the residuals by
i = yi yi
Analysing the residuals is essential as it is used to check the individual fitting
(outlier) and the global fitting of the model, for example by checking that
there is no structure.
134 R for Statistics

7.1.2 Example
Air pollution is currently one of the most serious public health worries world-
wide. Many epidemiological studies have proved the influence that some chem-
ical compounds such as sulphur dioxide (SO2 ), nitrogen dioxide (NO2 ), ozone
(O3 ) or other air-borne dust particles can have on our health.
Associations set up to monitor air quality are active all over France to
measure the concentration of these pollutants. They also keep a record of
meteorological conditions such as temperature, cloud cover, wind, etc.
Here we analyse the relationship between the maximum daily ozone level
(in g/m3 ) and temperature. We have at our disposal 112 observations col-
lected during the summer of 2001 in Rennes (France).

7.1.3 Steps
1. Read the data.
2. Represent the scatterplot (xi , yi ).
3. Estimate the parameters.
4. Draw the regression line.

5. Conduct a residual analysis.

6. Predict a new value.

7.1.4 Processing the Example

1. Reading the data:
Read the data and summarise the variable of interest, here maxO3 and T12:

> ozone <- read.table("ozone.txt",header=T)

> summary(ozone[,c("maxO3","T12")])
maxO3 T12
Min. : 42.00 Min. :14.00
1st Qu.: 70.75 1st Qu.:18.60
Median : 81.50 Median :20.55
Mean : 90.30 Mean :21.53
3rd Qu.:106.00 3rd Qu.:23.55
Max. :166.00 Max. :33.50

2. Representing the scatterplot (xi , yi ):

> plot(maxO3~T12,data=ozone,pch=15,cex=.5)
Simple Linear Regression 135

Each point on the graph (Figure 7.1) represents, for a given day, a temperature
measurement taken at midday, and the ozone peak for the day. From this
graph, the relationship between temperature and ozone concentration seems
to be linear.
160
140
120
maxO3

100
80
60
40

15 20 25 30

T12

Figure 7.1
Representation of pairs (xi , yi ) with a scatterplot.

3. Estimating the parameters:

The lm function (linear model) is used to fit a linear model:

> simple.reg <- lm(maxO3~T12,data=ozone)

> summary(simple.reg)
Call:
lm(formula = maxO3 ~ T12, data = ozone)

Residuals:
Min 1Q Median 3Q Max
-38.0789 -12.7352 0.2567 11.0029 44.6714

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -27.4196 9.0335 -3.035 0.003 **
T12 5.4687 0.4125 13.258 <2e-16 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 17.57 on 110 degrees of freedom

Multiple R-Squared: 0.6151, Adjusted R-squared: 0.6116
F-statistic: 175.8 on 1 and 110 DF, p-value: < 2.2e-16

Amongst other things, we obtain a Coefficient matrix which, for each pa-
136 R for Statistics

rameter (each row), has four columns: its estimation (Estimate column), its
estimated standard deviation (Std. Error), the observed value statistic for
the test H0 : j = 0 against H1 : j 6= 0. Finally, the p-value (Pr(>|t|))
yields the probability to obtain such a large statistic under H0 .
Coefficients 0 and 1 are estimated by 27.4 and 5.5. The test of significance
for the coefficients here gives p-values of 0.003 and around 2. 1016 . Thus, the
null hypothesis for each of these tests is rejected in favour of the alternative
hypothesis. The p-value of less than 5% for the constant (intercept) indicates
that the constant must appear in the model. The p-value less than 5% for the
slope indicates a significant link between maxO3 and T12.
The summary of the estimation step features the estimation of residual stan-
dard deviation , which here is 17.57, as well as the corresponding number of
degrees of freedom n 2 = 110.
The value of R2 and adjusted R2 are also given. The value of R2 is rather high
(R2 = 0.6151), thus supporting the suspected linear relationship between the
two variables. In other words, 61% of the variability of maxO3 is explained by
T12.
The final row, which is particularly useful for multiple regression, indicates
the result of the test of the comparison between the model used and the model
using only the constant as explanatory variable.
We can consult the list of different results (components of the list, see Sec-
tion 1.4.7, p. 20) of the object simple.reg using

> names(simple.reg)
[1] "coefficients" "residuals" "effects" "rank"
[5] "fitted.values" "assign" "qr" "df.residual"
[9] "xlevels" "call" "terms" "model"

We can then retrieve the coefficients using

> simple.reg$coef
(Intercept) T12
-27.419636 5.468685

or by using the coef function:

> coef(simple.reg)

To fit the model without the constant, we proceed as follows:

> reg.without.intercept <- lm(maxO3~T12-1, data=ozone)

4. Drawing the regression line:

We can simply apply the following command to obtain Figure 7.2:
Simple Linear Regression 137

> plot(maxO3~T12,data=ozone,pch=15,cex=.5)
> abline(simple.reg)

160
140
120
maxO3

100
80
60
40

15 20 25 30

T12

Figure 7.2
Scatterplot and regression line.

5. Conducting a residual analysis:

The residuals are obtained using the residuals function; however, these resid-
uals do not have the same variance (heteroscedastic). We therefore need to
use and plot (Figure 7.3) studentised residuals, which have the same variance.

> res.simple<-rstudent(simple.reg)
> plot(res.simple,pch=15,cex=.5,ylab="Residuals")
> abline(h=c(-2,0,2),lty=c(2,1,2))
3
2
1
Residuals

0
1
2
3

0 20 40 60 80 100

Index

Figure 7.3
Representation of residuals.
138 R for Statistics

In theory, 95% of the studentised residuals can be found in the interval [2, 2].
This is the case here because only four residuals are found outside this interval.

6. Predicting a new value:

As we now have a new observation xnew, we can simply use the estimations
to predict the corresponding Y value. However, the predicted value is of
little interest without its corresponding confidence interval. Let us examine
an example of this. Here, we have a new temperature observation T12 equal
to 19 degrees for 1st October 2001.

> xnew <- 19

> xnew <- as.data.frame(xnew)
> colnames(xnew) <- "T12"
> predict(simple.reg,xnew,interval="pred")
fit lwr upr
[1,] 76.48538 41.4547 111.5161

It must be noted that the xnew argument of the predict function must be a
data-frame in which the names of the explanatory variables are the same as in
the original dataset (here T12). The predicted value is 76.5 and the prediction
interval at 95% is [41.5, 111.5]. To represent, on one single graph, the confi-
dence interval of a fitted value along with that of a prediction, we calculate
these intervals for all the points which were used to draw the regression line.
We make the two appear on the same graph (Figure 7.4).

> gridx <- data.frame(T12=seq(min(ozone[,"T12"]),

max(ozone[,"T12"]),length=100))
> CIline <- predict(simple.reg,new=gridx,
interval="conf",level=0.95)
> CIpred <- predict(simple.reg,new=gridx,
interval="pred",level=0.95)
> plot(maxO3~T12,data=ozone,pch=15,cex=.5)
> matlines(gridx,cbind(CIline,CIpred[,-1]),
lty=c(1,2,2,3,3),col=1)
> legend("topleft",lty=2:3,c("pred","conf"))

7.1.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
Next specify that the field separator is the space and the decimal-point char-
acter is ..
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
Simple Linear Regression 139

conf

160
pred

140
120
maxO3

100
80
60
40

15 20 25 30

T12

Figure 7.4
Confidence and prediction intervals.

2. Representing the scatterplot graphically (xi , yi ):

Graphs Scatterplot... Then select variables x and y. It is possible to
identify specific points using the mouse. By default, a boxplot is drawn for
each variable x and y, the regression line is drawn as well as a smoothing
curve, but it is possible to uncheck these options in the dialogue box.
3. Estimating the parameters:
Statistics Fit models Linear regression... then select the ex-
planatory variable (x) and the response variable (y). It is then possible to
obtain a confidence interval for each parameter of the regression (0 and 1 ):
Models Confidence intervals...
4. Drawing the regression line:
The regression line is drawn when we create the cloud of points (see item 2).
5. Conducting a residual analysis:
Models Add observation statistics to data...
We can then add to the dataset the fitted values, the residuals, the studentised
residuals, the estimated values and Cooks distances. We can then construct
graphs using these new variables. Many graphs are available directly:
Models Graphs Basic diagnostic plots
6. Predicting a new value:
It is not possible to predict a new value automatically with Rcmdr.

7.1.6 Taking Things Further

Theoretical reminders about linear regression are available in many books,
such as Clarke and Cooke (2004), Moore et al. (2007) or Faraway (2005).
140 R for Statistics

7.2 Multiple Linear Regression

7.2.1 Objective
Multiple linear regression consists of explaining and/or predicting a quanti-
tative variable Y by p quantitative variables X1 , , Xp . The multiple re-
gression model is a generalisation of the simple regression model (see Worked
Example 7.1). We therefore suppose that the n data collected conform to the
following model:
yi = 0 + 1 xi1 + 2 xi2 + + p xip + i , i = 1, , n, (7.1)
where xij are known numbers measuring the explanatory variables which are
organised within a matrix X known as the experimental design matrix. The
parameters j of the model are unknown and need to be estimated. The
parameter 0 (intercept) corresponds to the constant of the model. The i
are random unknown variables and represent errors of measurement.
By writing the model 7.1 in the form of a matrix, we obtain
Y=X + (7.2)

y1 1 x11 ... x1p 0 1
Y = ... ,
.. .. .. .. ,
X = . = ... , = ...

. . .
yn 1 xn1 . . . xnp p n
From these observations, we estimate the unknown parameters of the model
by minimising the least-square criterion:
2
n
X Xp
= argmin yi 0 j xij = argmin(Y X)0 (Y X)
0 , ,p i=1 j=1 Rp+1

If matrix X is full rank, that is to say, if the explanatory variables are not
collinear, the least square estimator of is
= (X0 X)1 X0 Y
Once the parameters have been estimated, we can calculate the fitted
values:
yi = 0 + 1 xi1 + + p xip
or predict new values. The difference between the observed value and the
fitted value is, by definition, the residual:
i = yi yi
Analysing the residuals is essential as it is used to check the individual fitting
of the model (outlier) and the global fitting, for example by checking that
there is no structure.
Multiple Linear Regression 141

7.2.2 Example
We reexamine the ozone dataset introduced in Worked Example 7.1 (p. 133).
Here we analyse the relationship between the maximum daily ozone level (in
g/m3 ) and temperature at different times of day, cloud cover at different
times of day, the wind projection on the East-West axis at different times
of day and the maximum ozone concentration for the day before the day in
question. The data was collected during the summer of 2001 and the sample
size is 112.

7.2.3 Steps
1. Read the data.
2. Represent the variables.
3. Estimate the parameters.
4. Choose the variables.
5. Conduct a residual analysis.
6. Predict a new value.

7.2.4 Processing the Example

1. Reading the data:
Read the data and choose the variables of interest, here the response variable,
maximum ozone, denoted maxO3, and the explanatory variables: temperature,
cloud cover, wind projection on the East-West axis at 9am, midday, and 3pm,
as well as the ozone maximum from the previous day max03y.

> ozone <- read.table("ozone.txt",header=T)

> ozone.m <- ozone[,1:11]
> names(ozone.m)
[1] "maxO3" "T9" "T12" "T15" "Ne9" "Ne12"
[7] "Ne15" "Wx9" "Wx12" "Wx15" "maxO3y"

Names are attributed using the names or colnames function. It can be inter-
esting to summarise these variables (summary) or to represent them.
2. Representing the variables:
In order to make sure that there are no errors during input, it is important to
conduct a univariate analysis of each variable (histogram, for example). When
there are not many variables, they can be represented two by two on the same
graph using the pairs function (here pairs(ozone.m)). It is also possible to
explore the data using a principal component analysis with the illustrative
explanatory variable (see Worked Example 10.1, p. 209).
142 R for Statistics

3. Estimating the parameters:

To estimate the parameters, we must first write the model. Here, we use
lm (linear model) with a formula (see Appendix A.2, p. 266), as with simple
regression. Here, there are ten explanatory variables and it would therefore be
fastidious to write all of them. R software can be used to write them quickly,
using the command
> reg.mul <- lm(maxO3~.,data=ozone.m)
> summary(reg.mul)

Call:
lm(formula = maxO3 ~ ., data = ozone.m)

Residuals:
Min 1Q Median 3Q Max
-53.566 -8.727 -0.403 7.599 39.458

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 12.24442 13.47190 0.909 0.3656
T9 -0.01901 1.12515 -0.017 0.9866
T12 2.22115 1.43294 1.550 0.1243
T15 0.55853 1.14464 0.488 0.6266
Ne9 -2.18909 0.93824 -2.333 0.0216 *
Ne12 -0.42102 1.36766 -0.308 0.7588
Ne15 0.18373 1.00279 0.183 0.8550
Wx9 0.94791 0.91228 1.039 0.3013
Wx12 0.03120 1.05523 0.030 0.9765
Wx15 0.41859 0.91568 0.457 0.6486
maxO3y 0.35198 0.06289 5.597 1.88e-07 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 14.36 on 101 degrees of freedom

Multiple R-Squared: 0.7638, Adjusted R-squared: 0.7405
F-statistic: 32.67 on 10 and 101 DF, p-value: < 2.2e-16
As in simple linear regression (see Worked Example 7.1), outputs from the
software are in the form of a matrix (under the name Coefficients) with four
columns for each parameter: its estimate (column Estimate), its estimated
standard deviation (Std. Error), the observed value of the test statistic for
H0 : i = 0 against H1 : i 6= 0, and the p-value (Pr(>|t|)). The latter gives,
as a test statistic under H0 , the probability of exceeding the estimated value.
Here, when they are tested one at a time, the significant variables are maxO3y
and Ne9. However, since the regression is multiple and the explanatory vari-
ables are not orthogonal, it can be misleading to use these tests. Indeed,
Multiple Linear Regression 143

testing a coefficient means testing the significance of a variable whereas the

other variables are in the model. In other words, this means testing that the
variable does not provide any additional information, bearing in mind that
all the other variables are within the model. It is therefore important to use
procedures for choosing the models such as those presented below.
The summary of the estimation step features the estimation of residual stan-
dard deviation , which here is 14.36, as well as the number of associated
degrees of freedom n 11 = 101.
The last row indicates the result of the test of comparison between the model
used and the model which uses only the constant as the explanatory variable.
It is, of course, significant, as the explanatory variables provide information
about the response variable.
4. Choosing the variables:
It is possible to choose the variables by hand, step by step. We remove the least
significant variable, T9, and then recalculate the estimations, and so on. In R
the leaps package processes the choice of variables: The regsubsets function
yields, for different criteria (Bayesian information criterion or BIC, adjusted
R2 , Mallows Cp, etc.), the best model (if nbest=1) with one explanatory
variable, two explanatory variables, ... with nvmax explanatory variables.
The graphical representation helps to analyse the results.

> library(leaps)
> choice <- regsubsets(maxO3~.,data=ozone.m,nbest=1,nvmax=11)
> plot(choice,scale="bic")

140
140
130
130
120
bic

120
120
110
110
97
(Intercept)

T9

T12

T15

Ne9

Ne12

Ne15

Wx9

Wx12

Wx15

maxO3y

Figure 7.5
Choosing variables with BIC.

We can also determine the variables for the best model according to the BIC
criterion using the following line of code:
144 R for Statistics

> summary(choice)$which[which.min(summary(choice)$bic),]
(Intercept) T9 T12 T15 Ne9 Ne12
TRUE FALSE TRUE FALSE TRUE FALSE
Ne15 Wx9 Wx12 Wx15 maxO3y
FALSE TRUE FALSE FALSE TRUE

The criterion is optimal for the top row of the graph. For the Bayesian in-
formation criterion (argument scale="bic"), we retain the model with four
variables: T12, Ne9, Wx9 and maxO3y. We thus fit the model using the selected
variables:

> final.reg <- lm(maxO3~T12+Ne9+Wx9+maxO3y,data=ozone.m)

> summary(final.reg)

Call:
lm(formula = maxO3 ~ T12 + Ne9 + Wx9 + maxO3y, data = ozone.m)

Residuals:
Min 1Q Median 3Q Max
-52.396 -8.377 -1.086 7.951 40.933

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 12.63131 11.00088 1.148 0.253443
T12 2.76409 0.47450 5.825 6.07e-08 ***
Ne9 -2.51540 0.67585 -3.722 0.000317 ***
Wx9 1.29286 0.60218 2.147 0.034055 *
maxO3y 0.35483 0.05789 6.130 1.50e-08 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 14 on 107 degrees of freedom

Multiple R-Squared: 0.7622, Adjusted R-squared: 0.7533
F-statistic: 85.75 on 4 and 107 DF, p-value: < 2.2e-16

5. Conducting a residual analysis:

The observed residuals can be calculated using the residuals function on the
output of the lm function (residuals(final.reg)). However, these residuals
do not have the same variance (heteroscedastic). We therefore need to use
studentised residuals, which have the same variance:

> res.m <- rstudent(final.reg)

> plot(res.m,pch=15,cex=.5,ylab="Residuals")
> abline(h=c(-2,0,2),lty=c(2,1,2))

In theory, 95% of the studentised residuals can be found in the interval [2, 2].
Multiple Linear Regression 145

This is the case here where only three residuals are found outside this interval
(see Figure 7.6).

3
2
1
Residuals

0
1
2
3

0 20 40 60 80 100

Index

Figure 7.6
Representation of residuals.

6. Predicting a new value:

As we now have a new observation, we can simply use the estimators to predict
the corresponding Y value. However, the predicted value is of little interest
without its corresponding confidence interval. Let us examine an example
of this. For 1st October 2001, we obtain the following values: maxO3y=70,
T12=19, Ne9=8 and Wx9=2.05. It must be noted that the xnew argument
of the predict function must be a data-frame with the same names as the
explanatory variables of the original dataset.

> xnew <- matrix(c(19,8,2.05,70),nrow=1)

> colnames(xnew) <- c("T12","Ne9","Wx9","maxO3y")
> xnew <- as.data.frame(xnew)
> predict(final.reg,xnew,interval="pred")
fit lwr upr
[1,] 72.51437 43.80638 101.2224

The predicted value is 72.5 and the confidence interval at 95% for the predic-
tion is [43.8, 101.2].

7.2.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
Next, specify that the field separator is the space and the decimal-point char-
acter is ..
146 R for Statistics

To check that the dataset has been imported successfully:

Statistics Summaries Active data set

2. Representing the data on a graph:

Graphs Scatterplot... Then select variables x and y. Unfortunately,
the only graphs available only represent the variables two by two.
3. Calculating the parameters:
Statistics Fit models Linear regression... then select the ex-
planatory variables (x) and the response variable (y). It is then possible to
obtain a confidence interval for each parameter of the regression:
Models Confidence intervals...
4. Choosing the variables:
Many different models must be constructed and tested (the leaps package is not
implemented): Models Hypothesis tests Compare two models...
5. Conducting a residual analysis:
Models Add observation statistics to data...
We can add to the dataset the fitted values, the residuals, the studentised
residuals, the predicted values, and Cooks distances, and create graphs with
these new variables. Many graphs are available directly using
Models Graphs Basic diagnostic plots
6. Predicting a new value:
It is not possible to predict a new value automatically with Rcmdr.

7.2.6 Taking Things Further

Theoretical reminders and exercises on linear regression are available in many
books, such as Clarke and Cooke (2004), Moore et al. (2007), Faraway (2005)
or Fox and Weisberg (2011). Linear regression and many extensions are also
available in Hastie et al. (2009).
Partial Least Squares (PLS) Regression 147

7.3 Partial Least Squares (PLS) Regression

7.3.1 Objective
PLS regression consists of predicting a quantitative variable Y from p quanti-
tative variables X1 , . . . , Xp . Multiple regression (Section 7.2) is also dedicated
to this purpose. However, when the number p of explanatory variables is very
large, or if it exceeds the number n of individuals to be analysed, it is dif-
ficult, or even impossible, to use the least squares method. PLS regression
is then particularly suited for cases with data with large dimensions such as
spectrometry data or genetic data.
Put simply, PLS regression iteratively searches for a sequence of orthogo-
nal components (or underlying variables). These components, known as PLS
components, are chosen in order to maximise covariance with the explana-
tory variable Y . The choice of the number of components is important as it
greatly influences the quality of the prediction. Indeed, if the chosen num-
ber is not sufficient, important information can be forgotten. However, if
the chosen number is too big, the model may be overfitted. This choice is
generally conducted using cross-validation or training/validation. These two
procedures follow the same principle: The initial data is separated into two
distinct parts, a training part and a validation part. There are generally more
individuals in the training set than in the validation set. We construct the
PLS regression model from the training set and then calculate the predictions
on the validation set. We first construct the one-component PLS model and
then repeat this procedure for 2, 3, . . . , k PLS components. The number of
components retained corresponds to the model which leads to the minimum
prediction error. This procedure is implemented in the pls package.

7.3.2 Example
We would like to predict the organic carbon content in the ground from spec-
troscopic measurements. As measuring the organic carbon content of the
ground is more expensive than collecting spectroscopic measurements, we
would like to be able to predict organic carbon content from spectroscopic
measurements.
In order to construct a prediction model, we have access to a database1
with 177 soils (individuals) and both their organic carbon content (OC, re-
sponse variable) and the spectra acquired in the visible and near-infrared
range (400 nm2500 nm), which gives 2101 explanatory variables. We will
then predict the organic carbon content of three new soils.
1 Thank to Y. Fouad and C. Walter; Achi H., Fouad Y., Walter C., Viscarra Rossel R.A.,
Lili Chabaane Z. and Sanaa M. (2009). Regional predictions of soil organic carbon content
from spectral reflectance measurements. Biosystems Engineering, 104, 442446.
148 R for Statistics

7.3.3 Steps
1. Read the data.
2. Represent the data.
3. Conduct a PLS regression after choosing the number of PLS components.
4. Conduct a residual analysis.
5. Predict a new value.

7.3.4 Processing the Example

1. Reading the data:
While being imported, an X is added by default to the names of variables
that correspond to wavelength. Spectral data often needs to be preprocessed;
this is usually done by standardising the data from a spectrum. To do so, we
calculate the mean and the standard deviation for each spectrum (i.e. for each
row in the matrix, omitting the first column which corresponds to Y ). Then,
for all the data in a row, we subtract its mean and divide by its standard
deviation:

> Spectrum <- read.table("Spectrum_Breizh.txt", sep=";",

header=TRUE, row.names=1)
> row.mean <- apply(Spectrum[,-1], 1, mean)
> Spectrum[,-1] <- sweep(Spectrum[,-1], 1, row.mean, FUN="-")
> row.sd <- apply(Spectrum[,-1], 1, sd)
> Spectrum[,-1] <- sweep(Spectrum[,-1], 1, row.sd, FUN="/")

2. Representing the data:

The first step is to visualise the response variable Y , which corresponds to
organic carbon content OC, through the representation of its histogram:

> hist(Spectrum[,"OC"], freq=F, xlab="Organic Carbon content")

> lines(density(Spectrum[,"OC"]))
> rug(Spectrum[,"OC"])

The rug function is used to add vertical lines to the abscissa for the observa-
tions of variable Y . We can see in Figure 7.7 that one observation is extreme:
It is greater than 8, whereas all the others are less than 5. The following lines
can be used to show that individual 79 has a very high value (8.9) The other
values do not exceed 4.89.

> which(Spectrum[,1]>8)
[1] 79
> Spectrum[79,1]
Partial Least Squares (PLS) Regression 149

0.4
0.3
Density

0.2
0.1
0.0

0 2 4 6 8

Organic Carbon content

Figure 7.7
Representation of organic carbon content.

[1] 8.9
> max(Spectrum[-79,1])
[1] 4.89

This individual is unusual and can be removed when constructing the model:

> Spectrum <- Spectrum[-79,]

The explanatory variables are spectra and can be represented by curves. Each
spectrum can be represented by a different line (continuous, dashed, etc.)
and/or colour, depending on the value of the OC variable. To construct this
graph, we divide the OC variable into seven classes of almost equal size using
the cut function; the points at which they are cut are determined by the
quantiles (quantile). The factor obtained in this way is converted to a colour
and line type code using the as.numeric function and the arguments col and
lty. On the abscissa axis, we use the wavelengths:

> color <- as.numeric(cut(Spectrum[,1], quantile(Spectrum[,1],

prob=seq(0,1,by=1/7)), include.lowest = TRUE))
> matplot(x=400:2500, y=t(as.matrix(Spectrum[,-1])),type="l",
lty=color,col=color,xlab="Wavelength",ylab="Reflectance")

We obtain the graph in Figure 7.8 in which each individual is a different curve.
It would seem possible to distinguish groups of curves drawn with the same
code, which means that similar curves admit similar values for the OC variable.
3. Conducting a PLS regression after choosing the number of PLS compo-
nents:
To conduct this regression, we use the preinstalled pls package that must be
150 R for Statistics

2
1
Reflectance

0
1
2
3
4

500 1000 1500 2000 2500

Wavelength

Figure 7.8
Representation of individuals (after a standardisation by row).

loaded (see Section 1.6). By default, this package centres all the variables
but does not reduce them. We reduce the explanatory variables using the
argument scale. We must also fix a maximum number of PLS components
(the higher this number, the longer it will take to run the program). To be
safe, we fix this number at 100, which is already extremely high.

> library(pls)
> pls.model <- plsr(OC~., ncomp=100, data=Spectrum,
scale=TRUE, validation="CV")

With the pls package, the number of PLS components is, by default, deter-
mined by cross validation. We calculate the fitting error and the prediction
error obtained with 1, 2, . . . , 100 PLS components and we then draw the two
types of error using the plot.mvr function (Figure 7.9):

> msepcv.pls <- MSEP(pls.model, estimate=c("train","CV"))

> plot(msepcv.pls,col=1,type="l",legendpos="topright",main="")

The two error curves are typical: a continual decline for the fitting error de-
pending on the number of components and a decrease followed by an increase
for the prediction error. The optimal number of components for the prediction
corresponds to the value for which the prediction error is minimal:
Partial Least Squares (PLS) Regression 151

train
CV

1.0
0.8
MSEP

0.6
0.4
0.2
0.0

0 20 40 60 80 100

Number of components

Figure 7.9
Evolution of both errors depending on the number of PLS components.

> ncp <- which.min(msepcv.pls$val["CV",,])-1

> ncp
4 comps
4

We subtract 1 to obtain the number of PLS components directly (here there

are four). As previously indicated, it was not necessary to fix the maximum
number of PLS components at 100; we could have limited ourselves to 30. We
fit the final model using the four PLS components:

> reg.pls <- plsr(OC~., ncomp=ncp, data=Spectrum, scale=TRUE)

4. Conducting a residual analysis:

The residuals are obtained using the residuals function. Note that the resid-
uals are provided for the one-component PLS model, and then the two-
component model, . . . , until the model with the chosen number of compo-
nents (here, four). We therefore only draw the residuals for the last model
(Figure 7.10):

> res.pls <- residuals(reg.pls)

> plot(res.pls[,,ncp],pch=15,cex=.5,ylab="Residuals",main="")
> abline(h=c(-2,0,2), lty=c(2,1,2))

No individuals are so extreme that they need to be removed from the analysis.
It must be noted that if individual 79 had not been deleted previously, it would
have had a very large residual and would have been deleted at this point.
The variability of residuals is greater for the last individuals of the dataset
(individuals coded by rmqs) and smaller for the first (individuals coded by
Bt). The samples from the units analysed do not seem very homogeneous.
152 R for Statistics

2.0
1.5
1.0
0.5
Residuals

0.5
1.5

0 50 100 150

Index

Figure 7.10
Representation of residuals.

5. Predicting a new value:

We have obtained a set of three complementary curves. For each one, we
will compare the OC content predicted by the PLS regression to the observed
values.
First we read the data and apply the same preprocessing step as before:

> ReflecN <- read.table("Spectrum_new.txt", sep=";", header=TRUE,

row.names=1)
> row.meanN <- apply(ReflecN[,-1], 1, mean)
> ReflecN[,-1] <- sweep(ReflecN[,-1], 1, row.meanN, FUN="-")
> row.sdN <- apply(ReflecN[,-1], 1, sd)
> ReflecN[,-1] <- sweep(ReflecN[,-1], 1, row.sdN, FUN="/")

The predicted values are obtained using the predict function:

> pred <- predict(reg.pls, ncomp=ncp, newdata=ReflecN[,-1])

> pred
, , 4 comps
OC
3236 0.4065104
rmqs_726 2.6930854
rmqs_549 2.6035296

These predicted values can be compared with the observed values:

> ReflecN[,1]
[1] 1.08 1.60 1.85
Partial Least Squares (PLS) Regression 153

7.3.5 Rcmdr Corner

It is not possible to conduct PLS regression with Rcmdr.

7.3.6 Taking Things Further

PLS regression is often associated with specific graphs which can be used to
analyse the model more precisely and to describe the analysed sample.
To better understand the role of each of the variables in the regression
model, the coefficients can be drawn (Figure 7.11) for each of the wavelengths
using the following commands:
> plot(reg.pls, plottype="coef", comps=1:ncp, main="",
legendpos="topleft", xlab="Wavelength", labels=400:2500)

Comp 1
Comp 2
Comp 3
Comp 4
0.002
Regression coefficient

0.000
0.002

899 1399 1899 2399

Wavelength

Figure 7.11
Coefficients of the PLS regression for one to four-component models.

The interpretation of this type of graph goes beyond the objectives set out
in this work and we will simply underline the fact that there are coefficients
for each reflectance of a given wavelength. They can therefore be interpreted
in the same way as those of a multiple regression, the only difference being
that here they are all on the same scale, as the initial variables (reflectances)
are centred and reduced.
It is possible to represent the cloud of individuals on the PLS components
(called PLS scores) in order to visualise the similarities between individuals,
154 R for Statistics

as in PCA (see Worked Example 10.1). For example, we draw components 1

and 2 on an orthonormal index (asp=1) and colour all the samples from the
rmqs group in grey using the following commands:
> color <- rep(1,nrow(Spectrum))
> color[substr(rownames(Spectrum),0,4)=="rmqs"] <- "grey30"
> plot(reg.pls, plottype="scores", comps=c(1,2),
col=color, asp=1)
> abline(h=0,lty=2)
> abline(v=0,lty=2)
In Figure 7.12 we can see more clearly the division of the observed individuals
into two populations.
100

50

Comp 2 (32 %)

50

200 100 0 100

Comp 1 (52 %)

Figure 7.12
Analysis of the model by component: components 1 and 2.

Another classical analysis consists of representing the correlations between

the loadings and the variables as in PCA (see graph of variables, Worked
Example 10.1). We visualise the variables which are the most closely linked to
each of the components. The wavelengths can also be coloured using greyscale
(from black for the first variables to light grey for the last, see Figure 7.13).
> color <- grey(seq(0,.9,len=ncol(Spectrum)))
> plot(pls.model,plottype="cor",comps=1:2,col=color,pch=20)

Since the variables (here the wavelengths) are sorted, the evolution of their
relationship with the PLS components can be represented using the argument
plottype="loadings" (Figure 7.14):
> plot(pls.model,plottype="loadings",labels=400:2500,comps=1:ncp,
legendpos="topright",xlab="Wavelength",ylab="Loadings")
> abline(h=0,lty=2)
Partial Least Squares (PLS) Regression 155

1.0

0.5

















































Comp 2 (32 %)

0.0

0.5

1.0

1.5 1.0 0.5 0.0 0.5 1.0 1.5

Comp 1 (52 %)

Figure 7.13
Representation of the correlations with the loadings.

A brief presentation of the PLS regression can be found in Hastie et al.

(2009, p. 8082). For more details and some case studies, see Varmuza and
Filmozer (2009) or Vinzi et al. (2010).
156 R for Statistics

Comp 1 (51.8 %)
0.06

Comp 2 (32.1 %)
Comp 3 (7.4 %)
Comp 4 (2.5 %)
0.04
0.02
Loadings

0.00
0.02
0.04
0.06

899 1399 1899 2399

Wavelength

Figure 7.14
Representation of loadings.
8
Analysis of Variance and Covariance

8.1 One-Way Analysis of Variance

8.1.1 Objectives
One-way analysis of variance (or ANOVA) is a statistical method used to
model the relationship between an explanatory qualitative variable A with I
categories (or levels) and a quantitative response variable Y . The main objec-
tive of the one-way analysis of variance is to compare the empirical means of Y
for I levels of A. The one-way analysis of variance can therefore be considered
an extension of the two-way comparison of means (see Worked Example 6.3).

The one-way analysis of variance is therefore used to study the effect of a

factor A on a quantitative variable Y . To do this, we construct the following
model:

yij = i + ij , i = 1, ..., I j = 1, ..., ni

where ni is the sample size of category i, yij the observed value j for sub-
population i, i the mean of sub-population i, and ij the residual of the
model. An individual is therefore here defined by the pair (i, j). The analysis
of variance thus consists of testing the equality of the i s.
The model can also be written in the following classical format:

yij = + i + ij

where is the overall mean and i is the specific effect of category i. In this
latest formulation, there are I + 1 parameters to be estimated, of which only
I are identifiable. A linear constraint must therefore be imposed. There are
a number of different constraints and the most common are
One of the i is set at zero, which means that level i is considered the
reference level.
The sum of all i s is null, so the mean is taken as reference.
The numerical estimations of the parameters, obtained from the least
squares method, depend of course on the chosen constraint. Nevertheless,
158 R for Statistics

whatever the constraint, it is always possible to test the overall significance

of the factor. This test, which does not depend on the constraint, is Fishers
exact test. In this test, the variability explained by factor A (the between-
group variability) is compared with the residual variability (the within-group
variability). The hypotheses of this test are

H0 : i i = 0 compared with H1 : i i 6= 0.

These hypotheses require us to test the sub-model with the complete model:

yij = + ij , model under H0

yij = + i + ij , model under H1

Finally, a subsequent residual analysis is essential as it is used to check the

individual fit of the model (outlier) and the overall fit.
Remark
In analyses of variance, the collected data is often rewritten in table format
as in Table 8.1:
TABLE 8.1
Some Ozone Data Organized into Cells
Wind Ozone Values
North 87 82 114 79 101 ...
South 90 72 146 108 81 ...
East 92 121 146 106 45 ...
West 94 80 79 106 70 ...

In this table, the individuals are indexed conventionally in the format yij
where i is the row index and j the column index. This presentation of the
data indeed validates the determination of the model. Nevertheless, prior to
the analysis, the data must be transformed by constructing a classic table of
individuals variables by considering one variable for the factor and another
for the quantitative variable. For example, this yields Table 8.2.

8.1.2 Example
We reexamine in more detail the ozone dataset of Worked Example 7.1 (p. 133).
Here we analyse the relationship between the maximum daily ozone concen-
tration (in g/m3 ) and wind direction classed by sector (North, South, East,
West). The variable wind has I = 4 levels. We have at our disposal 112 pieces
of data collected during the summer of 2001 in Rennes (France).

8.1.3 Steps
1. Read the data.
One-Way Analysis of Variance 159
TABLE 8.2
Presentation of the Data Table
Wind Ozone
North 87
North 82
North 114
South 90
East 92
West 94
... ...

2. Represent the data.

3. Analyse the significance of the factor.
4. Conduct a residual analysis.
5. Interpret the coefficients.

8.1.4 Processing the Example

1. Reading the data:
Import the dataset and summarise the variables of interest, here maxO3 and
wind:

> ozone <- read.table("ozone.txt",header=T)

> summary(ozone[,c("maxO3","wind")])
maxO3 wind
Min. : 42.00 East :10
1st Qu.: 70.75 North:31
Median : 81.50 South:21
Mean : 90.30 West :50
3rd Qu.:106.00
Max. :166.00

During the summer of 2001, the most common wind direction was West, and
there were very few days with an East wind.
2. Representing the data:
Prior to an analysis of variance, boxplots are usually constructed for each level
of the qualitative variable. We therefore present the distribution of maxO3
according to wind direction:

> plot(maxO3~wind,data=ozone,pch=15,cex=.5)
160 R for Statistics

Looking at the graph in Figure 8.1, it would seem that there is a wind effect.
This impression can be validated by testing the significance of the factor.

160
140
120
maxO3

100
80
60
40

East North South West

wind

Figure 8.1
Boxplots of maxO3 according to the levels of the variable wind.

3. Analysing the significance of the factor:

First of all, we use the lm (linear model) function (or the aov function) with
a formula (see Appendix A.2, p. 266) in order to estimate the parameters of
the model. The anova function returns the analysis of variance table:

> reg.aov1 <- lm(maxO3~wind,data=ozone)

> anova(reg.aov1)
Analysis of Variance Table

Response: maxO3
Df Sum Sq Mean Sq F value Pr(>F)
wind 3 7586 2528.69 3.3881 0.02074 *
Residuals 108 80606 746.35
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

The first column indicates the factors associated degrees of freedom, the sec-
ond the sum of the squares, and the third the mean square (sum of the squares
divided by the degrees of freedom). The fourth column features the observed
value of the test statistic. The fifth column (Pr(>F)) contains the p-value,
that is to say, the probability that the test statistic under H0 will exceed the
estimated value. The p-value (0.02) is less than 5%, thus H0 is rejected and
we can accept the significance of wind at the 5% level. There is therefore at
least one wind direction for which the maximum level of ozone is significantly
different from the others.
One-Way Analysis of Variance 161

4. Conducting a residual analysis:

Residuals are available using the residuals function but they do not have the
same variance (heteroscedastic). We therefore use studentised residuals. In
order to plot the residuals according to the levels of the variable wind, we use
the package lattice which is presented in detail in Section 3.2 (p. 73).

> res.aov1 <- rstudent(reg.aov1)

> library(lattice)
> mypanel <- function(...) {
panel.xyplot(...)
panel.abline(h=c(-2,0,2),lty=c(3,2,3),...)
}
> trellis.par.set(list(fontsize=list(point=5,text=8)))
> xyplot(res.aov1~I(1:112)|wind,data=ozone,pch=20,
ylim=c(-3,3),panel=mypanel, ylab="Residuals",xlab="")

0 20 40 60 80 100

South West
L 3
L
L LL
L L 2
L L
L
L L 1
L
L
L LL L L L L
L L L L 0
L L L L LL
L LL L LL L L
L
L
L L L L L LLL
LLL LL L LL L
L L L LL L LLL L 1
L L
L 2
Residuals

East North
3
L
L
2
L
L
1 L
L L
L L
L L L L
L L L
0 L LL LL LL L
L
L
L L
L L LL LLL
L
L
1 LL
L
2
L
3

0 20 40 60 80 100

Figure 8.2
Representation of the residuals according to the levels of the variable wind.
162 R for Statistics

In theory, 95% of the studentised residuals can be found in the interval [2, 2].
Here there are nine residuals outside of the interval, that is, 8%, which is
acceptable. The distribution of the residuals seems to be comparable from
one category to another.
5. Interpreting the coefficients:
Now that an overall wind effect has been identified, we must examine how
direction influences maximum ozone levels. In order to do so, the coefficients
are analysed using the Students t-test.
As mentioned above, there are a number of different ways of writing the one-
way analysis of variance model. By default, R uses the constraint 1 = 0,
which means taking as reference the first label of the variable alphabetically,
which here is East.

> summary(reg.aov1)
Call:
lm(formula = maxO3 ~ wind, data = ozone)

Residuals:
Min 1Q Median 3Q Max
-60.600 -16.807 -7.365 11.478 81.300

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 105.600 8.639 12.223 <2e-16 ***
windNorth -19.471 9.935 -1.960 0.0526 .
windSouth -3.076 10.496 -0.293 0.7700
windWest -20.900 9.464 -2.208 0.0293 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 27.32 on 108 degrees of freedom

Multiple R-Squared: 0.08602, Adjusted R-squared: 0.06063
F-statistic: 3.388 on 3 and 108 DF, p-value: 0.02074

Amongst other things, we obtain a Coefficient matrix which, for each pa-
rameter, (each line), has four columns: its estimation (Estimate column), its
estimated standard deviation (Std. Error), the observed value of the test un-
der H0 : i = 0 against H1 : i 6= 0. Finally, the associated p-value (Pr(>|t|))
yields the probability of exceeding the estimated value.
The estimate of , here denoted Intercept, is the mean East wind concen-
tration in maxO3. The other values obtained correspond to the deviation from
this mean for each wind cell.
The last three lines of column Pr(>|t|) correspond to the test H0 : i = 0.
The following question can thus be answered: Is there a similarity between the
One-Way Analysis of Variance 163

wind considered and the wind from reference cell East? The South wind is not
different, unlike the West wind. The p-value associated with the North wind
is slightly more than 5% and thus we cannot confirm a significant difference
from the East wind.
If we want to select another specific control cell, for example the second label,
which in this case is North, it can be indicated as follows:

> summary(lm(maxO3~C(wind,base=2),data=ozone))
Call:
lm(formula = maxO3 ~ C(wind, base = 2), data = ozone)

Residuals:
Min 1Q Median 3Q Max
-60.600 -16.807 -7.365 11.478 81.300

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 86.129 4.907 17.553 <2e-16 ***
C(wind, base = 2)1 19.471 9.935 1.960 0.0526 .
C(wind, base = 2)3 16.395 7.721 2.123 0.0360 *
C(wind, base = 2)4 -1.429 6.245 -0.229 0.8194
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 27.32 on 108 degrees of freedom

Multiple R-Squared: 0.08602, Adjusted R-squared: 0.06063
F-statistic: 3.388 on 3 and 108 DF, p-value: 0.02074
P
Another possible constraint is i i = 0, and is used as follows:

> summary(lm(maxO3~C(wind,sum),data=ozone))
Call:
lm(formula = maxO3 ~ C(wind, sum), data = ozone)

Residuals:
Min 1Q Median 3Q Max
-60.600 -16.807 -7.365 11.478 81.300

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 94.738 3.053 31.027 <2e-16 ***
C(wind, sum)1 10.862 6.829 1.590 0.1147
C(wind, sum)2 -8.609 4.622 -1.863 0.0652 .
C(wind, sum)3 7.786 5.205 1.496 0.1376
---
164 R for Statistics

Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 27.32 on 108 degrees of freedom

Multiple R-Squared: 0.08602, Adjusted R-squared: 0.06063
F-statistic: 3.388 on 3 and 108 DF, p-value: 0.02074
P
By default, R provides the values of , 1 , 2 and 3 . As i i = 0, in order
to find the coefficient associated with the West wind (last level), we must
calculate

4 = 1 2 3 = 10.038

In all these analyses, it can be seen that the values of the estimates change
depending on the constraint. On the other hand, the overall test given in the
last row of the lists, which corresponds to the result of the analysis of variable
table, remains the same.
If within the same session, we want to carry out multiple analyses of variance
with the same constraint, it is preferable to use

> options(contrasts = c("contr.sum", "contr.sum"))

the model is therefore written:

> summary(lm(maxO3~wind,data=ozone))

8.1.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
Next specify that the field separator is the space and the decimal-point char-
acter is ..
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
2. Representing the data:
Graphs Boxplot... Select the response variable maxO3 and click Plot by
groups... then select the variable wind. It is possible to identify specific
points using the mouse.
3. Analysing the significance of the factor:
Statistics Fit models Linear model... then select the response
variable (y) and the explanatory variable (x).
The outputs here are estimations of Pparameters and i with constraint
1 = 0. To change this constraint to i i = 0, use: Tools Options...
then in contrasts, replace contr.Treatment and contr.poly by contr.sum.
Then click on Exit and Restart R Commander.
One-Way Analysis of Variance 165

To build the analysis of variance table:

Models Hypothesis tests ANOVA table...

4. Conducting a residual analysis:

Many graphs are available for residual analysis, but not that of the lattice
package. Nevertheless, we can obtain a few graphs using Models Graphs
5. Interpreting the coefficients:
See the item 3 Analysing the significance of the factor.

8.1.6 Taking Things Further

It is possible to construct multiple comparisons of means tests. In this case
we use the TukeyHSD function. This function can also be used for a one-
way analysis of variance, but only works for two-or-more factor analyses of
variance when the dataset is balanced.

Theoretical presentations and exercises on analysis of variance are available

in many books, such as Clarke and Cooke (2004), Moore et al. (2007), Sahai
and Ageel (2000) or Faraway (2005).
166 R for Statistics

8.2 Multi-Way Analysis of Variance with Interaction

8.2.1 Objective
Multi-way analysis of variance (or ANOVA) is the generalisation of one-way
ANOVA (see Worked Example 8.1). Using this method, it is possible to
model the relationship between a quantitative variable and many qualitative
variables. We examine a two-variable case, but the generalisation to multiple
variables is immediate. Let us consider two qualitative explanatory variables,
denoted A and B, and a quantitative response variable, denoted Y . We de-
note I as the number of levels of the variable A and J as that of variable B.

By analysing variance, when we have multiple measurements for each con-

frontation of a level of A with a level of B, it is possible and often interesting
to process the complete model using interaction. Interaction means examin-
ing the simultaneous effect of factors A and B on variable Y . The model is
classically written as:

yijk = + i + j + ij + ijk (8.1)

where is the overall mean, i is the effect due to the i-th level of the factor
A, j is the effect due to the j-th level of the factor B, ij is the interaction
effect when factor A is at level i and factor B is at level j, and ijk is the
residual term.

It is therefore interesting to test the hypotheses of an effect of factors A

and B, and the interaction AB. In order to do so, we construct Fishers exact
test to compare the explained variability with the residual variability. We
advise that you first test the significance of the interaction. Indeed, if the
interaction is significant, the two factors are influential via their interaction
and it is therefore not necessary to test their respective influence via the main
effect. The hypotheses of the interaction test are

(H0 )AB : (i, j) ij = 0 compared with (H1 )AB : (i, j) ij 6= 0.

These hypotheses involve testing the sub-model against the complete model:

yijk = + i + j + ijk , model under (H0 )AB (8.2)

yijk = + i + j + ij + ijk , model under (H1 )AB

This test, which identifies the overall influence of the interaction of the factors,
is simply a test between two nested models.

We can then proceed in the same manner to test the effect of a single
Multi-Way Analysis of Variance with Interaction 167

factor. For example, in order to test the effect of factor A, we test the sub-
model without factor A against the model with factor A but without the
interaction:

yijk = + j + ijk model under (H0 )A

yijk = + i + j + ijk model under (H1 )A

It may also be interesting to estimate the parameters of the model (, i ,

j , ij ). Nevertheless, as for the one-way analysis of variance, there are more
parameters in the model than there are calculable parameters: 1 + I + J + IJ
parameters to be estimated, whereas only IJ are calculable. It is therefore
necessary to impose 1 + I + J linearly independent constraints in order to
make the system reversible.
The classical constraints are
1. Cell-analysis constraints:

= 0, i, i = 0, j, j = 0

2. Cell reference constraints:

1 = 0, 1 = 0, i, i1 = 0, j, 1j = 0

3. Sum constraints:
X X X X
i = 0, j = 0, i, ij = 0, j, ij = 0
i j j i

8.2.2 Example
We reexamine in more detail the ozone dataset of Worked Example 7.1 (p. 133).
Here we will analyse the relationship between the maximum daily ozone con-
centration (in g/m3 ) and wind direction classed into sectors: North, South,
East, West and precipitation classed into two categories: Dry and Rainy. We
have at our disposal 112 pieces of data collected during the summer of 2001
in Rennes (France). Variable A admits I = 4 levels and variable B has J = 2
levels.

8.2.3 Steps
1. Read the data.
2. Represent the data.
3. Choose the model.
4. Interpret the coefficients.
168 R for Statistics

8.2.4 Processing the Example

1. Reading the data:
Import the dataset and summarise the variables of interest, here maxO3, wind
and rain:

> ozone <- read.table("ozone.txt",header=T)

> summary(ozone[,c("maxO3","wind","rain")])
maxO3 wind rain
Min. : 42.00 East :10 Dry :69
1st Qu.: 70.75 North:31 Rainy:43
Median : 81.50 South:21
Mean : 90.30 West :50
3rd Qu.:106.00
Max. :166.00

2. Representing the data:

A boxplot is presented for each cell (confrontation of a level of wind with a
level of rain) which yields IJ = 8 boxplots (Figure 8.3):

> boxplot(maxO3~wind*rain,data=ozone)
160
140
120
100
80
60
40

East.Dry North.Dry South.Dry West.Dry East.Rainy North.Rainy South.Rainy West.Rainy

Figure 8.3
Boxplot of max03 according to the confrontation of the levels of the variables
wind and rain.

It must be noted that the ozone level is generally higher during dry weather
than during rainy weather. Another way of graphically representing the in-
teraction is as follows (see Figure 8.4):

> par(mfrow=c(1,2))
> with(ozone,interaction.plot(wind,rain,maxO3))
> with(ozone,interaction.plot(rain,wind,maxO3))
Multi-Way Analysis of Variance with Interaction 169

rain wind
110

110
Dry South
Rainy West
100

100
mean of maxO3

mean of maxO3
East
North
90

90
80

80
70

70
East North South West Dry Rainy

wind rain

Figure 8.4
Interaction graph depicting wind (left) and rain (right) on the x-axis.

These graphs are interesting and can represent different interactions. We

therefore advise that you construct both graphs even if only the more explicit
of the two is analysed. The broken lines are almost parallel so the interaction
is not very pronounced. We thus conduct the test to compare models (8.1)
and (8.2).
3. Choosing the model:

> mod.int <- lm(maxO3~wind*rain,data=ozone)

> anova(mod.int)
Analysis of Variance Table

Response: maxO3
Df Sum Sq Mean Sq F value Pr(>F)
wind 3 7586 2528.7 4.1454 0.00809 **
rain 1 16159 16159.4 26.4910 1.257e-06 ***
wind:rain 3 1006 335.5 0.5500 0.64929
Residuals 104 63440 610.0

The first column indicates the factors associated degrees of freedom, the sec-
ond the sum of the squares, and the third the mean square (sum of the squares
divided by the degrees of freedom). The fourth column features the observed
value of the test statistic. The fifth column (Pr(>F)) contains the p-value,
that is to say, the probability that the test statistic under H0 will exceed the
estimated value. The p-value (0.65) is more than 5%, thus we retain H0 and
conclude that the interaction is not significant. We therefore consider that
the model has no interaction (8.2).

> mod.without.int <- lm(maxO3~wind+rain,data=ozone)

> anova(mod.without.int)
170 R for Statistics

Analysis of Variance Table

Response: maxO3
Df Sum Sq Mean Sq F value Pr(>F)
wind 3 7586 2528.7 4.1984 0.007514 **
rain 1 16159 16159.4 26.8295 1.052e-06 ***
Residuals 107 64446 602.3

The two factors are significant and there is therefore both a wind effect and a
rain effect on the maximum daily ozone level. We shall therefore retain this
model and analyse the coefficients.
4. Interpreting the coefficients:
The calculations depend on the constraints used. By default, the constraints
used by R are 1 = 0 and 1 = 0. However, here thereP is no reason
P to fix a
reference level. We therefore choose the constraints i i = 0 and j j = 0,
which yields

> summary(lm(maxO3~C(wind,sum)+C(rain,sum),data=ozone))
Call:
lm(formula = maxO3 ~ C(wind, sum) + C(rain, sum), data = ozone)

Residuals:
Min 1Q Median 3Q Max
-42.618 -15.664 -3.712 8.295 67.990

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 90.135 2.883 31.260 < 2e-16 ***
C(wind, sum)1 7.786 6.164 1.263 0.2093
C(wind, sum)2 -8.547 4.152 -2.059 0.0420 *
C(wind, sum)3 5.685 4.694 1.211 0.2285
C(rain, sum)1 12.798 2.471 5.180 1.05e-06 ***
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

Residual standard error: 24.54 on 107 degrees of freedom

Multiple R-Squared: 0.2692, Adjusted R-squared: 0.2419
F-statistic: 9.856 on 4 and 107 DF, p-value: 7.931e-07

Amongst other things, we obtain a Coefficient matrix with four columns for
each parameter (row). These include its estimation (column Estimate), its
estimated standard deviation (Std. Error), the observed value of the given
test statistic, and finally the p-value (Pr(>|t|)) which yields under H0 the
probability of exceeding the estimated value.
Multi-Way Analysis of Variance with Interaction 171

The estimator of , here denoted Intercept, corresponds to the mean ef-

fect.
PI The West wind effect is not included in the list. As the constraint is
i i = 0, we estimate 4 for (7.786 8.547 + 5.685) = 4.924. The ef-
fect of the category Rainy is 12.798. It is reassuring that the coefficient
which corresponds to dry weather exceeds that associated with wet weather,
as radiation is one of the catalysts of ozone production.
Like in one-way analysis of variance (see Worked Example 8.1), estimates
depend on the constraints used. However, the overall test, which corresponds
to the results of the analysis of variable table, remains the same.
If, within the same session, we want to carry out multiple analyses of variance
with the same constraint, it is preferable to use

> options(contrasts = c("contr.sum", "contr.sum"))

The models are therefore written as:

> summary(lm(maxO3~wind+rain+wind:rain,data=ozone))
> summary(lm(maxO3~wind+rain,data=ozone)) #without interaction

8.2.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
Next specify that the field separator is the space and the decimal-point char-
acter is ..
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
2. Representing the data:
It is not possible to represent the interaction.
3. Choosing the model:
Multiple models can be chosen by using Statistics Fit models Linear
model... and then writing the model. If multiple models have been con-
structed, it is possible to test one against another by using Model Hypothesis
tests Compare two models...
By default, the outputs are estimates of parameters , i , j and ij with
reference cell constraints. To change this constraint to a sum constraint, use
Tools Options... then in contrasts, replace contr.Treatment and
contr.poly by contr.sum. Then click on Exit and Restart R Commander.
To draw up the analysis of variance table:
Models Hypothesis tests ANOVA table
4. Interpreting the coefficients:
See previous item.
172 R for Statistics

8.2.6 Taking Things Further

During the construction of the model, we assume that the residuals follow
a normal distribution and are homoscedastic. It can also be interesting to
test the equality of variance of residuals for the different combinations of the
interaction. In order to do so, we can construct a Bartlett test using the
bartlett.test function. If we reject the equality of variance, we can construct
a Friedman test, which is based on ranks, using the friedman.test function.

Theoretical presentations and exercises on analysis of variance are available

in many books, such as Clarke and Cooke (2004), Faraway (2005), Moore et al.
(2007) or Sahai and Ageel (2000).
Analysis of Covariance 173

8.3 Analysis of Covariance

8.3.1 Objective
The analysis of covariance aims to model the relationship between a quanti-
tative response variable Y and quantitative and qualitative explanatory vari-
ables. This worked example is dedicated to examining the simplest case: that
is, a case in which there are only two explanatory variables, one of which is
quantitative and the other qualitative. The quantitative variable is denoted
X and the qualitative variable Z, which is considered to have I categories. As
the relationship between the response variable Y and the explanatory vari-
able X can depend on the categories of the qualitative variable Z, the natural
procedure involves conducting I different regressions, one for each level i of
Z. In terms of modelling, this yields:

y1,j = 1 + 1 x1,j + 1,j j = 1, . . . , n1 level 1 of Z

y2,j = 2 + 2 x2,j + 2,j j = 1, . . . , n2 level 2 of Z
.. ..
. .
yI,j = I + I xI,j + I,j j = 1, . . . , nI final level of Z

Put more simply

yi,j = i + i xi,j + i,j i = 1, . . . , I, j = 1, . . . , ni (8.3)

In this model the 2I parameters (i and i ) are calculated by the least

squares method by conducting a simple regression for each level of Z. In terms
of error, one typically considers that all the i,j have the same variance 2 ,
an assumption made both in regression and analysis of variance.

This model can also be written by identifying an overall mean and a

mean slope (as in analysis of variance):

yi,j = + i + ( + i )xi,j + i,j i = 1, . . . , I, j = 1, . . . , ni . (8.4)

Constraints must thus be added in order to obtain

P an identifiable
P model. As
in analysis of variance, we classically impose i i = 0 and i i = 0, or
indeed one i = 0 and one i = 0.

To know whether Z has an effect on Y , we need to test the equality of the

intercepts and that of the slopes between different models of linear regression.
A number of different cases could be considered (see Figure 8.5).
174 R for Statistics

Figure 8.5
Illustration of the models (8.4), (8.5) and (8.6).

The more general model (8.4), which we refer to as the complete model,
is illustrated on the left in Figure 8.5. In this case, the slopes and intercepts
are assumed to be different for each level of Z.
A first simplification of this model is to suppose that variable X intervenes
in the same manner, whatever the level of variable Z. It amounts to assuming
that the slopes are identical, but that the intercepts are not. This is the model
illustrated in the centre of Figure 8.5, which is written as

yi,j = i + xi,j + i,j i = 1, . . . , I, j = 1, . . . , ni (8.5)

It must be noted that here there is no interaction between X and Z as the

slopes are identical.
Another simplification of the complete model (8.4) is to assume that in
the categories of Z, the intercepts are the same but the slopes are not (see
right-hand graph in Figure 8.5). The model is therefore written as

yi,j = + i xi,j + i,j i = 1, . . . , I, j = 1, . . . , ni (8.6)

The choice of model (8.4), (8.5) or (8.6) depends on the problem addressed.
We advise starting with the most general model (8.4) and then, if the slopes
are the same, we turn to model (8.5); if the intercepts are the same, we turn
to model (8.6). As the models are nested, it is possible to test one model
against another.
Once the model is chosen, the residuals must be analysed. This analysis
is essential as it is used to check the individual fit of the model (outliers) and
the global fit, for example by checking that there is no structure.

8.3.2 Example
We are interested in the balance of flavours between different ciders and, more
precisely, the relationship between the sweet and bitter flavours according
to the type of cider (dry, semi-dry or sweet). The flavour evaluations were
provided by a tasting panel (mean marks, from 1 to 10, by 24 panel members)
for each of 50 dry ciders, 30 semi-dry ciders and 10 sweet ciders.
Analysis of Covariance 175

8.3.3 Steps
1. Read the data.
2. Represent the data.

3. Choose the model.

4. Conduct a residual analysis.

8.3.4 Processing the Example

1. Reading the data:
Import the data, check the number of samples and the number of variables in
the dataset, and summarise the variables of interest, here Type, Sweetness
and Bitterness:

> cider <- read.table("cider.csv",header=TRUE,sep=",")

> dim(cider)
[1] 90 5
> summary(cider[,c(1,2,4)])
Type Sweetness Bitterness
Dry :50 Min. :3.444 Min. :2.143
Semi-dry:30 1st Qu.:4.580 1st Qu.:3.286
Sweet :10 Median :5.250 Median :3.964
Mean :5.169 Mean :4.274
3rd Qu.:5.670 3rd Qu.:5.268
Max. :7.036 Max. :7.857

2. Representing the data:

Prior to an analysis of covariance, it may be useful to explore the data by rep-
resenting, for each level of the qualitative variable, the scatterplot confronting
the response variable Sweetness and the quantitative explanatory variable
Bitterness. It is possible to represent these scatterplots on the same graph
using different symbols (Figure 8.6). To do so, we use the argument pch and
convert the variable Type as numeric to define pch values from 1 to the num-
ber of levels according to the values of Type. The argument col can be used
exactly in the same way if we want to use a specific colour for each cider.
Then, we add a legend on the top right of the graph to specify the symbol
used for each type of cider (Figure 8.6).

> plot(Sweetness~Bitterness,pch=as.numeric(Type),data=cider)
> legend("topright",levels(cider$Type),pch=1:nlevels(cider$Type))
176 R for Statistics

3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0

Dry
Semidry
Sweet

Sweetness

2 3 4 5 6 7 8

Bitterness

Figure 8.6
Representation of the scatterplot.

It is also possible to represent the data using the xyplot function from the
lattice package (see Figure 8.7):

> library(lattice)
> xyplot(Sweetness~Bitterness|Type,data=cider)

To have a more precise idea of the effect of the qualitative variable, we could
represent the regression line for each category but it is difficult to get an
idea of the most appropriate model simply by looking at these graphs. It is
therefore preferable to analyse the models.
3. Choosing the model:
The complete model defined by (8.4) is adjusted using the lm (linear model)
function:

> global <- lm(Sweetness~-1+Type+Type:Bitterness, data=cider)

We remove the intercept by writing 1 but it must be specified that it is

necessary to calculate an intercept for each type of cider by adding the factor
Type. We must then specify that we require a different slope for each type of
cider by writing the interaction of the variables Sweetness and Type.
The one-slope model (8.5) can be written as

> slopeU <- lm(Sweetness~-1 + Type + Bitterness, data = cider)

and a model with only one intercept (8.6)

> interceptU <- lm(Sweetness ~ Type:Bitterness, data = cider)

Analysis of Covariance 177

Sweet
7

4
Sweetness

Dry Semidry
7

2 3 4 5 6 7 8

Bitterness

Figure 8.7
Scatterplots according to the categories of the variable Type.

In order to choose the best model, we can conduct multiple tests:

(a) slope or intercept equality: We conduct a test between the models
(8.5) and (8.4) and another between the models (8.6) and (8.4) using the
following commands:
> anova(slopeU,global)
Analysis of Variance Table

Model 1: Sweetness ~ -1 + Type + Bitterness

Model 2: Sweetness ~ -1 + Type + Type:Bitterness
Res.Df RSS Df Sum of Sq F Pr(>F)
1 86 24.019
2 84 23.118 2 0.90126 1.6374 0.2006

> anova(interceptU,global)
178 R for Statistics

Analysis of Variance Table

Model 1: Sweetness ~ Type:Bitterness

Model 2: Sweetness ~ -1 + Type + Type:Bitterness
Res.Df RSS Df Sum of Sq F Pr(>F)
1 86 25.861
2 84 23.118 2 2.7428 4.983 0.009015 **
As the p-value (0.20) of the first test of nested models is greater than 5%, we
can conclude that the slopes can be considered equal. As the p-value of the
second test of nested models is less than 5%, we conclude that the intercepts
are not equal. We therefore conserve the model without the interaction, which
corresponds to model (8.5). This means that whatever the type of cider, the
effect of the perception of bitterness and the perception of sweetness is the
same. Now that we are in the (8.5) configuration, we will test whether or not
it is possible to further simplify the model.
(b) null slope or intercept equality: We conduct a test between the model
(8.5) and the simple regression model (see Worked Example 7.1): This test will
reveal whether it is preferable to conserve the model with different intercepts,
or if a simple linear regression model between the two quantitative variables
is sufficient. We will then carry out a test between the model (8.5) and the
one-way analysis of variance (see Worked Example 8.1) to know whether the
latter is sufficient.
> simple <- lm(Sweetness ~ Bitterness,data = cider)
> anova(simple,slopeU)
Analysis of Variance Table

Model 1: Sweetness ~ Bitterness

Model 2: Sweetness ~ -1 + Type + Bitterness
Res.Df RSS Df Sum of Sq F Pr(>F)
1 88 30.945
2 86 24.019 2 6.9256 12.399 1.857e-05 ***

> anova1 <- lm(Sweetness ~ Type,data = cider)

> anova(anova1,slopeU)
Analysis of Variance Table

Model 1: Sweetness ~ Type

Model 2: Sweetness ~ -1 + Type + Bitterness
Res.Df RSS Df Sum of Sq F Pr(>F)
1 87 38.571
2 86 24.019 1 14.552 52.102 1.960e-10 ***
The two simplified models are rejected and the (8.5) model retained. The
perception of bitterness thus has an influence on the perception of sweetness
Analysis of Covariance 179

and this influence is the same for all types of cider (the slope is the same but
the intercepts are different).

4. Conducting a residual analysis:

The residual analysis is similar to that conducted in Worked Example 7.1.
Depending on the model chosen, we analyse the residuals using

> xyplot(rstudent(chosen_model)~Bitterness|Type, ylab="res",

data=cider)

8.3.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
Next specify that the field separator is a comma.
To check that the dataset has been imported successfully:
Statistics Summaries Active data set

2. Representing the data on a graph:

Graphs Scatterplot... Then select variables x and y. Then click on Plot
by groups and choose the qualitative variable. This automatically produces
a graph with a default colour for each level of the qualitative variable and the
corresponding regression lines.

3. Choosing the model:

Multiple models can be chosen by using Statistics Fit models Linear
model... and then writing the model. If multiple models have been con-
structed, it is possible to test one against another using Models Hypothesis
tests Compare two models...

4. Conducting a residual analysis:

Models Add observation statistics to data...
A number of graphs are made directly available. To construct them, use
Models Graphs Basic diagnostic plots
It is not possible to make the graph offered by lattice.

8.3.6 Taking Things Further

It is possible to construct an analysis of covariance model with all possible ef-
fects, and to test all the effects by decomposing the variability. This approach
is easier when there are multiple qualitative and/or quantitative variables.
The car package must be used in order to obtain a type III sum of squares
decomposition of variation (in this way the same sum squares are obtained
for any order of the effects in the model). The Anova function (with a capital
A) is used to obtain the analysis of variance table. We therefore write
180 R for Statistics

> library(car)
> global <- lm(Sweetness~Type+Bitterness+Type:Bitterness,
data=cider)
> Anova(global, type="III")
Anova Table (Type III tests)

Response: Sweetness
Sum Sq Df F value Pr(>F)
(Intercept) 144.139 1 523.7375 < 2.2e-16 ***
Type 2.743 2 4.9830 0.009015 **
Bitterness 6.879 1 24.9968 3.098e-06 ***
Type:Bitterness 0.901 2 1.6374 0.200634
Residuals 23.118 84

From these results, the effects with the highest p-value, and those over 5%
can be eliminated.
Once the model has been chosen, it is possible to estimate the coefficients
of the model using the summary function:
> global <- lm(Sweetness~Type+Bitterness,data=cider)
> summary(global)
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 6.30248 0.21230 29.687 < 2e-16 ***
TypeSemi-dry 0.41723 0.12293 3.394 0.00104 **
TypeSweet 0.83126 0.18797 4.422 2.84e-05 ***
Bitterness -0.31928 0.04423 -7.218 1.96e-10 ***
The coefficients are estimated using the constraint that the first coefficient
(that
P of the category Dry) is equal to 0. If we want to use the constraint
i i = 0, we use C(Type,sum) for the analysis of variance.

Theoretical presentations and exercises on analysis of covariance are avail-

able in many books, such as Faraway (2005).
9
Classification

9.1 Linear Discriminant Analysis

9.1.1 Objective
The aim of linear discriminant analysis is to explain and predict the values
of a qualitative variable Y from quantitative and/or qualitative explanatory
variables X = (X1 , . . . , Xp ). We here suppose that the variable Y is binary,
and the generalization to more than two categories is tackled towards the end
of the worked example.
Linear discriminant analysis (LDA) can be presented according to two
very different but equivalent aspects. Firstly, one point of view consists of
modelling the probability of belonging to a group using Bayes formula. To
fix the notations, the categories of the response variable Y are designated 0
and 1. Bayes rule estimates the posterior probability of allocation:

P(Y = 1|X = x) =
1 f1 (x)
0 f0 (x) + 1 f1 (x)
(9.1)

where f0 (x) and f1 (x) stand for the densities of X knowing that Y = 0 and
Y = 1, and 0 = P(Y = 0) and 1 = P(Y = 1) represent the prior probabili-
ties of belonging to classes 0 and 1, respectively. These two probabilities must
be fixed by the user or estimated from the sample. In order to calculate the
posterior probabilities P(Y = 0|X = x) and P(Y = 1|X = x), the discrimi-
nant analysis models the distributions of X knowing that Y = j (j = 0, 1) by
normal distributions. More specifically, we make the assumption

X|Y = 0 N (0 , ), X|Y = 1 N (1 , ).

The parameters 0 , 1 and are estimated by the method of maximum like-

lihood. For a new individual, we deduce from (9.1) the posterior probabilities
and allocate this new individual to the group with the greatest one.
LDA can also be seen as a dimensionality reduction technique. In this
case, as in principal component analysis, the principle consists of calculating
a new variable, known as the canonical variable w0 X = w1 X1 + . . . + wp Xp ,
as a linear combination of the initial variables. This variable is calculated in
such a way as to maximise the ratio of the between-group and within-group
182 R for Statistics

variances (for further details, see McLachlan (2004), Chapter 18). For an
individual x = (x1 , . . . , xp ), the canonical variable defines a score function
S(x) = w1 x1 + . . . + wp xp . An individual x is allocated to a given group by
comparing the value of the score S(x) with a threshold value s.
Remark
The two ways of presenting LDA do not allow to handle qualitative explana-
tory variables. However, as we shall see in the example, a disjunctive encoding
of the explanatory variables can be used to conduct a linear discriminant anal-
ysis in the presence of such variables. Each category of the variable is there-
fore processed as a quantitative variable with values 0 or 1. In order to avoid
problems of collinearity between the different categories, the column which
is associated with the factors first category is not taken into account in the
model. When a discriminant analysis is conducted with one or more qualita-
tive explanatory variables, the assumption of normality is clearly not verified:
It is therefore important to be wary when interpreting posterior probabilities.

9.1.2 Example
We would like to explain the propensity to snore from six explanatory vari-
ables. The file snore.txt contains a sample of 100 patients. The variables
considered in the sample are
age: of the individual in years
weight: of the individual in kilograms (kg)
size: height of the individual in centimeters (cm)
alcohol: number of units drunk per day (each equal to one glass of red
wine)
sex: of the individual (W = woman, M = male)
snore: propensity to snore (Y = snores, N = does not snore)
tobacco: smoking behaviour (Y = smoker, N = non-smoker)
The aim of this study is to try to explain snoring (variable snore) using the
six other variables presented above. The following table presents an extract
from the dataset:
age weight size alcohol sex snore tobacco
1 47 71 158 0 M N Y
2 56 58 164 7 M Y N
.
. .
. .
. .
. .
. .
. .
. .
.
. . . . . . . .
99 68 108 194 0 W Y N
100 50 109 195 8 M Y Y
Linear Discriminant Analysis 183

9.1.3 Steps
1. Read the data.
2. Construct the model.
3. Estimate the classification error rate.
4. Make a prediction.

9.1.4 Processing the Example

1. Reading the data:

> snore_data <- read.table("snore.txt",header=T)

Summarise the dataset:

> summary(snore_data)
age weight size alcohol
Min. :23.00 Min. : 42.00 Min. :158.0 Min. : 0.00
1st Qu.:43.00 1st Qu.: 77.00 1st Qu.:166.0 1st Qu.: 0.00
Median :52.00 Median : 95.00 Median :186.0 Median : 2.00
Mean :52.27 Mean : 90.41 Mean :181.1 Mean : 2.95
3rd Qu.:62.25 3rd Qu.:107.00 3rd Qu.:194.0 3rd Qu.: 4.25
Max. :74.00 Max. :120.00 Max. :208.0 Max. :15.00

sex snore tobacco

M:75 N:65 N:36
W:25 Y:35 Y:64

2. Constructing the model:

The lda function from the MASS package is used to conduct the LDA. The
MASS package is installed by default in R, and simply needs to be loaded. The
lda function uses a formula like the function lm (see Appendix A.2, p. 266):

Y X1 + X2 .

The prior probabilities also need to be specified. If there are no prior assump-
tions about these two quantities, there are two possible strategies:
The probabilities are chosen to be equal, that is, 0 = 1 = 0.5;
The probabilities are estimated by the proportion of observations in each
group.
The choice of prior probabilities is specified by the prior argument of the lda
function. If nothing is specified, the second strategy is chosen by default. We
first write the model which takes all the explanatory variables into account:
184 R for Statistics

> library(MASS)
> model <- lda(snore~.,data=snore_data)
> model

Call:
lda(snore ~ ., data = snore_data)

Prior probabilities of groups:

N Y
0.65 0.35

Group means:
age weight size alcohol sexW tobaccoY
N 50.26154 90.47692 180.9538 2.369231 0.3076923 0.6769231
Y 56.00000 90.28571 181.3714 4.028571 0.1428571 0.5714286

Coefficients of linear discriminants:

LD1
age 0.05973655
weight -0.01620579
size 0.01590170
alcohol 0.24058822
sexW -0.55413371
tobaccoY -1.14621434

In the output, we find the prior probabilities of the model, the centre of gravity
for each of the two groups, and the coefficients of the canonical variable. As a
rule, the software returns the coefficients of the canonical variable so that the
within-class variance equals 1. It is difficult to find the important variables for
the discrimination simply by comparing the centres of gravity; it is often more
relevant to study the influence of the coefficients on the score. For example,
it can be seen that, when all the other variables are equal, a man will have a
score 0.5541 higher than a woman. In the same way, a ten-year age difference
between two individuals who otherwise share the same characteristics will
contribute to an evolution in the score of 0.5974.
No statistical test can be used to test the significance of the coefficients of the
canonical variable. Nevertheless here we can see that the variables weight
and size are close to 0. It may therefore be interesting to delete these two
variables in the model as follows:

> model1 <- lda(snore~age+alcohol+sex+tobacco,data=snore_data)

> model1
Call:
lda(snore ~ age + alcohol + sex + tobacco, data = snore_data)
Linear Discriminant Analysis 185

Prior probabilities of groups:

N Y
0.65 0.35

Group means:
age alcohol sexW tobaccoY
N 50.26154 2.369231 0.3076923 0.6769231
Y 56.00000 4.028571 0.1428571 0.5714286

Coefficients of linear discriminants:

LD1
age 0.06062024
alcohol 0.24060629
sexW -0.54047702
tobaccoY -1.13039555

It can be seen that the coefficients of the canonical variable are only slightly
affected by the removal of these two variables.
To impose equal prior probabilities on the model, we use the prior argument:

> model2 <- lda(snore~.,data=snore_data,prior=c(0.5,0.5))

> model3 <- lda(snore~age+alcohol+sex+tobacco,data=snore_data,
prior=c(0.5,0.5))
> model2
Call:
lda(snore ~ ., data = snore_data, prior = c(0.5, 0.5))

Prior probabilities of groups:

N Y
0.5 0.5

Group means:
age weight size alcohol sexW tobaccoY
N 50.26154 90.47692 180.9538 2.369231 0.3076923 0.6769231
Y 56.00000 90.28571 181.3714 4.028571 0.1428571 0.5714286

Coefficients of linear discriminants:

LD1
age 0.05973655
weight -0.01620579
size 0.01590170
alcohol 0.24058822
sexW -0.55413371
tobaccoY -1.14621434

The modification of prior probabilities does not change the canonical variable.
186 R for Statistics

This choice only influences the way new individuals are allocated to a given
group.

3. Estimating the classification error rate:

The lda function is used to estimate the classification error rate by cross
validation. In order to do so, simply add the argument CV=TRUE to the function
call. We obtain the labels predicted by the model named model using the
following command:

> pred <- lda(snore~.,data=snore_data,CV=TRUE)$class

> table(pred,snore_data$snore)

pred N Y
N 53 22
Y 12 13

We thus obtain an estimation of the classification error rate:

> sum(pred!=snore_data$snore)/nrow(snore_data)
[1] 0.34

The classification error rate is estimated in the same way for the other models
studied:

> pred1 <- lda(snore~age+alcohol+sex+tobacco,data=snore_data,

CV=TRUE)$class
> table(pred1,snore_data$snore)
pred1 N Y
N 52 23
Y 13 12
> sum(pred1!=snore_data$snore)/nrow(snore_data)
[1] 0.36

> pred2 <- lda(snore~.,data=snore_data,prior=c(0.5,0.5),

CV=TRUE)$class
> table(pred2,snore_data$snore)
pred2 N Y
N 41 10
Y 24 25
> sum(pred2!=snore_data$snore)/nrow(snore_data)
[1] 0.34

> pred3 <- lda(snore~age+alcohol+sex+tobacco,data=snore_data,

prior=c(0.5,0.5),CV=TRUE)$class
> table(pred3,snore_data$snore)
Linear Discriminant Analysis 187

pred3 N Y
N 42 9
Y 23 26
> sum(pred3!=snore_data$snore)/nrow(snore_data)
[1] 0.32

The last model is the most effective in terms of misclassification. Furthermore,

we can see that the last two models (those for which the prior probabilities are
equal to 0.5) are also able to identify individuals who snore as they detect 25
and 26 out of 35, compared with 13 and 12 out of 35 for the first two models.
4. Making a prediction:
The models constructed previously can be used for prediction. The table be-
low contains the values for six explanatory variables measured on four new
patients:

Age Weight Size Alcohol Sex Tobacco

42 55 169 0 W N
58 94 185 4 M Y
35 70 180 6 M Y
67 63 166 3 W N

In order to predict the labels of these four new individuals, we first collect
the new data in a data-frame with the same structure as the initial data table
(and in particular the same names for variables). We thus construct a new
matrix for the quantitative variables and another for the qualitative variables
before grouping them together in the same data-frame:

> new_d1 <- matrix(c(42,55,169,0,58,94,185,4,35,70,180,6,67,63,

166,3),ncol=4,byrow=T)
> new_d2 <- matrix(c("W","N","M","Y","M","Y","W","N"),ncol=2,
byrow=T)
> new_d <- cbind.data.frame(new_d1,new_d2)
> names(new_d) <- names(snore_data)[-6]

The predict function can be used to allocate each of these four individuals a
group:

> predict(model,newdata=new_d)
$class
[1] N N N Y
Levels: N Y
188 R for Statistics

$posterior
N Y
1 0.7957279 0.2042721
2 0.6095636 0.3904364
3 0.7325932 0.2674068
4 0.3532721 0.6467279

$x
LD1
1 -0.6238163
2 0.3246422
3 -0.2586916
4 1.4140108

The predict function returns a list of three objects:

The first element (class) contains the predicted groups.
The second element (posterior) is a two-column matrix featuring the
posterior possibilities of belonging to groups N and Y for each individual.
The third element (x) designates the score value for each individual.
The model allocates the first three individuals to group N and the fourth
to Y. The rule for allocation to a group with an observation x is conducted
by comparing the posterior probabilities or, in the same way, comparing the
score with a threshold value. Modifying the prior probabilities does not change
the canonical variable; the only thing that will change is the threshold from
which we change the allocation to a group. Since the software does not return
this threshold value, the allocations are made by comparing the posterior
probabilities for each group.

9.1.5 Rcmdr Corner

It is not possible to conduct a linear discriminant analysis with Rcmdr.

9.1.6 Taking Things Further

To conduct an LDA, we assume that the variance-covariance matrices of the
explanatory variables are the same for each group. This hypothesis can be
avoided thanks to quadratic discriminant analysis. Posterior probabilities are
calculated supposing that

X|Y = 0 N (0 , 0 ), X|Y = 1 N (1 , 1 )

The qda function, available in the MASS package, can be used to conduct a
quadratic discriminant analysis in R. It is used in the same way as lda.
Linear Discriminant Analysis 189

In this chapter, we limited our analysis to a binary response variable.

LDA can easily be adapted to cases where the response variable carries J > 2
categories. The posterior probabilities are still calculated using Bayes formula
by assuming normality for each group X|Y = j N (j , ) for j = 1, . . . , J
(or X|Y = j N (j , j ) for the quadratic discriminant). In this case, we are
no longer looking for one canonical variable, but rather (J 1) variables. The
calculations can be implemented in R with no difficulty and are conducted in
the same way: the models are written in the same way and individuals are
allocated to the group with the highest posterior probability.
For further theoretical elements on linear discriminant analysis, refer to
McLachlan (2004).
190 R for Statistics

9.2 Logistic Regression

9.2.1 Objective
The aim of logistic regression is to explain and predict the values of a quali-
tative variable Y , which is generally binary, from qualitative and quantitative
explanatory variables X = (X1 , . . . , Xp ). If the categories of Y are denoted 0
and 1, the logistic model is expressed as
 
p(x)
log = 0 + 1 x 1 + . . . + p x p
1 p(x)

where p(x) denotes the probability P(Y = 1|X = x) and x = (x1 , . . . , xp ) is a

realization of X = (X1 , . . . , Xp ). The coefficients 0 , 1 , . . . , p are estimated
by the method of maximum likelihood.

9.2.2 Example
Treatment for prostate cancer changes depending on whether or not the lym-
phatic nodes surrounding the prostate are affected. In order to avoid an inva-
sive investigative procedure (opening the abdominal cavity), a certain number
of variables are considered explanatory variables for the binary variable Y : if
Y = 0, the cancer has not reached the lymphatic system; if Y = 1, the can-
cer has reached the lymphatic system. The aim of this study is therefore to
explain and predict Y from the following variables:
Age of the patient at the time of diagnosis (age)
Serum acid phosphatase level (acid)
Result of an X-ray analysis, 0 = negative, 1 = positive (Xray)
Size of the tumour, 0 = small, 1 = large (size)
State of the tumour as determined by a biopsy, 0 = medium, 1 = serious
(grade)
Logarithm of the acidity level (log.acid)
The following table presents an extract from the dataset:
age acid Xray size grade Y log.acid
1 66 0.48 0 0 0 0 -0.7339692
2 68 0.56 0 0 0 0 -0.5798185
.
. .
. .
. .
. .
. .
. .
. .
.
. . . . . . . .
52 64 0.89 1 1 0 1 -0.1165338
53 68 1.26 1 1 1 1 0.2311117
Logistic Regression 191

9.2.3 Steps
1. Read the data.
2. Construct the model.
3. Choose the model.
4. Make a prediction.

9.2.4 Processing the Example

1. Reading the data:
Reading the data from the file prostate.txt:

> prostate <- read.table("prostate.txt",header=T)

The variables Xray, size, grade and Y must be converted into factors:

> for (i in 3:6){prostate[,i] <- factor(prostate[,i])}

Summarise the dataset:

> summary(prostate)
age acid Xray size grade Y log.acid
Min. :45.00 Min. :0.4000 0:38 0:26 0:33 0:33 Min. :-0.9163
1st Qu.:56.00 1st Qu.:0.5000 1:15 1:27 1:20 1:20 1st Qu.:-0.6931
Median :60.00 Median :0.6500 Median :-0.4308
Mean :59.38 Mean :0.6942 Mean :-0.4189
3rd Qu.:65.00 3rd Qu.:0.7800 3rd Qu.:-0.2485
Max. :68.00 Max. :1.8700 Max. : 0.6259

2. Constructing the model:

In order to better explain the idea of logistic regression, we first consider the
elementary logistic model used to explain variable Y by variable log.acid.
Let us denote X the variable log.acid and x an occurrence of this variable.
In this case, the logistic model is simply written as
 
p(x)
log = 0 + 1 x,
1 p(x)
where p(x) designates the probability P(Y = 1|X = x).
Logistic regression belongs to the family of generalised linear models. An
adjustment of these models in R is conducted using the glm function. Use of
this function is similar to that of the lm function. A model must be written
with a formula of the type (see Appendix A.2, p. 266):

Y X1 + X2 .
192 R for Statistics

A family of probability distributions must also be chosen. Within the context

of logistic regression, we are interested in binomial families:

> model_log.acid<-glm(Y~log.acid,data=prostate,family=binomial)
> summary(model_log.acid)
Call:
glm(formula = Y ~ log.acid, family = binomial, data = prostate)

Deviance Residuals:
Min 1Q Median 3Q Max
-1.9802 -0.9095 -0.7265 1.1951 1.7302

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.404 0.509 0.794 0.4274
log.acid 2.245 1.040 2.159 0.0309 *
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 70.252 on 52 degrees of freedom

Residual deviance: 64.813 on 51 degrees of freedom
AIC: 68.813

Number of Fisher Scoring iterations: 4

The output yields a Coefficients matrix with four columns for each parame-
ter (row): estimations of parameters (column Estimate), estimated standard
deviations of the parameters (Std. Error), observed values of the test statis-
tic (z value), and p-values (Pr(>|z|)), which are the probability that, under
the null hypothesis, the test statistics will exceed the estimated values of the
parameters.
The coefficients 0 and 1 are estimated by the method of maximum likeli-
hood. Here 0 = 0.404 and 1 = 2.245; we therefore obtain the model:
 
p(x)
log = 0.404 + 2.245x
1 p(x)

or equivalently,

exp(0.404 + 2.245x)
p(x) =
1 + exp(0.404 + 2.245x)

It is possible to test the significance of the models coefficients and to ob-

tain confidence intervals for the same coefficients. For example, the p-value
Logistic Regression 193

associated with the Wald test for hypotheses

H0 : 1 = 0 against H1 : 1 6= 0

is 0.0309. The null hypothesis will therefore be rejected in favour of the

alternative hypothesis with a 5% risk: the variable log.acid is retained in
the model.
Figure 9.1 presents the estimated values of p(x) obtained with the model
model log.acid:

> beta <- coef(model_log.acid)

> x <- seq(-2,2,by=0.01)
> y <- exp(beta[1]+beta[2]*x)/(1+exp(beta[1]+beta[2]*x))
> plot(x,y,type="l",xlab="log.acid",ylab="p(x)")
> abline(h=0.5,lty=2)
> xlim <- -beta[1]/beta[2]
> abline(v=xlim,lty=2)
0.0 0.2 0.4 0.6 0.8 1.0
p(x)

2 1 0 1 2

log.acid

Figure 9.1
Estimated values of p(x) for the model model log.acid.

Note that

0.5 if x 0.18
p(x) = P(Y
b = 1|X = x)
> 0.5 if x > 0.18

For the values of x exceeding 0.18, the model will thus predict the value
Y = 1 (cancer has reached the lymphatic system) whereas for values less than
or equal to 0.18, it will predict Y = 0 (the cancer has not yet reached the
lymphatic system).
It is of course possible to consider other logistic models. For example, the
model with only the variable size as explanatory variable:

> model_size <- glm(Y~size,data=prostate,family=binomial)

194 R for Statistics

> model_size

Call: glm(formula = Y ~ size, family = binomial,

data = prostate)

Coefficients:
(Intercept) size1
-1.435 1.658

Degrees of Freedom: 52 Total (i.e. Null); 51 Residual

Null Deviance: 70.25
Residual Deviance: 62.55 AIC: 66.55

In this case, only the coefficient associated with the level 1 of the variable size
is estimated (the coefficient associated with the level 0 is by default taken to
be zero). We obtain
  
p(x) 1.435 if x = 0
log =
1 p(x) 1.435 + 1.658 = 0.223 if x = 1.

It can also be interesting to consider models comprising one or more inter-

actions, such as that made up of the variables age, grade as well as the
interaction age:grade:

> model_inter <- glm(Y~age+grade+age:grade,data=prostate,

family=binomial)

The model which includes all the explanatory variables can also be studied:

> model <- glm(Y~.,data=prostate,family=binomial)

> model

Call: glm(formula = Y ~ ., family = binomial, data = prostate)

Coefficients:
(Intercept) age acid Xray1 size1
10.08672 -0.04289 -8.48006 2.06673 1.38415
grade1 log.acid
0.85376 9.60912

Degrees of Freedom: 52 Total (i.e. Null); 46 Residual

Null Deviance: 70.25
Residual Deviance: 44.77 AIC: 58.77
Logistic Regression 195

3. Choosing the model:

Several models have already been estimated. Other models would of course
also be possible. The anova function can be used to compare two nested
models using the deviance statistic to test the nullity of the coefficients of the
bigger model. For example, it is possible to compare the models model and
model log.acid using the command

> anova(model_log.acid,model,test="Chisq")
Analysis of Deviance Table

Model 1: Y ~ log.acid
Model 2: Y ~ age + acid + Xray + size + grade + log.acid
Resid. Df Resid. Dev Df Deviance P(>|Chi|)
1 51 64.813
2 46 44.768 5 20.044 0.001226 **
---
Signif. codes: 0 *** 0.001 ** 0.01 * 0.05 . 0.1 1

As the p-value of the test is 0.001226, the null hypothesis is rejected. Thus
the model model log.acid is rejected and at least one of the supplementary
variables of model is considered to be relevant.
The anova function was used here to compare two models. Besides, the step
function is used to choose a model with the help of a step-by-step procedure
based on minimising the AIC value (Akaike Information Criterion):

> model_step <- step(model,direction="backward")

> model_step

Call: glm(formula = Y ~ acid + Xray + size + log.acid,

family = binomial,data = prostate)

Coefficients:
(Intercept) acid Xray1 size1 log.acid
9.067 -9.862 2.093 1.591 10.410

Degrees of Freedom: 52 Total (i.e. Null); 48 Residual

Null Deviance: 70.25
Residual Deviance: 46.43 AIC: 56.43

The step-by-step selection procedure conducted by the step function is here

a backward procedure: at each step the variable for which withdrawal from
the model leads to the greatest decrease of the AIC value is removed. The
direction of the step-by-step procedure is specified by the argument direction
(here direction="backward"). The process stops when all the variables are
removed or when removing a variable no longer decreases the criterion. In
196 R for Statistics

this example, the variables age and grade have been removed. In order to
use this backward method, it must first be ensured that the model used in
the argument for the step function contains a sufficient number of variables.
We could have used a more general model involving certain interactions as
the initial model for the procedure. The step function can also be used to
conduct forward or progressive (simultaneously forward and backward) step-
by-step procedures.

4. Making a prediction:
The logistic models constructed previously can be used for prediction. The
table below contains measurements from four new patients for the six explana-
tory variables:

Age Acid Xray Size Grade log.acid

61 0.60 1 0 1 0.51
49 0.86 0 0 1 0.15
67 0.72 1 0 1 0.33
51 0.95 1 1 1 0.05

To obtain the values p(x) = P(Y = 1|X = x) predicted by the model

model step for these four new individuals, we first bring together the new
data in a data-frame with the same structure as the initial data table (and
especially the same names for the variables):

> new_d <- data.frame(matrix(c(61,0.60,1,0,1,-0.51,49,0.86,0,0,1,

-0.15,67,0.72,1,0,1,-0.33,51,0.95,1,1,1,-0.05),ncol=6,byrow=T))
> names(new_d) <- names(prostate)[-6]

Remember, here the variables Xray, size, and grade must be converted into
factors:

> for (i in 3:5){new_d[,i] <- factor(new_d[,i])}

The predict function is used to predict the probabilities P(Y = 1|X = x) for
each of these new individuals:

> prediction <- predict(model_step,newdata=new_d,type="response")

> prediction
1 2 3 4
0.4835626 0.2737392 0.6512557 0.9459345

For the first two individuals, the predicted probabilities are less than 0.5.
We therefore predict Y = 0, that is to say, the cancer has not reached the
lymphatic system, whereas for the last two individuals we predict that the
cancer has indeed reached the lymphatic system.
Logistic Regression 197

The predict function can also be used to estimate the models misclassification
rate (proportion of errors made by the model when applied to the available
sample):
(a) First calculate the probabilities predicted for each individual from the
sample:
> prediction_prob <- predict(model_step,newdata=prostate,
type="response")
(b) Compare these probabilities to 0.5 in order to obtain the predicted label
for each individual:
> prediction_label <- as.numeric(prediction_prob>0.5)
(c) Draw up the contingency table with the predicted labels and the true
labels:
> table(prostate$Y,prediction_label)
prediction_label
0 1
0 28 5
1 6 14
We thus obtain the models misclassification rate:
> error <- sum(prediction_label!=prostate$Y)/nrow(prostate)
> error
[1] 0.2075472
The rate obtained in this way is generally optimistic as the same sample is
used to construct the model and to estimate the misclassification rate. We
can obtain a more precise estimation using cross-validation methods. In order
to do so, we use the function cv.glm. First of all, the package boot must be
loaded and a cost function created which admits the observed values of Y as
well as the predicted probabilities as input:

> library(boot)
> cost <- function(Y_obs,prediction_prob)
+ return(mean(abs(Y_obs-prediction_prob)>0.5))
> cv.glm(prostate,model_step,cost)$delta[1]
1
0.2830189

9.2.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL. The col-
umn separator (field separator) must then be specified.
198 R for Statistics

Remember, here the variables Xray, size, grade and Y must be converted
into factors: Data Manage variables in active data set Convert
numeric variables to factors....
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
2. Constructing the model:
Statistics Fit models Generalized linear model....
To write the model:
(a) Double-click on the response variable.
(b) Next choose the explanatory variables, separating them with a + and
a : when an interaction is required.

3. Choosing the model:

Models Stepwise model selection
The menu Models can also be used to obtain confidence intervals for the
parameters of the model, to conduct tests between nested models, or even to
represent residuals.

9.2.6 Taking Things Further

The residuals function can be used to calculate the different types of residuals
coming from a logistic regression (Pearson or deviance residuals, for example).
We might also use the function logLik to obtain the log-likelihood of a model
or to calculate the value of the statistic for a likelihood ratio test. Should
there be a response variable with more than two categories, two cases must
be distinguished. When the categories are ordinal, we build an ordinal poly-
tomous model using the polr function from the package MASS. When there
is no ordinal relation between the categories of Y , we turn to a nominal poly-
tomous model, also known as a multinomial model. In such cases we must
use the multinom function from the package nnet. We can also call upon the
vglm function of the package VGAM which can be used to build both ordinal
polytomous models and multinomial models.
Furthermore, in order to make the prediction, we compared the probability
p(x) to the 0.5 threshold. It can often be wise to change the value of this
threshold. In the example which we considered, we could choose a threshold
lower than 0.5 if we wanted to minimise the risk of mistakenly predicting that
cancer has not reached the lymphatic system. The techniques for scoring and
tracing the set graphs such as ROC and lift curves (see package ROCR) can
assist us in our choice of threshold. For further theoretical elements on logistic
regression, refer to Fox and Weisberg (2011), Hosmer and Lemeshow (2000)
and Collett (2003).
Decision Tree 199

9.3 Decision Tree

9.3.1 Objective
Decision trees are tools for data exploration and to assist in decision-making,
which can both explain and predict a quantitative variable (regression tree) or
a qualitative variable (classification tree) from quantitative and/or qualitative
explanatory variables. They consist of a sequential algorithm based on clas-
sification and regression tree (CART, Breiman et al., 1984), which constructs
individual classes. The classes are generated using binary rules constructed
from explanatory variables so that the individuals from a given class might
be as homogeneous as possible in terms of the response variable.
The main advantages of this method are the simplicity of and ease with
which the results can be both understood and interpreted due to the tree
diagram produced. These trees are composed of nodes and leaves and
are easy to read. Understanding is thus facilitated by labelling the nodes,
which is made possible by producing decision rules. They are also used to
select the relevant variables in a dataset with a great number of potentially
interesting variables.
Two packages, rpart and tree, can be used to build CART-based trees.
Here we will simply use the default package rpart.

9.3.2 Example
As an illustration we shall study the example of treating prostate cancer which
is explained in detail in Worked Example 9.2. The objective remains the
same: to predict Y (has the cancer reached the lymphatic system?) using the
variables age, acid, Xray, size, grade and log.acid. The response variable
is qualitative and we will therefore construct a classification tree.

9.3.3 Steps
1. Read the data.
2. Construct and analyse the decision tree.
3. Choose the size of the tree.
4. Estimate the classification error rate.

5. Predict Y for new individuals.

9.3.4 Processing the Example

1. Reading the data:
200 R for Statistics

Import the data from file prostate.txt and convert the variables Xray, size,
grade and Y into factors:

> prostate <- read.table("prostate.txt",header=T)

> for (i in 3:6) prostate[,i] <- factor(prostate[,i])

Summarise the dataset:

> summary(prostate)
age acid Xray size grade Y
Min. :45.00 Min. :0.4000 0:38 0:26 0:33 0:33
1st Qu.:56.00 1st Qu.:0.5000 1:15 1:27 1:20 1:20
Median :60.00 Median :0.6500
Mean :59.38 Mean :0.6942
3rd Qu.:65.00 3rd Qu.:0.7800
Max. :68.00 Max. :1.8700
log.acid
Min. :-0.9163
1st Qu.:-0.6931
Median :-0.4308
Mean :-0.4189
3rd Qu.:-0.2485
Max. : 0.6259

2. Constructing and analysing the decision tree:

We use the rpart function of the rpart package, which uses a formula like the
lm function (see Appendix A.2, p. 266):

> library(rpart)
> prostate.tree <- rpart(Y~.,data=prostate)
> prostate.tree
n= 53

node), split, n, loss, yval, (yprob)

* denotes terminal node

1) root 53 20 0 (0.6226415 0.3773585)

2) acid< 0.665 28 4 0 (0.8571429 0.1428571) *
3) acid>=0.665 25 9 1 (0.3600000 0.6400000)
6) Xray=0 16 7 0 (0.5625000 0.4375000) *
7) Xray=1 9 0 1 (0.0000000 1.0000000) *

Each node of the tree corresponds to a group of individuals, for which R yields
Node number: node, for example, 3)
Decision Tree 201

Cut rule: split, for example, acid>=0.665

Number of individuals in the group: n, for example, 25
Number of misclassified individuals: loss, for example, 9
Predicted value: yval, for example, 1
Probability of belonging to each class: (yprob), for example (0.3600
0.6400).

Thus in the case of the first node (called the root of the tree), it can be
seen that there are 53 observations (all the individuals), and that for 62.3%
of those individuals, cancer did not spread to the lymphatic system (Y = 0).
This first node is segmented using the variable acid. The individuals are
divided into two sub-groups: the first (node 2, acid < 0.665) is composed of
28 individuals, 85.7% of which correspond to Y = 0 and 14.3% to Y = 1. The
second (node 3, acid >= 0.665) is composed of 25 individuals (36% Y = 0,
64% Y = 1). Node 2 is not segmented whereas node 3 is segmented according
to the variable Xray, which yields nodes 6 and 7.
When the algorithm can no longer be segmented into nodes, we end by a leaf.
Thus, in our example we obtain three leaves which correspond to nodes 2, 6
and 7, and which can be identified in the script by the symbol *. For a pure
leaf, we obtain a group of individuals from just one class (example: in leaf 7,
100% of the individuals have a value of Y = 1).
Once the tree has been constructed, the class assigned to each leaf must of
course be defined. For a pure leaf, the answer is obvious. If the leaf is not pure,
a simple rule is to decide to assign it the class which is the most represented
in the leaf. Thus, if we examine leaf 6 (56.25% Y = 0, 43.75% Y = 1), we can
assume the following attribution rule: if acid > 0.665 and Xray = 0 then
Y = 0.
The plot function is used to obtain the tree which represents this segmentation
graphically (Figure 9.2). A number of graphical parameters are used to obtain
a tree which is easier to read:

> plot(prostate.tree,branch=0.2,compress=T,margin=0.1,
main="Prostate.tree")
> text(prostate.tree,fancy=T,use.n=T,pretty=0,all=T)

This diagram summarises all the abovementioned information. There are 53

data items (33 Y=0 and 20 Y=1) and therefore the predicted value at the first
node is 0. We then read the segmentation rules (acid < 0.665, etc.) and the
distribution of Y=0 and Y=1 in each node and each leaf. We can also calculate
the number of misclassified data: here, 4 + 7 + 0 = 11.
202 R for Statistics
Prostate.tree

0
|
33/20

acid< 0.665
acid>=0.665

0 1
24/4 9/16

Xray=0
Xray=1
0 1
9/7 0/9

Figure 9.2
Decision tree: predict variable Y (prostate dataset).

To analyse a tree, it is possible to read all the results using the summary
function. The advantage of this summary is that it lists the variables that
compete with those that are chosen. For each node, the summary therefore
uses Primary splits to return the primary variables which could have been
chosen (as second, third or fourth choice depending on the maxcompete argu-
ment).
Here is an extract of the summary:

> prostate.tree <- rpart(Y~.,data=prostate,maxcompete=2)

> summary(prostate.tree)

Call:
rpart(formula = Y ~ ., data = prostate, maxcompete = 2)
n= 53
Decision Tree 203

CP nsplit rel error xerror xstd

1 0.35 0 1.00 1.00 0.1764428
2 0.10 1 0.65 0.90 0.1723861
3 0.01 2 0.55 0.85 0.1699056

Node number 1: 53 observations, complexity param=0.35

predicted class=0 expected loss=0.3773585
class counts: 33 20
probabilities: 0.623 0.377
left son=2 (28 obs) right son=3 (25 obs)
Primary splits:
acid < 0.665 to the left, improve=6.528518, (0 missing)
log.acid < -0.407996 to the left, improve=6.528518, (0 missing)
Xray splits as LR, improve=5.302152, (0 missing)
Surrogate splits:
log.acid < -0.407996 to the left, agree=1.00, adj=1.00, (0 split)
Xray splits as LR, agree=0.585, adj=0.12, (0 split)
size splits as LR, agree=0.585, adj=0.12, (0 split)
age < 47.5 to the right, agree=0.566, adj=0.08, (0 split)

Node number 2: 28 observations

predicted class=0 expected loss=0.1428571
class counts: 24 4
probabilities: 0.857 0.143
...

At the first node level, the two competing variables are log.acid and Xray,
and the substitute variable is log.acid. It is clear that for the trees, the
choice of the level at which the cuts of continuous variables are made is de-
cided according to rank, which explains why the variables acid and log.acid
provide exactly the same information. The surrogate splits can be used to
handle missing values.

3. Choosing the size of the tree:

As the aim is to construct homogeneous groups, it may seem natural to choose
the tree with the greatest number of pure leaves. However, with a view to
predicting the Y value of a new individual, the challenge is to construct the
tree which best represents reality. However, the more leaves a tree has, the
greater the risk of misclassifying new individuals. On the contrary, the smaller
the tree, the more stable its future predictions will be. There is thus a com-
promise to construct a tree that well represents the reality and is sufficiently
stable. The rpart function always chooses trees through cross-validation us-
ing the argument xval: by default, this argument is fixed at 10 (the function
constructs 10 samples of size n/10). Other arguments, such as the minimum
number of data items found in a node, may interfere with this one (argument
minsplit).
204 R for Statistics

> prostate.tree2 <- rpart(Y~.,data=prostate,minsplit=5)

> plot(prostate.tree2,branch=0.2,compress=T,margin=0.1,
main="Prostate.tree2")
> text(prostate.tree2,pretty=0)

Prostate.tree2

acid< 0.665
|

age>=58.5 Xray=0
Xray=0
0
acid>=0.705
0 1 1

acid< 0.8
1
acid>=0.825
0

0 1

Figure 9.3
Tree built using minsplit=5.

By modifying the minsplit argument, we obtain a bigger tree (Figure 9.3)

which must then be pruned.
The tree is pruned using the snip.rpart function, either directly with the mouse
in the graph window (which must be active) or by specifying the nodes to be
removed. Another method is to analyse the results of the cross-validation
using the plotcp function. We therefore conduct a classical cross-validation:
each individual is put aside and the tree is constructed with the remaining
Decision Tree 205

n 1 individuals and the initial individual is then predicted (the leave-one-

out method).

> prostate.tree2 <- rpart(Y~.,data=prostate,minsplit=5,xval=53)

> plotcp(prostate.tree2)

We therefore choose a tree with a complexity value of cp=0.087 and which

also corresponds to a tree with four leaves (Figure 9.4).
size of tree

1 2 4 6 8
1.2
1.0
Xval Relative Error

0.8
0.6
0.4

Inf 0.23 0.087 0.035 0.016

cp

Figure 9.4
Choosing the size of the tree using plotcp.

Let us now construct the final tree:

> prostate.fin <- rpart(Y~.,data=prostate,minsplit=5,cp=0.087)

> plot(prostate.fin,branch=0.2,compress=T,margin=0.1,
main="Prostate.fin")
> text(prostate.fin,fancy=T,use.n=T,pretty=0,all=T)

It must be noted here that only the variables acid and Xray are involved in
predicting Y (Figure 9.5). Misclassifications are now equal to 4+3+0+0 = 7.
4. Estimating the classification error rate:
To estimate the rate of misclassification of the final tree, we can read the
number of classification errors on the graph in Figure 9.5, or calculate it
directly using the object prostate.fin:
206 R for Statistics
Prostate.fin

0
|
33/20

acid< 0.665
acid>=0.665

0 1
24/4 9/16
Xray=0
Xray=1
0 1
9/7 0/9

acid>=0.705
acid< 0.705

0 1
9/3 0/4

Figure 9.5
Final tree obtained using cross-validation.

> pred <- predict(prostate.fin,newdata=prostate,type="class")

> tab <- table(prostate$Y,pred)
> tab
pred
0 1
0 33 0
1 7 13
> error <- sum(pred!=prostate$Y)/nrow(prostate)
> error
[1] 0.1320755

Here the error rate of around 13.2% is calculated from the data which was used
to construct the tree and therefore underestimates the real error rate even if
the tree was chosen by cross-validation. To estimate the true rate, a forecast
is made using cross-validation (the aforementioned leave-one-out method):

> pred <- prostate$Y

> for (i in 1:length(pred)){
+ tree <- rpart(Y~.,data=prostate[-i,],minsplit=1,cp=0.087)
Decision Tree 207

+ pred[i] <- predict(tree,newdata=prostate[i,],type="class")

+ }
> tab <- table(prostate$Y,pred)
> tab
pred
0 1
0 31 2
1 9 11
> error <- sum(pred!=prostate$Y)/nrow(prostate)
> error
[1] 0.2075472

The estimated error rate is 20.8%. It can be seen that the rate is indeed greater
than the misclassification rate estimated without cross-validation (around
13.2%).
5. Predicting Y for new individuals:
The above decision tree can be used in the context of forecasting. The follow-
ing table contains measurements from four new patients for the six explanatory
variables:

Age Acid Xray Size Grade log.acid

61 0.60 1 0 1 0.51
49 0.86 0 0 1 0.15
67 0.72 1 0 1 0.33
51 0.95 1 1 1 0.05

To obtain the values of Y predicted by the decision tree for these four new
individuals, we first bring together the new data in a data-frame with the
same structure as the initial data table (and especially the same names for
the variables):

> new.d <- data.frame(matrix(c(61,0.60,1,0,1,-0.51,49,0.86,0,0,1,

-0.15,67,0.72,1,0,1,-0.33,51,0.95,1,1,1,-0.05),ncol=6,byrow=T))
> names(new.d) <- names(prostate)[-6]
> for (i in 3:5){new.d[,i] <- factor(new.d[,i])}
> summary(new.d)

age acid Xray size grade log.acid

Min. :49.0 Min. :0.6000 0:1 0:3 1:4 Min. :-0.510
1st Qu.:50.5 1st Qu.:0.6900 1:3 1:1 1st Qu.:-0.375
Median :56.0 Median :0.7900 Median :-0.240
Mean :57.0 Mean :0.7825 Mean :-0.260
3rd Qu.:62.5 3rd Qu.:0.8825 3rd Qu.:-0.125
Max. :67.0 Max. :0.9500 Max. :-0.050
208 R for Statistics

> pred <- predict(prostate.fin,newdata=new.d,type="class")

> pred
1 2 3 4
0 0 1 1
Levels: 0 1
Remark
Here, we must list all the explanatory variables for the new individuals. An-
other solution is to reconstruct the tree (with the rpart function) using only
those variables which have previously been retained, prior to making the pre-
diction.

For the first two individuals, the tree predicts Y = 0, which means that
the cancer has not yet reached the lymphatic system, whereas for the last
two individuals, the tree predicts Y = 1, that is, the cancer has reached the
lymphatic system.
More precisely, for these individuals, we can also examine the probability of
predicting Y = 1 (rather than Y directly):

> pred <- predict(prostate.fin,newdata=new.d)

> pred
0 1
1 0.8571429 0.1428571
2 0.7500000 0.2500000
3 0.0000000 1.0000000
4 0.0000000 1.0000000

We can consider that the risk of being mistaken when we predict Y = 0 for
the first two individuals is rather low as, according to the tree, there is less
than a 25% risk that the cancer has reached the lymphatic system.

9.3.5 Rcmdr Corner

It is not possible to construct a decision tree with Rcmdr.

9.3.6 Taking Things Further

It is possible to construct a tree when the variable Y is multivariate, again
using rpart.
It is well known that decision trees are unstable. For this reason, we
recommend the use of the argument maxcompete which displays the concurrent
variables. It is also possible to combine a great number of trees using the
randomForest package.
For more details on decision tree, refer to Breiman et al. (1984), Devroye
et al. (1996) or Hastie et al. (2009).
10
Exploratory Multivariate Analysis

10.1 Principal Component Analysis

10.1.1 Objective
Principal component analysis (PCA) summarises a data table where the in-
dividuals are described by (continuous) quantitative variables. PCA is used
to study the similarities between individuals from the point of view of all the
variables and identifies individuals profiles. It is also used to study the lin-
ear relationships between variables (from correlation coefficients). The two
studies are linked, which means that the individuals or groups can be charac-
terised by the variables and the links between variables illustrated from typical
individuals.
We use the PCA function from the FactoMineR package (see Appendix A.4,
p. 269) because it allows one to obtain numerous indicators and graphical
outputs. Using this function, it is possible to add supplementary elements
and construct the graphs simply and automatically.

10.1.2 Example
The dataset involves the results of decathlon events at two athletics compe-
titions which took place one month apart: the Athens Olympic Games (23rd
and 24th August 2004) and Decastar (25th and 26th September 2004). On
the first day, the athletes compete in five events (100 m, long jump, shot put,
high jump, 400 m) and in the remaining events on the second day (110 m
hurdles, discus, pole vault, javelin, 1500 m). In Table 10.1, we have grouped
together the performances for each athlete at each of the ten events, their
overall rank, their final number of points and the competition in which they
participated.
The aim of conducting PCA on this dataset is to determine profiles for
similar performances: are there any athletes who are better at endurance
events or those requiring short bursts of energy? And are some of the events
similar? If an athlete performs well in one event, will he necessarily perform
well in another?
210 R for Statistics

TABLE 10.1
Athletes Performances in ten Decathlon Events

1500m

Points

Comp
100m

400m

110m

Rank
Long

High
Shot

Jave
Disc

Pole
Sebrle 10.85 7.84 16.36 2.12 48.36 14.05 48.72 5.00 70.52 280.01 1 8893 OG
Clay 10.44 7.96 15.23 2.06 49.19 14.13 50.11 4.90 69.71 282 2 8820 OG
Karpov 10.5 7.81 15.93 2.09 46.81 13.97 51.65 4.60 55.54 278.11 3 8725 OG
Macey 10.89 7.47 15.73 2.15 48.97 14.56 48.34 4.40 58.46 265.42 4 8414 OG
Warners 10.62 7.74 14.48 1.97 47.97 14.01 43.73 4.90 55.39 278.05 5 8343 OG
Zsivoczky 10.91 7.14 15.31 2.12 49.40 14.95 45.62 4.70 63.45 269.54 6 8287 OG
Hernu 10.97 7.19 14.65 2.03 48.73 14.25 44.72 4.80 57.76 264.35 7 8237 OG
.. .. .. .. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . . . . .
SEBRLE 11.04 7.58 14.83 2.07 49.81 14.69 43.75 5.02 63.19 291.70 1 8217 Deca
CLAY 10.76 7.40 14.26 1.86 49.37 14.05 50.72 4.92 60.15 301.50 2 8122 Deca
.. .. .. .. .. .. .. .. .. .. .. .. .. ..
. . . . . . . . . . . . . .

10.1.3 Steps
1. Read the data.
2. Choose the active individuals and variables.
3. Choose if the variables need to be standardised.

4. Choose the number of dimensions.

5. Analyse the results.
6. Automatically describe the dimensions of variability.

7. Back to raw data.

10.1.4 Processing the Example

1. Reading the data:
It is here important to specify that the first column corresponds to the names
of the individuals (row.names=1):

> decath <- read.table("decathlon.csv",sep=";",dec=".",

header=TRUE, row.names=1)
> summary(decath)

2. Choosing the active individuals and variables:

As we want to determine the performance profiles, the variables corresponding
to the athletes performance are considered as active (the first ten variables).
Principal Component Analysis 211

The choice of active variables is very important: these variables, and these
variables alone, will be taken into account when constructing the PCA dimen-
sions. In other words, these are the only variables which are used to calculate
the distances between individuals. As supplementary variables, we can add
the quantitative variables number of points and rank, and the qualitative
variable competition. Supplementary variables are very useful in helping to
interpret the dimensions.
We can also choose the individuals which participate in constructing the di-
mensions, known as active individuals. Here, as is often the case, all the
individuals are considered active.
3. Choosing if the variables need to be standardised:
We can centre and standardise the variables or simply centre them. For the
dataset processed in this example, we do not have the choice. Standardisation
is essential as the variables are of different units. When the variables are of the
same units, both solutions are possible and lead to two separate analyses. This
decision is therefore crucial. Standardisation means that every variable has
the same importance. Not standardising the data therefore means that each
variable has a weight corresponding to its standard deviation. This choice is
all the more important as the variables have different variances.
To conduct PCA using the FactoMineR package (see Appendix A.4, p. 269),
the preinstalled package must be loaded (see Section 1.6). We then use the
PCA function.

> library (FactoMineR)

> res.pca <- PCA(decath, quanti.sup=11:12, quali.sup=13)

Variables 11 and 12 are quantitative supplementary variables and variable 13

is a qualitative supplementary variable. By default, the variables are cen-
tred and standardised (this is known as standardised PCA). To avoid scaling
the variables, use the argument scale.unit=FALSE. By default, results are
given for the first five dimensions. To specify another number of dimen-
sions, use the argument ncp. The list res.pca contains all the results (see
names(res.pca)).
4. Choosing the number of dimensions:
There are many solutions available in order to determine the number of di-
mensions to be analysed in PCA. The most common is to represent them
using a bar chart of the eigenvalues or percentage of variance associated with
each dimension (Figure 10.1):

> barplot(res.pca$eig[,2],names=paste("Dim",1:nrow(res.pca$eig)))
212 R for Statistics

10 15 20 25 30
5
0

Dim 1 Dim 2 Dim 3 Dim 4 Dim 5 Dim 6 Dim 7 Dim 8 Dim 9 Dim 10

Figure 10.1
Percentage of variance associated with each dimension of the PCA.

We are therefore looking for a visible decrease or gap in the bar chart. Here,
we can analyse the first four dimensions. Indeed, after four dimensions, we
remark a regular decrease in the percentage of variance and can see a small
jump between dimensions 4 and 5. To know the percentage of variance
explained by the first four dimensions, we use the following line of code:

> round(res.pca$eig[1:4,],2)
eigenvalue percentage of cumulative percentage
variance of variance
comp 1 3.27 32.72 32.72
comp 2 1.74 17.37 50.09
comp 3 1.40 14.05 64.14
comp 4 1.06 10.57 74.71

The first two dimensions express 50% of the total percentage of variance, that
is to say, 50% of the information in the data table is contained within the first
two dimensions. This means that the diversity of the performance profiles
cannot be summarised by two dimensions. Here, the first four dimensions are
sufficient to explain 75% of the total variance.
5. Analysing the results:
By default, the PCA function yields the graph of individuals (Figure 10.2)
and the graph of variables for the first two dimensions (Figure 10.3). The
supplementary qualitative variable is represented on the graph of individu-
als. Each category of the qualitative variable competition is represented at
the barycentre of the individuals which take this category. The supplemen-
tary quantitative variables are represented on the graph of variables. The
correlation between these variables and the dimensions can be interpreted.
Principal Component Analysis 213

Individuals factor map (PCA)

Casarsa
4

Korkizoglou YURKOV
Parkhomenko
2

Sebrle
Zsivoczky
Smith Macey
Pogorelov Clay
Dim 2 (17.37%)

MARTINEAU Terek SEBRLE

HERNU Barras CLAY
BOURGUIGNON Uldal McMULLEN KARPOV
Turi Karpov
Decastar Bernard
OG
0

BARRAS Qi Schoenbeck
Karlivans
Ojaniemi Hernu
BERNARD
Smirnov ZSIVOCZKY
Gomez Schwarzl
Lorenzo
Nool Averyanov Warners

WARNERS
-2

NOOL

Drews

-6 -4 -2 0 2 4

Dim 1 (32.72%)

Figure 10.2
Decathlon PCA: graph of individuals (results of Decastar shown in capitals).

To interpret the results of a PCA, it is common to simultaneously study the

results of the individuals and of the variables. Indeed, we want to characterise
the different performance profiles from the events. The general rules for in-
terpreting the two graphs simultaneously are the following ones: individuals
which have high (resp. low) coordinates on a dimension take high (resp. low)
values for the variables which are highly correlated to this dimension.
Overall, the first dimension confronts the consistently high performance pro-
files (i.e. of good all-round athletes), such as Karpov at the Olympics, and
profiles of athletes who are consistently weak (relatively!), such as Bour-
guignon at Decastar.
It must be noted that in Figure 10.3, the variable 100m is negatively correlated
with the variable long jump as athletes who run fast in the 100 metres, and
therefore have a short time, jump a long way.
It must here be noted that the first principal component is the linear combina-
tion of the variables which best summarise all the variables. In this example,
the automatic summary provided by PCA almost coincides with the number
of points (the correlation is almost equal to 1), that is to say, the official
214 R for Statistics

Variables factor map (PCA)

1.0

Discus
400m Shot.put
1500m
0.5

Javeline
High.jump
110m.hurdle
Dim 2 (17.37%)

100m
Rank
0.0

Points

Pole.vault
Long.jump
-0.5
-1.0

-1.0 -0.5 0.0 0.5 1.0

Dim 1 (32.72%)

Figure 10.3
Decathlon PCA: graph of variables.

summary. Of course, the number of points was not used in the automatic
summary, as this variable is purely illustrative.
We can interpret the position of the barycentres of the competition variable on
the graph of individuals. The Olympics category has a positive coordinate on
dimension 1 whereas the Decastar category has a negative coordinate. We can
therefore consider that, on average, athletes perform better at the Olympics
than at Decastar.
Dimension 2 is positively correlated to endurance variables (1500 m, 400 m)
and power variables (discus, shot put). It means that dimension 2 opposes
tough athletes (at the bottom since their time for 1500 m and 400 m is small)
to powerful athletes on the top. It can be noted that both the categories
of the qualitative variable competition have null coordinates on dimension
2, showing that overall there is no evolution between the Olympics and the
Decastar from the second dimension point of view: this means that the ath-
letes improved their overall performance (see dimension 1) but their profile
did not change.
These interpretations can be refined using numerical results. The individuals
Principal Component Analysis 215

numerical results are contained within the object res.pca$ind. We then

obtain a list with the individuals coordinates on the dimensions (the principal
components or scores), the squared cosines (which measure the quality of
the projection of the individuals on the dimensions and are used to know if
distances between individuals can be interpreted), and the contributions of
the individuals to the construction of the dimensions. The results can be
visualised by concatenating them; for example,

> round(cbind(res.pca$ind$coord[,1:3],res.pca$ind$cos2[,1:3],
res.pca$ind$contrib[,1:3]), digits=2)
Dim.1 Dim.2 Dim.3 Dim.1 Dim.2 Dim.3 Dim.1 Dim.2 Dim.3
Sebrle 4.04 1.37 -0.29 0.70 0.08 0.00 12.16 2.62 0.15
Clay 3.92 0.84 0.23 0.71 0.03 0.00 11.45 0.98 0.09
Karpov 4.62 0.04 -0.04 0.85 0.00 0.00 15.91 0.00 0.00
Macey 2.23 1.04 -1.86 0.42 0.09 0.29 3.72 1.52 6.03
.
. .
. .
. .
. .
. .
. .
. .
. .
. .
.
. . . . . . . . . .

The list res.pca$var yields similar results for the variables: the coordinates,
the squared cosines and the contributions of the variables. Loadings can be
obtained by dividing the coordinates of the variables on a dimension by the
square root eigenvalue of the corresponding dimension:

> loadings<-sweep(res$var$coord,2,sqrt(res$eig[1:5,1]),FUN="/")

Results of the supplementary quantitative variables are contained within the

object res.pca$quanti.sup and the qualitative variables are contained within
the object res.pca$quali.sup.
To improve the readability of the results, we can draw the graph of individuals
by, for example, colouring the individuals according to a qualitative variable.
In this case we use the plot.PCA function (which can be called indifferently
using plot or plot.PCA, see Appendix A.1, p. 257). In the plot.PCA function,
we provide the object which contains all the results of the PCA (res.pca),
the graph that we want to construct (for the individuals choix="ind") and
we specify that the points are coloured according to the thirteenth variable
(habillage=13). We can also alter the size of the font for the labels ("cex =
0.7" instead of 1 by default).

> plot(res.pca, choix="ind", habillage=13, cex=0.7)

We were only interested in the results of the first two dimensions, but it is
possible to construct graphs of individuals and variables on dimensions 3 and
4:

> plot(res.pca, choix="ind", habillage=13, axes=3:4, cex=0.7)

> plot(res.pca, choix="var", habillage=13, axes=3:4)
216 R for Statistics

To save a graph in pdf format, for example, we use the argument new.plot=F:

> pdf("mypath/mygraph.pdf")
> plot(res.pca, choix="var", habillage=13, axes=3:4, new.plot=F)
> dev.off()

6. Automatically describing the dimensions of variability:

In the FactoMineR package there is a function by which the dimensions of
PCA can be described automatically: dimdesc. This function is a new aid to
interpretation. It sorts the quantitative variables according to their coordi-
nates on a dimension, that is to say, the correlation coefficient between the
quantitative variable and the PCA dimension. Only the significant correlation
coefficients are retained. This function also sorts the qualitative variables as
well as the categories of the qualitative variables. In order to do so, a one-way
analysis of variance model (see Section 8.1) is conducted for each qualitative
variable: the response variable Y corresponds to the principal component (the
coordinates of the individuals or scores) and is explained according to a quali-
tative variable. The p-value of the overall test (F -test) is calculated as well as
the
P p-values of the tests of each category (test H0 : i = 0 with the constraint
i i = 0). Only the significant results are given. The p-values associated
with the F -tests are sorted in ascending order. Thus, the qualitative variables
are sorted from most to least characteristic. In the same way, the p-values of
the t-tests by categories are sorted.

> dimdesc(res.pca)
$Dim.1
$Dim.1$quanti
correlation P-value
Points 0.9561543 2.099191e-22
Long.jump 0.7418997 2.849886e-08
Shot.put 0.6225026 1.388321e-05
High.jump 0.5719453 9.362285e-05
Discus 0.5524665 1.802220e-04
Rank -0.6705104 1.616348e-06
X400m -0.6796099 1.028175e-06
X110m.hurdle -0.7462453 2.136962e-08
X100m -0.7747198 2.778467e-09

$Dim.2
$Dim.2$quanti
correlation P-value
Discus 0.6063134 2.650745e-05
Shot.put 0.5983033 3.603567e-05
X400m 0.5694378 1.020941e-04
X1500m 0.4742238 1.734405e-03
Principal Component Analysis 217

High.jump 0.3502936 2.475025e-02

Javeline 0.3169891 4.344974e-02
Long.jump -0.3454213 2.696969e-02

This function becomes very useful when there are a lot of variables. Here
we can see that the first dimension is primarily linked to the variable Points
(correlation coefficient of 0.96), and then to the variable 100m (negative cor-
relation of 0.77), etc. The second dimension is principally described by the
quantitative variables Discus and Shot put. None of the qualitative variables
can be used to characterise dimensions 1 and 2 with a confidence level of 95%.
The confidence level can be changed (proba = 0.2) and, for the qualitative
variables, yields

> dimdesc(res.pca, proba = 0.2)

$Dim.1$quali
P-value
Competition 0.1552515

$Dim.1$category
Estimate P-value
OG 0.4393744 0.1552515
Decastar -0.4393744 0.1552515

7. Going back to raw data:

It is always important to refine the interpretation by going back to the raw
data. As the raw data table is often difficult to read, to help us we can use
the data table which has been centred-standardised:

> round(scale(decath[,1:12]),2)
100m Long Shot High 400m ...
Sebrle -0.56 1.83 2.28 1.61 -1.09 ...
Clay -2.12 2.21 0.91 0.94 -0.37 ...
Karpov -1.89 1.74 1.76 1.27 -2.43 ...
Macey -0.41 0.66 1.52 1.95 -0.56 ...
Warners -1.44 1.52 0.00 -0.08 -1.43 ...
.
. .
. .
. .
. .
. .
. .
.
. . . . . . .

For example, Sebrle throws the shot put further than average (positive stan-
dardised value) somewhat remarkably (extreme value as it is greater than
2).
In the same way we can calculate the matrix of the correlations to know the
exact correlation coefficient between the variables rather than the approxima-
tions provided by the graph:
218 R for Statistics

> round(cor(decath[,1:12]),2)
100m Long Shot High 400m ...
100m 1.00 -0.60 -0.36 -0.25 0.52 ...
Long -0.60 1.00 0.18 0.29 -0.60 ...
Shot -0.36 0.18 1.00 0.49 -0.14 ...
High -0.25 0.29 0.49 1.00 -0.19 ...
400m 0.52 -0.60 -0.14 -0.19 1.00 ...
.
. .
. .
. .
. .
. .
. .
.
. . . . . . .

10.1.5 Rcmdr Corner

A graphical interface for FactoMineR is also available in Rcmdr which makes
it possible to conduct the previous analyses with an easy-to-use drop-down
menu. To do this, simply install the drop-down menu once by launching the
following command in an R window (internet connection required):
> source("http://factominer.free.fr/install-facto.r")
In following sessions, the FactoMineR drop-down menu will be present in
Rcmdr and you will simply need to type library(Rcmdr).

1. Import the data:

FactoMineR Import data from text file, clipboard, or URL ...
You must specify that the individuals identifiers are available in the dataset.
In order to do so, we select Row names in the first column. This option
is available in the FactoMineR tab but is not available from the Data tab.
26. We will not discuss the choice of active elements and whether or not to
standardise the variables, and will only show how to conduct PCA with the
FactoMineR drop-down menu. Launch Rcmdr and click on the FactoMineR
tab. Choose Principal Component Analysis to open the main window from
the PCA drop-down menu (Figure 10.4).
By default, all the variables are considered as active, but it is possible to
select some of them. It is also possible to select supplementary qualita-
tive variables (Select supplementary factors), supplementary quantita-
tive variables (Select supplementary variables) and supplementary indi-
viduals (Select supplementary individuals). By default, the results on
the first five dimensions are provided in the object res, the variables are
centred and standardised, and the graphs are provided for the first two di-
mensions. It is preferable to click Apply rather than Submit, as it launches
the analysis whilst also keeping the window open, and it is thus possible to
modify certain options without having to reset all the parameters.
The graphical options window (Figure 10.5) is separated into two parts. The
left part concerns the graph of individuals and the right part the graph of vari-
ables. It is possible to represent only the supplementary qualitative variables
Principal Component Analysis 219

Figure 10.4
PCA main window in the FactoMineR menu.

(without the individuals Hide some elements: check ind); it is also possible
to omit the labels for the individuals (Label for the active individuals).
The individuals can be coloured according to a qualitative variable (Coloring
for individuals: choose the qualitative variable).
The window for the different output options can be used to visualise the
different results (eigenvalues, individuals, variables, automatic description of
the dimensions). All the results can also be exported into a csv file (which
can be opened in spreadsheet such as Excel, for example).

10.1.6 Taking Things Further

1. Constructing confidence ellipses around categories:
It is possible to construct confidence ellipses around the categories of a sup-
plementary qualitative variable (see Figure 10.6). These ellipses thus give
220 R for Statistics

Figure 10.5
PCA graphical options window.

confidence regions for the position of each of the categories. To construct

these ellipses, we write

> plotellipses(res.pca,keepvar=13)

2. Handling missing values in PCA:

It is possible to perform PCA with missing values using the missMDA package.
The first step consists of imputing the missing values using the imputePCA
function and the second step consists of performing PCA on the completed
dataset:

> res <- imputePCA(my.incomplete.dataset)

> res.pca <- PCA(res$completeObs)

Overviews of PCA are available in many books such as Jolliffe (2002),

Lebart et al. (1984) or Govaert (2009). In Husson et al. (2010), many studies
conducted with FactoMineR are also described in detail.
Principal Component Analysis 221

Individuals factor map (PCA)

Decastar
OlympicG
Casarsa
4

Parkhomenko YURKOV
Korkizoglou
2

Sebrle
Zsivoczky Macey
Smith
Dim 2 (17.37%)

Clay
MARTINEAU HERNU TerekPogorelov SEBRLE CLAY
Barras
BOURGUIGNON Uldal KARPOV
Turi Decastar McMULLEN Karpov
BARRAS OlympicG
0

Qi Schoenbeck Bernard
Karlivans Hernu
BERNARD Ojaniemi
Smirnov ZSIVOCZKY
Lorenzo Gomez Averyanov
Schwarzl Nool Warners
WARNERS
-2

NOOL

Drews
-4

-6 -4 -2 0 2 4 6

Dim 1 (32.72%)

Figure 10.6
Confidence ellipses for the categories of a qualitative variable in PCA.
222 R for Statistics

10.2 Correspondence Analysis

10.2.1 Objective
Correspondence analysis (CA) is used to summarise and visualise contingency
tables, that is, a table confronting two qualitative variables. At the junction
between row i and column j, we find the number of individuals carrying the
category i of the first variable and j of the second. The aims of CA are as
follows:
Compare the row profiles with one another.
Compare the column profiles with one another.
Interpret the proximity between the rows and the columns, in other words,
visualise the associations between the categories of the two variables.
To conduct a CA, we use the CA function from the FactoMineR package
devoted to multivariate exploratory data analysis (see Appendix A.4, p. 269).

10.2.2 Example
The dataset represents the number of students from French universities by
subject (or discipline) and by course by sex for the academic year 2007
2008. This is a table confronting the two qualitative variables Discipline
and Level-sex. In the rows it features ten disciplines offered by the uni-
versity and in the columns the junctions of the variables level (bachelors,
masters and PhD) and sex (male or female). The CA is thus applied between
one qualitative variable Discipline and a qualitative variable Level-sex cor-
responding to the confrontation of these two variables. This situation occurs
frequently in correspondence analysis. Furthermore, for discipline, we also
have the total number of students by level, by sex and an overall total (see
Table 10.2).
The aim of this study is to have an overall image of the university. Which
disciplines have the same student profiles? Which disciplines are favoured by
women (and men respectively)? Which disciplines tend to have the longest
courses of study?

10.2.3 Steps
1. Read the data.
2. Choose the active rows and columns.
3. Conduct the CA.
4. Choose the number of dimensions.
Correspondence Analysis 223
TABLE 10.2
University Datafile
Bachelors Masters ... Sum
F M F M
Law, Political Science 69373 37317 42371 21693 ... 179125
Social Science, Economics, Management 38387 37157 29466 26929 ... 136474
Economic and Social Administration 18574 12388 4183 2884 ... 38029
Literature, Linguistics, Arts 48691 17850 17672 5853 ... 96998
Languages 62736 21291 13186 3874 ... 103833
Social Sciences and Humanities 94346 41050 43016 20447 ... 213618
Joint Humanities 1779 726 2356 811 ... 5700
Mathematics, Physics 22559 54861 17078 48293 ... 158689
Life Science 24318 15004 11090 8457 ... 69742
Sport Science 8248 17253 1963 4172 ... 32152

5. Analyse the results.

10.2.4 Processing the Example

1. Reading the data:
It is here important to specify that the first column corresponds to the names
of the rows (row.names=1). Check that the data has been imported correctly
using the summary function or by visualising an extract of the dataset.

> University <- read.table("University.csv",sep=";",

header=T, row.names=1)
> summary(University)
> University[1:5,1:3]
Bachelors.F Bachelors.M Masters.F
Law, Political Science 69373 37317 42371
Social Sc., Eco., Management 38387 37157 29466
Economic and Social Admin. 18574 12388 4183
Literature, Linguistics, Arts 48691 17850 17672
Languages 62736 21291 13186

2. Choosing the active rows and columns:

The choice of active rows and columns is vital as only these rows and columns
are taken into account when constructing the dimensions. As we want to
compare all the disciplines, we will consider them as active rows. The columns
for the variable Level-sex will be considered active and those corresponding
to the various totals will be considered illustrative (the information contained
within these variables is already taken into account by the active columns).
These variables (from 7 to 12) are only used for interpreting the results.
3. Conducting the CA:
224 R for Statistics

The correspondence analysis can now be conducted using the FactoMineR

package (see Appendix A.4, p. 269). In order to do so, the package must be
installed and then loaded (see Section 1.6). We then use the CA function.

> library(FactoMineR)
> res.ca <- CA(University, col.sup=7:12)

As mentioned above, columns 7 to 12 are considered illustrative. The list

res.ca contains all the outputs. By default, a chi-square test is constructed
to test the independence of the variables Disciplines and Level-sex. The
test is constructed using only the active categories (rows and columns). In
our example the independence hypothesis is largely rejected as the p-value
is near 0: The chi-square of independence between the two variables
is equal to 170789.2 (p-value = 0).
This means, as might be expected, that there are associations between certain
disciplines and sex-level combinations. It is therefore interesting to construct
a CA in order to visualise these associations between categories.
4. Choosing the number of dimensions:
In order to determine the number of dimensions to be analysed in CA, it can
be useful to represent the eigenvalues or the percentage of variance associated
with each dimension in bar charts (Figure 10.7).

> barplot(res.ca$eig[,2],names=paste("Dim",1:nrow(res.ca$eig)))
70
50
30
0 10

Dim 1 Dim 2 Dim 3 Dim 4 Dim 5 Dim 6

Figure 10.7
Percentage of variance associated with each dimension of the CA.

We can also directly examine the percentages of variance associated with each
dimension using the following command:

> round(res.ca$eig,3)
eigenvalue percentage of cumulative percentage
variance of variance
Correspondence Analysis 225

dim 1 0.117 70.718 70.718

dim 2 0.026 15.508 86.226
dim 3 0.018 10.901 97.127
dim 4 0.004 2.627 99.754
dim 5 0.000 0.246 100.000
dim 6 0.000 0.000 100.000

The percentages of variance associated with the first dimensions are high, and
we can simply describe the first three dimensions, as confirmed by the jump
seen in Figure 10.7. The first three dimensions express nearly 97% of the
total variance: In other words, 97% of the information in the data table is
summarised by the first three dimensions.
5. Analysing the results:
By default, the CA function yields a graph simultaneously representing (active
and illustrative) rows and columns on the first two dimensions. In Figure 10.8

CA factor map

PhD.F
PhD
0.4

PhD.M
Joint humanities
Life science
Masters.F
0.2

Masters
Social sciences and Humanities
Literature, Masters.M
Law, political science
Linguistics, Sum.F
Dim 2 (15.51%)
0.0

Arts Sum Social science, Mathematics,

Bachelors.F Sum.M Physics
Economics,
Languages Bachelors Management
-0.2

Bachelors.M

Economic and Social Administration

-0.4

Sport science
-0.6

-0.4 -0.2 0.0 0.2 0.4 0.6 0.8

Dim 1 (70.72%)

Figure 10.8
CA on university data.
226 R for Statistics

we can view the major tendencies emanating from the analysis. If the graph
contains too many points and becomes difficult to read, we can, for example,
represent the rows alone using the invisible argument of the plot.CA:

> plot(res.ca, invisible=c("col","col.sup"))

First of all it must be noted that, by design, there can be no size effect in
CA as each row (and column respectively) is divided by its margin. Thus, in
the example, we do not see on one side the disciplines which attract a lot of
students and on the other those which attract the fewest.
We can also examine proximities between the disciplines. Two disciplines
are considered similar if they have the same profiles (they attract students
of the same sex and for the same course length). For example, the graph
shows that the disciplines Languages and Literature, Linguistics, Arts
mainly attract women at Bachelors level. The supplementary columns can
help in interpreting the CA graph. Thus, the first dimensions opposes men
and women, whereas the second dimensions classes the levels from Bachelors
(at the bottom) to PhD (at the top). The disciplines on the left of the graph
(or right, respectively) are mainly studied by women (or men, respectively):
literary disciplines are towards the left of the graph (and therefore studied
mainly by women) and scientific disciplines are towards the right of the graph
(and thus studied by men). The disciplines towards the bottom of the graph
are those with short courses (Bachelor level over-represented by social and
economic administration and sport, for example), whereas the disciplines to-
wards the top of the graph have longer courses (PhD level over-represented
for biological and life sciences, for example). However, it is not always easy
to interpret the dimensions of the CA. In this case, we focus on the proximity
between the categories.
To refine these interpretations, we can go on to examine all the numerical
results. The numerical results of the rows are contained within the object
res.ca$row. We thus obtain a table with the row coordinates on the dimen-
sions, a table with the cosines squared (which represent the quality of the
projection of a row category), and a table with the contributions of the rows
to the construction of the dimensions (not given here):

> round(cbind(res.ca$row$coord[,1:3],res.ca$row$cos2[,1:3]),2)
Dim 1 Dim 2 Dim 3 Dim 1 Dim 2 Dim 3
Law, political science -0.10 0.07 -0.13 0.30 0.13 0.55
Social sc., Eco., Management 0.18 -0.02 -0.20 0.46 0.00 0.52
Economic and Social Admin. -0.19 -0.37 0.01 0.20 0.80 0.00
Literature, Ling., Arts -0.32 0.05 0.08 0.91 0.02 0.06
Languages -0.45 -0.18 0.09 0.79 0.13 0.03
Social sc. and Humanities -0.19 0.08 0.01 0.84 0.15 0.00
Joint humanities -0.13 0.28 -0.58 0.04 0.18 0.77
Mathematics, Physics 0.67 0.01 0.08 0.98 0.00 0.01
Correspondence Analysis 227

Life science 0.03 0.25 0.28 0.01 0.41 0.52

Sport science 0.32 -0.57 0.08 0.21 0.67 0.01

Similarly, the column results are obtained using

> round(cbind(res.ca$col$coord[,1:3],res.ca$col$cos2[,1:3],
res.ca$col$contrib[,1:3]),2)
Dim 1 Dim 2 Dim 3 Dim 1 Dim 2 Dim 3 Dim 1 Dim 2 Dim 3
Bachel.F -0.35 -0.04 0.04 0.96 0.01 0.01 39.72 2.27 3.65
Bachel.M 0.23 -0.20 0.03 0.55 0.39 0.01 11.51 37.49 1.21
Masters.F -0.11 0.17 -0.21 0.14 0.33 0.49 1.99 20.90 43.75
Masters.M 0.58 0.06 -0.07 0.95 0.01 0.01 40.43 1.97 3.26
PhD.F -0.06 0.43 0.39 0.01 0.49 0.42 0.08 21.02 25.41
PhD.M 0.47 0.36 0.35 0.46 0.26 0.26 6.27 16.36 22.71

We can then examine the results concerning the supplementary elements con-
tained within the object res.ca$col.sup:

> round(cbind(res.ca$col.sup$coord[,1:3],
res.ca$col.sup$cos2[,1:3]),2)
Dim 1 Dim 2 Dim 3 Dim 1 Dim 2 Dim 3
Sum.F -0.26 0.05 -0.02 0.44 0.02 0.00
Sum.M 0.37 -0.07 0.02 0.60 0.02 0.00
Bachelors -0.12 -0.10 0.04 0.13 0.09 0.01
Masters 0.19 0.12 -0.15 0.23 0.10 0.13
PhD 0.22 0.39 0.37 0.11 0.35 0.32
Sum 0.00 0.00 0.00 0.00 0.00 0.00

As expected, the coordinates of the total are at the barycentre of the cloud.
The clouds centre of gravity corresponds to the mean profile. This shows
that if the position of a Level-sex combination is close to the centre of the
cloud, this combination has the same discipline profile as that of the students
overall.
We can also construct the graph for dimensions 3 and 4:

> plot(res.ca, axes = 3:4)

10.2.5 Rcmdr Corner

A graphical interface is also available in Rcmdr which makes it possible to con-
duct the previous analyses with an easy-to-use drop-down menu. To do this,
simply install the drop-down menu once by launching the following command
in an R window (Internet connection required):
> source("http://factominer.free.fr/install-facto.r")
228 R for Statistics

In following sessions, the FactoMineR drop-down menu will be present in

Rcmdr and you will simply need to type library(Rcmdr).

1. Reading the data from a file:

FactoMineR Import data from text file, clipboard, or URL ...
You must specify that the row names are available in the dataset. In order to
do so, we select Row names in the first column. This option is available
in the FactoMineR tab but is not available from the Data tab.
25. We will not describe the choice of active elements but simply show how
to conduct CA in Rcmdr.
To conduct CA, click on the FactoMineR tab and then on Correspondence
analysis. In the main window of the CA drop-down menu, it is possible to se-
lect supplementary rows (Select supplementary rows), or columns (Select
supplementary columns). By default, results on the first five dimensions
are provided in the object res. It is preferable to click Apply (rather than
Submit), which launches the analysis whilst keeping the window open, and it
is thus possible to modify certain options without having to reconfigure all
the parameters.
In graph options it is possible to omit the representation of certain elements
(rows or columns, active or supplementary) or labels. The additional out-
put options window is used to obtain the different results (eigenvalues, rows,
columns). All the results can also be exported into a csv file (which can be
opened in spreadsheet such as Excel, for example).

10.2.6 Taking Things Further

1. Relationship with the 2 :
It can be interesting to go back to the raw data to analyse in more detail the
relationship between two variables and particularly the proximity between
the categories of these variables. We can also construct a chi-square test (see
Worked Example 6.2) and calculate the matrix of deviations from indepen-
dence.

> resultat <- chisq.test(University[,1:6])

> round(100 * resultat$residuals^2 / resultat$stat, 2)
> round(resultat$residuals,1)

This matrix of deviation from independence is visualised using CA.

2. Text mining:
Correspondence analysis is an extremely useful tool for conducting text analy-
ses. Text analysis consists of comparing texts, for example, from two different
authors, based on the words used in the texts. The principle is as follows: We
Correspondence Analysis 229

list all the words used in the texts and then draw up a table with the texts in
the rows and the words in the columns. Within a cell, we specify the number
of times a word has been used in a text. Applying a CA to this kind of table
makes it possible to compare the two authors and to see which words are over-
or under-used by which texts.

Overviews of CA can be found in Greenacre and Blasius (2006), Greenacre

(2007), Lebart et al. (1984), Murtagh (2005) or Govaert (2009) as well as
in Husson et al. (2010). The last book also details different case studies
conducted using FactoMineR.
230 R for Statistics

10.3 Multiple Correspondence Analysis

10.3.1 Objective
The aim of multiple correspondence analysis (MCA) is to summarise and visu-
alise a data table where the individuals are described by qualitative variables.
MCA is used to study the similarities between individuals from the point of
view of all the variables and identify individuals profiles. It is also used to as-
sess the relationships between the variables and study the associations of the
categories. Finally, as with PCA and CA, the individuals or groups of individ-
uals (rows) can be characterised by the categories of the variables (columns).
If certain variables from the data table are quantitative, it is possible to in-
clude them in the analysis by dividing them into classes (Section 2.3.2, p. 35).
To conduct MCA, we use the MCA function from the FactoMineR package spe-
cially designed for multivariate exploratory data analysis (see appendix A.4,
p. 269).

10.3.2 Example
The dataset contains information from sixty-six clients who took out loans
from a credit company. The eleven qualitative variables and the associated
categories are as follows:
Loan: Renovation, Car, Scooter, Motorbike, Furnishings, Sidecar. This
variable represents the item for which clients took out a loan.
Deposit: yes, no. This variable indicates whether or not clients paid a
deposit before taking out the loan. A deposit represents a guarantee for
the loan organisation.
Unpaid: 0, 1 or 2, 3 and more. This variable indicates the number of
unpaid loan repayments for each client.
Debt load: 1 (low), 2, 3, 4 (high). This variable indicates the clients
debt load. The debt load is calculated as the ratio between costs (sum of
expenses) and income. The debt load is divided into four classes.
Insurance: No insurance, DI (Disability Insurance), TPD (Total Perma-
nent Disability), Senior (for people older than 60). This variable indicates
the type of insurance the client has taken out.
Family: Common-law, Married, Widower, Single, Divorcee.
Children: 0, 1, 2, 3, 4 and more.
Accommodation: Home owner, First-time buyers, Tenant, Lodged by fam-
ily, Lodged by employer.
Multiple Correspondence Analysis 231

Profession: Technician, Manual Labourer, Retired, Management, Senior

Management.

Title: Mr., Mrs., Miss.

Age: 20 (18 to 29 years old), 30 (30 to 39), 40 (40 to 49), 50 (50 to 59),
60 and over.
The aim of this study is to characterise the credit companys clientele.
Firstly we want to identify the different banking behaviour profiles, that is,
to typify the individuals. We then study the relationship between the la-
bels (profession, age, etc.) and the dimensions of variability for the banking
behaviour profiles (i.e. characterise the clients with specific behaviours).

10.3.3 Steps
1. Read the data.
2. Choose the active individuals and variables.

3. Choose the number of dimensions.

4. Analyse the results.
5. Automatically describe the dimensions of variability.
6. Back to raw data in cross tabulation.

10.3.4 Processing the Example

1. Reading the data:

> Credit <- read.table("Credit.csv",sep=";", header=TRUE)

By examining the results yielded by the summary function, we remark that the
Age variable is not considered qualitative. It therefore needs to be converted:

> Credit[,"Age"] <- factor(Credit[,"Age"])

In MCA it is important to check that there are no rare categories, that is,
those with very small frequencies, as MCA attributes a lot of importance to
these categories. To do this, we recommend representing each variable:

> for (i in 1:ncol(Credit)){

+ par(ask=TRUE) # plot graphs one-by-one
+ plot(Credit[,i])
}
232 R for Statistics

If certain variables admit categories with small frequencies, there are a number
of ways to deal with it:

Naturally grouping together certain categories; we recommend this solu-

tion for ordinal categories (see Exercise 2.7).
Ventilation (random allocation) of the individuals associated with rare
categories in the other categories (see Exercise 2.6); certain programs
offer automatic ventilation (see below).
Elimination of individuals with rare categories.
In this example, one single individual carries the category Sidecar for the
Loan variable. It is therefore only natural to group this category along with
Motorbike (see Section 2.3.3, p. 37):

> levels(Credit[,"Loan"])[6] <- "Motorbike"

2. Choosing the active individuals and variables:

We want to determine banking behaviour profiles and will thus indicate as ac-
tive those variables corresponding to banking information (the first five). The
choice of active variables is very important as only these variables contribute
to the construction of the MCA dimensions, and thus only these variables
are used to calculate the distances between individuals. As supplementary
variables, we add the other qualitative variables (corresponding to the ques-
tions about general information for the clients). These variables are useful
in interpreting the results. It is also possible to add quantitative variables as
supplementary (without having to divide them into classes in advance).
We can also choose the individuals who participate in constructing the di-
mensions, known as active individuals. Here, as is often the case, all the
individuals are considered active.
To conduct the MCA using the FactoMineR package (see Appendix A.4, p. 269),
the preinstalled package must be loaded (see Section 1.6). We then use the
MCA function.

> library(FactoMineR)
> res.mca <- MCA(Credit, quali.sup=6:11, level.ventil=0)

The MCA is constructed from the first five variables (the active variables)
whereas the variables 6 to 11 are supplementary. The argument level.ventil
is 0 by default, which means that no ventilation is conducted. If the argument
has a value of 0.05, this means that the categories with frequency equal to or
less than 5% of the total number of individuals are ventilated automatically
by the MCA function prior to constructing the MCA. By default, results are
given for the first five dimensions. To specify another number of dimensions,
use the argument ncp. The list res.mca contains all the results.
Multiple Correspondence Analysis 233

3. Choosing the number of dimensions:

There are many solutions available in order to determine the number of dimen-
sions to be analysed in MCA. The most well-known is to represent the barplot
of eigenvalues or of percentages of variance associated with each dimension.

> barplot(res.mca$eig[,2],names=paste("Dim",1:nrow(res.mca$eig)))
14
12
10
8
6
4
2
0

Dim 1 Dim 3 Dim 5 Dim 7 Dim 9 Dim 11 Dim 13

Figure 10.9
Percentage of variance associated with each dimension of the MCA.

The decrease in percentages of variance (Figure 10.9) is regular. There is no

particularly noticeable gap or break in the diagram, and it is therefore difficult
to choose the number of dimensions. However, these low levels of percentage
of variance and the decrease are common in MCA. To obtain the numerical
results, we use the following instruction:

> round(res.mca$eig[1:5,],2)
eigenvalue percentage of cumulative percentage
variance of variance
dim 1 0.40 15.33 15.33
dim 2 0.32 12.38 27.71
dim 3 0.29 11.25 38.96
dim 4 0.27 10.47 49.43
dim 5 0.24 9.14 58.57

The first two dimensions express 28% of the total variance. In other words,
28% of the information in the data table is summarised by the first two di-
mensions, which is relatively high in MCA. We will interpret only the first two
dimensions even though it may be interesting to analyse dimensions 3 and 4.
4. Analysing the results:
The MCA function yields a graph simultaneously representing individuals and
categories (active and illustrative) on the first two dimensions. This repre-
sentation quickly becomes saturated and it is therefore necessary to conduct
234 R for Statistics

separate representations for the individuals and the categories of the vari-
ables. For this we use the invisible argument of the plot.MCA function. To
represent the individuals only, we use

> plot(res.mca, invisible=c("var","quali.sup"))

MCA factor map

1.5

33
L
1.0

66 55 30
49
35 L
L10
7 6 L
24 31 L
L L L L

4
56
0.5

8
361
L
34 39 32
36 L47
L L
22 40
L L L L
46
2
L
L 44
25
L
Dim 2 (12.38%)

L 1
23 L
51
13 5
28 L 12
14
15 26
17
L
58 65
L
0.0

L
L 425345 L
54
38 LL L
L LL
L64
27 21 11 L
9
18 37
L L L
LL L 29 48
59 L L
L
0.5

16
L
63 57
L L

19 43
62 L L
52
1.0

L
41 L
L
50
20
L
60
L
1.5

1.5 1.0 0.5 0.0 0.5 1.0 1.5

Dim 1 (15.33%)

Figure 10.10
MCA of credit data: representation of the individuals.

This graph of individuals (Figure 10.10) illustrates the general shape of the
scatterplot to see if, for example, it identifies particular groups of individuals.
This is not the case here. To make the results easier to interpret, it can
be helpful to colour the individuals according to a variable, with one colour
used for each category of the variable. We thus use the habillage argument
specifying either the name or the number of the qualitative variable (graph
not provided):

> plot(res.mca, invisible=c("var","quali.sup"),habillage="Loan")

> plot(res.mca, invisible=c("var","quali.sup"),habillage=1)

The graph with all the categories of the active and illustrative qualitative
Multiple Correspondence Analysis 235

variables (Figure 10.11) can be used to get an idea of the overall tenden-
cies emanating from the analysis. To construct this graph and represent the
categories alone, we use

> plot(res.mca, invisible="ind")

MCA factor map

Unp_1_ou_2
Child_4
1.0

Senior Scooter
60
Retired Renovation
Widower Debt_3
Child_1
Home owner
0.5

Married Motorbike Senior management

DI
Unp_0 Mrs. Deposit
Mr. Child_2
Dim 2 (12.38%)

50 First-time buyers
Debt_2 No insurance Common-law
Management Technician Child_3
0.0

Debt_4
Child_0 30 Lodged by the employer
40 20
no_deposit Tenant Car
Divorcee
Debt_1
-0.5

Single
Manual Labourer
Furnishings
Lodged by the family
-1.0

Unp_3 and +
Miss TPD
-1.5

-2 -1 0 1 2
Dim 1 (15.33%)

Figure 10.11
MCA of credit data: representation of active and illustrative categories.

The first dimension of Figure 10.11 opposes a young profile (on the right)
with an old profile (on the left). We thus find older people, home-owners
who have taken out a loan to finance renovation work, opposing younger people
who have taken out a loan, for example, to buy a scooter. Young people
tend to have a relatively high debt load.
Dimension 2 mainly opposes people in great financial difficulty (below) with
the others (above). We can therefore identify people who have difficulty paying
back their loan (three or more payments missed) and those who have taken
out TPD insurance. This dimension can thus be qualified as the dimension of
financial difficulty.
Using a number of examples, we evoke the general rules for interpreting the
proximity between categories of two variables, be they the same or differ-
ent. For example, the category Senior of the variable Insurance is close to
236 R for Statistics

the category Retired of the (illustrative) variable Profession, thus indicat-

ing that many of the individuals who carry the category Senior also carry
the category Retired. We can also see that the category Lodged by the
employer and Tenant for the Accommodation variable are close, which means
that these people generally carry the same categories for the other variables.
The individuals who carry these categories thus have the same profile.
To interpret a dimension in more detail, we recommend analysing the nu-
merical data of the variables contained within the object res.mca$var. We
thus obtain a list with the coordinates of the categories on the dimensions
res.mca$var$coord, the squared cosines res.mca$var$cos2 (which mea-
sure the quality of the projection of the categories on the dimensions) and
the contributions of the categories to the construction of the dimensions
res.mca$var$contrib. We also obtain res.mca$var$eta2 which, for each
dimension, corresponds to the correlation ratio 2 between the principal com-
ponent (the coordinates of the individuals on the dimension) and each qual-
itative variable. The individuals numerical results are contained within the
object res.mca$ind. The coordinates of the individuals (scores or principal
components) are contained within the object res.mca$ind$coord.
To know which variables are most related to the dimensions, that is, to in-
terpret the dimensions from the variables, we draw the graph of correlation
ratios between the dimensions and the variables (Figure 10.12):
> plot(res.mca, choix="var")

Variables representation
1.0
0.8
0.6
Dim 2 (12.38%)

Unpaid Insurance

Loan
0.4

Profession
Family Debt load
Accommodation Age
0.2

Title Children
Deposit
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Dim 1 (15.33%)

Figure 10.12
MCA of credit data: representation of active and illustrative variables.
Multiple Correspondence Analysis 237

The variable which contributes the most to the creation of dimension 1 is the
Loan variable, and that which contributes the most to the creation of dimen-
sion 2 is the Unpaid variable (Figure 10.12). This information summarises
the overall influence of all the categories of each of the variables to the con-
struction of the dimensions. In this way, the categories of the Loan variable
contribute greatly to the creation of dimension 1 and the categories of the
Unpaid variable to the creation of dimension 2. We can then examine the
results concerning the supplementary qualitative variables contained within
the object res.mca$quali.sup.
It is possible to construct the graphs of individuals and variables on dimensions
3 and 4:

> plot(res.mca, invisible="ind", axes = 3:4)

> plot(res.mca, invisible=c("var","quali.sup"), axes=3:4)
> plot(res.mca, choix="var", axes=3:4)

To save a graph in pdf format for example, we use the argument new.plot=F:

> pdf("mypath/mygraph.pdf")
> plot(res.mca,invisible=c("var","quali.sup"),axes=3:4,
new.plot=FALSE)
> dev.off()

5. Automatically describing the dimensions of variability:

In the FactoMineR package there is a function by which the dimensions of
MCA can be described automatically: dimdesc. This function is a new aid to
interpretation. It sorts the supplementary quantitative variables according to
their coordinates on a dimension, that is to say, the correlation coefficient be-
tween the quantitative variable and the MCA dimension. Only the significant
correlation coefficients are retained. This function also sorts the qualitative
variables as well as the categories of the qualitative variables. In order to do
so, a one-way analysis of variance model (see Section 8.1) is constructed for
each qualitative variable: variable Y corresponds to the principal component
(the coordinates of the individuals or scores) and is explained according to a
qualitative variable. The p-value of the overall test (F -test) is calculated as
well as theP p-values of the tests of each category (test H0 : i = 0 with the
constraint i i = 0). Only the significant results are given. The p-values
associated with the F -tests are sorted in ascending order. Thus, the qualita-
tive variables are sorted from most to least characteristic. In the same way,
the p-values of the t-tests by categories are sorted (in ascending order when
the coefficient is positive, and in descending order when it is negative). This
will identify the most characteristic categories. Below is a description of the
second dimension of MCA from the variables and the categories:

> dimdesc(res.mca)
238 R for Statistics

$Dim 2
$Dim 2$quali
R2 p.value
Unpaid 0.4679498 2.330816e-09
Insurance 0.4620452 1.979222e-08
Loan 0.3895658 3.732164e-06
Profession 0.2440563 1.661355e-03
Debt.load 0.1865282 4.857269e-03
Deposit 0.1033137 8.498296e-03
Family 0.1945394 9.447136e-03
Age 0.1781788 1.618904e-02
Accommodation 0.1734846 1.883183e-02
Title 0.1101503 2.532061e-02

$Dim 2$category
Estimate p.value
Unp_1_ou_2 0.5997078 6.148708e-07
Senior 0.5149935 2.212144e-05
Retired 0.4477685 1.257211e-03
60 0.4556995 1.538945e-03
Debt_3 0.3550533 1.640259e-03
Renovation 0.3127094 4.448861e-03
Home owner 0.3432251 6.371855e-03
Deposit 0.1826945 8.498296e-03
Scooter 0.4031618 1.373824e-02
DI 0.1652700 4.212931e-02
Widower 0.3640249 4.252309e-02
Lodged by the family -0.4156119 3.553903e-02
Child_0 -0.2990457 2.172157e-02
Single -0.3322199 1.086364e-02
Debt_1 -0.2911376 1.009127e-02
no_deposit -0.1826945 8.498296e-03
Miss -0.4824068 8.440543e-03
Car -0.2954665 8.029752e-03
Manual Labourer -0.4605202 1.323161e-03
Furnishings -0.5179085 1.049530e-05
TPD -0.6985157 3.360963e-09
Unp_3 and + -0.6634535 7.988517e-10
This function becomes very useful when there are a lot of categories. Here
we find that the second dimension is mainly linked to the categories TPD and
Unp 3 and +.
6. Going back to raw data in cross-tabulation:
It is interesting to go back to the raw data to analyse more closely the re-
lationship between two variables and particularly the proximity between the
Multiple Correspondence Analysis 239

categories of these variables. We can thus construct cross-tabulations using

the table command, calculate row and column percentages and conduct a
chi-square test (see Worked Example 6.2).

10.3.5 Rcmdr Corner

A graphical interface for FactoMineR is also available in Rcmdr which makes
it possible to conduct the previous analyses with an easy-to-use drop-down
menu. To do this, simply install the drop-down menu once by launching the
following command in an R window (Internet connection required):
> source("http://factominer.free.fr/install-facto.r")
In following sessions, the FactoMineR drop-down menu is present in Rcmdr
and you simply need to type library(Rcmdr).

1. Import the data:

Data Import data from text file, clipboard, or URL ...
26. We will not describe the choice of active elements but simply show how
to conduct MCA in Rcmdr.
Click on the FactoMineR tab and then Multiple Correspondence analysis.
In the main MCA window, it is possible to select supplementary qualita-
tive variables (Select supplementary factors), supplementary quantita-
tive variables (Select supplementary variables) and supplementary indi-
viduals (Select supplementary individuals). By default, results on the
first five dimensions are provided in the object res. It is preferable to click
Apply rather than Submit, as it launches the analysis whilst also keeping the
window open, and it is thus possible to modify certain options without having
to reset all the parameters.
The graph options window and the other output options window are similar
to those for PCA. In graph options it is possible to omit the representation of
certain elements (individuals or categories) or labels.
The window for the output options can be used to visualise the different results
(eigenvalues, individuals, variables, automatic description of the dimensions).
All the results can also be exported into a csv file (which can be opened in a
spreadsheet such as Excel, for example).

10.3.6 Taking Things Further

1. MCA as a pre-processing method to transform qualitative variables into
quantitative variables:
The principal components (coordinates of the individuals or scores) of the
MCA are continuous quantitative variables which summarise a set of qualita-
tive variables and which can be used as input of unsupervised classification,
240 R for Statistics

for example. In this way, MCA can be seen as a pre-processing step that
can convert qualitative variables into quantitative ones. To implement unsu-
pervised classification on the principal components of the MCA, see Worked
Example 11.1.
2. Characterising the categories of one specific qualitative variable:
It can also be interesting to use the catdes function (see Worked Example 11.1)
to describe one specific qualitative variable according to quantitative and/or
qualitative variables.
3. Constructing confidence ellipses around categories:
Confidence ellipses can also be drawn around the categories of a categorical
variable (i.e. around the barycentre of the individuals carrying that category).
These ellipses allow one to visualise whether two categories are significantly
different or not. It is possible to construct confidence ellipses for all the
categories for a number of categorical variables using the plotellipses function:

> plotellipses(res.mca,keepvar=c("Loan","Unpaid","Insurance",
"Profession"))

4. Handling missing values in MCA:

It is possible to perform MCA with missing values using the missMDA pack-
age. The first step consists of imputing the missing values in the indicator
matrix of dummy variables (or disjunctive data table) using the imputeMCA
function and the second step consists of performing MCA using this completed
indicator matrix of dummy variables:

> tab.disj.comp <- imputeMCA(my.incomplete.dataset)

> res.mca<-MCA(my.incomplete.dataset, tab.disj=tab.disj.comp)

Theoretical reminders on MCA are available Greenacre and Blasius (2006),

Greenacre (2007), Lebart et al. (1984) or Govaert (2009) as well as in Husson
et al. (2010). The last book also details different case studies using Fac-
toMineR.
11
Clustering

11.1 Ascending Hierarchical Clustering

11.1.1 Objective
Ascending hierarchical clustering (AHC) constructs a hierarchy of individuals
that is graphically represented by a hierarchical tree also named a dendro-
gram. Pruning this tree yields groups (clusters) of individuals. Hierarchical
clustering requires to define a distance and an agglomerative criterion. Many
distances are available (Manhattan, Euclidean, etc.) as well as several ag-
glomeration methods (Ward, single, centroid, etc.).
To perform AHC, we use the agnes function from the cluster package. Us-
ing this function, it is possible to have as an input a dissimilarity matrix or
a data table individuals quantitative variables (this is not possible with
the hclust function). We here present an example in which the individuals
are characterised by quantitative variables (the case of qualitative variables is
mentioned in Section 11.1.6, Taking Things Further). Since we often com-
bine principal component methods and hierarchical clustering to enrich the
description of the data, we chose the Euclidean distance and the Wards ag-
glomerative criterion. Indeed, principal component methods are based on the
notion of variance, which is related to Euclidean distance and Ward criterion.

11.1.2 Example
Let us again examine the dataset of the athletes participating in the decathlon
(see Worked Example 10.1) and construct an ascending hierarchical cluster-
ing of the athletes performance profiles. We are only interested in the ten
performance variables.

11.1.3 Steps
1. Read the data.
2. Standardise the data if necessary.

3. Construct the ascending hierarchical clustering.

242 R for Statistics

4. Prune the hierarchical tree.

5. Characterise the clusters.

11.1.4 Processing the Example

1. Reading the data:

> decath <- read.table("decathlon.csv",sep=";",header=TRUE,

row.names=1)
2. Standardising the data if necessary:
Prior to performing an AHC, the variables can be standardised. For the
decathlon dataset, we do not have the choice. Standardisation is essential as
the variables are of different units. When the variables are of the same units,
both solutions are possible and lead to two separate analyses. This decision is
therefore crucial. Standardisation means attributing the same importance to
each of the variables when calculating the distance between individuals. Not
standardising the data means attributing more importance to the variables
with a high standard deviation. Here, the data is standardised using the scale
function.
3. Constructing the ascending hierarchical clustering:
To perform an AHC using the cluster package, the preinstalled package must
be loaded (see Section 1.6). We then use the agnes function:

> library(cluster)
> res.ahc <- agnes(scale(decath[,1:10]), method= "ward")
> plot(res.ahc, which.plots=2, main="Dendrogram",
xlab="Individual")

The agnes function takes as an input the (standardised or non-standardised)

dataset. The argument method specifies the aggregation method and is set
here to Ward. By default, the argument metric is set to the Euclidean dis-
tance. The plot.agnes function draws the hierarchical tree (or dendrogram)
(see Figure 11.1).
4. Pruning the hierarchical tree:
Given the form of the hierarchical tree (Figure 11.1), we can prune it to con-
struct clusters of individuals. In order to do so, we must first define the height
at which the tree will be cut. For example, a cut height of 10 (Figure 11.1)
defines two clusters whereas a height of 8.1 defines four clusters.
Different methods are available to choose the number of clusters. Here, we
use the heights at which two clusters join together as a guide. To obtain these
heights in a numerical vector, change the object res.ahc into a hclust-type
object and extract its height. As these heights are ranked in ascending order,
we inverse this order (rev):
Ascending Hierarchical Classification 243
Dendrogram

10
8
6
Height

Casarsa
4

Karpov

MARTINEAU
Smith

BOURGUIGNON
Turi
Terek
Korkizoglou
2

CLAY
KARPOV
Macey

Parkhomenko

HERNU
Warners

Schoenbeck

Pogorelov
SEBRLE

Uldal
Drews
Sebrle
Clay

Barras
Gomez
Hernu
ZSIVOCZKY

Karlivans
BARRAS
Lorenzo
NOOL
Averyanov
Ojaniemi

Zsivoczky
YURKOV
Bernard
McMULLEN

Smirnov
Qi

Schwarzl
WARNERS

Nool
BERNARD
0

Individual
Agglomerative Coefficient = 0.76

Figure 11.1
Hierarchical tree.

> res.ahc2 <- as.hclust(res.ahc)

> plot(rev(res.ahc2$height),type="h",ylab="Height")

The height of the fortieth and final bar (Figure 11.2) gives an idea of how
difficult it would be to bring together the first individuals and form a clustering
of forty clusters. The height of the thirty-ninth bar also gives an idea of the
difficulty in changing from this forty-cluster clustering to a thirty-nine-cluster
clustering. This change is obtained either by merging two individuals, or by
aggregating an individual to an existing cluster. In the same way, the height
of the first bar gives an idea of the difficulty in aggregating the two remaining
clusters to group the n = 41 individuals together.
As we are looking for a small number of homogeneous clusters, we are inter-
ested in the heights of the first bars. The most marked jump separates the
fifth and sixth bars: it is therefore difficult to go from six to five clusters.
One possibility for the cut would therefore be to take six groups. Between
the fourth and the third bar there is another less pronounced jump. In order
to have enough individuals in each cluster, we chose to use this jump which
defines four clusters.
We now need to find out which individuals are in each of these four clusters.
244 R for Statistics

10
8
Height

6
4
2

0 10 20 30 40

Index

Figure 11.2
Barplot of the heights as a help to choose the number of clusters.

To do this, we use the cutree function and indicate the cut, that is, the number
of clusters into which the individuals are divided (for example, k=4).

> clusters.hac <- cutree(res.ahc, k=4)

> clusters.hac
[1] 1 1 1 2 2 3 2 2 2 2 3 2 2 2 2 2 2 2 2 3 3
[22] 2 3 4 4 3 4 3 3 3 3 2 3 2 2 2 4 4 4 4 4

5. Characterising the clusters:

Having grouped the individuals into clusters, it is interesting to describe these
clusters in order to interpret the similarities and differences between these
groups of individuals. It is possible to describe the clusters using variables with
the catdes function of the FactoMineR package (see Appendix A.4, p. 269).
This function allows one to describe a qualitative variable (more precisely,
its categories) from quantitative and qualitative variables. Here, we want
to describe the clusters which correspond to the categories of the qualitative
variable clusters.hac. This variable is thus transformed into a factor. Then,
to use the catdes function, we construct a unique dataset with the raw data
and the cluster variable named "Cluster". The function is then applied on
the new dataset, and we specify with the argument num.var the index of the
variable which will be described.
Ascending Hierarchical Classification 245

> library(FactoMineR)
> clusters.hac <- as.factor(clusters.hac)
> decath.comp <- cbind.data.frame(decath,clusters.hac)
> colnames(decath.comp)[14] <- "Cluster"
> catdes(decath.comp, num.var = 14)
$quanti
$quanti$2
Mean in Overall sd in Overall
v.test category mean category sd
100m -2.265855 10.89789 10.99805 0.1701572 0.259796
110m.H -2.397231 14.41579 14.60585 0.3097931 0.466000
400m -2.579590 49.11632 49.61634 0.5562394 1.139297
1500m -2.975997 273.18684 279.02488 5.6838942 11.530012

$quanti$3
Mean in Overall sd in Overall
v.test category mean category sd
1500m 3.899044 290.76364 279.02488 12.627465 11.530012
400m 2.753420 50.43546 49.61634 1.272588 1.139297
Long. -2.038672 7.09364 7.26000 0.283622 0.312519

The catdes function sorts the quantitative variables from the most to the least
characteristic as positives (variables for which the individuals of the cluster
carry values which are significantly higher than the mean for all the individu-
als), then from the least to the most characteristic as negatives (variables for
which the individuals of the cluster carry values which are significantly lower
than the mean for all the individuals). The variables are sorted according to
the v-tests defined as
xq x
v-test = r  
s2 IIq
Iq I1

where xq is the average of variable X for the individuals of cluster q, x is

the average of X for all the individuals, and Iq is the number of individuals
within the cluster q. A v-test with an absolute value of greater than 2 here
indicates that the cluster mean is significantly different from the general mean:
a positive sign (or negative, respectively) for the v-test indicates that the mean
of the cluster is superior (or inferior, respectively) to the general mean. In
the example, the results are given for the description of clusters 2 and 3.
The individuals from cluster 2 run the 1500 m faster (in a shorter time) than
the others. For the individuals in this cluster, the average time taken to
run 1500 m is 273.2 s compared with 279.0 s for all the individuals. Three
other variables characterise the individuals in this cluster significantly (v-test
greater than 2 in absolute value): the 400 m, the 110-m hurdles and the 100 m.
These athletes run faster (shorter times) in these events.
For the qualitative variables, it is the categories of the variables which are
246 R for Statistics

sorted. These categories are sorted from most to least characteristic when the
category is over-represented in the cluster (compared to other clusters) and
from least to most characteristic when the category is under-represented in
the cluster. In the example, none of the qualitative variables characterise the
clusters 2 and 3.

11.1.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
It must next be specified that the first column contains the individuals iden-
tifiers, which is not possible using the drop-down menu. We therefore need to
copy the line of code generated by Rcmdr (top window) and add row.names=1
to the reading instruction before clicking on Submit.
2. Standardising the data if necessary:
Data Manage variables in active data set
Standardize variables... Then select the variables to be standardised;
these variables are added to the raw dataset.
3. Constructing the tree:
Statistics Dimensional analysis Cluster analysis
Hierarchical cluster analysis...
We must then select the variables which will enable us to calculate the dis-
tances between the individuals (here the standardised variables). Be aware
that Rcmdr uses the hclust function to conduct the ascending hierarchical clus-
tering. We then need to choose the agglomerative criterion (for example, the
Wards criterion) and the associated distance (with Ward the Euclidean dis-
tance squared, which corresponds to the argument of the Euclidean distance
for the agnes function).
4. Pruning the tree:
Statistics Dimensional analysis Cluster analysis
Add hierarchical clustering to data set...
Specify the number of clusters (groups) of individuals to be constructed (here
4). After validation, the dataset contains a supplementary qualitative variable
which corresponds to the affiliation to the clusters (by default, the name of
the variable is hclus.label).
5. Characterising the clusters:
First, load the FactoMineR package and drop-down menu (see Appendix A.4
p. 269). It is thus possible to characterise the clusters of individuals by click-
ing FactoMineR Description of categories. Then simply choose the
variable hclus.label to be characterised. By default, all the quantitative
and qualitative variables are used to characterize the cluster variable.
Ascending Hierarchical Classification 247

11.1.6 Taking Things Further

In this section we describe some extensions of the clustering.

1. Representing the clusters on a principal component graph:

It is possible to colour the individuals on a principal component map according
to the cluster with which they are affiliated.
As a reminder, we have performed PCA on the decathlon data with the first
ten variables as active (those used to perform the clustering), variables 11 and
12 as quantitative supplementary, and variable 13 as qualitative supplemen-
tary. To colour the individuals according to the cluster, we perform a PCA
specifying that variable 14 is also qualitative supplementary (variable 14 cor-
responds to the affiliated cluster). We then construct the graph (plot.PCA
function) specifying that the individuals are coloured according to variable
14.

> res.pca <- PCA(decath.comp, quanti.sup = 11:12,

quali.sup = 13:14, graph=F)
> plot(res.pca, choix = "ind", habillage=14)

2. Clustering from qualitative variables:

If the individuals are described by qualitative variables, there are two possible
solutions:
Construct an appropriate distance or dissimilarity like the Jaccard simi-
larity coefficient, for example (see Kaufman and Rousseeuw, 1990);
Transform the qualitative variables into quantitative variables using a
multiple correspondence analysis (see Worked Example 10.3). Then re-
trieve the coordinates of the individuals (principal components or scores)
and perform the AHC on these coordinates (see below for the example
of running together PCA-AHC).
3. Clustering from the principal components:
It can be interesting to perform a principal component method (PCA or MCA)
prior to performing a clustering in order to omit the information contained
within the last dimensions which can be considered as noise. In the present
example, as all data is quantitative, the principal component method is a PCA.
We here preserve the information contained within the first eight dimensions
using the argument ncp=8 of the PCA, which represents 96% of the total
variance. We then perform the AHC on the coordinates of the individuals:

> library(FactoMineR)
> res.pca <- PCA(decath[,1:10], graph=FALSE, ncp=8)
> res.ahc <- agnes(res.pca$ind$coord, method="ward")
248 R for Statistics

Traditionally, the aggregation method used when running an AHC on prin-

cipal components is the Wards agglomerative criterion (criterion based on
variance). The same procedure can be used with MCA.
Remark
When we conduct a clustering on the principal components, the new data
table (the coordinates of the individuals) must under no circumstances be
standardised or the distances between individuals will be distorted.

4. Consolidating a clustering:
It can be interesting to consolidate the clustering with a partitioning method
such as k-means (see Worked Example 11.2), specifying the number of clusters
to be constructed. In order to do this, we initialise the centres of the clusters
with the barycentres of the constructed clusters using ascending hierarchical
clustering. We thus calculate the means for each of the variables (centred and
standardised as the AHC is conducted on centred and standardised data) for
each cluster using the aggregate function:

> centre.clusters <- aggregate(scale(decath[,1:10]),

by=list(clusters.hac),FUN=mean)
> centre.clusters
Group.1 100m Long.jump Shot.put High.jump ...
1 1 -1.526034 1.9279293 1.653179 1.272289 ...
2 2 -0.380781 0.2312184 -0.123492 -0.106361 ...
3 3 0.138758 -0.5257989 0.485427 0.393352 ...
4 4 1.285824 -0.5491438 -0.994111 -0.765361 ...

The kmeans function takes as its argument the dataset and the number of
clusters or the centres of the clusters.

> clusters.kmeans <- kmeans(scale(decath[,1:10]),

centers=centre.clusters[,-1])
> clusters.kmeans$cluster
Sebrle Clay Karpov Macey Warners ...
1 1 1 1 2 ...

Here the argument centers=centre.clusters[,-1] initialises the centres of

the clusters (the first column of centre.clusters is omitted as this column
corresponds to the number of the group and must not be used). We retrieve
the new partition within the object cluster.
5. Clustering with the HCPC function:
The four-point process detailed above are integrated in the HCPC (for hi-
erarchical clustering on principal components) function of the FactoMineR
package. This function combines principal component methods, hierarchical
Ascending Hierarchical Classification 249

clustering and partitioning, to better describe and visualise the data. The
HCPC function performs a hierarchical clustering (with the Euclidean dis-
tance and the Wards agglomerative criterion) on the principal components
obtained with principal component methods (PCA for quantitative variables
and MCA for qualitative ones). The first step consists of performing the prin-
cipal component method, here a PCA, and then applying the HCPC function
on the object res.pca:

> res.pca <- PCA(decathlon, quanti.sup=11:12, ncp=Inf,

quali.sup=13,graph=F)
> res.hcpc <- HCPC(res.pca, consol=FALSE)

If the argument ncp=Inf, it means that all the principal components are
preserved, which amounts to constructing the clustering from the (scaled)
raw data. When the argument ncp is less than the number of variables, some
principal components are removed which remains to denoise the data. The
consol=FALSE argument is used to specify that no consolidation (see item 4)
is done after pruning the tree. The HCPC function yields an interactive graph
(Figure 11.3) with the hierarchical tree from which it is possible to choose a
cut height by clicking on the tree. An optimal cut height is also suggested
by a horizontal line (here at a height of around 0.80). On the same device,
another graph is drawn illustrating increases in within-cluster variability from
one partition with k + 1 clusters and one partition with k clusters.
Remark that the graphs provided by the agnes function and the HCPC func-
tion are different. Indeed, compared to the graph provided by the agnes func-
tion, the individuals are sorted according to their similarity (i.e. according
to their coordinates on the first principal component). It avoids positioning
very different individuals side-by-side on the hierarchical tree if they belong
to very different clusters. Furthermore, the individuals and the clusters are
grouped together according to variance, whereas with the agnes function they
are grouped according the square root of the variance. Thus, the hierarchical
tree for the HCPC function tends to be flattened: it is thus more difficult
to distinguish between the clusters for the first groupings of individuals but
easier for the last groups of clusters (i.e. the top of the hierarchical tree) and
it is thus easier to choose the height of the cut.
After having clicked on the graph, the hierarchical tree is represented in three
dimensions on the first two principal components (Figure 11.4). Individuals
are coloured according to their cluster.
The clusters (obtained after the hierarchical tree has been pruned) are de-
scribed by the variables, the dimensions of the principal component method
and by the individuals using the following commands:

> res.hcpc$desc.var
> res.hcpc$desc.axes
> res.hcpc$desc.ind
250 R for Statistics

1.2
Hierarchical Clustering

0.8
0.4
0.0
Click to cut the tree
inertia gain
1.5
1.0
0.5
0.0

BOURGUIGNON
Uldal
Karlivans
BARRAS
HERNU
MARTINEAU
Lorenzo
NOOL
Casarsa
Turi
Pogorelov
SEBRLE
Parkhomenko
YURKOV
Zsivoczky
Korkizoglou
Terek
CLAY
KARPOV
BERNARD
Nool
Schoenbeck
Drews
Schwarzl
WARNERS
Warners
Averyanov
Ojaniemi
Smirnov
Qi
Gomez
Barras
ZSIVOCZKY
Hernu
Smith
McMULLEN
Bernard
Macey
Clay
Sebrle
Karpov

Figure 11.3
Hierarchical tree provided by HCPC.

We find here the same description of the clusters as the one given by the catdes
function. Moreover, the description by the individuals gives, in the object
res.hcpc$desc.ind$para, the paragon of each cluster (i.e. the individual
closest to the centre of each cluster) and its distance to the centre of the
cluster.

Reminders of AHC are available in Kaufman and Rousseeuw (1990), Lebart

et al. (1984), Hastie et al. (2009, p. 520528), Murtagh (2005) or Govaert
(2009), as well as in Husson et al. (2010). This last book details case studies
and the results of a clustering performed after a principal component method.
Ascending Hierarchical Classification 251

Hierarchical clustering on the factor map

cluster 1
cluster 2
cluster 3
cluster 4
2.0
1.5

Casarsa 4
3
Korkizoglou
Parkhomenko YURKOV 2
height
1.0

HERNU Zsivoczky Macey Sebrle

MARTINEAU Pogorelov Smith Clay 1

Dim 2 (17.37%)
Uldal SEBRLE
Turi Barras CLAY KARPOV Karpov
BOURGUIGNON Terek McMULLEN 0
BARRAS Schoenbeck Bernard
Karlivans Qi Hernu -1
0.5

Lorenzo BERNARD Ojaniemi

Smirnov Nool Warners
WARNERS -2
NOOL
-3
Drews
-4
0.0

-4 -2 0 2 4 6
Dim 1 (32.72%)

Figure 11.4
Three-dimensional tree on the first two dimensions of the PCA provided by
HCPC.
252 R for Statistics

11.2 The k-Means Method

11.2.1 Objective
The k-means method provides a partition of individuals, that is, K clusters of
individuals which are as homogeneous as possible. The partitioning method
k-means is based on the Euclidean distance between individuals and uses the
barycentre of each cluster as a representative individual. k-means requires
to define the number of clusters K. In order to do so, we can, amongst
other things, construct an ascending hierarchical clustering and determine a
natural cut level (see Worked Example 11.1).
The method first consists of choosing (randomly or not) K initial cluster
centres and then it consists of repeating (until the centres no longer change)
the two steps: (1) (re)assign each individual to the cluster to which the indi-
vidual is closest; (2) update the cluster means, that is, calculate the barycentre
of the individuals for each cluster. To perform k-means, we use the kmeans
function. The input of this function is a data table individuals quantita-
tive variables (the case of qualitative variables is mentioned in Section 11.2.6,
Taking Things Further).

11.2.2 Example
Let us again examine the dataset of the athletes participating in the decathlon
(see Worked Example 10.1) and construct a partition of the athletes per-
formance profiles. We are therefore only interested in the ten performance
variables.

11.2.3 Steps
1. Read the data.
2. Standardise the variables if necessary.

3. Construct the partition.

4. Characterise the clusters.

11.2.4 Processing the Example

1. Reading the data:

> decath <- read.table("decathlon.csv",sep=";",header=TRUE,

row.names=1)

2. Standardising the variables if necessary:

The k-Means Method 253

Prior to performing a k-means algorithm, the variables can be standardised.

For the decathlon dataset, we do not have the choice. Standardisation is
essential as the variables are of different units. When the variables are of
the same units, both solutions are possible and lead to two separate analyses.
This decision is therefore crucial. Standardisation means attributing the same
importance to each of the variables when calculating the distance between
individuals. Not standardising the data means attributing more importance
to the variables with a high standard deviation. Here, the data is standardised
using the scale function.
3. Constructing the partition:
Taking into account the ascending hierarchical clustering performed on this
dataset (see Worked Example 11.1), we here partition into K = 4 clusters:

> results.kmeans <- kmeans(scale(decath[,1:10]), centers=4)

The kmeans function takes as an input the (standardised or non-standardised)

dataset. Then, the argument centers can specify either the number of clusters
or the centre of each cluster. In the former case, the initialisation of the k-
means algorithm is random, that is, four individuals are chosen randomly as
starting centre of cluster whereas in the latter case, the centre of each cluster
is specified (see item 4 in Section 11.1.6, Taking Things Further).

> results.kmeans
$k$-means clustering with 4 clusters of sizes 11, 13, 5, 12

Cluster means:
X100m Long.jump Shot.put High.jump X400m ...
1 1.058138862 -0.8935708 -0.18280501 -0.27095958 1.1885477 ...
2 -0.419569671 0.4278885 -0.04683446 -0.29297229 -0.2955972 ...
3 -1.232016906 1.4285640 1.37419773 1.47464827 -0.9747632 ...
4 -0.002086435 -0.2396743 -0.35427380 -0.04867052 -0.3631204 ...

Clustering vector:
Sebrle Clay Karpov Macey Warners ...
3 3 3 3 2 ...

Within cluster sum of squares by cluster:

[1] 78.18352 73.25410 28.56075 63.17204
(between_SS / total_SS = 39.2 %)

In output, for each cluster and each variable, the function yields the means
of the individuals of the cluster, a vector featuring the cluster number of each
individual, and the within-cluster variability.
254 R for Statistics

4. Characterising the clusters:

Having grouped the individuals into clusters, it is interesting to describe these
clusters in order to interpret the similarities and differences between these
groups of individuals. It is possible to describe the clusters using variables with
the catdes function of the FactoMineR package (see Appendix A.4, p. 269).
This function allows one to describe a qualitative variable (more precisely,
its categories) from quantitative and qualitative variables. Here, we want to
describe the clusters which correspond to the categories of the qualitative
variable results.kmeans$cluster. This variable is thus transformed in a
factor. Then, to use the catdes function, we construct a unique dataset with
the raw data and the cluster variable named "Cluster". The function is then
applied on the new dataset, and we specify with the argument num.var the
index of the variable which will be described.

> library(FactoMineR)
> decath.comp <- cbind.data.frame(decath,
factor(results.kmeans$cluster))
> colnames(decath.comp)[14] <- "Cluster"
> catdes(decath.comp, num.var = 14)
$quanti
$quanti$2
Mean in Overall sd in Overall
v.test category mean category sd
Pole.vault 4.452686 5.046154 4.76244 0.176354 0.274589

$quanti$4
Mean in Overall sd in Overall
v.test category mean category sd
1500m -3.247660 269.82083 279.02488 5.941016 11.530012
Pole.vault -3.320412 4.538333 4.76244 0.158000 0.274589

The catdes function sorts the quantitative variables from the most to the least
characteristic as positives (variables for which the individuals of the cluster
carry values which are significantly higher than the mean for all the individu-
als), then from the least to the most characteristic as negatives (variables for
which the individuals of the cluster carry values which are significantly lower
than the mean for all the individuals). The variables are sorted according to
the v-tests defined as
xq x
v-test = r  
s2 IIq
Iq I1

where xq is the average of variable X for the individuals of cluster q, x is

the average of X for all the individuals, and Iq is the number of individuals
within the cluster q. A v-test with an absolute value of greater than 2 here
indicates that the cluster mean is significantly different from the general mean:
The k-Means Method 255

a positive sign (or negative, respectively) for the v-test indicates that the mean
of the cluster is superior (or inferior, respectively) to the general mean.
In the example, the results are given for the description of clusters 2 and 4.
The individuals in cluster 2 jump high in the pole vault. For the individuals
in this cluster, the average height is 5.05 m compared with 4.76 m for all the
individuals (including those from cluster 2). The individuals in cluster 4 are
characterised by the fact that they do not jump as high as the others in the
pole vault (v-test less than 2) and that they run the 1500 m faster (in a
shorter time) than the others.
For the qualitative variables, it is the categories of the variables which are
sorted. These categories are sorted from most to least characteristic when
the category is over-represented in the cluster (compared to other clusters)
and least to most characteristic when the category is under-represented in
the cluster. In the example, none of the qualitative variables characterise the
clusters 2 and 4.

11.2.5 Rcmdr Corner

1. Reading the data from a file:
Data Import data from text file, clipboard, or URL ...
We must specify that the first column contains the individuals identifiers,
which is not possible using the drop-down menu. It is therefore necessary to
retrieve the line of code generated by Rcmdr (upper window) and to add
row.names=1 to the reading data from file instruction before clicking on
Submit.
2. Standardising the variables if necessary:
Data Manage variables in active data set
Standardize variables... Then select the variables to be standardised;
these variables are added to the raw dataset.
3. Constructing the partition:
Statistics Dimensional analysis Cluster analysis
K-means cluster analysis...
We must then select the variables which will enable us to calculate the dis-
tances between the individuals (here the standardised variables) and spec-
ify the number of clusters. It is possible to add the cluster variable to the
dataset by specifying Assign clusters to the data set and to name the
Assignment variable (by default, the name is KMeans).
4. Characterising the clusters:
First, load the FactoMineR package and drop-down menu (see Appendix A.4,
p. 269). It is thus possible to characterise the clusters of individuals by click-
ing FactoMineR Description of categories. Then simply choose the
256 R for Statistics

variable KMeans to be characterised. By default, all the quantitative and

qualitative variables are used to characterize the cluster variable.

11.2.6 Taking Things Further

1. Representing the clusters on a principal component graph:
It is possible to colour the individuals on a principal component map according
to the cluster to which they are affiliated.
As a reminder, we have performed PCA on the decathlon data with the first
ten variables as active (those used to perform the clustering), variables 11 and
12 as quantitative supplementary and variable 13 as qualitative supplemen-
tary. To colour the individuals according to the cluster, we perform a PCA
specifying that variable 14 is also qualitative supplementary (variable 14 cor-
responds to the affiliated cluster). We then construct the graph (plot.PCA
function) specifying that the individuals are coloured according to variable
14.

> res.pca <- PCA(decath.comp, quanti.sup = 11:12,

quali.sup = 13:14, graph=F)
> plot(res.pca, choix = "ind", habillage=14)

2. Partitioning from qualitative variables:

If the individuals are described by qualitative variables, it is impossible to
construct a clustering using the method detailed above. However, as with
AHC, it is possible to come back to quantitative data by conducting a prior
multiple correspondence analysis (see Worked Example 10.3) and retrieving
the coordinates of the individuals on the principal components. Be careful:
in that case, data must in no circumstances be standardised or the distances
between individuals will be distorted.
3. Other partition methods:
Other partitioning methods are available since the distance between individ-
uals can be non-Euclidean and the representative object of each cluster can
be modified. For example, the pam function from the cluster package per-
forms a partition around a median individual in each group rather than the
barycentre. This function can have as an input a data table or a dissimilarity
matrix.

Theoretical reminders on partitioning methods are available in Lebart et al.

(1984) and Hastie et al. (2009).
Appendix

A.1 The Most Useful Functions

In this section we offer a non-exhaustive list of the most useful functions in
R. These functions are divided under different headings: generic functions,
numerical functions, data handling functions, distributions of probabilities,
basic statistical functions, advanced statistical functions, graphical functions,
reading and exportation functions, text management and finally a few other
miscellaneous useful functions. For each function we provide a concise defini-
tion and an example of its use.

A.1.1 Generic Functions

In R there are generic functions, that is, functions which can be called by the
same order but which yield different results depending on the class of object
to which they are applied. The main generic functions are as follows:

Function Description
print Writes the results (either all results or an extract)
plot Constructs a graph
summary Summarises the results of model fitting functions

There are many functions which begin with print, plot or summary, for
example, print.lm, print.PCA, print.rpart, etc. They can all be called using
the generic instruction print rather than by print.lm, print.PCA, print.rpart,
etc. However, to obtain help with a function which writes an rpart object
for example, we must write help("print.rpart").

A.1.2 Numerical Functions

Function Description
abs(x) Absolute value
sqrt(x) Square root
ceiling(x) Gives the smallest following integer: ceil-
ing(5.24)= 6, ceiling(5)= 5, ceiling(-5.24)=
5
258 R for Statistics

Function Description
floor(x) Gives the largest previous integer: floor(5.24)=
5, floor(5)= 5, floor(-5.24)= 5
trunc(x) Truncates the value of x to 0 decimal digits:
trunc(-5.24)= 5
round(x, Rounds to n decimal digits: round(5.236,
digits=n) digits=2)= 5.24
signif(x, Rounds to n total digits: signif(5.236,
digits=n) digits=2)= 5.2
cos(x), sin(x), Trigonometric functions
tan(x), acos(x),
cosh(x), acosh(x),
etc.
log(x) Natural logarithm
log10(x) Base 10 logarithm
exp(x) Exponential

A.1.3 Data Handling Functions

Function Description
c Concatenates within a vector
cbind Concatenates tables one next to the other (jux-
taposition in columns), see Section 2.5
cbind.data.frame Juxtaposes data-frames into columns, see Sec-
tion 2.5
rbind Juxtaposes tables in rows (Caution! This func-
tion puts one row on top of another without tak-
ing column names into account), see Section 2.5
rbind.data.frame Juxtaposes data-frames in rows; the column
names of the data-frames must be the same (the
columns are sorted in the same order for all the
tables in order to link the variables prior to con-
catenation), see Section 2.5
merge Merges the tables according to a key, see Sec-
tion 2.5
sort Sorts vectors in ascending order (or descending
order if decreasing = TRUE)
order Sorts a table according to one or more columns
(or rows): x[order(x[,3], -x[,6]), ] ranks
the table x depending on (ascending) the third
columns of x then, in the case of equality in the
third column of x, depending on (descending) the
sixth column of x
by(data, Applies the FUN function to each level of the vec-
INDICES, FUN) tor INDICES in the data table
Appendix 259

Function Description
dimnames Yields the names of the dimensions of an object
(list, matrix, data-frame, etc.)
rownames Yields the row names of a matrix
row.names Yields the row names of a data-frame
colnames Yields the column names of a matrix
col.names Yields the column names of a data-frame
names Yields the names of an object (list, matrix, data-
frame, etc.)
dim Yields the dimensions of an object
nrow or NROW Yields the number of rows in a table (in capitals,
yields a response even if the object is a vector)
ncol or NCOL Yields the number of columns in a table (in capi-
tals, yields a response even if the object is a vec-
tor)
factor Defines a vector as a factor (if ordered=TRUE the
levels of the factors are taken to be ordinal)
levels Yields the levels of a factor
nlevels Yields the number of levels of a factor
as.data.frame(x) Converts x into a data-frame
as.matrix(x) Converts x into a matrix
as.list(x) Converts x into a list
as.vector(x) Converts x into a vector
is.data.frame(x) Tests whether x is a data-frame
is.matrix(x) Tests whether x is a matrix
is.vector(x) Tests whether x is a vector
is.list(x) Tests whether x is a list
class(x) Yields the class of object x (matrix, data-frame,
list, etc.)
mode(x) Yields the mode of the object x (numeric, logic,
etc.)
as.character(x) Converts x into a character
as.numeric(x) Converts x into a numeric
as.integer(x) Converts x into an integer
as.logical(x) Converts x into a Boolean
is.character(x) Tests whether x is a chain of characters
is.numeric(x) Tests whether x is a numeric
is.integer(x) Tests whether x is an integer
is.logical(x) Tests whether x is a Boolean
260 R for Statistics

Function Description
which Yields the positions of the true values of
a vector or a logic table: the param-
eter arr.ind=TRUE yields the numbers of
the rows and columns in the table (Sec-
tion 2.4.1, p. 41): which(c(1,4,3,2,5,3) ==
3) yields 3 6; which(matrix(1:12,nrow=4)
==3,arr.ind=TRUE) yields (row 3, column 1)
which.min Yields the index of the minimum of a vector
which.max Yields the index of the maximum of a vector
is.na Tests whether the piece of data is missing
is.null Tests whether the piece of data is null
length Yields the length of a list or a vector
any Tests whether at least one value of a logic vector
is true: any(is.na(x)) yields TRUE if at least one
piece of data is missing in x
split(x,fac) Divides the table x according to the levels of fac

A.1.4 Probability Distributions

There are many probability distributions available in R. All the functions
begin with the letters d, p, q or r and end with the name of the distribution.
The letter d indicates that the function calculates the distribution, the
letter p indicates that the function calculates a probability, the letter q indi-
cates that the function calculates a quantile, and the letter r indicates that
the function generates a random number.
When we generate a random series of numbers it can be interesting to
again generate the same series of numbers at a later date. In this case, we
must indicate the random seed generator used to trigger the series, using
the command set.seed(1234) (the seed can be changed by using a different
integer than 1234).

Function Description
pnorm(q) Yields the probability P(X q) for X N (0, 1):
pnorm(1.96) = 0.975. To calculate P(X > q) we
use the argument lower.tail = FALSE
qnorm(p) Yields the quantile of order p of the N (0, 1) dis-
tribution: qnorm(0.975) = 1.96
pbinom(q, size, Yields the probability P(X q) for binomial dis-
prob) tribution B(size,prob): pbinom(5, 10, .5) =
0.623
ppois(q, lamda) Yields the probability P(X q) for X following
a Poisson distribution with parameter lambda
punif(q, min, Yields the probability P(X q) for X according
max) to a uniform distribution on [min, max]
Appendix 261

Function Description
pchisq(q, df) Yields the probability P(X q) when X follows
a chi-square distribution with df degrees of free-
dom
pt(q, df) Yields the probability P(X q) when X follows
a Students t distribution with df degrees of free-
dom
pf(q, df1, df2) Yields the probability P(X q) when X follows
a Fishers distribution with df1 and df2 degrees
of freedom
sample(x, size, This function is used to select a sample of size el-
replace = FALSE) ements of the vector x without replacement (with
replacement if replace=TRUE)
set.seed(n) Is used to choose a seed for the random number
generator; n must be an integer

A.1.5 Basic Statistical Functions

The following statistical functions are used to describe a quantitative variable
x. For all these functions, the parameter na.rm=TRUE enables us to eliminate
missing data prior to calculation. If na.rm=FALSE and if there is missing data,
the function yields an error message.

Function Description
mean(x, Mean of x calculated from the present data
na.rm=TRUE)
sd(x) Standard deviation of x
var(x) Variance of x if x is a vector, or variance-
covariance matrix if x is a matrix
cor(x) Matrix of the correlations of x
median(x) Median of x
quantile(x, Quantiles of x for the given probabilities probs
probs)
range(x) Range of x
sum(x) Sum of the elements of x
min(x) Minimum of x
max(x) Maximum of x
sign(x) Yields the sign of x (positive or negative)
scale(x, Centres (center=TRUE) and standardises
center=TRUE, (scale=TRUE) x
scale=TRUE)
colMeans(x) Calculates the mean for each column in table x
rowMeans(x) Calculates the mean for each row in table x
262 R for Statistics

Function Description
apply(x,MARGIN, Applies the function FUN to the rows or columns
FUN) in the table x: apply(x, 2, mean) calculates the
mean of each column of x; apply(x, 1, sum) cal-
culates the sum for each row of x
aggregate(x,by, Applies the function FUN to x according to the
FUN) list of factors offered in by: aggregate(x, by
=list(vec),mean) calculates the means of x for
each level of vec

A.1.6 Advanced Statistical Functions

Function Description
t.test Constructs a confidence interval for a mean, tests
the equality of a mean to a given value or con-
structs a test to compare the means of two sub-
sets; t.test(x) constructs a confidence interval
for the mean of x and tests the equality of the
mean of x to the value mu (by default mu=0);
t.test(x~fac) constructs the test of equality of
means in two given sub-populations defined by
the factor fac (which must have two categories).
By default, the test is constructed with unequal
variances var.equal = FALSE; the test is by de-
fault bilateral. We can construct a unilateral test
using alternative = "less" or alternative =
"greater"; see Worked Examples 6.1 and 6.3
var.test Constructs a test of comparison of variances; see
Worked Example 6.3
chisq.test Constructs a chi-square test; see Worked Exam-
ple 6.2
prop.test Constructs a test of equality of proportions; see
Worked Examples 6.4 and 6.5
lm(formula) Constructs a linear model, that is, a multiple re-
gression, an analysis of variance or covariance de-
pending on the nature of the explanatory vari-
ables; see Worked Examples 7.1, 7.2, 8.1, 8.2,
and 8.3
anova Yields the analysis of variance table
aov(formula) Constructs the analysis of variance model defined
by formula; if formula =y~x1 + x2 + x1:x2,
constructs a model with the main effects x1 and
x2 and with the interaction of x1 with x2; see
Worked Examples 8.1 and 8.2
Appendix 263

Function Description
glm Constructs a generalised linear model; see
Worked Example 9.2
dist(x) Constructs a matrix of distances between the
rows of the matrix x
PCA Function of the FactoMineR package which con-
structs a principal component analysis with the
possibility of adding supplementary individuals
and quantitative and/or qualitative variables; see
Worked Example 10.1
CA Function of the FactoMineR package which con-
structs a correspondence analysis; see Worked
Example 10.2
MCA Function of the FactoMineR package which con-
structs a multiple correspondence analysis with
the possibility of adding supplementary individu-
als and quantitative and/or qualitative variables;
see Worked Example 10.3
dimdesc Function of the FactoMineR package which de-
scribes the dimensions of principal component
methods
catdes Function of the FactoMineR package which de-
scribes a qualitative variable according to quan-
titative and/or qualitative variables; see Worked
Examples 10.3 and 11.1
agnes Function which constructs an ascending hierar-
chical classification; see Worked Example 10.2
kmeans(x, Constructs a K-means classification from the
centers) data table x; centers corresponds to the number
of classes or the centres of classes used to initiate
the algorithm; see Worked Example 11.2
HCPC Constructs a classification from the results of a
principal component methods; see Section 11.1.6
lda Function of the MASS package used to conduct
a linear discriminant analysis, see Worked Exam-
ple 9.1
rpart Function of the rpart package which constructs
a regression (or segmentation) tree, see Worked
Example 9.3

A.1.7 Graphical Functions

Function Description
barplot Yields a bar chart
hist Yields a histogram
264 R for Statistics

Function Description
boxplot Yields a boxplot; boxplot(y~fac) yields a graph
with a boxplot for each category of the factor fac
pie Yields a pie chart
points Plots points on a preexisting graph
lines Plots lines on a preexisting graph
curve Draws the curve for a function
abline(a,b) Draws the line with slope b and intercept a on a
preexisting graph
legend Adds a legend to a preexisting graph; leg-
end("topleft", legend = ... ) draws the
legend at the top left
scatterplot(y~x) Constructs a scatterplot y according to x; by
default, a regression line (reg.line=TRUE) and
a non-parametric adjustment curve (by default,
smooth=TRUE) are drawn. Boxplots are also pro-
vided for x and y (by default, boxplots="xy").
scatterplot(y~x|z) constructs a scatterplot for
each category of z
pairs Constructs scatterplots for each pair of variables
of a table
persp Draws graphs in perspective, or response surfaces
image Constructs three-dimensional response surfaces
locator Reads the position of the cursor on the graph
identify Searches for the individuals with the coordinates
closest to the position of the cursor
colors() Provides the list of the 657 colours defined by
default in R
graphics.off() Closes all the graphs which are open
X11() Creates a new empty graph window
pdf, postscript, Used to save a graph in pdf, postscript, jpeg,
jpeg, png, bmp png or bmp format; all the functions are used in
the same way: pdf("mygraph.pdf"); graph or-
der ; dev.off(); see Section 3.1.5

A.1.8 Import and Export Functions

Function Description
read.table Reads a file in table format and creates a data-
frame
read.csv Reads a file with a csv extension containing a
table and creates a data-frame
scan Reads the data originating in a vector or list from
a console or file
write Writes the data in a file
Appendix 265

Function Description
write.table Writes a table in a file
save Saves R objects in a file .Rdata
load Loads objects saved using the save function
history Retrieves the most recent command lines
savehistory Saves the command lines in a .Rhistory file
loadhistory Reads the command lines saved in a .Rhistory
file

A.1.9 Text Management

Function Description
substr(x, Extracts or replaces a sub-chain of characters:
start=n1, substr("abcdef", 2, 4) yields bcd
stop=n2)
grep(pattern, x, Gives the indices for the elements of list x for
fixed=FALSE) which the sub-chain pattern is present. If
fixed=FALSE then pattern is an expression, if
fixed=TRUE then pattern is a chain of charac-
ters
sub(pattern, Finds, within the chain x, the sub-chain pattern
replacement, x, and replaces it with replacement: sub("man","R
fixed=FALSE) useR","Hello man") returns Hello R useR.
Replacement only occurs once. See also gsub for
multiple replacement.
strsplit(x, split) Cuts a chain of characters into multiple sub-
chains according to the character split: str-
split("abedtedr","e") returns "ab" "dt" "dr"
paste(..., Concatenates multiple chains of characters by
sep="") separating them with sep
toupper(x) Writes x in uppercase characters
tolower(X) Writes X in lowercase letters
apropos(what) Yields objects containing the chain of characters
what in their name

A.1.10 Some Other Useful Functions

Function Description
seq(from, to, Generates a sequence: seq(1,9,2) gives 1 3 5 7
by) 9
rep(x, ntimes) Repeats sequence x a given number of times:
rep(4:6,2) gives 4 5 6 4 5 6. It is also possible to
repeat each term of the object: rep(4:6,each=2)
gives 4 4 5 5 6 6
266 R for Statistics

Function Description
cut(x, n) Divides a continuous variable into a qualitative
variable with n levels
solve Inverts a matrix or resolves a linear system; see
Section 1.4.5
eigen Gives the eigenvalues and eigenvectors of a ma-
trix; see Section 1.4.5
svd Compute the singular-value decomposition of a
rectangular matrix

A.2 Writing a Formula for the Models

Many methods, such as linear regression, analysis of variance, analysis of
covariance, logistic regression, etc., require a model to be written. For all
these functions, the models are written in similar ways. In this section we
show how to do this using a number of examples:

Y ~ .: model with the variable Y explained by all the other variables in

the dataset

Y ~ x1+x2: model with the variable Y explained by the variables x1, x2;
equivalent to Y ~ 1+x1+x2
Y ~ -1+x1+x2: model with the variable Y explained by the variables x1,
x2 without the constant (-1 eliminates the constant)
Y ~ x1+x2 + x1:x2: model with the variable Y explained by the variables
x1, x2 and the interaction between x1 and x2
Y ~ x1*x2: equivalent to the the previous model
Y ~ x1*x2*x3: is equivalent to the model with all the main effects and
the interactions between x1, x2 and x3, thus Y ~ x1+x2+x3 + x1:x2 +
x1:x3 + x2:x3 + x1:x2:x3
Y ~ (x1+x2):x3: is equivalent to the model Y ~ x1:x3+x2:x3
Y ~ x1*x2*x3 - x1:x2:x3: is equivalent to the model Y ~ x1+x2+x3 +
x1:x2 + x1:x3 + x2:x3

Y ~ x1+x2 %in% x1: model with the effects of x1 and of x2 organised

into a hierarchy (or nested) in x1
Y ~ sin(x1)+sin(x2): model of Y by sin(x1) and sin(x2)
Y ~ x1 | fac: model where the effect of x1 is possible for each level of
fac
Appendix 267

Y ~ I(x1^ 2): model with the variable Y explained by x12 ; I(.) protects
the expression x1^2, else x1^2 is interpreted as x1*x1 = x1+x1+x1:x1 =
x1
Y ~ I(x1+x2): model where the variable Y is explained by the constant
and the variable resulting from the sum (individual by individual) of the
variables x1 and x2; I(.) protects the expression x1+x2, otherwise x1+x2
is interpreted as two explanatory variables

A.3 The Rcmdr Package

The R Commander graphical user interface is available in the Rcmdr package.
With this interface, R can be used with an easy-to-use drop-down menu. The
educational advantage of this package is that it also provides the lines of code
which correspond to the analyses carried out: users can therefore familiarise
themselves with programming in R by seeing which functions are used. The
Rcmdr interface does not contain all the functions available in R, nor does
it feature all the options for the different functions, but the most common
functions are programmed and the most classical options are available.
The package only needs to be installed once (see Section 1.6, p. 24). The
interface is then loaded using
> library(Rcmdr)
The interface (Figure A.5) opens automatically. This interface features a
drop-down menu, a script window and an output window. When the drop-
down menu is used, the analysis is launched and the lines of code which were
used to generate the analysis are written in the script window.
The simplest way to import data with Rcmdr is to use an Excel file:
Data Import data from Excel, Access or dBase data set ...
For a txt or csv file:
Data Import data from text file, clipboard, or URL ... Next
specify that the field separator is the space and the decimal-point character
is ..
To check that the dataset has been imported successfully:
Statistics Summaries Active data set
If we want to read a dataset in csv format which contains the individuals
names, it is not possible to specify in the Rcmdr drop-down menu that the
first column contains the identifier. We can therefore import the dataset
considering the individuals names as a variable. We thus modify the line of
code written in the script window by adding the argument row.names=1 and
then clicking on Submit.
268 R for Statistics

Figure A.5
Main window of Rcmdr.

To change the active dataset, simply click on the menu Data. If we modify
the active dataset (for example by converting a variable), the dataset must be
refreshed using
Data Active data set Refresh active data set
The output window features the code lines in red and the results in blue.
The graphs are drawn in R. At the end of an Rcmdr session, the script window,
and therefore all the instructions, can be saved along with the output file that
is, all the results. R and Rmcdr can be closed simultaneously using File
Exit From Commander and R.

Remarks
Writing in the script window of Rcmdr or in the R window amounts to the
Appendix 269

same thing. If an instruction is launched from Rcmdr, it is also recognised

in R and vice versa. The objects created by Rcmdr can thus be used in R.

Rcmdr windows may not open correctly, or may be hidden behind other
open windows. In this case, if using Windows, right-click on the R icon or
on the shortcut used to launch R, then click on Properties, and change
the target by adding --sdi after the file access path Rgui.exe, which,
for example, gives

"C:\Program Files\R\R-2.14.1\bin\Rgui.exe" --sdi

A.4 The FactoMineR Package

A.4.1 The FactoMineR Package
FactoMineR is a package dedicated to exploratory multivariate data analysis
in a French way. The most common exploratory multivariate data analy-
sis methods are programmed within it: Principal Component Analysis (PCA
function), Correspondence Analysis (CA function), Multiple Correspondence
Analysis (MCA function). More advanced methods are also available and
can be used to take the structure of the variables or individuals into account:
Multiple Factor Analysis (MFA function) in the sense of Escofier and Pages,
Hierarchical Multiple Factor Analysis (HMFA function) or Dual Multiple Fac-
tor Analysis (DMFA function). The catdes function can be used to describe
a qualitative variable according to quantitative and/or qualitative variables.
The condes function can be used to describe a quantitative variable according
to quantitative and/or qualitative variables.

In each of these methods, it is also possible to add illustrative elements:

supplementary individuals and supplementary quantitative and/or qualitative
variables. For each of these analyses, there are many features which help to
interpret the data: representation quality, contribution for individuals and for
variables. Graphical representations are the core of each of the analyses, and
there are many graphical options available in the package: colouring the indi-
viduals according to a qualitative variable, representing only those variables
which are best projected on the factorial planes, etc.

The package only needs to be installed once (see Section 1.6, p. 24). The
package is then loaded using
> library(FactoMineR)
A website entirely dedicated to the FactoMineR package can be found at
270 R for Statistics

http://factominer.free.fr. The methods are described and examples are

given in detail.
Remark
Another data analysis package is available in R: the ade4 package. Another
website dedicated to this package which also provides a great number of de-
tailed examples can be found at http://pbil.univ-lyon1.fr/ADE-4.

A.4.2 The Drop-Down Menu

A graphical interface is also available and can be installed within the Rcmdr
package interface (see Appendix A.3). There are two ways of loading the
FactoMineR package:

Install the FactoMineR drop-down menu in Rcmdr permanently. To do

this, simply write or copy and paste the following line of code into an R
window:

> source("http://factominer.free.fr/install-facto.r")

For subsequent use of the FactoMineR drop-down menu, simply launch

Rcmdr with the command library(Rcmdr), and the drop-down menu will
appear by default.
Install the FactoMineR drop-down menu in Rcmdr for the current session.
To do this, install RcmdrPlugin.FactoMineR (see Section 1.6). Then, each
time that you want to use the FactoMineR drop-down menu, launch Rcmdr
and then click Tools Load Rcmdr plug-in(s) .... Choose the Fac-
toMineR Plug-in from the list, and then relaunch Rcmdr in order to take
this new plug-in into account. This option is more complicated, which is
why we advise users to opt for the first possibility. An explanation of the
use of the drop-down menu can be found in the PCA section, p. 220.

A.5 Answers to the Exercises

A.5.1 Exercises: Chapter 1
Exercise 1.1 (Creating Vectors)
1. The three vectors are created using the rep function with the arguments
each and times:

> vec1 <- rep(1:5,3)

> vec1
[1] 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Appendix 271

> vec2 <- rep(1:5,each=3)

> vec2
[1] 1 1 1 2 2 2 3 3 3 4 4 4 5 5 5
> vec3 <- rep(1:4,times=(2:5))
> vec3
[1] 1 1 2 2 2 3 3 3 3 4 4 4 4 4

2. The paste function concatenates vectors:

> vec4 <- paste("A",0:10,")",sep="")

> vec4
[1] "A0)" "A1)" "A2)" "A3)" "A4)" "A5)"
[7] "A6)" "A7)" "A8)" "A9)" "A10)"

3. The position of the letter q is first calculated. Then the vector of letters
and the vector of indices are pasted:

> pos.q <- which(letters=="q")

> vec5 <- paste(letters[1:pos.q],1:pos.q,sep="")
> vec5
[1] "a1" "b2" "c3" "d4" "e5" "f6" "g7" "h8" "i9"
[10] "j10" "k11" "l12" "m13" "n14" "o15" "p16" "q17"

Exercise 1.2 (Working with NA)

1. The vector is created, and the mean and variance are computed:

> set.seed(007)
> vec1 <- runif(100,0,7)
> mean(vec1)
[1] 3.564676
> var(vec1)
[1] 3.94103

2. Missing values are allocated at random:

> vec2 <- vec1

> ind <- sample(1:100,10)
> vec2[ind] <- NA
> indNA <- which(is.na(vec2))
[1] 4 19 33 38 40 49 62 71 90 99

3. It is necessary to use the argument na.rm=TRUE to compute the mean and

variance.

> mean(vec2)
[1] NA
272 R for Statistics

> mean(vec2,na.rm=T)
[1] 3.539239
> var(vec2)
[1] NA
> var(vec2,na.rm=T)
[1] 3.879225

4. We delete the missing values and find again the mean and variance previ-
ously calculated with the argument na.rm=TRUE:

> vec3 <- vec2[-indNA]

> mean(vec3)
[1] 3.539239
> var(vec3)
[1] 3.879225

5. If the missing values are replaced by the mean of the variable, then the
mean is the same as before but the variance is under-estimated:

> vec4 <- vec2

> vec4[indNA] <- mean(vec3)
> mean(vec4)
[1] 3.539239
> var(vec4)
[1] 3.487384

6. The missing values are replaced by values drawn from a normal distribution
with mean and standard deviation of the variable:

> vec5 <- vec2

> vec5[indNA] <- rnorm(length(indNA),mean(vec3),sd(vec3))
> mean(vec5)
[1] 3.545158
> var(vec5)
[1] 3.665774

7. The missing values are replaced by values drawn from a Uniform distribu-
tion from the minimum to the maximum of the observed values:

> vec6 <- vec2

> vec6[indNA] <- runif(length(indNA),min(vec3),max(vec3))
> mean(vec6)
[1] 3.726103
> var(vec6)
[1] 3.863905
Appendix 273

8. The missing values are replaced by values randomly drawn from the ob-
served values:

> vec7 <- vec2

> vec7[indNA] <- sample(vec3,10)
> mean(vec7)
[1] 3.477317
> var(vec7)
[1] 3.89452
Exercise 1.3 (Creating, Manipulating and Inverting a Matrix)
1. The matrix mat is created prior to attributing the row and column names:

> mat <- matrix(c(1,0,3,4,5,5,0,4,5,6,3,4,0,1,3,2),ncol=4)

> rownames(mat) <- paste("row",1:4,sep="-")
> colnames(mat) <- paste("column",1:4)

2. The diagonal elements of the matrix mat are obtained as follows:

> vec <- diag(mat)

> vec
[1] 1 5 3 2

3. The matrix containing the first 2 rows of mat is obtained as follows:

> mat1 <- mat[c(1,2),]

> mat1
column 1 column 2 column 3 column 4
row-1 1 5 5 0
row-2 0 5 6 1

4. The matrix containing the last 2 columns of mat is obtained as follows:

> mat2 <- mat[,(ncol(mat)-1):ncol(mat)]

> mat2
column 3 column 4
row-1 5 0
row-2 6 1
row-3 3 3
row-4 4 2

5. To calculate the determinant and invert the matrix, simply use the func-
tions det and solve:

> det(mat)
[1] 60
274 R for Statistics

> solve(mat)
row-1 row-2 row-3 row-4
column 1 0.5 -0.5 0.1666667 -5.551115e-17
column 2 -0.6 0.4 -0.4666667 5.000000e-01
column 3 0.7 -0.3 0.4333333 -5.000000e-01
column 4 -1.2 0.8 -0.2666667 5.000000e-01
Exercise 1.4 (Selecting and Sorting in a Data-Frame)
1. The iris data is loaded and a new dataset is created by selecting only the
rows carrying the value "versicolor" for the Species variable:

> data(iris)
> iris2 <- iris[iris[,"Species"]=="versicolor", ]

2. The dataset is sorted according to the first variable using the order func-
tion:

> iris2[order(iris2[,1],decreasing=TRUE),]
Sepal.Length Sepal.Width Petal.Length Petal.Width Species
51 7.0 3.2 4.7 1.4 versicolor
53 6.9 3.1 4.9 1.5 versicolor
77 6.8 2.8 4.8 1.4 versicolor
...
61 5.0 2.0 3.5 1.0 versicolor
94 5.0 2.3 3.3 1.0 versicolor
58 4.9 2.4 3.3 1.0 versicolor
Exercise 1.5 (Using the apply Function)
1. To calculate the benchmark statistics, simply use the summary function:

> library(lattice) # load the package

> data(ethanol)
> summary(ethanol)
NOx C E
Min. :0.370 Min. : 7.500 Min. :0.5350
1st Qu.:0.953 1st Qu.: 8.625 1st Qu.:0.7618
Median :1.754 Median :12.000 Median :0.9320
Mean :1.957 Mean :12.034 Mean :0.9265
3rd Qu.:3.003 3rd Qu.:15.000 3rd Qu.:1.1098
Max. :4.028 Max. :18.000 Max. :1.2320

2. To calculate the quartiles, we can use the apply function:

> apply(X=ethanol,MARGIN=2,FUN=quantile)
NOx C E
0% 0.3700 7.500 0.53500
25% 0.9530 8.625 0.76175
50% 1.7545 12.000 0.93200
75% 3.0030 15.000 1.10975
100% 4.0280 18.000 1.23200
Appendix 275

3. The instruction for the previous question by default yields the quartiles.
Indeed, since we have not specified the argument probs for the quantile func-
tion, the argument used by default is: probs=seq(0,0.25,0.5,0.75,1) (see
the help section for the quantile function). To obtain deciles, we have to spec-
ify probs=seq(0,1,by=0.1) as the argument. The help section for the apply
function indicates the optional arguments via ...: optional arguments
to FUN. It is therefore possible to pass probs=seq(0,1,by=0.1) as an
argument to the FUN=quantile function:

> apply(ethanol,2,quantile,probs=seq(0,1,by=0.1))
NOx C E
0% 0.3700 7.5 0.5350
10% 0.6000 7.5 0.6496
20% 0.8030 7.5 0.7206
30% 1.0138 9.0 0.7977
40% 1.4146 9.0 0.8636
50% 1.7545 12.0 0.9320
60% 2.0994 12.6 1.0104
70% 2.7232 15.0 1.0709
80% 3.3326 15.0 1.1404
90% 3.6329 18.0 1.1920
100% 4.0280 18.0 1.2320

Exercise 1.6 (Selection in a Matrix with the apply Function)

1. The matrix containing the columns of mat having all values smaller than
6 is obtained as follows:

> mat <- matrix(c(1,0,3,4,5,5,0,4,5,6,3,4,0,1,3,2),ncol=4)

> mat3 <- mat[,apply((mat<6),2,all)]
> mat3
[,1] [,2] [,3]
[1,] 1 5 0
[2,] 0 5 1
[3,] 3 0 3
[4,] 4 4 2

2. Because there is only one row which do not contain 0, we have to use
drop=FALSE such that the output is a matrix and not a vector:

> mat4 <- mat[apply((mat>0),1,all),,drop=FALSE]

> mat4
[,1] [,2] [,3] [,4]
[1,] 4 4 4 2
Exercise 1.7 (Using the lapply Function)
1. The MASS package and the dataset Aids2 are loaded:
276 R for Statistics

> library(MASS) # load the package

> data(Aids2)
> summary(Aids2)
state sex diag death status
NSW :1780 F: 89 Min. : 8302 Min. : 8469 A:1082
Other: 249 M:2754 1st Qu.:10163 1st Qu.:10672 D:1761
QLD : 226 Median :10665 Median :11235
VIC : 588 Mean :10584 Mean :10990
3rd Qu.:11103 3rd Qu.:11504
Max. :11503 Max. :11504

T.categ age
hs :2465 Min. : 0.00
blood : 94 1st Qu.:30.00
hsid : 72 Median :37.00
other : 70 Mean :37.41
id : 48 3rd Qu.:43.00
haem : 46 Max. :82.00
(Other): 48

2. The function is.numeric returns a boolean: TRUE when the object on which
it is applied is numeric. We have to apply this function to each column of the
data-frame Aids2 and take the negation (operator !). As the data-frame is a
list, applying a function to each column is (usually) equivalent to applying a
function to each component of the list; this is the scope of the lapply function:

> ind <- !unlist(lapply(Aids2,is.numeric))

3. We just have to select the variables of the data-frame using ind:

> Aids2.qual <- Aids2[,ind]

4. We use the levels function on each element of the data-frame Aids2.qual:

> lapply(Aids2.qual,levels)

Exercise 1.8 (Levels of the Qualitative Variables in a Subset)

1. The package MASS and the dataset Aids2 are loaded:

> library(MASS) # load the package

> data(Aids2)

2. The selection is done as follows:

> res <- Aids2[(Aids2[,"sex"]=="M")&(Aids2[,"state"]!="Other"),]

Another method consists in using the subset function.

Appendix 277

3. The summary indicates that the levels are still the same, M and F, but no
individuals take the category F.

> summary(res)
state sex diag death status
NSW :1726 F: 0 Min. : 8302 Min. : 8469 A: 947
Other: 0 M:2518 1st Qu.:10155 1st Qu.:10671 D:1571
QLD : 217 Median :10662 Median :11220
VIC : 575 Mean :10583 Mean :10987
3rd Qu.:11104 3rd Qu.:11504
Max. :11503 Max. :11504

T.categ age
hs :2260 Min. : 0.00
hsid : 68 1st Qu.:30.00
other : 55 Median :37.00
blood : 54 Mean :37.36
haem : 40 3rd Qu.:43.00
id : 21 Max. :82.00
(Other): 20

4. The attributes of sex are:

> attributes(res[,"sex"])
$levels
[1] "F" "M"

$class
[1] "factor"

5. We transform the sex variable into a character object and print the at-
tributes of the resulting object:

> sexc <- as.character(res[,"sex"])

> attributes(sexc)
NULL

6. Transform the sexc into a factor:

> sexf <- as.factor(sexc)

> attributes(sexf)
$levels
[1] "M"

$class
[1] "factor"
278 R for Statistics

7. We select the indices of the non-numeric variables:

> ind <- !unlist(lapply(res,is.numeric))

> ind
state sex diag death status T.categ age
TRUE TRUE FALSE FALSE TRUE TRUE FALSE

8. We transform the selected variables into character:

> res[,ind] <- lapply(res[,ind],as.character)

9. We transform the selected variables into factors:

> res[,ind] <- lapply(res[,ind],as.factor)

> summary(res)
state sex diag death status
NSW:1726 M:2518 Min. : 8302 Min. : 8469 A: 947
QLD: 217 1st Qu.:10155 1st Qu.:10671 D:1571
VIC: 575 Median :10662 Median :11220
Mean :10583 Mean :10987
3rd Qu.:11104 3rd Qu.:11504
Max. :11503 Max. :11504

T.categ age
hs :2260 Min. : 0.00
hsid : 68 1st Qu.:30.00
other : 55 Median :37.00
blood : 54 Mean :37.36
haem : 40 3rd Qu.:43.00
id : 21 Max. :82.00
(Other): 20

A.5.2 Exercises: Chapter 2

Exercise 2.1 (Robust Reading of Data)
1. The scan function allows us to read data from file in the form of charac-
ter vectors. Since the first row contains characters, this type is compulsory.
Moreover, two types of decimal are used: , and ..

> myvector <- scan("mydata.csv",what="",sep=";")

Read 20 items
> myvector
[1] "surname" "height" "weight" "feet_size" "sex"
[6] "tony" "184" "80" "9.5" "M"
[11] "james" "175,5" "78" "8.5" "M"
[16] "john" "158" "72" "8" "M"

2. Replace commas by dots:

Appendix 279

> myvector <- gsub(",",".",myvector)

3. Raw data matrix (with row and column names):

> mymatrix <- matrix(myvector,nrow=4,ncol=5,byrow=TRUE)

> mymatrix
[,1] [,2] [,3] [,4] [,5]
[1,] "surname" "height" "weight" "feet_size" "sex"
[2,] "tony" "184" "80" "9.5" "M"
[3,] "james" "175.5" "78" "8.5" "M"
[4,] "john" "158" "72" "8" "M"

4. Retrieval of column and row names:

> namecol <- mymatrix[1,-1]

> namecol
[1] "height" "weight" "feet_size" "sex"
> namerow <- mymatrix[-1,1]
> namerow
[1] "tony" "james" "john"
> mymatrix <- mymatrix[-1,-1]

5. Data-frame of data with the row and column names:

> mydata <- as.data.frame(mymatrix)

> colnames(mydata) <- namecol
> rownames(mydata) <- namerow
> summary(mydata)
height weight feet_size sex
158 :1 72:1 42:1 M:3
175.5:1 78:1 43:1
184 :1 80:1 44:1

The first three variables are factors and must therefore be converted to nu-
merics (Section 2.3.1, p. 35):

> for (i in c(1,2,3)){

+ mydata[,i] <- as.numeric(as.character(mydata[,i]))
+ }
> summary(mydata)
height weight feet_size sex
Min. :158.0 Min. :72.00 Min. :42.0 M:3
1st Qu.:166.8 1st Qu.:75.00 1st Qu.:42.5
Median :175.5 Median :78.00 Median :43.0
Mean :172.5 Mean :76.67 Mean :43.0
3rd Qu.:179.8 3rd Qu.:79.00 3rd Qu.:43.5
Max. :184.0 Max. :80.00 Max. :44.0
280 R for Statistics
Exercise 2.2 (Reading Data from File)
> test1 <- read.table("test1.csv",sep=",",header=TRUE)
> summary(test1)
CLONE B IN HT19
1-105 :9 Min. :1 Min. :1.000 Min. : 3.000
1-41 :9 1st Qu.:1 1st Qu.:2.750 1st Qu.: 7.968
18-428:9 Median :1 Median :4.500 Median : 9.055
18-429:5 Mean :1 Mean :4.688 Mean : 8.556
3rd Qu.:1 3rd Qu.:7.000 3rd Qu.: 9.925
Max. :1 Max. :9.000 Max. :11.300
C19 HT29
Min. : 5.00 Min. : 4.50
1st Qu.:18.75 1st Qu.:11.38
Median :21.50 Median :12.75
Mean :21.75 Mean :11.91
3rd Qu.:26.00 3rd Qu.:13.75
Max. :34.00 Max. :15.25

> test1prn <- read.table("test1.prn",header=TRUE)

> summary(test1prn)
CLONE B IN HT19
1-105 :9 Min. :1 Min. :1.000 Min. : 3.000
1-41 :9 1st Qu.:1 1st Qu.:3.000 1st Qu.: 8.150
18-428:9 Median :1 Median :5.000 Median : 9.110
18-429:6 Mean :1 Mean :4.727 Mean : 8.595
3rd Qu.:1 3rd Qu.:7.000 3rd Qu.: 9.900
Max. :1 Max. :9.000 Max. :11.300
C19 HT29
Min. : 5.00 Min. : 4.50
1st Qu.:19.00 1st Qu.:11.50
Median :22.00 Median :12.75
Mean :21.85 Mean :11.97
3rd Qu.:26.00 3rd Qu.:13.75
Max. :34.00 Max. :15.25

> test2 <- read.table("test2.csv", sep=";", header=TRUE,

na.strings="")
> str(test2)
data.frame: 999 obs. of 6 variables:
$ CLONE: Factor w/ 87 levels "","1-105","1-41",..: 2 2 2 2 2 2 2 2 2 3 ...
$ B : int 1 1 1 1 1 1 1 1 1 1 ...
$ IN : int 1 2 3 4 5 6 7 8 9 1 ...
$ HT19 : num 8.95 9.11 9.85 10.82 9.4 ...
$ C19 : int 21 21 26 29 22 19 20 20 24 33 ...
$ HT29 : num 12.5 13 14.2 15 13 ...

> test3 <- read.table("test3.csv", sep=";", header=TRUE,

na.strings=".")
Appendix 281

> summary(test3)
CLONE B IN HT19
1-105 : 18 Min. :1.000 Min. :1 Min. : 0.89
1-41 : 18 1st Qu.:1.000 1st Qu.:3 1st Qu.: 7.35
18-428 : 18 Median :1.000 Median :5 Median : 9.32
18-429 : 18 Mean :1.234 Mean :5 Mean : 10.47
18-430 : 18 3rd Qu.:1.000 3rd Qu.:7 3rd Qu.: 10.50
18-438 : 18 Max. :2.000 Max. :9 Max. :1142.00
(Other):891 NAs : 15.00
C19 HT29
Min. : 3.00 Min. : 1.96
1st Qu.:17.00 1st Qu.:10.75
Median :23.00 Median :12.88
Mean :22.08 Mean :12.20
3rd Qu.:28.00 3rd Qu.:14.00
Max. :37.00 Max. :17.50
NAs :20.00 NAs :15.00

Exercise 2.3 (Reading Data from File with Date Format)

1. Read the dataset, skipping the first two lines:

> ski <- read.table("test4.csv", sep="|", skip=2, header=TRUE,

row.names=1)
> summary(ski)
age gender first.time.skiing
Min. :24.00 Min. :0.00 1980-05-01:1
1st Qu.:28.00 1st Qu.:0.00 1982-01-31:1
Median :32.00 Median :0.00 1992-01-15:1
Mean :30.38 Mean :0.25 2003-03-16:1
3rd Qu.:33.00 3rd Qu.:0.25 2005-02-26:1
Max. :33.00 Max. :1.00 2006-03-04:1
(Other) :2

2. Use the format Date for the last variable. The format POSIXct could be
interesting if the time was specified.

> ski2<-read.table("test4.csv",sep="|",skip=2,header=TRUE,
row.names=1,colClasses=c("character","numeric",
"factor","Date"))
> summary(ski2)
age gender first.time.skiing
Min. :24.00 0:6 Min. :1980-05-01
1st Qu.:28.00 1:2 1st Qu.:1989-07-20
Median :32.00 Median :2004-03-06
Mean :30.38 Mean :1998-06-01
3rd Qu.:33.00 3rd Qu.:2006-12-02
Max. :33.00 Max. :2009-03-06
282 R for Statistics
Exercise 2.4 (Reading Data from File and Merging)
1. Read the datasets:

> state1 <- read.table("state1.csv",sep=";",header=TRUE)

> state2 <- read.table("state2.csv",sep=",",header=TRUE)
> state3 <- read.table("state3.csv",row.names=1,header=TRUE)

2. Merge by key (common variable region for state1 and state3, then com-
mon variable state for state2 and the previous table):

> state13 <- merge(state1,state3,by="region")

> state123 <- merge(state2,state13,by="state")
Exercise 2.5 (Merging and Selection)
1. Open the two files in Excel (or OpenOffice), save them in text format .csv
(with OpenOffice, choose field separator ; ). Then read these two files using

> fusion1 <- read.table("fusion1.csv",sep=";",header=TRUE)

> summary(fusion1)
yhat1 yhat3 yhat2
Min. :-0.46520 Min. :-0.6267 Min. :-2.1563
1st Qu.:-0.37217 1st Qu.: 0.2906 1st Qu.:-1.1357
Median :-0.26458 Median : 1.1447 Median :-0.8284
Mean :-0.26641 Mean : 0.9956 Mean :-0.9200
3rd Qu.:-0.17349 3rd Qu.: 1.5200 3rd Qu.:-0.6037
Max. :-0.03419 Max. : 2.8496 Max. :-0.2903
yhat4 yhat5
Min. :0.5374 Min. :0.07706
1st Qu.:0.9553 1st Qu.:0.17804
Median :1.4079 Median :0.25966
Mean :1.3332 Mean :0.27391
3rd Qu.:1.6578 3rd Qu.:0.37683
Max. :2.1196 Max. :0.48926
> fusion2 <- read.table("fusion2.csv",sep=";",header=TRUE)
> summary(fusion2)
Rhamnos Fucos Arabinos
Min. :-0.5348 Min. :-2.3026 Min. :-0.77403
1st Qu.:-0.3513 1st Qu.:-2.0090 1st Qu.:-0.24294
Median :-0.1010 Median :-1.4524 Median :-0.03542
Mean : 0.0925 Mean :-1.4028 Mean : 0.15860
3rd Qu.: 0.5184 3rd Qu.:-0.7409 3rd Qu.: 0.50681
Max. : 1.3447 Max. :-0.2829 Max. : 1.39122
Xylos Mannos
Min. :0.4243 Min. :0.02623
1st Qu.:0.9477 1st Qu.:0.15794
Median :1.1553 Median :0.49551
Mean :1.1117 Mean :0.51873
3rd Qu.:1.3142 3rd Qu.:0.81277
Max. :1.7882 Max. :1.14627
Appendix 283

2. Retain two columns for each table and create the data-frame:

> fusion1 <- fusion1[,c("yhat1","yhat3")]

> fusion2 <- fusion2[,c("Rhamnos","Arabinos")]
> data <- cbind(fusion1,fusion2)

3. Create two variables and add these to the data-frame data:

> yres1 <- data[,"yhat1"]-data[,"Rhamnos"]

> yres2 <- data[,"yhat3"]-data[,"Arabinos"]
> data <- cbind.data.frame(data,yres1,yres2)
> names(data)
[1] "yhat1" "yhat3" "Rhamnos" "Arabinos" "yres1" "yres2"

Exercise 2.6 (Ventilation)

1. Calculate the frequencies:

> tabl <- table(Xfac)

> tabl/sum(tabl)
Xfac
A B C D
0.60 0.20 0.17 0.03

2. Display the categories with a sample size of less than 5%:

> lev <- levels(Xfac)

> select <- (tabl/sum(tabl))<0.05
> lev[select]
[1] "D"

3. Frequencies of the categories without category D:

> mod.select <- lev[!select]

> prov <- factor(Xfac[(Xfac%in%mod.select)],levels=mod.select)
> prov <- table(prov)
> freq <- prov/sum(prov)
> freq
prov
A B C
0.6185567 0.2061856 0.1752577

4. Select individuals carrying the category D, random draw (ventilation) of

their new category. The factor Xfac has categories which are described in
levels but which are no longer represented (i.e. the categories which have
been ventilated, here D). The last row updates the list of factor levels.
284 R for Statistics

> for (j in lev[select]) {

+ ## sampling in the levels at random with replacement
+ if (length(mod.select)==1) stop("only one category\n") else
+ Xfac[Xfac==j] <- sample(mod.select,sum(Xfac==j),
replace=T, prob = freq)
+ }
> Xfacvent <- factor(as.character(Xfac))
> Xfacvent
[1] A A A A A A A A A A A A A A A A A A A A A A A A A A A
[28] A A A A A A A A A A A A A A A A A A A A A A A A A A A
[55] A A A A A A B B B B B B B B B B B B B B B B B B B B C
[82] C C C C C C C C C C C C C C C C B A C
Exercise 2.7 (Ventilation by Ordinal Factors)
1. Calculate the frequencies:

> Xfac <- ordered(c(rep("0-10",1),rep("11-20",3),rep("21-30",5),

rep("31-40",20),rep("41-50",2),rep("51-60",2),
rep("61-70",1),rep("71-80",31),rep("+ de 80",20)),
levels=c("0-10","11-20","21-30","31-40","41-50","51-60",
"61-70","71-80","+ 80"))
> tabl <- table(Xfac)
> tabl/sum(tabl)
Xfac
0-10 11-20 21-30 31-40 41-50
0.01176471 0.03529412 0.05882353 0.23529412 0.02352941
51-60 61-70 71-80 + 80
0.02352941 0.01176471 0.36470588 0.23529412

2. Display the categories to be ventilated:

> p <- 0.05

> select <- (tabl/sum(tabl))<p
> lev <- levels(Xfac)
> lev[select]
[1] "0-10" "11-20" "41-50" "51-60" "61-70"

3. As long as there is a category with a sample size of less than 5%, it must
be ventilated (and deleted from the list of levels): this is the object of the
loop while.

> while(any((tabl/sum(tabl))<p)) {
+ ## take the first category whose numbers are too low
+ j <- which(((tabl/sum(tabl))<p))[1]
+ K <- length(lev) # effective of the categories updated
+ ## merge the next or the previous category for the last one
Appendix 285

+ if (j<K) {
+ if ((j>1)&(j<K-1)) {
+ levels(Xfac) <- c(lev[1:(j-1)],paste(lev[j],
+ lev[j+1],sep="."),paste(lev[j],lev[j+1],sep="."),
+ lev[j+2:K]) }
+ if (j==1) {
+ levels(Xfac) <- c(paste(lev[j],lev[j+1],sep="."),
+ paste(lev[j],lev[j+1],sep="."),lev[j+2:K]) }
+ if (j==(K-1)) {
+ levels(Xfac) <- c(lev[1:(j-1)],paste(lev[j],
+ lev[j+1],sep="."),paste(lev[j],lev[j+1],sep=".")) }
+ } else {
+ levels(Xfac) <- c(lev[1:(j-2)],paste(lev[j-1],
+ lev[j],sep="."),paste(lev[j-1],lev[j],sep="."))
+ }
+ tabl <- table(Xfac) ## table updated and ...
+ lev <- levels(Xfac) ## ... categories updated
+ }
> tabl
Xfac
0-10.11-20.21-30 31-40 41-50.51-60.61-70
9 20 5
71-80 + 80
31 20
Exercise 2.8 (Cross-Tabulation Data Table)
1. A contingency table can simply be constructed by creating the following
matrix:

> tabl <- matrix(c(2,1,3,0,0,4),2,3)

> colnames(tabl) <- c("Ang","Mer","Tex")
> rownames(tabl) <- c("Low","High")

2. The tabl matrix is not of table type. There are missing attributes. Also,
the opposite simple operation (using as.data.frame) is not directly possible.
One solution is to add attributes to this matrix. Instead, more simply, we
construct this table by hand:

> tabmat <- matrix("",length(tabl),3)

> tabmat[,3] <- as.vector(tabl)
> tabmat[,2] <- rep(rownames(tabl),ncol(tabl))
> tabmat[,1] <- rep(colnames(tabl),each=nrow(tabl))

3. Implement a data-frame tabframe. Column 3, which is converted to a

factor using the data.frame function, must be converted into a numeric (Sec-
tion 2.3.1, p. 35). Then, calculate the total sample size and the number of
factors:
286 R for Statistics

> tabframe <- data.frame(tabmat)

> tabframe[,3] <- as.numeric(as.character(tabframe[,3]))
> tabframe
X1 X2 X3
1 Ang Low 2
2 Ang High 1
3 Mer Low 3
4 Mer High 0
5 Tex Low 0
6 Tex High 4
> n <- sum(tabframe[,3])
> nbfac <- ncol(tabframe)-1

4. Create the tab matrix and the counter:

> tabcomplete <- matrix("",n,nbfac)

> iter <- 1

5. On the first two rows, we operate the loop on all the rows of the tabframe
table and control it using the sample size (not null). In the new table, we
then repeat the allocation of categories, for as many times as required by the
sample size (and therefore not the last column of tabmat which contains these
sample sizes). The row index for this table is managed by the iter counter.

> for (i in 1:nrow(tabframe)) {

+ if (tabframe[i,3]>0) {
+ for (j in 1:tabframe[i,3]) {
+ tabcomplete[iter,] <- tabmat[i,-ncol(tabframe)]
+ iter <- iter+1
+ }
+ }
+ }
> data.frame(tabcomplete)
X1 X2
1 Ang Low
2 Ang Low
3 Ang High
4 Mer Low
5 Mer Low
6 Mer Low
7 Tex High
8 Tex High
9 Tex High
10 Tex High

The tabmat matrix is used when attributing the rows of tabcomplete. It is

Appendix 287

not possible to allocate to rows of the data-frame directly. Indeed, a data-

frame is a list, as is a row of a data-frame, and it is impossible to allocate a
list within a row of a matrix which is itself a vector.

A.5.3 Exercises: Chapter 3

Exercise 3.1 (Draw a function)
1. Define a variation grid x and then draw the sine curve:

> x <- seq(0,2*pi,length=1000)

> plot(sin(x)~x,type="l")

2. To add the title, we use the title function (or we could also directly use
the main argument within the plot function):

> title("Plot of the sine function")

Plot of the sine function

1.0
0.5
sin(x)

0.0
0.5
1.0

0 1 2 3 4 5 6

Figure 3.6
Plot of the sine function.

Exercise 3.2 (Comparison of Distributions)

1. To illustrate a normal distribution, simply draw the density using the
dnorm function. We can then improve the graph by drawing the abscissa
axis and then a segment between 0 and the maximum of the normal distribu-
tion.

> plot(dnorm,-4,4)
> abline(h=0)
> segments(0,0,0,dnorm(0),lty=2)

2. To draw new curves we use the curve function with the argument add=TRUE.
To differenciate between curves, use a different colour for each distribution.
288 R for Statistics

> curve(dt(x,5),add=TRUE,col=2)
> curve(dt(x,30),add=TRUE,col=3)

3. Simply use the legend function and position it at the top left:

> legend("topleft",legend=c("normal","Student(5)","Student(30)"),
col=1:3,lty=1)

Exercise 3.3 (Plotting Points)

1. The scatterplot is imported and constructed immediately:

> ozone <- read.table("ozone.txt",header=T)

> plot(maxO3~T12,data=ozone)

2. To connect the points, simply use type="l"; this graph is illegible as we

must first sort the data in ascending order on the abscissa axis.
3. The order function makes it possible:

> sorted <- order(ozone[,"T12"])

> plot(maxO3~T12,data=ozone[sorted,],type="b")

Exercise 3.4 (Law of Large Numbers)

1. Create a vector X of length 1000:

> set.seed(123)
> X <- rbinom(1000, size=1, prob=0.6)

2. The cumsum function is used to construct a vector of cumulated sums:

> Sl <- cumsum(X)

> Ml <- Sl/(1:1000)
> plot(Ml, type="l")
> abline(h=0.6, col=2)

The resulting graph illustrates the law of large numbers.

Exercise 3.5 (Central Limit Theorem)

1. SN follows a binomial distribution with parameters N and p, with mean
p
N p and standard deviation N p (1 p).
2. Set the random seed generator and then simulate a vector of 1000 occur-
rences of a binomial distribution with the parameters N and p:

> set.seed(123)
> p <- 0.5
> N <- 10
Appendix 289

> U10 <- (rbinom(1000, size = N, p=p) - N*p) /sqrt(N*p*(1-p))

> N <- 30
> U30 <- (rbinom(1000, size = N, p=p) - N*p) /sqrt(N*p*(1-p))
> N <- 1000
> U1000 <- (rbinom(1000, size = N, p=p) - N*p)/sqrt(N*p*(1-p))

3. Before drawing the curve for the standard normal distribution, create a
grid for x varying between 4 and 4. Then, divide the graph window into one
row and three columns, plot a histogram and overlay the curve for the normal
distribution.

> gridx <- seq(-4, 4, by = 0.01)

> par(mfrow=c(1,3))
> hist(U10, xlim=c(-4,4), ylim=c(0,0.6), prob=T)
> lines(gridx, dnorm(gridx), col=4)
> hist(U30, xlim=c(-4,4), ylim=c(0,0.6), prob=T)
> lines(gridx, dnorm(gridx), col=4)
> hist(U1000, xlim=c(-4,4), ylim=c(0,0.6), prob=T)
> lines(gridx, dnorm(gridx), col=4)

Exercise 3.6 (Drawing Sunspots)

1. The separator is a comma in the file.

> spot<-read.table("sunspots.csv",sep=",",header=T)
> summary(spot)
nb_spot year month day
Min. : 0.00 Min. :1749 Min. : 1.00 Min. :1
1st Qu.: 15.70 1st Qu.:1807 1st Qu.: 3.75 1st Qu.:1
Median : 42.00 Median :1866 Median : 6.50 Median :1
Mean : 51.27 Mean :1866 Mean : 6.50 Mean :1
3rd Qu.: 74.92 3rd Qu.:1925 3rd Qu.: 9.25 3rd Qu.:1
Max. :253.80 Max. :1983 Max. :12.00 Max. :1

2. Create the qualitative variable thirty:

> thirty <- floor((spot[,2]-1749+30)%/%30)

> thirty <- factor(thirty)

3. Check that the colours mentioned do indeed feature in the colour palette.

> color<-c("yellow","magenta","orange","cyan","grey","red",
"green","blue")
> all(color%in%colors())
[1] TRUE
290 R for Statistics

4. To draw the chronological series of Figure 3.34, we first construct the graph
without a curve and without points (argument type="n"). This makes it
possible to define the variation ranges of x and y along with the axes labels.
We thus draw parts of the curve one by one, changing the colour for each
category of thirty:

> palette(color)
> coordx <- seq(along=spot[,1])
> plot(coordx,spot[,1],xlab="Time",ylab="Number of sunspots",
type="n")
> for (i in levels(thirty)) {
> select <- thirty==i
> lines(coordx[select],spot[select,1],col=i)
> }

Exercise 3.7 (Plotting a Density)

1. To plot the curve for the normal distribution, first define the variation
interval of x:

> x <- seq(-3.5,3.5,length=1000)

> plot(x,dnorm(x),type="l",ylab="Density")

2. To draw a horizontal line, use abline and the argument h:

> abline(h=0)

35 For questions 3 to 5, use the functions polygon, arrows and text. In

order to be able to write mathematics using the text function, use expression:

> select <- x>=qnorm(0.95)

> absci <- c(x[select],rev(x[select]))
> ordon <- c(rep(0,sum(select)),rev(dnorm(x[select])))
> polygon(absci,ordon,col="blue")
> arrows(2.7,0.2,2,dnorm(2),len=0.1)
> text(2.7,0.2,expression(paste(alpha==5,"%")),pos=3)

Exercise 3.8 (Multiple Graphs)

1. To generate the graphs in Figure 3.36, we need to redefine the margins of
each graph using the par function. We also use the layout function to define
the layout of the three graphs:

> par(mar=c(2.3,2,0.5,0.3))
> layout(matrix(c(1,1,2,3), 2, 2, byrow = TRUE))
> plot(1:10,10:1,pch=0)
> plot(rep(1,4),type="l")
> plot(c(2,3,-1,0),type="b")
Appendix 291

2. The argument widths of layout is used to define the width of each column:

> par(mar=c(2.3,2,0.5,0.3))
> layout(matrix(c(1,1,2,3), 2, 2, byrow = TRUE),widths=c(4,1))
> plot(1:10,10:1,pch=0)
> plot(rep(1,4),type="l")
> plot(c(2,3,-1,0),type="b")
Exercise 3.9 (Plotting Points and Rows)
1. Read data from a file and check reading:

> ozone <- read.table("ozone.txt",header=T)

> summary(ozone)

2. Draw the graph using the xyplot function of the lattice package:

> library(lattice)
> xyplot(maxO3~T12,data=ozone)

3. Draw the graph and connect the points using the argument type=c("p","l")
(the graph looks disorganised as the data is not sorted):

> xyplot(maxO3~T12,data=ozone,type=c("p","l"))

4. Draw the graph using the argument type="a" which sorts the abscissa:

> xyplot(maxO3~T12,data=ozone,type="a")

The curve does not go through all the points, as the function smoothes the
curve before drawing it.
Exercise 3.10 (Panel)
1. Read the data and construct a data-frame (denoted data) with the first
three rows of maxO3 and T12:

> ozone <- read.table("ozone.txt",header=T)

> data <- ozone[1:3,c("T12","maxO3")]

2. Repeat the same data six times using a loop:

> prov <- data

> for (i in 1:5) prov <- rbind(prov,data)

3. Construct the ordinal factor, specifying the order of the levels using levels:

> type <- ordered(c(rep("h",3),rep("b",3),rep("p",3),rep("l",3),

rep("s",3),rep("S",3)),levels=c("p","l","b","h","s","S"))
292 R for Statistics

4. Add this ordinal factor to the prov data: here, a data-frame must be
created as some variables are quantitative and others qualitative.

prov <- cbind.data.frame(prov,type)

5. Reproduce the graph in Figure 3.10 using xyplot and the appropriate panel
function:

> mypanel <- function(x,y,subscripts) {

+ panel.xyplot(x,y,type=unique(type[subscripts]))
+ }
> xyplot(maxO3~T12|type,data=prov,panel=mypanel,xlab="",ylab="",
scales=list(draw=FALSE),layout=c(6,1))

6. We leave the reader to reproduce this graph using the factor function rather
than the ordered function in question 3.

A.5.4 Exercises: Chapter 4

Exercise 4.1 (Factorial)
1. Construction of the factorial function using the prod function:

> my.factorial <- function(n) {

+ if (n<0) stop("the integer must be positive")
+ if (n==0) return(1)
+ if (floor(n)!=n) warning(paste(n,"rounded to",floor(n)))
+ result <- prod(1:n)
+ return(result)
+ }

Here are three examples of the use of this function:

> my.factorial(4)
[1] 24
> my.factorial(4.2)
[1] 24
Message davis :
In my.factorial(4.2) : 4.2 rounded to 4
> my.factorial(-3)
Erreur dans my.factorial(-3) : the integer must be positive

2. Construction of the factorial function using a loop for:

> my.factorial <- function(n) {

+ if (n<0) stop("the integer must be positive")
+ if (n==0) return(1)
Appendix 293

+ if (floor(n)!=n){
+ warning(paste(n,"rounded to",floor(n)))
+ n <- floor(n)
+ }
+ result <- 1
+ for (i in 1:n) result <- result*i
+ return(result)
+ }

Exercise 4.2 (Ventilation)

1. Construct the ventilation function:

> ventilation <- function(Xfac,p=0.05) {

+ if (!is.factor(Xfac)) stop("Xfac must be a factor \n")
+ categories <- levels(Xfac)
+ if (length(categories)<=1) stop("not enough levels \n")
+ tabl <- table(Xfac)
+ select <- (tabl/sum(tabl))<p
+ if (!any(select)) return(Xfac) else {
+ sel <- categories[!select]
+ prov <- factor(Xfac[(Xfac%in%sel)],levels=sel)
+ prov <- table(prov)
+ freq <- prov/sum(prov)
+ for (j in categories[select]) {
+ ## sampling in the levels at random with replacement
+ if (length(sel)==1){
+ warning("1 level\n")
+ Xfac[Xfac==j]<-sel
+ } else Xfac[Xfac==j]<-sample(sel,sum(Xfac==j),
+ replace=T,prob=freq)
+ }
+ Xfacvent <- factor(as.character(Xfac))
+ }
+ return(Xfacvent)
+ }

2. We apply the function from the previous question to each column of the
table which is a factor:

> ventil.tab <- function (tab, threshold=0.05) {

+ for (i in 1:ncol(tab)) {
+ if (is.factor(tab[,i])) tab[,i]<-ventilation(tab[,i],p=threshold)
+ }
+ return(tab)
+ }
294 R for Statistics
Exercise 4.3 (Ventilation on an Ordered Factor)
1. The function is almost or exactly the same as that seen in the answer for
Exercise 2.7. To the output we add (return) and a few verifications:

> ventilation.y <- function(Xfac,p=0.05) {

+ if (!is.ordered(Xfac)) stop("Xfac must be ordered \n")
+ categories <- levels(Xfac)
+ if (length(categories)<=1) stop("not enough levels \n")
+ tabl <- table(Xfac)
+ selecti <- (tabl/sum(tabl))<p
+ if (!any(selecti)) return(Xfac) else {
+ numero <- which(selecti)
+ while(any((tabl/sum(tabl))<p)) {
.. paste here the inside of the while loop from Exercise 2.7
+ }
+ }
+ return(Xfac)
+ }

2. We apply the function from the previous question to each column of the
table which is an ordinal factor:

> ventil.y.tab <- function (tab, threshold=0.05) {

+ for (i in 1:ncol(tab)) {
+ if (is.ordered(tab[,i])) tab[,i]<-ventilation.y(tab[,i],threshold)
+ }
+ return(tab)
+ }
Bibliography

Breiman L., Friedman J., Olshen R. and Stone C. (1984). Classification and
Regression Trees. Wadsworth & Brooks, Monterey, CA.

Clarke G.M. and Cooke D. (2004). A Basic Course in Statistics. Wiley, New
York, 5 ed.
Collett D. (2003). Modelling Binary Data. Chapman & Hall/CRC, Boca
Raton, FL, 2 ed.

Devroye L., Gyorfi L. and Lugosi G. (1996). A Probabilistic Theory of Pattern

Recognition. Springer, New York.
Faraway J.J. (2005). Linear Models with R. Chapman & Hall/CRC, London.
Fox J. and Weisberg S.H. (2011). An R Companion to Applied Regression.
Sage Publications, London, 2 ed.
Govaert G. (2009). Data Analysis. Wiley, New York.
Greenacre M. (2007). Correspondence Analysis in Practice. Chapman &
Hall/CRC, Boca Raton, FL.

Greenacre M. and Blasius J. (2006). Multiple Correspondence Analysis and

related methods. Chapman & Hall/CRC, Boca Raton, FL.
Hastie T., Tibshirani R. and Friedman J. (2009). The Elements of Statistical
Learning: Data Mining, Inference, and Prediction. Springer, New York, 2
ed.

Hosmer D. and Lemeshow S. (2000). Applied Logistic Regression. Wiley, New

York, 2 ed.
Husson F., Le S. and Pages J. (2010). Exploratory Multivariate Analysis by
Example Using R. Chapman & Hall/CRC, London.

Jolliffe I. (2002). Principal Component Analysis. Springer, New York.

Kaufman L. and Rousseeuw P.J. (1990). Finding Groups in Data: An Intro-
duction to Cluster Analysis. Wiley, New York.

295
296 R for Statistics

Lebart L., Morineau A. and Warwick K.M. (1984). Multivariate Descriptive

Statistical Analysis: Correspondence Analysis and Related Techniques for
Large Matrices. Wiley series in probability and mathematical statistics:
Applied probability and statistics. Wiley, New York.
McLachlan G. (2004). Discriminant Analysis and Statistical Pattern Recog-
nition. Wiley, New York, 2 ed.

Moore D.S., McCabe G.P. and Craig B. (2007). Introduction to the Practice
of Statistics. Freeman, W. H. & Company, New Zealand, 6 ed.
Murrel P. (2005). R Graphics. Chapman & Hall/CRC, New Zealand.
Murtagh F. (2005). Correspondence Analysis and Data Coding with R and
Java. Chapman & Hall/CRC Press, London.

Sahai H. and Ageel M.I. (2000). The Analysis of Variance: Fixed, Random
and Mixed Models. Birkhauser Publisher, Boston, MA.
Varmuza K. and Filmozer P. (2009). Introduction to Multivariate Statistical
Analysis in Chemometrics. CRC Press, Boca Raton, FL.

Vinzi V., Chin W., Henseler J. and Wang H.e. (2010). Handbook of Partial
Least Squares. Springer, Boston, MA.
Welch B.L. (1951). On the comparison of several mean values: an alternative
approach. Biometrika, 38, 330336.
Statistics

R for Statistics
Although there are currently a wide variety of software packages suitable for
the modern statistician, R has the triple advantage of being comprehensive,
widespread, and free. Published in 2008, the second edition of Statistiques
avec R enjoyed great success as an R guidebook in the French-speaking world.
Translated and updated, R for Statistics includes a number of expanded and
additional worked examples.

Organized into two sections, the book focuses first on the R software, then on the
implementation of traditional statistical methods with R.

Focusing on the R software, the first section covers:

Basic elements of the R software and data processing
Clear, concise visualization of results, using simple and complex graphs
Programming basics: pre-defined and user-created functions

The second section of the book presents R methods for a wide range of traditional
statistical data processing techniques, including:

Regression methods Exploratory multivariate analysis

Analyses of variance and covariance Clustering methods
Classification methods Hypothesis tests

After a short presentation of the method, the book explicitly details the R
command lines and gives commented results. Accessible to novices and
experts alike, R for Statistics is a clear and enjoyable resource for any
scientist.

Datasets and all the results described in this book are available on the books webpage at
http://www.agrocampus-ouest.fr/math/RforStat

K13834