This is a repository copy of Research Methods Using R: Advanced Data Analysis in the

Behavioural and Biological Sciences.

White Rose Research Online URL for this paper:

https://eprints.whiterose.ac.uk/181926/

Version: Submitted Version

Book:
Baker, Daniel Hart orcid.org/0000-0002-0161-443X (2022) Research Methods Using R:
Advanced Data Analysis in the Behavioural and Biological Sciences. Oxford University
Press , (360pp).

Reuse
Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless
indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by
national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of
the full text version. This is indicated by the licence information on the White Rose Research Online record
for the item.

Takedown
If you consider content in White Rose Research Online to be in breach of UK law, please notify us by
emailing [email protected] including the URL of the record and the reason for the withdrawal request.

[email protected]
https://eprints.whiterose.ac.uk/
 Research Methods Using R: advanced data
analysis in the behavioural and biological sciences

Daniel H. Baker, University of York

March 2022
2

Acknowledgements
First and foremost, I would like to thank all of the students who have taken my
Advanced Research Methods module at the University of York over the past
7 years. You have all made an invaluable contribution to this work, by asking
questions, by suggesting new ways to think about methods, and through your
enthusiasm and perseverance on a difficult topic. This book is for you, and for
all future students who want to learn these techniques.
Many of my colleagues have had a huge influence on developing my own under-
standing of the topics covered in the book, especially Tim Meese, who as my PhD
supervisor taught me more than I can remember about Fourier analysis, signal
detection theory, and model fitting! I am further indebted to Tom Hartley, Alex
Wade, Harriet Over and Shirley-Ann Rueschemeyer, who provided invaluable
feedback on the material at various stages.
I am also very grateful to Martha Bailes and the team at Oxford University
Press, who have guided this project from an initial proposal through to a finished
product. Martha coordinated reviews from many anonymous reviewers from
diverse fields, whose comments and suggestions have strengthened the manuscript
immeasurably. I am further indebted to Bronte McKeown for checking the
technical content of the book, including the code and equations.
Finally, a huge thank you to my family, Gemma, Archie and Millie, who have
put up with me being buried in a laptop writing this book and tinkering with
the figures for far too long!

Note on versions and permissions

This document is a preprint, posted with permission of the publisher to the White
Rose institutional repository as ‘green’ open access. The text is the submitted
manuscript pre-copyediting and pre-typesetting, and therefore differs slightly
from the published version. The page numbers also differ between versions. The
citation to the version of record is:
Baker, D.H. (2022). Research Methods Using R: advanced data analysis in the be-
havioural and biological sciences, Oxford University Press, ISBN: 9780192896599.
Contents

1 Introduction 9
1.1 What is this book about? . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Who will find this book useful? . . . . . . . . . . . . . . . . . . . 10
1.3 Topics covered . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Some words of caution . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Implementation in R . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2 Introduction to the R environment 15

2.1 What is R? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 RStudio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Finding your way around RStudio . . . . . . . . . . . . . . . . . 17
2.4 Customising RStudio for your own needs . . . . . . . . . . . . . . 18
2.5 Scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Data objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.7 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.8 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.9 Conditional statements . . . . . . . . . . . . . . . . . . . . . . . . 26
2.10 Loops . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.11 Importing data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.12 How to find out how to do something . . . . . . . . . . . . . . . 30
2.13 Table of useful core R functions . . . . . . . . . . . . . . . . . . . 31
2.14 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3 Cleaning and preparing data for analysis 35

3.1 Why do data need to be ‘cleaned’? . . . . . . . . . . . . . . . . . 35
3.2 Organising data in wide and long formats . . . . . . . . . . . . . 35
3.3 Inspecting data: histograms and scatterplots . . . . . . . . . . . 37
3.4 Identifying outliers objectively . . . . . . . . . . . . . . . . . . . 40
3.5 Identifying missing and out of range values . . . . . . . . . . . . 47
3.6 What should we do with outliers? . . . . . . . . . . . . . . . . . . 50
3.7 Normalizing and rescaling data . . . . . . . . . . . . . . . . . . . 51
3.8 Transforming data and testing assumptions . . . . . . . . . . . . 52
3.9 Recoding categorical data and assigning factor labels . . . . . . . 56

3
4 CONTENTS

3.10 Putting it all together - importing and cleaning some real data . 57
3.11 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4 Statistical tests as linear models 65

4.1 Many statistical tests involve comparing different models . . . . . 65
4.2 Regression (and correlation) . . . . . . . . . . . . . . . . . . . . . 66
4.3 T-tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4 ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5 Assumptions of linear models . . . . . . . . . . . . . . . . . . . . 77
4.6 Is everything just a linear model then? . . . . . . . . . . . . . . . 77
4.7 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Power analysis 81
5.1 What is statistical power? . . . . . . . . . . . . . . . . . . . . . . 81
5.2 Effect size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3 How can we estimate effect size? . . . . . . . . . . . . . . . . . . 83
5.4 Power curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.5 Problems with low power . . . . . . . . . . . . . . . . . . . . . . 87
5.6 Problems with high power . . . . . . . . . . . . . . . . . . . . . . 88
5.7 Measurement precision impacts power . . . . . . . . . . . . . . . 89
5.8 Reporting the results of a power analysis . . . . . . . . . . . . . . 89
5.9 Post-hoc power analysis . . . . . . . . . . . . . . . . . . . . . . . 91
5.10 Doing power analysis in R . . . . . . . . . . . . . . . . . . . . . . 92
5.11 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 95

6 Meta analysis 97
6.1 Why is meta analysis important? . . . . . . . . . . . . . . . . . . 97
6.2 Designing a meta analysis . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Conducting and summarising a literature search . . . . . . . . . 99
6.4 Different measures of effect size . . . . . . . . . . . . . . . . . . . 101
6.5 Converting between effect sizes . . . . . . . . . . . . . . . . . . . 103
6.6 Fixed and random effects . . . . . . . . . . . . . . . . . . . . . . 104
6.7 Forest plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.8 Weighted averaging . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.9 Publication bias and funnel plots . . . . . . . . . . . . . . . . . . 107
6.10 Some example meta analyses . . . . . . . . . . . . . . . . . . . . 109
6.11 Calculating and converting effect sizes in R . . . . . . . . . . . . 110
6.12 Conducting a meta analysis in R . . . . . . . . . . . . . . . . . . 112
6.13 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 114

7 Mixed-effects models 117

7.1 Different types of mixed-effects model . . . . . . . . . . . . . . . 118
7.2 Mixed-effects regression example: lung function in bottlenose
dolphins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3 Factorial mixed-effects example: lexical decision task for nouns
and verbs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
CONTENTS 5

7.4 How can I decide if an effect is fixed or random? . . . . . . . . . 129

7.5 Dealing elegantly with missing data and unequal groups . . . . . 130
7.6 Reporting and comparing mixed-effects models . . . . . . . . . . 131
7.7 Practical problems with model convergence . . . . . . . . . . . . 132
7.8 Running mixed-effects models in R with lmerTest . . . . . . . . . 132
7.9 Further resources for mixed-effects models . . . . . . . . . . . . . 137
7.10 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 138

8 Stochastic methods 141

8.1 What does stochastic mean? . . . . . . . . . . . . . . . . . . . . . 141
8.2 Ways of generating random numbers . . . . . . . . . . . . . . . . 142
8.3 Part 1: Using random numbers to find stuff out . . . . . . . . . . 143
8.4 Part 2: Resampling methods . . . . . . . . . . . . . . . . . . . . 150
8.5 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.6 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 160

9 Nonlinear curve fitting 163

9.1 Fitting models to data . . . . . . . . . . . . . . . . . . . . . . . . 163
9.2 Linear models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.3 Nonlinear models . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.4 A practical example: exponential modelling of disease contagion 166
9.5 Parameter spaces and the combinatorial explosion . . . . . . . . 168
9.6 Optimization algorithms . . . . . . . . . . . . . . . . . . . . . . . 169
9.7 The Downhill Simplex algorithm . . . . . . . . . . . . . . . . . . 170
9.8 Local minima . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
9.9 Some practical considerations . . . . . . . . . . . . . . . . . . . . 173
9.10 Two philosophical approaches in model fitting . . . . . . . . . . . 174
9.11 Tools to help with model development . . . . . . . . . . . . . . . 174
9.12 How to fit curves to data in R . . . . . . . . . . . . . . . . . . . . 175
9.13 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 180

10 Fourier Analysis 183

10.1 Joseph Fourier: polymath . . . . . . . . . . . . . . . . . . . . . . 183
10.2 Example applications . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.3 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
10.4 The amplitude spectrum . . . . . . . . . . . . . . . . . . . . . . . 186
10.5 The phase spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.6 Limitations on sampling: the Nyquist limit and frequency resolution190
10.7 Example: bat species identification by frequency . . . . . . . . . 190
10.8 Fourier analysis in two dimensions . . . . . . . . . . . . . . . . . 191
10.9 Example: using 2D Fourier analysis to measure goosebumps . . . 194
10.10Filtering: altering signals in the Fourier domain . . . . . . . . . . 197
10.11Stimulus construction in the Fourier domain . . . . . . . . . . . . 198
10.12Doing Fourier analysis in R . . . . . . . . . . . . . . . . . . . . . 203
10.13Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6 CONTENTS

11 Multivariate t-tests 211

11.1 Thinking about and visualising bivariate data . . . . . . . . . . . 212
11.2 The one-sample and paired Hotelling’s T 2 statistic . . . . . . . . 212
11.3 Example: multivariate analysis of periodic EEG data . . . . . . . 217
11.4 The two-sample (independent) Hotelling’s T 2 statistic . . . . . . 217
11.5 Example: visual motor responses in zebrafish larvae . . . . . . . 219
2
11.6 The Tcirc statistic . . . . . . . . . . . . . . . . . . . . . . . . . . 221
11.7 Mahalanobis distance as an effect size measure for multivariate
statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
11.8 Calculating Hotelling’s T 2 in R . . . . . . . . . . . . . . . . . . . 222
11.9 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 225

12 Structural equation modelling 227

12.1 How are different mental abilities related? . . . . . . . . . . . . . 227
12.2 Testing hypotheses using data with structural equation modelling 231
12.3 SEM stage 1: model specification . . . . . . . . . . . . . . . . . . 232
12.4 SEM stage 2: model identification . . . . . . . . . . . . . . . . . 233
12.5 SEM stage 3: model evaluation . . . . . . . . . . . . . . . . . . . 234
12.6 SEM stage 4: model modificiation . . . . . . . . . . . . . . . . . 237
12.7 Comparing different models . . . . . . . . . . . . . . . . . . . . . 238
12.8 Cross-validation on fresh data sets . . . . . . . . . . . . . . . . . 239
12.9 Power and SEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
12.10Dealing with missing data . . . . . . . . . . . . . . . . . . . . . . 239
12.11Doing SEM in R using the lavaan package . . . . . . . . . . . . . 240
12.12Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 245

13 Multidimensional scaling and k-means clustering 247

13.1 The k-means clustering algorithm . . . . . . . . . . . . . . . . . . 249
13.2 Comparing different numbers of clusters . . . . . . . . . . . . . . 249
13.3 Example: k-means clustering of dinosaur species . . . . . . . . . 253
13.4 Variants of k-means clustering . . . . . . . . . . . . . . . . . . . . 254
13.5 The multidimensional scaling algorithm . . . . . . . . . . . . . . 255
13.6 Metric vs non-metric MDS . . . . . . . . . . . . . . . . . . . . . . 258
13.7 Example: multidimensional scaling of viruses . . . . . . . . . . . 258
13.8 Combining k-means clustering and MDS . . . . . . . . . . . . . . 258
13.9 Normalizing multivariate data . . . . . . . . . . . . . . . . . . . . 262
13.10Examples from the literature of k-means clustering and MDS . . 262
13.11Doing k-means clustering in R . . . . . . . . . . . . . . . . . . . 263
13.12Doing multidimensional scaling in R . . . . . . . . . . . . . . . . 266
13.13Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 268

14 Multivariate pattern analysis 271

14.1 Why use machines? . . . . . . . . . . . . . . . . . . . . . . . . . . 271
14.2 Predicting group membership . . . . . . . . . . . . . . . . . . . . 272
14.3 Different types of classifier algorithm and pattern analysis . . . . 273
14.4 Situations with more than two categories . . . . . . . . . . . . . 277
CONTENTS 7

14.5 Preparing data for analysis . . . . . . . . . . . . . . . . . . . . . 277

14.6 Assessing statistical significance . . . . . . . . . . . . . . . . . . . 278
14.7 Ethical issues with machine learning and classification . . . . . . 278
14.8 Doing MVPA in R using the Caret package . . . . . . . . . . . . 279
14.9 Categorising cell body segmentation . . . . . . . . . . . . . . . . 280
14.10Decoding fMRI data . . . . . . . . . . . . . . . . . . . . . . . . . 284
14.11Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 286

15 Correcting for multiple comparisons 289

15.1 The problem of multiple comparisons . . . . . . . . . . . . . . . . 289
15.2 The traditional solution: Bonferroni correction . . . . . . . . . . 291
15.3 A more exact solution: Sidak correction . . . . . . . . . . . . . . 292
15.4 A higher power solution: the Holm-Bonferroni correction . . . . . 292
15.5 Adjusting α reduces statistical power . . . . . . . . . . . . . . . . 293
15.6 Controlling the false-discovery rate . . . . . . . . . . . . . . . . . 294
15.7 Comparison of FDR vs FWER correction . . . . . . . . . . . . . 295
15.8 Cluster correction for contiguous measurements . . . . . . . . . . 297
15.9 Example of cluster correction for EEG data . . . . . . . . . . . . 298
15.10Correcting for multiple comparisons in R . . . . . . . . . . . . . 304
15.11Implementing cluster correction in R . . . . . . . . . . . . . . . . 307
15.12Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 312

16 Signal detection theory 315

16.1 The curious world of chicken sexing . . . . . . . . . . . . . . . . . 315
16.2 Hits, misses, false alarms and correct rejections . . . . . . . . . . 316
16.3 Internal responses and decision criteria . . . . . . . . . . . . . . . 318
16.4 Calculating sensitivity and bias . . . . . . . . . . . . . . . . . . . 321
16.5 Radiology example: rating scales and ROC curves . . . . . . . . 325
16.6 Removing bias using forced choice paradigms . . . . . . . . . . . 327
16.7 Manipulating stimulus intensity . . . . . . . . . . . . . . . . . . . 329
16.8 Type II signal detection theory (metacognition) . . . . . . . . . . 331
16.9 Practical problems with ceiling performance . . . . . . . . . . . . 331
16.10Calculating d′ in R . . . . . . . . . . . . . . . . . . . . . . . . . . 332
16.11Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 333

17 Bayesian statistics 337

17.1 The philosophy of frequentist statistics . . . . . . . . . . . . . . . 337
17.2 An unfortunate consequence of the fixed Type I error rate . . . . 338
17.3 Bayesian statistics: an alternative approach . . . . . . . . . . . . 340
17.4 The legacy of Thomas Bayes: Reverend . . . . . . . . . . . . . . 340
17.5 Bayes’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
17.6 Probability distributions . . . . . . . . . . . . . . . . . . . . . . . 346
17.7 The Bayesian approach to statistical inference . . . . . . . . . . . 348
17.8 Heuristics for Bayes Factor scores . . . . . . . . . . . . . . . . . . 349
17.9 Bayes factors can support the null hypothesis . . . . . . . . . . . 349
17.10Bayesian implementations of different statistical tests . . . . . . . 350
8 CONTENTS

17.11So which approach is the best? . . . . . . . . . . . . . . . . . . . 351

17.12Software for performing Bayesian analyses . . . . . . . . . . . . . 351
17.13Calculating Bayes Factors in R using the BayesFactor package . 352
17.14Further resources on Bayesian methods . . . . . . . . . . . . . . 357
17.15Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 357

18 Plotting graphs and data visualisation 361

18.1 Four principles of clear data visualisation . . . . . . . . . . . . . 362
18.2 Choosing colour palettes . . . . . . . . . . . . . . . . . . . . . . . 373
18.3 Building plots from scratch in R . . . . . . . . . . . . . . . . . . 378
18.4 Adding data to plots . . . . . . . . . . . . . . . . . . . . . . . . . 380
18.5 Setting alpha transparency for points and shapes . . . . . . . . . 382
18.6 Choosing, checking and using colour palettes . . . . . . . . . . . 383
18.7 Saving graphs automatically . . . . . . . . . . . . . . . . . . . . . 387
18.8 Combining multiple graphs . . . . . . . . . . . . . . . . . . . . . 390
18.9 Adding raster images to a plot . . . . . . . . . . . . . . . . . . . 393
18.10Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 395

19 Reproducible data analysis 399

19.1 Version control of analysis scripts . . . . . . . . . . . . . . . . . . 400
19.2 Writing code for others to read . . . . . . . . . . . . . . . . . . . 402
19.3 Using R Markdown to combine writing and analysis . . . . . . . 404
19.4 Automated package installation . . . . . . . . . . . . . . . . . . . 404
19.5 Identifying a robust data repository . . . . . . . . . . . . . . . . 405
19.6 Automatically uploading to and downloading from OSF . . . . . 406
19.7 Future proofing with open data formats . . . . . . . . . . . . . . 408
19.8 Providing informative meta-data . . . . . . . . . . . . . . . . . . 409
19.9 Practice questions . . . . . . . . . . . . . . . . . . . . . . . . . . 410

20 Endnotes 413
20.1 A parting note . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
20.2 Answers to practice questions . . . . . . . . . . . . . . . . . . . . 414
20.3 Alphabetical list of key R packages used in this book . . . . . . . 425

21 References 429
Chapter 1

Introduction

1.1 What is this book about?

Welcome to Research Methods Using R. The aim of this book is to take techniques
that often seem inaccessible, and explain the basic concepts in plain language.
This is, in essence, a statistics textbook. However, it is somewhat different in
flavour from other statistics texts, which in my experience fall into two camps.
Most numerous are the introductory texts, which cover a range of standard
methods (descriptive statistics, t-tests, ANOVA, correlation, regression). There
are many excellent examples, including Learning Statistics with R by Danielle
Navarro (Navarro 2019), and the Discovering Statistics series by Andy Field
and colleagues (Field, Miles, and Field 2012). These introductory books are
appropriate for instruction at undergraduate and masters level in many disciplines
within the broad umbrella of the Natural and Social Sciences (for example
biology, chemistry, neuroscience, psychology and sociology), and knowledge of
the methods they explain is a requirement of accreditation for some courses.
The methods covered in this book do not necessarily require an understanding
of these techniques, but I anticipate that most readers will at some point have
completed an introductory statistics course that will have explained them. It
would certainly not hurt to have a copy of an introductory text to hand whilst
working through this book.
At the other extreme are specialised advanced statistics texts, filled with complex
equations and aimed at trained statisticians. These books (and papers) are
the absolute authority on most methods, yet they frequently assume a level of
understanding and mathematical competence which frankly I do not possess!
It is worth confessing from the outset that I have no advanced mathematical
training, and usually aim to understand techniques at a conceptual level, rather
than understand every detail of the underlying mathematics. I will attempt
to do the same here, substituting formal equations for diagrams and verbal

9
10 CHAPTER 1. INTRODUCTION

explanations wherever possible.

What I have aimed to produce here is a text that can be used as a starting
point for the reader’s use of particular techniques - a ‘way in’ to a method, or
‘primer’. We will introduce each method by explaining the theory behind it, and
then walk through some example computer code demonstrating how it can be
implemented. Often seeing a working implementation of a technique is the best
way to understand how to apply it to your own research. In a sense this is a bit
like a recipe book, with sections of code that can be adapted into the reader’s
own scripts to solve a particular problem.

1.2 Who will find this book useful?

I anticipate that advanced undergraduate and postgraduate students in the life
sciences will be the most obvious audience for this book. Many of the tools are
widely used in cognition, perception, neuroscience and related topics, but there
are examples from many different disciplines throughout the book. Students
beginning a PhD in one of these topics will find much that is relevant - I definitely
would have when I started my own PhD. It certainly doesn’t hurt to already
have some statistical training, e.g. from an undergraduate level stats course.
However at the point I started my PhD, I had forgotten most of the things I
learned as a undergraduate, so this is probably not essential!
Because all of the methods are implemented in a statistical programming language
called R, the reader should expect to develop some programming expertise
throughout the book. Chapters 2 - 4 go through some of the basics of the
R language, and demonstrate how to do some of the things I expect most
readers to already be familiar with. For this reason, I think the text is suitable
for a complete beginner who has never programmed before, though previous
programming experience in any language should make the learning process faster.
Programming often seems a daunting prospect if you have never attempted it
before, and there are many myths and stereotypes that frequently put people
off. One of the most pernicious is that programming is a ‘male’ activity. This
could not be further from the truth, and the history of computer programming,
from Ada Lovelace (the first programmer) to the black female programmers who
contributed to the American Space Program (see the film Hidden Figures), is
filled with exemplary female programmers. The R community is explicitly and
deliberately inclusive and supportive - see for example the brilliant R-Ladies
organisation (https://rladies.org/) that organise meetups across the world, as
well as online events, to promote gender diversity in the R community.
I think of learning programming as like learning a foreign language. It takes
some time at first to understand the basic grammar and learn the vocabulary,
but once you get the hang of the basics, programming is a very creative, and
sometimes even enjoyable, skill. In the modern world it is also an extremely
marketable skill, and well worth including on one’s CV. So, to any readers who
1.3. TOPICS COVERED 11

feel worried about the prospect of learning to program: persevere, and believe in
yourself. Anyone can learn to program if they put in the time and effort. It is
not magic, and it is not beyond you, but it is a skill that takes practice to get
the hang of.

1.3 Topics covered

The scope of this book is a set of computational techniques that are useful in
many domains of research. Because my background is in experimental psychology
and perception research, most of the methods are applicable those disciplines.
However many of the techniques are also very relevant to other fields of science.
My expectation is that most readers will dip into and out of different topics,
rather than read the whole book from start to finish. To this end, chapters
attempt to be as self-contained as possible, though most will build on (and
assume knowledge of) the basic programming skills covered in Chapter 2. Most
chapters begin by explaining the conceptual and theoretical uses of a technique,
and end with an example implementation in the statistical programming language,
R.
Chapter 2 introduces the R environment, and covers basic programming con-
cepts including data objects, loops, conditional statements and functions. Chap-
ter 3 discusses how to inspect and process raw data so that it is ready for further
analysis. Chapter 4 demonstrates how basic statistics that I expect to be famil-
iar to most readers (e.g. t-tests, ANOVAs, correlations etc.) can be implemented
in R.
Chapter 5 discusses statistical power analysis - a technique that can be used
to determine how many participants should be included in an experiment. This
method is becoming increasingly important in the context of the ‘replication
crisis’, where many influential effects have failed to replicate, in part because
the original studies were underpowered.
Chapter 6 introduces meta analysis - a method for combining the results of
previous studies on a single topic to estimate the overall effect size. Meta
analysis is most widely used in biomedical research, for example to synthesise
the evidence regarding the effectiveness of different medicines. However the same
tools can be used in other types of research, and are increasingly applied to
address fundamental empirical questions.
Linear mixed-effects models are described in Chapter 7. These methods are
an extension of the general linear model used in ANOVA and regression, but
they are more flexible and can deal more robustly with missing data, as well as
explicitly modelling variance at multiple levels.
Chapter 8 introduces a class of techniques I refer to as stochastic methods.
These involve using random numbers to simulate experiments, to estimate the
robustness of data by resampling, and to model systems that have random
12 CHAPTER 1. INTRODUCTION

components. This is an important core skill, useful in many areas of research,

but the most common application is to ‘bootstrap’ some statistical test by
repeating it many times on resampled versions of the original data.
Chapter 9 focuses on nonlinear curve fitting and function minimization; in
plain language, we will see how to fit models to data that are not well described
by a straight line. This is also a useful generic technique, with applications in
many areas of science, such as fitting models to empirical data, and predicting
results in conditions that have not yet been tested.
Chapter 10 explains Fourier analysis, a mathematical technique for analysing
periodic signals. This is widely used to analyse time-varying data (such as
electromagnetic brain activity), and in the analysis of sound samples (e.g. of
speech or music). Fourier analysis can be applied to images and even movies,
and filtering operations are described that can be used to smooth noisy data, or
modify experimental stimuli.
Next follow four chapters on multivariate statistics - methods that can be used
when you have more than one dependent measure. In Chapter 11 we discuss
multivariate versions of the t-test such as Hotelling’s T 2 statistic that can be
used to compare conditions for which multiple dependent variables have been
measured. Structural Equation Modelling (Chapter 12) is an extension of
factor analysis and correlation, and is used to understand the structure and
relationships between sets of variables. Multidimensional scaling and k-means
clustering (Chapter 13) are methods for grouping complex multidimensional
data and visualising it in simpler ways. Finally, multivariate pattern analysis
(MVPA; Chapter 14) is a machine learning technique that is used to categorise
data into two or more groups, and is increasingly important in many research
areas, including neuroimaging research, and personalised medicine.
Correction for multiple comparisons is discussed in Chapter 15. We introduce
traditional methods for adjusting the threshold for significance to avoid detecting
effects that are not real (false positives). We then introduce two newer approaches
- false-discovery rate correction and cluster correction. These avoid overly
stringent correction of significance criteria, which can reduce statistical power
and obscure real effects.
In Chapter 16, we explain signal detection theory, a fundamental theory that
underlies research in many disciplines, including perception and memory research,
but is now increasingly applied in the machine learning and artificial intelligence
literature. We particularly focus on forced choice methods, which are used
extensively in empirical work, but often sidelined in textbooks.
Bayesian statistics are introduced in Chapter 17. The Bayesian approach uses
a fundamentally different philosophy from traditional ‘frequentist’ statistics. In
general it is better at dealing with null results and Type I errors (false positives).
There are Bayesian versions of many common statistical tests.
In the final chapters, we discuss some practical considerations. In Chapter 18
1.4. SOME WORDS OF CAUTION 13

we cover some techniques for graph plotting and data visualisation. Unlike many
works on graph plotting in R, we use the ‘base’ plotting functions instead of the
popular ggplot2 package. In Chapter 19, we discuss issues around reproducible
data analysis, including methods of version control, open data formats, and
automatically downloading and uploading data from public repositories.

1.4 Some words of caution

Reading a book on statistics makes you an expert statistician the same way that
reading a book on juggling makes you an expert juggler. It doesn’t. The only way
to learn is by doing, by practising, by making mistakes, and by working things
out for yourself. The same goes for programming, and most other worthwhile
skills. This book aims to give you a grounding in the theory of how the methods
work, and some examples of how to implement them. Applying this to your own
research will be challenging and it will take time, but that is the only way to
really learn.

A further caveat: this book is not meant to be the final word on any method
or technique. It is intended to be an introduction (a primer) to the methods,
and a starting point for your own reading and learning. So there will be lots
of things I don’t mention, and probably many faster or more efficient ways of
programming something. Also, the R community is very fast moving, and new
packages are being created all the time. So it is worthwhile checking online for
package updates with new features, and also for more recent packages for a given
method.

1.5 Implementation in R
Most of the methods covered in this book are not implemented in commercial
statistics packages such as SPSS and SAS. Instead, we use a statistical pro-
gramming language called R for all examples (R Core Team 2013). R has the
advantage that it is an open source language, so anyone can develop their own
packages (collections of code that implement statistical tests) for others to use. It
is now standard practise for papers describing a new statistical technique to have
an associated R package for readers to download. R is free to download, and can
be installed on most operating systems. There are also some online R clients that
can be accessed directly through a web browser without needing any installation.
I mostly used another language, Matlab, throughout my PhD and Postdoctoral
years, and many of the methods we discuss here can also be implemented in
Matlab, or other programming languages such as Python. However, in the
interests of consistency, we use a single language throughout, which is introduced
in Chapter 2. I have also made the R code for all examples and figures available
on a GitHub repository at: https://github.com/bakerdh/ARMbookOUP.
14 CHAPTER 1. INTRODUCTION

1.6 History
This book has grown from a lecture course that I developed in 2014, and
deliver twice each year at the University of York. The course is taught to
third year undergraduates, and is often audited by postgraduate students on
MSc and PhD courses. Many former students have gone on to have successful
careers in numerate areas. I try to make the material as accessible as possible,
and sometimes use props like 3D printed surfaces to demonstrate function
minimisation, and a bingo machine to explain bootstrapping! The biggest
concern students have is about whether they will understand the mathematical
content for advanced statistical methods. For this reason I deliberately avoid
equations where possible, preferring instead to explain things at a conceptual
level, and I have done the same in this book.
The topics taught on the module go beyond the core undergraduate research
methods syllabus, and so are rarely discussed in typical textbooks. Up until now,
I have mostly used tutorial papers and package manuals as the recommended
reading. But this is not ideal, and I always felt that students on the module
would benefit from a single text that presents everything in a common style.
Producing a written explanation of the course content seemed the natural next
step, and my publishers agreed! The original lecture content equates to about
half of the topics included here, with other content added that seemed related
and useful.
Chapter 2

Introduction to the R
environment

This chapter introduces the R programming language. It can be safely skipped by

those already well-versed in R. However it is worth at least skim-reading for those
new to the language, even if they have other programming experience. We will
first introduce the language (section 2.1) and developer environment (sections
2.2 and 2.3), and explain how to adjust the appearance to suit your needs
(section 2.4). We will then go through examples of R syntax for fundamental
programming concepts such as scripts (section 2.5), data objects (section 2.6),
functions (section 2.7), packages (section 2.8), conditional statements (section
2.9) and loops (section 2.10). Finally, we will discuss how to import data
(section 2.11) and solve problems (section 2.12), finishing with a table of useful
R functions (section 2.13) and some practice questions (section 2.14).
Note that this chapter is intended to cover the basics of R in enough detail
for you to understand the rest of the book. However, if it seems overwhelming
you might also find it helpful to look at some online resources that teach the
basics of R at a slower pace. There are many free resources available (including
instructional videos), as well as some paid-for courses, and both online and print
textbooks. This is a rapidly evolving area, so it is worth seeing what is available
at the time of reading. I would suggest starting with this chapter, however, as it
may be sufficient to progress through rest of the material in this book.

2.1 What is R?
R is a statistical programming language. It has been around since the early
1990s, but is heavily based on an earlier language called S. It is primarily an
interpreted language, meaning that we can run programs directly, rather than
having to translate them into machine code first (though a compiler which does

15
16 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

this is also available). Its particular strengths lie in the manipulation, analysis
and visualisation of data of various kinds. Because of this focus, there are other
tasks that it is less well-suited to - you probably wouldn’t use it to write a
computer game, for example. However, it is rapidly increasing in popularity and
is currently the language of choice for statisticians and data analysts. We will
use it exclusively throughout this book for all practical examples.
The core R language is maintained by a group of around 20 developers known
as the R Development Core Team. The language is freely available for all major
operating systems (Windows, Mac and various flavours of Linux), and the R
Foundation is a not-for-profit organisation. This means that R is an inherently
free software project - nobody has to pay to use it, and there is no parent company
making huge profits from it. This is quite different from many other well-known
programming languages and statistical software packages that you may have
come across. It means that R is available to anyone with a computer and
internet connection, anywhere in the world, regardless of institutional affiliation
or financial circumstances. This seems to me an inherently good thing.
You can download R from the R project website. It is hosted by the Compre-
hensive R Archive Network (CRAN), a collection of around 100 servers, mostly
based at Universities across the world. The CRAN servers all mirror the same
content, ensuring that if one goes down the software is still available from the
rest. The CRAN mirrors also contain repositories of R packages, which we will
discuss further in section 2.8. If you do not have R installed on your computer
and plan to start using it, now would be a good time to download it. Exactly
how the installation works depends on your computer and its operating system,
but instructions are available for all systems at the R project website.

2.2 RStudio
In parallel with the development of the core R language, a substantial amount
of work has been done by a company called RStudio. This is a public benefit
corporation that makes some money from selling things like ‘pro’ and enterprise
versions of its software, web hosting and technical support. However its primary
product, the RStudio program, is free and open source. RStudio (the program)
is an integrated development environment (IDE) for R. It has a number of user-
friendly features that are absent in the core R distribution, and is now the most
widely-used R environment (again it is available for all major operating systems).
I strongly recommend downloading and installing it from the RStudio website.
Note that RStudio requires that you already have a working R installation on
your computer, as it is a separate program that sits ‘on top of’ R itself. So
you need to first install R, and then install RStudio in order for it to work. If
you have insurmountable problems installing R on your computer, there are
now web-based versions that run entirely through a browser and do not require
installation, such as RStudio Cloud and rdrr.io, though these services may not
be free.
2.3. FINDING YOUR WAY AROUND RSTUDIO 17

The RStudio company has also produced a number of very well-written and
useful R packages designed to do various things. For example, the Shiny package
can produce dynamic interfaces to R code that can run in web browsers. The
RMarkdown package can produce documents that combine text, computer code
and figures (discussed further in section 19.3). A version of it (bookdown) was used
to create the first draft of this book. Finally, a suite of tools collectively referred
to as the Tidyverse offer a uniform approach to organising and manipulating
data (essentially storing even very complex data structures in a spreadsheet-like
format). Although many introductions to R now focus on these tools, they are
advanced-level features that are not required to implement the examples in this
book, and so we will not discuss them further.

2.3 Finding your way around RStudio

On my (Apple Macintosh) computer, the default window for RStudio looks like
that shown in Figure 2.1. In the lower left corner is the Console. You can use
this section to type in commands, which R will execute immediately. One way
to think of it is a bit like a calculator. In fact, you can type in basic sums and it
will give you the answer. Try it now - type in 1 + 1 and press return. Hopefully
it will give you the answer, 2. All code that is executed during an R session is
echoed to the console, along with any output such as the results of statistical
tests.

Figure 2.1: Example screenshot of the default RStudio window.

An alternative to entering commands directly at the console is to write a script

containing instructions. We will go into more detail about scripts in section
18 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

2.5, but for now note that they appear in the upper left corner of the RStudio
window.

In the upper right corner is the Environment tab, which contains a summary of
all of the information currently stored in R’s memory. This will usually be empty
when you first launch R, but will fill up with things like data sets, and the results
of various analyses. There are other tabs in the upper right corner, including the
History tab which keeps track of all commands executed in the current R session.
The Environment and History tabs can be saved to files (using the disk icon),
and also loaded back in from a previous session. The broom icon, which appears
in several panels, empties the contents of the Environment (or other section).

The lower right corner of the RStudio window contains several more tabs. The
first is the Files tab, which contains a file browser, allowing you to open scripts
and data files from within RStudio. The second tab is the Plots tab, which will
display any graphs you create. You can also export graphs from this window,
and navigate backwards and forwards through multiple graphs using the left
and right arrow buttons.

The next tab in the lower right pane is the Packages list, which contains all
of the packages of R code currently installed on your computer. We discuss
packages in more detail in Section 2.8. The Help tab is also in this section of
the window, and is used to display help files for functions and packages when
you request them.

As with most other desktop applications, there are a number of drop down
menus that allow you to do various things, such as load and save files, copy and
paste text, and so on. I use these much less frequently than in other programs,
largely because many of the functions are duplicated in the various toolbars in
the main RStudio window. Overall, the RStudio package provides a user-friendly
environment for developing and running R code.

2.4 Customising RStudio for your own needs

In common with other contemporary software, RStudio has several customisable
options to improve accessibility. For example, in the Preferences menu (accessed
through the File menu) there is a section called Appearance. This allows you
to change the font and text size, and also to zoom the entire window in and
out. You can choose between a wide range of different colour themes for the
application. These are helpful because they change the text colour for different
components of an R script, which can make code easier to navigate. It is also
possible to change the background colour, for users who find reading easier
against a particular background. Finally, there are options to change the layout
of the different panels in the RStudio window, and features for spell checking
and autocompleting text. Since different users will have different requirements,
it is best to have a look at these settings and see what works best for you.
2.5. SCRIPTS 19

2.5 Scripts
A script is a text file that contains computer code. Instead of entering commands
one at a time into the R console, we can type them into a script. This allows
us to save our work and return to it later, and also to share scripts with others.
Although this sounds like a simple concept, it has enormous benefits. First of
all, we have an exact record of exactly what we did in our analysis that we can
refer back to later. Second of all, someone else could check the code for errors,
reproduce your analysis themselves and so on. This is very different from the
way most people use statistical packages with graphical interfaces like SPSS
(although technically it is possible to record the analysis steps in these packages,
researchers rarely do it in practise). Sharing scripts online through websites like
the Open Science Framework is becoming commonplace, as part of a drive for
openness and reproducibility in research, as we will discuss in greater detail in
Chapter 19.
In RStudio we can create a new script through the File menu. These traditionally
have the file extension .R - however they are just plain text files, so you can
open and edit them in any normal text editor if you want to. It is a good idea
to create scripts for everything you do in R and save them somewhere for later
reference. R works slightly differently from some other programming languages,
in that you can easily run chunks of code from a script without having to run
the whole thing. You do this by highlighting the section of code you wish to
execute (using the mouse) and then clicking the Run icon (the green arrow at
the top of the script panel). You will see the lines of code reproduced in the
Console, along with any output.
Another way to run a script is to use the source command, and provide it with
the location and file name of the script you wish to run. This will execute the
entire script with a single command, as follows:
source('~/MyScripts/script.R')

You can use the source command from the console, and also include it in a script
so that you execute the code from multiple scripts in sequence.

2.6 Data objects

A key concept in all programming languages is the data object (sometimes called
a variable, though we will not this term here because of the potential for confusion
with variables in the context of statistics). You can think of a data object as
a container that you can keep information in. The information will usually be
numbers or text. In some programming languages it is important to specify in
advance what the data object is going to contain, but R will usually sort this out
for you without you having to worry about it. Every data object has a name,
which can be anything you like as long as it doesn’t start with a number, or
contain any spaces or reserved characters (some symbols such as @). We will
20 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

discuss choosing good names in section 19.2.

In general it is advisable to give data objects a meaningful name. A data object
containing your birthday could be called MyBirthday, rather than something
arbitrary such as a. To create a data object and store some information in it,
we traditionally use the <- operator, though current versions of R also permit
the = operator. Here is a simple example:
a <- 10

In the above line of code, we have created a data object called a and stored the
number 10 in it. We read the <- opearator as ‘is given the value’. So the above
example means ‘a is given the value 10’. This data object will then appear in
the Environment pane of the RStudio window. If we want the data object to
contain a string of text, we wrap the text in inverted commas (quotes) so that R
knows it is text data, rather than a reference to another data object:
b <- 'Hello'

Once a data object has been created, we can use it in calculations, in much the
same way that mathematicians use letters to represent numbers in algebraic
expressions. For example, having stored the number 10 in the object a, we can
multiply it by other numbers as follows:
a*2 # multiply the contents of 'a' by 2

## [1] 20
Data objects can contain more than one piece of information. A list of numbers
is called a vector, and can be generated in several ways. A sequence of integers
can be defined using a colon:
numvect <- 11:20
numvect

## [1] 11 12 13 14 15 16 17 18 19 20
Or, we can combine several values using the c (concatenate) operation, which
here has the same result:
numvect <- c(11,12,13,14,15,16,17,18,19,20)
numvect

## [1] 11 12 13 14 15 16 17 18 19 20
Just as for a single value, we can perform operations on the whole vector of
numbers. For example:
numvect^2 # raise the contents of 'numvect' to the power 2

## [1] 121 144 169 196 225 256 289 324 361 400
The above code calculates the square of each value in the data object numvect
2.6. DATA OBJECTS 21

(the ˆ symbol indicates raising to a power). We can also request (or index)
particular values within a vector, using square brackets after the name of the
data object. So, if we want to know just the fourth value in the numvect object,
we can ask for it like this:
numvect[4] # return the 4th value in 'numvect'

## [1] 14
If we want a range of values we can index them using the colon operator:
numvect[3:8] # just entries 3 to 8 of the vector

## [1] 13 14 15 16 17 18
And if we want some specific entries that are not contiguous, we can again use
the c (concatenate) function:
numvect[c(1,5,7,9)] # some specific entries from the vector

## [1] 11 15 17 19
Finally, we can use other data objects as our indices. For example:
n <- 6
numvect[n]

## [1] 16
Data objects can also have more than one dimension. A two-dimensional data
object is like a spreadsheet with rows and columns, and is referred to as a matrix.
We need to tell R how big a matrix is going to be so that it can reserve the right
amount of memory to store it in. The following line of code generates a matrix
with ten rows and ten columns, storing the values from 1 to 100:
d <- matrix(1:100, nrow=10, ncol=10)
d

## [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
## [1,] 1 11 21 31 41 51 61 71 81 91
## [2,] 2 12 22 32 42 52 62 72 82 92
## [3,] 3 13 23 33 43 53 63 73 83 93
## [4,] 4 14 24 34 44 54 64 74 84 94
## [5,] 5 15 25 35 45 55 65 75 85 95
## [6,] 6 16 26 36 46 56 66 76 86 96
## [7,] 7 17 27 37 47 57 67 77 87 97
## [8,] 8 18 28 38 48 58 68 78 88 98
## [9,] 9 19 29 39 49 59 69 79 89 99
## [10,] 10 20 30 40 50 60 70 80 90 100
Again, we can index a particular value. For example, if we want the number
from the 8th row and the 4th column, we can ask for it by adding the indices
22 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

8,4 in square brackets after the data object name:

d[8,4]

## [1] 38
This is very similar to the way you can refer to rows and columns in spreadsheet
software such as Microsoft Excel, except that in R rows and columns are both
indexed using numbers (whereas Excel uses letters for columns). If we want
all of the rows or all of the columns, omitting the number (i.e. leaving a blank
before or after the comma) will request this, for example:
d[8,] # row 8 with all columns

## [1] 8 18 28 38 48 58 68 78 88 98
d[,4] # column 4 with all rows

## [1] 31 32 33 34 35 36 37 38 39 40
We can also request a range of values for rows and/or columns using the colon
operator:
d[1:3,5:7] # rows 1:3 of columns 5:7

## [,1] [,2] [,3]

## [1,] 41 51 61
## [2,] 42 52 62
## [3,] 43 53 63
Data objects are not limited to having only two dimensions. In R, objects with
three or more dimensions are called arrays, and will be introduced when required
throughout the book.
A particularly useful class of data object in R is the data frame. This is very
similar to a matrix, but each column can contain a different type of data (whereas
in a matrix we cannot combine numbers and text in different columns). Columns
and rows can have headings to help identify what they contain. This is very
similar indeed to spreadsheets in software packages like Excel, and this is a helpful
way to think about them. The following code produces a simple data frame
containing the top 5 greatest songwriters according to Rolling Stone magazine:
Position <- 1:5
Songwriter <- c('Bob Dylan','Paul McCartney',
'John Lennon','Chuck Berry','Smokey Robinson')
chart <- data.frame(Position, Songwriter)
chart

## Position Songwriter
## 1 1 Bob Dylan
## 2 2 Paul McCartney
## 3 3 John Lennon
2.7. FUNCTIONS 23

## 4 4 Chuck Berry
## 5 5 Smokey Robinson
In the above code, we first created two vectors called Position and Songwriter.
Then, we used the data.frame function to combine these together into a single
data object, that I’ve called chart. When we look at the chart object (by typing
its name), we can see that the data are organised into two columns and five rows.
The first column contains the chart position, and the second column contains
the name of the songwriter. The Environment panel will also contain the two
vectors and the data frame, as shown in Figure 2.2.

Figure 2.2: Example screenshot of the Environment panel, showing two vectors,
and also a data frame that contains both vectors.

Data frames can have as many rows and columns as you like, so they are a
very general and flexible way to store and manipulate data of different kinds.
There are numerous other classes of data object in R, and new classes can be
defined when required, so we will discuss any other data types as they come up
throughout the book.
It is important to have a good conceptual understanding of data objects, because
this is how R stores information. If you load in a data set from an external file,
this will be stored in a data object. You might, for example, have a spreadsheet
file containing questionnaire responses, which you could load into R and store in
a data frame called questionnaire. Similarly, when you run a statistical test, the
results will typically be stored in a data object as well. If you run a t-test (see
Chapter 4 for details of how to do this), this will produce a data object that you
might call ttestresults, and will contain the t-statistic, the p-value, the degrees of
freedom, and lots of other useful information.

2.7 Functions
The real power of high-level programming languages is the use of functions. A
function is a section of code that is wrapped up neatly. It accepts one or more
inputs, and produces one or more outputs. Just like data objects, every function
also has a name (subject to similar restrictions, in that they cannot start with a
number, or contain any reserved characters or spaces). Indeed, any functions
that you create will also appear in the Environment, and can be considered
a type of object. There are many hundreds of functions built into R, which
24 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

do different useful things. A very simple example is the mean function, which
calculates the mean (average) of its inputs. The following code calculates the
mean of the numbers from 1 to 10:
mean(1:10)

## [1] 5.5
Notice that the inputs to the function are inside of the brackets after the
function name. Functions can also take data objects as their inputs, and store
their outputs in other data objects. Earlier on, we stored the numbers from 11
to 20 in a data object called numvect. So we can pass these values into the mean
function, and store the output in a new object as follows:
averageval <- mean(numvect)
averageval

## [1] 15.5
You can find out more about a function using the help function, and passing it
the name of the function you are interested in:
help(mean)

In RStudio the documentation will appear in the Help panel in the lower right
corner of the main window.
It is possible to define new functions yourself to do things that you find useful.
To show you the syntax, here is a very simple example of a function that takes
three numbers as inputs. It adds the first two together, and divides by the third:
addanddivide <- function(num1,num2,num3){
output <- (num1 + num2)/num3
return(output)
}

The name of the function (addanddivide) is defined as being a function with

three inputs (num1, num2 and num3 ). These inputs become data objects inside
the function, and can be used in the same way as any other data object, though
they do not appear in the Environment, and are available only from within the
function.
All of the operations involved in the function then appear inside the curly
brackets. In this example, we have a line of code that adds two numbers together
and divides by the third number, and stores this in a new data object called
output. The return argument tells R to pass the output object back out of
the function so it is available in the Environment. We create the function by
executing (running) the code that defines it. Once it has been created, it appears
in the Environment, and we can call our new function just like we’d call any
other built in function:
2.8. PACKAGES 25

addanddivide(7,5,3)

## [1] 4
With these inputs (7, 5 and 3) we get the output four, because (7+5)/3 = 4.
But now that we have created the function, we can call it as many times as we
like with any inputs. You can also call functions from within other functions,
meaning that operations can be nested within each other, producing sequences
of arbitrary complexity.

2.8 Packages
Sets of useful functions with a common theme are collected into packages. There
are many of these built into the basic R distribution, which you can see by
clicking on the Packages tab in the lower right panel of the main RStudio window
(an example is shown in Figure 2.3a). You can also install new packages to
perform specific functions. Those meeting some basic quality standards are
available through CRAN (at time of writing over 10,000 packages), but it is also
possible to download other packages and install them from a package archive
file. In RStudio, the package manager has a graphical interface for installing and
updating packages (click the Install button in the Packages tab). The window
that pops up (see Figure 2.3b) allows you to install any packages by typing the
package name into the dialogue box. However you can also install packages
from the console or within a script using the install.packages command. For
example, the following code will download and install the zip package used for
compressing files:
install.packages('zip')

Once installed, packages need to be activated before they are visible to R and
their functions become available for use. This can be done manually by clicking
the checkbox in the packages list (see examples in Figure 2.3a - the Matrix and
methods packages are active), or using the library function in the console or in a
script. For example, we can activate the zip package that we just downloaded
like this:
library(zip)

One consequence of the large number of available packages is that sometimes

there are name conflicts, where two packages contain different functions with the
same name. To specify which package a function comes from, we can include the
package name before the function call, separated by two colons. For example,
the following code specfies the median function from the stats package (though
it is unlikely that this particular function would have a name conflict):
stats::median(1:10)

## [1] 5.5
26 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

Figure 2.3: Example screenshot of the Packages panel (a) and the Install Packages
dialogue box (b).

We will use several packages throughout the book, and some of the more
important ones are summarised in section 20.3.

2.9 Conditional statements

Sometimes in a computer program, we want to run some lines of code only if
particular conditions are met, and skip them otherwise. We achieve this using
something called a conditional statement. The most common is the if statement,
which has the form ‘if X then run this code’. ‘X’ in this context will be a logical
statement, like (a > 5, meaning ‘a is greater than 5’), or (a == 1, meaning ‘a is
equal to 1’). If the statement is evaluated as TRUE, the subsequent code will
be run. If the statement is FALSE, the code will be ignored. Here is a short
example:
a <- 1

if (a==1){
print('a is equal to 1')
}

## [1] "a is equal to 1"

If a has a different value, the code inside the if statement (inside the curly
brackets) will not execute (note there is no output being printed):
a <- 0
2.9. CONDITIONAL STATEMENTS 27

if (a==1){
print('a is equal to 1')
}

We can augment an if statement by telling R what to do if the condition is not

met (the else argument):
a <- 0

if (a==1){
print('a is equal to 1')
} else {
print('a is not equal to 1')
}

## [1] "a is not equal to 1"

This form is known as an if. . . else statement, and can even be extended with
many other conditional statements as follows:
a <- -1

if (a>0){
print('a is a positive number')
} else if (a<0) {
print('a is a negative number')
} else {
print('a is zero')
}

## [1] "a is a negative number"

This sequence first checks if the value of a is positive (a>0). If it is it produces

a message to tell us. Next it checks if it is negative (a<0), and tells us if it is.
If neither of these conditions are met, it concludes that a must equal zero, and
tells us that. Of course, in a real computer program, we would do something
more meaningful inside of our if statement. I often use them to select whether
to export figures as files rather than draw them to the plot window, or whether
to load in raw data or data that have already been processed.

The logical statements can involve calls to various functions. Two useful ones
are the is.na and is.infinite functions. These check if data are classed as not
a number (for example if values are missing, or are irrational numbers such as
the square root of -1), or infinite values. The functions return TRUE or FALSE
values, which are interpreted appropriately by the if statement. These functions
are useful for preventing operations that will cause a script to crash, for example
if a missing or infinite number is used in a calculation.
28 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

2.10 Loops
Another powerful feature of programming languages is the loop. A loop instructs
the computer to repeat the same section of code many times, which it will
typically do extremely fast. The simplest type of loop is the for loop. This
repeats the operations inside the loop a fixed number of times. For example, the
following code will print out the word ‘Hello’ ten times:
for (n in 1:10){
print('Hello')
}

## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"
## [1] "Hello"

The terms in the brackets (n in 1:10) define the behaviour of the loop. They
initialise a counter, which can be called anything, but here is called n. The value
of n increases by one each time around the loop, between the values of 1 and
10. We could change these numbers to span any range we like. The instructions
within the curly brackets {print(‘Hello’)} tell R what to do each time around
the loop.

We can also incorporate the value of the counter into our loop instructions,
because it is just a data object containing a single value. In the following
example, we print out the square of the counter:
for (n in 1:10){
print(n^2)
}

## [1] 1
## [1] 4
## [1] 9
## [1] 16
## [1] 25
## [1] 36
## [1] 49
## [1] 64
## [1] 81
## [1] 100
2.10. LOOPS 29

A really useful trick is to use the counter to index another data object. In the
following example, we store the result of an equation in the nth position of the
data object called output:
output <- NULL # create an empty data object to store the results
for (n in 1:10){
output[n] <- n*10 + 5
}
output

## [1] 15 25 35 45 55 65 75 85 95 105

Of course we can have as many lines of code inside of a loop as we like, and
these can call functions, load in data files and so on. It is also possible to embed
loops inside of other loops to produce more complex code structures.

There are two other types of loop available in R, called while and repeat loops.
These do not always repeat a fixed number of times. Their termination criteria
are defined so that the loop exits when certain conditions are met. For example,
a while loop will continue repeating as long as a particular conditional statement
is satisfied. Below is a while loop that adds a random number to a counter object
on every iteration. The loop continues while the counter value is less than five,
and exits when the value of the counter is greater than (or equal to) five. If you
run this code several times, the loop will repeat a different number of times on
each execution.
counter <- 0
while (counter<5){
counter <- counter + runif(1)
print(counter)
}

## [1] 0.7136377
## [1] 1.014727
## [1] 1.62777
## [1] 2.279565
## [1] 3.093874
## [1] 3.776449
## [1] 4.106962
## [1] 4.554587
## [1] 5.174488

A repeat until loop is very similar, except that the conditions for terminating the
loop are evaluated (checked) at the end of the loop rather than at the start (as
with a while loop). These are used less frequently but are appropriate in some
circumstances.
30 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

2.11 Importing data

Most of the time, you will want to use R to process and analyse your own
data. There are many functions for importing data into R, including packages
to deal with specific file formats. For example the readxl package allows you
to load Microsoft Excel files, and the R.matlab package reads the Matlab .mat
file format. A really useful built in function is read.csv, which reads in comma-
separated-values text files. These are a plaintext spreadsheet format, that can
be exported from various software packages, where the distinct values (cells in
the spreadsheet) are separated by commas. By default, the data you read in are
stored in a data frame, with the option of treating the first row as the column
names. The following line of code will do this:
data <- read.csv('filename.csv',header=TRUE)

Whereas if you want to treat the first row as data values (and specify your
column names separately), you would enter:
data <- read.csv('filename.csv',header=FALSE)

The spreadsheet contents will be stored as a data frame in the Environment (the
memory) of R. As noted above, there are so many R packages now that you will
be most likely able to find a function to read in virtually any file format you
need, even including specialist data formats like MRI images.
A helpful feature of RStudio is the Import Dataset option from the File menu.
This provides a graphical interface for importing data that is stored in widely-
used formats (including Excel and SPSS). What is especially clever is that the R
code for loading in the data is automatically generated and sent to the console.
This means you can copy the code into a script so that loading data is automated
in the future. In Chapter 3, we will discuss how to inspect and clean up data
that you have imported.

2.12 How to find out how to do something

As mentioned above, the help function is a great way of finding out more about
a specific R function. However this is not much use if you don’t already know
the name of the function, or the package it is part of. Back in the early years of
programming, people used printed coding manuals for a given language. These
were pretty hard to navigate, and could only give you information about the
core functions of a language. But that was before the internet. All of the
documentation for every R package is available online, and there are many
support forums (websites like Stack Overflow) where you can read the answers
to coding questions posted by others. All of this information means it is now
possible to just type a question into a search engine and get a useful answer
almost immediately.
People who are new to programming often feel like this is cheating somehow,
2.13. TABLE OF USEFUL CORE R FUNCTIONS 31

Table 2.1: Table of useful core R functions with brief descriptions.

Function Description
mean Calculates the average of a vector of numbers
sd Calculates the standard deviation of a vector of numbers
sqrt Calculates the square root of its inputs
nrow Returns the number of rows in a matrix
ncol Returns the number of columns in a matrix
dim Returns the dimensions of a data object
rowMeans Calculates the mean of each row in a matrix
colMeans Calculates the mean of each column in a matrix
seq Generates a sequence of numbers with specified spacing
rep Generates a repeating sequence of numbers
abs Calculates the absolute value (removes the sign)
sign Returns the sign of the input (-1, 0 or 1)
pmatch Partial matching of strings
which Returns the indices of items satisfying a logical condition
is.nan Returns TRUE for any values that are not a number
is.infinite Returns TRUE for any values that are infinite
apply Applies a function over specified dimensions of a matrix
%% Returns the remainder (modulus) for integer division
%/% Returns the quotient (the bit that’s not the remainder)
unique Removes duplicate values from a vector
paste Combines two or more strings into a single string

and that they should magically know the answer to their question already. This
is not true. Everyone uses search engines to find out how to do something, or to
remind themselves the name of a function, or the specific syntax they need. I
do this all the time - almost everything in this book I have worked out how to
do by reading about it online. Expert professional programmers do it all the
time as well. Often if you ask someone who is a more experienced programmer
for help, they will actually just search for the answer on the internet. This is
such a trueism that there is a whole sub-genre of internet memes about how
programmers all have to Google things all the time. There is no shame in this -
it’s the best way to learn.

2.13 Table of useful core R functions

Table 2.1 lists some core R functions that I find particularly useful. This table
omits plotting functions, which are discussed in Chapter 18. In section 20.3
there is also a list of the main packages we use throughout the book, with some
key functions highlighted.
32 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT

2.14 Practice questions

1. What is the name of the language that R is based on?

A) SPSS
B) S
C) Matlab
D) BASIC
2. In RStudio, you can see the data objects currently held in memory using
the:
A) Environment tab
B) History tab
C) Viewer tab
D) Console
3. Which of the following is not a benefit of using scripts?
A) You can share your code with others
B) Your code will run in other programming languages such as Python
C) You can reproduce your own analysis in the future
D) You can save your work and return to it later
4. Which of the following is not a legal object name in R?
A) Jonathan
B) a
C) var125
D) 9thNumber
5. In R, a data object with rows and columns can be called a:
A) Vector or scalar
B) Array or list
C) Matrix or data frame
D) Spreadsheet variable
6. To refer to a particular value within a data object in R, we index it using:
A) Normal brackets ()
B) Square brackets []
C) Curly brackets {}
D) Pointy brackets <>
7. Collections of functions on a common theme are called:
A) Repositories
B) Functions
C) Toolboxes
D) Packages
8. How many times will the following loop repeat? for (n in 31:35){}
A) 1
B) 5
C) 31
D) 35
a <- 0
for (n in 1:3){
2.14. PRACTICE QUESTIONS 33

a <- a + 4
}
if (a>10){print('Bananas')}
if (a<10){print('Apples')}
if (a==10){print('Pears')}
if (a==12){print('Oranges')}

9. What will be the output of the above lines of code?

A) Bananas
B) Apples
C) Oranges
D) Bananas and Oranges
10. What is the name of the R function used for transposing a matrix? (Hint:
the answer is not given anywhere in this chapter, but the section on how
to find out how to do something might help!)
A) tp
B) trans
C) tr
D) t
Answers to all questions are provided in section 20.2.
34 CHAPTER 2. INTRODUCTION TO THE R ENVIRONMENT
Chapter 3

Cleaning and preparing

data for analysis

3.1 Why do data need to be ‘cleaned’?

Real data are messy. In any study a number of factors, such as equipment
malfunction, participant variability, experimenter error, environmental noise,
data entry mistakes or file corruption, can have an effect on your results. Hope-
fully such problems are not too severe, and will affect relatively few of your
measurements. But there is no way to know this without somehow checking your
data. In this chapter we recommend initial visual inspection, and illustrate two
common methods for checking data visually (histograms and scatterplots). We
will then discuss how to identify and deal with outliers more formally, and discuss
several useful methods of checking distributions and rescaling data. However, we
begin with the practical issue of how to format data appropriately for use in R.

3.2 Organising data in wide and long formats

There are two main conventions for storing data in R. They are known as wide
and long format. The idea of wide format is that each cell in a matrix stores
a single observation (i.e. a measurement of the dependent variable), and the
location in the matrix gives us information about where that observation has
come from. So, consider an experiment where 8 participants each complete
5 repetitions of some task. Their data could be stored in an 8 (rows) by 5
(columns) matrix, much as we might arrange them in a spreadsheet. Each row
will represent a different participant, and each column will represent a different
repetition. The data might look something like this:
## [,1] [,2] [,3] [,4] [,5]

35
36 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

## [1,] 92 134 96 124 175

## [2,] 179 167 81 185 72
## [3,] 15 95 94 159 132
## [4,] 114 108 102 206 33
## [5,] 71 105 152 144 65
## [6,] 41 112 136 3 116
## [7,] 94 112 84 67 153
## [8,] 118 22 49 129 109
An alternative way to lay out the data is long format, in which each observation
has its own row. In each row, one column will store the value of the observation,
and any other columns will store associated condition information about that
observation. When stored in a data frame, the column headings should indicate
the information stored in them (much like variable names in SPSS or Excel).
The data set from the above example laid out in long format would look (for the
first 10 entries) like this:
## Participant Repetition Measurement
## 1 1 1 92
## 2 2 1 179
## 3 3 1 15
## 4 4 1 114
## 5 5 1 71
## 6 6 1 41
## 7 7 1 94
## 8 8 1 118
## 9 1 2 134
## 10 2 2 167
Notice that the measurement values are the same as those in the wide table.
The numbers in the participant and repetition columns correspond to the row
and column numbers from the wide version of the data set.
Readers familiar with running ANOVA in the widely-used SPSS package may
recall that in that software, repeated measures data are stored in wide format,
whereas between subjects data are stored in long format. In R, the decision
about how to lay out your data will vary depending on your experimental design,
and also how you are planning to analyse the data. Many R functions assume
long format, as we will see in Chapter 4, and so for running tests like regression
and ANOVA this may be preferred. However, some of the more sophisticated
methods in later chapters, which require the user to store data in an efficient
way, might tend to work better in wide format. A good example might be an
image - representing each pixel intensity in a two-dimensional matrix is much
more natural than representing the whole image as a long list of intensities.
To facilitate switching between long and wide formats, there are several helper
functions. The reshape2 package contains functions called melt (for converting
from wide to long) and acast (for going the other way). Alternatively, the
3.3. INSPECTING DATA: HISTOGRAMS AND SCATTERPLOTS 37

newer tidyverse framework offers functions called gather and spread which fulfil
analogous roles as part of the tidyr package. Since requirements will differ
depending on idiosyncracies of a data set, the help documentation for these
functions will likely prove useful if your data needs restructuring. A final useful
function for restructuring data is the transpose function, t. This function turns
rows into columns, and columns into rows, much more straightforwardly than
can be done using spreadsheet software, where copying and pasting with special
options is required.

3.3 Inspecting data: histograms and scatter-

plots
The best way to understand what is happening in a data set is usually to look
at it. There are many methods of visualisation that are useful, but two of the
most commonly used are histograms (for each individual dependent variable),
and scatterplots (for pairs of variables). A histogram counts the number of
observations within bins of a particular size. Figure 3.1a,b shows two histograms
with different bin sizes, generated using the hist function with the following code:
hist(data,breaks=seq(-3,3,length=11),col='grey')
hist(data,breaks=seq(-3,3,length=21),col='cornflowerblue')

The entry to the breaks argument defines the start and end points of each bin.
The hist function will add up the total number of values in the data set between
the lower and upper boundaries of each bin. To define these boundaries, the
seq function generates a sequence of numbers between two points, here with a
specified length. For example:
seq(-3,3,length=11)

## [1] -3.0 -2.4 -1.8 -1.2 -0.6 0.0 0.6 1.2 1.8 2.4 3.0
So this line of code generates a sequence of numbers between -3 and 3, with 11
evenly spaced values. The hist function will then use this sequence to bin the
data. Note that there will always be one less bin than the length of the sequence,
and that the mid-points of the bins are the means of successive pairs of break
points.
A popular alternative to showing histograms with discrete bins is to plot a
smoothed kernel density function. The smoothing can sometimes obscure discon-
tinuities in the data, but smoothed functions are often more visually appealing.
An example is shown in Figure 3.1c, created using the density function as follows:
a <- density(data)
plot(a$x,a$y,type='l',lwd=2)

The histograms in Figure 3.1a-c show data that are approximately normally
distributed, and have no obvious outliers. But real data are often much less clean.
38 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

70
120 (b)
(a) 60
100
50
80
Frequency

Frequency
40
60
30
40 20
20 10

0 0

−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

data data

0.4 60
(d)
(c) 50
0.3
40
Frequency
a$y

0.2 30

20
0.1
10

0.0 0

−3 −2 −1 0 1 2 3 −2 0 2 4 6

a$x data

Figure 3.1: Example histograms. Panel (a) has ten bins, and panel (b) has
20 bins. Panel (c) shows a kernel density function, and panel (d) includes two
extreme outliers.
3.3. INSPECTING DATA: HISTOGRAMS AND SCATTERPLOTS 39

For example, in Figure 3.1d there are two outlier points, at x-values of 4.5 and
6.1, as indicated by the arrows. These data points are far outside of the range
of values in the rest of the sample, and are likely to have been caused by some
kind of error. In Section 3.4 we will discuss some methods for identifying such
outliers more objectively, and also consider how to replace them in an unbiased
way.
Histograms are informative regarding the shape of a distribution, and will usually
make any outliers clear. However it is often also helpful to plot each individual
data point using a scatterplot. Usually the term scatterplot makes us think
of bivariate plots such as that shown in Figure 3.2b, which show two variables
plotted against each other. But if we have only a single dependent variable, it is
still important to inspect the individual data points. We can use a univariate
scatterplot to do this, in which the x-position is arbitrary, as shown in Figure
3.2a. This plot shows every individual observation on some measure (shown on
the y-axis), but within a given condition the x-position of a data point is not
informative.

10 3
(a) (b)
9
2
8
7
Measure B

1
Measure

6
●
5 ● 0
●
● ●
4 ●●● ● ●
●
●●
●
●

● ●● −1
3 ●● ●●●●● ●●
● ●●
●
●
●
●● ●●
●
● ●
2 ● ● ●● ●
● ●● −2 ●
●
1
0 −3
A B −3 −2 −1 0 1 2 3
Condition Measure A

Figure 3.2: Examples of univariate (a) and bivariate (b) scatterplots. Kernel
density functions are shown for each measure along the margins of panel (b)
(grey curves).

The scatterplot in Figure 3.2b contains three outliers, which are highlighted in
blue. These values are not particularly remarkable in either their x- or y-values,
so the grey kernel density histograms along the margins do not reveal them.
However they clearly differ from the rest of the population (grey points), and
might therefore be considered to be outliers. Both of the graphs in Figure
3.2 were produced using the generic plot functions (see Chapter 18 for more
detailed discussion of creating plots in R), though alternative methods are also
40 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

available, for example using the geom_point function in the ggplot2 library.
Remember that the code to produce all figures is also available in the book’s
GitHub repository at: https://github.com/bakerdh/ARMbookOUP.

3.4 Identifying outliers objectively

Plotting data will often reveal obvious outlying values. But in many situations,
we would ideally like a more quantitative method for identifying outliers that
are more ambiguous. There have been many approaches proposed for this, some
of which we will describe in the following sections. The goal of outlier detection
is to identify values that were generated by a different process from the rest of
the data set. This is most often due to some form of equipment malfunction or
other unforseen event. For example, imagine that you set up a camera to count
the number of times a bird leaves its nest each day. Most days the bird leaves
between 20 and 30 times. But on one day a leaf fell on the camera early in the
morning and blocked the lens - on this day only the 5 exits before the leaf fell
were recorded. The value 5 is a clear outlier, and is far outside the range for the
other days. But how might we identify such an outlier objectively?

3.4.1 Tukey’s ‘fence’ method

A widely used method for outlier identification was proposed by the American
mathematician John Tukey (1977). This involves calculating two fences using the
interquartile range of a data set. The interquartile range (sometimes abbreviated
to IQR) is the distance between the two points that encompass 50% of the
values in a data set. In other words, it is the distance between the 25% and 75%
quantiles of the data set (see Figure 3.3). The inner fence is placed 1.5 times the
IQR below the first quantile, and 1.5 times the IQR above the third quantile.
Data points falling outside of these values are considered outliers. The outer
fence is defined similarly using 3 times the IQR, and values exceeding these
limits are considered extreme outliers.
A major practical advantage of Tukey’s method is that the fences are determined
based on the interquartile range. This is calculated using only the central 50% of
data points, so it is not influenced by extreme outliers. For example, imagine we
had a sample of 50 data points with values in the range 0 to 150. We calculate
the interquartile range as lying between 63 and 85, with the inner fence extending
from 28 to 119. These estimates would be unaffected if an equipment malfunction
meant that two or three data points took on extreme values of >2000. And of
course these extreme values would certainly be identified as outliers by Tukey’s
method!
Outliers identified in this way are often illustrated visually using box and whisker
plots (see Figure 3.3). The box encompasses the interquartile range, and the
whiskers (error bars) show the limits of the inner fence. Individual outliers
are traditionally plotted using different symbols from the rest of the data set,
3.4. IDENTIFYING OUTLIERS OBJECTIVELY 41

Q1 IQR Q3

● ●
●
●●
● ● ● ●
●
Outer fence

Outer fence
Inner fence

Inner fence
50%

− 5σ − 4σ − 3σ − 2σ − 1σ 0σ 1σ 2σ 3σ 4σ 5σ

Figure 3.3: Illustration of the inner and outer fences, and the interquartile range.
The x-axis is scaled in standard deviation (sigma) units, relative to the mean.
The cloud of grey data points show representative data sampled from the black
population distribution, but also features several outliers (blue points). The
vertical dotted lines, horizontal bar, and blue shaded region of the distribution
illustrate the interquartile range, between which 50% of data points lie. The
dashed lines and error whiskers indicate the inner fence, which extends 1.5 times
the IQR above and below Q1 and Q3 respectively. The outer fence extends 3
times the IQR above and below Q1 and Q3, and is shown by the solid vertical
lines.
42 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

and in some conventions non-outlying points are not plotted at all. Sometimes
extreme outliers (beyond the outer fence) use a different symbol from near
outliers (between the inner and outer fences). Rudimentary boxplots can be
created automatically using the boxplot function from the graphics package in R.
For example, the following line of code produces the boxplot in Figure 3.4:
boxplot(ndata)
4

●
●
2
0
−2

●
●

Figure 3.4: Example rudimentary boxplot, generated using the boxplot function.

3.4.2 Standard deviations and related criteria

An alternative to calculating fences is to exclude data points that lie beyond some
multiple of the sample standard deviation (actually, for normally distributed data,
the inner fence is at around 2.7 times the standard deviation, so this method
is similar to just shifting the fence!). Conventions vary, and should ideally be
decided (and preregistered) in advance before seeing the data. Classifying as
outliers any data points that lie beyond 3 standard deviations of the mean
would, for normally distributed data, count less than 0.3% of data points as
being outliers. So for a sample size below 100, any data points that exceed this
threshold are very likely to be outliers.
In the following R example, we generate 100 random samples from a normal
distribution (how computers generate random numbers is explained in Chapter
8), and add a clear outlier at a known location (entry 57). We then calculate our
3.4. IDENTIFYING OUTLIERS OBJECTIVELY 43

threshold criterion as 3 times the standard deviation of the sample. To work out
which values exceed this distance from the mean, we subtract the mean from
each data point, and for convenience take the absolute values (i.e. any negative
numbers become positive). Any values in this normalised data set that exceed
the criterion value will be picked up as outliers. We can find the indices of these
values using the which function.
data <- rnorm(100) # generate some random data
data[57] <- 100 # replace the value in row 57 with an outlier
criterion <- 3*sd(data) # calculate 3 times the standard deviation
normdata <- abs(data-mean(data)) # subtract the mean and take the absolute value
which(normdata>criterion) # find the indices of any outlier values

## [1] 57
Two more formal variants of this approach include Chauvenet’s criterion (Chau-
venet 1863) and Thompson’s Tau (Thompson 1985). These methods use the
total sample size to determine the threshold for deciding that a given data point
is an outlier, based on properties of either the normal (Gaussian) distribution or
the T-distribution.
The procedure for Chauvenet’s criterion is to convert all values to absolute
z-scores (subtracting the mean and scaling by the standard deviation) and
again identifying data points that exceed a criterion. This time the criterion is
calculated by taking the quantile of the normal distribution at 1/(4N), where
N is the sample size. This means that the criterion becomes more stringent
(i.e. larger) as the sample size increases, because with larger samples we expect
a greater number of extreme values. A rudimentary implementation is given by
the following function:
d_chauv <- function(data){
i <- NULL # initialise a data object to store outlier indices
m <- mean(data) # calculate the mean of the data
s <- sd(data) # calculate the standard deviation
Zdata <- abs(data-m)/s # convert data to absolute z-scores
dmax <- abs(qnorm(1/(4*length(data)))) # determine the criterion
i <- which(Zdata>dmax) # find indices of outliers
return(i)}

The above function uses the qnorm function to estimate the appropriate quantile,
and the which function to return the indices of any outlier values. Using this
function with the example data from above (where we added an outlier at entry
57) again correctly identifies the outlier:
d_chauv(data)

## [1] 57
The modified Thompson’s Tau method is conceptually similar, except that the
44 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

critical value is obtained from the T-distribution using a given α value, and then
converted to the tau statistic with the equation:

tα/2 (N − 1)
τ=√ q (3.1)
N (N − 2 + t2α/2 )

where tα/2 is the critical t-value with N -2 degrees of freedom using the significance
criterion α (typically α = 0.05), and N is the sample size. Another difference is
that the tau method is iterative, with only the value that deviates most from
the mean being compared with the critical value on each iteration. The τ value
is recalculated on each iteration, using only the data remaining after removing
outliers on previous iterations.
Figure 3.5 illustrates the results of simulations that show how sensitive each
test is for detecting a single outlier of known value. The Thompson’s Tau test
identifies the known outlier with a similar sensitivity to methods that reject
values exceeding 1.5 times the sample standard deviation. The Chauvenet
criterion is more conservative, and similar to rejecting values exceeding about
2.8 times the standard deviation. Note that these values are dependent on the
sample size, for example the Chauvenet criterion becomes more conservative as
sample size increases.
One distinct danger posed by the availability of several different outlier detection
algorithms is that they provide an experimenter with hidden degrees of freedom
in their analysis. An unscrupulous researcher could easily engineer a desired
result by choosing an outlier rejection algorithm after their data have been
collected. This is highly unethical and strongly discouraged. To avoid such
issues, analysis plans should ideally be preregistered before the data have been
collected, and must detail how outliers will be handled. In general, a criterion of
around 2.5 or 3 standard deviations is usually appropriate for univariate data,
though it is worth checking standard practice in a given research area.

3.4.3 The Mahalanobis distance for multivariate data

Methods based on the properties of the normal distribution are usually sufficient
for identifying outliers in univariate data sets, where we have only a single
dependent variable. But in multivariate data sets, outliers can sometimes occur
that are within the mid-range of each individual variable, yet occupy a distinct
part of the possible space of values compared with the rest of the data set (see
the example in Figure 3.2b).
In such cases, a useful statistic is the Mahalanobis distance, D, proposed by the
prolific Indian mathematical physicist and statistician Prasanta Mahalanobis
(1936). Intuitively, this is similar to the Euclidean distance (shortest straight
line) between each point and the centre of mass of the whole data set. However,
it is more sophisticated than the Euclidean distance because the Mahalanobis
3.4. IDENTIFYING OUTLIERS OBJECTIVELY 45

1
Tukey
Proportion detected

Thompson
0.8
Chauvenet

0.6

0.4
1 SD
0.2 2 SD
3 SD
0
0 1 2 3 4
Outlier position (SD from true mean)

Figure 3.5: The proportion of outliers detected as a function of outlier position,

expressed in standard deviation units. The grey curves were calculated by
rejecting all data points that exceed 1, 2 or 3 standard deviations from the sample
mean. Tukey’s fence method produces a curve equivalent to approximately 2.7
standard deviations (light blue dash-dotted curve). The blue dashed and dotted
curves were calculated using Thompson’s Tau and the Chauvenet criterion. All
simulations involved 99 values drawn from a normal distribution, and an added
‘outlier’ with values given along the x-axis. Note that changes in sample size
affect the relative positions of the curves.
46 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

distance for a given data point is scaled by the variance of the data set in the
direction of that data point. This allows us to take into account any correlations
between the two (or more) dependent variables.

To understand how this works, consider first the data set shown in Figure 3.6a.
Here, the two variables are uncorrelated, and the variable plotted on the x-axis
has a smaller variance than the one plotted on the y-axis. The two white outlier
points are both the same Euclidean distance from the centroid of the data (black
point), because the blue vector lines are the same length. But if we express their
distance in standard deviation units along the appropriate axis, the square outlier
is clearly a more extreme outlier than the triangle outlier (because the standard
deviation is larger in the y direction). This is what the Mahalanobis distance
calculates, and for this example the square point would have a Mahalanobis
distance of around D = 6, and the triangle point a distance of around D = 3.
So, the square is a more extreme outlier than the triangle when expressed with
this metric, which fits with our intuitions.

8 8
(a) (b)

4 4

● ●
0 0
y

−4 −4

−8 −8
−8 −4 0 4 8 −8 −4 0 4 8
x x

Figure 3.6: Illustration of outliers in two dimensional (bivariate) data sets. The
black point is the centre of mass, and the white triangle and square symbols are
outliers. The blue lines depict the Euclidean distances between the centroid and
each outlier, and are of identical length.

Next let’s consider the data set shown in Figure 3.6b. This time, there is a
clear correlation between the two variables. Again, the two outliers are the same
Euclidean distance from the centroid. As before, the Mahalanobis distance is
greater for the square point, because the variance is estimated in the direction of
the blue vector line. This is somewhat harder computationally than calculating
the standard deviation for one or other dependent measure. Happily, we do not
need to perform these calculations by hand. The mahalanobis function is part of
3.5. IDENTIFYING MISSING AND OUT OF RANGE VALUES 47

the core stats package in R. Imagine we have a data frame containing around
500 rows of bivariate data (from Figure 3.2b) structured as follows:
head(bdata)

## x y
## 1 0.685092067 0.6039170
## 2 -0.005550195 0.2395705
## 3 -0.777641329 -1.0976698
## 4 1.875702830 1.5417293
## 5 -0.377129105 0.3195294
## 6 -0.454686991 0.3052273
We can pass this data frame into the mahalanobis function, along with the
means of each variable (calculated inline using the colmeans function), and
the covariance matrix (calculated inline using the cov function). The function
returns a squared distance estimate for each x-y pair of points:
D2 <- mahalanobis(bdata, colMeans(bdata), cov(bdata))

D2[1:6]

## [1] 0.4343403 0.2757397 0.9850249 3.5338519 1.8662220 2.2249707

Plotting these distance estimates (in Figure 3.7) reveals the three clear outliers
that we highlighted in blue in Figure 3.2b. Because the Mahalanobis distance
is calculated in units of the standard deviation, we can apply similar threshold
criteria as described in Section 3.4.2. However, the output of the mahalanobis
function is a squared distance, so we must either first take the square root
(with the sqrt function), or square our standard deviation threshold criterion. A
threshold of 3 times the standard deviation is a squared distance value of 32 = 9,
shown by the vertical dotted line in Figure 3.7. Our three outliers (indicated by
the arrows) clearly exceed this criterion. Note also that the squaring avoids any
negative distance values.
One final benefit of the Mahalanobis distance is that it generalises easily to
any number of dimensions. If you have a large multivariate data set with more
than two dependent variables, it is still a useful method for identifying outliers.
Importantly, regardless of the number of dimensions, a single distance statistic
will be computed for each case (e.g. participant). If you plan on using any of
the multivariate methods discussed in Chapters 11 - 14, a working knowledge of
the Mahalanobis distance will prove useful.

3.5 Identifying missing and out of range values

In addition to the above techniques, it is also worth being aware of some useful
R functions for identifying completely missing values, or those that are outside
of some specified range. The is.na function returns TRUE for any values stored
48 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

200
150
Frequency

100

Outliers
50
0

0 10 20 30 40

Mahalanobis D2

Figure 3.7: Histogram of squared Mahalanobis distance measures. The dotted

vertical line indicates a threshold equivalent to 3 standard deviations. The
positions of three clear outliers are indicated by arrows.
3.5. IDENTIFYING MISSING AND OUT OF RANGE VALUES 49

in a data object that are classed as ‘not a number’. This is especially useful as
missing values loaded in from spreadsheets are assigned NA by default. We can
combine the is.na function with the which function to return the indices of the
missing values.
To give an example, suppose we have a data set comprising 10 values, 3 of which
are missing:
## [1] 4 6 NA 1 7 8 NA 3 NA 5
The is.na function will return TRUE for the missing values, and FALSE for the
others:
is.na(nandata)

## [1] FALSE FALSE TRUE FALSE FALSE FALSE TRUE FALSE TRUE FALSE
Wrapping the which function around this will return the indices of the NA values,
at positions 3, 7 and 9:
which(is.na(nandata))

## [1] 3 7 9
One way to remove the NA values is to invert the output of the is.na function
using the ! operator (making TRUE become FALSE, and vice versa), to return
only the rows that contain real numbers:
trimmeddata <- nandata[which(!is.na(nandata))]
trimmeddata

## [1] 4 6 1 7 8 3 5
A related function is the is.infinite function, which returns TRUE for positive
and negative infinity - these often appear if values are inadvertently divided by
zero. However the which function can take any logical argument, so it can also
be used to restrict values to within a certain range, for example:
trimmeddata[which(trimmeddata<5)] # return only values < 5

## [1] 4 1 3
It might be useful to do this if one’s data have a natural range. For example,
many real-world measurements such as height, weight, heart rate and so on, must
take on positive values. If some sort of error has resulted in negative values for
some observations of these types of variables, it would be reasonable to consider
these observations as out of range.
Finally, we can combine these logical statements with other functions, such as
the mean, sum or sd functions. To calculate the mean of the nandata object,
excluding the NA values, we can nest the which statement inside the indexing of
the variable as follows:
50 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

mean(nandata[which(!is.na(nandata))])

## [1] 4.857143
Although this is a good general solution, it can be rather cumbersome. In
some functions, there is an alternative implementation specifically for NA values
known as the na.rm flag. By setting this to TRUE, we instruct the function to
remove the NA values automatically. The mean, sum and sd functions (along
with other core functions) have this facility, for example:
mean(nandata, na.rm=TRUE)

## [1] 4.857143
Notice that this returns the same result as the more involved solution. Not all
functions have an na.rm flag, but you can check the help files to work out which
ones do.
The above tools are extremely useful for cleaning up data and dealing with
missing or problematic values. They can be used in combination with several of
the other outlier detection criteria that were described in Sections 3.4 - 3.4.3.
However this leads us to a more philosophical question of what we should be
doing with outliers and missing values in the first place.

3.6 What should we do with outliers?

Once outliers and missing values have been identified, we must decide what to
do with them. There are three main possibilities - we can delete the outliers, we
can replace them with estimated values, or we can leave them in the data set.
Opinions differ regarding the best course of action, and this will typically depend
on the nature of your data set and what you are trying to do with it. In general,
it is not thought to be a good idea to exclude outliers simply because they look
different from the other values. One should instead have some independent
rationale for removing them, for example if there is evidence that the outliers
were caused by a substantively different data generating process from the rest of
the data.
Deleting a small number of outliers from a large data set is unlikely to make
much difference to our ability to detect an effect. However, when sample sizes
are smaller, removing several outliers will substantially reduce our statistical
power (see Chapter 5). Outlier removal is a particular issue with multivariate
data sets, because it can often be the case that many individual participants
will be missing observations from at least one of the dependent variables. As we
discuss further in Chapter 12, this can drastically reduce effective sample sizes for
methods such as structural equation modelling. On the other hand, approaches
such as mixed effects models (see Chapter 7) are able to cope well with missing
values, and might be a better alternative to more traditional methods such as
3.7. NORMALIZING AND RESCALING DATA 51

analysis of variance (ANOVA), for which a single missing value usually requires
removal complete removal of a participant.
Alternatively, techniques exist to estimate (impute) plausible values to replace
those that are missing, which can make some statistical models more stable.
These include mean imputation (replacing a missing value with the sample
mean), Windsorization (replacing an outlying value with a neighbouring value),
and clipping (setting extreme values to a prespecified maximum or minimum
threshold). If you have a solid rationale for replacing missing or outlying values,
and have ideally specified this in advance through preregistration, then these
methods avoid many of the shortcomings of outlier deletion.
Finally we can decide to leave outliers in a data set. This might affect our choice
of statistical test, as data sets with substantial outliers are unlikely to meet the
assumptions of parametric statistics. We could instead use methods that make
fewer assumptions, such as nonparametric alternatives, or the ‘bootstrapping’
approach we will introduce in Chapter 8. There is also a class of statistics
called robust statistics that are designed to be used with noisy data. A simple
example of a robust statistic is the trimmed mean, in which some percentage of
extreme values is removed from the data set, and the mean calculated using the
remainder. For example, a 10% trimmed mean would involve rejecting the lowest
and highest 10% of values from a data set, and using the remaining 80% of
values to estimate the mean. Other variants forego the assumption that data are
normally distributed, and instead use other distributions such as t-distributions,
which have longer tails (this approach is also common in Bayesian statistics, see
Kruschke 2014). A full discussion of robust statistics is beyond the scope of the
current text, but the interested reader is referred to the book Robust Statistics
by Peter Huber (2004).

3.7 Normalizing and rescaling data

For some analysis methods, it is appropriate to normalize data, either by sub-
tracting the mean, and/or by scaling by the variance (this is also sometimes
called standardising). This is especially important for multivariate techniques
such as k-means clustering (see Chapter 13) and multivariate pattern analysis
(see Chapter 14), where different dependent variables might have very different
ranges. By rescaling, we level the playing field, so that each variable is weighted
equally in the analysis. Of course for other types of analysis this would be very
inappropriate, and might even remove any legitimate differences between groups!
When appropriate, we can rescale data using using the scale function:
rescaleddata <- scale(data)

By default this function performs the following operations on each column of a

matrix:
1. subtracts the column mean from each data point
52 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

2. divides each data point by the standard deviation of the column

It is possible to perform only one or the other of these operations by setting
either the center or scale arguments to FALSE as follows:
rescaleddata <- scale(data,scale=FALSE) # only subtract the mean
rescaleddata <- scale(data,center=FALSE) # only divide by the SD

It is important to remember that the units of rescaled data will not be the same
as the units of the original measurements. One way to think of normalized data
is as being akin to z-scores, where each data point is expressed in standard
deviation units. The univariate scatterplots in Figure 3.8 illustrate the effect of
rescaling a data set.

20

●
●
● ● ●● ●
● ●
● ● ● ●●● ●
● ●● ●● ●
10 ●●● ● ●
●
●
●●● ●●●● ●
● ●
●● ●●●
●
● ●
● ●● ●
●
● ● ● ●
● ● ● ●●● ● ●● ●● ●● ● ●
●● ● ●●
●●●●●
● ●
●●
●●
●●
● ●
●
● ●
●
●●●
●● ●●
●●
●●
●●
0 ● ●● ● ● ●●
●
●● ● ●● ●●● ●●●●●
●●●●●
●
●
● ● ●● ●
●●●
● ●● ●●● ● ●● ●●● ●
●
● ● ●
●
● ●●●
●

Raw Centred Scaled Centred &

−10 Scaled

Figure 3.8: Illustration of rescaling using an example data set (black points).
The dark blue points show centering, with no change in variance. The grey
points illustrate scaling by the standard deviation - this also affects the mean,
but does not centre the mean on 0. Finally, the light blue points show the effects
of centering and then scaling.

3.8 Transforming data and testing assumptions

A key assumption underlying parametric statistical tests is that data should be
normally distributed - in other words a histogram of the data should follow the
bell-shaped curve of a Gaussian distribution. But what if, even after dealing
3.8. TRANSFORMING DATA AND TESTING ASSUMPTIONS 53

with outliers, this is still not the case? If this happens, it is sometimes possible
to transform the data by applying mathematical operations to the full data set.
The most common of these are logarithmic transforms, which pull in the long
tail of a positively skewed distribution, and squaring, which has a similar effect
on negatively skewed distributions (exponential transforms can also be used for
dealing with negative skew). Examples of how each of these transforms can
make skewed data conform more closely to a normal distribution are shown in
Figure 3.9. A deeper discussion on the advantages of data transforms is given
by Bland (2000).

As many introductory statistics texts will explain in detail, there are two main
tests to assess whether data are consistent with a normal distribution. These
are the Kolmogorov-Smirnov test (Smirnov 1948; Kolmogorov 1992) and the
Shapiro-Wilk test (Shapiro and Wilk 1965). The Kolmogorov-Smirnov test
involves comparing the cumulative distribution functions of the data and a
reference distribution (i.e. a normal distribution), and is implemented in R’s
built in stats package by the ks.test function. For example:
ks.test(data1,'pnorm',mean(data1),sd(data1))

##
## One-sample Kolmogorov-Smirnov test
##
## data: data1
## D = 0.17671, p-value = 0.003879
## alternative hypothesis: two-sided

Note that the ks.test function requires as inputs the sample of data and the name
of the cumulative distribution you wish to compare it to (pnorm is a cumulative
normal). By default the data will be compared to a distribution with a mean of
0 and a standard deviation of 1. So it is necessary to either provide the actual
mean and standard deviation of your data as additional arguments (as above),
or to first normalize your data using the scale function (see Section 3.7):
ks.test(scale(data1),'pnorm')

##
## One-sample Kolmogorov-Smirnov test
##
## data: scale(data1)
## D = 0.17671, p-value = 0.003879
## alternative hypothesis: two-sided

The data we have tested here were generated using a lognormal distribution
(the grey distribution in the top panel of Figure 3.9), so the test is signifi-
cant, indicating a deviation from normality. As with most assumption tests, a
non-significant Kolmogorov-Smirnov would mean that there was no significant
difference between the data and reference distribution, so we could assume the
54 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

Log transform Original data

Squaring Original data

Figure 3.9: Examples of data transforms. The upper panel shows some positively
skewed data (grey), and a more normal distribution following a log transform
(blue). The lower panel shows some negatively skewed data (grey), and a more
normal distribution following a squaring transform (blue).
3.8. TRANSFORMING DATA AND TESTING ASSUMPTIONS 55

data are approximately normal.

In common with other statistical tests, increasing sample size will increase
statistical power (see Chapter 5), so small data sets are less likely to be sig-
nificant (and thus more likely to pass the test) than larger ones, all else being
equal. Simulation studies (e.g. Yap and Sim 2011) have demonstrated that the
Kolmogorov-Smirnov test is lower in power than the Shapiro-Wilk test, which
works in a similar way but using a regression framework. In R, the shapiro.test
function is conducted as follows:
shapiro.test(data1)

##
## Shapiro-Wilk normality test
##
## data: data1
## W = 0.886, p-value = 3.285e-07
Again, a significant result (p < 0.05) implies a deviation from normality. Notice
that the p-value for the Shapiro-Wilk test is generally smaller than that for the
Kolmogorov-Smirnov test using the same data, illustrating its greater power.
However, most implementations of Shaprio-Wilk are limited to samples of less
than 5000 data points, whereas the Kolmogorov-Smirnov test has no such
restrictions.
Finally, Q-Q plots can be very informative in visually assessing and understanding
deviations from normality. These are constructed by plotting the quantiles of the
data against the quantiles of a reference distribution (e.g. a normal distribution).
The qqnorm function generates a plot, and the qqline function adds a reference
line. Examples are shown in Figure 3.10 for normally distributed data (left) and
positively skewed data (right). Note the substantial deviation from the diagonal
reference line at the upper end of the skewed distribution.
Typically, one would use tests of normality and Q-Q plots on a raw data set
first. If this deviates from normality, it is worth trying an appropriate transform,
and then running the normality test again. Unfortunately, even after applying
a transform, some data sets simply cannot be manipulated sufficiently to meet
the assumptions of parametric statistics. This might rule out the use of some
tests, however there are also many non-parametric alternatives that can be
used instead, just like for data sets with outliers. In general, non-parametric
methods involve rank-ordering a data set, and performing calculations on the
rankings rather than the raw data. This avoids any problems with outliers and
deviations from normality, but at the expense of statistical power (the ability of
a test to detect a true effect). Common examples are the Spearman correlation
coefficient, the Mann-Whitney U test, and Friedman’s ANOVA, all of which are
implemented in R. Non-parametric methods are not the main focus of the current
text, however in Chapter 8 we will discuss bootstrap resampling techniques,
which can be used as a flexible non-parametric approach to statistical testing.
56 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

Normal Q−Q Plot Normal Q−Q Plot

● ●
4
2 ●
●●
●●●●
● ●
● ●●
●●
●●
● 3
Sample Quantiles

Sample Quantiles
●
●●
1 ●
●●
●●
●
●
● ●●
●● ●
●●
●
●
●●
●
●●
●
●
●
●
●● ●
●
● ●
●●
●
●
●
● 2 ●●
●
●●
●
0 ●
●
●
●
●
●●
●●
●●
● ●
●●
●
●●
● ●
●●
●
●
●●
●● ●
● ●●
●
●
●●
● ●●
●
● ●
●●
●
●● ●
●●
●
●●
●
● ●
●●
●
●●
● ●
● 1 ●
●
●
●
●
●
−1 ●
●●●●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●●
●
●
●● ●
●●
●
●●
●
●●
●●
●
●
●●● ●●●
●●
●●
●●●
● ●●●●
● ●●
● ● ●●

−2 −1 0 1 2 −2 −1 0 1 2

Theoretical Quantiles Theoretical Quantiles

Figure 3.10: Example Q-Q plots for normally distributed data (left) and skewed
data (right).

3.9 Recoding categorical data and assigning fac-

tor labels
So far we have mostly talked about data that are numerical and continuous. But
often we also need to deal with categorical data, for example reflecting group
membership, sex, condition information, or responses on a categorical scale (such
as A, B, C, or low, medium, high). In R, such data are usually stored as factors -
data objects with a set of possible text labels, which are also (behind the scenes)
given numerical values. Factor variables are particularly useful when conducting
statistical tests to compare different groups. For example in Analysis of Variance
(ANOVA), the independent variables are categorical, and these should be stored
as factors (see section 4.4 for more details on how to run ANOVA in R).

To demonstrate how factors work, let’s create a data object containing sex
information for seven rats:
ratsex <- c('M','M','F','M','F','F','F')
ratsex

## [1] "M" "M" "F" "M" "F" "F" "F"

By default, this data object contains a list of character strings. We can convert
it to a factor using the factor function:
ratfactor <- factor(ratsex)
ratfactor
3.10. PUTTING IT ALL TOGETHER - IMPORTING AND CLEANING SOME REAL DATA57

## [1] M M F M F F F
## Levels: F M
Now we see that R has defined two levels for the factor, F and M. An integer
code is also assigned to each level, by default in alphabetical order. We can see
these numerical values using the as.numeric function:
as.numeric(ratfactor)

## [1] 2 2 1 2 1 1 1
Notice that all values of M are coded as 2, and all values of F are coded as 1. If
we wanted a particular ordering of the numerical values associated with each
level of the factor, we can specify this when we create the data object using the
factor command:
ratfactor <- factor(ratsex, levels = c('M','F'))
as.numeric(ratfactor)

## [1] 1 1 2 1 2 2 2
Now we have coded M as 1, and F as 2. What if we wanted to change the labels
from ‘M’ and ‘F’ to ‘male’ and ‘female’? We can do this using the levels function
as follows:
levels(ratfactor)[levels(ratfactor)=='M'] <- 'male'
levels(ratfactor)[levels(ratfactor)=='F'] <- 'female'
ratfactor

## [1] male male female male female female female

## Levels: male female
Notice that in the above code we used a logical statement inside the square
brackets to find the levels of the ratfactor object that were either ‘M’ or ‘F’. For
example, the code levels(ratfactor)==‘M’ is interpreted as ‘find the levels of the
ratfactor object that have the value M’. This is the most straightforward way to
recode a factor using base R commands, though there are also functions that
achieve the same effect in other packages, including the revalue and mapvalues
functions in the plyr package, and the recode_factor function in the dplyr
package.

3.10 Putting it all together - importing and

cleaning some real data
In this section we will load in and analyse some data based on a real experiment
conducted using the online Qualtrics platform. I have actually replaced the
original data with some simulated values, to better illustrate the points we have
made in this chapter. However it is close enough to the sort of thing one might
download from an online platform for collecting data (many of which now exist).
58 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

The experiment was quite lighthearted - it was a general knowledge quiz with
ten questions. Before seeing the questions, participants rated their own general
knowledge ability on a scale from 0-100. The idea was to see if these ratings
predicted actual performance in the quiz.
We can load in the data using the read.csv function (the data can be downloaded
from the book’s GitHub repository), and take a look at the first few rows as
follows:
quizdata <- read.csv('data/qualtricsexample.csv')
head(quizdata)

## StartDate EndDate Progress Duration..in.seconds. Finished

## 1 17/03/2021 03:08 17/03/2021 03:08 100 34 TRUE
## 2 17/03/2021 03:08 17/03/2021 03:09 100 20 TRUE
## 3 23/03/2021 11:05 23/03/2021 11:05 100 54 TRUE
## 4 23/03/2021 11:05 23/03/2021 11:14 100 564 TRUE
## 5 23/03/2021 11:05 23/03/2021 11:14 100 555 TRUE
## 6 23/03/2021 11:05 23/03/2021 11:14 100 568 TRUE
## RecordedDate ResponseId DistributionChannel UserLanguage Q1 Q2 Q3
## 1 17/03/2021 03:08 R_1jx28z4EYzdLmNd anonymous EN 84 D A
## 2 17/03/2021 03:09 R_20U6cF4f7pFfbxb anonymous EN 77 C D
## 3 23/03/2021 11:05 R_oXHNzSI4aAISjxn anonymous EN 51 D C
## 4 23/03/2021 11:14 R_29iR3XxZbzDWizy anonymous EN 83 B A
## 5 23/03/2021 11:14 R_1dBvXJRckeXpMrZ anonymous EN 88 C A
## 6 23/03/2021 11:14 R_1qURia8GbrDPfAE anonymous EN 68 A B
## Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11
## 1 C A A A B A C C
## 2 A A B C A A C
## 3 D B A D D A A A
## 4 D A C A C A A A
## 5 A D A A B C A A
## 6 A B A C D A C A
The first 10 columns are things we aren’t really that interested in - they contain
metadata about the running of the experiment that we can ignore. We can strip
these columns out by restructuring the data frame to only include the final 11
columns as follows (sometimes it is also necessary to remove the first few rows
in a similar way):
quizdata <- quizdata[,10:20]
head(quizdata)

## Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11
## 1 84 D A C A A A B A C C
## 2 77 C D A A B C A A C
## 3 51 D C D B A D D A A A
## 4 83 B A D A C A C A A A
3.10. PUTTING IT ALL TOGETHER - IMPORTING AND CLEANING SOME REAL DATA59

## 5 88 C A A D A A B C A A
## 6 68 A B A B A C D A C A
After restructuring, the first column contains the ratings from 0 - 100. Let’s take
a closer look at these values using a histogram, shown in Figure 3.11a as follows:

Histogram of quizdata$Q1 Histogram of quizscores

15 10 8 ● ●

(a) (b) (c)

7 ● ● ●

6 ● ●● ● ● ● ●

10
6

quizscores
5
Frequency

Frequency

● ● ● ●●●● ●

4 ●● ●● ● ● ● ● ●
4
5
3 ● ● ● ● ●● ● ●

2
2 ● ● ● ● ● ● ●

0 0 1 ● ●

0 20 40 60 80 100 1 2 3 4 5 6 7 8 40 50 60 70 80 90

quizdata$Q1 quizscores quizdata$Q1

Figure 3.11: Histograms and scatterplot for the example qualtrics data. Panel
(a) shows the histogram of self-ratings of general knowledge, panel (b) shows
the histogram of actual quiz performance, and panel (c) shows the correlation
between the two measures.

hist(quizdata$Q1)

Two features are clear from the histogram: there are two outlier points with a
value of 0, and overall the distribution looks negatively skewed. Perhaps the
ratings of 0 were genuinely participants who thought they had very poor quiz
ability. But it could also be that 0 was the default rating on the scale used, and
these participants did not change it for whatever reason. We can use the code
from earlier in the chapter to identify data points that are more than 3 standard
deviations from the mean as follows:
criterion <- 3*sd(quizdata$Q1) # calculate 3 times the standard deviation
normdata <- abs(quizdata$Q1-mean(quizdata$Q1)) # subtract the mean and take the absolute value
which(normdata>criterion) # find the indices of any outlier values

## [1] 21 30
This code identifies the participants in rows 21 and 30, and indeed these are the
ones that produced ratings of 0:
quizdata$Q1[c(21,30)]

## [1] 0 0
Given our concerns about the possibility that these participants did not use the
rating scale correctly, and their distance from the rest of the scores, we might
be justified in removing them from the data set. We can do this using another
60 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

which statement, but this time to include only the participants whose ratings
are less than 3 standard deviations from the mean, as follows:
quizdata <- quizdata[which(normdata<criterion),]

To see whether the data are normally distributed, we can run the Shapiro-Wilk
test as follows:
shapiro.test(quizdata$Q1)

##
## Shapiro-Wilk normality test
##
## data: quizdata$Q1
## W = 0.96021, p-value = 0.08514
Although the data show some evidence of negative skew, the test is (just)
non-significant (p = 0.085), so we can proceed assuming a normal distribution.
Next we can look at the answers to the quiz questions themselves. These are
all four-option multiple choice questions, with the answers stored as A,B,C and
D. When we loaded in the data, R converted the responses to factors. However,
some of the questions have not been answered by certain participants, and we
have some missing values. Just as we might in an exam, we will mark the
questions with missing data as incorrect.
To score the quiz, we will use two loops (see section 2.10), one inside the other.
The outer loop will run through each participant in turn, and the inner loop
will run through each question for a given participant. For this example, we will
assume that the correct answer for each question was ‘A’, and we will count up
the number of questions each participant got right and store this in a new data
object called quizscores:
# make a list of zeros to store the scores
quizscores <- rep(0,nrow(quizdata))

for (participant in 1:nrow(quizdata)){ # loop through participants

for (question in 1:10){ # loop through questions
# check if the answer to this question was correct (A)
if (as.character(quizdata[participant,question+1])=='A'){
# if so, increase the score for this participant by 1
quizscores[participant] <- quizscores[participant] + 1}
}
}

We can inspect a histogram of these scores, again using the hist function, as
shown in Figure 3.11b, and also run the Shapiro-Wilk test:
hist(quizscores)
shapiro.test(quizscores)
3.11. PRACTICE QUESTIONS 61

##
## Shapiro-Wilk normality test
##
## data: quizscores
## W = 0.95898, p-value = 0.07544
Despite the somewhat unusual form of the data, with evidence of a floor effect
at the lower end, the Shapiro-Wilk test is again non-significant. Finally, we
can inspect a scatterplot of the two variables plotted against each other, to
see if there is evidence that participants were able to predict their own general
knowledge ability:
plot(quizdata$Q1,quizscores,type='p')

The graph produced by the above code is shown in Figure 3.11c. It does appear
to be the case that individuals who rated their ability more highly also obtained
generally higher test scores, which we might go on to test using correlation, or
other statistical tests described in Chapter 4. We can also check for multivariate
outliers using the Mahalanobis distance as follows:
bothscores <- data.frame(quizdata$Q1,quizscores)
D <- mahalanobis(bothscores,colMeans(bothscores),cov(bothscores))
sort(round(D,digits=2))

## [1] 0.02 0.18 0.18 0.23 0.49 0.51 0.51 0.51 0.61 0.64 0.72 0.74 0.74 0.79 0.90
## [16] 0.99 1.01 1.04 1.07 1.12 1.17 1.19 1.30 1.30 1.32 1.35 1.47 1.63 1.66 1.72
## [31] 1.84 1.86 1.89 2.21 2.21 2.30 2.47 2.52 2.55 3.11 3.15 3.34 3.46 3.91 3.96
## [46] 3.98 4.36 4.57 5.21 5.71 8.31
By sorting the distances using the sort function, we can see that the largest
value is 8.31, which does not exceed our criterion of D2 = 9 (or D = 3).
The Qualtrics quiz data is a deliberately simple example - most data sets
would involve multiple conditions, experimental manipulations, or independent
variables. However this has hopefully given a good indication of how we can
import and clean a data set in preparation for further analysis.

3.11 Practice questions

1. Two conventions for arranging data in R are known as:
A) wide and tall
B) wide and long
C) melted and spread
D) spread and long
2. The whiskers on a boxplot conventionally show:
A) The standard error
B) The interquartile range
C) The inner fence
62 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS

D) 95% confidence intervals

3. Which of the following is not a method for replacing outliers?
A) Windsorization
B) Mean imputation
C) Clipping
D) The interquartile range
4. Which method would be most suitable for identifying outliers in a correla-
tion analysis?
A) The modified Thompson’s Tau statistic
B) Tukey’s outer fence method
C) The Chauvenet criteria
D) The Mahalanobis distance
5. For normally distributed data, Tukey’s inner fence is approximately equiv-
alent to which multiple of the standard deviation?
A) 1.5
B) 1.96
C) 2.7
D) 3
6. A 20% trimmed mean would be calculated using:
A) The central 60% of values
B) The central 20% of values
C) The central 80% of values
D) The central 90% of values
7. Which of the following is not a method that can be used when the assump-
tions of parametric statistics are violated?
A) Bootstrap tests
B) The Kolmogorov-Smirnov test
C) Non-parametric statistics
D) Robust statistics
8. Which command would normalize a data set by its standard deviation
only:
A) scale(data)
B) scale(data,scale=FALSE)
C) scale(data,center=FALSE)
D) scale(data)/sd(data)
9. To assess deviations from a normal distribution, you could use:
A) The Kolmogorov-Smirnov test
B) A Q-Q plot
C) The Shapiro-Wilk test
D) All of the above
10. To make positively skewed data more normal, you could use a:
A) Logarithmic transform
B) Exponential transform
C) Squaring transform
D) Additive transform
3.11. PRACTICE QUESTIONS 63

Answers to all questions are provided in section 20.2.

64 CHAPTER 3. CLEANING AND PREPARING DATA FOR ANALYSIS
Chapter 4

Statistical tests as linear

models

This chapter will demonstrate R implementations of statistical tests that are

commonly taught on introductory statistics courses. I would anticipate that most
readers will be familiar with such tests, including t-tests, regression, ANOVA
and correlation. If you have not come across these before, you might find it
helpful to read an introductory statistics text (e.g. Navarro 2019; Field, Miles,
and Field 2012), which will explain them in much more detail. I will introduce
them using a slightly different approach from that traditionally taken - we will
see in this chapter that many such tests can be considered in terms of explaining
data using a linear model (i.e. a straight line).

4.1 Many statistical tests involve comparing dif-

ferent models
Statistical tests were developed by different people at different times. For
example, the t-test (which compares the means of two groups) was developed by
William Sealey Gosset (writing as Student) who was employed by the Guinness
corporation during the early 20th century to analyse data relating to brewing
beer. Student needed to establish whether different brewing methods affected
the end product, despite having only small samples of noisy data to work with.
Analysis of Variance (ANOVA) was developed by Ronald Fisher for analysing
data on crop yields (Fisher unfortunately held some abhorrent views on issues
such as race and eugenics). The way these methods are usually taught makes
them seem like totally distinct techniques, with different underlying assumptions
and operations. But really, they are doing fundamentally the same thing - fitting
straight lines to data, and checking how much variance (or error) is left over
that the straight lines can’t explain.

65
66 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

Considered from this perspective, we will see that many statistical tests involve
fitting a model to try and explain our data. Usually the model assumes that
our measurements (known as the dependent variable) can be predicted to some
extent by one or more other factors (known as independent variables). We
can then compare the fit of this statistical model to a ‘null model’, in which
those other factors do not predict our measurements. If our model explains the
data better than the null model (according to some criterion), we consider it to
be statistically significant (we will discuss these assumptions in more detail in
Chapter 17). The clearest way to demonstrate the model comparison approach is
by starting with linear regression (where it is most explicit), and then applying
the same logic to other situations.

4.2 Regression (and correlation)

In linear regression, we look at the influence of a predictor (independent variable)
on some sort of outcome measure (dependent variable). Both variables are
continuous, rather than being split into discrete categories. We will demonstrate
regression using an example data set from the literature, on growth rates of
moso bamboo. Bamboo is a fast growing plant, and so has the potential for
rapid carbon sequestration, which has important environmental implications. A
study by Yen (2016) measured the relationship between the diameter at breast
height (DBH - the width of the plant at around 140cm from the ground) and the
overall culm (stem) height. There is a very straightforward linear relationship
across a sample of 30 plants, shown by the data points in Figure 4.1.
The data (which I extracted from a figure in the original paper) are stored in
a data frame called bamboodata. It has two columns, containing the DBH and
culm height data for 30 plants. The first few rows look like this:
head(bamboodata)

## DBH culmheight
## 1 6.39 7.92
## 2 6.19 8.39
## 3 6.88 8.03
## 4 6.69 8.74
## 5 6.48 9.06
## 6 7.18 8.33
It is clear from Figure 4.1 that the data are highly correlated, which we can
confirm by calculating a correlation coefficient by passing the data to the cor
function:
cor(bamboodata)

## DBH culmheight
## DBH 1.0000000 0.9314495
## culmheight 0.9314495 1.0000000
4.2. REGRESSION (AND CORRELATION) 67

10 11 12 13 14

●
● ●
● ●
●
● ● ●
Culm Height (m)

●
● ●
●
●
●

● ●
● ●
● ●
● ●
●
●
9

●
● ●
●
8

●
7

5 6 7 8 9 10 11 12

Diameter at Breast Height (cm)

Figure 4.1: Linear regression between diameter at breast height and culm height
for moso bamboo, based on Figure 1 of Yen (2016).
68 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

If we were interested in statistical significance, there is also a test version

(cor.test), which requires us to pass the two variables (columns) in as separate
arguments:
cor.test(bamboodata$DBH,bamboodata$culmheight)

##
## Pearson's product-moment correlation
##
## data: bamboodata$DBH and bamboodata$culmheight
## t = 13.545, df = 28, p-value = 8.122e-14
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## 0.8596580 0.9671648
## sample estimates:
## cor
## 0.9314495
The output gives us the same correlation coefficient on the final line (along with
95% confidence intervals just above), and also calculates a t-statistic and p-value
to assess statistical significance (on the fifth line). Notice that the p-value is
expressed in scientific notation as 8.122e-14. This is how R represents very
small numbers, in this case it means 8.122 × 10−14 , or 0.00000000000008122 (the
easiest way to think of this is that you shift the main number by the value given
after the e - here 14 places to the right).
In regression, we want to fit a linear model (i.e. a straight line) that allows us
to predict values of the dependent variable (culm height) using the values of
the independent variable (DBH; note that in this example these are both things
that have been measured, and so DBH might not meet a strict criteria for being
an independent variable because it will involve measurement error, but this is
just an example). To do this, we will use the lm (linear model) function in R.
The lm function (as well as other related functions, including those for running
ANOVA) uses a syntax to specify models, that has the general form DV ~ IV.
The tilde symbol (~) means is predicted by. In other words, we’re saying that the
dependent variable is predicted by the independent variable. For our example,
we want to run the model culmheight ~ DBH, and we will also tell the lm function
the name of the data frame containing our data (bamboodata). Finally, we will
store the output of the model in a new data object called bamboolm. We do this
with a single line of code, and then have a look at the output using the generic
summary function:
bamboolm <- lm(culmheight ~ DBH, data=bamboodata)
summary(bamboolm)

##
## Call:
## lm(formula = culmheight ~ DBH, data = bamboodata)
4.2. REGRESSION (AND CORRELATION) 69

##
## Residuals:
## Min 1Q Median 3Q Max
## -1.26249 -0.50913 -0.02087 0.45275 1.43376
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.27726 0.71223 1.793 0.0837 .
## DBH 1.12081 0.08274 13.545 8.12e-14 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.6488 on 28 degrees of freedom
## Multiple R-squared: 0.8676, Adjusted R-squared: 0.8629
## F-statistic: 183.5 on 1 and 28 DF, p-value: 8.122e-14

The output tells us what we have done (repeating the function call), and then
gives us a table of (unstandardised) coefficients. These are the intercept (1.28,
given in the Estimate column for the (Intercept) row) and the slope of the fitted
line (1.12, given in the Estimate column for the DBH row). We can use these to
plot a straight line with the equation y = β0 + β1 x (where β0 and β1 are the
model coefficients that correspond to the y-intercept and gradient), which in
this case is culmheight = 1.28 + 1.12*DBH. That is the line that is shown in
Figure 4.1, and which gives an excellent fit to the data. If we need to obtain
standardised regression coefficients we can either standardise the data first (see
section 3.7), or run the lm.beta function from the QuantPsyc package on the
model output object (e.g. lm.beta(bamboolm) for the above example).
The summary table also provides some other useful statistics. There is an
R-squared value (R2 = 0.87) telling us the proportion of the variance explained,
and an overall F-statistic (F = 183.5) and p-value for the regression model. The
p-value is the same as the one for the correlation, and for the DBH coefficient
(in the Coefficients table), because we only have one predictor. For multiple
regression models the table will indicate the significance of each predictor, and the
F-statistic will tell you about the full model. We are also given some information
about the residuals (the left over error that the model can’t explain) which can
be used to check the assumptions of the test (more on this in section 4.5).
So that’s how to run a straightforward linear regression in R. But to set things
up for the rest of this chapter, we should dig a little bit deeper into what linear
regression is actually doing. We have fitted a line to describe our data, but
what does the p-value indicate? In regression, we are effectively comparing two
different models. In the null model, the line has a slope of 0, which means there
is no effect of the value of DBH on culmheight. The best we can do to predict
culmheight is to use the overall mean (sometimes known as an intercept-only
model). In the alternative model, the slope of the line is allowed to vary to try
and fit the data better. Usually in regression we don’t bother to show both lines,
70 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

but it’s worth making them explicit so that you can see the difference - they are
plotted in Figure 4.2.

Null model Alternative model

14 14
● ●
● ●
● ● ● ● ● ●
13 ● 13 ●

● ● ● ● ● ●
12 ● 12 ●
Culm Height (m)

Culm Height (m)

● ● ● ●
● ●
● ●
11 ● 11 ●

● ● ● ●
● ● ● ●
10 10
● ●
● ●● ● ●●
● ●
9 ● 9 ●
● ●
● ● ● ●
8 ● ● 8 ● ●

7 7

5 6 7 8 9 10 11 12 5 6 7 8 9 10 11 12

Diameter at Breast Height (cm) Diameter at Breast Height (cm)

Figure 4.2: Null and alternative model fits for the bamboo data. The null model
has a slope of 0, the alternative model can have any slope value. Thin vertical
lines show the residuals between model and data.

In both panels of Figure 4.2 I have also added some thin vertical lines, that join
the (thick black) fitted line to each individual data point. These are called the
residuals - they are the error between the data and the model prediction. One
way of thinking about residuals is that they represent how well the model (thick
line) is able to describe the data (points). If the fit is poor, the residual lines will
be long (as in the null model). If the fit is good, the residual lines will be short
(as in the regression model). The proportion of the variance explained (R2 ) and
the statistical comparison between the null and alternative models are based on
the lengths of these lines (though we will not go into further details here about
precisely how this works). I think of it as the left over variance (i.e. differences
between points) that the models cannot explain. The p-value in regression is
really telling us whether the alternative model can explain significantly more of
the total variance than the null model.

4.3 T-tests
It is rarely made explicit in introductory texts that this basic idea, of assessing the
fits of two models by comparing the left over variances, is also what is happening
in other tests such as t-tests. The t-test is used to compare the means of two
groups to see if they differ from each other. Usually, this is explained as taking
the mean difference and dividing by the pooled standard error (which is the
4.3. T-TESTS 71

equation of the t-statistic). However, we can also think of a t-test as comparing

two models: a null model where the two means are the same, and an alternative
model where they differ. Moreover, we can consider group membership to be
a predictor variable just like the independent variable in regression - the only
difference being that it takes on discrete values (1 and 2), rather than continuous
values.
To demonstrate this, we will split our example data set in half. All of the bamboo
plants with a DBH value greater than 8.5 cm will be in the wide group, and all
of the plants with a DBH value below 8.5 cm will be in the narrow group. In
general it is a bad idea to throw away information in this way, but it might be
justified in some contexts, and besides this is just an example. We can perform
this split by creating a grouping variable in R, and adding it to the bamboodata
data frame:
sizegroup <- NULL # make an empty data object to store the group labels
sizegroup[which(bamboodata$DBH<8.5)] <- 1 # set the narrow plants to be in group 1
sizegroup[which(bamboodata$DBH>8.5)] <- 2 # set the wide plants to be in group 2
bamboodata$sizegroup <- as.factor(sizegroup) # add to data frame as a factor
sizegroup

## [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
The which function in the above code returns the indices of the DBH vector
that satisfy the conditional statement (e.g. it tells us which entries in the DBH
vector are less than 8.5, or greater than 8.5). The as.factor command tells R
that the data should be treated as categorical (factor) labels for the purposes of
conducting statistical tests (see section 3.9). The data with the group split are
plotted in Figure 4.3.
In R, we can then run a t-test using the t.test function. The standard way to do
this is to split the culmheight data into two separate data objects for the narrow
and wide groups, and then plug them into the t-test function:
group1 <- bamboodata$culmheight[bamboodata$sizegroup==1]
group2 <- bamboodata$culmheight[bamboodata$sizegroup==2]
t.test(group1, group2, var.equal=TRUE)

##
## Two Sample t-test
##
## data: group1 and group2
## t = -10.016, df = 28, p-value = 9.298e-11
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## -3.66897 -2.42303
## sample estimates:
## mean of x mean of y
## 9.267333 12.313333
72 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

14
● ●●
●● ●
12 ●
●
● ● ●
● ●●
●●
Culm Height (m)

●● ●
10

●
●
●● ●●
●
● ●
● ●
● ●
8
6
4
2
0

Narrow (1) Wide (2)

Group

Figure 4.3: Culm height data split into narrow and wide groups by DBH value.

The 5th line of the output gives us a large t-value (-10), and a very small p-value,
indicating a highly significant group difference. An alternative syntax would be
to use the same formula structure that we used for regression, which the t.test
function also accepts. This time we are predicting the height values using group
membership, so the appropriate formula is culmheight ~ sizegroup:
t.test(culmheight ~ sizegroup, data=bamboodata, var.equal=TRUE)

##
## Two Sample t-test
##
## data: culmheight by sizegroup
## t = -10.016, df = 28, p-value = 9.298e-11
## alternative hypothesis: true difference in means is not equal to 0
## 95 percent confidence interval:
## -3.66897 -2.42303
## sample estimates:
## mean in group 1 mean in group 2
## 9.267333 12.313333

This is a different way of achieving exactly the same result, and you can see that
the outcomes are identical. But, we could also run the test explicitly as a linear
model (using the lm function), again with the sizegroup variable as the predictor
4.3. T-TESTS 73

and the culmheight variable as the outcome.

lmttest <- lm(culmheight ~ sizegroup, data=bamboodata)
summary(lmttest)

##
## Call:
## lm(formula = culmheight ~ sizegroup, data = bamboodata)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.35333 -0.74333 0.06967 0.79667 1.18667
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 9.2673 0.2150 43.09 < 2e-16 ***
## sizegroup2 3.0460 0.3041 10.02 9.3e-11 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8329 on 28 degrees of freedom
## Multiple R-squared: 0.7818, Adjusted R-squared: 0.774
## F-statistic: 100.3 on 1 and 28 DF, p-value: 9.298e-11

The output looks different from the output of the t-test function, as it has the
same layout as the regression output we saw in the previous section. But you
can see that the values of the t-statistic and p-value in the table of coefficients
are exactly the same as the ones we got from the t-test function (the minus sign
is missing from the t-statistic, but this is arbitrary anyway because it depends
on the order in which the groups are entered).

Now, this consistency across methods prompts us to think about the t-test in
the context of regression. Just like with regression, we can conceptualise the
t-test as comparing two models. The null model is one in which the means do
not vary with group, given by the horizontal black line in the left panel of Figure
4.4. The alternative model is one where the means can vary with group, given
by the diagonal black line in the right panel of Figure 4.4.

Just as with regression, we can calculate the residual error between each data
point and the accompanying model prediction for its group. The model prediction
for the null model is the grand mean (horizontal black line). The model prediction
for the alternative model is the group mean for each condition. Then we
compare the two model fits statistically to see if the alternative model describes
significantly more of the variance than the null model. This is another way of
thinking about what a significant t-test means: conceptually it is exactly the
same as linear regression.
74 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

Null model Alternative model

14 ●
14 ●
●● ● ●● ●
● ● ● ●
12 ●●
● ● ●
● 12 ●●
● ● ●
●
●● ●●
●● ●●
● ●●● ● ●●●
Culm Height (m)

Culm Height (m)

10 ●●
10 ●●
●● ● ●● ●
●
●● ●
●●
● ● ● ●
8 ●● 8 ●●

6 6

4 4

2 2

0 0

Figure 4.4: T-tests conceptualised a comparison between a null model where the
group means do not differ (left) and an alternative model where they do (right).

4.4 ANOVA
Finally, we can extend the same regression logic to analysis of variance (ANOVA),
where the independent variable has more than two levels. In the bamboo paper by
Yen (2016), the data set is split into 5 groups by DBH value, in 1cm increments
as shown in the two graphs in Figure 4.5.

Again, we can think of ANOVA as comparing a null model where the predicted
values are not affected by group (left panel), with an alternative model where the
predicted values change across group (right panel). Notice that the alternative
model involves specifying four separate lines (thick lines joining the means),
which can have different slopes. This is why the number of degrees of freedom
for the independent variable in a one-way ANOVA is always one less than the
number of levels. To conduct the ANOVA in R, we can use the aov function
(the DBHgroup column contains the groupings):
anovamodel <- aov(culmheight ~ DBHgroup, data=bamboodata)
summary(anovamodel)

## Df Sum Sq Mean Sq F value Pr(>F)

## DBHgroup 4 75.42 18.854 34.68 7.28e-10 ***
## Residuals 25 13.59 0.544
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Or we can achieve the same result using the linear model (lm) function:
4.4. ANOVA 75

Null model Alternative model

14 14
● ●
● ● ● ● ● ●
● ●
● ●
● ●

● ●● ● ● ●
● ●
12 ● 12 ●
● ● ● ●
Culm Height (m)

Culm Height (m)

● ●
● ●
●
●
●
●

● ● ● ●
● ● ● ●
10 ●
10 ●
● ●●●
● ●●●

● ●
● ●
●
●
●
●

● ● ● ●
8 ●● 8 ●●

6 6

1 2 3 4 5 1 2 3 4 5

DBH Group DBH Group

Figure 4.5: ANOVA conceptualised as a comparison between a null model where

the group means do not differ (left) and an alternative model where they do
(right).

anovalm <- lm(culmheight ~ DBHgroup, data=bamboodata)

summary(anovalm)

##
## Call:
## lm(formula = culmheight ~ DBHgroup, data = bamboodata)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.2283 -0.5837 0.1033 0.4988 1.5800
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 8.5850 0.3010 28.520 < 2e-16 ***
## DBHgroup2 0.9733 0.4257 2.286 0.031 *
## DBHgroup3 2.1850 0.4257 5.133 2.64e-05 ***
## DBHgroup4 3.6100 0.4257 8.480 7.99e-09 ***
## DBHgroup5 4.2583 0.4257 10.003 3.19e-10 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.7373 on 25 degrees of freedom
## Multiple R-squared: 0.8473, Adjusted R-squared: 0.8229
## F-statistic: 34.68 on 4 and 25 DF, p-value: 7.277e-10
76 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

Table 4.1: Table summarising how common statistical tests can be implemented
in R, using both a dedicated function and a linear model. The terms DV and
IV (IV1, IV2) are assumed to be column names of the dependent (DV) and
independent (IV) variables in a data object called ‘dataset’, and also to exist
as independent data objects. The term ID indicates a participant identification
variable for repeated measures tests. The lme function is part of the nlme package,
and the ezANOVA function is part of the ez package. Note that alternative
implementations for more complex designs can produce different results, and do
not necessarily test appropriate assumptions, or make the same corrections.
Test Generic function call
One-sample t-test t.test(DV)
Independent t-test t.test(DV[which(IV==1)], DV[which(IV==2)], var.equal=TRUE)
Dependent (paired) t-test t.test(DV[which(IV==1)], DV[which(IV==2)], paired=TRUE)
Linear regression summary(lm(DV ~ IV, data=dataset))
Multiple regression summary(lm(DV ~ IV1 + IV2, data=dataset))
One-way independent ANOVA summary(aov(DV ~ IV, data=dataset))
One-way repeated measures ANOVA ezANOVA(dataset, dv=DV, wid=ID, within=IV)
Factorial independent ANOVA summary(aov(DV ~ IV1 * IV2, data=dataset))
Factorial repeated measures ANOVA ezANOVA(dataset, dv=DV, wid=ID, within=c(IV1,IV2))
Mixed design ANOVA ezANOVA(dataset, dv=DV, wid=ID, within=IV1, between=IV2)

Test Linear model call

One-sample t-test summary(lm(DV ~ 1, data=dataset))
Independent t-test summary(lm(DV ~ IV, data=dataset))
Dependent (paired) t-test summary(lm(DV[which(IV==1)] - DV[which(IV==2)] ~ 1, data=dataset))
Linear regression summary(lm(DV ~ IV, data=dataset))
Multiple regression summary(lm(DV ~ IV1 + IV2, data=dataset))
One-way independent ANOVA summary(lm(DV ~ IV, data=dataset))
One-way repeated measures ANOVA anova(lme(DV ~ IV, random = ~1|ID/IV, data=dataset))
Factorial independent ANOVA anova(lm(DV ~ IV1 * IV2, data=dataset))
Factorial repeated measures ANOVA anova(lme(DV ~ IV1 * IV2, random = ~1|ID/IV1/IV2, data=dataset))
Mixed design ANOVA anova(lme(DV ~ IV1 * IV2, random = ~1|ID/IV1, data=dataset))

Notice that the F-statistic in the ANOVA summary table, and the final line
of the regression output, are identical (F = 34.68). This is because the un-
derlying calculations for the tests we call ANOVAs, t-tests and regressions are
fundamentally the same thing (a linear model).

If required, it is also possible to conduct factorial and repeated measures ANOVAs

in R. The formula syntax for linear models can be easily extended to include
multiple independent variables, using the format: DV ~ IV1 * IV2. The asterisk
denotes factorial combination, so that interaction terms will also be calculated.
Repeated measures components can be included as random factors, which we will
discuss in more detail in Chapter 7. This can be achieved using the ezANOVA
function from the ez package, or the lme function from the nlme package, by
providing a column of participant ID numbers and specifying which independent
variables are repeated measures.

Table 4.1 provides example R functions for popular tests, using both the generic
function and the linear model form. Traditionally, we would use regression when
our independent variable is continuous, a t-test when it is discrete with two
levels, and ANOVA when it has more levels. But as Table 4.1 illustrates, these
4.5. ASSUMPTIONS OF LINEAR MODELS 77

separate tests are really all part of the wider family of the general linear model,
and can all be implemented within the same framework.

4.5 Assumptions of linear models

It is only appropriate to fit linear models when their assumptions are met. A
key assumption of parametric statistics is that the data are normally distributed
(conforming to a Gaussian, or bell-shaped, curve), which we can test using
the Kolmogorov-Smirnov and Shapiro-Wilk tests we encountered in section 3.8.
Actually for linear models, it is more accurate to say that we test whether the
residuals are normally distributed, as this allows us to make the comparison
on a full data set, taking the effect of the independent variable(s) into account.
For tests comparing different groups, it is also necessary to check whether the
variances are equivalent for each group (the homogeneity of variances assumption),
which is typically achieved using Levene’s test (the leveneTest function in the car
package). Amazingly even Levene’s test can be conceptualised as a linear model,
but one using the absolute residuals rather than the original data points. Finally,
for repeated measures designs we should test whether the pairwise differences
between conditions also have equal variance (the sphericity assumption) in a
similar way using Mauchly’s test (the mauchly.test function). If the assumptions
are not met, it is possible to run alternative versions of many statistical tests
instead, including nonparametric tests. As it turns out, many nonparametric
statistics can even be thought of as linear models, but using the ranks of the
data instead of the data themselves!

4.6 Is everything just a linear model then?

Despite the ubiquity of linear models throughout this chapter, not everything
can be considered in this framework. For example, in Chapter 9, we will discuss
fitting nonlinear curves to data, and in Chapter 10 we will describe data as the
sum of multiple sine waves. However, the general concept of linear models is
relevant to much of the other material in the book, particularly Chapter 7 in
which we will introduce (linear) mixed effects models, and Chapter 11 where
we extend the t-test to cope with multiple dependent variables. More generally,
linear regression can be extended to the case of multiple independent (predictor)
variables, which is called multiple regression. We will not consider multiple
regression further in this book, but there are many good texts available that do
so in R, including Fox and Weisberg (2018), Lilja (2016), and Field, Miles, and
Field (2012).

4.7 Practice questions

1. T-tests were originally developed to analyse data relating to:
A) Brewing beer
78 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS

B) Bamboo growth
C) Crop yields
D) Counting students
2. To test the effect of age on brain volume, the appropriate linear model
formula would be:
A) age ~ brainvolume
B) brainvolume ~ age
C) age - brainvolume
D) brainvolume - age
3. In R, the function to run a t-test is called:
A) ttest
B) t-test
C) t.test
D) Ttest
4. In regression, the residuals indicate:
A) The total variance in the data set
B) The differences between each pair of data points
C) The amount of the variance explained by a model
D) The error between the data and the model fit
5. The null hypothesis produces a model line with a slope of:
A) 1
B) -1
C) 0
D) It depends on the data
6. The alternative hypothesis produces a model line with a slope of:
A) 1
B) -1
C) 0
D) It depends on the data
7. For a one-way ANOVA with three levels, how many regression coefficients
would we expect (not including the intercept)?
A) 1
B) 2
C) 3
D) 4
8. The as.factor function is used to:
A) Define a dependent variable
B) Turn numeric data into text
C) Tell R that a data object is categorical
D) Round number so that it is an integer
9. For a categorical independent variable with four levels, which R functions
could you use to analyse the data?
A) lm or t.test
B) lm or aov
C) aov or t.test
D) lm only
4.7. PRACTICE QUESTIONS 79

10. For factorial ANOVA designs, we indicate an interaction between two

independent variables using:
A) An asterisk symbol (∗)
B) A plus symbol (+)
C) A tilde symbol (~)
D) A slash symbol (/)
Answers to all questions are provided in section 20.2.
80 CHAPTER 4. STATISTICAL TESTS AS LINEAR MODELS
Chapter 5

Power analysis

5.1 What is statistical power?

The power of a statistical test is its ability to detect an effect when it is actually
there. Another way of putting this is to say it is the ability of a test to correctly
reject the null hypothesis. Power is a probability and so it always has a value
between 0 and 1, though it is also sometimes expressed as a percentage. One
way to think about it is to imagine that we ran the same experiment lots of
times, and then counted how often the experiments produced significant results.
Consider an experimental design with a very strong effect and a very large
sample size. The Stroop effect (Stroop 1935) is a neat example. Participants
view coloured letters that are arranged to spell out the names of colours in the
participant’s native language. They are asked to press a button to indicate
the colour of the letters. On average, people are slower to report the colour of
the letters when it is incongruent with the word the letter spells (i.e. the word
‘green’ written in red letters), than when they are congruent (i.e. the word ‘green’
written in green letters).
If we were to repeat this experiment 100 times on 100 different people each time
(fortunately this is just hypothetical!), and 98 of the repetitions produced a
significant difference in reaction time between congruent and incongruent words,
we would say that this experimental design had a power of 0.98 (or 98%). Now
let’s imagine that we reduced the sample size to 10 participants, and ran the
experiment 100 more times. Because the sample size is smaller, we are much
less likely to find statistically significant effects. Perhaps now only 40 of the
repetitions are significant: our power would be 0.4 (or 40%).
Rather than laboriously running hundreds of replications of an experiment, we
can estimate the power of a study design if we know three things: the sample
size, the effect size, and our criteria for statistical significance. Sample size is

81
82 CHAPTER 5. POWER ANALYSIS

straightforward - it is the number of separate observations we are planning to

make. These observations might be human participants, animal subjects, cells,
plants, or higher-order entities like schools or countries. Traditionally in the life
sciences, p < 0.05 is our criterion for statistical significance (called the α level).
So the last thing we need to know about is the effect size. What’s that?

5.2 Effect size

An effect size is a statistic that quantifies, in standardised units, some effect
of interest. There are two main types of effect size: those that summarise the
amount of variance explained by a statistical model, and those that quantify the
difference between group means. Examples in the former category include the
correlation coefficient, r, the coefficient of determination R2 calculated during
linear regression, and the η 2 and partial-η 2 effect sizes often used when reporting
ANOVA. All of these measures are based on the proportion of the total variance
in the data set that we can explain using a linear statistical model (see Chapter
4). For example, an R2 value of 0.9 tells us that 90% of the variance in our
dependent variable is explained by the independent variable. These measures
are standardised, so they can be easily compared across different studies, giving
us an idea of how important an effect is in both relative and absolute terms. In
Chapter 6 we will introduce meta analysis - a method for pooling effect sizes
across studies.
The other type of effect size is based on the standardised difference in means.
The general idea is that we want to be able to quantify the difference of group
means in a way that also considers the variability of the data. This results in
values that are independent of the units of measurement, and so can again be
easily compared across studies. For t-tests and related statistics, Cohen’s d is a
widely used measure of effect size, which we will focus on here. It is defined as
the mean difference scaled by the standard deviation (Cohen 1988):

x̄1 − x̄2
d= , (5.1)
σ

where x̄1 and x̄2 are the group means, and σ is the pooled standard deviation.
This is a standardized score, conceptually similar to the z-score, but for means
rather than individual observations. Because the denominator is the standard
deviation (and not the standard error), the effect size is independent of sample
size, although effect size estimates do become more accurate as sample size
increases. Other related effect sizes include Hedge’s g and Glass’s δ, which
slightly vary the denominator term. For multivariate statistics, the Mahalanobis
distance (Mahalanobis 1936) extends Cohen’s d to the multivariate case (see
section 3.4.3).
As a heuristic, Cohen (1988) suggested that effect sizes (d) of 0.2, 0.5 and 0.8
correspond to small, medium and large effects respectively. Let’s think about how
5.3. HOW CAN WE ESTIMATE EFFECT SIZE? 83

this applies to some hypothetical data sets. Imagine we want to know if tortoises
can run faster than hares. We time a group of hares and a group of tortoises
running along a racetrack. In the first race, the mean times are 57 seconds for
the hares, and 108 seconds for the tortoises. The difference (51 seconds) seems
large. But when we look at the raw data, we see that the individual animals are
quite variable in how long they take - maybe some of them get distracted eating
grass, whereas others are more on task. We calculate the standard deviation
as being 200 seconds, meaning that our effect size is d = (108-51)/200 = 0.26 -
quite a small effect according to Cohen’s heuristics.
Next suppose we re-run the race but we remove all of the distractions so that
the animals stay focussed. The mean times are rather shorter overall, 32 seconds
for the hares, and 80 seconds for the tortoises. Notice that the mean difference
is about the same as it was before - 48 seconds this time. But this time the
standard deviation is much smaller, at just 54 seconds. The smaller standard
deviation means that our effect size ends up being much larger: d = (80-32)/54
= 0.89. So even though the raw difference in means has stayed the same, the
precision of the measurement has improved.
One way to think about Cohen’s d is to consider the underlying population
distributions implied by different values of the statistic. Figure 5.1 shows four
pairs of distributions, with various mean differences and standard deviations.
The figure illustrates that one can increase d either by increasing the difference
between the group means, or by reducing the variance (spread) of the distri-
butions. Although in many situations these properties are fixed, and we must
simply measure them as best we can, in some experimental contexts it may be
possible to influence them to increase an effect size and boost statistical power.
For example, to increase the effect of a drug or intervention, one could apply a
higher drug dose, or a longer intervention duration. In section 5.7 we will discuss
how collecting more data for each individual in a study can often reduce the
overall standard deviation.

5.3 How can we estimate effect size?

In order to estimate the power of a study we plan to run, we need to know the
likely effect size. But how can we estimate this before we’ve run the study? If
we are running a replication study, we can use the effect size from the original
study. However we should note that this is likely to be inflated if the original
study was underpowered (Ioannidis 2008). We can understand why this should
be so by thinking about the distribution of observed effect sizes across many
hypothetical repetitions of a study.
When power is low, effect size estimates are very variable. Only the very largest
effect sizes are significant, and these are much bigger than the true effect size,
so must be an overestimate. This is illustrated in the left panel of Figure 5.2,
which shows the results of a simulation (see Chapter 8 for discussion of how to
84 CHAPTER 5. POWER ANALYSIS

d=1 d=2

−4 −2 0 2 4 −4 −2 0 2 4

d=2 d=4

−4 −2 0 2 4 −4 −2 0 2 4

Figure 5.1: Example pairs of distributions that correspond to different effect

sizes. The Cohen’s d statistic for each pair is given above the panel. The shaded
blue region highlights the overlap between the curves; more shading implies a
smaller effect size.
5.3. HOW CAN WE ESTIMATE EFFECT SIZE? 85

run such simulations) in which t-tests were run on 1000 synthetic data sets, each
comprising N=10 participants. The true (generative) effect size is given by the
white diamond, and the individual points are the estimated effect sizes for each
data set. Grey points below the horizontal line are non-significant, and blue
points above it are significant. If we imagine that only the significant ‘studies’
were published (an effect known as publication bias), we might estimate a mean
effect size around d = 1 (blue diamond), much higher than the true effect size of
d = 0.5 (white diamond).
If the study design has a larger sample size (N = 50), the estimates of effect
size become more accurate and regress to the mean. The spread of effect sizes
becomes tighter about the true value, and most repetitions return an effect size
close to the actual effect size (see right panel of Figure 5.2). Note that because
of the larger sample size, the power is higher, and a smaller observed effect
size is required for statistical significance (i.e. the horizontal line moves down).
Effect size estimates from underpowered studies should therefore be treated with
caution because they are more likely to overestimate the true effect size.

3 3
True effect size
Estimated effect size
Effect size (d)

Effect size (d)

2 2

1 1

0 0

N=10 N=50
−1 −1

Figure 5.2: Simulations to demonstrate effect size inflation resulting from un-
derpowered studies. 1000 data sets were generated using a mean of 1 and a
standard deviation of 2, for 10 participants per data set (left) or 50 participants
per data set (right). Points above the horizontal lines indicate significant effects,
and points below the lines indicate non-significant effects. White diamonds are
the true effect sizes, and blue diamonds are effect size estimates calculated only
from significant studies. The position of each point along the x-axis is arbitrary.

Another approach is to find or do a meta analysis on the topic in question. We

86 CHAPTER 5. POWER ANALYSIS

will cover meta analysis in more detail in Chapter 6, but in brief it is a technique
for calculating the average effect size across a number of studies. Because this
increases the overall power, the effect size estimate is likely to be more accurate.
Note that the effects of publication bias can still influence meta analyses, as
non-significant results are less likely to be available for inclusion in the analysis.
When conducting novel or exploratory research, there may not be any suitable
studies on which to base our effect size estimates. One common solution is to
run a pilot study with a smaller sample. This is commonplace in clinical trials,
where the eventual sample size is very large indeed (hundreds or thousands
of participants), and large cost savings can potentially be made by running a
smaller scale pilot study first, perhaps on a few dozen participants. Although
for many lab-based studies this might not be practical, piloting a new experi-
mental paradigm is always worthwhile if possible, as it provides much additional
information besides a possible effect size.
If there really is no existing data to estimate the likely effect size, one can
use Cohen’s heuristics for small, medium or large effect sizes to perform power
calculations. An important concept is the smallest effect size of interest. The
idea here is that effects smaller than some value would be of no theoretical or
practial importance. For example, if a drug treatment had an effect size of d
= 0.01, it would provide no meaningful benefit to patients and would not be
worth the expense of developing. So it might be practical to power a study
to detect a larger effect size that we think would be clinically meaningful. Of
course, as expected effect sizes get smaller, the sample size required to achieve
adequate power will increase, and there is a balance to be struck with practical
considerations around resource allocation for a given study. A useful way to
think about these issues is to plot power curves, as we describe in the next
section.

5.4 Power curves

For a given effect size and α level, we can calculate power as a function of sample
size (N). The resulting power curves are shown in the left panel of Figure 5.3,
and are instructive and perhaps surprising to many the first time they see them.
Although the general trend fits with our intuitions (i.e. that we are more likely
to find a significant effect if we test more individuals, at least in cases where
there is a real effect), the actual numbers involved depart from the sample sizes
of typical studies in many areas of research. It is generally considered desirable
to aim for a power of 0.8 (shown by the horizontal dashed line). For a large effect
(of d = 0.8) this will require testing at least 26 individuals per group (e.g. 52 in
total in a between subjects design). For a medium effect (of d = 0.5) we would
need to test 64 individuals per group. For a small effect (of d = 0.2), our sample
size would be almost 400 per group. Relatively few experimental studies have
sample sizes this large, yet most effects in the life sciences tend to be in the
small to medium range. We discuss the consequences of underpowered studies
5.5. PROBLEMS WITH LOW POWER 87

in the following section.

1 1

0.8 0.8

0.6 0.6
Power

Power
0.4 0.4

N = 20
d = 0.2
N = 40
0.2 d = 0.5 0.2
N = 80
d = 0.8

0 0
0 50 100 150 200 0 0.2 0.4 0.6 0.8 1

Sample size (N) Effect size (d)

Figure 5.3: Power curves for different combinations of sample size and effect size.
The criterion for statistical significance was fixed at 0.05 in all cases.

The right panel of Figure 5.3 shows analogous functions for fixed sample sizes
as a function of effect size. These echo the results of the left panel, showing for
example that a sample of 20 individuals (per group) can only detect effect sizes
of d > 0.9 with satisfactory power. Such large effect sizes are unusual in most
areas of life science research (indeed, many consider large effects to be trivial
and not worth investigating at all), yet many published studies across diverse
disciplines tend to have samples around this size. Similarly, effect sizes in the
small-to-medium range will always have low power with double-digit sample
sizes. A consequence of all of this is that many studies are underpowered.

5.5 Problems with low power

Why are low powered studies bad? First of all, an underpowered study is unlikely
to detect an effect even when one exists. This means returning a null result
that, within the framework of frequentist inferential statistics, is inconclusive
(see Chapter 17 for further discussion of this point). So, we spend time and
resources collecting data, but get back a result that doesn’t tell us much. This
is disappointing, but it often interacts with a second issue, publication bias, to
produce even more serious problems.
Publication bias is the general tendency for significant results to be more likely to
be published than non-significant ones. This is a systemic issue with the current
scientific publishing model, whereby journals are more likely to accept papers
reporting significant results (which are seen as more ‘interesting’). Researchers
88 CHAPTER 5. POWER ANALYSIS

are more likely to write up significant findings, with non-significant results

languishing in a notional ‘file drawer’ (hence the moniker, the file drawer effect).
Now, any study with α = 0.05 has a fixed false positive rate of 0.05, meaning
that one in twenty studies in which there is no real effect will incorrectly identify
significant effects (known as a Type I error). If our power is low and we observe
a positive result, the likelihood that it is a false one increases dramatically. In
the extreme case where there is no true effect (when d=0), any ‘significant’ effect
we find is by definition a false positive (see right panel of Figure 5.3 - at the
far left, the curves converge at a power of 0.05). If these results are more likely
to get published than true null findings, the literature will quickly fill up with
spurious effects.
There is substantial evidence that this has been happening for many years. For
example, an attempt to replicate many preclinical cancer biology results found
that only 11% of findings were still significant upon replication (Begley and Ellis
2012). In neuroscience, most studies have surprisingly low power, averaging
around 0.3 (Button et al. 2013). Recent attempts to replicate large numbers
of studies in experimental psychology (Open Science Collaboration 2015) have
found that around 65% of reported effects cannot be replicated, even when using
much larger sample sizes than the original studies. This situation is termed
the replication crisis; because spurious effects tend not to replicate, and many
reported effects do not replicate, it is very likely that many reported effects are
spurious. One solution to these issues is to increase statistical power, and so
power analysis has become increasingly important, and is now often required
when seeking ethical approval and grant funding.

5.6 Problems with high power

The above discussion might lead us to conclude that the more power we have the
better. But there can be issues with high statistical power as well. One potential
problem of overpowered studies is that many effects will become significant, even
if they have trivially small effect sizes. A very small effect size might have no
practical or theoretical implications, and yet with a huge sample size it will still
be statistically significant. This point is related to the idea of the smallest effect
size of interest - any smaller effect sizes may simply not be interesting to anyone.
Additionally, for complex factorial designs it can be rather tedious to write up
an overpowered study, as the expectation is that every significant effect will need
to be commented on and discussed (for a humorous take on this issue, with a
more serious Appendix, see Friston 2012).
Historically, overpowered studies were not something that many researchers
needed to worry about. But in the era of big data, it is quite possible to amass
enormous data sets, perhaps over the internet, or through national testing
programs such as the UK Biobank (see https://www.ukbiobank.ac.uk/). When
data sets involve tens of thousands of individuals, standard frequentist inferential
5.7. MEASUREMENT PRECISION IMPACTS POWER 89

tests can become essentially meaningless, as almost everything is significant. In

such cases, standardised effect sizes become a key statistic for understanding the
structure of the data, and modelling approaches that are not primarily concerned
with null hypothesis significance testing may be more appropriate (see Chapters
9 and 12 for examples).

5.7 Measurement precision impacts power

Most work on power analysis implicitly assumes that effect sizes are fixed.
However, experimenters often have some control over effect size via the precision
of their measurement of the dependent variable. When the measurements we
make are noisy (as is typically the case in most experiments), it is intuitively
understood that running more repetitions (trials) for each individual delivers
‘better’ data (i.e. data that are less noisy, and more likely to produce statistically
significant results). Formally, the effect of running more trials is to reduce
the contribution of within-participant variance to the overall sample standard
devation. The sample standard deviation is defined as:
r
σw2
σs = σb2 + , (5.2)
k

where σw and σb are the within and between participant standard deviations,
and k is the number of trials (D. Baker et al. 2021). Because the sample standard
deviation appears on the denominator of the effect size equation for Cohen’s d
(equation (5.1)), running more trials on each individual increases effect size, and
therefore drives up statistical power. Of course this is only meaningful when
the within-participant variance is high compared to the between-participants
variance, but this appears to be the case for many experimental paradigms in
psychology and neuroscience, and by extension other areas of the life sciences.
A useful way to assess the joint impact of trial number (k) and sample size (N)
on statistical power is to produce a two-dimensional contour plot, as shown in
Figure 5.4. Each curve illustrates the combinations of N and k that lead to a
given level of statistical power. Researchers can therefore optimise their study
design by trading off these factors - if individual participants are hard to recruit,
each one could be tested for longer. If individual participants are plentiful, each
one could be tested for less time. A Shiny application to generate power contours
is available at: https://shiny.york.ac.uk/powercontours

5.8 Reporting the results of a power analysis

When reporting the results of a power analysis, it is important to include
the values you have used to perform the calculations, and also to give some
justification or rationale for why you have chosen those specific numbers. For
example, imagine we were applying for ethical approval to compare a new drug
90 CHAPTER 5. POWER ANALYSIS

Mean difference = 1; Within SD = 10; Between SD = 2

100
Power
10%
90
20%
30%
40%
80
50%
60%
Trials per participant (k)

70 70%
80%
90%
60 100%

50

40

30

20

10

0
0 20 40 60 80 100 120 140 160 180 200
Sample size (N)

Figure 5.4: Power contour plot. Curves show combinations of N and k that
give a constant level of statistical power. This example assumes a true group
mean of 1, within-participant standard deviation of 10, and between participants
standard deviation of 2.
5.9. POST-HOC POWER ANALYSIS 91

treatment with a placebo in a group of mice, and need to conduct a power

analysis to decide how many subjects to test. The literature on similar drugs
reports a range of effect sizes from d = 0.2 to d = 0.52. We might decide that
we should power our study to detect the largest effect that is likely to be of
interest, because we only care if our drug is better than those already available.
So we’ll go with the upper bound here (d = 0.52). The typical α level in this
field is likely to be 0.05, so we would probably just stick with this (though it is
a good idea to think more deeply about our choice of criterion, see Lakens et al.
2018). Finally, we choose the standard power of 0.8 as our target. The power
calculations tell us we need 60 subjects in each of the control and experimental
groups. We might report these results as follows:
“We conducted a power analysis based on data from previous studies.
Our main objective is determining whether this new treatment is
superior to the best-in-class treatment, which has an effect size of
d = 0.52. Power calculations with α = 0.05 indicate that a sample
size of 60 mice per group are required to achieve 80% power for this
effect size. Our total sample size is therefore 120.”
For a second example, let’s suppose we wished to replicate a reaction time
experiment from the literature (for examples of replication studies on a large
scale, see Open Science Collaboration 2015). In the original study, participants
made responses in two conditions, and there was a significant difference between
the conditions using a repeated measures design: participants were 30 ms faster
to respond in condition A than in condition B. The original study reported a
significant effect size of d = 0.44, with 95% confidence intervals ranging from
0.38 to 0.51. We wish to power our replication study to be able to detect effects
at the lower end of this range, so we use d = 0.38 for power calculations, along
with a power of 0.9 and α = 0.05. Because we have an a priori expectation
about the direction of the effect, it is reasonable to use a one-sided statistical
test. The power calculations tell us that 61 participants would be required for
this replication. Note that the design and assumptions are important here -
using a two sided test and a between participants design would require almost
five times as many participants! We might report the power analysis like this:
“To design a well-powered replication of the experiment reported by
Roy and Kim (2018), we conducted a power analysis. The lower
bound effect size from the original study was d = 0.38. We planned
to detect this using a repeated measures design, assuming a one-sided
test in the direction of the original effect. For a criterion of α =
0.05, this will require N = 61 participants to reach 90% power.”

5.9 Post-hoc power analysis

So far, we have talked about trying to determine how many individuals we should
test in order to achieve a target level of power. Another use of power analysis is
92 CHAPTER 5. POWER ANALYSIS

to calculate the power of a study that has already been conducted. This is known
as post-hoc power analysis, or observed power, and is available as an option in
commercial statistics packages such as SPSS. Indeed, this is the method that
has been used to estimate the level of power in particular areas of the literature.
It is also sometimes used to interpret null results - determining whether an effect
was likely to be non-significant because a study was underpowered. However,
there are some theoretical concerns with interpreting observed power (Hoenig
and Heisey 2001). Most of these centre around the fact that observed power is
inversely proportional to the p-value (with low p-values equating to high power).
In other words, a non-significant result is likely to have low power because it
is non-significant. This means that calculating observed power provides no
additional information beyond a properly reported p-value and is therefore
misleading. A more fruitful approach to interpreting null results is offered by
Bayesian statistics, as discussed in Chapter 17.

5.10 Doing power analysis in R

Power analysis can be conducted using a number of online tools, as well as
specialist software such as the free G*Power application (Faul et al. 2007, 2009).
There is also an R package called pwr that we will use for the examples here. It
contains dedicated functions for different types of statistical test. They all have
a similar structure: we specify all the values apart from the one we want the
function to calculate, and it works out the missing value. If we want to know
the power, we specify the effect size, the sample size and the α level. If we want
to know the sample size, we specify the power, the effect size and the α level,
and so on. You can install the pwr package either using the Packages tab in
RStudio (see section 2.8), or by typing install.packages(‘pwr’) in the console.
We will begin with the function for a t-test. Let’s conduct a power analysis to
estimate the required sample size to detect an effect of d = 0.5 with power of
0.8 and α = 0.05. Note that we do not specify the sample size, as this is what
we want the function to calculate for us.
library(pwr)
pwr.t.test(d=0.5, power=0.8, sig.level=0.05)

##
## Two-sample t test power calculation
##
## n = 63.76561
## d = 0.5
## sig.level = 0.05
## power = 0.8
## alternative = two.sided
##
## NOTE: n is number in *each* group
5.10. DOING POWER ANALYSIS IN R 93

The output returns all of the values we have just entered, and also tells us that
we will require a sample size of N = 63.77. Of course it is not practical to
test 0.77 of a participant, so we always round up to the nearest whole number.
Therefore a sample size of N = 64 per group is required (so N = 128 in total).
Variants for one-sample and paired t-tests are also available, as detailed in the
help.
A similar function can conduct power analysis for correlations. Let’s say we
want to know the smallest correlation coefficient that can be detected with a
power of 0.8 and a sample size of 30 participants.
pwr.r.test(n=30, power=0.8, sig.level=0.05)

##
## approximate correlation power calculation (arctangh transformation)
##
## n = 30
## r = 0.4866474
## sig.level = 0.05
## power = 0.8
## alternative = two.sided
This output tells us that a correlation coefficient of r = 0.49 or larger can be
detected with 80% power.
Next let’s look at power calculations for a one-way ANOVA. Here we use a
different measure of effect size, called f (also known as Cohen’s f ). Note that
importantly, this is very different from the F-ratio usually reported in an ANOVA
summary table. The effect size f is closely related to d, such that f = d2 . It is
calculated by taking the standard deviation across the population (i.e. group)
means and dividing it by their pooled standard deviation (i.e. across participants).
We also need to know how many groups there are in the study design (the input
k). Let’s calculate the power of a study design with an effect size of f = 0.1, and
N = 30 participants in each of five groups:
pwr.anova.test(k=5, n=30, f=0.1, sig.level=0.05)

##
## Balanced one-way analysis of variance power calculation
##
## k = 5
## n = 30
## f = 0.1
## sig.level = 0.05
## power = 0.1342476
##
## NOTE: n is number in each group
This design has a very low power indeed, only 0.13.
94 CHAPTER 5. POWER ANALYSIS

We can use these functions to produce power curves, such as those shown in
Figure 5.3 by entering a range of effect sizes or sample sizes in a loop (see
section 2.10). For example, we can produce a very instructive power curve for
correlations as follows:
N <- 4:100
r <- NULL
for (n in 1:length(N)){
output <- pwr.r.test(n=N[n], power=0.8, sig.level=0.05)
r[n] <- output$r
}

plot(N,r,type='l',lwd=3,xlim=c(0,100),ylim=c(0,1))
1.0
0.8
0.6
R

0.4
0.2
0.0

0 20 40 60 80 100

Figure 5.5: Curve showing the minimum correlation coefficient that can be
detected at 80% power, as a function of sample size.

This curve (plotted in Figure 5.5) shows the minimum r value that can be
detected with 80% power at a range of sample sizes. Even studies with N = 100
cannot reliably detect small correlations where r < 0.25.
Power calculations for more complex and sophisticated designs, or those using
statistical techniques not covered by the pwr package, are best done by simulation.
An excellent introduction to power analysis by simulation is given by Colegrave
and Ruxton (2020), and we will discuss some of the stochastic methods required
for this in Chapter 8.
5.11. PRACTICE QUESTIONS 95

5.11 Practice questions

1. If we ran 10000 simulations of an experiment, and 2000 were significant,
what would the statistical power be?
A) 0.8
B) 0.2
C) 0.02
D) 0.08
2. Calculate Cohen’s d for group means of 15 and 12, with a standard deviation
of 20.
A) 0.75
B) 0.6
C) 0.15
D) 0.42
3. As the sample size increases, what happens to effect size estimates?
A) They increase
B) They decrease
C) They become more accurate
D) It depends on the magnitude of the effect
4. A low powered study:
A) Is very likely to produce a significant effect
B) Has a higher false positive rate than a high powered study
C) Typically has a large sample size
D) Is very unlikely to produce a significant effect
5. Effect sizes are often inflated in published studies with low power because:
A) For a result to be significant, it must have a large effect size
B) Low power reduces the standard deviation
C) Low power increases the standard deviation
D) Replications are usually more accurate
6. How many participants would be required to reach 80% power for a one-
sample t-test with an effect size of 0.8, assuming an alpha level of 0.05?
A) 20
B) 25
C) 26
D) 15
7. What is the power of a study testing 24 participants to detect a correlation
of r = 0.3?
A) 0.30
B) 0.29
C) 0.23
D) 0.26
8. What is the observed power of a one-way ANOVA with an effect size of
f =0.33, 8 groups, and 30 participants per group, assuming an alpha level
of 0.01?
A) 0.99
B) 0.91
96 CHAPTER 5. POWER ANALYSIS

C) 0.98
D) 1.00
9. What is the smallest effect size (w) that can be detected using a Chi-
squared test with 12 participants, 10 degrees of freedom and a power of
0.5 (assume an alpha level of 0.05)?
A) 1.16
B) 0.91
C) 0.88
D) 0.99
10. The function pwr.f2.test calculates power for factorial ANOVAs using the
general linear model. Assuming numerator and denominator degrees of
freedom of 2 and 12, what is the smallest effect size that can be detected
with 80% power and alpha of 0.05?
A) 0.83
B) 24.4
C) 0.69
D) 0.60
Answers to all questions are provided in section 20.2.
Chapter 6

Meta analysis

Meta analysis is a method for combining the results of several studies computa-
tionally. Usually, it is some measure of effect size (see section 5.2) that we choose
to combine, such as Cohen’s d, or the r value from a correlation. Of course,
the simplest way to do this is just to average the effect sizes from a bunch of
studies that all measure the same thing. But often there are differences in study
design, sample size, and other features, that make a straightforward average
inappropriate. Imagine combining three studies. Two of them are high quality,
testing hundreds of participants using state-of-the-art methods. The other is a
rather shoddy affair that should probably never have been published in the first
place. It would hardly seem fair to give them all equal weight in our calculations.
The tools of meta analysis allow us to take factors such as this into account.
The main outcome of a meta analysis is an aggregate effect size estimate, which
is used to determine whether, on the balance of evidence, a real effect exists.

6.1 Why is meta analysis important?

The scientific literature is huge, with over a million new articles published every
year (e.g. in 2020, PubMed lists 1.6 million publications, just in the journals they
index). It is often the case that studies aim to replicate the key findings from
previous work, and that many researchers tackle the same problem using similar
methods. These individual studies do not always come to the same conclusions,
and are individually limited by constraints on sample size and other aspects of
the methodology. By combining results across multiple studies, we can increase
the overall statistical power of our observations, and hopefully come up with a
more reliable answer to important questions.
The importance of meta analysis first became apparent in the early 1990s. The
classic example is the use of corticosteroids to treat complications arising from
premature birth. During the 70s and 80s, several studies were published on the

97
98 CHAPTER 6. META ANALYSIS

topic, but the evidence from reading them individually appeared mixed. For this
reason, corticosteroids were not routinely prescribed in cases of premature birth.
In 1990 a meta analysis was published (Crowley, Chalmers, and Keirse 1990)
that showed a clear benefit of the drugs (a reduction in mortality of 30-50%),
and their use became mainstream clinical practice.
There are two ways of interpreting this story. On the one hand, many thousands
of babies suffered and died unnecessarily during the years when the evidence
supporting corticosteroid use was available but had not been synthesised together.
On the other hand, over the past three decades, many thousands of babies have
been treated using this method, and many lives have been saved. Either way, the
importance of meta analysis is clear - unambiguous answers to medical questions
can save lives.
The corticosteroid example led to the creation in 1993 of the Cochrane Collab-
oration, an international charity organisation dedicated to coordinating meta
analyses on a range of topics. These are freely available in the Cochrane Library
(https://www.cochranelibrary.com/). Most of the Cochrane meta analyses are on
medical topics, including a substantial number on mental health and psychiatric
conditions. They do not focus only on medications - many analyses are concerned
with dietary and lifestyle factors, and other therapeutic techniques. The logo of
the Cochrane Collaboration is a stylised version of the corticosteroid data.
In addition to medical reviews, the tools of meta analysis can be applied to
other topics, including basic experimental laboratory science. These might be
less obviously life saving, but they have become increasingly important in recent
years for establishing whether reported effects are robust. This has led to some
interesting conclusions about entire subfields of research, and is an important
aspect of the replication crisis being widely discussed in many fields. Overall, a
meta analysis should represent the strongest form of evidence on a particular
topic, as it synthesises all of the available data in a systematic and quantitative
way.

6.2 Designing a meta analysis

The stages involved in conducting a meta analysis (particularly on clinical topics)
can be quite prescribed. For example the Cochrane Handbook (Higgins and
Green 2011) is a substantial official guide that describes the process in detail. I
will not attempt to replicate too much of this content here. However, the key
stages are:
1. Decide the scope of the analysis - what are you interested in?
2. Decide on search terms and inclusion/exclusion criteria
3. Search online databases using the key terms and build a database of articles
4. Sift through the articles and apply the inclusion/exclusion criteria
5. Extract an effect size from each included study
6. Combine the effect sizes and produce a summary plot
6.3. CONDUCTING AND SUMMARISING A LITERATURE SEARCH 99

The literature search (stages 1-4) can be succinctly summarised using a PRISMA
diagram, which we will introduce in the next section. The remainder of the
chapter will mostly focus on the final two stages, as these comprise the numerical
and computational parts of the process.

6.3 Conducting and summarising a literature

search
A meta-analysis is a quantitative analysis that results from a systematic review
of the literature. It is therefore important to be very clear about the scope of the
underlying review. We might be interested in a particular disease, or a particular
experimental paradigm, for example. We implement these decisions by deciding
on a set of search terms that we will use when searching online databases, and
also some inclusion and exclusion criteria. Let’s imagine we want to conduct a
meta-analysis looking at the effectiveness of chlorophyllin (chlorophyll dissolved
in water) for treating fish parasites. Our search terms might be chlorophyllin
and fish parasites, and we might decide only to include experimental studies
that look at the effect on parasitic algae - studies looking at crustaceans and
protozoa would be excluded. It is advisable to preregister the protocol for your
meta analysis online, for example using the OSF website (https://osf.io/), or
the AsPredicted website (https://aspredicted.org/). This gives your readers
confidence that you are sticking to your original intentions, and not changing
your inclusion criteria or other protocols based on the results.

When conducting the literature search, it is important to be systematic and

record all decisions about including or excluding a study. A useful tool to guide
this process is the PRISMA diagram. This is a flow chart detailing how many
studies were involved at each stage of the literature search. An example PRISMA
diagram is shown in Figure 6.1, and it is good practice to include one in the
writeup of any meta-analysis. PRISMA stands for the Preferred Reporting Items
for Systematic Reviews and Meta-Analyses, which are a set of standard protocols
that have been widely adopted. They include the diagram, as well as a 27-item
statement described by Moher et al. (2009). Templates are available from
the PRISMA website (http://www.prisma-statement.org/), and the diagram
can also be created using several R packages including PRISMAstatement and
prismadiagramR.

The PRISMA diagram is helpful because it proposes a structure for conduct-

ing a meta-analysis. We would start by searching online databases, such as
PubMed, Google Scholar, Web of Science, and other subject-specific databases
(e.g. PsychInfo, ArXiv, CINAHL). Deciding key terms for literature searches is
quite idiosyncratic, and so will differ widely depending on the topic (see Field
and Gillett (2010) for some suggestions). Occasionally, there will also be relevant
studies that these searches do not return, but which you already know about -
these can be included as ‘other sources’. One example might be very recently
100 CHAPTER 6. META ANALYSIS

Records identifed through Additional records identifed

database searching through other sources
(n = 401) (n = 13)

Records after duplicates removed

(n = 412)

Records screened Records excluded

(n = 412) (n = 239)

Full-text articles assessed Full-text articles excluded,

for eligibility with reasons
(n = 173) (n = 108)

Studies included in qualitative synthesis

(n = 65)

Studies included in
quantitative synthesis
(meta-analysis)
(n = 65)

Figure 6.1: Example PRISMA diagram, reporting the number of studies included
and excluded at each stage of a literature review.
6.4. DIFFERENT MEASURES OF EFFECT SIZE 101

published work, including preprints, that is yet to be indexed. Another example

might be if you contact researchers in the field to ask about unpublished data
that you subsequently include - this is a good way to help reduce the ‘file-drawer’
effect that can cause publication bias (see section 5.5).
The various sources of information will usually have overlapping content, so
it is important to develop a principled method for removing duplicate records.
Ideally this can be automated using unique identifiers such as DOI (digital object
identifier) numbers or PubMed IDs. Where this is not possible, identifying
duplicates using titles, author names and journal page or issue numbers might
be necessary. For very large literatures there will be some benefit to automating
this process, and in general, scripting analyses is preferred as it is reproducible,
and less prone to data entry errors. However for a smaller corpus of studies it
might be better done manually, perhaps using a spreadsheet or some reference
manager software.
The next stage is to screen all non-duplicate records, usually by reading the brief
description of each study given in the abstract. There will be many records that
are clearly not within the remit of the analysis, and these should be excluded at
this stage without needing to inspect the full text. Good reasons for exclusion
might be studies that use a completely different method, species or paradigm,
or that are review articles rather than primary research. All of the remaining
records that pass this screening process should then be obtained and inspected.
At this stage it might again become clear that some studies are not within
the remit of the meta-analysis. However this time a note should be made of
the reasons for exclusion. These reasons need not be lengthy or particularly
detailed, but they should reflect the inclusion criteria you originally specified
before starting to look through the literature.
The results should then be extracted from each of the remaining studies. Normally
effect sizes are used in meta-analysis to summarise the findings of each study, and
we will discuss these in the next section. It is always worth keeping detailed notes
of the above process. This will help anyone trying to reproduce the decisions that
were taken during the analysis, including your future self if you have forgotten
what you did! In particular, you should keep a record of the numbers of studies
involved at each stage, which you can then use to generate the PRISMA diagram.

6.4 Different measures of effect size

As mentioned in section 5.2, there are several types of effect size. Many studies
use ordinal scales or continuous data. Examples of ordinal data include rating
scales and questionnaire data - these can take on fixed values, but the order
is meaningful (e.g. a pain rating of 7 means more pain than a rating of 2).
Examples of continuous data include physical properties of the world or of an
organism; things like heart rate, temperature, brain activity, reaction time and
so on. These can take on any value within some reasonable range. These types
102 CHAPTER 6. META ANALYSIS

of data are well described by effect size measures based on differences in means
(such as Cohen’s d), or those indicating the proportion of the overall variance
explained by some predictor (e.g. the correlation coefficient r, and the ANOVA
effect size measure η 2 ).

In much of the clinical literature, other types of effect size are common, which
you will come across if you read materials on meta analysis. These are based on
the concepts of risk and odds, which are important ideas to know about. They
are used most often for dichotomous (binary) data, which have obvious relevance
in medicine - is the patient dead or alive; are they infected or cured? In fact any
type of data can be arbitrarily made dichotomous, for example by deciding on a
criterion or cut off. For example, continuous measurements of blood pressure
can be categorised into high and low blood pressure groups by choosing some
threshold (currently 120/80 for stage 1 hypertension). So the risk and the odds
can in principle be calculated for any type of data, though this should only be
done when it is a theoretically meaningful thing to do.

The risk is defined as the number of events divided by the sample size. This is a
familiar concept - if one out of every thousand people get a particular disease,
the risk is 1/1000 or 0.001. The odds is very closely related, but subtly different.
It is the number of events divided by the number of non-events. So, in the
example of one in a thousand people, the odds would be 1/999, which will be
very similar indeed to the risk. However the numbers start to diverge as events
become more common. Consider a condition that affects half of a sample of 100
people: the risk will be 50/100 = 0.5, but the odds will be 50/50 = 1. A risk
score can never exceed 1, but an odds score can take on any positive number.
Figure 6.2 shows how the risk and odds diverge as events become more common.

Although raw risk and odds scores are sometimes clinically meaningful, in clinical
trials it is more common to report the risk ratio or the odds ratio. These are the
ratios of risk or odds values comparing a treatment group and a control group
(e.g. the risk for the treatment group, divided by the risk for the control group).
They will tell you, for example, how much a treatment or drug changes your risk
of some outcome, such as recovering from a disease. Because these are ratios of
event counts, they will always be positive numbers, and a value of 1 will always
mean there is no difference between the treatment and control groups. However,
whether values above or below 1 indicate a positive outcome will depend on
exactly what is being measured. An odds ratio of 3 might be good news if it
means a drug makes you more likely to recover from an illness, but very bad
news if it makes you more likely to have a heart attack!

Lastly, you will also see that some studies report the log odds ratio. This is just
a log transform of the odds ratio, which is a sensible thing to do given the range
of possible odds ratios. After the log transform, odds ratios >1 will have positive
values, and odds ratios <1 will have negative values. Ratios in general are often
more appropriately represented in log units, and thinking ‘logarithmically’ is
something that gets easier with practice.
6.5. CONVERTING BETWEEN EFFECT SIZES 103

10 2
Risk
Odds

Log10(Risk or odds)
8
1
Risk or odds

6
0
4

−1
2

0 −2

0 20 40 60 80 100 0.0 0.5 1.0 1.5 2.0

Number of events Log10(Events)

Figure 6.2: Comparison of odds and risk scores for events of different probabilities,
assuming a total sample of 100. The right hand panel shows the log transform
of the same values.

6.5 Converting between effect sizes

To conduct our meta analysis, we need the same type of effect size from each
study we include. But not all studies report the same measures of effect size,
and some report no effect size measures at all. For this reason, we need tools to
calculate effect sizes, and to convert between them. We have already encountered
(in section 5.2) the equation for Cohen’s d (Cohen 1988), which is the difference
in means scaled by the standard deviation:

x̄1 − x̄2
d= (6.1)
σ

where x̄1 and x̄2 are the group means, and σ is the pooled standard deviation.
If these values are reported in a source paper, we can use them to calculate d. If
the means and standard deviation are not available, we can also convert from a
correlation coefficient (r):

2r
d= √ , (6.2)
1 − r2

or from a t-statistic (where df is the degrees of freedom):

2t
d= √ (6.3)
df
104 CHAPTER 6. META ANALYSIS

In section 6.11 we will discuss how to convert between effect sizes and com-
mon statistics using R. However tools also exist online to perform these cal-
culations, for example a number of tools are available at the website: http:
//www.psychometrica.de/effect_size.html (Lenhard and Lenhard 2016).

6.6 Fixed and random effects

There are two main types of meta analysis, that involve slightly different as-
sumptions and calculations. In a fixed effects meta analysis, the underlying effect
size is assumed to be constant across all studies included. This is particularly
appropriate for things like clinical trials, where the dosage of a drug, and the
outcome measure, might be the same across all studies. The alternative is a
random effects meta analysis. This has provision for the underlying effect size
to vary across studies, and so is more flexible, and often more applicable to
lab-based research topics that are less standardised. It is usually clear which
type of meta analysis is appropriate for a given topic, though this can become
more apparent as the data are collected. If in doubt, the random effects design
is a safer bet, as it has fewer assumptions.

6.7 Forest plots

Once we have collated the effect sizes from all of the studies we are including in
a meta analysis, it is typical to represent them using a forest plot. This graph
plots the effect size along the x-axis, and the authors of each study (usually in
chronological or alphabetical order) along the y-axis. Figure 6.3 shows a simple
example, using a built in data set from the rmeta package in R (these are the
corticosteroid data that form the Cochrane Collaboration logo). Forest plots
can become much more elaborate than this, for example, I recently published a
meta analysis (Baker et al. 2018) involving 65 studies, for which the forest plot
was much larger.
Forest plots have several additional features. The vertical dashed line at an odds
ratio of 1 is known as the line of no effect (recall that this means equal odds in
the treatment and control groups). If we were using other types of effect size
measure, such as d or r, this would be at x = 0. The individual studies can then
be compared to this value. Each individual study is represented by a rectangle.
The size of the rectangle is typically proportional to the sample size in the study,
such that larger sample sizes produce larger rectangles. We can therefore place
more confidence in studies represented by larger rectangles, as they should give
us a more reliable estimate of the true effect size.
Each rectangle has an accompanying error bar. These are traditionally 95%
confidence intervals. Confidence intervals are useful because they work a bit like
a t-test. If the confidence intervals overlap the line of no effect for a particular
study, the chances are it is not significant. Also, if the confidence intervals
6.7. FOREST PLOTS 105

Auckland
Block
Study Reference

Doran
Gamsu
Morrison
Papageorgiou
Tauesch

Summary

0.02 0.06 0.16 0.40 1.00 2.51

Odds Ratio

Figure 6.3: Example forest plot of corticosteroid data from Crowley, Chalmers,
and Keirse (1990), available as part of the rmeta package.
106 CHAPTER 6. META ANALYSIS

from one study do not overlap the mean effect size for another study, the
two studies can be considered to have different effect sizes. If our data meet
parametric assumptions, the confidence intervals are usually calculated using the
approximation 1.96*SE, though they can also be derived by bootstrap resampling
(see Chapter 8) if the original data are available.
At the foot of the plot is the summary effect - this is the grand average effect
size across all studies. It is traditionally represented by a diamond. The middle
of the diamond corresponds to the mean, and the left and right corners are the
95% confidence intervals. The effect is deemed to be significant if the error bars
do not overlap the line of no effect, as is clearly the case for the example in
Figure 6.3. As mentioned before, the grand average effect size is not simply the
arithmetic mean of the individual studies. To understand how it is calculated,
we need to discuss the concept of weighted averaging.

6.8 Weighted averaging

To average three numbers, we add them up and divide by how many numbers
we have, for example:

x1 + x2 + x3 2+7+3
mean = = =4
3 3

Another way of thinking about this is to assume that each number has a weight of
1, which it is multiplied by, with the denominator being the sum of the weights:

ω1 × x1 + ω2 × x2 + ω3 × x3 1×2+1×7+1×3
wmean = = =4
ω1 + ω2 + ω3 1+1+1

In this example, because the weights are all set to 1, the end result is the same.
But if we weight some values differently from others it will change the outcome.
For example, if we assign the second number a higher weight it will bring the
average up (because it is a big number than the others):

1×2+3×7+1×3
wmean = = 5.2
1+3+1

Note that the weight appears twice: to multiply the value on the numerator,
and also as part of the sum of the weights on the denominator. In general terms
a weighted average is defined as:

Σ(ω1:i × x1:i )
wmean =
Σω1:i
6.9. PUBLICATION BIAS AND FUNNEL PLOTS 107

where ωi is the ith weight, and xi is the ith value in the list of numbers we wish
to average. This is the procedure typically used to calculate the grand mean
effect size. But what are the weights?
In meta analysis, we want to use weights that give an indication of the quality
of each study. One very simple way to do this is to use the sample size as the
weights - a study testing ten participants would have a weight of 10, and a
study testing 100 participants would have a weight of 100. This treats each
study as though it were part of a single monolithic study (i.e. it is a fixed effects
approach). Other alternatives are to use uniform weights (e.g. to give each
study a weight of 1), or to choose some predetermined criteria based on the
methodology used. For example, one might decide to weight studies using a
state-of-the-art recording device more highly than those using older technology
(for example in neuroscience, MEG has lower recording noise than EEG; in
genetics, PCR is better than older methods like RFLP).
These options are reasonable and defensible in some situations, but they are not
what is typically done in meta analysis. Instead, the weights are derived from
the variance for each study. Specifically, we use the inverse variance, 1/σ 2 . This
will be a large value for studies with small variance (i.e. very reliable studies),
and small for studies with large variances (i.e. unreliable studies). Note that the
σ term represents the standard deviation of the sampling distribution, which is
the sample standard error. As such, the sample size contributes to the inverse
variance weights (because the standard error calculation includes the sample
size).
As well as calculating the weighted average of the effect sizes, we also need to
derive confidence intervals (to tell us where the corners of the diamond should
go). One way to do this is to calculate the variance of the weighted average using
the squares of the weights to combine the variances. Another option is to use
stochastic simulations (see Chapter 8) to estimate the variance (Sánchez-Meca
and Marin-Martinez 2008). Fortunately these rather complex calculations are
done automatically in meta analysis software, so we will not consider them
further here.

6.9 Publication bias and funnel plots

Once the forest plot has been created, we can conclude whether there is an overall
effect. But there is a problem here: the results of the meta analysis depend
entirely on the studies we include, and we know (from biases in the scientific
publication process) that statistically significant studies are more likely to be
published (see section 5.5). So does this mean that meta analysis is pointless,
because publication bias means we are almost guaranteed to get a positive result,
given that the majority of our source studies will report significant effects?
Fortunately, meta analysis provides a useful tool to assess the likely impact of
publication bias - the funnel plot. Like the forest plot, the funnel plot represents
108 CHAPTER 6. META ANALYSIS

effect size on the x-axis. But this time, the studies are ordered along the y-axis,
usually according to either sample size or inverse variance. Studies with large
samples appear towards the top, and studies with small samples appear near the
bottom. An ideal funnel plot looks like the example in the left panel of Figure
6.4.

100 ●
●
●
●
●
●
100 ●
●
●
●

●
● ●
●
● ●
● ●
●
● ●
●
●
●
● ●
80 ●
●
80 ●
Sample size (N)

Sample size (N)

●●
● ● ● ●●
●
● ●
● ● ●
●
● ●
●●
● ●
● ●
● ● ● ●
● ● ●
● ●
● ● ●
● ●
● ●
● ● ● ●
● ● ●
60 ●
●
●
●
●
● 60 ●
●
●
●
● ●●
●
●
●

● ●
● ●●
● ●●
● ●
● ● ●●
● ● ● ●
● ● ● ● ●
● ● ● ● ●●
●
●● ●
● ●●
●
● ●● ●
● ● ● ●
● ●
● ●● ●●
40 ●
●
●● ●
●
●
●
● ●
●
●
40 ●
● ●●

●●● ● ●●
●●● ●●
●
● ●●
●●
●
● ● ● ●
● ● ●
● ●
● ● ● ●● ●
●
●● ● ● ● ●
● ●● ●● ●
●● ● ●● ●● ●
● ● ● ●●
● ● ●●
●● ● ●● ●●
● ● ● ● ● ● ● ● ●● ●
● ●● ● ●
● ● ● ● ●●
20 ●
● ●●
● ●
●
●

●
●
●
●
●
●
●
● 20 ●

●●●
● ●●

●
●
● ●
● ●
●
●

● ●
●
●
●● ● ● ● ● ●● ●
●● ● ● ● ● ● ●
● ● ● ●
● ● ●● ● ● ● ●● ● ●
●
● ● ●
● ● ● ● ●●●● ●●
● ● ● ● ●
● ● ● ●
● ● ● ● ●● ● ●● ● ● ● ●
● ● ● ●● ● ● ●
● ● ● ● ● ● ● ● ●
● ● ● ●● ● ● ● ●
●

0 0

0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Effect size (d) Effect size (d)

Figure 6.4: Example funnel plots, in which each point represents a simulated
study. In the left panel, a symmetrical funnel (triangle) shape is apparent, with
large-sample studies producing estimates close to the true effect size (top) and
small-sample studies producing more variable estimates (bottom). In the right
plot, studies with estimates below the true mean are suppressed (i.e. remain
unpublished). The plot becomes asymmetric, and the mean effect size (dashed
line) is overestimated.

The symmetrical funnel plot gets its triangular shape because studies (points)
with large sample sizes (at the top) are more likely to produce estimates of effect
size close to the true mean (solid line), whereas studies with small sample sizes
(at the bottom) will produce more variable effect sizes. If a funnel plot looks
like this, it is unlikely that publication bias is a big problem for the area under
study.
Now let’s think about what would happen if studies that were non-significant
did not get published. This might happen for nefarious reasons (such as a
pharmaceutical company deliberately suppressing a study that shows their drug
is ineffective), but it is also likely just as a consequence of human nature, and
the current incentive system in scientific publication. Non-significant results are
much harder to get published, as many journals will simply reject them out of
hand as being ‘uninteresting’. Furthermore, most researchers have limited time,
and will often prioritise publishing studies with significant results, which might
6.10. SOME EXAMPLE META ANALYSES 109

be more likely to get published in prestigious journals, and so be better for their
career.
The right hand panel of Figure 6.4 shows an asymmetrical funnel plot, in which
all studies with effects below the true mean of d=1 are omitted. The effects of
this are clearest in the small sample studies, which now skew out to the right.
Studies with a larger effect size were close to the true mean anyway, so these
look much the same as before. One consequence is that the mean effect size
across all the studies (shown by the dashed line) now overestimates the true
mean (solid line).
Funnel plots can be used to test for publication bias, and this is routinely done
as part of a meta analysis. In situations where publication bias is detected,
techniques exist to estimate what the true underlying effect size is likely to be.
One striking example is a meta analysis by Shanks et al. (2015) that looked at
priming studies of consumer choice. The funnel plot they produced (see their
Figure 2) was highly asymmetrical, whereas a funnel plot of replication studies
was symmetrical about an effect of d = 0. This is strong evidence for publication
bias, or other types of questionable research practise (such as p-hacking) in this
particular paradigm.

6.10 Some example meta analyses

To illustrate the diverse range of topics that are amenable to meta analysis, this
section will discuss three examples from the literature. The first is a study that
investigated whether sexual orientation is a risk factor for suicidal behaviour in
young people (Miranda-Mendizábal et al. 2017). The PRISMA diagram from
this study is quite remarkable, as it shows how an initial set of over 30,000
database records was reduced to only 7 that were suitable for inclusion in the
quantiative analysis. The odds ratios from these seven studies all exceeded
one, indicating an increased risk of suicidal behaviour in individuals identifying
as homosexual, though three of the original studies were non-significant. The
random effects meta analysis resulted in an overall odds ratio of 2.26 - this
means that lesbian, gay and bisexual adolescents have more than twice the
risk of suicidal behaviours compared with their heterosexual counterparts. A
funnel plot showed no evidence of publication bias in the seven included studies,
which was supported by a non-significant Harbord test (a quantitative test of
asymmetry (Harbord, Egger, and Sterne 2006)). The study concludes that public
health strategies should be developed to explicitly support the LGB community,
given that the analysis indicates they are a high-risk group. (Note that the use
of the acronym LGB here reflects the identities of the participants included in
the study. This is not intended to exclude other groups (for example, from the
extended form of the acronym LGBTQIA+), however the extent to which the
results might generalise beyond the population studied is unclear).
Glasziou and Mackerras (1993) investigated the effect of vitamin A supplemen-
110 CHAPTER 6. META ANALYSIS

tation on the likelihood of dying from infectious diseases. They included only
controlled trials in their meta analysis, involving measles, respiratory diseases or
diarrhoea, in children in developing countries. They pooled odds ratios across
studies, with aggregate effects being calculated separately for different diseases,
and in community studies. The largest result was for measles: across three
studies, the average odds ratio was 0.34 (i.e. a reduced risk of death of 66%
following supplementation). A subset of five community studies also found a
mortality reduction of 30%. To assess the potential impact of publication bias,
the authors calculated a statistic called the ‘failsafe N’ (Rosenthal 1979) for the
community study result (with 5 studies). This statistic tells us the number of
non-significant studies that would need to exist (yet remain unpublished) for
there to be no effect overall. For this example it was 53, which is an implausibly
large number of studies to remain unpublished. Overall, these results suggest a
strong benefit of either vitamin A supplementation, or a well-balanced diet, in
reducing the risk of death from infectious disease.

An amazing online resource is the website Neurosynth. This uses an automated

meta analysis approach to neuroimaging data, combining brain activation across
thousands of studies (Yarkoni et al. 2011). One can choose a search term, such
as acoustic or emotion, and the system will show the brain regions most strongly
associated with that term in the fMRI literature. It also works the other way
around - if you choose a brain region, it will list studies reporting significant
activation there. Finally, there are many thousands of whole-brain maps showing
gene-expression, that link back to the term-based meta analyses. So, if you are
interested in a particular gene, you can generate hypotheses about the sorts
of cognitive functions it might affect. The Neurosynth website showcases the
enormous potential of expanding the basic concept of meta analysis to large
modern datasets, as well as automating much of the computation.

6.11 Calculating and converting effect sizes in

R
The compute.es package in R provides a set of 16 tools for calculating and
converting between effect sizes. All of these have a common output, so the only
thing that differs is the information you feed in. The simplest is the mes function,
which converts means to effect sizes:
mean1 <- 14
mean2 <- 12
sd1 <- 5
sd2 <- 6
n1 <- 100
n2 <- 100

mes(mean1, mean2, sd1, sd2, n1, n2)

6.11. CALCULATING AND CONVERTING EFFECT SIZES IN R 111

## Mean Differences ES:

##
## d [ 95 %CI] = 0.36 [ 0.08 , 0.64 ]
## var(d) = 0.02
## p-value(d) = 0.01
## U3(d) = 64.14 %
## CLES(d) = 60.11 %
## Cliff's Delta = 0.2
##
## g [ 95 %CI] = 0.36 [ 0.08 , 0.64 ]
## var(g) = 0.02
## p-value(g) = 0.01
## U3(g) = 64.09 %
## CLES(g) = 60.07 %
##
## Correlation ES:
##
## r [ 95 %CI] = 0.18 [ 0.04 , 0.31 ]
## var(r) = 0
## p-value(r) = 0.01
##
## z [ 95 %CI] = 0.18 [ 0.04 , 0.32 ]
## var(z) = 0.01
## p-value(z) = 0.01
##
## Odds Ratio ES:
##
## OR [ 95 %CI] = 1.93 [ 1.16 , 3.2 ]
## p-value(OR) = 0.01
##
## Log OR [ 95 %CI] = 0.66 [ 0.15 , 1.16 ]
## var(lOR) = 0.07
## p-value(Log OR) = 0.01
##
## Other:
##
## NNT = 8.64
## Total N = 200

For a single line of code, the output is very extensive. The idea is to provide all
the various measures of effect size you might need to use. These include Cohen’s
d in the first section (which has a value of 0.36), Hedge’s g in the second section,
the equivalent correlation coefficient (r) and z-score, odds ratios and number
needed to treat (NNT). Similar functions exist if you have a p-value, t-value, r
value and so on, for example:
112 CHAPTER 6. META ANALYSIS

pvalue <- 0.03

pes(pvalue, n1, n2)

tvalue <- 2.5

tes(tvalue, n1, n2)

rvalue <- 0.24

res(rvalue, n = n1)

I have hidden the output from the above functions, but it is in exactly the same
format as for the previous example. What we usually want though, is to just
extract the single effect size we are interested in. We can do this by assigning
the output of the function call to a data object as follows:
output <- mes(mean1, mean2, sd1, sd2, n1, n2)

The object called output then contains all of the numbers you might need in
fields with sensible names. For example, you can request Cohen’s d and its
variance as follows:
output$d

## [1] 0.36
output$var.d

## [1] 0.02
And we can convert the variance to a standard deviation by taking the square
root:
sqrt(output$var.d)

## [1] 0.1414214
The effect size, its standard deviation and the sample size are the values you will
need to enter into a meta analysis. It will often help to store them in another
data object so they can be easily accessed and entered into the meta analysis
functions.

6.12 Conducting a meta analysis in R

Imagine that we have compiled effect sizes (and their standard errors) from a
list of studies as follows:
effectsizes <- c(0.7, 0.4, 2.1, 0.9, 1.6)
standarderrors <- c(0.2, 0.3, 0.9, 0.3, 0.5)

The effect sizes might be values of Cohen’s d, and the standard errors will be
the square root of the variance estimates that are returned when the effect size
6.12. CONDUCTING A META ANALYSIS IN R 113

is calculated (see previous section). The rmeta package contains functions that
will use these values to conduct a meta analysis. There are several varieties of
meta analysis available, but we will use the meta.summaries function to conduct
a random effects meta analysis using the effect size measures.
meta.summaries(effectsizes,standarderrors,method='random')

## Random-effects meta-analysis
## Call: meta.summaries(d = effectsizes, se = standarderrors, method = "random")
## Summary effect=0.852 95% CI (0.456, 1.25)
## Estimated heterogeneity variance: 0.078 p= 0.149

The output from this function tells us the summary effect size (0.852) and its 95%
confidence intervals. This is useful, but it’s more helpful if we save the output of
the function into a data object, which we can then pass into the metaplot and
funnelplot functions to produce graphical summaries of the results as follows:
metaoutput <- meta.summaries(effectsizes,standarderrors,method='random')

# this line of code tells R to put the next two plots side by side
par(mfrow=c(1,2), las=1)

metaplot(effectsizes,standarderrors,summn=metaoutput$summary,
sumse=metaoutput$se.summary,sumnn= metaoutput$se.summary^-2,
xlab='Effect size (d)',ylab="Study",summlabel='')

funnelplot(metaoutput, plot.conf=TRUE)

● ●
Study

Size

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0 1 2 3 4

Effect size (d) Effect

Figure 6.5: Auto-generated forest and funnel plots, using the metaplot and
funnelplot functions.
114 CHAPTER 6. META ANALYSIS

These plots (shown in Figure 6.5) are quite rudimentary, but can be improved
by specifying additional input arguments. For example, the author names
can be specified using the labels argument, and different colours chosen with
the colors argument (see further details in the help files). The funnel plot
can be automatically mirrored about its mid-point by adding the argument
mirror=TRUE to the function call.
It is also helpful to use the generic summary function to get a more detailed
summary of the meta analysis, which includes everything you would need to
generate your own forest plot manually (e.g. in another plotting package):
summary(metaoutput)

## Random-effects meta-analysis
## Call: meta.summaries(d = effectsizes, se = standarderrors, method = "random")
## ----------------------------------------------------
## Effect (lower 95% upper) weights
## 1 0.7 0.31 1.09 1.7
## 2 0.4 -0.19 0.99 1.2
## 3 2.1 0.34 3.86 0.2
## 4 0.9 0.31 1.49 1.2
## 5 1.6 0.62 2.58 0.6
## ----------------------------------------------------
## Summary effect: 0.85 95% CI ( 0.46,1.25 )
## Estimated heterogeneity variance: 0.078 p= 0.149
Those are the basics of doing a meta analysis in R. There is much more function-
ality in the rmeta package, and other packages are available for specific types of
meta analysis and other variations on the analysis.

6.13 Practice questions

1. Calculate Cohen’s d for a study with means of 0.3 and 0.34, standard
deviations of 0.1, and sample sizes of 24 per group.
A) 0.8
B) 0.9
C) 0.3
D) 0.4
2. Calculate the equivalent Cohen’s d for a correlation coefficient of r=0.7
with a sample size of 10.
A) 0.70
B) 1.96
C) 0.87
D) 0.33
3. What is the odds ratio that corresponds to a t-test with a p-value of 0.01,
and 30 participants per group?
A) 3.48
6.13. PRACTICE QUESTIONS 115

B) 1.25
C) 0.69
D) 4.19
4. What is the value of Hedge’s g for an ANOVA with an F-ratio of 13.6 and
17 participants in each group?
A) 1.26
B) 0.53
C) 0.60
D) 1.24
5. What is the log odds ratio for comparing proportions of 0.7 and 0.6 with 3
participants per group?
A) 0.12
B) 1.56
C) 0.44
D) 0.24
6. Conduct a random effects meta analysis using effect sizes of d = 0.1, 0.6,
-0.2, 0.9 and 1.1, with standard deviations of 0.2, 0.3, 0.1, 0.4 and 0.5.
What is the aggregate effect size?
A) 0.80
B) -0.01
C) 0.37
D) 0.20
7. What is the aggregate effect size using the values from question 7, but
conducting a fixed effects analysis instead?
A) 0.80
B) -0.01
C) 0.37
D) 0.20
8. Produce a forest plot using the data from question 6 (assuming random
effects). Is there a significant effect overall?
A) No, because the diamond overlaps the line of no effect
B) Yes, because the diamond overlaps the line of no effect
C) Yes, because most individual studies do not overlap the line of no
effect
D) No, because one of the individual studies has a negative effect
9. Which of the following is not a plausible explanation for an asymmetrical
funnel plot?
A) Small sample studies being more likely to produce significant effects
B) Random sampling
C) P-hacking
D) Publication bias
10. If 10 members of a treatment group of 500 recover from an illness, whereas
only 5 members of a control group of 400 recover, what is the odds ratio?
A) 0.020
B) 0.013
C) 1.60
116 CHAPTER 6. META ANALYSIS

D) 1.61
Answers to all questions are provided in section 20.2.
Chapter 7

Mixed-effects models

Mixed-effects models (also sometimes referred to as linear mixed models, or

hierarchical linear models) are a class of statistical tests that build upon simpler
tests that you may already be familiar with, such as regression, t-tests and
ANOVA. As readers of Chapter 4 will know, all of these tests are based on the
general linear model, in that they all involve fitting straight lines to data to
explain a portion of the total variance in a systematic way. The key difference
with mixed-effects models is the inclusion of one or more random effects - variables
where our observations are grouped into sub-categories that have a systematic
effect on the outcome. In many situations, a mixed-effects model could be used
instead of one of the simpler tests. The main advantages are (i) the mixed-
effects approach models more of the variance, including item-level variance, (ii)
mixed-effects models can incorporate group and individual differences, and (iii)
mixed-effects models can cope well with missing data points and unequal group
(cell) sizes. The aim of this chapter is to introduce the basic concepts of mixed
effects models, and illustrate them with examples and code.
It is important to be clear about the distinction between fixed and random effects.
A fixed effect is can be either a continuous or categorical independent variable,
much as we might use in standard regression or ANOVA designs. We include
these in our design because we anticipate that they may be able to explain a
portion of the variance in our dependent variable (outcome measure), and we
would like to know if this effect is statistically significant. A random effect is a
grouping variable that we might also expect to have an effect on the dependent
variable, but usually one that we are not interested in, and therefore want to
control for. Examples of groups include different individuals (each of whom is
tested multiple times), different testing sites (schools, countries, forests, oceans,
farms, bee colonies, etc.), or any other plausible grouping variable such as species,
family, breed (e.g. of dog) or genotype. Whereas the repeated measures ANOVA
framework can cope with a single grouping variable, mixed effects models can
include multiple random effects.

117
118 CHAPTER 7. MIXED-EFFECTS MODELS

In common with other versions of the general linear model (see Chapter 4), the
general idea of mixed effects models is to try to account for as much of the overall
variance in the data set as possible, using our various predictors. We usually
want to know if our fixed effects are able to account for a significant proportion
of the variance, but we also want to account for the variance due to our random
effects. Sometimes this is because random effects are ‘nuisance’ variables that
we need to control for, but are not really interested in. Including the random
effect in our model means that we can remove this variance, reducing the noise
in our estimate of the fixed effects. Sometimes accounting for a random effect
can reveal structure in a data set that is otherwise masked by group differences,
as we will demonstrate with our first example.

7.1 Different types of mixed-effects model

In our first example of mixed effects modelling, we will use a simulated data set
with one continuous predictor variable (a fixed effect or independent variable),
one outcome measure (dependent variable), and one random effect (grouping
variable) with five levels. Figure 7.1 plots the outcome against the predictor;
this just looks like cloud of points without much overall structure. If we try
conducting traditional linear regression, we get a quite a flat line overall (dashed
line), which does not differ much from a line that we constrain to be flat (solid
line). In regression terminology, the solid line is an intercept only model, which
has its slope fixed at zero, and can only shift up and down. The dashed line
has both an intercept and a slope that are free to vary. If there were an overall
positive or negative relationship in this data set, linear regression would give us
a fitted (dashed) line with a slope that differs from zero.
Using linear regression, we can ask statistically if the slope of the fitted line
differs significantly from zero. In this case it does not (p > 0.05):
## Analysis of Variance Table
##
## Response: DV
## Df Sum Sq Mean Sq F value Pr(>F)
## IV 1 28 28.34 0.0674 0.7953
## Residuals 498 209475 420.63
However, what if the data points came from five different groups (which is how
they were actually simulated)? When we plot them this way (different colours
in Figure 7.2a) there is much a clearer structure. Within each group, the IV is
quite strongly predictive of the DV. Mixed-effects models allow us to characterise
these relationships in three different ways.
The first approach is to allow each group to have its own intercept, referred to as
a random intercepts model. This means we would fit a regression line across all
five groups that has a constant slope, but allow it to shift up and down for each
group. This is what is represented by the individual coloured lines in Figure 7.2b.
7.1. DIFFERENT TYPES OF MIXED-EFFECTS MODEL 119

Linear regression

100
●

● ●
●
●●
●
80 ● ●
●
●
●

●
●
●
●
●
●
●
●
●
● ● ● ● ●
● ●● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●● ● ●
●● ● ● ●● ●
●
●
●● ●● ● ● ● ●● ●
● ●
● ●
● ● ●●●● ● ●
●
● ● ● ●●●
● ●
● ● ● ●● ● ●● ● ●● ● ●
60 ● ●
● ● ●● ● ● ●
● ●●●● ● ● ● ●
●
●● ● ● ●
●
● ● ●
● ● ● ● ●
●●
●
●
● ●●● ● ●
●
● ● ●● ● ● ● ● ● ●
● ● ● ●
DV

● ● ● ● ● ● ●
●
● ● ● ● ● ●● ● ● ● ● ● ● ● ●
● ● ● ● ● ●● ●
● ● ● ● ● ●● ●
● ●● ● ●
● ● ● ●● ●● ●● ● ●● ●● ● ● ● ●
●
● ●● ● ● ● ● ●●● ● ● ● ●●
● ● ● ● ●● ●
● ●● ● ●
● ● ● ● ● ● ● ●● ● ● ● ●● ● ●
●
●
● ● ● ● ●● ● ●
● ●● ● ●

40
● ●
●● ● ● ● ● ● ●● ● ● ●
● ● ●● ● ● ●
● ●
● ● ● ● ● ●
●
●● ● ●● ●
● ● ●● ●
● ●● ● ●● ● ● ● ● ● ●
●● ●● ●●● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ● ●● ● ●●
● ●● ●● ●●
● ● ● ● ● ● ● ●
●●●
● ●
●
● ● ● ● ●● ●● ● ●●
● ● ● ● ●
● ● ●
● ● ● ● ●● ●●
●
● ● ●

20 ●
● ●
●
●● ●
●
●
● ●
●●
●
●● ●
●
● ● ●●●
●
●●
● ● ● ●●
●
● ●

● ● ● ● ●
● ●
● ● ● ● ● ●
● ●
● ● ● ●
●
● ●● ● ●
●
● ●
●● ●
● ● ●
● ●●
0 ●
●
●
●●

●
●

0 20 40 60 80 100
IV

Figure 7.1: Simulated data showing the relationship between one independent
variable (IV) and one dependent variable (DV). The solid grey line is an intercept-
only regression, with slope constrained to be 0. The dashed black line is the best
fitting regression line, with slope and intercept free to vary.
120 CHAPTER 7. MIXED-EFFECTS MODELS

Data tagged by group Random intercepts

100 100
(a) ● (b) ●

● ●
● ● ● ●
●● ●●

● ●

80 ● ●
●
●
●

●
●
●
●
●

●
●
●
80 ● ●
●
●
●

●
●
●
●
●

●
●
●
● ● ● ● ● ●
● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ●● ● ● ●● ●
●● ● ● ● ● ●● ● ● ● ●
●● ● ● ● ● ●● ● ● ● ●
●● ● ● ●● ● ●● ● ● ●● ●
●● ● ● ● ●● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ●
● ● ● ●
● ● ●● ● ● ● ● ● ● ●● ● ● ● ●
● ● ●● ●● ● ● ●● ●●
● ● ● ●
60
●

60
● ● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ● ●●
● ● ● ● ● ●● ● ● ●●
● ● ● ●
● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ●
● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
●● ●
●
● ●● ● ●● ● ●● ●
●
● ●● ● ●● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
DV

DV
● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ●●● ● ● ● ● ● ●●● ●
● ● ●● ● ● ● ●
● ●● ● ● ● ●● ● ● ● ●
● ●● ●
● ● ●● ● ● ● ● ● ● ●● ● ● ● ●
●● ● ● ● ● ●● ● ● ● ●
● ● ● ● ●● ● ● ● ● ● ●● ●
●● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●
●
● ● ● ●●● ● ● ● ●● ●
● ● ● ●●● ● ● ● ●●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ●● ● ● ● ● ● ●● ● ●
● ● ● ● ● ● ● ●
● ●

40
●

40
● ● ● ● ●
● ● ● ● ● ● ● ●
●
● ● ● ● ● ● ● ● ● ●
●
● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●
●● ●●
● ● ● ● ● ● ● ●
● ● ●● ● ●
● ● ●● ● ● ● ●
● ● ●● ● ●
● ● ●● ● ● ● ●
● ● ● ● ● ● ● ●
●● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●
● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ●● ●
● ● ●● ● ● ●●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ●
● ●
● ● ● ●● ● ● ● ● ●● ●
● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●
● ● ● ● ● ● ● ● ● ●
● ● ● ● ●● ● ● ● ● ●●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ●

20 20
● ● ● ● ● ● ● ●
● ● ● ●●● ● ● ● ● ● ●●● ● ●
●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ●
●● ● ● ● ● ● ●
●● ● ●
● ● ● ● ● ●
● ●● ● ●●
● ● ● ● ●● ● ● ● ● ● ●● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ●
● ● ● ●● ● ● ● ● ●● ●
● ●
● ● ● ●
● ●
● ● ● ●
● ● ● ● ● ●
● ●● ● ●●

0 ●
●

●
●●
● 0 ●
●

●
●●
●

0 20 40 60 80 100 0 20 40 60 80 100
IV IV

Random slopes Random intercepts and slopes

100 100
(c) ● (d) ●

● ●
● ● ● ●
●● ●●

● ●

80 ● ●
●
●
●

●
●
●
●
●

●
●
●
80 ● ●
●
●
●

●
●
●
●
●

●
●
●
● ● ● ● ● ●
● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ● ● ●
● ●● ● ● ●● ●
●● ● ● ● ● ●● ● ● ● ●
●● ● ● ● ● ●● ● ● ● ●
●● ● ● ●● ● ●● ● ● ●● ●
●● ● ● ● ●● ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ●
● ● ● ●
● ● ●● ● ● ● ● ● ● ●● ● ● ● ●
● ● ●● ●● ● ● ●● ●●
● ● ● ●
60
●

60
● ● ● ● ● ● ● ● ● ● ● ● ●
● ●
● ● ●● ● ● ● ● ●●
●●
● ● ●
●
● ● ● ● ● ●● ● ● ● ● ●●
●●
● ● ●
●
● ● ●
● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ●
● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●
●● ● ● ●● ● ●● ● ● ●● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ● ●● ●
DV

DV

● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ●●●
● ● ● ● ● ●●●
● ●
● ●● ●● ● ● ● ●● ●● ● ●
● ● ● ●●
●●
●
● ● ●● ●
●
● ● ●
● ● ● ● ●●
●●
●
● ● ●● ●
●
● ● ●
●
● ● ● ●● ● ● ● ● ●● ●
● ● ● ● ● ● ●● ● ● ● ● ● ● ●●
● ●● ● ● ●● ●
●
● ● ●●● ● ● ● ●● ●
● ● ●●● ● ● ● ●●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ●● ● ● ● ● ● ●● ● ●
● ● ● ● ● ● ● ●
●● ●●

40
●

40
● ● ● ● ●
● ● ● ● ● ● ●
●
● ● ● ● ● ● ● ● ●
●
● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●
● ●● ● ● ● ● ● ● ● ●
●● ● ● ● ●● ● ● ●
● ●
● ● ●
● ● ●
● ● ●● ● ● ● ●
● ● ●
● ● ●
● ● ●● ● ● ● ●
● ● ● ● ● ● ● ●
●● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●
● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ●● ● ●● ● ● ● ● ● ●● ● ●● ●
● ● ●● ● ● ●●
● ● ● ● ● ● ●
● ● ● ● ● ● ● ● ●
● ●
● ● ● ●
● ● ● ● ● ●
● ●
● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ●
● ● ● ● ● ● ● ● ● ●
● ● ● ● ●● ● ● ● ● ●●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ●

20 20
● ● ● ● ● ● ● ●
● ● ● ●●● ● ● ● ● ● ●●● ● ●
●
● ● ● ● ● ● ● ●
● ● ● ● ● ● ●
● ● ● ●
●● ● ● ● ● ● ●
●● ● ●
● ● ● ● ● ●
● ●● ● ●●
● ● ● ● ●
● ● ● ● ● ● ●
● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ● ● ● ● ●
● ● ● ●● ● ● ● ● ●● ●
● ●
● ● ● ●
● ●
● ● ● ●
● ● ● ● ● ●
● ●● ● ●●

0 ●
●

●
●●
● 0 ●
●

●
●●
●

0 20 40 60 80 100 0 20 40 60 80 100
IV IV

Figure 7.2: The same data as shown previously, but with groups tagged in
different colours (a). Mixed effects models with random intercepts (b), random
slopes (c), and random intercepts and slopes (d), are shown by the regression
lines.
7.1. DIFFERENT TYPES OF MIXED-EFFECTS MODEL 121

The black dashed line is the grand average regression line. It is quite different
from our traditional regression fit (in Figure 7.1), which was essentially flat. If
we now perform a statistical comparison, we see that there is a highly significant
effect of the IV:

## Type III Analysis of Variance Table with Satterthwaite's method

## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## IV 40802 40802 1 496.48 344.39 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

An alternative to the random intercepts model is the random slopes model,

shown in Figure 7.2c. In this model, the slopes are permitted to vary, but the
intercept is fixed across all groups. Recall that the intercept is the value of y
when x=0 (in the regression equation y = β0 + β1 x), which is why all the lines
in Figure 7.2c meet at this point. For our current data set this model also does
a reasonable job of describing the data.

Finally, we can allow both the intercepts and slopes to vary between groups,
as shown in Figure 7.2d. This captures the shallower slope of the bottom
group (in purple), as well as the vertical offsets between groups. Overall this
type of model has more degrees of freedom than the other two. We could
alternatively have run five completely independent linear regressions (or used
multiple regression) instead of our single mixed effects model, but the mixed
effects approach additionally gives us the grand average regression line, that
takes account of the sample size of each group and has greater overall power
than for any individual group. In other words it tells us about the overall effect
of the independent variable, rather than its effect only within each group. For
models where either the slope or intercept are fixed, the mixed effects framework
lets us jointly estimate the value of the fixed parameter across all of our groups.

Mixed-effects regression models are enormously flexible, and we will learn about
the syntax to implement them later in the chapter. The decision of whether to
include random intercepts, random slopes, or both will depend heavily on the
hypothesis you are trying to test. However it is quite rare to find a situation where
only random slopes are required - most models either involve random intercepts,
or allow both parameters to vary. Sometimes it might also be advisable to test
more than one model, and use goodness of fit indicators such as R2 to decide
which model describes the data best (see section 7.6 for more detail). Note
that because a random intercepts model requires fewer degrees of freedom than
a model in which both parameters vary, it can sometimes be fit to data sets
with fewer observations. In the next section, we will run through an example of
mixed-effects regression using data from the literature.
122 CHAPTER 7. MIXED-EFFECTS MODELS

7.2 Mixed-effects regression example: lung

function in bottlenose dolphins
A study by Fahlman et al. (2018) measured resting lung function in 32 bottlenose
dolphins. The study used mixed-effects regression models to understand the
relationships between several variables. The main dependent variable was the
tidal volume (VT ), measured in litres, which is an index of lung capacity. Here we
focus on the relationship between VT and body mass (in kg), as shown in Figure
2 of the original paper. The raw data are available online (https://osf.io/6wjh8/)
and are replotted in Figure 7.3 (I have slightly tweaked the inspiration data by
adding one to each data point for the purposes of this example).

12
●
● Expiration ● ●
10 ● Inspiration
Tidal volume (l)

● ● ●
● ●
● ● ●
8 ●● ●
●● ● ●● ●
●
●
● ●● ● ●● ●
●
● ● ●
● ●
● ●● ● ●
6 ● ●
● ●
● ● ●● ● ●●
● ●
●
● ● ●● ●
● ● ● ● ● ● ● ● ●● ● ●
● ● ●● ●● ● ● ●● ●
● ●
● ● ● ●
4 ●● ●
●
●
●● ● ● ●
● ● ●
● ●
●
● ●
2

0
50 100 150 200 250 300
Body mass (kg)

Figure 7.3: Tidal volume for bottlenose dolphins as a function of body mass,
modified and replotted from Fahlman et al. (2018).

The traditional approach to analysing these data might be to conduct a single

regression, incorporating all data points and fitting one straight line. However,
this is not ideal in the current situation for two reasons. First, each dolphin
contributes between 1 and 4 observations (notice that there are 32 animals, but
more than 32 data points of each condition). If we averaged across measurements
for each animal, this would mean our measurement precision changed for different
animals. If we included each data point, we would be double-counting (or
triple or quadruple counting) some animals, meaning they would contribute
disproportionately to estimating the regression line. Second, we have two
measures here (expiration and inspiration - breathing out and in), and in
7.3. FACTORIAL MIXED-EFFECTS EXAMPLE: LEXICAL DECISION TASK FOR NOUNS AND VERBS123

principle the relationship with body mass might differ between them.
The mixed-effects approach allows us to deal with both of these issues in a coher-
ent way. We treat body mass and breath direction (expiration and inspiration)
as fixed effects, and animal as a random effect. This allows us to have separate
regression lines for the two directions of breath (expiration and inspiration), and
test if there is evidence overall for an effect of body mass. The two lines in
Figure 7.3 show these fits, and the regression output tells us there is a significant
effect of body mass, and also of breath direction:
## Type III Analysis of Variance Table with Satterthwaite's method
## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## bodymass 30.719 30.719 1 26.720 26.521 2.096e-05 ***
## direction 34.799 34.799 1 77.414 30.044 5.095e-07 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The significant effect of body mass indicates that the overall regression slope
is significantly steeper than 0. The significant effect of direction tells us that
the two breath directions involve different regression lines (though note this is
largely because I modified the data for this example).
To account for the multiple observations from some animals, we treat animal as
a random effect. In the current context, this is similar to a repeated measures
design for a t-test or ANOVA. But crucially, those tests expect a balanced design,
where each individual contributes the same number of observations. The mixed-
effects approach relaxes this assumption, so it is more flexible when dealing with
real data sets.
As you might recall from Chapter 4, the general linear model underlying regression
can also be used in situations where the independent variable is categorical rather
than continuous. The most familiar instances of this are t-tests and ANOVAs.
We can bring the benefits of mixed-effects models to factorial experimental
designs, and also include multiple random effects, as we will see in our next
example.

7.3 Factorial mixed-effects example: lexical de-

cision task for nouns and verbs
Mixed-effects models have become very popular the area of psycholinguistics -
the study of how the brain processes language. In this field, it is necessary to
have multiple different stimulus examples that are all drawn from a particular
category - for example nouns or verbs. Some of these examples will be easier
than others for a given task, and each example can be presented to the same
participant only once so that memory processes do not interfere with the task.
In traditional analyses (e.g. ANOVAs), this item-level variance would be ignored,
which reduces statistical power (Westfall, Kenny, and Judd 2014). Mixed-effects
124 CHAPTER 7. MIXED-EFFECTS MODELS

Table 7.1: Example stimuli for a lexical decision task. I ask the forgiveness
of any real psycholinguists reading this, who would no doubt have numerous
objections to using these stimuli in an actual experiment.

Example Nouns Non-nouns Verbs Non-verbs

1 pancake poncake speak speek
2 eyeball eyepall break bweak
3 doorway doornay drill driil
4 computer compuler stretch stredch
5 watermelon wadermelon manipulate manipulake
6 husband hubsand pinch pinsh
7 chicken chisken spell srell
8 television telerision grow mrow
9 flagpole flagmole watch watsh
10 railway rainway grab greb
11 tractor tragtor listen listun
12 fertilizer fertinizer fly flei
13 rectangle rectangel swim swib
14 lettuce lestuce yawn yorn
15 flower flowor scrape scyape
16 rabbit rabpit wink wimk
17 banana banama fight finht
18 firefly firafly jump fump
19 cheesecake cheeseqake punch punxh
20 grape grage scratch scratsh

models allow us to treat such variance in a principled way, much as a repeated

measures t-test or ANOVA deals with variance between individual participants.

For our next example, we consider a lexical decision task in which participants
are presented with a string of characters, and must decide if they are a word
or a non-word. Examples of non-words are often based on real words, but with
some errors introduced, for example “bekause”, but they can also be nonsense
strings of letters, for example “okjsdfj”. In our experiment, 20 participants each
respond to 20 nouns and 20 verbs, and also 20 non-words based on the original
nouns, and 20 non-words based on the original verbs. The dependent variable is
the reaction time, measured in milliseconds. Our example stimuli are shown in
Table 7.1.

For our first participant, we will show these stimuli in a random order, and
measure reaction times for each decision (word vs non-word). Their data might
look something like the values shown in Table 7.2.

In a traditional analysis, we would average across all of the examples in each

category, and use the participant means (i.e. the numbers in the final row of
7.3. FACTORIAL MIXED-EFFECTS EXAMPLE: LEXICAL DECISION TASK FOR NOUNS AND VERBS125

Table 7.2: Example reaction times for a lexical decision task, for one participant
(times in ms).

Example Nouns Non-nouns Verbs Non-verbs

1 453 634 495 509
2 494 547 614 557
3 477 553 473 611
4 414 587 567 602
5 476 537 530 614
6 441 597 480 600
7 496 550 516 599
8 533 652 519 585
9 477 590 526 582
10 457 473 498 556
11 446 616 544 624
12 497 593 556 584
13 477 510 603 600
14 474 586 553 539
15 456 620 600 623
16 546 591 512 640
17 403 638 477 579
18 452 566 476 616
19 450 585 514 526
20 517 663 504 541
Mean 472 584 528 584
126 CHAPTER 7. MIXED-EFFECTS MODELS

Table 7.2). These would then be entered, along with the means of the other
19 participants, into a 2x2 repeated measures ANOVA. The two factors for the
ANOVA are word type (noun or verb) and word validity (word or non-word),
and each participant would contribute a single mean reaction time for each of
those four conditions (meaning 80 data points in total - 20 participants × 4
conditions). The ANOVA results might look something like this:
##
## Error: subject
## Df Sum Sq Mean Sq F value Pr(>F)
## Residuals 1 5470 5470
##
## Error: Within
## Df Sum Sq Mean Sq F value Pr(>F)
## wordtype 1 20654 20654 18.010 6.21e-05 ***
## validity 1 103705 103705 90.431 1.61e-14 ***
## wordtype:validity 1 10878 10878 9.485 0.00289 **
## Residuals 75 86009 1147
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
We see significant effects for word type, validity, and their interaction. We can
also plot the means for each condition, along with individual data points, as
shown in Figure 7.4.
Figure 7.4 is a conventional plot, in which the error bars show the standard
deviation across participants, and each point corresponds to a different individual
person (N=20). Note that some individual participants are generally fast or
generally slow, regardless of the condition. For example, you can see that the
highest points stay near the top in each condition, as linked by the faint grey lines.
It is this between-participant variance (the tendency for individuals to differ
systematically across conditions) that the repeated measures design can discard,
and which gives it a greater statistical power relative to a between-participants
design (where we test different individuals in each condition).
However, there is another way to think about the results of this experiment.
Notice that some of the words in Table 7.1 are likely to be easier than others to
identify. For example, in the non-nouns set, poncake might be quite a challenging
word (because the o looks like and a, and pancake is a noun). On the other
hand, lestuce might be an easier example to identify correctly as a non-word.
We can produce an alternative plot to Figure 7.4, by averaging reaction times
across participants for each item (instead of averaging across items for each
participant). Again, this will involve there being 80 observations: 20 items × 4
conditions. For the current example, we can see from Figure 7.5 that the ‘By
items’ plot has a key similarity with the ‘By participants’ plot (Figure 7.4): the
group means (horizontal black lines) are the same in both graphs. This has to be
the case because these are the grand averages across both items and participants.
7.3. FACTORIAL MIXED-EFFECTS EXAMPLE: LEXICAL DECISION TASK FOR NOUNS AND VERBS127

800
By participants
Reaction time (ms)

● ●
700 ●
●● ● ●
●●
●● ●●●
●●● ● ● ●
●
●● ● ●● ●● ●● ● ●
●● ● ● ● ●●
●●
● ● ●
● ● ●
● ● ●
600 ●● ●● ●●
●
●●
●
●●
●
●
● ●
●● ●
●
●●
●● ●
●● ●
500 ●
●

400
Nouns

Non−nouns

Verbs

Non−verbs

Figure 7.4: Graph showing condition means (black bars) and individual data
points (symbols) for the four conditions. Error bars indicate the standard
deviation across participants, and thin grey lines join points for an individual
participant.
128 CHAPTER 7. MIXED-EFFECTS MODELS

However the group variances differ between the plots, as do the individual points.

800
By items
Reaction time (ms)

700

600

500

400
Nouns

Non−nouns

Verbs

Non−verbs
Figure 7.5: Graph showing condition means (black bars) and data points for
each item (squares) for the four conditions. Error bars indicate the standard
deviation across items, and thin grey lines join points for an individual stimulus
item.

Something to notice about the ‘By items’ plot (Figure 7.5 is that the standard
deviations are much smaller than in the ‘By participants’ plot (Figure 7.4). This
suggests that the responses to the different items are more similar to each other
than are the responses of different individuals. Now, we could in principle do
another ANOVA, but this time treating item as the unit we average within,
instead of participant. Note that item is repeated within the word class of noun
or verb, but not across classes (see Table 7.1), which is why the faint grey lines
only link within a word type category in Figure 7.5. The (mixed) ANOVA output
looks like this:
##
## Error: item
## Df Sum Sq Mean Sq
## wordtype 1 77105 77105
##
## Error: Within
## Df Sum Sq Mean Sq F value Pr(>F)
## wordtype 1 37483 37483 67.41 4.75e-12 ***
## validity 1 12804 12804 23.02 7.95e-06 ***
7.4. HOW CAN I DECIDE IF AN EFFECT IS FIXED OR RANDOM? 129

## wordtype:validity 1 10878 10878 19.56 3.25e-05 ***

## Residuals 75 41707 556
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
This analysis produces larger F-ratios (sometimes referred to F2, where the
ANOVA based on participants is referred to as F1), and smaller p-values, because
for our example the item-wise variance is lower than the participant-wise variance.
However this feels somewhat uncomfortable - running the same analysis two
ways on the same data set has a rather underhand feel. Most experimentalists
would balk at the idea of counting items as the unit of observation, because
it divorces the number of observations from the number of participants. And
indeed, there are more principled statistical objections to this practise (Rietveld
and Hout 2007).
Mixed-effects models offer an alternative. We can instead calculate a model that
takes into account both the participant variance and the item-level variance,
treating them both as random effects. Word type and validity - the independent
variables from our ANOVAs - are included as fixed effects. Instead of averaging
over items (or participants), we give the model all of our data points without
any averaging. In this example, that consists of 20 participants × 20 items ×
4 conditions = 1600 data points. The output of such a model might look as
follows:
## Type III Analysis of Variance Table with Satterthwaite's method
## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## wordtype 180483 180483 1 100.96 115.15 < 2.2e-16 ***
## validity 806248 806248 1 1539.00 514.38 < 2.2e-16 ***
## wordtype:validity 217552 217552 1 1539.00 138.80 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The mixed-effects approach is modelling each individual effect - each participant’s
change in means across condition, and each item’s change in means across condi-
tion. You will see by comparing the output tables that this generally produces
larger F-ratios, smaller p-values, and that there are many more (denominator)
degrees of freedom than in the ANOVAs. In the ANOVAs, we had averaged
across either items or participants to reduce the total variance. Modelling the
individual effects increases our statistical power, and also deals sensibly with
any correlations within participants or items.

7.4 How can I decide if an effect is fixed or ran-

dom?
One common question in mixed effects modelling is how to decide if an effect
should be considered as fixed or random. For example, in Figure 7.3, we treated
130 CHAPTER 7. MIXED-EFFECTS MODELS

breath direction as a fixed effect with two levels - why wasn’t this a random effect
instead? This turns out to be quite a complicated question, to which there is not
always a definitive answer, and these decisions are often left to the person doing
the analysis. One good heuristic is to think about whether you are interested in
the effect in question: if you are, it should probably be a fixed effect. In addition,
there are two important factors that prevent a variable from being treated as
a random effect. First, if the variable is continuous, it cannot be used as a
random effect; only categorical variables can be treated in this way. This rules
out variables like age and weight from being treated as random factors, unless
they are discretised into categories first. Second, random effects should have at
least five levels, as with fewer levels the estimate of the standard deviation for
that variable will be inaccurate. This is the reason why breath direction needed
to be a fixed effect in the dolphins example, and it also means that variables
such as sex/gender, handedness, blood group, and (in genetics) single-nucleotide
polymorphism must either be treated as fixed factors or ignored.

7.5 Dealing elegantly with missing data and un-

equal groups
An added bonus with mixed-effects models is that they cope well with missing
values, and situations where we have groups of different sizes. Imagine that we
decided to exclude trials from our lexical decision task on which the participant
got the task wrong (i.e. they indicated that a stimulus was a word when it was a
non-word, or vice versa). The excluded trials will involve different words for each
participant. This has the effect of making the experimental design unbalanced.
Depending on the research design, there could be many other reasons for missing
observations, such as equipment malfunction, participants missing sessions in
longitudinal work, and so on.

For traditional ANOVAs, we usually deal with missing data by either excluding
participants (known as listwise deletion, this can substantially reduce power),
or by averaging across only the data points we have (which means that some
participants contribute more observations than others). In mixed-effects models,
we are estimating the properties of an underlying regression line as best we can
for each comparison. Critically, if some estimates are missing, we can still come
up with a sensible parameter estimate. We can see how this might work by
randomly removing some of the observations from our psycholinguistics data set.
For example, if we remove 5% of the data points, the analysis runs fine, and the
summary table changes only very slightly:

## Type III Analysis of Variance Table with Satterthwaite's method

## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## wordtype 187295 187295 1 103.89 120.63 < 2.2e-16 ***
## validity 801521 801521 1 1459.20 516.22 < 2.2e-16 ***
## wordtype:validity 225751 225751 1 1459.17 145.39 < 2.2e-16 ***
7.6. REPORTING AND COMPARING MIXED-EFFECTS MODELS 131

## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Another situation that causes problems for ANOVA is when different groups
have very different sample sizes, also resulting in an unbalanced design. The
main issues are that it is difficult to accurately test whether groups of very
different sizes meet the homogeneity of variances assumption (i.e. that they
have equal variances), and also that if sample size covaries with an independent
variable, it can confound the main effect (in factorial designs). These issues
can cause particular problems when conducting research on rare conditions and
diseases, or in groups that comprise a minority of a population. Designs that
aim to sample the population at random (such as polling research) will tend
to select relatively few people in such categories, leading to highly unbalanced
designs. Again, mixed effects models are able to deal with this situation more
appropriately than ANOVA, because they correctly account for the variance
structure of the underlying data. This makes them a good choice for analysis of
data that relates to equality and diversity of underrepresented groups.

7.6 Reporting and comparing mixed-effects

models
As we have seen in the above examples, the predictors in mixed-effects models
can be assessed using F-tests, just like in regression and ANOVA. We can report
the main effects and interactions in much the same way as we would report
regression or ANOVA, along with p-values. However, it is critical that the design
of the model is properly described, so that it is clear to the reader which variables
are fixed effects and which are treated as random effects.
In some situations, it can also be helpful to demonstrate how well models with
different assumptions (e.g. with a random effect added or removed) are able
to describe the data. This is best achieved by fitting models in a systematic
way, much like in multiple regression where additional predictors are added in
sequence. Models can then be compared using a variety of statistics. The R2
statistic tells us the proportion of the total variance explained by the model,
just like in regression models, with larger values indicating a better model fit.
The model with the largest log likelihood statistic also indicates the best fit in
absolute terms (regardless of number of free parameters). Finally, the Akaike
Information Criterion (AIC) and Bayesian Information Criterion (BIC) scores
are based on the likelihood but also take the number of free parameters into
account, with the best model being the one with the smallest value.
Finally, it is important to display your data graphically in an intuitive way, so
that the reader can visualise the important effects for themselves and understand
what they mean. As well as showing main effects and interactions, by item
plots (e.g. Figure 7.5) are often useful, as are graphs that allow inspection
of the residual variance that the model cannot explain (such as Q-Q plots).
132 CHAPTER 7. MIXED-EFFECTS MODELS

More detailed recommendations on reporting mixed-effects models are given by

Meteyard and Davies (2020), though ultimately there are likely to be discipline-
specific norms and study-specific priorities for what to report. A useful guide is
to read relevant papers on the same topic that have used mixed-effects models,
to see how the results have been reported.

7.7 Practical problems with model convergence

Values for the free parameters of mixed effects models are determined using an
optimization algorithm similar to those that we will discuss in Chapter 9. In this
context, the free parameters are estimates of intercepts, slopes and variances in
the model. Sometimes the algorithm can fail to converge on the best parameter
estimates, which will produce a convergence error. This is very similar to the
problems with local minima that we will cover in section 9.8. In such situations,
the fitted parameters might not be the best possible estimates, though they will
often be acceptable. There are several possible solutions, including running the fit
again from different starting parameter values, or trying alternative optimzation
algorithms. It will sometimes also help to normalize any continuous fixed effects
so that they have a mean of zero and a standard deviation of one (see section
3.7). This does not change the overall structure of your data, but it places the
likely parameter values closer to the starting point for the algorithm.

7.8 Running mixed-effects models in R with

lmerTest
This section will demonstrate how to build and run mixed-effects models using the
lmer function in the lmerTest package (Kuznetsova, Brockhoff, and Christensen
2017). This builds upon the older lme4 package (Bates et al. 2015), and in
particular adds tests for significance (i.e. p-values) that were previously not
included. There are other R packages that can be used to run mixed-effects
models, including the nlme package (which can also run nonlinear mixed-effects
models) and the brms package (Bürkner 2017, 2018) for running Bayesian
versions of mixed-effects models. Outside of R, an extension to the SAS package
called PROC MIXED, and bespoke software such as MLwiN can also be used.
Before we can run a mixed-effects model, we need to get our data into the correct
format. The function expects data to be stored in a data frame, which we first
introduced in section 2.6. In brief, a data frame is a special type of data structure
that can store numbers, strings and factor variables, much like a spreadsheet.
We can create one using the data.frame function to combine a series of vectors,
or by reading data in from an external file such as a csv file (with the read.csv
function).
The syntax for the lmer function is based on the linear modelling formula used
in other R functions such as lm (linear model), and aov (to run ANOVAs), that
7.8. RUNNING MIXED-EFFECTS MODELS IN R WITH LMERTEST 133

we introduced in Chapter 4. These formulae have the general form: DV ~ IV,

where DV represents the dependent variable, IV is the independent variable,
and the tilde symbol (~) is read as ‘is predicted by’. So, if we wanted to predict
children’s heights based on their age, we might use a command such as:
output <- lm(height ~ age, data=dataset)

The above code would run a linear model using the data stored in a data frame
called dataset, trying to predict values of the height column using the values in
the age column. If we have additional independent variables, such as sex, we
can include them either as single predictors (additive, as in regression formulae),
or as factors that interact with the other independent variables (multiplicative,
as for ANOVA formulae):
# regression notation (no interaction)
output <- lm(height ~ age + sex, data=dataset)

# ANOVA notation (with interaction)

output <- lm(height ~ age * sex, data=dataset)

Formulae for the lmer function follow similar rules, but there is an additional
piece of syntax to consider. A random effect is always entered after the fixed
effects (i.e. independent variables), it is entered in brackets, and it is entered
after a vertical slash symbol. For example, if we wanted to include nationality
as a grouping variable to predict height, we might do so as follows:
# mixed-effects model call with random intercepts
output <- lmer(height ~ age + (1|nationality), data=dataset)

The above line of code will run a mixed-effects model with random intercepts
(see Figure 7.2b), using age as a predictor and nationality as a grouping variable
(random effect). Alternatively, we can specify a random slopes model (see
Figure 7.2c) by incorporating the independent variable into the random effects
specification, and specifying (using a 0) that the intercept is not included as a
random effect:
# mixed-effects model call with random slopes
output <- lmer(height ~ age + (0 + age|nationality), data=dataset)

Finally, we can specify a model with random slopes and intercepts (see Figure
7.2d) by allowing the intercepts to vary again:
# mixed-effects model call with random slopes and intercepts
output <- lmer(height ~ age + (1 + age|nationality), data=dataset)

A more generic specification for these three models is as follows:

# mixed-effects model call with random intercepts
output <- lmer(DV ~ IV + (1|group), data=dataset)
134 CHAPTER 7. MIXED-EFFECTS MODELS

# mixed-effects model call with random slopes

output <- lmer(DV ~ IV + (0 + IV|group), data=dataset)

# mixed-effects model call with random slopes and intercepts

output <- lmer(DV ~ IV + (1 + IV|group), data=dataset)

Just as we can have more than one independent variable, random effects can be
defined for multiple grouping variables, depending on the structure of our data
set. For example, the call for our factorial mixed-effects model for the lexical
decision task was:
model <- lmer(RT ~ wordtype * validity + (1|subject) + (1|item), data=RTlmm)

This line of code specifies that reaction time (RT) is predicted by two factorially
combined independent variables (word type and validity), with random intercepts
on subject and item.
As with many R functions, the output of the model fit is stored in another data
object (the one to the left of the <- assignment). This data structure contains
a lot of information, and we can extract it in several ways. Simply inspecting
the object (by typing its name) will give us some helpful numbers, such as the
number of observations, but it is not generally very informative:
model

## Linear mixed model fit by REML ['lmerModLmerTest']

## Formula: RT ~ wordtype * validity + (1 | subject) + (1 | item)
## Data: RTlmm
## REML criterion at convergence: 16473.13
## Random effects:
## Groups Name Std.Dev.
## item (Intercept) 22.66
## subject (Intercept) 33.53
## Residual 39.59
## Number of obs: 1600, groups: item, 40; subject, 20
## Fixed Effects:
## (Intercept) wordtype validity wordtype:validity
## 357.22 102.10 141.97 -46.64
Alternatively, we can use the generic summary function to get some more useful
numbers out:
summary(model)

## Linear mixed model fit by REML. t-tests use Satterthwaite's method [

## lmerModLmerTest]
## Formula: RT ~ wordtype * validity + (1 | subject) + (1 | item)
## Data: RTlmm
##
7.8. RUNNING MIXED-EFFECTS MODELS IN R WITH LMERTEST 135

## REML criterion at convergence: 16473.1

##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -3.2488 -0.6845 -0.0095 0.6634 3.1224
##
## Random effects:
## Groups Name Variance Std.Dev.
## item (Intercept) 513.5 22.66
## subject (Intercept) 1124.1 33.53
## Residual 1567.4 39.59
## Number of obs: 1600, groups: item, 40; subject, 20
##
## Fixed effects:
## Estimate Std. Error df t value Pr(>|t|)
## (Intercept) 357.220 16.809 117.499 21.25 <2e-16 ***
## wordtype 102.099 9.515 100.961 10.73 <2e-16 ***
## validity 141.973 6.260 1539.000 22.68 <2e-16 ***
## wordtype:validity -46.643 3.959 1539.000 -11.78 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Correlation of Fixed Effects:
## (Intr) wrdtyp valdty
## wordtype -0.849
## validity -0.559 0.592
## wrdtyp:vldt 0.530 -0.624 -0.949

This provides a summary table, which includes coefficient estimates, t-statistics

and p-values for the fixed effects, as well as some helpful summary information
about the residuals. It is more typical to report F-ratios for ANOVA-type designs,
which we can obtain by using the anova function:
anova(model)

## Type III Analysis of Variance Table with Satterthwaite's method

## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## wordtype 180483 180483 1 100.96 115.15 < 2.2e-16 ***
## validity 806248 806248 1 1539.00 514.38 < 2.2e-16 ***
## wordtype:validity 217552 217552 1 1539.00 138.80 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Similarly, we can request a summary table for our random effects terms with
the ranova function:
ranova(model)
136 CHAPTER 7. MIXED-EFFECTS MODELS

## ANOVA-like table for random-effects: Single term deletions

##
## Model:
## RT ~ wordtype + validity + (1 | subject) + (1 | item) + wordtype:validity
## npar logLik AIC LRT Df Pr(>Chisq)
## <none> 7 -8236.6 16487
## (1 | subject) 6 -8611.1 17234 749.12 1 < 2.2e-16 ***
## (1 | item) 6 -8402.6 16817 332.16 1 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The values included here are measures of how the model fit changes when a term
is removed. So, a significant p-value here implies that the random effects term is
making a meaningful contribution to the fit of the model.
We may also wish to report R2 values for the model fit. Calculating these
requires a function from a different package - the r.squaredGLMM function from
the MuMIn package, which we can provide with the model output as follows:
library(MuMIn)
r2 <- r.squaredGLMM(model)
r2

## R2m R2c
## [1,] 0.3454568 0.6798848
2
Two values are calculated and reported by this function. The first (Rm ) is the
2
marginal R value, which represents the proportion of the variance explained by
our fixed effects (i.e. traditional independent variables), excluding any random
effects. The second (Rc2 ) is the conditional R2 value, which is the proportion
of the variance explained by the full model, including both fixed and random
effects. It is helpful to report both of these statistics for each model that you
run.
If we want to compare two (or more) models, we can again use the anova
function to produce a table of useful statistics, including AIC, BIC and log
likelihood scores. The model with the smallest AIC and BIC scores, and the
largest log-likelihood score gives the best account of the data. Here is an example
comparing the models from Figure 7.2b-d:
anova(simmodel3,simmodel4,simmodel5)

## refitting model(s) with ML (instead of REML)

## Data: groupdata
## Models:
## simmodel3: DV ~ IV + (1 | group)
## simmodel4: DV ~ IV + (0 + IV | group)
## simmodel5: DV ~ IV + (1 + IV | group)
## Df AIC BIC logLik deviance Chisq Chi Df Pr(>Chisq)
7.9. FURTHER RESOURCES FOR MIXED-EFFECTS MODELS 137

## simmodel3 4 3843.7 3860.5 -1917.8 3835.7

## simmodel4 4 3901.6 3918.4 -1946.8 3893.6 0.000 0 1
## simmodel5 6 3830.1 3855.4 -1909.0 3818.1 75.507 2 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
It is sometimes necessary to find out the coefficients for our random effects
groups (perhaps to plot individual regression lines as in Figures 7.2b-d). The
coef function allows us to extract these values (shown here for the model displayed
in Figure 7.2d, and stored in the simmodel5 variable):
coef(simmodel5)

## $group
## (Intercept) IV
## 1 32.911341 0.9268825
## 2 -1.536704 1.0272352
## 3 4.649459 0.9156914
## 4 -11.237636 0.4863508
## 5 -25.522081 1.0810784
##
## attr(,"class")
## [1] "coef.mer"
The first column contains the intercept values, and the second column the slope
values.
Finally, we can inspect the residuals using a Q-Q plot, which we first encountered
in section 3.8. These graphs show the expected quantiles (based on a normal
distribution) along the x-axis, and the actual residuals (from the data) along
the y-axis. Substantial deviations from the major diagonal line (usually at the
extremes) indicate that the normality of residuals assumption has been violated,
and the model results should be treated with some caution. We can generate a
Q-Q plot (see Figure 7.6) using the qqnorm function, after first extracting the
residuals from the model object using the resid function:
modelresiduals <- resid(model) # extract the residuals
qqnorm(modelresiduals) # create the plot
qqline(modelresiduals) # add the diagonal line

7.9 Further resources for mixed-effects models

To learn more about mixed-effects models, many resources are available. The
book Linear mixed-effects models using R (Galecki and Burzykowski 2013) is a
comprehensive source that discusses a number of different designs with detailed
examples. The documentation and paper (Bates et al. 2015) for the lme4
package is helpful and contains example implementations. For item-level analysis
specifically, the paper by Baayen, Davidson, and Bates (2008) is an influential
138 CHAPTER 7. MIXED-EFFECTS MODELS

Normal Q−Q Plot

100
●
●●●
●
●●
●●●●●
●
●
●
●●
●
●●
●
●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
Sample Quantiles

●
●
●●
●
●●
●
50

●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
0

●
●
●
●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
−100 −50

●●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●
●●
●●
●●
●
●●
●
●●
●
●●●
●
●●●●
● ●●

−3 −2 −1 0 1 2 3

Theoretical Quantiles

Figure 7.6: Q-Q plot showing the residuals.

source, Westfall, Kenny, and Judd (2014) discuss issues relating to statistical
power, and Meteyard and Davies (2020) provide recommendations on reporting.
Readers convinced by the arguments for Bayesian statistics (Chapter 17) are
advised to read about Bayesian hierarchical models, which have similar properties
(see e.g. Kruschke 2014). It is also worth identifying papers in your own area
of research that use the mixed-effects approach, to find examples of common
practice. Finally, there are many helpful blog posts and discussion board threads
online that are well worth reading when troubleshooting specific issues.

7.10 Practice questions

1. In mixed-effects models, independent variables are called:
A) Random effects
B) Fixed effects
C) Mediators
D) Outcomes
2. A random effect is best described as:
A) A variable that is noisier than other variables
B) An effect that cannot be predicted based on group membership
C) A variable where sub-groups influence the outcome
D) Choosing slope values with a random number generator
7.10. PRACTICE QUESTIONS 139

3. Which of the following is not an advantage of mixed-effects models?

A) They can cope with situations where individuals contribute different
numbers of observations
B) They can handle group-level differences in the outcome
C) They can deal with variation due to different stimulus items
D) They are guaranteed to give significant results when ANOVAs do not
4. Mixed-effects models cope well with missing data because:
A) They can still estimate parameters even when some values are missing
B) They delete all participants with missing data
C) They set all missing values to 0
D) They delete all items with missing data
5. The formula DV ~ IV + (0 + IV|group) will fit a model with:
A) Random slopes
B) Random intercepts
C) Random slopes and random intercepts
D) An intercept fixed at 0
6. Which formula would fit a model with random slopes and intercepts?
A) DV ~ IV + (0 + IV|group)
B) DV ~ IV + (1 + IV|group)
C) DV ~ IV + (1|group)
D) DV ~ IV + (DV|group)
7. To check the assumption of normally distributed residuals, we usually plot
a:
A) Histogram
B) Normal distribution
C) Q-Q plot
D) By-items plot
8. Which summary statistic tells us the proportion of the variance explained
by a full mixed-effects model?
A) Rc2
B) AIC
C) Log likelihood
2
D) Rm
9. When comparing two models, which summary statistic will be smallest for
the best fitting model?
A) Rc2
B) AIC
C) Log likelihood
2
D) Rm
10. Imagine a study in which participants respond to a sequence of 100 images
of celebrity faces. Half of the participants are from a country in which
the celebrities are well-known, but the rest of the participants are from a
country where they are not famous. What would be the most appropriate
design to model these data?
A) Both image and country are random effects
B) Both image and country are fixed effects
140 CHAPTER 7. MIXED-EFFECTS MODELS

C) Image is a fixed effect and country is a random effect

D) Country is a fixed effect, and image is a random effect
Answers to all questions are provided in section 20.2.
Chapter 8

Stochastic methods

8.1 What does stochastic mean?

In lay terms, stochastic means random. The methods described in this chapter
involve using random numbers in one way or another. For various historical
reasons there are a number of more specific terms that are often used to refer to
particular methods within this general area, including:
• Monte Carlo methods
• Resampling
• Bootstrapping
• Jackknifing
• Probabilistic simulations
• Permutation testing
• Plug-in principle
This plethora of terms hints at how flexible and widespread stochastic methods
are. In this chapter, we will focus on running simulations with a stochastic
component, and bootstrap resampling methods to calculate confidence intervals
and test hypotheses.
I mostly find stochastic techniques useful because they offer a shortcut that
allows one to avoid lengthy mathematical derivations, and tend not to require
parametric assumptions. My favourite quote that illustrates their utility comes
from the Numerical Recipes books:
“Offered the choice between mastery of a five-foot shelf of analytical
statistics books and middling ability at performing statistical Monte
Carlo simulations, we would surely choose to have the latter skill.”
(Press et al. 1986)
The quote pithily expresses the main advantage of stochastic methods - they

141
142 CHAPTER 8. STOCHASTIC METHODS

allow us to work out useful stuff without requiring formal equations. Note that
the methods involving random numbers we discuss in this chapter are distinct
from the concept of a random effect, that we introduced in Chapters 6 and 7.
Random effects are where individuals, groups or studies differ in their means,
whereas here we use random numbers for several other purposes.

8.2 Ways of generating random numbers

Almost everyone who needs random numbers for something will generate them
using a computer. However computers are the opposite of random - they are
deterministic machines. This means that for a given input, they should always
produce the same output. For most applications, like calculating your tax return
or running a statistical test, we want them to behave in this way. It would cause
utter chaos if such calculations involved a stochastic component: you’d get a
different answer every time! But when we do want random numbers, we need
some way to create them.
The most common way, and the method used by R, is to use a complex algorithm
to generate sequences of numbers that appear random to a first approximation.
These are called pseudo-random numbers, and they are predictable if you run
the same algorithm again with the same starting conditions. The algorithm
takes as an input a number called a seed, and changing the value of the seed will
dramatically affect the output. Even two consecutive seed values will produce
completely different sequences of numbers. It is common practice to use the
current time from the computer’s clock as the seed, meaning that the same two
sequences of pseudo-random numbers should never reoccur. But, if we were to
store the seed, and use it again in the future, we will get exactly the same set of
numbers out.
Pseudo-random numbers are absolutely fine for all of the techniques we will
discuss here. But they do have clear weaknesses for any applications where
privacy or security is important. A good example is online gambling websites. If
these used a pseudo-random number generator based on the computer’s clock,
and we knew the algorithm, we could predict in advance what cards would
be drawn in an online poker game, or what numbers would be picked by a
simulated roulette wheel, and perhaps win large amounts of money fraudulently.
Apparently this has actually happened in the past, though gambling companies
have since wised up and now mostly use something called a true random number
generator.
Truly random numbers require some sort of input from the physical world. This
might come from the weather, the decay of a radioactive material, temperature
fluctuations in a piece of metal, and so on. Hardware true random number
generators using this sort of information are commercially available, now usually
as small USB devices. There are also online services, such as the website
www.random.org, that will generate truly random numbers for you. However,
8.3. PART 1: USING RANDOM NUMBERS TO FIND STUFF OUT 143

these are not necessary for anything we will discuss here, and are only worth the
bother and expense if you really need them.
The remainder of this chapter is divided into two parts. In part 1, we will
describe how stochastic methods can be used to model different situations, in
order to gain insights into how a system, model, or experiment might behave. In
part 2, we will introduce the concept of resampling. This is a way of analysing
data that can be used to estimate confidence intervals, and also to conduct
statistical hypothesis testing.

8.3 Part 1: Using random numbers to find stuff

out
In the spirit of the Press quote above, let’s do some simulations using random
numbers to demonstrate their usefulness. I have included the R code here to
show that there’s no sleight of hand or funny business going on - you can run it
yourself if you want to. Imagine that we have two sets of uniformly distributed
random numbers (where each sample has an equal probability of taking on any
value between 0 and 1), as shown in Figure 8.1:
par(mfrow=c(1,2), las=1)

a <- runif(100000)
b <- runif(100000)
hist(a, breaks = 100, col = 'white')
hist(b, breaks = 100, col = '#8783CF')

Histogram of a Histogram of b

1000 1000

800 800
Frequency

Frequency

600 600

400 400

200 200

0 0

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

a b

Figure 8.1: Two populations of uniformly distributed random numbers.

144 CHAPTER 8. STOCHASTIC METHODS

What would we expect the distribution to look like if we added these two samples
together? Intuitively, we might guess that the distribution of the summed values
should also be uniform. However this intuition would be incorrect. In fact, the
summed distribution has a clear peak in the centre, as shown in Figure 8.2,
generated by the following code:
hist(a+b, breaks = 100, col = '#CFCDEC')

Histogram of a + b
2000
1500
Frequency

1000
500
0

0.0 0.5 1.0 1.5 2.0

a+b

Figure 8.2: The sum of two populations of uniformly distributed random numbers.

Why does this happen? It is a consequence of something called Central Limit

Theorem, which is a mathematical theorem that states that the sum of several
non-normal distributions will tend to approximate a normal distribution. The
reason for this is that it is very unlikely that the two biggest numbers, or the
two smallest numbers, of the two samples will be paired together (of course, this
assumes that the two sets of numbers are not sorted before being added). It is
8.3. PART 1: USING RANDOM NUMBERS TO FIND STUFF OUT 145

much more likely that the summed numbers will be of middling value. If we
kept on summing lots and lots of uniform distributions, we would eventually end
up with a beautiful normal distribution, as shown in Figure 8.3, and generated
using the following code:
bigsum <- runif(100000)
for (n in 1:99){bigsum <- bigsum + runif(100000)}
hist(bigsum, breaks = 100, col = 'grey')

Histogram of bigsum
1000 1500 2000 2500
Frequency

500
0

40 45 50 55 60

bigsum

Figure 8.3: The sum of 100 populations of uniformly distributed random numbers.

With this simple example, using random numbers, we have demonstrated Central
Limit Theorem in action. Of course, there are a whole load of complex mathe-
matical equations that explain how it works in detail (see e.g. the Wikipedia
entry on Central Limit Theorem). But, in keeping with the spirit of the quote
at the start of the chapter, we have been able to show that the theorem works
146 CHAPTER 8. STOCHASTIC METHODS

without needing to think about the underlying mathematics.

Incidentally, Central Limit Theorem explains why the assumption of normality
holds for so many variables in scientific research. If a dependent variable is
determined by a combination of underlying factors, it will tend to be normally
distributed even if the underlying factors are not. Consider a physiological
variable, like heart rate. For each individual, this will be determined by factors
such as current arousal, fitness level, age, sex, and various genetic influences.
Across a population of individuals, the distribution of heart rates should be
approximately normal because central limit theorem says that all of these different
influences will combine at random. The same is likely to be true of high level
psychological constructs (like IQ), brain activity measured using neuroimaging
techniques, pollen yield of flowers, phenotypes dependent on many separate
genes, and so on.

8.3.1 Stochastic simulations: models and synthetic data

If stochastic simulations can be used to find stuff out, what sorts of things
might we use them for? One common application is to model systems which are
themselves stochastic - a key example being the human brain. Many dynamic
neural models incorporate a stochastic component, which often helps us to better
understand how the brain works.
A recent example from my own research is a study in which my colleague Bruno
Richard and I modelled perception in binocular rivalry (Baker and Richard
2019). Rivalry is a curious phenomenon where our percept of conflicting stimuli
shown to the left and right eyes fluctuates over time in an unpredictable way.
Since the process itself is stochastic, we needed to use a model containing a
random component to properly understand it. In the paper, we were particularly
concerned with working out the magnitude and characteristics of the neural
noise (i.e. inside the participant’s brain) that governs rivalry alternations. We
approached this by adding different amounts of dynamic noise to the contrasts
of the rivalling stimuli in a psychophysical experiment to provide a rich data set
with many conditions. We then attempted to reproduce the pattern of human
data using models with different types and amounts of internal noise. The
key point is that this analysis could probably not have been done analytically
(i.e. with noise-free equations) - we needed to be able to use stochastic methods
to simulate human perception convincingly.
Another related and important use of stochastic methods is to simulate what
we think might happen in an experiment before we collect the data. This
is a form of modelling, but one that is prospective rather than retrospective.
For simple designs, involving only one or two variables, it is not necessarily
very informative. However for complex factorial designs with many interacting
variables (see Chapter 7 on mixed effects models, and Chapter 12 on Structural
Equation Modelling) data simulation can reveal how interdependencies might
affect the results. This can help researchers to develop their intuitions about how
8.3. PART 1: USING RANDOM NUMBERS TO FIND STUFF OUT 147

the results of an experiment might turn out. Power calculations (see Chapter 5)
can also be done by simulation, affording greater flexibility (e.g. for complex or
unbalanced deisgns) than analytic approximations (Colegrave and Ruxton 2020).
An added advantage of simulating data is that one can construct an analysis
pipeline in advance of running the experiment. This saves time later, is useful
for clarifying and making explicit one’s assumptions and expectations, and can
also be included in preregistration materials.

8.3.2 Generating random numbers in R from different dis-

tributions
The stats package in R provides access to a number of different statistical
distributions that can be used to generate random numbers. A full list is
available by typing help(Distributions). Each distribution has four functions,
that are shown in Figure 8.4 for a normal distribution. The following code chunk
demonstrates their use:
par(mfrow=c(2,2), las=1) # divide the plot into four panels

hist(rnorm(1000, mean=0, sd=1), breaks = 50,

main='rnorm',xlab='x',ylab='Frequency',xlim=c(-4,4))

plot(seq(-4,4,0.001),dnorm(seq(-4,4,0.001),mean=0,sd=1),
type='l',lwd=3, main='dnorm',xlab='x',ylab='Density')

plot(seq(-4,4,0.001),pnorm(seq(-4,4,0.001),mean=0,sd=1),
type='l',lwd=3, main='pnorm',xlab='x',ylab='Cumulative probability')

plot(seq(0,1,0.001),qnorm(seq(0,1,0.001),mean=0,sd=1),
type='l',lwd=3, main='qnorm',xlab='Quantile',ylab='x')

The top left panel of Figure 8.4 shows the output of the rnorm function, which
generates a sequence of n random numbers drawn from a normal distribution,
with mean and standard deviation defined by the function call (defaults are
mean = 0 and sd = 1). The rnorm function is the most useful function for our
current purposes, but for reference we will also describe the outputs of the other
three related functions.
In the top right panel of Figure 8.4, the dnorm function produces a probability
density plot for the same normal distribution. This gives the probability of
drawing a number with value x from a normal distribution with the mean and
standard deviation specified by the function call. Note that this function does
not produce random numbers directly, but it has many uses, such as plotting
smooth curves to summarise distributions.
In the lower left panel of Figure 8.4, the pnorm function provides the cumulative
distribution function. A good way to understand this is to imagine that at each
148 CHAPTER 8. STOCHASTIC METHODS

rnorm dnorm

0.4
40

0.3
30
Frequency

Density
0.2
20

10 0.1

0 0.0

−4 −2 0 2 4 −4 −2 0 2 4

x x

pnorm qnorm

1.0 3

0.8 2
Cumulative probability

1
0.6
0
x

0.4
−1

0.2 −2

0.0 −3

−4 −2 0 2 4 0.0 0.2 0.4 0.6 0.8 1.0

x Quantile

Figure 8.4: Example distribution function outputs for random numbers, density
function, cumulative density and quantiles, for a normal distribution with mean
= 0 and sd = 1.
8.3. PART 1: USING RANDOM NUMBERS TO FIND STUFF OUT 149

value of x on the curve, you are adding up the probabilities for every number
between −∞ and x. For any input value of x, this function will tell you the
probability that a number drawn from a normal distribution will have a value
smaller than x. Equivalently, subtracting the probability from 1 will tell you
the probability of a number having a value larger than x. This is particularly
important for calculating p-values in statistical testing. For example, if we run
an ANOVA and calculate an F-ratio, we compare this to the (inverse) cumulative
F distribution (from the pf function) with an appropriate number of degrees of
freedom. This provides the ubiquitous p-value that is used to determine if a test
is statistically significant.

Finally, in the lower right panel of Figure 8.4, the qnorm function provides
quantiles from the normal distribution. This is the reverse of the cumulative
distribution - notice that the x and y axes are switched between the lower two
panels. So it can be used to reverse engineer a test statistic if we know the
p-value. This is used in some of the effect size conversion tools we discussed in
section 6.5.

Equivalent functions are available for other distributions with a consistent naming
pattern. For example the rgamma, dgamma, pgamma and qgamma functions
generate a gamma distribution. This has a positive skew, and is sometimes
used for modelling prior distributions in Bayesian statistics. Other particularly
useful distributions include the uniform distribution (runif, dunif, punif and
qunif ) we encountered in the central limit theorem example above, the log-normal
distribution (rlognorm, dlognorm, plognorm and qlognorm), the F distribution (rf,
df, pf and qf ) used in ANOVA and related statistics, and the Poisson distribution
(rpois, dpois, ppois, qpois) that is used to model event probabilities such as the
spiking of neurons.

All of the above functions use the same underlying random number generator.
We can set the seed to a specific (integer) value, and be confident that the
sequence of pseudo-random numbers we generate will always be the same. For
example, setting the seed to 100 and asking for 5 random numbers from a normal
distribution produces the following output:
set.seed(100)
rnorm(5)

## [1] -0.50219235 0.13153117 -0.07891709 0.88678481 0.11697127

If we change the seed to a different value (99), we will get a different sequence:
set.seed(99)
rnorm(5)

## [1] 0.2139625 0.4796581 0.0878287 0.4438585 -0.3628379

Crucially, if we set the seed back to 100, we should get our first sequence of
numbers out again:
150 CHAPTER 8. STOCHASTIC METHODS

set.seed(100)
rnorm(5)

## [1] -0.50219235 0.13153117 -0.07891709 0.88678481 0.11697127

If the user does not specify a seed, then one is generated automatically based on
the computer clock the first time any functions involving random numbers are
called. If it is important to reproduce the random seed again in future, it can be
saved to a data object as follows:
seed <- .Random.seed

This needs to be done before any random numbers are actually generated, as
each time we sample from the random number generator we change its state.
The seed can then be restored by setting .Random.seed <- seed, which should
permit full reproducibility of the original random sequence.
The set.seed function can also be used to specify the random number generator
algorithm to be used (with the kind argument). The default is the exciting-
sounding Mersenne-Twister (Matsumoto and Nishimura 1998), which is a widely-
used algorithm implemented in a number of programming languages and software
packages (including SPSS). There are half a dozen alternatives with similarly
exotic names, and also the option for users to specify their own algorithms if
required.

8.4 Part 2: Resampling methods

A second widespread application of stochastic techniques is an approach known as
resampling. This is a class of methods that is widely used to estimate confidence
intervals and other measures of precision, and can also be used to test statistical
hypotheses. Resampling is generally a nonparametric method that can be used
on any type of data, and makes use of random sampling. The basic idea is that we
take some data and perform a statistic or calculation upon it. A simple example
would be to calculate the mean, but any method that produces a summary
statistic that we are interested in is suitable: other examples might include a
t-statistic or correlation coefficient, or the parameters of a fitted model. Next, we
repeatedly resample the original data by drawing randomly chosen values from it.
With each set of resampled data we calculate the same test statistic. Over many
iterations, we build up a population of resampled test statistics (e.g. resampled
means, or t-values, or correlation coefficients).
This resampled distribution has several uses. One thing we can do is to take
the confidence intervals of the population, usually at the upper and lower 2.5%
of the distribution, as these points will provide 95% bootstrapped confidence
intervals. Confidence intervals can be used as error bars, giving an indication of
the precision of our original estimate. We could also compare the distribution
to a particular point, such as the mean of another condition, or a suitable
8.4. PART 2: RESAMPLING METHODS 151

benchmark (such as a value of 0). The technique is often called bootstrapping as

a reference to the phrase ‘pulling yourself up by your bootstraps’, which means
to improve your situation by your own efforts. In the context of data analysis,
the idea is that we derive additional information directly from the data itself
without making further observations or assumptions.
The confidence intervals are a statement about the likelihood that the true
population value falls within a particular range. When used with parametric
data to calculate the confidence intervals of a mean, bootstrapping produces
very similar estimates to analytic approximations (95% confidence intervals are
well approximated by 1.96*SE when parametric assumptions hold). This is
shown in Figure 8.5a. The grey histogram bars show a sample of data derived
from an underlying normal distribution (black curve). The white point and
black error bars show the mean and analytic approximation of 95% confidence
intervals. The blue shaded region shows the distribution of bootstrapped means,
calculated by resampling the data 10,000 times. The blue error whiskers give
the 2.5% and 97.5% points of this distribution, which are the bootstrapped 95%
confidence intervals of the mean. Evidently, the error bars on the two points
are very similar, showing that for parametric data the two methods are equally
useful.
The usefulness of bootstrapping becomes more apparent when we consider non-
parametric data, such as the skewed data shown in Figure 8.5b. Here the
appropriate measure of central tendency is the median, and approximating the
confidence intervals from the standard error does not make sense (the error bars
on the white point are symmetrical for a start, which must be wrong given the
skewness of the data). By bootstrapping, we produce more plausible estimates
of confidence about where the true (population) median might lie (error bars on
the blue point). These are asymmetrical, consistent with the skew in the data.

8.4.1 Resampling with and without replacement

There are two varieties of resampling that can potentially be used to bootstrap.
Resampling without replacement means that on any given bootstrapping iteration
(repeat), each value from the original data set can be included in the resampled
data set only once. This type of bootstrapping only makes sense if the resampled
data set is smaller than the original data set - otherwise we would just end
up with thousands of identical data sets! Reducing the sample size (known as
subsampling) can cause problems in some situations because it reduces statistical
power (see Chapter 5), for example if a statistical test is being conducted.
However there are some situations where sampling without replacement is the
right thing to do. Lottery draws are a good example: if we ended up drawing
the same number twice it would cause a lot of confusion.
The second type of resampling, which is more widely used, is resampling with
replacement. Here, each value from the original data set can be included in the
resampled data set more than once. Assuming the resampled data set is the
152 CHAPTER 8. STOCHASTIC METHODS

4
(a)
3
●
2

●
1

−10 −5 0 5 10

8
(b)
6
●
4

●
2

0 10 20 30 40

Figure 8.5: Examples of bootstrapping confidence intervals on measures of

central tendency. Panel (a) shows a sample of data (histogram) from a normal
distribution (black curve). The white point and error bars show the mean and
analytically derived confidence intervals (1.96*SE). The blue point and error
bars show the mean and 95% confidence intervals derived by bootstrapping. The
blue distribution is the distribution of bootstrapped means. Panel (b) has the
same format, but the data are from a gamma distribution, and the measure of
central tendency is the median rather than the mean.
8.4. PART 2: RESAMPLING METHODS 153

same size as the original data set, this also means that some values from the
original set might not be included at all in the resampled data. I always think of
bootstrapping using the analogy of a bag of ping pong balls, each of which has a
number from the original data set written on it. Resampling involves drawing a
ball from the bag, and noting down the number. In resampling with replacement,
the ball then goes back in the bag, meaning there is the possibility it will be
pulled out again. In resampling without replacement, the ball stays out of the
bag until the end of this bootstrap iteration.

8.4.2 Using resampling for hypothesis testing

It is also possible to use resampling to test statistical hypotheses. The basic
idea is that we build up a population of resampled test statistics to represent
the null hypothesis, and compare the test statistic from the original data to
this distribution. Several specific variants have been proposed, but the general
approach for a two-sample test is to randomize the group assignment of each
data point on every iteration. This will generate a distribution of resampled
test statistics that we might expect to observe if there were no true effect - in
other words a null distribution. The null distribution is typically centred on 0,
with a spread determined by the characteristics of the data. If the original test
statistic lies close to the middle of this distribution then it is unlikely to indicate
a real effect. On the other hand if the original test statistic is in one of the tails
of the distribution, it is more likely to indicate a true effect. This is because if
the original test statistic is extreme, it means that group membership matters,
and implies that there is a genuine difference between the groups. Figure 8.6
illustrates this procedure.

Once the null distribution has been generated, we can calculate a p-value by
determining the proportion of resampled test statistics that are more extreme
than the original test statistic. For a one-sided test this is the proportion of
resampled statistics that are either smaller than or larger than the original
statistic. For a two-sided test, the absolute values are used instead. Resampling
approaches are inherently non-parametric, so can be used in situations where
the assumptions of more traditional parametric statistics are not met. Crucially,
this method works for any test statistic one might come up with, even if the
expected distribution is unknown. Some variants include the bootstrap test, in
which data are resampled with replacement, and the permutation test, in which
all possible permutations (i.e. combinations of group ordering) of the data are
included in the resampled distribution. A similar approach can also be taken
with correlations by randomly reshuffling the pairings of the two dependent
variables on each iteration (see below). For a more elaborate use of resampling
methods, see section 15.8, which describes a related method for controlling for
multiple comparisons.
154 CHAPTER 8. STOCHASTIC METHODS

A B A' B'
60 94 58 50
Resampled data

73 60 62 58
Original data

87 86 97 60
t = −0.11
62 18 25 86
t = 1.97

99 50 62 87
97 44 18 60
58 83 94 83
62 29 73 99
93 25 93 29
58 62 62 44

−6 −4 −2 0 2 4 6
t−statistic

Figure 8.6: Illustration of the bootstrap test. The original data (left) consists
of two groups, A and B (columns), which produce a t-statistic of 1.97. These
data are resampled by randomly reshuffling the group allocations, and a new
t-statistic is calculated (here -0.11) using the resampled groups, A’ and B’ (blue
squares indicate values that originated in group A). The distribution of resampled
t-statistics from 10,000 such resampling iterations is shown in the right hand
plot, along with the original t-statistic (blue line). For this example, 3.4% of
the population lies to the right of the blue line, implying a one-sided p-value
of 0.034, or a two-sided p-value of 0.069. This is close to the value from the
original t-test of p = 0.067.
8.4. PART 2: RESAMPLING METHODS 155

8.4.3 How to do resampling in R

Resampling is a sufficiently useful technique that it is now built into some
commercial statistics packages (including SPSS) for some tests. In R there is a
package and a function that are both called boot that can be used to bootstrap
other functions. However in the interests of making the stages of bootstrapping
explicit, the examples here use two built-in functions: sample and quantile. The
sample function resamples a data set. If we resample the values from 1 to 10 with
replacement, we generally get a set of numbers that includes several duplicates,
and also several missing values:
sample(1:10,10,replace=TRUE)

## [1] 7 10 5 3 5 8 9 6 5 9
If we resample without replacement, we get a random permutation of the numbers,
which is useful in some situations (e.g. for randomising the order of conditions
in an experiment).
sample(1:10,10,replace=FALSE)

## [1] 10 3 9 4 6 5 2 7 1 8
Finally, we can also resample either with or without replacement but produce
a smaller data set by specifying how many values we need with the second
argument to the sample function. This is known as subsampling:
sample(1:10,5,replace=FALSE)

## [1] 8 6 7 1 3
The sample function becomes particularly useful when it is embedded in a loop
(see section 2.10) that repeats an operation many times on the resampled data.
The following code resamples the mean of some data, and plots the distribution
of resampled means in Figure 8.7:
# generate some random synthetic data
data <- rnorm(100, mean=1, sd=3)

# create an empty data object to store the resampled means

allmeans <- NULL
# repeat the resampling lots of times
for (n in 1:10000){
# use the sample function to resample the data, and store in rsdata
rsdata <- sample(data,replace=TRUE)
# calculate the mean of the resampled data
allmeans[n] <- mean(rsdata)
}

# plot a histogram of the resampled means

156 CHAPTER 8. STOCHASTIC METHODS

b <- hist(allmeans,breaks=20)
# add a vertical line showing the true mean
lines(c(mean(data),mean(data)),c(0,max(b$counts)),col='black',lwd=8,lty=2)

Histogram of allmeans
1000
Frequency

600
200
0

−1.0 −0.5 0.0 0.5 1.0 1.5

allmeans

Figure 8.7: Histogram of bootstrapped means. The black dashed line is the true
mean, and the dotted lines are the 95% confidence intervals.

The data object allmeans now contains 10,000 bootstrapped means. We can
estimate the confidence intervals from this population using the quantile function.
This function returns values at a specific proportion of a distribution. To get the
95% confidence intervals, we request proportions of 0.025 for the lower bound,
and 0.975 for the upper bound, because 95% of the values will lie between these
points.
# use the quantile function to get the confidence intervals
# from the population of bootstrapped means
CIs <- quantile(allmeans, c(0.025,0.975))
CIs

## 2.5% 97.5%
## 0.2493662 1.4879034

We can add the limits to our histogram (vertical dotted lines) to visualise them
as follows:
8.4. PART 2: RESAMPLING METHODS 157

# add vertical lines showing the confidence intervals

lines(CIs[c(1,1)],c(0,max(b$counts)/2),lty=3,lwd=4)
lines(CIs[c(2,2)],c(0,max(b$counts)/2),lty=3,lwd=4)

These upper and lower confidence intervals can then be used to plot error bars
for the mean in other figures. Of course, we are not limited to bootstrapping the
mean. We can bootstrap any test we are interested in, and obtain confidence
intervals on the test statistic. For example, we could bootstrap confidence
intervals on the t-statistic of a one-sample t-test using the same data (see Figure
8.8).
maint <- t.test(data,mu=0) # calculate a t-statistic instead of a mean

allT <- NULL

for (n in 1:10000){
allT[n] <- t.test(sample(data,replace=TRUE),mu=0)$statistic}
CIs <- quantile(allT,c(0.025,0.975))

b <- hist(allT,breaks=20)
lines(c(maint$statistic,maint$statistic),c(0,max(b$counts)),lty=2,lwd=8)
lines(CIs[c(1,1)],c(0,max(b$counts)/2),lty=3,lwd=4)
lines(CIs[c(2,2)],c(0,max(b$counts)/2),lty=3,lwd=4)

Histogram of allT
1500
Frequency

1000
500
0

−2 0 2 4 6

allT

Figure 8.8: Distribution of resampled t-statistics, showing the true mean (dashed
line) and 95% confidence intervals (dotted lines).
158 CHAPTER 8. STOCHASTIC METHODS

These confidence intervals could be used to compare t-statistics across different

data sets (e.g. in meta-analysis, see Chapter 6), or to compare the test statistic
to a fixed point (e.g. to zero). One can also produce confidence intervals on
p-values, effect size measures, and any other statistic one might be interested
in. The key thing to remember is that we repeat whatever calculation we are
interested in many times on sets of randomly resampled data.

Finally, let’s conduct a bootstrap test on some weakly correlated data. We’ll
generate these ourselves so that we have control over the extent of the correlation:
# generate a vector of 50 random values
var1 <- rnorm(50)
# generate a vector of 50 values that includes a fraction of var1
var2 <- rnorm(50) + 0.25*var1

# calculate the correlation coefficient for these two vectors

truecor <- cor(var1,var2)
truecor

## [1] 0.2186587

To generate a null distribution, we need to randomly re-order both of the vectors

to destroy the correspondence between them (so that the value in row 1 of var1
is unlikely to be paired with the value in row 1 of var2, and so on). We can do
this using the sample function as follows:
nullR <- NULL
for (n in 1:10000){
var1r <- sample(var1,50,replace=TRUE)
var2r <- sample(var2,50,replace=TRUE)
nullR[n] <- cor(var1r, var2r)
}

The data object nullR now contains the null distribution of 10,000 resampled
correlation coefficients. We can calculate a one-sided p-value by working out
the proportion of this distribution that is larger than our original correlation
coefficient (stored in the data object truecor):
length(which(nullR>truecor))/10000

## [1] 0.062

The which function here returns the indices of any entries in nullR that are larger
than the value of truecor. Then the length function counts how many indices
have been returned by the which function; this is converted to a proportion by
dividing by the number of resampling iterations (10,000). Finally, it is worth
visualising both the null distribution and the original correlation coefficient, as
shown in Figure 8.9.
8.5. FURTHER READING 159

hist(nullR,breaks=20)
lines(c(truecor,truecor),c(0,1200),col=pal2tone[1],lwd=6)

Histogram of nullR
1000
Frequency

600
200
0

−0.4 −0.2 0.0 0.2 0.4

nullR

Figure 8.9: Null distribution from a bootstrap test of a correlation. Each corre-
lation coefficient in the distribution is calculated using independently resampled
data from each variable. The vertical blue line shows the correlation coefficient
from the original data.

Notice that the number of resampling iterations determines the precision of the
resulting p-value. Running 10,000 iterations gives us a precision of 1/10,000 =
0.0001. If we only ran 100 iterations, we would only have a precision of 1/100 =
0.01. Values smaller than this will default to 0.

8.5 Further reading

If you want to read more about stochastic methods, a very comprehensive source
is the book called An Introduction to the Bootstrap by Efron and Tibshirani
(1993). Bradley Efron essentially invented the technique of bootstrapping in
the late 1970s, which had an enormous impact on the field, and this book is
the definitive source on the method. Alternatively, for a tutorial-style paper on
resampling methods with ecology-related examples in R, see Fieberg, Vitense,
and Johnson (2020). For a general example of data simulation, the book on
power simulations by Colegrave and Ruxton (2020) is a useful resource. If you
160 CHAPTER 8. STOCHASTIC METHODS

need to simulate a particular type of data, there may well be specific R packages
and online tutorials designed with this in mind.

8.6 Practice questions

1. Pseudo-random numbers can be generated by:
A) Radioactive decay
B) Weather patterns
C) The computer’s clock
D) A mathematical algorithm
2. According to Central Limit Theorem, the sum of many different distribu-
tions will be approximately:
A) Normal
B) Uniform
C) Positively skewed
D) Triangular
3. A sequence of pseudo-random numbers can be re-created if we know:
A) The date the numbers original numbers were generated
B) The seed value
C) The kernel number
D) The first number in the sequence
4. Bootstrapping methods allow us to calculate:
A) The mean or median value of a test statistic
B) A t-statistic for any data set
C) Confidence intervals on any statistic
D) Whether a distribution is normal or not
5. Stochastic simulations are NOT useful for:
A) Understanding dynamic systems such as the brain
B) Modelling complex processes without using analytic equations
C) Checking whether a sequence of numbers is truly random
D) Testing our expectations about statistical power
6. Use the rgamma function to generate 100000 random numbers from a
gamma distribution with shape and scale parameters of 2. The median is
approximately:
A) 3.35
B) 2.85
C) 4.52
D) 1.11
7. Which R function is used to extract confidence intervals from a population
of bootstrapped values?
A) qnorm
B) hist
C) mean
D) quantile
8. 95% confidence intervals of a population indicate:
8.6. PRACTICE QUESTIONS 161

A) The points which 97.5% of values lie between

B) The points which 95% of values lie between
C) The points which 1.96% of values lie between
D) The points which lie beyond an alpha level of 0.05
9. Which sequence of numbers could have been resampled without replacement
from the following set? [5, 3, 7, 12, 15]
A) 7 15 3 12 5
B) 3 15 5 3 12
C) 5 3 15 12 8
D) 12 15 3 12 5
10. Use the rpois and hist functions to generate and plot 10000 samples from
a Poisson distribution with a lambda value of 2. The histogram is best
described as:
A) Bimodal
B) Negatively skewed
C) Positively skewed
D) Symmetrical
Answers to all questions are provided in section 20.2.
162 CHAPTER 8. STOCHASTIC METHODS
Chapter 9

Nonlinear curve fitting

9.1 Fitting models to data

Real empirical data can be quite complicated. Using a model of some sort can
often help us to make sense of our data. This might allow us to reduce the
complexity of the data by summarising it using one or more parameter estimates.
Or, it might help us to understand the processes that led to the data being
produced. Actually, many familiar statistical tests are based on fitting linear
models (straight lines) to data, including linear regression and ANOVA (see
Chapter 4). But sometimes our data do not involve linear trends, and we need
a more elaborate model to describe them. In this context, a model might be
a mathematical equation that can predict values of a dependent variable for
given levels of an independent variable. Such models will usually have some
parameters (often called coefficients). These are numbers in the equation that
control the model’s behaviour, and which we can alter to try and improve the fit
to the data. But how can we find appropriate values of these parameters that
give the best description of our data?

The general problem here is called parameter optimization, and there is a whole
class of computational techniques designed to solve it. This chapter will discuss
some of the issues involved in fitting models to data, and introduce a well-
established optimization algorithm called the Downhill Simplex Algorithm. We
will first outline the idea of linear and nonlinear models, and how to calculate the
error of a model fit. Next we will introduce the possible parameter space for a
model, and discuss how this depends on the number of parameters in the model.
Then we will introduce the simplex algorithm, and describe some problems it
can encounter during optimization. Finally we will go through an illustrative
example of function fitting in R.

163
164 CHAPTER 9. NONLINEAR CURVE FITTING

9.2 Linear models

Most readers will be familiar with the idea of linear regression. In regression,
we aim to fit a straight line to a data set, such that we can predict how our
dependent variable (i.e. the thing we have measured) changes as a function of
another variable. An intuitive example of this might be how height increases as
a function of age during childhood, as shown in Figure 9.1.
90

●
85
80

●
Height (cm)

●
75

●
70

●
65

●
60

10 15 20

Age (months)

Figure 9.1: Straight line fit to age vs height data.

To the extent that height will continue to increase approximately linearly with
age, we could use the fitted line to predict how tall this particular child might
be in another 3 months. In regression notation, the equation of a straight line is:

y = β0 + β1 x, (9.1)

where the β1 parameter determines the slope (gradient) of the line, and the β0
parameter is a vertical offset that determines the value of y when x = 0 (often
called the y-intercept). Performing regression involves finding the values of the
two parameters (β0 and β1 ) that give the best description of the data. For the
above example this turns out to be β0 = 55.4 and β1 = 1.2. The slope value is
telling us that every month the child grows another 1.2 cm on average.
For real data, there will always be some amount of error between the straight
line fit and the data points (shown by the thin vertical lines in Figure 9.1). One
9.3. NONLINEAR MODELS 165

way of thinking about fitting is that we are trying to make this error as small as
we possibly can. If the parameter estimates were completely wrong this would
give a very poor fit. For example if the slope parameter were negative, the model
(thick blue line) would predict that children should shrink as they age! It follows
that the best fitting parameter values (of β0 and β1 ) are the ones that produce
the smallest error between model and data.

9.3 Nonlinear models

Sometimes data are not best described by a straight line. In principle we could
fit some other equation to these data. For example, if the data followed a square
law, we could adapt our equation as follows:

y = β0 + β1 x2 , (9.2)

This transforms our straight line into a curve, as shown in Figure 9.2.
90

●
85
80

●
Height (cm)

●
75

●
70

●
65

●
60

10 15 20

Age (months)

Figure 9.2: Quadratic model fit to age vs height data (a quadratic function is
one that involves squaring).

The quadratic curve gives a slightly better fit to the data (the vertical lines are
shorter than before). More generally, the exponent might not be exactly 2, and
so its value could become another free parameter in the equation, much like β0
and β1 :
166 CHAPTER 9. NONLINEAR CURVE FITTING

y = β0 + β1 xγ , (9.3)

This means that when we fit the model to our data, we would have three different
parameters to adjust (β0 , β1 and γ) instead of just two. In fact, we could in
principle fit any equation to any set of data if we had reason to do so, and these
equations would have as many parameters as they might need. As we will see
later in the chapter, models with a large number of parameters quickly become
very difficult to fit.

9.4 A practical example: exponential modelling

of disease contagion
This book was mostly written during 2020, when the world experienced a
pandemic outbreak of a novel coronavirus. At the time of writing, the news
is full of graphs showing exponential increase in cases and deaths. A critical
question during the early weeks was how quickly the disease would spread within
different countries. One way in which this can be predicted is to fit nonlinear
growth curve models to the existing data for a country, and try to extrapolate
forward into the future. Figure 9.3 shows some example data for the United
Kingdom, with two different fitted curves. The dashed black curve is a fit to
the first 30 days of data (indicated by the arrow). It provides a good fit to
the data up to this point, but substantially overestimates the increase in cases
afterwards. The curve shown in blue is a fit to all 44 days of data, which gives a
much better overall description, despite overshooting slightly between days 20
and 30. The data on cases were aggregated by researchers at Johns Hopkins
University (Dong, Du, and Gardner 2020), and made publicly available online
(https://github.com/CSSEGISandData/COVID-19).

This example illustrates two important points. First, that fitting curves to data
is an important research skill that can factor into critical life or death decision
making at the highest levels. Models of infection were used to guide government
policy about how to control the virus. Second, that extrapolating from curves
fitted to limited data can be extremely misleading - the dashed curve in Figure
9.3 does not give accurate future predictions, and basing important decisions
on it at the end of March would have been a bad idea. Of course most of
the coronavirus modelling was rather more sophisticated than a 3-parameter
exponential function, but the same caveats apply no matter how elaborate the
model.
9.4. A PRACTICAL EXAMPLE: EXPONENTIAL MODELLING OF DISEASE CONTAGION167

Fit first 30 days

100

Fit all data

Thousands of cases

●
●
80

●
●
●
60

●
●
●
●
40

●
●
●
●
●
●
20

●●
●
●●
●●●
●●●●
●●●●●●●●●●●●●●●●●
0

0 10 20 30 40 50

Days from 1st March 2020

Figure 9.3: UK coronavirus cases for a 6 week period in early 2020. The dashed
black curve was fit to the first 30 days of data, and predicts the following two
weeks. The blue curve was fit to the full data set.
168 CHAPTER 9. NONLINEAR CURVE FITTING

9.5 Parameter spaces and the combinatorial ex-

plosion
When we fit a straight line with two free parameters, it is feasible to work out
how good the fit is for all possible combinations of these parameters within a
reasonable range (and with some specified resolution). For the example in Figure
9.1, we could work out how good the fit is for all intercept values from β0 = 0
to β0 = 100, and all slope values from β1 = -2 to β1 = 2. A useful measure of
goodness of fit is the root mean squared (RMS) error. This involves taking the
lengths of all the thin vertical lines in Figure 9.1, squaring them and then taking
the square root of their mean (average). We could then plot the error for each
combination of β0 and β1 as shown in Figure 9.4.

100 20 30 40 50

10
80
Intercept (β0)

60

40
20
20
30

40
100 90 80 70 60 50
0

Slope (β1)
−2 −1 0 1 2

Figure 9.4: Parameter space for a linear model fit. The star indicates the best
fitting parameters, which give the smallest error. Blue shading indicates the
depth of the surface.

This visualisation tells us that the region of the parameter space that gives the
best fit is somewhere around β1 = 1 and β0 = 60, which corresponds well to our
original parameter estimates from the regression fit (shown by the black star).
The parameters that give the best fit are those that produce the smallest error
between the line and the data points, and so this is the lowest point in a virtual
three dimensional ‘space’ consisting of one dimension for each parameter (β0
and β1 ) and a third dimension for the value of the error (the height, indicated
by the contours and shading). This error surface will be different for every data
9.6. OPTIMIZATION ALGORITHMS 169

set, and for each model equation we might attempt to fit. Note that although in
this chapter we plot several error surfaces, we would not typically visualise them,
as the computational cost of evaluating the model for all possible parameter
combinations is usually too great.

If we only have two free parameters, testing all plausible combinations of pa-
rameter values is possible on a modern computer (assuming some sensible level
of sampling resolution). But as we add more free parameters, the amount
of time required to do this will increase exponentially. This is known as the
combinatorial explosion - the rapid growth in the complexity of a problem as
more dimensions are added. Also, the error surface will have more and more
dimensions (always n+1, where n is the number of free parameters, and the extra
dimension represents the error between the model and the data). Spaces with
more than three dimensions are pretty much impossible to represent graphically
or to imagine in our dimensionally-limited brains. Clearly, we need an algorithm
to do this for us.

9.6 Optimization algorithms

A class of computer algorithms exist that aim to find the lowest point on a
multidimensional error surface without sampling every possible combination of
parameters. These optimization algorithms are used in most areas of scientific
research, as well as in many other fields, most notably economics and finance.
Most algorithms involve iterating many times through some basic operations
that are intended to improve the fit of the model. When the algorithm believes it
has found the lowest point on the error surface, it returns the parameter values
that give this best fit. Algorithms are often inspired by real-world processes,
such as the cooling of metal, or the swarming of insects. Examples include:

• Gradient descent (Curry 1944)

• Simulated annealing (Pincus 1970)
• Genetic algorithms (Holland 1992)
• Particle swarm optimization (Kennedy and Eberhart 1995)
• Artificial bee colony optimization (Karaboga 2005)

Different algorithms have various strengths and weaknesses on features such as

computational efficiency, speed, accuracy, memory requirements, and the type
of problem they are best suited to solving. It is therefore often the case that
particular subdisciplines prefer a specific algorithm. In the next section, we
will introduce a widely-used, generic optimization algorithm called the Downhill
Simplex method. Because this is one of the oldest methods around, it has been
implemented in many programming languages, including R.
170 CHAPTER 9. NONLINEAR CURVE FITTING

9.7 The Downhill Simplex algorithm

The simplex algorithm, sometimes also called the amoeba method, was developed
by Nelder and Mead (1965). It involves constructing a virtual shape (called the
simplex) that navigates around the multidimensional error surface (e.g. Figure
9.4) in search of the lowest point. The shape has a number of vertices (corners),
each of which consists of an estimate of all of the model parameters. When the
algorithm is first started, the initial guess for these parameters is usually fed in
by the programmer. On each iteration of the algorithm, the error between the
model and the data (e.g. the height of the error surface) is calculated for the set
of parameters at each vertex of the simplex.
Next, the simplex performs geometric operations to try and get closer to the
surface minimum. These operations can include expansion, contraction, shrinkage
and reflection, or combinations of these operations. Expansion means that one
corner of the simplex (usually the one giving the worst fit) moves away from
the shape’s centroid by some amount, so the shape gets larger. Contraction is
the opposite - one corner moves towards the centroid making the shape smaller.
Shrinkage is similar, except that all points except for the one that is currently
giving the best solution (i.e. the one at the lowest point on the surface) move
in towards the best one. Reflection is a little more complicated, as it involves
shifting one vertex to the opposite side of the mean of the other vertices. Note
that reflection is the main way for the simplex to move around the surface, as
this shifts the centroid of the shape quite substantially. An example of the
progress of a simplex algorithm (every five iterations) across a three dimensional
error surface is shown by the blue triangles in Figure 9.5.
Over time, the simplex navigates around the error surface, and eventually returns
a set of parameters corresponding to the best solution (i.e. lowest point) it has
found. The algorithm decides when to stop based on a set of termination criteria
- rules that tell it when it should finish. Several termination criteria are possible,
including the total number of iterations of the algorithm, the total number of
times the error function has been evaluated, and how close the vertices of the
simplex can get before terminating. Usually all of these criteria can be adjusted
by the programmer if the algorithm has either been taking too long or is not
providing a very good fit. Once the set of best fitting parameters has been
returned, these can be plugged back into the model equation to create a curve
that fits the data points.

9.8 Local minima

One common problem with the simplex algorithm (and other optimization
methods) is that it can sometimes get stuck in a local minimum. This is a region
of the error surface that is lower than all surrounding points, but is not the
overall lowest possible point (the global minimum). The algorithm gets stuck
because once it is in the local minimum, any direction it tries to move in makes
9.8. LOCAL MINIMA 171

100 20 30 40 50

10
80
Intercept (β0)

60

40
20
20
30

40
100 90 80 70 60 50
0

Slope (β1)
−2 −1 0 1 2

Figure 9.5: Path taken by the simplex, proceeding left to right from the starting
position (white triangle), to the final solution (black star). The blue triangles
are intermediate instances, sampled every 5 iterations.
172 CHAPTER 9. NONLINEAR CURVE FITTING

the model fit worse. So it stays put, and cannot find the global minimum. A
good example of a surface with multiple minima is the Himmelblau function,
shown by the contour plot in Figure 9.6.

0 250 350 400 450 500 55

0 30 250 300 0 60
35 200 0
4 50
100
150
50

0
300

100

−2
200
250
−4 50 350
400
150
300
200 550 500 450 500

−4 −2 0 2 4

Figure 9.6: The Himmelblau function. This is a mathematical function that

has four minima of slightly different depths (shaded blue), and is often used for
benchmarking optimization algorithms.

The black star around (x = 3, y = 2) indicates the true global minimum, but
often an optimization algorithm will get stuck in one of the two local minima
on the left hand side of the plot, and return parameter values from one of these
two locations instead. These are also good solutions, and sometimes they will
be sufficient for whatever purpose we have in fitting our model (the model curve
will likely follow the data quite closely). However they are not quite as good as
the global minimum, which we would ideally like to find.

There are two main ways to fix the local minima problem. The first is to restart
the simplex algorithm from many random starting points, in the hope that
one version finds the global minimum. The other is to alter one of the fitted
parameter values by a large amount, and then restart the algorithm from this
new location. This method, referred to as casting the stone, assumes that if the
original solution is the global minimum, the algorithm will not find a better
solution in the new region of the search space to which it has been ‘cast’, and
will reconverge to the original solution.
9.9. SOME PRACTICAL CONSIDERATIONS 173

9.9 Some practical considerations

Complex models can often take a long time to fit. One way to speed things
up is to optimize the code that calculates the model predictions as much as
possible. This can involve removing extraneous commands, replacing loops with
matrix operations, pre-allocating memory, compiling code, and making use of
parallel processing capabilities. Because the model code will be called hundreds
or thousands of times during optimization, even small increases in efficiency can
often translate to time savings of several hours.

Another way to speed up fitting is to constrain the range of values that one or
more parameters can take. Often this can be determined on practical grounds.
In the baby height example from Figure 9.1, we could constrain the slope value
to always be positive (because babies shrinking as they get older doesn’t make
sense). When model parameters represent real-world properties of a physical
system, it is sometimes reasonable to constrain them to lie within a sensible
range. For example, a model parameter representing body temperature in live
humans could be constrained to lie between 10 and 50◦ C because temperatures
outside of this range would be fatal.

Models that contain a stochastic (random) element (see Chapter 8) are not
typically suitable for use with optimization algorithms. This is because the
model itself returns a different solution each time it is run, even with a fixed
set of parameters. This means that the error surface changes on every iteration,
causing obvious problems for fitting. One solution to this is to use the same
seed value for the random number generator on each iteration. This freezes the
surface and allows the minimum to be found. However, it will be important to
rerun the fitting with different values of the random seed to check that similar
parameter values are found each time.

Real data sets often contain some data points that are more reliable than others.
This might be because more observations were made in some conditions than in
others. In such situations, it can be useful to weight the data points by some
measure of their reliability when calculating the error of the fit. Doing this
might prevent very noisy data points from having an undue influence on the
fit. The precise values used for the weights will depend on the type of data
you are fitting, and the general idea of weighting was introduced in section 6.8.
As with many aspects of computational modelling, the precise details of the
implementation are left to the modeller, and as you gain more experience you
will usually develop heuristics that work for the type of data you are interested
in.
174 CHAPTER 9. NONLINEAR CURVE FITTING

9.10 Two philosophical approaches in model fit-

ting
There are many reasons why it is useful to fit models to data. One advantage
is that complex data can often be summarised into one or more parameter
estimates. These parameters can be taken forward to a further stage of analysis -
for example if a model is fitted to the data for each individual in an experiment,
and then the model parameters are compared between groups or conditions. In
such cases, the model can be essentially descriptive, perhaps involving some
convenient mathematical function (e.g. an exponential decay function). This can
be a valuable thing to do, and is a widespread and perfectly legitimate use of
model fitting.
However, another type of model is one that attempts to represent, on some
level, the underlying processes involved in producing the data. This type of
model is called a functional (or sometimes generative) model, and can often be
more informative than the descriptive models outlined above. For example, by
comparing different functional models, it is often possible to infer the processes
involved in the system under study. Models that do not fit the data well can be
rejected, whereas more accurate models are preferred. Furthermore, if a model is
more than just a description, it can often generate predictions for new conditions
and situations, that can subsequently be tested empirically. This might be
particularly useful for predicting future events, e.g. for climate conditions, stock
market activity, or neural responses to a novel stimulus. Thinking about the
type of model you wish to construct is an important first step in undertaking
any form of computational modelling.
To illustrate the distinction between descriptive and functional models, let’s
think again about the coronavirus case data shown in Figure 9.3. We fitted a
descriptive model here, using an exponential function with three free parameters.
As we saw, the model was not very good at predicting the future when we fitted it
to the first 30 days of case data. And of course it is just a simple equation, so the
model has no concept of things like typical transmission rates, legal restrictions,
vaccination, and mask wearing, that might affect cases. If instead we had built
a functional model, it might have parameters representing these sorts of factors,
and would therefore be more robust in its future predictions (for an example see
Friston et al. 2020). We could also see what effect changing these parameters
would be likely to have, perhaps letting us predict the impact of different public
health policy interventions.

9.11 Tools to help with model development

Given the above, it is reasonable to ask how one should decide upon the model
one is trying to fit. This is a very domain-specific question, and so it is not
sensible to try to be prescriptive here. However, there are some general rules
of thumb that may be of use. First, there is no substitute for visualising the
9.12. HOW TO FIT CURVES TO DATA IN R 175

data you are trying to model (and see Chapters 3 and 18 for some guidance on
plotting). If the data have a clear form, for example a Gaussian-like distribution,
this might suggest, or rule out, particular mathematical functions. Second, the
simpler a model is, the easier its behaviour will be to understand. One way to
simplify a model is to reduce the number of free parameters as far as possible.
Statistics have been proposed to mathematically compare the performance of
models with different numbers of parameters. The Akaike Information Criteria
(AIC; Akaike 1974) is a widely used example, and contains a penalty term that
increases with the number of free parameters. This can help to avoid ‘overfitting’,
by excluding parts of a model that may not be necessary to provide an acceptable
fit. In general though, reading existing studies on a similar topic is the best
way to get a feel for the types of models that might be suitable for a particular
data set. Some more detailed practical suggestions for model development are
proposed by Blohm, Kording, and Schrater (2020) - although the authors focus
on neuroscience, the points they make are generally applicable to modelling in
other domains.

9.12 How to fit curves to data in R

The R language contains many packages that implement different optimization
algorithms. We will focus on the pracma package, which contains a version of
the Nelder-Mead (1965) downhill simplex algorithm that we described earlier in
the chapter. The package has dozens of useful functions, but the key function
for our present purpose is called nelder_mead. However this function requires as
an input the name of the function that we want to optimize, which we will need
to write ourselves. The function we optimize must:
• take a vector of parameters as its input
• return the error between the the model and data as its output
Let’s fit something that’s a bit more interesting than a straight line. A Gaussian
function can be described with two free parameters as follows:

−(x−a)2
f (x) = e 2σ 2 , (9.4)

where a and σ are free parameters, and x is the value along the x-axis. The
parameter a controls the horizontal offset of the Gaussian, and σ controls the
spread (width). Gaussian functions can be used to characterise many biological
processes, such the tuning functions of neurons. We can implement the equation
as the first line of an R function (see section 2.7 for a refresher on how function
definitions work in R). I have named the function errorfit, because it calculates
the error between the model and some data (in other words, the error of the fit):
errorfit <- function(p){ # define a new function called 'errorfit'
# equation of a gaussian, with parameters from the input, p
gaus <- exp(-((x-p[1])^2)/(2*p[2]^2))
176 CHAPTER 9. NONLINEAR CURVE FITTING

# root mean square error between model and data

rms <- sqrt(sum((gaus-ydata)^2)/length(ydata))
return(rms)} # the function returns the rms error

The second line of the function calculates the root-mean-squared (RMS) error
by taking the differences between the model (stored in the gaus data object)
and the data (stored in the ydata data object), squaring the differences, and
calculating the square root of the mean. We will generate some synthetic data
for the model to fit, where we know the true values of the free parameters, and
see how well the simplex algorithm can recover them. Let’s set the values to be
a = 2 and σ = 3, and generate data using the Gaussian function for a range of
x-values, adding a bit of noise (to simulate measurement error):
x <<- seq(-10,15,1) # sequence of x-values from -10 to 15
p <- c(2,3) # true parameter values used to simulate some data
ydata <- exp(-((x-p[1])^2)/(2*p[2]^2)) # simulated data from gaussian
ydata <<- ydata + 0.05*rnorm(length(x)) # add noise to simulated data
plot(x,ydata,type='p') # plot simulated data
1.0

●
●

● ●
0.8

●
0.6

●
ydata

●
●
0.4

● ●
0.2

● ● ● ●
● ● ● ●
● ● ●
0.0

●
●
● ●

−10 −5 0 5 10 15

Figure 9.7: Simulated data from a Gaussian function

The above code generates the graph shown in Figure 9.7. One small point to
notice. When we define the data objects x and ydata, we use the double arrow
assignment (<<-) to specify that they are global variables. This means they are
available from within the errorfit function, so we do not need to explicitly pass
9.12. HOW TO FIT CURVES TO DATA IN R 177

them to it as inputs.
If we tried to guess the model parameters without knowing them in advance, we
might estimate that the middle of the function was around 5, and the spread
was around 1. This would produce the (very poor) fit shown in Figure 9.8, and
given by the following code:
p <- c(5,1) # a guess at some possible parameter values
pred <- exp(-((x-p[1])^2)/(2*p[2]^2)) # model prediction using these parameters
plot(x,ydata,type='p') # plot the data again
lines(x,pred,lwd=2,col='#8783CF') # add the model prediction
1.0

●
●

● ●
0.8

●
0.6

●
ydata

●
●
0.4

● ●
0.2

● ● ● ●
● ● ● ●
● ● ●
0.0

●
●
● ●

−10 −5 0 5 10 15

Figure 9.8: Simulated data with a poorly fitting Gaussian function

Notice that we store the parameters in a data object called p, which we could pass
to the errorfit function to get a numerical estimate of how good (or otherwise)
the fit is:
# calculate the error of the fit with our first guess parameter values
errorfit(p)

## [1] 0.3870659
The fit is obviously poor, and hopefully by optimizing our parameters we will
be able to improve on the RMS error of 0.4. We will do this using the downhill
simplex algorithm, that is called using the nelder_mead function. We provide
the function with the name of the errorfit function and a starting ‘guess’ for
178 CHAPTER 9. NONLINEAR CURVE FITTING

what the parameter values might be. It returns a data object, which contains
the estimated parameters. We can then plug those parameters into our equation,
and generate a curve that fits the data well (see Figure 9.9).
library(pracma) # load the pracma package

sout <- nelder_mead(errorfit, c(1,1)) # fit the model to the data

p <- sout$xmin # extract the parameter estimates

# get the model predictions for these parameters

pred <- exp(-((x-p[1])^2)/(2*p[2]^2))

plot(x,ydata,type='p')
lines(x,pred,lwd=2,col='#8783CF') # plot the model fit as a line
1.0

●
●

● ●
0.8

●
0.6

●
ydata

●
●
0.4

● ●
0.2

● ● ● ●
● ● ● ●
● ● ●
0.0

●
●
● ●

−10 −5 0 5 10 15

Figure 9.9: Simulated data with the best-fitting Gaussian function (curve).

The data object produced by the nelder_mead function (the sout object) contains
several pieces of information besides the final parameter estimates. For example
it includes the value of the function, and the number of iterations that were run.
We don’t need to look at this information now, but it is there if you ever need it.
The estimated parameter values should be close to the original values used to
generate the data, and as you can see (from Figure 9.9) the curve provides a
good fit to the data points. We have extracted the estimated parameters from
sout$xmin and stored them in the data object p:
9.12. HOW TO FIT CURVES TO DATA IN R 179

## [1] 2.080850 2.970019

Notice that our estimated parameter values are close to the original values (2
and 3) that we used to generate the data in the first place. If we want to know
how good the fit is numerically, we can call the errorfit function ourselves, and
it will tell us the RMS error:
errorfit(p)

## [1] 0.04361436
The RMS error is much smaller than for our non-optimized best guess parameters
that we started with. This means that the simplex algorithm has done a good
job of fitting the model and finding some good parameter values. Something we
could potentially do next is to bootstrap this whole process (see Chapter 8) to
obtain confidence intervals on our parameter values.
This example of function fitting contains all of the same steps as we would go
through for a more sophisticated model fit. We need to create an R function that
calculates the model predictions for a given set of parameters, and calculates the
error between the model predictions and the data. We then pass this function
to the simplex algorithm (nelder_mead function), along with an initial guess
about the parameters. This initial guess will often influence the end result, so
it is sensible to repeat the fitting process many times using random starting
parameters, and choose the model parameters with the best overall fit. For the
example above, such a procedure might look something like this:
# first initialise data objects to store the best error value and parameters
bestrms <- 10000
bestp <- c(0,0)

# now loop through some number of repetitions

for (n in 1:100){

# run the simplex with some random starting values

sout <- nelder_mead(errorfit, 10*rnorm(2))
p <- sout$xmin # extract the fitted parameters
thiserror <- errorfit(p) # work out the error of the fit

# if this is the best fit we've found so far, store the parameters
if (thiserror<bestrms){
bestrms <- thiserror
bestp <- p}
}

This general approach can be used to fit models of arbitrary complexity to any
type of data, and is an enormously flexible and useful scientific tool. Of course,
180 CHAPTER 9. NONLINEAR CURVE FITTING

more complex models will take longer to fit, and might require additional lines
of code (or extra functions) to specify. You could use different measures of error,
and specify some additional options in the simplex fit, such as the maximum
number of iterations or evaluations, as described in the help files for nelder_mead.
Alternatively, it is also possible to use a Bayesian approach (see Chapter 17)
to fitting models. This involves sampling a version of the error surface using a
stochastic process (see Kruschke 2014). Happy fitting!

9.13 Practice questions

1. If a model has 12 free parameters, how many dimensions does the error
surface have?
A) 3
B) 12
C) 13
D) 24
2. The ‘height’ of the error surface represents:
A) The value of the parameters
B) The error between model and data
C) The number of degrees of freedom
D) The number of free parameters
3. Which of the following is not an operation performed by a simplex?
A) Circulation
B) Expansion
C) Reflection
D) Contraction
4. A good solution that is not the global minimum is often called:
A) An iteration
B) The RMS error
C) An error surface
D) A local minima
5. One straightforward way to speed up model fitting is to:
A) Optimise the code that calculates the model predictions
B) Delete all the data objects in the Environment
C) Start the simplex from a global minimum
D) Collect more data before fitting your model
6. Which of the following would be the most sensible constraint on model
parameters?
A) Constraining a date parameter to always be in the future
B) Constraining a parameter that represents blood pressure to be negative
C) Constraining a parameter that represents heart rate to be positive
D) Keeping a parameter value as large as possible
7. The nelder_mead function must operate on:
A) A function that has less than three free parameters
B) A function that is built into the core R toolbox
9.13. PRACTICE QUESTIONS 181

C) A function that returns the best fitting parameters

D) A function that calculates the error between model and data
8. A useful tool for comparing models with different numbers of free parame-
ters is:
A) The Downhill Simplex algorithm
B) Akaike’s Information Criteria
C) Cohen’s d
D) The p-value of a t-test
9. The best fitting model for a particular data set should have:
A) A bigger RMS error value than other models
B) A smaller RMS error value than other models
C) The largest number of free parameters
D) The smallest number of free parameters
10. A functional model attempts to:
A) Model the underlying processes that produced the data
B) Describe the data using a convenient mathematical equation
C) Incorporate as many free parameters as possible
D) Predict all possible states of the system under study
Answers to all questions are provided in section 20.2.
182 CHAPTER 9. NONLINEAR CURVE FITTING
Chapter 10

Fourier Analysis

Fourier analysis is a technique that allows us to determine the frequency content

of waveforms. It is used in a variety of signal-processing situations, and is a core
analysis method in many areas of science and engineering. In this chapter we
will explain the theory behind Fourier analysis in one and two dimensions, and
illustrate its usefulness with several examples. After finishing the chapter, you
should have an understanding of how to conduct Fourier analysis in R, and how
to interpret and manipulate the frequency content of signals.

10.1 Joseph Fourier: polymath

Figure 10.1: Portrait of Joseph Fourier. By Louis-Léopold Boilly,

https://commons.wikimedia.org/w/index.php?curid=3308441

Fourier Analysis takes its name from a 19th Century French mathematician
called Joseph Fourier (see Figure 10.1). Fourier was a polymath: an expert on

183
184 CHAPTER 10. FOURIER ANALYSIS

many topics. Of particular note, he is generally credited as being the first person
to desribe the greenhouse effect - the process by which carbon dioxide traps heat
near the surface of a planet and causes global temperatures to rise.
The basic idea behind the technique that bears his name is that any waveform
can be decomposed into a bunch of sine waves of different frequencies (you may
have encountered the sine and cosine functions when learning trigonometry). Of
course, in Fourier’s era there were no computers, meaning that this procedure
had to be carried out by hand. This was a prohibitively slow process, and so
Fourier Analysis was not widely used until long after his death. However modern
computers make the calculations straightforward and very fast, and Fourier
Analysis has been used in a wide variety of signal processing applications, as we
will describe below.

10.2 Example applications

So what’s the point in breaking waveforms down into sine waves? The main
purpose is to describe the waveform in terms of the frequencies it contains. In
the time domain, frequency refers to how often a signal repeats in a given period
of time. The unit for this is the Hertz (Hz) - one Hz means one repetition per
second, ten Hz means ten repetitions per second (see Figure 10.2). Describing a
sound recording in terms of frequencies might allow us to identify any interference
in a particular frequency range, or perhaps alter the signal to remove certain
frequencies. Indeed, most methods for compressing audio files (such as MP3
compression) involve removing frequencies that are inaudible to human ears.
Many biological processes are also periodic or semi-periodic, and can usefully be
summarised using Fourier analysis. Examples include heart rate and breathing, as
well as slower processes like the circadian rhythm and the cell cycle. Furthermore,
many modern scientific methods, such as crystallography and magnetic resonance
imaging, crucially depend on Fourier analysis to work.
In psychology and neuroscience, an important application of Fourier Analysis
is to analyse time-varying brain signals, of the type recorded using Electroen-
cephalography (EEG) and Magnetoencephalography (MEG). These ‘brainwaves’
often have activity at characteristic frequencies that are associated with specific
mental states or cognitive operations. For example, when participants in an
experiment become tired, they produce more activity in the alpha band of
frequencies from 8-12Hz. Synchronisation and suppression between brain areas
involves oscillations at higher frequencies in the gamma band, above 30Hz. To
quantify this type of activity, it is typical to use Fourier Analysis to decompose
the signal into its component frequencies.
Finally, as we will see later, it is also possible to apply Fourier Analysis to
two dimensional images (or even movies). This has numerous applications
in image processing, filtering, and synthesis that are useful for designing and
understanding experimental stimuli.
10.2. EXAMPLE APPLICATIONS 185

1Hz

10Hz

0 0.2 0.4 0.6 0.8 1

Time (s)

Figure 10.2: Example sine waves of different frequencies. The 1Hz wave at the
top goes through a single cycle (it increases, then decreases, then returns to
baseline) during the one second of time depicted. The 10Hz wave below it goes
through ten cycles in the same period of time.
186 CHAPTER 10. FOURIER ANALYSIS

10.3 Terminology
A key first step in understanding Fourier Analysis is to get your head around some
important terms. If we take a waveform and calculate the Fourier transform,
this will break the waveform down into its component frequencies. The result
is referred to as the Fourier spectrum, and has two parts as we will describe
in a moment. If we want to convert from the spectrum back to the waveform,
we perform the inverse Fourier transform (see Figure 10.3 for an example).
These operations are also referred to as Fourier analysis and Fourier synthesis
respectively, and the underlying mathematics are known as the Fourier theorem.
Understanding the maths is not required to use these methods, which are
implemented as core functions in most computer programming languages.

Waveform Fourier spectrum

Fourier analysis
(Fourier transform)

Fourier synthesis
(Inverse transform)

0 0.2 0.4 0.6 0.8 1 0 100 200 300 400 500

Time (s) Frequency (Hz)

Figure 10.3: Illustration of the Fourier transform between a waveform (left) and
its Fourier spectrum (right), for one second of brain activity measured using
electroencephalography (EEG).

Fourier analysis is an invertible linear process, which means that no information

is lost when converting between the waveform and its spectrum. A consequence
is that you can perform these operations as many times as you like without
distorting your original signal. I think of the Fourier spectrum as an alternative
way of representing the same information. Graphically, this is most often
represented by the amplitude spectrum.

10.4 The amplitude spectrum

The amplitude spectrum is typically plotted as a graph, in which the x-axis
shows the frequency, and the y-axis shows the amplitude. The frequency, as
10.4. THE AMPLITUDE SPECTRUM 187

mentioned above (see Figure 10.2), is the number of cycles per unit of time. The
amplitude is the vertical difference between the peaks and the troughs of the
sine wave (see Figure 10.4). A low amplitude means a very small change, and a
high amplitude means a large change.

Low amplitude

High amplitude

0 0.2 0.4 0.6 0.8 1

Time (s)

Figure 10.4: Example sine waves of different amplitudes. The sine wave at the
top has a low amplitude, the one below it has a high amplitude.

Figure 10.5 shows an example amplitude spectrum. This is based on the same
data as in Figure 10.3, but here we have zoomed in on the portion of the
spectrum from 0 - 30 Hz. This is where most of the action is in human brain
activity (because of the intrinsic timescales at which neurons operate), and so is a
worthwhile frequency range to focus on. The highest amplitude is at 5Hz, which
is consistent with the clearly periodic nature of the waveform in Figure 10.3,
showing 5 peaks and 5 troughs in one second. The amplitude spectrum therefore
has a direct mapping to features of the waveform it represents. The units of
amplitude are the same as the units used to measure the original waveform.
For the EEG data used in the examples here, these are microvolts (µV ), but
the units will correspond to whatever dependent variable you have chosen to
measure.
188 CHAPTER 10. FOURIER ANALYSIS

0 5 10 15 20 25 30
Frequency (Hz)

Figure 10.5: Example amplitude spectrum, zoomed in from 0 - 30Hz.

10.5 The phase spectrum

The amplitude spectrum contains only half of the information produced by the
Fourier transform. The remainder is found in the phase spectrum. The property
of phase corresponds to the horizontal alignment a waveform. It is always a
relative measure, and is usually expressed with reference to some meaningful
event, such as the onset of a stimulus or the start of a recording. Phase is
a circular (periodic) term, with phase angles being measured in degrees (or
sometimes radians, which are an alternative angular unit). Figure 10.6 shows
two example sine waves with different phases. The upper waveform has a phase
angle of 0◦ relative to the vertical dashed line. This is referred to as being
in sine phase. The lower waveform has a phase angle of 90◦ relative to the
vertical dashed line. This is referred to as being in cosine phase. Two other
common phases are negative sine (180◦ ) and negative cosine (270◦ ) phase, which
correspond to a vertical flip (peaks become troughs and vice versa) of the two
waveforms shown in Figure 10.6.
The Fourier component at each frequency in the amplitude spectrum has an
accompanying phase term. This is difficult to show graphically, and so is not
generally represented, at least for many frequencies at once. However the phase
information is critical, as without it the original waveform cannot be reproduced.
Indeed, phase scrambled stimuli are often used in experiments as a control (or
baseline) condition because recognisable information (e.g. a speech signal) is
10.5. THE PHASE SPECTRUM 189

Sine phase

Cosine phase

0 0.2 0.4 0.6 0.8 1

Time (s)

Figure 10.6: Example sine waves of different phases, relative to the vertical
dashed line. The waveform at the top is in sine phase with the line (i.e. the line
is mid-way through a cycle). The waveform below is in cosine phase with the
line (i.e. the line is at a peak)
190 CHAPTER 10. FOURIER ANALYSIS

destroyed by the scrambling process.

10.6 Limitations on sampling: the Nyquist limit

and frequency resolution
Fourier analysis is not magic - it has two limitations. The first limitation is the
maximum frequency that can be included in the spectrum, which is determined
by the sample rate of the original waveform. The sample rate is how often the
dependent measure is taken, which is usually determined by the hardware used
to record the data. Consumer video cameras usually record around 25 or 30
frames per second, giving a sample rate of 25-30Hz. Microphones and other
audio hardware often record at a very high sample rate, around 44100Hz, so that
they can capture high frequency sounds. EEG data is often sampled at a rate of
1000Hz (e.g. a thousand measurements per second). fMRI data is sampled much
more slowly, at around one sample every three seconds (i.e. 1/3Hz). Eye tracking
hardware usually operates somewhere between 30Hz and 1000Hz, depending on
the make and model of equipment.
The maximum frequency that can be resolved is known as the Nyquist limit, and
it is always exactly half of the sample rate. So, a signal recorded at 1000Hz has
a Nyquist limit of 500Hz. This is because a frequency of half the sample rate
has sufficient resolution for one high sample (e.g. a peak) and one low sample
(e.g. a trough).
The second limitation is the frequency resolution - e.g. the granularity of steps
along the frequency axis of the amplitude spectrum. Surprisingly this is deter-
mined by the duration of the sample, and not the sample rate. A one-second
sample has a frequency resolution of 1Hz. That means the spectrum will contain
information at 1Hz, 2Hz, 3Hz and so on, but not at any intermediate frequencies
1
(1.5Hz for example). A ten second sample has a frequency resolution of 10 Hz,
meaning that there are intermediate frequency bins between the integer frequen-
cies in steps of 0.1Hz. The frequency resolution is therefore given by 1/(duration
in seconds), regardless of the sample rate. Only the Nyquist limit is determined
by the sample rate.

10.7 Example: bat species identification by fre-

quency
A good applied example of how Fourier analysis can be used is in the identification
of bat echolocation signals. Different species of bats produce calls at distinct
frequencies, which are generally well above the limits of human hearing. By
combining the Fourier spectra with classification techniques (see Chapter 14), it
is possible to identify individual species by their calls. For example, waveforms
and Fourier spectra for two bat species are shown in Figure 10.7. The waveforms
10.8. FOURIER ANALYSIS IN TWO DIMENSIONS 191

look broadly similar - the offset along the x-axis is arbitrary, and determined only
by when in the recording the call began. However the amplitude spectra in the
lower plot have peaks at very different frequencies. The Common Pipistrelle’s
call (black) peaks at around 5000 Hz, whereas the Noctule’s call (blue) peaks at
around 2000 Hz.
There have been many different classification systems proposed that use Fourier
transformed echolocation signals to identify bat species. For example, Walters
et al. (2012) trained an artificial neural network to discriminate between 34
different bat species. The calls were first assigned to one of five different groups;
this classification had an extremely high accuracy of around 98%. Calls were
subsequently assigned to individual species, which had a slightly lower accuracy
of around 84% (but still far above chance performance of 100/34 = 2.9%).
Online tools are available to classify bat calls, and mobile phone applications
and dedicated handheld devices are now available that can perform classification
in real time out in the field. These tools are all based on Fourier analysis.

10.8 Fourier analysis in two dimensions

One potential use of Fourier analysis (that is sometimes surprising when first
encountered) is that it can be applied in more than one dimension. This means
that images (which we can think of as two dimensional signals) and movies (three
dimensional signals) can also be analysed in this way. The mathematics turn
out to be equivalent to the one dimensional case, albeit slightly more complex.
Because Fourier analysis uses the sine wave as its basis function, a consequence is
that we can consider all images to be reducible to a combination of sine waves of
various spatial frequencies and orientations. See Weisstein (1980) for an excellent
and detailed tutorial on two-dimensional Fourier analysis.
So what is spatial frequency? For the waveforms we have considered so far, the
signal (usually a measure like voltage, or sound pressure level) changes as a
function of time, and the frequency units are Hertz (cycles per second). Because
the signal is changing over time, we call this property temporal frequency. For
an image, the ‘signal’ is the change in luminance as a function of space, and we
call the rate of this change the spatial frequency. The units of spatial frequency
are cycles per degree of visual angle, or sometimes (for simplicity) cycles per
image. Figure 10.8 shows example sine wave grating images, with different spatial
frequencies and orientations. By applying Fourier analysis in two dimensions, it
becomes apparent that all photographs can be described as the sum of multiple
sine waves of different spatial frequencies, orientations, phases and amplitudes.
Just as we took the Fourier transform of a 1D waveform, so we can Fourier
transform a 2D image. The Fourier spectrum will then be two-dimensional also,
and is referred to as Fourier Space (or sometimes the Fourier Domain). The left
panel of Figure 10.9 shows an image of a bug hotel. Although it is hard to think
of a photograph as a ‘waveform’, consider that an image is made up of pixels
192 CHAPTER 10. FOURIER ANALYSIS

(a)
Pipistrelle
Amplitude

Noctule

0 100 200 300 400 500

Time (ms)

(b)
Noctule
Amplitude

Pipistrelle

0 2000 4000 6000 8000 10000

Frequency (Hz)

Figure 10.7: Waveforms (a) and Fourier spectra (b) for example calls from two
bat species: Pipistrellus pipistrellus (black), and Nyctalus noctula (blue).
10.8. FOURIER ANALYSIS IN TWO DIMENSIONS 193

Figure 10.8: Example sine wave gratings of different spatial frequencies and
orientations. The left grating has a low spatial frequency (3 cycles per image),
the middle grating has a higher spatial frequency (10 cycles per image). The
right grating has an oblique orientation.

that vary in intensity across space, and that if we plot those values for one row
of the image (superimposed in blue), they look very much like a waveform. In
the right panel is the Fourier spectrum of the image. This has been zoomed into
the central low spatial frequency portion, where most of the energy resides.

The two-dimensional Fourier spectrum is hard to interpret without some guidance.

Figure 10.10 uses small patches of grating to illustrate the layout graphically.
The lowest spatial frequencies are represented in the centre of the plot, with
higher spatial frequencies towards the edges. Orientation is represented by the
angle from vertical. However, because of a somewhat confusing convention,
horizontal orientations are traditionally represented along the vertical axis, and
vertical orientations are represented along the horizontal axis. Oblique (diagonal)
orientations are represented in between. The spectrum is symmetrical (mirrored)
about the mid-point (illustrated by the vertical gratings along the horizontal
axis).

Of course in a real Fourier spectrum, the small grating icons are not shown.
Instead the value (brightness) at each point (i.e. each x,y coordinate) represents
the amplitude at that particular combination of orientation and spatial frequency.
The right hand panel of Figure 10.9 shows an example in which most of the
energy is concentrated at low spatial frequencies (as is typical for natural images),
with dominant vertical energy (along the horizontal axis), caused by the vertical
contours (wooden poles) in the original image.
194 CHAPTER 10. FOURIER ANALYSIS

Figure 10.9: Greyscale image of a bug hotel (left), and its Fourier spectrum
(right).

10.9 Example: using 2D Fourier analysis to

measure goosebumps
If you are cold, tiny hairs on your arms and down your back will stand on end
and you will get goosebumps - small raised lumps at the base of each hair. The
same response, called the piloerection response, can also be triggered by other
experiences, such as being frightened, or listening to a particular piece of music
that gives you ‘chills’. A group of researchers in Germany and Austria developed
an elegant technique to quantify the piloerection response using Fourier analysis.
Benedek et al. (2010) built a device that involved attaching a small video camera
to a patch of skin, to continuously film a close up view of the hair follicles.
An LED light source illuminates the patch of skin from one direction. Most
of the time no goosebumps are visible, and the output of the camera might
look something like the image in Figure 10.11a. However, when a participant
experiences goosebumps, they cast shadows in the opposite direction to the light
source, which are clearly visible in the images (see Figure 10.11b).
Hair follicles are distributed with approximately even spacing across the skin.
That means the pattern of shadows will have a consistent spatial frequency,
which turns out to be around 0.4 cycles per mm (in other words, you have a hair
follicle about every 2.5 mm). If we look at the ratio of the Fourier spectra of the
two images, there is a clear peak corresponding to the goosebumps (see Figure
10.11c). Benedek et al. (2010) created software (available at www.goosecam.de)
that takes the Fourier transform of the camera images, and sums the amount
of signal around this frequency to give a continuous, objective estimate of the
piloerection response. This method has many potential applications in several
areas of research, particularly in understanding our experience of music and
10.9. EXAMPLE: USING 2D FOURIER ANALYSIS TO MEASURE GOOSEBUMPS195

10

10
10 8 6 4 2 0 2 4 6 8 10
Frequency (cycles per image)

Figure 10.10: Illustration of Fourier Space.

196 CHAPTER 10. FOURIER ANALYSIS

(a) Baseline (b) Goosebumps

(c)
Ratio of spectra

Figure 10.11: Illustration of using Fourier analysis to quantify the piloerection

response. The images on the top row show a small patch of skin without (a)
and with (b) goosebumps. Panel (c) shows the ratio of the Fourier spectra of
the two images (smoothed), which exhibits a strong peak corresponding to the
presence of goosebumps. The blue trace at the right hand margin shows the
polar average of the ratio. Image credit: Rémi de Fleurian.
10.10. FILTERING: ALTERING SIGNALS IN THE FOURIER DOMAIN197

film. Amazingly, the researchers also found an individual who has direct control
over their own piloerection response, and could give themselves goosebumps on
demand!

10.10 Filtering: altering signals in the Fourier

domain
One common reason for using Fourier analysis is that signals can often be
manipulated in the Fourier domain in various ways that would be challenging
to achieve using only the original signal. A simple example is to remove high
frequency ‘noise’ from a signal. This is achieved by filtering: multiplying the
Fourier spectrum by a filter constructed to include some frequencies and exclude
others, and then taking the inverse transform. An example of this for removing
noise from an EEG waveform is shown in Figure 10.12. The filter here is referred
to as a low pass filter, because it passes (allows through) low frequencies (those
within the blue shaded region), but blocks higher frequencies (those outside
it). Because high frequencies here contain mostly noise, this has the effect of
smoothing the waveform (shown by the black curve in the right panel of Figure
10.12).

Fourier spectrum Waveform

0 20 40 60 80 100 0 0.2 0.4 0.6 0.8 1

Frequency (Hz) Time (s)

Figure 10.12: Illustration of low-pass filtering. The left plot shows the Fourier
spectrum, with superimposed low-pass filter (blue), which excludes the high
frequency components outside of the shaded region. The right panel shows the
original waveform (blue) and the filtered waveform (black) which lacks the high
frequency noise and is therefore visibly smoother.

We can also apply filters in two dimensions. Figure 10.13 shows low pass and high
pass filters, and their effect on the bug hotel image. The low pass filtered image
198 CHAPTER 10. FOURIER ANALYSIS

looks blurry, as the fine detail is stored at the higher spatial frequencies which
have been removed by the filter. The high pass filtered image lacks extended
light and dark regions (represented by the lower spatial frequencies), and retains
only edges at higher frequencies. Note also how the overlaid pixel intensities for
the central row (shown in dark blue) are smooth in the low pass filtered version,
and jagged in the high pass filtered version.

Finally, we can filter in the orientation domain. Figure 10.14 shows filters and the
resulting images in which either horizontal (left) or vertical (right) information
is removed, leaving information at the orthogonal orientation. Notice how in
the image where horizontal information is removed (left) we can still clearly see
the vertical poles at either side of the image, and the vertical white parts of the
little drawer unit in the centre at the bottom. On the other hand, in the image
where vertical information is removed (right) these features are missing, but a
central horizontal bar, and the horizontal parts of the drawer unit are visible.

10.11 Stimulus construction in the Fourier do-

main
As well as altering existing stimuli in Fourier space, we can construct stimuli
from scratch in this way also. The classic example of this is to synthesise a
square wave from the Fourier series of odd harmonic sine waves. We begin
with a sine wave at the lowest frequency, which for this example is 1Hz. This is
referred to as the fundamental, often denoted 1F. It is generated by setting the
amplitude of the Fourier spectrum at 1Hz to 1, and taking the inverse transform
to visualise the waveform (see top row of Figure 10.15). We then add energy at
each of the odd harmonic frequencies in turn - the harmonics are the multiples
of the fundamental frequency, and we just want the odd numbered ones: 3F,
5F, 7F and so on. The amplitude of each harmonic is given by h1 , where h is
the harmonic (e.g. 3, 5, 7 etc.). As you can see from the example, as more
components are added, the transition from high to low values in the synthesised
waveform become sharper, and the peaks and troughs become flatter. Eventually,
we have a complete square wave with sharp edges.

The same approach can be taken to generate images. For example, we can create
sinusoidal stimuli with very tightly defined properties (specified bandwidths) in
the Fourier domain. A popular stimulus in computer vision research is the Gabor
pattern, which is a spatially localised sine wave grating. We can generate these
in the Fourier domain by shifting a two-dimensional Gaussian blob (like the
low pass filter in Figure 10.13) away from the origin of Fourier space. This will
produce a Gabor pattern in the spatial domain (see Figure 10.16 for examples).
An interesting observation is that patterns with a small footprint in the Fourier
domain have a large spatial extent in the spatial domain, and vice versa. This
means that small patches of grating have a broader frequency bandwidth than
large ones, and so their orientation and spatial frequency are less clearly defined.
10.11. STIMULUS CONSTRUCTION IN THE FOURIER DOMAIN 199

Figure 10.13: Example of low and high pass filtering on the bughouse image.
The top row shows low pass and high pass filters, in which frequencies in the
lighter regions pass the filter, but frequencies in the darker regions are attenuated.
The lower row shows the resulting filtered images: low pass filtering produces a
blurred image, high pass filtering produces a sharp looking image but without
coarse changes in light and dark.
200 CHAPTER 10. FOURIER ANALYSIS

Figure 10.14: Example of filtering in the orientation domain. The left column
shows a filter that blocks horizontal information, but retains vertical information.
The right column shows the opposite filter.
10.11. STIMULUS CONSTRUCTION IN THE FOURIER DOMAIN 201

1F

1F+3F

1F+3F+5F

1F+3F+5F+7F

1F+3F+5F+7F+9F

1F to 49F

1 3 5 7 9 0 1 2 3 4 5
Frequency (Hz) Time (s)

Figure 10.15: Illustration of Fourier synthesis of a square wave. Left column

shows the Fourier spectrum, right column shows the synthesised waveform
created by inverse transforming the spectrum. Successive rows add additional
components, at odd harmonics of the fundamental frequency. By the final row,
with 25 harmonics, the square wave is well-defined.
202 CHAPTER 10. FOURIER ANALYSIS

Figure 10.16: Gabor stimuli synthesised in Fourier space. The upper row shows
the Fourier spectra, and the lower row the spatial transforms.
10.12. DOING FOURIER ANALYSIS IN R 203

10.12 Doing Fourier analysis in R

So far in this chapter, we have seen that we can:
• Represent waveforms and images by their frequencies
• Filter the Fourier spectrum to change the original signal
• Construct new signals by synthesising a spectrum
In this section we will discuss how to implement these operations in R. Most
of the examples use base R functions, though we will also use some custom
functions, and a function from the signal library. The waveform in Figure 10.3
is loaded in from an external data file as follows:
load('data/EEGdata.RData')
thiswave <- allwaves[1,]

The key function is the fft (Fast Fourier Transform) function. This takes a
vector or matrix as its input, and returns a complex-valued Fourier spectrum of
the same dimensions. Complex numbers are a mathematical convenience, and
contain ‘real’ and ‘imaginary’ components. It is not necessary to fully grasp
the mathematics of complex numbers, but in contemporary implementations
of Fourier analysis, the amplitude and phase information are represented in
Cartesian coordinates by the real and imaginary components of the number. An
optional argument to the fft function, inverse = TRUE, will request the inverse
transform. By convention, we also scale the output of the function by the length
of its input. The following lines of code perform the Fourier transform on the
waveform and confirm (using the is.complex function) that we have a complex
valued output:
output <- fft(thiswave)/length(thiswave)
is.complex(output)

## [1] TRUE
We can determine the frequencies for plotting the amplitude spectrum if we
know the duration of the signal (here it was 1 second) and the sample rate (here
1000Hz). The amplitudes can then be plotted as a function of frequency by
taking the absolute values of the Fourier spectrum (e.g. forcing any negative
values to be positive using the abs function) as follows:
samplerate <- 1000
duration <- 1
frequencies <- ((1:(samplerate*duration))-1)/duration
plot(frequencies[2:500],abs(output[2:500]),type='l',lwd=2)

Note that in the above code (see Figure 10.17 for the output), we plot values
only up to the Nyquist limit of 1000/2 = 500Hz. The spectrum is mirrored
about its midpoint, so the values from an index of 501 onwards are a reflection of
the spectrum plotted in Figure 10.17. Notice also that we begin plotting at the
second index of the vectors containing the frequency and Fourier spectrum data.
204 CHAPTER 10. FOURIER ANALYSIS

1.2
abs(output[2:500])

0.8
0.4
0.0

0 100 200 300 400 500

frequencies[2:500]

Figure 10.17: Example amplitude spectrum from 0 - 500Hz

This is because the first entry in the spectrum, known as the DC component (by
analogy to direct current), often has a much larger amplitude than the other
frequencies. The DC component has a frequency of 0 Hz, that corresponds to
the vertical offset of the waveform (a bit like the intercept term in regression
and ANOVA). Since it is often uninteresting, we have omitted it from the plot
above, but it can be included if required.

To apply a filter to the amplitude spectrum, we can construct one using a

function from the signal package. The fir1 function produces a type of filter
called a finite impulse response (or FIR) filter. This is a commonly used type of
filter in signal processing, that has several convenient properties such as being
symmetrical and also very stable. We can create and plot a FIR filter in the
temporal (e.g. non-Fourier) domain as follows (see Figure 10.18:
cutfrequency <- 15
filter1 <- fir1(samplerate-1,2*cutfrequency/(samplerate/2),type='low')
plot(filter1,type='l',lwd=2)

Rendering the filter in the temporal domain is sometimes useful, as it can be

used for convolution with a signal. Convolution is equivalent to multiplying the
filter with the signal at each consecutive time point. However this can be quite a
slow and inefficient process, especially for long signals. A more efficient approach
is to take the Fourier transform of the filter, and multiply this by the Fourier
10.12. DOING FOURIER ANALYSIS IN R 205

0.05
0.03
filter1

0.01
−0.01

0 200 400 600 800 1000

Index

Figure 10.18: Waveform of a finite impulse response filter.

transform of the signal. This produces the same result, because convolution
in the temporal domain is the same as multiplication in the Fourier
domain. So, we can apply the filter in the Fourier domain, and then take the
inverse transform to view the filtered signal as follows:
# multiply the fourier spectra of the waveform and filter
filteredspectrum <- output*abs(fft(filter1))
# inverse transform and take the Real values
filteredwave <- Re(fft(filteredspectrum,inverse=TRUE))
plot(1:1000,filteredwave,type='l',lwd=2)

The filtered waveform in Figure 10.19 is much smoother than the original, shown
in Figure 10.3.

Next we will demonstrate Fourier analysis in two dimensions. However there is a

small issue that needs dealing with first. Recall that the 1D amplitude spectrum
is mirrored about its centre, with the lowest frequencies at the extremes, and the
highest frequencies in the centre. Well, this is the opposite of how 2D amplitude
spectra are conventionally plotted - with the lowest frequencies in the centre (see
Figure 10.9). To represent the 2D spectrum in the conventional way, we need to
perform an operation called the quadrant shift. In brief, this involves switching
the top left and bottom right quadrants of the spectrum, and the top right and
bottom left quadrants. This means that the low spatial frequencies previously
206 CHAPTER 10. FOURIER ANALYSIS

4
2
filteredwave

0
−2
−4

0 200 400 600 800 1000

1:1000

Figure 10.19: The filtered waveform.

represented in the corners of the spectrum are now represented in the centre.
In many programming languages there is a built in function to implement the
quadrant shift, but in R we need to define the following single line function:
fftshift <- function(im) {im * (-1)^(row(im) + col(im))}

My grasp of imaginary numbers is not sufficient to understand exactly what this

function is doing, so I defer to the mathematical wizards who came up with it
and trust that it does the job.

We first load the image in from a file using the readJPEG function from the jpeg
package. The image is stored as a 512x512x3 matrix. The 512x512 is the size of
the image in pixels (in the x and y directions), and the third dimension contains
three colour channels: red, green and blue. We will just use the information in
the red colour channel and discard the others, so that our image is black and
white.
library(jpeg)
bughouse <- readJPEG('images/bughouse.jpg')
bughouse <- bughouse[,,1]

The image is now stored as a 512x512 matrix of pixel intensities. We can take
the Fourier transform, applying the quadrant shift, as follows:
10.13. PRACTICE QUESTIONS 207

bugspectrum <- fft(fftshift(bughouse))

Now that we have Fourier transformed the image, we can do some more agressive
filtering. Perhaps we could include only oblique orientations within a narrow
range of spatial frequencies, using an oriented bandpass filter like those in Figure
10.16. These are created in the Fourier domain using a short function called
offsetgaus as follows:
offsetgaus <- function(n,std,x,y){
i <- matrix(data = (1-(n/2)):(n/2), nrow=n, ncol=n)
j <- t(apply(i,2,rev))
h <- exp(-(((i+x)^2) / (2 * std^2)) - (((j+y)^2) / (2 * std^2)))
return(h)}

# create a Gabor filter using two Gaussian functions, offset from the origin
g <- offsetgaus(512,8,20,20) + offsetgaus(512,8,-20,-20)

The filter and its spatial transform will look very similar to those shown in Figure
10.16. We then multiply the filter by the Fourier spectrum of the image, and take
the inverse transform, with a bit of quadrant shifting sleight of hand. Finally,
we rescale the luminances to between 0 and 1, and then plot the resulting image
(see Figure 10.20).
# apply the filter and inverse transform
filteredimage <- Re(fftshift(fft((bugspectrum*g), inverse=TRUE)))
filteredimage <- filteredimage - min(filteredimage) # scale the luminances
filteredimage <- filteredimage/max(filteredimage) # to between 0 and 1

# plot the filtered and original image side by side

plot(x=NULL,y=NULL,xlim=c(0,4.5),ylim=c(-1,1),axes=FALSE, ann=FALSE, lwd=2)
rasterImage(filteredimage,0,-1,2,1)
rasterImage(bughouse,2.5,-1,4.5,1)
points(3.75,0.25,pch=1,col=pal2tone[2],cex=8,lwd=8)

The clearest feature in the filtered image is a diagonal plank of wood, which has
the most left-oblique energy. This is circled in blue in the original image (right
panel of Figure 10.20).
This section has provided example code for performing Fourier analysis and
filtering in both one and two dimensions. The practice questions below test your
understanding with further examples.

10.13 Practice questions

1. What operation will convert the Fourier spectrum to a waveform?
A) The Fourier transform
B) The quadrant shift
208 CHAPTER 10. FOURIER ANALYSIS

Figure 10.20: Oblique filtered bug hotel image (left). The strongest feature
corresponds to a diagonal plank of wood, circled in blue in the original image
(right).

C) The inverse Fourier transform

D) The phase spectrum
2. What determines the Nyquist limit of a signal?
A) The sample rate
B) The frequency with the largest amplitude
C) The signal duration
D) The sample rate multiplied by the duration
3. What determines the frequency resolution of the Fourier spectrum?
A) The sample rate
B) The number of samples in the signal
C) The signal duration
D) The sample rate multiplied by the duration
4. What units do we use to measure Fourier phase?
A) Hertz
B) The units the dependent variable is measured in
C) Cycles per degree
D) Degrees
5. Which type of filter would we use to remove only high frequencies from a
signal?
A) A high pass filter
B) A low pass filter
C) A bandpass filter
D) A notch filter
6. Which line of code will convert the data object waveform to its Fourier
amplitude spectrum?
10.13. PRACTICE QUESTIONS 209

A) angle(fft(waveform))
B) abs(fft(waveform,inverse=TRUE))
C) abs(fft(waveform))
D) angle(fftshift(waveform))
7. Which pair of operations are equivalent?
A) Convolution in the temporal domain and division in the Fourier
domain
B) Squaring in the temporal domain and subtraction in the Fourier
domain
C) Addition in the temporal domain and convolution in the Fourier
domain
D) Convolution in the temporal domain and multiplication in the Fourier
domain
8. In Fourier space, the highest spatial frequencies are traditionally repre-
sented:
A) In the corners
B) In the upper half
C) In the centre
D) In the lower half
9. What will the following line of code do? angle(fft(waveform))
A) Return the amplitude spectrum
B) Return the phase spectrum
C) Return the full Fourier spectrum
D) Return a smoothed waveform
10. Which line of code will return a filtered version of the data object signal?
A) abs(fft(fft(signal)*filter,inverse=TRUE))
B) Re(fft(fft(signal)*filter,inverse=TRUE))
C) Re(fft(fft(signal,inverse=TRUE)*filter))
D) abs(fft(signal*filter,inverse=TRUE))
Answers to all questions are provided in section 20.2.
210 CHAPTER 10. FOURIER ANALYSIS
Chapter 11

Multivariate t-tests

Many widely-used statistics are univariate in nature, in that they involve a single
dependent variable (outcome measure). If you have more than one dependent
variable, a number of alternative statistical tests are available that can deal with
all of the dependent variables at once, rather than running a series of univariate
tests. The next four chapters will introduce a selection of these methods, which
are referred to as multivariate techniques.

Multivariate statistics have some advantages over their univariate cousins. In

particular, because they consider more than one dependent variable at a time,
they will typically have greater statistical power for detecting an effect (see
Chapter 5 for an explanation of power). For some research designs, using
multivariate methods also means that a single (omnibus) test can be conducted,
rather than a series of several univariate tests, one for each dependent variable.
This helps to avoid issues with multiple comparisons, and the required corrections,
that we will discuss in more detail in Chapter 15.

In this chapter we will consider a multivariate extension of the t-test, first

introduced by Hotelling (1931). Known as Hotelling’s T 2 , this statistic allows
us to compare a set of multivariate observations to a particular value, or to
compare two sets of multivariate observations (e.g. from two groups). We will
also consider a variant of the T 2 statistic proposed by Victor and Mast (1991),
2
called Tcirc , and discuss situations where this might be used. Because visualising
multivariate data can be challenging, we will introduce the T 2 statistic using
bivariate data, i.e. data with two dependent variables, but the mathematics of
T 2 extends to any number of dependent variables.

211
212 CHAPTER 11. MULTIVARIATE T-TESTS

11.1 Thinking about and visualising bivariate

data
We will begin by considering some possible forms that bivariate data might take.
For these examples, we will refer to our two dependent variables as x and y,
but in principle these can be any two things that we might care to measure.
They could be weight and height, for example, or reaction times and accuracy,
or any other two dependent measures. They can be in any units, and there is no
requirement that the units of the two measures be the same. One key situation
that we will discuss later in the chapter is when the components are the real
and imaginary terms from a Fourier transform (see Chapter 10).
Figure 11.1a shows some example bivariate data that are normally distributed
about zero in both the x and y dimensions, and show no correlation between
the two measures. These data do not differ significantly from the reference point
of 0 (at the origin), and we would ideally like a statistical test that can indicate
this. In Figure 11.1b, the data are displaced in the x direction, but still have a
mean of 0 in the y direction, and in Figure 11.1c they are similarly displaced in
the y direction. In principle, we could analyse the data in these panels using
a univariate test, simply by discarding the uninformative variable. But this is
not ideal, as we have no principled way of knowing in advance which variable
(x or y) is going to be the informative one (if we knew, we probably wouldn’t
bother to measure both!). Finally, in Figure 11.1d the data points are displaced
in both directions. One can envisage a situation where this offset does not quite
reach statistical significance in either the x or y directions when assessed using a
univariate test, but if we could somehow take both variables into account we
might be more likely to detect an effect.
A further complication is illustrated in Figure 11.2. Here, we see similar ar-
rangements as before, but this time with a positive correlation between the two
variables (it could equally well be negative). We will need our statistical test
to be able to take any such correlations into account when assessing statistical
significance, as they account for a proportion of the variance of each measure.
Put another way, when two variables are correlated, we can partly predict one
variable from scores on the other. This effectively reduces the number of degrees
of freedom, which needs to be considered when estimating statistical significance.

11.2 The one-sample and paired Hotelling’s T 2

statistic
A one-sample Student’s t-test calculates the difference between the sample mean
and a fixed point (often zero), scaled by the sample variance. Hotelling’s T 2 takes
a similar approach but with multivariate data. The sample mean is the average
across the x and y (and any additional) dimensions, also known as the centroid.
This is shown by the large point in Figure 11.3. The centroid is compared with
11.2. THE ONE-SAMPLE AND PAIRED HOTELLING’S T 2 STATISTIC213

(a) y (b) y
● ●
●
● ● ●●
● ●
● ●● ● ●
●● ● ● ● ● ● ●●● ● ●●
●● ●
● ● ● ●● ● ● ● ●
●●● ● ● ● ●● ● ●●●● ● ●●● ● ●●●●
●● ●● ●
●●
● ● ● ● ● ● ●●●
● ● ●● ●● ● ● ●●●
● ● ●● ● ●
●●●
● ●●●● ● ●●
●
● ● ●●● ●●●●● ● ●●●● ● ●● ● ●
● ● ● ● ● ● ●
●●●
●
●
● ●● ● ● ●
● ●
● ● ● ● ●●● ● ●● ● ●
●●
● ●● ● ●●
●●● ●●● ●●● ● ●● ● ● ●●● ●●● ●●●● ● ●●
●● ●● ●●● ●●● ●●
● ●●● ● ● ● ●● ●
●●● ●
●●
●●● ●●● ●
● ●●●●
● ● ●
●●● ●● ● ●●● ●
●● ● ● ● ●● ●●● ●●●
●●● ● ●●●
●
●●●
● ● ●● ●● ●● ●●
●
●● ●
●
● ●●●●
● ● ●● ● ●
●●●●● ● ●
●● ●●●
● ●●
● ●● ●●● ●
● ●● ● ●● ● ●●●●
●●
●● ●●●
● ●
●●● ●●●●●●
●●●●●
● ●●● ● ●
● ● ● ● ● ●● ● ●● ● ●● ● ●● ●●●● ●● ●●●● ● ●

x
●● ●●
x ●●
●●●●● ●
● ● ●● ●●●
● ●● ● ●●
●
● ●● ●●
●●●
●●
● ●●
●●●●
●
●
●
●●
●
●●
● ●●●
●
●
●● ●
●●●●
●●●
●●
●●
● ●●●●●●●●
● ●●●
●
●
●
● ●● ●●
●● ●●●●●
●●●
● ●● ●●
●
●●●●●
● ● ●● ● ●
●
●●●●● ●●●●
●●
●
●●
●●● ●● ●
● ●
●
●●●●●●
●
●
●●
●●
● ● ● ●●●
● ● ● ● ● ●● ●
● ●● ●
●
●●●
●●●
●
● ●
●●
●● ●
●●
●
●● ●
● ●●
●●
●●●●●
● ●
● ●● ●
●● ●●●
●● ●● ●● ●●●●
●●
●●● ●●● ● ● ●●●●● ●● ●● ● ●● ● ●●●● ●●●●
● ●●●●●● ●●● ●
●●
● ●● ● ●
●
●●
● ●●●
● ●●
● ●●
● ●● ● ●
●●
●
●●
●
●
●●
●●
●
●●
●●●
●
●
●●
●●●●●●
●●●●●●
● ● ● ●
●
● ● ●● ●
● ●●●
● ●●●
●● ●●●●
●● ●●●
●
●
●●
●●
● ●
●●●
●●
●
●
●
● ●
● ●
●●
●
●●
●●●
●
●●
●
●●●
●
●●
●
●● ● ● ●● ●
●
●● ●●●●●● ● ● ●
●● ●
●●●●●
●●● ● ● ● ● ● ●● ● ● ●● ●
● ●●● ●● ● ●●●●●● ●●●
●● ●● ● ●●●●●● ● ●● ●●●● ● ●
●●●
● ●● ●●●●● ● ● ●●●●●●●●●● ●●●●●● ●
● ●● ●● ●
● ●
●● ●●●●
●●●●● ● ●●● ● ●● ●●
● ●●●
●●● ● ●●
● ● ●
● ●● ● ●●● ●● ●● ●●●
●●
●●●
●● ● ●
●●
●●●
●●● ●
●●
●● ● ● ●
●●●●
●●
●● ●●●●●● ● ●
●● ● ● ●●●
●
●●●●
●
●●●
●
●
●
●
●
●● ●
●●
●
●
●
●●
●●
●
●
●
● ●
●●
●●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●●●
●
●●●
● ●
●●●
●●
●
●
●
● ●
● ●
●
●● ●●●
●●
● ●
● ● ●● ● ● ●● ● ●
●●
● ●
● ●●●
● ●
●
●
● ●●
●●●
●● ●●●
●
●
●
●
●●●●●●●●
●●
●●
●
●● ●● ● ●● ●
● ● ● ●●● ● ●●●●●● ● ●
●●●●●●●●
● ●
●
●●
● ●
●● ●
●●●●
●
●●
●●
●
● ●●●
●
●
●
●
●●●
● ●
●
●●
●
●
●
●
●
●
●
●
●
● ●●●
●●● ●●
●
●
●
●●●
● ●
●● ●●●● ●
●●●●●●
● ●●
●●
●
●●●●●
●● ●●
●
●
●
●●●● ●●
●●
●
●
●
●
●
●●
●●
●●●●●
●●
●
●
●●
●●●
●
●
●●
●●●●●●● ●
●● ●●● ●
●●● ● ●●
● ●●● ●●●●
●● ●
●●● ● ●● ●● ●● ● ●●●●
●
●●● ●● ● ●● ● ●●
●
● ● ● ● ●● ●●●●
● ●● ●● ●● ●●●
●● ●● ● ● ●● ● ●
●●● ●● ●● ●● ●
●
●●
●●●● ●●
● ●●
●●● ●
●●● ●● ●●
● ● ●
●●
●
● ● ●● ●
●
●●●●● ● ●●
● ● ●●●● ●●●●● ●●●
●
●● ●●● ●● ● ●● ● ● ●● ● ●
● ●●●● ●● ● ●●
● ● ●●
●
● ●●●
●
● ●
●
●●●●●
●
●●
●
● ●● ●
●●●● ●●
● ●●
●●●●
●● ● ● ● ●●●● ●●
●●● ●●
●●●●
●
●● ● ●●●
●●● ● ●●●●● ●●● ● ●● ●● ●● ●● ● ●●
● ●● ● ●●
● ●●●● ●●●● ●
●●
● ●●●
● ●●● ● ● ●●● ●● ● ●
●● ●● ●●●
●●
●●●● ●●●●● ●●
●
● ●●● ●
●
● ●● ●●● ●● ●● ●
●● ● ●
● ● ●● ● ●
● ●● ●● ● ● ●
● ● ● ● ● ● ● ●● ● ● ●●●● ● ●●
● ● ●● ● ● ●●● ●●● ● ●●● ● ●● ●● ●● ●●● ●●●● ●
●
● ● ● ●●
● ●● ● ●● ● ● ● ●
● ● ● ●
● ●● ● ● ● ●
● ●
●
● ●
● ● ●
● ●● ● ● ● ●●
● ● ● ●● ● ●
● ● ● ●● ●
● ●●
●
● ●
●
●

●
●● ● ●

(c) ● ●
●

●
●
●● ●

●●●
●
● ●●
●●●
●
●●
●
●
●●
y ●

●
● ●
● ●
●●
●
●
(d) ● y
● ●
● ●
●

●●
●● ●
●
●
●●

● ●●
● ●●
●
●

●
●●
●●
●

● ●
●
●
●
● ●●●●
●

●●
●● ●●●●● ●
●●●
● ●●●
● ●● ● ● ● ●● ● ●●●
●●
●●
●●●
● ●● ●●● ● ● ●● ●●
● ●●● ● ● ●● ●
● ● ● ●
● ● ●● ● ●● ● ●● ●
●● ● ●
● ● ●●
●● ● ● ● ●●●● ●● ●
● ●●
●
●
●● ●●●●●● ●●●● ● ● ● ● ● ●● ● ●●● ●●● ● ●●
● ●●●●●●● ● ●● ● ● ●● ● ●
● ●● ●
● ● ● ● ● ●● ●
●●● ● ●● ●● ● ● ● ●● ●●●●● ● ● ● ●
● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ●● ●● ● ●
● ● ● ● ●● ● ● ● ●●● ● ●
● ● ●●● ●● ● ●● ● ●
● ●● ● ● ●● ●
●● ●●●●●●●●
●●●●●●●●
●●●●
●●
●●● ●●● ● ●● ● ●● ●
● ●●●
● ●●
●●●●●●
●
●
●
●●●●●
● ● ● ●
● ● ●● ●
●● ●●● ●●
●
●●
●●
●●
● ● ● ●●●● ● ●● ●●●● ●● ●● ●●●●● ●
●●●●●
●●● ●
●●
●●● ●●
●
●●●●●● ● ● ● ●●
●
● ● ●● ●●●●● ●
● ●●● ●●●● ●
●● ●● ●●●●● ● ● ● ● ●●● ●●●●
● ●●● ●●● ● ●
●●
●●●●
●● ● ●● ● ●
● ●
●● ● ●●●●●●
● ●●
●
● ●●●●●
●●●●
●● ●● ●● ●●●
● ●●● ● ●● ● ● ● ●● ● ●●
●● ● ●
● ● ●●● ●●●
●
●
●●●
●
●
●●●●●
●●●●
● ●●● ●●●● ●
●
●
● ●● ●●
● ● ●
● ●
●●●
● ●●
● ●
●●●●● ●●●●●● ●
●●● ●●
●●
● ●●● ● ● ●
●●● ●● ●●●
●●●● ● ● ●
●●●
● ●
●●
●●●●
●
● ●●●●●●
● ●● ●●
●●●●
●● ●
● ● ●●●●
●●
●●
●● ● ●●●●● ●● ●●●●●
●
●●
●● ● ●● ●● ●● ● ●● ● ●
●●● ●
●●● ●●
● ●●● ●●
●● ●●
● ● ●●
● ●●● ● ●●● ● ●●● ●● ●●
●● ●●
●●
●
●●
●●●●
●●●●
●●
●● ●●● ●
●●●● ● ●
●
● ●● ●●
● ●●●●●
● ●
●●●
● ●
●●
●
● ●● ● ●●● ●
●●●●●● ●
●●●●● ●● ● ●●
● ●●●●●●● ●●●
●
●
●●● ●●
● ●
●
● ●●●
●●
●●
●
●
●
●
●
●●
●●●●●●●●●
● ●
●●
●●●
● ●●● ● ●
● ●● ● ● ●●
● ●●
●●● ● ●
●
●
●●●
●●●
●
●●
●
●
●
●
●
●
●● ● ●
●
●
●
●
● ●● ●●
● ●●
●
●
●●● ●
● ● ●●●●● ●● ●● ● ●●●●●●
● ●●● ●● ● ● ● ●● ● ● ●● ● ● ●● ●●●● ● ●● ● ●●●● ● ●
● ●● ●● ●●●●●
●
●
●●
●●
●●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●● ●
●●●●●
●●
●
● ●●●●●●●●
● ●● ●
●
●
● ● ●● ● ●
●
● ● ●●
●
●●
●●
● ●
●●
●●●
●●●●●
●●● ●
●●●
●●●
●●●● ●
●
●● ●
●●● ●
●●●● ● ●
● ●● ●●●
●
● ●● ●
●●
●
●●
●●●●●
●
●●●
●
● ●●●
● ●● ●●
●●●
●
●●
●●●●●
● ●●●
●
● ● ● ●●●● ● ●●●●●●●
● ●●● ●●
●
● ●
●
● ●● ●●
●
●●● ●●●
●● ●●● ● ●
● ● ●
●
● ● ●
● ● ● ● ●●●●●
●●●● ●●
●
● ●
●
●●●●●
● ●
●●
● ● ●●●● ● ●●●● ●●
●● ● ● ●●●● ● ●●●●●
●●●●
● ●●●
●
●●
● ●●●● ● ●●
● ●● ● ●
●●
●●● ●
●● ● ● ●● ●● ●●● ● ●● ●● ● ● ●●● ●●
●●● ● ● ● ●●●
● ●● ●
●●●●
●●●●●●●●●●
●●
●●●●
●
●●●●
●● ● ●
● ● ● ●● ●●●
● ●● ●
●● ●●
●
●● ●
● ●●● ●●● ●●●●●● ● ● ●
●● ● ● ● ●●●● ●●●●●● ●●
●● ●●●● ●● ● ●● ●
● ● ●● ●● ● ●● ●
●● ●
●● ●● ● ●●●
●● ● ●●●●●●●●●●● ●
●● ●●
● ●●● ●● ● ●●
●
● ●●●●●●
● ●●● ●
● ●
●● ● ●● ●
● ●● ●● ●● ●● ● ●●● ● ●● ● ●●
●●●●● ●●● ● ● ● ●● ● ●● ●
●
●●● ●● ●●
● ● ● ● ●●
● ● ●● ●● ●● ●● ●● ● ●●
●●● ●● ● ● ●●● ●
●
● ● ●● ● ●●●● ● ●● ●● ● ●●● ●● ●● ●● ● ● ● ●
● ●●●
●●●● ●
● ● ● ● ● ● ● ● ● ● ●● ● ● ●●●●●● ● ●

x x
● ●● ● ●
● ●
● ● ●●●●● ● ● ●●● ● ●●●
●● ● ●● ●● ●● ● ● ● ●
●● ● ● ●● ● ●●
● ●
● ● ● ● ●● ● ● ●● ●●
●
● ● ●● ● ● ● ●
● ● ●● ● ● ● ●● ● ●
● ●● ●
● ● ● ● ● ●● ●●
● ● ●
● ●●
● ● ● ●
● ●
●
●

Figure 11.1: Example scatterplots of 1000 samples of uncorrelated bivariate data,

with different offsets in the x and y directions.
214 CHAPTER 11. MULTIVARIATE T-TESTS

(a) y (b) y
● ●
●
● ●
● ● ● ●●
●● ● ● ●
● ●
● ● ●● ●
●● ●
● ●
● ●
●
● ● ●●
●●●● ● ● ●● ●●
●●
●●● ● ●●●
● ●
●
●●
●●
●
●●●●●● ●
●●●● ●●●
●●
●
●● ●●●●
●
●●
●● ●
● ●
●
● ● ●● ●●● ●● ●●
● ●● ●●● ●
●● ● ●●
●
●
●● ●●●● ●
● ● ● ●●
● ●● ● ●●●
● ●●● ●●
●●●● ●
●●
● ●
●●●
●●●
● ●●●●
●●●● ● ● ● ● ●● ●●
●
●
●
●
●●
●
●
●●
●●
●●
●
●
●
● ●● ●●
● ●● ●● ●●● ●● ●● ● ● ● ●
● ●●
●
● ● ●●●●●●
●● ●
●●
●●●●●●●●
● ●●●●
● ●●● ● ● ● ●●
●●●
●●
● ●●
●●
●
●●●
●●
● ●
●● ● ●●
● ●●● ● ●● ●
●
●
● ●●●●●●
● ●●●●
● ●●● ●
●●●● ● ●●●
● ●●●
●●●● ●●
●
●
●●
●●●●
●●
●
●●●
● ●● ●●●
●● ●
●●
● ●●
●●●● ●●● ●●●●●●
●●● ●●
●
●●●
● ●
●●●● ● ●● ●●● ●● ● ● ●●

x x
● ● ●●●
●
●● ●●●●
● ● ●
●
●●
●
●●● ● ●
●●● ●●●●
●
● ●
● ●
●●●
●●●●●●● ●
●● ●
● ●
●●
● ● ●●●
● ● ● ● ●
●
●●● ●● ●●
●
●● ●●● ●
●●●
●●●
●●
●
●●●
●
●●
●●
●
●
●● ●●●●●
● ●● ● ●● ●●● ●●
●
●●●
●●
●●
●
●●
●
●●
●
●●
●
●●
●●●
●● ●●
● ●
●●●●
●
●
●
●
●●
●
●
●
●
●
● ●●
●
●
●●
●
● ●●
● ●●
●●●
●●● ● ●● ●●●
●●
●● ●
●●●
●
●●●
●●●
●
●●●●
●
●●●
●
● ●●●●●●
●●●● ●
●●●
●
● ●●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
● ●●●●●● ●● ●●●● ●●
●●●
● ●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●●●
● ●●
●●● ●
● ●
●●
●●
●
● ●●
●●●
●●●●
● ● ● ● ●●
●●
●
●●
●●●
●●
●●●
●●● ● ●●
● ●
● ●● ●
●
●
●●●
●
●
●
●●
●
●
●●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●●●●●
●
●
●●●
●●● ●● ●
●
●
●
●●
●●●
●
●
●
●●
●
●
●●
●
●
●●●
●●
●
●
● ●
●
●●●
●
●
●
●●
●
●
●
●●
●●●●
●
●
●●
●
●
●●
●●●
●
●●●
●●
●
●●
●●
●● ●
●●
●●●
●
● ●
●●●
●
●
●● ●●● ●● ● ●● ●●
●
●●
●●
●●
●
●
●
●
●
●●●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●●●●●
●
●
●●
●
●●
●
●● ●
●
●●● ● ●●
● ●●●●●
●
●●●●●● ● ● ● ●●● ●●●
●●
●●
● ●
●●●
● ● ●● ● ●
●
●●●●●
●●
●●● ●●
●●●
●
●
●●●●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
● ●●●●
●
●● ● ●●● ● ●●
●● ●●
●
●
●
●
● ●
●●●
● ●
●●
●●●
●
●●
●●
●●
●●
●
●
●●
●●●
●
●●●●●●●●
● ●●●●
●● ●
●● ●●● ●● ●
●● ● ●
●● ●
●●
●●●
●●
●●
●
●●
●
●●
●●●
●
● ●
●
●●
●●
●●●● ●
● ●● ● ●
●● ●●
●
●●●
●
●●
●
●
●
● ●● ● ● ●●● ●●●●
●●●
●
●
●
●
●
●
●
●
●
●●●
●●
●●
●
●●●●●
● ●●●
● ● ●
●●
●
●● ●
●●
●
●●●
●
●●●●●
●●●●●
●●
●●● ● ● ●● ● ●●
● ●●●
● ●●
●
●
●
●
●
●●
●●
●
●
●
●●
●●
●
●●
●●●● ●●
●● ●●
●●● ●●
●●●
● ●●
● ● ●●● ● ●●●●
●
●
●●●●
●
● ●●●●
●● ●●●● ● ●●● ● ●●●●
● ●● ●
●●●●●● ● ●
● ● ●● ● ●●●
●●●●
● ●●●●●
●● ●
●
●●●
●●●●●●
●
●● ● ●●
●
● ●●● ●
●●
●●●●● ●●● ●● ●● ●●●●● ●●●● ● ●
● ●●●●
● ●● ●●●● ●● ● ● ●●●● ●●● ●●
● ●●
●●●●
●
●
●
●●● ●
● ●●●
● ● ●●●●
●●●● ●●
●
●
●
●
●●
●●●● ●● ●● ●● ● ●
●
●
● ● ●● ● ●
● ●● ● ●●●
● ●●●●●
● ●● ● ●● ● ●●●●● ●
●
● ● ● ●●●● ●●●● ●●
● ● ● ●
●●
● ●● ●
●●
● ●●
● ●
● ● ●● ● ● ●●●●
● ● ●
● ● ●●
● ●●
● ●
● ●
●

● ● ●●
● ● ● ●●●

y ●● ● ●
y
● ●

(c) (d)
●● ●● ●●
●●●●●● ●● ● ● ● ●
● ●●
●● ●● ●● ● ●●
● ● ● ●
●● ●●●●●●●● ●
● ● ● ● ● ●● ●● ● ●
●●● ●●●
● ●●●● ●
●●●
●●● ● ● ● ●●●●
● ●● ●●
●
●●
●
●● ●
●●● ●● ● ●●●● ●●●●●●●●●
●●●
●●● ●●●● ●●
●●●●
●●●●●●
●●
●●
● ● ●● ●
●
● ●
●●●
●
●●● ●● ●
●●●● ●
● ●●
●● ●●
●●
●●●● ●● ●●● ●●● ●●●
● ●
●●
●●
● ●●●●●
●
●●
●●●
●● ●●●●●●
●● ●
●●
●●
●●
●
●●
●
●●●
● ●●
●
●● ●●● ●●
●●
●
●●
●●
●
●
●●
●●
●
●
●
● ●●
●●●●●●●●
● ●● ●● ●●●●● ● ● ● ●● ●● ●
● ●●● ● ●
●
● ● ●
● ●●●●
●●●● ●
●●● ●● ● ● ● ● ●●
●●
●●● ●
●● ●●●●
● ●● ●●●●
●
●● ●●●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●●●
●●
●●
●
●
●
●
●
●●
●
●
●●
●
●
● ● ●●●
●●●● ●
● ●●●●●●
● ●
●
●
●●
●●
●
●●
●
●
●
●
●
●●
● ●
●●
●●●●
●●
●●●●
●●●● ● ●●
● ●●●●
● ● ●
●● ● ●●
●●●●
●●
●
●●●●
● ●●●
●●●●●●●
● ●●●
● ● ●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●
●
●
● ●
●
●●
●
●●
●●● ●
●
●● ●●●
●●●●
● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●●
●●●●
●●
●●
●●●●
●●● ●
●●●
● ● ●
●
● ●●
●
●●
●
●●●●●
●● ●
●●●
●●● ●●
● ●
●●●●
●
● ●
●
●●●●
●
●
●●
●
●●
●
●●
● ● ●●●
● ●●
● ●●●●●●
●
●
●●
● ●● ● ●● ●●●
●
● ● ●●●●● ●●
●
●●● ●●●●● ●●
● ●
●
●
●●●
● ●●
●●
● ●● ●
●
●●●● ● ●● ●●●
● ●
●●
●
●●●●
●
●●
●
●●
●●
●●
●
●●
●
● ●● ●
● ●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●●●●
● ●●●●● ● ● ●●●●
●●
●●
●
●
●●
●
●
●●●
●●
●
●
●●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●●
●●● ●
● ●● ● ●●●●●
●● ●
●
●●
● ● ●
●
● ●
●●●
● ●● ●● ● ●●
●● ●●●
●●
●
● ●
●
●●●●●
●● ●
●
●● ●●
● ●●●●
●●● ●
●
●●●
●
●
●
●
●●
●●● ●
●●●
●
●●
●●●
●
●
●●
●●
●
●
●●● ● ●● ● ● ● ●
●●●●●●
●
● ●
●●
●●
●●
●
●
●
●● ●
●
●●
●●
●●● ●●●
●
●
●●●
●
●●●
●●● ●
●●
●
●
●● ●
●
●
●●
● ●●● ● ●● ●
●
●
●●●
●●
●
●
●●
●●●
● ●
●●
●●
●●
●●●
●
● ● ●
●●●
●●● ●
●
●
●●●
●●
● ●
● ● ● ● ●●●● ●●
●
●●●●●
●●
● ●●●
● ●●●
●
●● ●●
●●●
●●● ●
●●●
● ●
●●●
●
●
●●
●
●
● ●
●
●●●● ●● ●● ●●●●●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●
●
●●●● ●
●● ●
● ●●●
●●
● ●
●●●
●●
● ●
●●
● ●
●● ●
●●
●
●●● ●
●● ● ●
● ●●●●
●●●
●●
●●
●
●●●●
●●●●● ● ●
●●●●●
●● ●●
●●●●
● ●●● ●
●
●●
● ●●●●●
●●● ● ● ●●
● ●●● ●
●
●●●
●●
● ●
●
●●●●
●
● ●●●●●
●●●●
●
●●
●●
● ● ●●
●●●
●●●●●●
●
● ●● ●● ●●●
●
●● ●●
●●●●●●
●●●
●●●●● ●
●●●●
● ●● ●
●● ●
●●● ●●●●● ●
●●
●● ● ●●●●●
● ● ●
● ●● ●●●●●
●
●●●● ●
●●●
● ●●
● ● ● ●●●●●●●
● ●●
●
●●●● ●
● ●● ●●●●
●● ● ● ●●● ● ● ●
●● ●
●● ● ●●●●●●●
●●● ●●
● ●
● ●
● ● ●● ●● ● ●
● ●
●● ● ●

x ●●
x
● ● ● ●
●●●●●
● ●● ● ●●●●●●● ●
●
●●● ●
● ●●
● ● ● ●● ● ●
● ● ●●●● ●●●
● ●● ● ● ●● ● ● ●● ●●● ●● ● ●
●●● ●●●●● ●
● ● ● ●● ●●● ● ● ●
●● ● ● ●●● ●● ● ●
● ●●●
●●
● ● ● ● ●● ●●●
● ●● ●
● ● ●
● ●
●
●
● ●

Figure 11.2: Example scatterplots of 1000 samples of positively correlated

bivariate data, with different offsets in the x and y directions.
11.2. THE ONE-SAMPLE AND PAIRED HOTELLING’S T 2 STATISTIC215

some other point in the space, for example the origin (x = 0, y = 0) in this
example. The distance between the two points is the length of the vector that
joins them, shown by the black line. The variance term is calculated from the
lengths of the residuals. These are the thin grey lines that join the mean to each
data point. Also included in the variance term is the covariance between the two
variables, which is best thought of conceptually as the correlation between them.

y ●
●
●
● ●
● ●
●● ●● ●● ●
● ●
● ● ●
● ●
● ● ● ●
● ● ● ●
● ● ●●
●● ● ●●
● ●
● ●
●
● ●
●
●
● ●
●
●
● ●● ●
● ● ● ● ● ● ● ●
●
●● ●
● ●● ●
● ●
● ●● ●
● ●
● ● ● ● ●
●
●● ● ● ●
●
● ●
● ●
●

x ●

Figure 11.3: Example scatterplot showing the sample mean (black point), vector
line between the sample mean and the origin (black line), and residual lines
joining each data point to the sample mean (grey lines).

The equation for calculating T 2 is:

T 2 = N (x̄ − µ)′ C −1 (x̄ − µ), (11.1)

where N is the sample size, (x̄ − µ) is a vector of differences between the sample
mean (x̄) and the point we are comparing it to (µ; i.e. the black point and the
origin in Figure 11.3), C is the covariance matrix (and C −1 its inverse). The
tick symbol (′ ) indicates transposition of the vector. Calculating the inverse
covariance matrix is impractical by hand, so it is always done by computer.
216 CHAPTER 11. MULTIVARIATE T-TESTS

However I have included the equation here so that you can see the role the
covariance matrix plays in calculating the test statistic.

As a reminder, covariance matrices have the following structure:

## x y
## x 0.904 -0.924
## y -0.924 1.221

The values on the diagonal of the matrix (x,x and y,y) give the variance for each
of the two variables (which must always be positive). The off-diagonal values
(x,y and y,x) give the covariance between the two variables (note that both these
values are identical, and may be negative as in the above example). All of the
values are in the original units of measurement - if the matrix is standardised, it
becomes a correlation matrix. The covariance matrix fully describes the variance
and covariance of a multivariate data set.

By including the inverse covariance matrix in the calculation, the T 2 statistic is

effectively decorrelating the dependent variables, and rescaling them to have equal
variance. This means that increases or decreases in the amount of correlation
between the variables does not affect the test’s statistical power (all else being
equal). It also means that the test is assessing only the difference between the
centroid and the comparison point - it is not telling you if the variables are
significantly correlated. Of course you could find this out by running a standard
correlation test if you need to know if your variables are correlated.

To determine statistical significance, the T 2 statistic is converted to an equivalent

F-statistic by multiplying by a scaling factor based on the number of dependent
variables (m) and the sample size (N ):

N −m 2
F = T (11.2)
m(N − 1)

The expected F-distribution then has m and N-m degrees of freedom. A p-

value can be estimated by comparing the calculated F statistic to the expected
F-distribution, in much the same way as for ANOVA.

For repeated measures designs (where the same participants complete two
different conditions), a paired-samples version of T 2 is achieved by subtracting
each participant’s scores across the two conditions, and performing the one
sample test comparing to zero. (It is not always appreciated that for univariate
t-tests, a paired samples test is identical to a one-sample test conducted on the
differences between the conditions). Furthermore, the same approach works with
an arbitrary number of dependent variables (m > 2), making T 2 a multivariate
(rather than a bivariate) statistic.
11.3. EXAMPLE: MULTIVARIATE ANALYSIS OF PERIODIC EEG DATA217

11.3 Example: multivariate analysis of periodic

EEG data
To demonstrate how to use the T 2 test on real data, we will re-analyse two con-
ditions from a data set reported by Vilidaite et al. (2018). This is an EEG study
that used the steady-state evoked potential method. In this paradigm a sensory
stimulus oscillates (flickers) at a fixed frequency, and neurons responsive to the
stimulus modulate their firing at the same frequency. These modulations can
be detected as electrical fluctuations at the scalp using electroencephalography
(EEG). Steady-state methods are widely used in research into visual and auditory
processing, and have the advantage that they do not require participants to
make responses, and so can be used in infants, animals, and patients who are
nonverbal.
The typical approach to analysing steady-state data is to take the Fourier
transform (see Chapter 10) of the EEG waveform, and look at activity at the
stimulus flicker frequency. The signal is represented using complex numbers,
that can be plotted as x (Real) and y (Imaginary) coordinates in a Cartesian
space, and analysed using multivariate statistics. Data for 100 participants at
two different stimulus levels are shown in Figure 11.4. In panel (a) the stimulus
level was 0 (the baseline condition), so we do not expect to see a signal. In panel
(b) the stimulus level was 32% contrast, so we anticipate a measurable signal.
We will conduct a one-sample T 2 test on each data set, and then a paired samples
T 2 test comparing the two conditions. For the baseline condition there was no
significant effect (T 2 = 1.12, F(2,98) = 0.56, p = 0.575). For the 32% contrast
condition there was a highly significant effect (T 2 = 40.19, F(2,98) = 19.89, p
< 0.001). We can also conduct a paired samples T 2 test to compare these two
conditions, which again produces a highly significant difference (T 2 = 42.29,
F(2,98) = 20.93, p < 0.001).

11.4 The two-sample (independent) Hotelling’s

T 2 statistic
An independent two-sample version of the T 2 statistic is also possible, using
a slightly modified formula. For this version of the test, our two groups can
have different numbers of observations and should be comprised of different
individuals, though they must always involve the same dependent variables. The
equation for the two-sample T 2 statistic is defined as:

N1 ∗ N2
T2 = (x¯1 − x¯2 )′ C −1 (x¯1 − x¯2 ), (11.3)
N1 + N2

where x¯1 and x¯2 are the vectors of sample means for the two groups, and N1
and N2 are the sample sizes. The covariance matrix (C ) is the pooled covariance
218 CHAPTER 11. MULTIVARIATE T-TESTS

Baseline 32% target

●
●

(a) 0.3 (b) 1 ●

●
●
● ● ●●
●
● ● ●
● ●● ●
●
● ●
● ●●
● ●
●
● ●● ●
● ●
●
●
● ●
●
● ● ● ●
● ● ●
● ●
●
● ● ● ● ● ●●
●●
●● ● ● ●
●● ●● ● ●
Re ●●
●
●●
●
● ●●●
●
●●
●
●● ● ●
0.3 Re ●
●
●
●
● ●
● ●● ●
●
●
●
● ●
●
● 1
● ●●● ●● ● ● ●● ● ●● ●
● ●●● ●
● ● ● ●
● ● ●
● ● ●● ●● ●●
●
●
● ●
●
● ●
●
●●
● ● ● ●
● ● ●
● ● ● ●
●● ●●● ● ●
● ● ●
● ● ● ●
● ● ●
● ●
● ●● ●
●
● ●
●
●
●
● ●

Im Im

Figure 11.4: Example SSVEP data. Each blue point is an individual participant
(N=100), the black points are the group means, and the orthogonal lines show
the eigenvectors of the bounding ellipse. Panel (a) shows data from the baseline
condition where no stimulus was shown, panel (b) shows data from a condition
where 32% contrast sine wave grating patches flickered at 7Hz. Both data sets
are from the 7Hz frequency bin of the Fourier spectrum of the EEG data recorded
at the occipital pole. The x-axis represents the real component, and the y-axis
the imaginary component of the complex number.
11.5. EXAMPLE: VISUAL MOTOR RESPONSES IN ZEBRAFISH LARVAE219

matrix across the two samples, taking sample size into account:

(N1 − 1)C1 + (N2 − 1)C2

C= , (11.4)
N1 + N2 − m

where m is the number of dependent variables, and C1 and C2 are the covariance
matrices of the two groups. For the two-sample version, the F-ratio is calculated
as:

N1 + N2 − m − 1 2
F = T , (11.5)
m(N1 + N2 − 2)

and the degrees of freedom are given by df1 = m and df2 = N1 + N2 − m − 1.

Conceptually, this test determines the distance between the centroids of the
two groups, taking into account their variances and covariances. It therefore
allows pairs of conditions to be compared statistically. If you have more than
two conditions, it is possible to run a MANOVA (multivariate analysis of
variance) - the multivariate extension of ANOVA. This method is covered in
many introductory statistics texts, so we will not discuss it further here. However
if you have conducted a MANOVA, the T 2 test can be used for post hoc tests,
assuming appropriate correction for multiple comparisons is applied (see Chapter
15).

11.5 Example: visual motor responses in ze-

brafish larvae
A study by Liu et al. (2015) measured the locomotor responses of zebrafish larvae
to a sudden onset or offset of light. They used an infra-red camera to record
arrays of 96 larvae simultaneously, between 3 and 9 days after fertilization, and
for different wild-type genetic strains. The visual motor response is a widely-used
assay of neural function that can be used to study development, or assess the
effects of different drugs or genetic mutations on the nervous system. The data
collected are rich and high dimensional, and can be analysed in many different
ways. In the Liu et al. study, the authors calculated a burst duration index for
each one second of time, which summarised the proportion of that time window
that the animal was moving. They then conducted two-sample Hotelling’s T 2
tests using the burst duration index for 30 second time windows to compare
different developmental timepoints and different genetic strains (note that the
use of 30 time points means that there were 30 dependent variables for most of
their tests).
Figure 11.5 shows a reanalysis of a subset of the zebrafish data, that are publicly
available at: https://doi.org/10.7910/DVN/HTXXKW. The left panel shows
the timecourse for two developmental stages (6 and 9 days post fertilization), for
220 CHAPTER 11. MULTIVARIATE T-TESTS

(a) (b)
6 DPF
0.06
9 DPF 1
BDI

0.04

Activity count
●
0.02
●
0
0.6
Count

0.4
Dark
0.2 0 ● Light

0
−30 −20 −10 0 10 20 30 0 0.1
Time (s) BDI

Figure 11.5: Zebrafish larvae visual motor reflex data from Liu et al. (2015).
Panel (a) shows the timecourse for the burst duration index (BDI; upper) and
the average activity count (lower), for larvae 6 (black) and 9 (blue) days post
fertilization (DPF). Shaded regions indicate 95% confidence intervals across 192
individuals, and the vertical dashed line indicates light onset. Panel (b) shows
the bivariate means across both variables (BDI and activity count) for the 1
second period before (dark, squares) and after (light, circles) the light stimulus
onset, again for 6 (black) and 9 (blue) days post fertilization.
 2
11.6. THE TCIRC STATISTIC 221

two measurement indices - the burst duration index (top) and the burst count
(bottom). The right panel illustrates the two measures plotted against each
other at two time points (one second before or after light onset). Two-sample
T 2 tests indicate no difference at the time point immediately before stimulus
onset (squares; T 2 = 3.45, F(2,381) = 1.72, p = 0.18), but a significant effect
one second after the light was presented (circles; T 2 = 8.85, F(2,381) = 4.41, p
= 0.01). This suggests that older larvae have a slightly weaker initial response
to light, though it is clear from Figure 11.5a that movement persists for longer
in the 9 day old larvae. Overall, the Liu et al. (2015) study is a good example
of how multivariate statistics can be used to analyse complex data sets.

2
11.6 The Tcirc statistic
Victor and Mast (1991) proposed a variant of the T 2 statistic called Tcirc
2
(the
circ is short for circular). This was intended specifically for analysing complex
Fourier components like those we encountered in Figure 11.4. The test has some
additional assumptions - specifically that the units of the dependent variables
have equal variance, and that there is no correlation between them. In other
words, the data should conform to a circular cloud of points (as in Figure 11.1)
and not an ellipsoidal one (as in Figure 11.2). If these conditions are met, the
one-sample verison of the statistic is calculated as:

2 |x̄ − µ|2
Tcirc = (N − 1) (11.6)
Σ|xj − x̄|2

where N is the sample size, x̄ is the sample mean, µ is the point of comparison,
and xj represents individual observations. The vertical slash symbols ( | | )
denote the absolute value of the numbers inside (i.e. the vector lengths). In
words, this equation takes the squared length of the line joining the sample mean
to the comparison point (i.e. the black line in Figure 11.3), and divides by the
sum of the squared residuals (i.e. the grey lines in Figure 11.3).
Note that crucially there is no covariance term in this equation, which makes it
substantially simpler to calculate. As with the original T 2 statistic, statistical
significance is estimated by comparison with an F-distribution, which for two
2
dependent variables has 2 and 2N-2 degrees of freedom for F = N Tcirc . Repeated
measures and two-sample versions are also possible.
2
Victor and Mast (1991) demonstrate that the Tcirc statistic can be more sensitive
2
(i.e. have greater power) than Hotelling’s T when its assumptions are met.
However there is an issue with the false positive rate when the assumptions
are violated (i.e. when the variables are correlated or have different variances).
I recently (D. Baker 2021) proposed a method for testing the assumptions,
that involves comparing the condition index of a data set to that expected
by chance. The condition index is the square root of the ratio of eigenvector
lengths (eigenvectors are the axes of a bounding ellipse, see examples given by
222 CHAPTER 11. MULTIVARIATE T-TESTS

the grey lines in Figure 11.4). This functions like other assumption tests, in that
2
a significant result means that the Tcirc should not be used, and Hotelling’s T 2
is a safer alternative.

11.7 Mahalanobis distance as an effect size mea-

sure for multivariate statistics
Back in section 3.4.3, we encountered a statistic called the Mahalanobis distance
(Mahalanobis 1936). This was like a multivariate version of a z-score, in that it
told us the distance between a single data point and the sample mean, taking
into account the sample variance and covariance. For the one-sample situation,
we can again use this statistic to calculate the distance between the multivariate
sample mean (centroid) and a comparison point (such as the origin). This gives
us a standardised measure of distance that is a multivariate generalisation of
the Cohen’s d statistic (see section 5.2). It is therefore an appropriate effect size
measure to include when reporting the results of the T 2 and Tcirc 2
statistics.
For the two-sample situation, there is a variant of the Mahalanobis distance that
can be applied for two independent groups, often called the pairwise Mahalanobis
distance. Just as the two-sample T 2 test combines the variances from the two
groups, we must do the same for the pairwise Mahalanobis distance. Indeed, the
equation is closely related to the two-sample T 2 equation (eqn. (11.3)), and is
given by:

p
D= (x¯1 − x¯2 )′ C −1 (x¯1 − x¯2 ), (11.7)

where all terms are as defined previously, and C is the pooled covariance matrix
calculated using eqn. (11.4). Note that some implementations of the Mahalanobis
distance actually return D2 , which can be converted back to D by taking the
square root (as in eqn. (11.7)). As with Cohen’s d, the D statistic is standardised
so it can be compared across different data sets, studies, and dependent variables,
and could in principle be used as an effect size for meta analysis (see Chapter 6).
I strongly recommend reporting it alongside the results of any T 2 or Tcirc2
test.

11.8 Calculating Hotelling’s T 2 in R

It is surprisingly hard to find a working implementation of a one-sample
Hotelling’s T 2 test in R. Several packages exist that contain one, including
the ICSNP and MVTests packages, however both of these are deprecated at
the time of writing and do not work with recent versions of R. As part of a
paper on analysing periodic data using multivariate statistics (D. Baker 2021), I
created my own package that implements all of the tests we have discussed in
this chapter.
11.8. CALCULATING HOTELLING’S T 2 IN R 223

The FourierStats package is hosted on the code repository GitHub, at https:

//github.com/bakerdh/FourierStats (for an overview of GitHub, see section 19.1).
Because it is on GitHub rather than the CRAN repository, we need to install it
in a slightly different way from normal, using a function called install_github
from the devtools package as follows:
install.packages('devtools')
library(devtools)
install_github("bakerdh/FourierStats")
library(FourierStats)

The package contains a function called tsqh.test that can calculate one-sample,
two-sample and repeated measures versions of Hotelling’s T 2 test. Let’s assume
that our first data set is stored in an N × 2 array called data:
head(data)

## [,1] [,2]
## [1,] 0.9796330 1.83289050
## [2,] 0.5984391 -0.65855164
## [3,] 2.3306366 1.44780388
## [4,] -0.2141754 -0.06080895
## [5,] -0.1746285 -0.28229951
## [6,] -0.7105413 -0.68882363
We can conduct a one-sample T 2 test using the tsqh.test function as follows:
tsqh.test(data)

## tsq Fratio df1 df2 pval method

## 1 40.18827 19.89116 2 98 5.618178e-08 One-sample T-squared test
If we wish to compare to a specific point in the two-dimensional space, we can
define this by adding the optional argument mu:
tsqh.test(data,mu=c(0.25,0.25))

## tsq Fratio df1 df2 pval method

## 1 7.933029 3.926449 2 98 0.02288985 One-sample T-squared test
For the one-sample case, we can calculate the Mahalanobis distance using the
built-in mahalanobis function from the stats package:
D2 <- mahalanobis(c(0,0),center=colMeans(data),cov=cov(data))
sqrt(D2)

## [1] 0.6339422
We are passing to the function the two points we wish to compare - the data
centroid (calculated using the colMeans function), and the comparison point
(0,0). We also provide the covariance matrix from the data (cov(data)). Note
224 CHAPTER 11. MULTIVARIATE T-TESTS

that the function returns the squared distance, so we must take the square root
to find D. If we want to compare to a different point, we can change the input
to the first argument, for example:
D2 <- mahalanobis(c(0.25,0.25),center=colMeans(data),cov=cov(data))
sqrt(D2)

## [1] 0.2816563
To compare two groups, we can again use the tsqh.test function, providing it
with both data sets, and specifying either a paired or unpaired test:
tsqh.test(data,y=baseline,paired=TRUE)

## tsq Fratio df1 df2 pval method

## 1 42.28931 20.93107 2 98 2.696304e-08 Paired T-squared test
tsqh.test(data,y=baseline,paired=FALSE)

## tsq Fratio df1 df2 pval method

## 1 41.21264 20.50225 2 197 8.152384e-09 Independent samples T-squared test
Calculating the Mahalanobis distance for the paired samples case again uses
the mahalanobis function, but this time on the difference between the data sets,
comparing to (0,0) (recall that a paired test is identical to a one-sample test on
the differences):
diff <- data-baseline
D2 <- mahalanobis(c(0,0),center=colMeans(diff),cov=cov(diff))
sqrt(D2)

## [1] 0.6503023
For the independent samples (unpaired) case, we instead use the pairwisemahal
function from the FourierStats package. The function expects the data to be
stored in a single matrix, with an additional grouping variable to identify which
group each observation belongs to. We can combine our two data objects using
the rbind function, and generate the group indices with the rep function:
# combine both data sets into a single 200x2 matrix
alldata <- rbind(data,baseline)
# create group labels of 100 1s and 100 2s
grouplabels <- rep(1:2,each=nrow(data))

Then both of these new data objects are passed to the pairwisemahal function:
pairwisemahal(alldata,grouplabels)

## 1 2
## 1 0.0000000 0.9078837
## 2 0.9078837 0.0000000
11.9. PRACTICE QUESTIONS 225

Note that this function returns D (like it should!) and not D2 , so there is no
need to take the square root. It returns a data object that is structured like a
correlation matrix, showing the pairwise distance between each pair of groups.
This allows you to pass in any number of groups, and obtain a full matrix of
distances.
The FourierStats package also contains a function called tsqc.test, that imple-
2
ments the Tcirc test. The syntax is identical to that for tsqh.test, so these
functions can be used interchangeably (though note that tsqc.test only works
for bivariate data, whereas tsqh.test can cope with any number of dependent
2
variables). However, in order to justify running a Tcirc test, we should first test
the condition index of each data set. The function CI.test runs the condition
index test as follows:
CI.test(data)

## CI N criticalCI pval
## 1 1.484294 100 1.282 0.0005631189
A full explanation of how this test works is given by D. Baker (2021). However
you can think of it as being similar to other assumption tests you might be
familiar with (see section 3.8), such as Mauchly’s test of sphericity that is
used to test the assumptions of repeated measures ANOVA, or Levene’s test of
homogeneity of variances. Just like these other assumption tests, if the condition
index test is significant at p < 0.05 (as it is above), then the assumptions of
2
Tcirc are violated, and we should instead run the T 2 test.
These are the basics of how to calculate the T 2 and Tcirc
2
statistics, and the
Mahalanobis distance in R. They are quite rarely used tests, and my hope is
that by including them here more people will know about and use them in the
future. Readers interested in the implementation of the tests are welcome to
inspect the code underlying the FourierStats package for further insights.

11.9 Practice questions

1. Multivariate tests are necessary when you have:
A) A single independent variable
B) A single dependent variable
C) More than one independent variable
D) More than one dependent variable
2. When is it appropriate to use Hotelling’s T 2 statistic?
A) With two or more dependent variables in any units
B) With only two dependent variables in any units
C) With two or more dependent variables which must have the same
units
D) With only two dependent variables which must have the same units
3. Hotelling’s T 2 statistic takes into account the:
226 CHAPTER 11. MULTIVARIATE T-TESTS

A) Means and covariances

B) Means, variances and covariances
C) Means and variances
D) Variances and covariances
4. The significance of a T 2 statistic is determined using:
A) A normal distribution
B) An F-distribution
C) A t-distribution
D) A Poisson distribution
5. The degrees of freedom for Hotelling’s T 2 depend on:
A) The sample size and number of groups
B) The sample size only
C) The sample size and number of dependent variables
D) The number of groups and the number of dependent variables
2
6. The Tcirc statistic assumes that the dependent variables:
A) Are uncorrelated and have equal variance
B) Are correlated and have different variances
C) Are uncorrelated but have different variances
D) Are correlated and have equal variance
7. The lengths of the major and minor axes of an ellipse are called the:
A) Eigenvectors
B) Eigenvalues
C) Condition index
D) Eigenmatrices
8. Which effect size would be the most appropriate for summarising the
difference in means between two independent groups of multivariate data?
A) Cohen’s d
B) The Mahalanobis distance
C) The pairwise Mahalanobis distance
D) The z-score
9. When running a two-sample Hotelling’s T 2 test, how is the covariance
matrix calculated?
A) It is the covariance matrix of the first sample
B) It is the covariance matrix of the second sample
C) It is the covariance matrix of the difference between the samples
D) It is the pooled covariance matrix across both samples
10. If the dependent variables are correlated, what effect does this have on the
power of the Hotelling’s T 2 test?
A) There is no effect on the power
B) The power will generally increase
C) The power will generally decrease
D) It will depend on the specific data set
Answers to all questions are provided in section 20.2.
Chapter 12

Structural equation
modelling

Structural equation modelling is a technique that allows us to make sense of

the relationships between different variables. This is achieved by creating a
model that specifies how the variables are connected to each other, and how
they relate to hidden internal constructs (called latent variables) that we cannot
measure directly (without any latent variables, the method is called path analysis).
Examples of latent variables are things like intelligence, belief, political leanings,
personality, quality of life, and nationality or other group affiliation. These
are concepts that we might be interested in understanding, but can only infer
by measuring more explicit observable variables - things like responses on a
test or questionnaire, income, health records, or voting behaviour. Structural
equation modelling is often what researchers really want to do when they run
multiple correlations between several different pairs of variables. It allows us
to explicitly compare different models that might reflect competing theories or
hypotheses about a situation. Because structural equation modelling is used in
situations with many different measures (e.g. multiple dependent variables), it is
an example of a multivariate technique. It is best introduced using a concrete
example.

12.1 How are different mental abilities related?

Our example uses a classic data set from the literature on the testing of mental
abilities. The Holzinger and Swineford (1939) data set contains test scores from
around 300 teenagers at two different schools, each of whom completed a series
of tests measuring performance for different tasks. There were 26 different tests
measuring different facets of mental ability, including visual ability, literacy, and
performance under timed conditions. For our example, we will consider a subset

227
228 CHAPTER 12. STRUCTURAL EQUATION MODELLING

Table 12.1: Summary of variables in the Holzinger and Swineford data set.

Variable name Test content

x1 Visual perception
x2 Cubes
x3 Lozenges
x4 Paragraph comprehension
x5 Sentence completion
x6 Word meaning
x7 Speeded addition
x8 Speeded counting
x9 Speeded discrimination

of 9 of the tests from the full study. Here is a snippet of the data set:
## id sex ageyr agemo school grade x1 x2 x3 x4 x5 x6
## 1 1 1 13 1 Pasteur 7 3.333333 7.75 0.375 2.333333 5.75 1.2857143
## 2 2 2 13 7 Pasteur 7 5.333333 5.25 2.125 1.666667 3.00 1.2857143
## 3 3 2 13 1 Pasteur 7 4.500000 5.25 1.875 1.000000 1.75 0.4285714
## 4 4 1 13 2 Pasteur 7 5.333333 7.75 3.000 2.666667 4.50 2.4285714
## 5 5 2 12 2 Pasteur 7 4.833333 4.75 0.875 2.666667 4.00 2.5714286
## 6 6 2 14 1 Pasteur 7 5.333333 5.00 2.250 1.000000 3.00 0.8571429
## x7 x8 x9
## 1 3.391304 5.75 6.361111
## 2 3.782609 6.25 7.916667
## 3 3.260870 3.90 4.416667
## 4 3.000000 5.30 4.861111
## 5 3.695652 6.30 5.916667
## 6 4.347826 6.65 7.500000
In the above output, the first six columns give demographic data about the
participants, including age, sex, school year, and school attended. These are not
of particular interest for the analysis we have in mind. The remaining columns
contain the nine dependent measures, which correspond to the tests described in
Table 12.1.
The nine tests probe different aspects of mental ability, from basic perception
through to numerical and linguistic functions. We can summarise the relation-
ships between the variables by generating a covariance matrix:
round(cov(HolzingerSwineford1939[,7:15]),digits=2)

## x1 x2 x3 x4 x5 x6 x7 x8 x9
## x1 1.36 0.41 0.58 0.51 0.44 0.46 0.09 0.26 0.46
## x2 0.41 1.39 0.45 0.21 0.21 0.25 -0.10 0.11 0.24
## x3 0.58 0.45 1.28 0.21 0.11 0.24 0.09 0.21 0.38
12.1. HOW ARE DIFFERENT MENTAL ABILITIES RELATED? 229

## x4 0.51 0.21 0.21 1.36 1.10 0.90 0.22 0.13 0.24

## x5 0.44 0.21 0.11 1.10 1.67 1.02 0.14 0.18 0.30
## x6 0.46 0.25 0.24 0.90 1.02 1.20 0.14 0.17 0.24
## x7 0.09 -0.10 0.09 0.22 0.14 0.14 1.19 0.54 0.37
## x8 0.26 0.11 0.21 0.13 0.18 0.17 0.54 1.03 0.46
## x9 0.46 0.24 0.38 0.24 0.30 0.24 0.37 0.46 1.02
The covariance matrix quantifies the relationships between the dependent vari-
ables. Covariance measures joint variability, or the tendency for two variables to
increase or decrease together. Actually, the covariance matrix is often hard to
interpret, because the values are not standardised and so they depend on the
units of each dependent variable. Instead, it can be more helpful to look at the
correlation matrix, which is a standardised version of the covariance matrix:
round(cor(HolzingerSwineford1939[,7:15]),digits=2)

## x1 x2 x3 x4 x5 x6 x7 x8 x9
## x1 1.00 0.30 0.44 0.37 0.29 0.36 0.07 0.22 0.39
## x2 0.30 1.00 0.34 0.15 0.14 0.19 -0.08 0.09 0.21
## x3 0.44 0.34 1.00 0.16 0.08 0.20 0.07 0.19 0.33
## x4 0.37 0.15 0.16 1.00 0.73 0.70 0.17 0.11 0.21
## x5 0.29 0.14 0.08 0.73 1.00 0.72 0.10 0.14 0.23
## x6 0.36 0.19 0.20 0.70 0.72 1.00 0.12 0.15 0.21
## x7 0.07 -0.08 0.07 0.17 0.10 0.12 1.00 0.49 0.34
## x8 0.22 0.09 0.19 0.11 0.14 0.15 0.49 1.00 0.45
## x9 0.39 0.21 0.33 0.21 0.23 0.21 0.34 0.45 1.00
Figure 12.1 shows the same correlation matrix in a graphical format. The
matrix shows generally positive correlations between different combinations of
variables. The strongest of these (r = 0.73) is between x4 and x5 - the paragraph
comprehension and sentence completion tasks - and there appears to be a cluster
of high correlations involving x4, x5 and x6 in the centre of the matrix. But even
so, just from inspecting the correlation matrix it is rather hard to understand
the structure of the data set.
An alternative approach is to construct a hypothetical model of the potential
relationships. One very simple model is that a single underlying factor determines
performance on all tasks. This general intelligence, or g, factor is widely discussed
in the literature on human cognitive ability (Spearman 1904). It is the classic
example of a latent variable - a construct that we hypothesise might exist, but
we cannot measure directly. This model can be expressed diagrammatically, as
shown in Figure 12.2.
The path diagram shown in Figure 12.2 has several key features. The nine
dependent variables from the Holzinger-Swineford dataset are shown in square
boxes. In the centre is the latent variable g, shown in a circle. These shapes
are the accepted conventions in SEM - squares or rectangles contain measured
variables, and circles or ovals contain latent variables. The arrows joining the
230 CHAPTER 12. STRUCTURAL EQUATION MODELLING

x1

x2

x3

x4

x5

x6

x7

x8

x9
1
x1 1 0.3 0.44 0.37 0.29 0.36 0.07 0.22 0.39
0.8
x2 0.3 1 0.34 0.15 0.14 0.19 −0.08 0.09 0.21
0.6
x3 0.44 0.34 1 0.16 0.08 0.2 0.07 0.19 0.33 0.4

x4 0.37 0.15 0.16 1 0.73 0.7 0.17 0.11 0.21 0.2

x5 0.29 0.14 0.08 0.73 1 0.72 0.1 0.14 0.23 0

x6 0.36 0.19 0.2 0.7 0.72 1 0.12 0.15 0.21 −0.2

−0.4
x7 0.07 −0.08 0.07 0.17 0.1 0.12 1 0.49 0.34
−0.6
x8 0.22 0.09 0.19 0.11 0.14 0.15 0.49 1 0.45
−0.8
x9 0.39 0.21 0.33 0.21 0.23 0.21 0.34 0.45 1
−1

Figure 12.1: Graphical illustration of the correlation coefficients in the Holzinger

& Swineford dataset.
12.2. TESTING HYPOTHESES USING DATA WITH STRUCTURAL EQUATION MODELLING231

x9 x1

x8 x2

x7 x3

x6 x4

x5

Figure 12.2: Example structural equation model with a single latent variable.

variables indicate the relationships between them, such as factor loadings, or

covariances. Notice that the arrow connecting g to x1 is dashed - this indicates
that the model weights are standardised relative to this covariance. Finally, the
double-headed arrows that loop round on each variable represent the residual
error that cannot be explained by the other relationships in the model.

An alternative model might be to propose that there are several latent variables,
which map on to specific abilities that are probed by more than one test. For
example, we might propose a latent variable for the visual tasks (x1-x3), another
for the literacy tasks (x4-x6) and a final one for the timed tasks (x7-x9). We could
allow interdependencies (i.e. correlations) between these three latent variables,
and represent the model with the diagram in Figure 12.3.

12.2 Testing hypotheses using data with struc-

tural equation modelling
We can think of the two model diagrams shown in Figures 12.2 and 12.3 as
explicit hypotheses about how different variables might be related to each other.
Structural equation modelling is a statistical technique that allows us to test how
well such models explain a given data set. This is a very powerful framework
that can uncover (and quantify) the underlying relationships between different
232 CHAPTER 12. STRUCTURAL EQUATION MODELLING

x9 x1

x8 x2

spd vsl

x7 x3

txt

x6 x4

x5

Figure 12.3: Example structural equation model with three latent variables.

measures. We could, for example, see which of the above models gives the best
quantitative description of the data set. There might also be a case for altering
the connections between different nodes in a model to obtain a better fit; that
could change our views on how different variables are related. The following
sections will go through four stages involved in SEM, before discussing some
general issues worth being aware of when conducting this type of analysis.

12.3 SEM stage 1: model specification

The model specification stage is broadly what we have just done. We consider the
possible relationships between the variables we have measured (or are planning to
measure), and one or more hypothesised latent variables. Note that at this stage
we do not say anything about the magnitude of any relationships - these are
estimated later. It can often be helpful to create diagrams like the ones in Figures
12.2 and 12.3 when designing a model, as a guide to thinking about possible
relationships. This is one of the real strengths of structural equation modelling,
as it allows us to make theories and hypotheses explicit by instantiating them in
a model, which we then go on to test empirically.
The models we create are limited only by our imagination and our theoretical
understanding of the topic (though it is important to have a large enough sample
size, and to fulfil some other constraints on model design that we will describe
12.4. SEM STAGE 2: MODEL IDENTIFICATION 233

later). It might seem that there are many possible degrees of freedom when
designing a model like this. However, usually we will be guided by previous
studies, and our intuitions about how different variables might be related. If we
have designed the study that generated the data set being analysed, it is likely
that we included measures because we had some sort of expectation about how
they would be related. If we really have no idea about how to design a model,
there is a technique called Exploratory Factor Analysis that can try to derive
the relationships for us. However this is beyond the scope of this chapter, and is
perhaps less well-suited to hypothesis-driven research.

12.4 SEM stage 2: model identification

Once we have specified our model, we then check that it is suitable for conducting
SEM. This involves a process called model identification, where we check that the
degrees of freedom in the model (known as the number of free parameters) does
not exceed the degrees of freedom in the data set (known as the number of data
points). If there are more free parameters than data points, the model is under
identified. This is a problem, because we cannot calculate a unique numerical
solution for each of the model parameters, and SEM cannot proceed. If there
are more data points than free parameters, the model is over identified: this is
what we are hoping for and can proceed to the next stage. If the two numbers
are equal, the model is referred to as just identified. This means that the model
is describing the data set, but not simplifying it at all, and so may be less able
to generalise beyond the current data set. It is also generally recommended that
there are at least three measured variables for each latent variable (for a detailed
explanation, see Chapter 9 of Kline 2015).
The number of data points can be determined from the number of measured
variables in our data set, according to: N = m(m + 1)/2, where m is the number
of variables. So, for our example data set we have nine variables (x1-x9), and
N = 9 ∗ (9 + 1)/2 = 45. This turns out also to be the number of unique entries
in the covariance matrix (the matrix that calculates the covariance between each
pair of variables, which we inspected earlier).
The number of free parameters in a model is the sum of:
• The total number of latent variables
• The number of error terms on the measured variables
• Any covariances between measured variables
It does not include links between the latent variables and the measures, or the
error terms on the latent variables. Our single latent variable model (Figure
12.2) has one latent variable, and nine error terms on the measured variables,
so it has 10 free parameters. The three-latent variable model (Figure 12.3) has
three latent variables, and nine error terms, so it has 12 free parameters. We
could potentially specify further covariances between data points, which would
add to the number of free parameters. But for now, both of these models have
234 CHAPTER 12. STRUCTURAL EQUATION MODELLING

far more data points than free parameters, so both are safely over identified (as
will typically be the case for data sets with many measures). Notice that model
identification does not depend on the number of cases (i.e. participants) included
in the data set, only on the structure of the data set and the model.

12.5 SEM stage 3: model evaluation

Once the model has been specified and identified, it is fitted to the data set. This
process involves estimating values of the parameters (variances and covariances)
that give the best description of the data set, and is done automatically using
computer software. Formally, we are looking for values that produce a model
covariance matrix that is as close as possible to the empirical covariance matrix
(calculated from the data). This involves conceptually very similar procedures
to those described in Chapter 9 on function optimization techniques.
There are several different methods of parameter optimization. The most common
is called Maximum Likelihood estimation, which involves finding the parameter
values that are most likely (in a probabilistic sense) to have resulted in the
observed data. This approach generally rests on the parametric assumption that
the data are normally distributed. There are other methods, including least
squares fitting, asymptotically distribution free (which has fewer assumptions)
and various other scaled and corrected measures.
Once fitted, the model is traditionally assessed using a chi-square statistic.
Somewhat counterintuitively, a non-significant chi-square statistic indicates a
good fit, because the statistic is comparing model and data, so the null hypothesis
(that they do not differ) indicates a good fit. However, as with other significance
tests, this turns out to be highly dependent on sample size (see Chapter 5),
and with large samples (e.g. N>400) will often be statistically significant even
when the model fit is actually quite good. To address this, several alternative fit
indices have been developed. Some of these (for example, the Bentler-Bonett
index, Comparative fit index, Bollen index, and McDonald index) indicate a
good fit when they have values near 1. Conventions about exactly what values
are considered ‘good’ will differ across disciplines, but 0.9 is often acceptable.
For other fit estimates, such as the Root Mean Square Error or measures of
residual variance, a low value near 0 indicates a good fit. Direct comparisons of
different fit indices are discussed by Cangur and Ercan (2015). When reporting
the outcome of SEM, it is typical to report several fit indices to give a complete
picture of the model’s performance.
A summary of SEM output might look something like this:
## lavaan 0.6-5 ended normally after 35 iterations
##
## Estimator ML
## Optimization method NLMINB
## Number of free parameters 21
12.5. SEM STAGE 3: MODEL EVALUATION 235

##
## Number of observations 301
##
## Model Test User Model:
##
## Test statistic 85.306
## Degrees of freedom 24
## P-value (Chi-square) 0.000
##
## Model Test Baseline Model:
##
## Test statistic 918.852
## Degrees of freedom 36
## P-value 0.000
##
## User Model versus Baseline Model:
##
## Comparative Fit Index (CFI) 0.931
## Tucker-Lewis Index (TLI) 0.896
##
## Loglikelihood and Information Criteria:
##
## Loglikelihood user model (H0) -3737.745
## Loglikelihood unrestricted model (H1) -3695.092
##
## Akaike (AIC) 7517.490
## Bayesian (BIC) 7595.339
## Sample-size adjusted Bayesian (BIC) 7528.739
##
## Root Mean Square Error of Approximation:
##
## RMSEA 0.092
## 90 Percent confidence interval - lower 0.071
## 90 Percent confidence interval - upper 0.114
## P-value RMSEA <= 0.05 0.001
##
## Standardized Root Mean Square Residual:
##
## SRMR 0.065
##
## Parameter Estimates:
##
## Information Expected
## Information saturated (h1) model Structured
## Standard errors Standard
##
236 CHAPTER 12. STRUCTURAL EQUATION MODELLING

## Latent Variables:
## Estimate Std.Err z-value P(>|z|)
## visual =~
## x1 1.000
## x2 0.554 0.100 5.554 0.000
## x3 0.729 0.109 6.685 0.000
## textual =~
## x4 1.000
## x5 1.113 0.065 17.014 0.000
## x6 0.926 0.055 16.703 0.000
## speed =~
## x7 1.000
## x8 1.180 0.165 7.152 0.000
## x9 1.082 0.151 7.155 0.000
##
## Covariances:
## Estimate Std.Err z-value P(>|z|)
## visual ~~
## textual 0.408 0.074 5.552 0.000
## speed 0.262 0.056 4.660 0.000
## textual ~~
## speed 0.173 0.049 3.518 0.000
##
## Variances:
## Estimate Std.Err z-value P(>|z|)
## .x1 0.549 0.114 4.833 0.000
## .x2 1.134 0.102 11.146 0.000
## .x3 0.844 0.091 9.317 0.000
## .x4 0.371 0.048 7.779 0.000
## .x5 0.446 0.058 7.642 0.000
## .x6 0.356 0.043 8.277 0.000
## .x7 0.799 0.081 9.823 0.000
## .x8 0.488 0.074 6.573 0.000
## .x9 0.566 0.071 8.003 0.000
## visual 0.809 0.145 5.564 0.000
## textual 0.979 0.112 8.737 0.000
## speed 0.384 0.086 4.451 0.000

There is quite a lot of information in the above output, so we will go through it

one section at a time. The first line tells us the version of the lavaan software
that we are using (lavaan is an R package we will describe later in the chapter),
and how the fitting proceeded, here with 35 iterations to optimise the model
(see Chapter 9 for more details on model fitting). The next section tells us
that a maximum likelihood (ML) estimator was used, along with the nlminb
optimization method. We also learn that the model had 21 free parameters, and
there were 301 observations (i.e. participants in the data set).
12.6. SEM STAGE 4: MODEL MODIFICIATION 237

Next, the sections headed Model Test User Model and Model Test Baseline Model
give us the results of chi-square tests for the model fit and for a baseline (null)
model in which covariances are all fixed at 0. We should expect the model
we designed to do better than the baseline model, and indeed we see that it
has a smaller chi-square test statistic, indicating a closer fit to the data. For
this example both tests are significant; recall that a significant chi-square test
can indicate a poor fit to the data, but that as discussed above this is hard to
evaluate because of the confounding effect of sample size on significance. The
following section of the output compares the model to the baseline using the
Comparative Fit Index and the Tucker-Lewis Index. Both of these values are
quite high, around 0.9, indicating that the model we designed gives a better fit
than the baseline model.
The three subsequent sections of the output report additional measures of
goodness of fit, including the log likelihood, the Akaike Information Criterion,
the Bayesian Information criterion, and the root mean square (RMS) error.
These values are particularly useful for comparing between different possible
models, as we will describe in more detail later in this chapter.
The final sections of the output show parameter estimates for the latent variables,
covariances and variances. These are somewhat difficult to interpret in table
format, so we can add the parameter estimates to the path diagram to give a
numerical indication of the strength of the links between variables (see Figure
12.4). This can be done using standardised or unstandardised values. In general,
standardised values are more useful, as the values are then similar to correlation
coefficients. The fitted parameters show high loading of individual measures onto
the three latent variables (coefficients between 0.42 and 0.86), and somewhat
smaller correlations between the latent variables (0.28 to 0.47).

12.6 SEM stage 4: model modificiation

The final stage of SEM is to consider possible modifications to the model that
might improve its description of the data. To do this, parameters can be added
or removed (or both). The change in fit when parameters are added is assessed
by the Lagrange Multiplier test (sometimes called the score test). This is based
on evaluating the gradient of the likelihood function (the function that maps
between model parameter values and the probability of the data), and if the test
is significant for a particular parameter we should consider adding it. Be aware
that this can cause problems for theory testing and interpretation if the new
parameter does not make sense. The Wald test does something similar, but to
check if any parameters can be removed from the model without significantly
reducing the fit quality (again, the Wald test is based on the likelihood function,
but evaluates its gradient slightly differently from the Lagrange Multiplier).
These procedures are conceptually similar to step-wise and backward step-wise
entry of predictors in multiple regression, and come with similar pitfalls. Adding
238 CHAPTER 12. STRUCTURAL EQUATION MODELLING

0.56 0.40

x9 x1

0.48 0.67 0.77 0.82

x8 x2

0.72 0.42
spd 0.47 vsl
0.57 1.00 1.00
1.00 0.58
0.28 0.46
x7 x3
0.68 0.66
txt

0.84 0.85
0.86
x6 x4

0.30 x5
0.27

0.27

Figure 12.4: Example structural equation model with three latent variables,
showing standardised parameter estimates.

many new parameters at once is not advisable, as the parameters may be highly
correlated (and therefore not very informative). The order in which parameters
are added and removed can also affect the outcome, so care is advised when
attempting changes to the model.

12.7 Comparing different models

At the start of this chapter we designed two possible models, with different
numbers of latent variables (see Figures 12.2 and 12.3). It is natural to ask
which of these models gives the best description of the data. We can do this by
comparing some of the fit indices between the models. The AIC (Akaike 1974)
and BIC (Schwarz 1978) scores take into account the fit quality and the number
of free parameters, so the model with the smallest score on these statistics gives
the best overall fit, given the number of degrees of freedom. For our example,
the model with a single latent variable has an AIC score of 7738, and the model
with three latent variables has an AIC score of 7517 (the BIC statistic behaves
similarly). This suggests that the model with three latent variables explains
the data better. We can also calculate a chi-square difference test to assess
whether the difference in model fits is significant. For our two example models,
the difference statistic is χ2 = 227, with a p-value of p < 0.001.
12.8. CROSS-VALIDATION ON FRESH DATA SETS 239

12.8 Cross-validation on fresh data sets

One concern in SEM is ensuring that the model generalises beyond the data set
it was constructed to explain. This can be confirmed by cross-validating the
model on fresh data. If the data set is large enough, we might split the data in
two (male and female is sometimes used, assuming these are expected involve
the same relationships between variables). If the data set is not sufficiently large
to allow this, collecting additional data might be required. To cross-validate, one
approach is to fit the model to the two data sets separately, and then compare the
model parameters (coefficient estimates) to check that they are similar. Another
option is to fit the model to one data set, and then assess how well the fitted
model (with all parameters fixed) describes the other data set.

12.9 Power and SEM

SEM is a large sample size technique, and parameter estimates will only be
stable with N>200 participants. A further standard recommendation is to test at
least 5 (and ideally 10) participants per measured variable (while still requiring
at least 200 observations). This is because the parameters we are estimating are
effectively correlation coefficients, and these are very hard to estimate precisely
with small sample sizes. It is possible to conduct power analyses (see Chapter
5) for study designs, often using stochastic simulation (see Satorra and Saris
1985; Wolf et al. 2013). However often these decisions will be limited more by
practical concerns such as the resources and time available. Versions of SEM
that perform better with small sample sizes have also been developed (Bollen
et al. 2007). For a detailed recent treatment of power in SEM see Wang and
Rhemtulla (2021).

12.10 Dealing with missing data

One issue that can dramatically reduce power is when observations are missing
from a data set. If we excluded all participants with at least one missing data
point, for some data sets this would substantially decrease the overall sample
size. To avoid this situation, it is common practise to replace missing data points
with an estimated value. This maintains the sample size and keeps the model
as robust as possible. A simple method to do this is to replace a missing data
point with the mean score for that variable (referred to as unconditional mean
imputation). More sophisticated approaches have also been proposed to impute
(i.e. estimate) the missing values, using methods such as regression or a technique
called expectation maximization that tries to calculate the most likely estimates
of the missing data points (Allison 2003). A popular option is full information
maximum likelihood, which uses all of the available data to estimate the most
likely values of missing data points. It assumes that the dependent variables are
continuous and normally distributed, and performs well when these assumptions
are met (see e.g. Cham et al. 2017).
240 CHAPTER 12. STRUCTURAL EQUATION MODELLING

12.11 Doing SEM in R using the lavaan package

Several packages exist for conducting SEM in R, but we will focus on the lavaan
package (Rosseel 2012), which is an acronym of latent variable analysis. Lavaan
contains tools for specifying models, fitting them to data, and assessing the
fit. It is very well documented, and there is a detailed tutorial available at:
http://lavaan.ugent.be/tutorial/index.html.
To specify a model, we must decide on the relationships between the variables,
and define any latent variables, using a special syntax. To define a latent variable,
we use the =~ operator. For example, defining the visual latent variable from
the Holzinger & Swineford example as being based on the first three dependent
variables looks like this:
visual =~ x1 + x2 + x3
If two variables are correlated, we can define this using the double tilde:
x1 ~~ x2
It is also possible to specify intercepts (~1) and regression (~) if desired. The
model definition is stored in a single text string. So, the three-factor model
syntax for the model shown in Figure 12.3 is defined as follows:
library(lavaan)
HS.model2 <- 'visual =~ x1 + x2 + x3
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9'

HS.model2

## [1] "visual =~ x1 + x2 + x3\n textual =~ x4 + x5 + x6\n speed =~ x7 + x8 + x9"

The function that parses these model definitions is relatively insensitive to
spacing, but note in the output that the carriage returns (i.e. new lines) have
been replaced by backslash-n. Each new definition should appear on a separate
line, and the variable names need to correspond to the variable names in the data
set (i.e. the column names of a data frame). The lavaan syntax is sufficiently
flexible that almost any conceivable structural equation model can potentially be
expressed. It is not necessary to specify relationships between latent variables,
or to include the error terms - these are added automatically.
Note that the default model specification in lavaan is to standardize the factor
loading of the first dependent variable for each latent variable to 1 (known as the
marker method). It is possible to change this behaviour in the model definition,
but this is only advised for advanced users and we will not consider it further
here.
The model can then be fitted to the data using the cfa (confirmatory factor
analysis) function, provided that the data set is stored in a data frame. For our
example data set, this is achieved as follows:
12.11. DOING SEM IN R USING THE LAVAAN PACKAGE 241

fit <- cfa(HS.model2, data = HolzingerSwineford1939)

The fit object stores the model definition, a summary of the fitting process, and
all of the various indices and test statistics. We can request a summary like
the example earlier in the chapter using the generic summary function, and
specifying that we want to see the fit indices as follows (I have suppressed the
output of this command in order to save space, but it is identical to that shown
previously):
summary(fit, fit.measures = TRUE)

From the output, we can extract the various statistics we might want to report.
If we want to compare the fits of two models statistically, we can use the anova
function as follows:
anova(fit,fitG)

## Chi-Squared Difference Test

##
## Df AIC BIC Chisq Chisq diff Df diff Pr(>Chisq)
## fit 24 7517.5 7595.3 85.305
## fitG 27 7738.4 7805.2 312.264 226.96 3 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The output of the anova function includes the AIC and BIC scores, and the
chi-square difference statistic with its accompanying p-value (here fit is the model
with 3 latent variables, and fitG is the model with a single latent variable).
If we want to view the path diagram for the model, we can use an automated
plotting function. We need to pass the fit object into the semPaths function,
which is part of a separate package called semPlot. Again, I have suppressed
the output, which is identical to that shown in Figure 12.4.
library(semPlot)
semPaths(fit,layout="circle",whatLabels="stand",edge.label.cex=1)

There are numerous plotting options, explained in the help file for the semPaths
function. These can be used to change the layout and style of the plot. In these
examples I have used the circle layout, as this shows the latent variables in
the middle of the diagram. Other options include tree and spring - it is worth
checking several of these alternatives to find the most natural and appropriate
way to present a given model. For more general discussion of producing attractive
and informative figures, see Chapter 18.
Model modification can then be conducted. We first calculate modification
indices for the factor loadings, which will tell us the effect of removing one
parameter on the other parameters in the model. The modindices function
calculates this information for all possible operators. Since our model does not
242 CHAPTER 12. STRUCTURAL EQUATION MODELLING

have any covariances between dependent variables, we will only inspect the
links to latent variables (though this does not mean that covariances between
dependent variables do not exist, we are just not considering them here).
mi <- modindices(fit)
mi[mi$op == "=~",1:4] # display only the indices involving latent variables

## lhs op rhs mi
## 25 visual =~ x4 1.211
## 26 visual =~ x5 7.441
## 27 visual =~ x6 2.843
## 28 visual =~ x7 18.631
## 29 visual =~ x8 4.295
## 30 visual =~ x9 36.411
## 31 textual =~ x1 8.903
## 32 textual =~ x2 0.017
## 33 textual =~ x3 9.151
## 34 textual =~ x7 0.098
## 35 textual =~ x8 3.359
## 36 textual =~ x9 4.796
## 37 speed =~ x1 0.014
## 38 speed =~ x2 1.580
## 39 speed =~ x3 0.716
## 40 speed =~ x4 0.003
## 41 speed =~ x5 0.201
## 42 speed =~ x6 0.273

The largest modification index (in the mi column) is 36.4, and corresponds to
the link between the visual latent variable and the speeded discrimination task.
This isn’t part of our original model, but we could consider an updated model
that includes such a link (see Figure 12.5:
HS.model3 <- ' visual =~ x1 + x2 + x3 + x9
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9 '

fit3 <- cfa(HS.model3, data = HolzingerSwineford1939)

semPaths(fit3,layout="circle",whatLabels="stand",edge.label.cex=1)

Note that the new link between visual (vsl) and x9 is now included, and has a
substantial coefficient (0.38). We can assess the improvement in fit statistically
using the Lagrange Multiplier test in the lavTestScore function as follows:
a <- lavTestScore(fit, add = 'visual =~ x9')
a$uni

##
12.11. DOING SEM IN R USING THE LAVAAN PACKAGE 243

0.37 0.42

x8 x1

0.63 0.80 0.76 0.81

x7 x2

0.61 0.44
spd 0.30 vsl
1.00 1.00
1.00 0.59
0.21 0.45
x6 x3
0.30 0.65
0.84
txt 0.45 0.38

0.86
0.85
x5 x9

0.27 x4
0.55

0.28

Figure 12.5: Updated structural equation model with an additional link between
variable x9 and the visual latent variable.

## univariate score tests:

##
## lhs op rhs X2 df p.value
## 1 visual=~x9 == 0 36.411 1 0
Note that we have passed in our original fit (the fit object), and not our updated
fit (the fit3 object). This test gives us a very small p-value, suggesting that the
modification has significantly improved the model fit. We can also compare the
RMSE values of the two models:
fitmeasures(fit,'rmsea')

## rmsea
## 0.092
fitmeasures(fit3,'rmsea')

## rmsea
## 0.065
These statistics show us that the root mean square error value is smallest for
the updated model (fit3 ), indicating a better fit to the data.
A similar approach can be taken for removing parameters using the Wald
244 CHAPTER 12. STRUCTURAL EQUATION MODELLING

test (lavTestWald function). This time, let’s remove the link with the lowest
standardised coefficient - the one between the visual latent variable and x2. We
achieve this by introducing a weight term onto this parameter in the model
definition, and then checking what happens when the weight is set to zero:
HS.model4 <- ' visual =~ x1 + b1*x2 + x3
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9 '

fit4 <- cfa(HS.model4, data = HolzingerSwineford1939)

lavTestWald(fit4, constraints = 'b1 == 0')

## $stat
## [1] 30.84248
##
## $df
## [1] 1
##
## $p.value
## [1] 2.79844e-08
##
## $se
## [1] "standard"
The Wald test also produces a significant p-value, suggesting this change to the
model should be investigated more thoroughly. However, on further inspection,
it actually produces a larger RMS error (and therefore a worse fit) than our
original model:
fitmeasures(fit,'rmsea')

## rmsea
## 0.092
HS.model5 <- ' visual =~ x1 + x3
textual =~ x4 + x5 + x6
speed =~ x7 + x8 + x9 '

fit5 <- cfa(HS.model5, data = HolzingerSwineford1939)

fitmeasures(fit5,'rmsea')

## rmsea
## 0.099
The above examples provide a basic introduction to the capabilities of structural
equation modelling. Of course, as with most of the techniques in this book,
there is much more to learn, and many excellent resources are available to help.
The book Principles and Practice of Structural Equation Modelling by Kline
(2015) is an authoritative but readable text that goes into much more detail than
12.12. PRACTICE QUESTIONS 245

we have had space for in this chapter. Another useful resource is the journal
Structural Equation Modeling, which publishes technical papers on this topic.
It is also worthwhile reading some empirical papers that use the methods, to
see how they are implemented and reported in your area of interest. Outside of
the R ecosystem, there are several commercial software packages designed for
structural equation modelling, including LISREL, Stata, Mplus, and the Amos
extension to IBM’s SPSS.

12.12 Practice questions

1. A latent variable is:
A) Something we measure in an experiment
B) Something we manipulate in an experiment
C) A hypothetical construct that we cannot directly observe
D) A measured variable that mediates the relationship between other
variables
2. When we fit a structural equation model, we are trying to:
A) Reproduce the precise values of each measurement
B) Model the covariances between the variables
C) Find the largest correlation coefficient in the data
D) Maximise the error between model and data
3. In a path diagram, squares and rectangles represent:
A) Latent variables
B) Measured variables
C) Independent variables
D) Error terms
4. In a path diagram, a double-headed arrow with both heads pointing to the
same measure indicates:
A) Covariance
B) Correlation
C) Regression
D) Residual error
5. How many data points are there in a data set with 5 measures?
A) 5
B) 13
C) 15
D) 25.5
6. In the lavaan syntax, the operator ~~ indicates:
A) A correlation between two dependent variables
B) A latent variable definition
C) An intercept
D) Regression
7. Consider a data set with 4 measured variables. We attempt to model this
using a single latent variable, and significant covariance between two pairs
of variables. The model is:
246 CHAPTER 12. STRUCTURAL EQUATION MODELLING

A) Over-identified
B) Under-identified
C) Just identified
D) It is impossible to say without seeing the data
8. Which of the following fit indices indicates a good fit when it has a value
near zero?
A) Bentler-Bonnett
B) Chi-square
C) RMSEA
D) McDonald
9. To assess whether a parameter can be removed from a model, we should
use the:
A) Chi-square test
B) Lagrange Multiplier test
C) Comparative fit index
D) Wald test
10. Structural equation modelling is typically unstable with sample sizes less
than:
A) N=200
B) N=300
C) N=400
D) N=1000
Answers to all questions are provided in section 20.2.
Chapter 13

Multidimensional scaling
and k-means clustering

In this chapter we will discuss two multivariate statistical techniques called

k-means clustering, and multidimensional scaling (MDS). The purpose of k-
means clustering is to partition multivariate data into a number of clusters, such
that observations within a given cluster are more similar to each other than
to observations from other clusters. The purpose of multidimensional scaling
is to reduce complex multivariate data sets to a smaller number of dimensions
(usually two or three) to facilitate graphical representation. These two methods
are included in the same chapter because they can often be used together -
for example MDS can be used to visualise the results of a multidimensional
k-means clustering analysis. Both methods are unsupervised machine learning
techniques, meaning that the algorithms involved try to find the best solution
without knowing any ground truth group or category labels.

To give an example of how these methods might be used together, let’s imagine
that we discover some new varieties of insect in an underground cave. The
insects are all about 10 mm long, but vary in their colouring from grey to blue,
and in the thickness and angle of the characteristic stripes that cover their backs
(see Figure 13.1a). You suspect that there might be three distinct species of
insect, but how might we test this hypothesis? One option might be to measure
all of the key variables from the insects (stripe thickness and angle, colour) and
use k-means clustering to try to determine the underlying structure of the data
set. Because it is challenging to visualise multivariate data with more than
two dimensions (i.e. variables), we could then use multidimensional scaling to
collapse the data into a two dimensional space for plotting. The end result might
look something like the graph shown in Figure 13.1b. There is evidence of three
primary clusters, for which the example insects in Figure 13.1a are prototypical
examples.

247
248CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

(a)

●●
●

(b) ● ● ●
●
● ●
●
●
●

● ●
● ●
●
● ●● ●
● ●
● ●
● ●● ● ●● ●
● ●● ● ●
● ● ● ● ●
●
● ● ●● ● ●
●● ● ●●
● ● ● ●● ●
● ●●
● ●● ●● ●●● ● ● ● ● ●●
● ●● ●● ● ●● ● ●
● ● ●● ● ● ● ● ●
● ●● ●● ●
● ● ●
● ● ● ●●
●●● ●
● ●
●●
●● ●●● ●●
●
●●●● ● ●●
●
● ●●
● ● ●●●● ●● ● ●
●
● ●●● ● ●
●
●
● ● ● ● ●● ●●
●●●● ●●● ●●
●● ●
●● ● ●
● ● ●
●
●●●
● ● ●●●
● ● ●● ●
●
Dimension 2

● ●● ●
●
●● ● ● ● ●● ● ●
● ●● ●●● ●●
● ● ● ● ● ●
● ● ●●
●●●
● ● ●
● ●
●● ● ●
● ● ●
●
●
● ● ●
● ●
● ●
●
● ● ●
● ●
● ● ● ●●● ●
● ●
●
● ● ●
●● ● ●
● ● ● ● ● ●
● ● ●
● ● ● ●● ● ● ● ●
● ● ●
● ●● ● ●●
● ● ●
● ● ●● ●●●● ● ●
●
● ● ● ●●● ●
● ●
●
● ●● ● ●
● ●● ●
● ●
●● ● ●
● ●●●●
●
● ● ● ●
● ● ● ● ●
● ● ● ● ● ●
● ●● ●
● ●
● ● ●● ●
● ● ●●
● ●
● ●● ● ●
● ●
●● ●
●
●

Dimension 1

Figure 13.1: Illustration of ficticious insects (a), and a multidimensional scaling

of their different traits (b) that suggests three distinct clusters. Note that the
dimensions on the x and y axes are produced by the multidimensional scaling
algorithm, and do not necessarily map directly onto the specific traits of the
insects such as colour and stripe thickness/orientation.
13.1. THE K-MEANS CLUSTERING ALGORITHM 249

13.1 The k-means clustering algorithm

To conduct k-means clustering, we must decide in advance how many clusters we
are dividing our data into (i.e. the value of k). This will depend heavily on the
data set we are working with, and what our aims are. The standard algorithm
then defines k centroids (mid-points of a cluster) using k randomly selected
observations from the data set (coloured points in Figure 13.2a). The Euclidean
distance (i.e. shortest straight line) between a data point and each centroid is
calculated, and the data point is allocated to its nearest cluster (Figure 13.2b).
Once all data points have been allocated to a cluster, the mid-point of each
cluster is then re-calculated, using the mean location of all points within the
cluster (Figure 13.2c). The cluster allocation process repeats iteratively to find a
solution with the smallest within-cluster variance, which you can think of as the
summed squared error between data points and centroids. Figure 13.2d shows
the cluster centres on each iteration, as they move from the start points to their
eventual locations.

Figure 13.3a shows some more complex simulated data (see Chapter 8) generated
from five two-dimensional Gaussian distributions. The colours of the points
indicate the true groupings, and you can see that there is some overlap between
the groups in either the x or y directions. Figure 13.3b shows the k-means
solution with k = 5, where each black point indicates a cluster centroid. The
algorithm has identified sensible clusters, though you can see that some data
points have been grouped with other points that come from a different generating
distribution (i.e. true group). The lines are the residuals that are used to calculate
the distance between each data point and its centroid. We can also see what
happens if we choose different values of k. Figure 13.3c shows clustering with k
= 2, and Figure 13.3d shows clustering with k = 10. These do produce plausible
clusterings, though the original (generating) groupings are not preserved.

13.2 Comparing different numbers of clusters

What can we do if we don’t know how many clusters there should be in our
data set? One option is to repeat the clustering for a range of different values of
k, and pick the one that best describes the data. There are various figures of
merit we can use for this. One possibility is to calculate R2 which tells us the
proportion of the total variance explained by cluster membership (in much the
same way as is done with regression or ANOVA). However this type of statistic
does not take the number of degrees of freedom into account, and so the ‘best’
fit will be when N = k, and each data point is its own cluster!

A better alternative is to compute a statistic that penalises model complexity

(here, the number of clusters). The Akaike Information Criterion (AIC; Akaike
1974) and Bayesian Information Criterion (BIC; Schwarz 1978) are two such
statistics. Both of them take an error term such as the residual sums of squares
250CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

(a) (b)
● ●
● ●●
●
●
● ● ●
●
● ●
●
● ●
● ● ●
●
● ●
● ● ● ● ● ●

●
●
●
●
● ●
● ●● ●
●
●
●
● ●
● ●●
● ● ● ● ● ●
● ●● ● ●●
● ● ● ●
● ● ●● ● ● ● ● ● ●● ● ● ●
● ● ● ● ● ●
● ● ● ●
● ●
● ● ● ●
●● ●●
● ● ● ●

● ● ● ●

(c) (d)
● ●
●
●
●
●
●
● ●
● ● ● ●
● ●
● ● ●
●
● ●
●

●
●
●
● ●
●
● ●

●
● ● ●
● ●
●●
●

●
● ● ●● ● ●
● ●
●● ●
●
● ●
● ●
● ●● ● ● ●
●●
●
● ● ●
●
● ●● ● ● ●
●●
● ● ● ● ● ● ● ●
●
●
●
● ●
●
●●
●
●● ● ●● ●
● ● ● ●

● ● ● ●

Figure 13.2: Illustration of the k-means clustering algorithm. In panel (a), the
data points are shown in grey, with the initial centroid estimates in black and
white. Panel (b) shows the initial cluster assignments, and residual vectors
(lines). Panel (c) shows the revised centroid locations and cluster assignments on
the second iteration of the algorithm. Panel (d) shows the path of each centroid
across four iterations of the algorithm, with data points assigned to their final
clusters.
13.2. COMPARING DIFFERENT NUMBERS OF CLUSTERS 251

(a) ●
●
● (b) ●
●
●
● ●
● ● ● ●
● ●
● ● ●●

●
●
●
● ●
●
●
●
●● ●
● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ● ●● ● ● ●
● ●● ● ● ● ●● ● ●
●
●●
● ●●
●
●●● ● ● ●
● ● ● ● ●●● ● ● ●
● ● ● ●
● ● ●● ● ●● ●●● ● ● ●● ● ●● ●●●
●
● ●●
● ● ●●
● ● ●
● ●
● ●● ●
● ● ●●
● ● ●
●
●
●
●
●
●
● ●● ● ●
●● ●
●●
● ● ●
●
●
● ●● ● ●
●● ●
●●
●● ● ●● ● ●● ● ●● ●
● ● ●● ● ● ● ●● ●
● ●●● ● ● ●● ● ●●● ● ● ●●
● ●
● ●● ●● ● ● ● ●● ●● ● ●
● ●
●● ● ● ● ●● ● ● ●
● ●
●
●●●●
● ●
●●●●●
●
●● ●●
● ●
● ●
● ●
● ●

(c) ●
●
● (d) ●
●
●
● ●
● ● ● ●
● ●● ● ●●
●
●
● ● ● ●● ●
● ●
●
●
● ● ●
●
●
●
●
● ●
●
● ●
● ●
● ● ●● ● ● ●●
● ●● ● ● ●● ●
● ●
●
● ●
●
●
●
●
● ● ● ● ● ●
●
● ●
●
●
●
● ● ● ●
●
●●
● ● ●● ● ●● ●●● ● ●
●●
● ● ●● ●
● ●● ●●● ● ●
● ●● ●
● ● ●●
● ● ●
●● ● ●
● ●● ●
● ● ●●
● ● ●
●● ●
● ●
● ●● ● ● ● ●
● ●● ● ●
● ●● ● ●●
● ●● ● ● ●
●●
● ● ●● ● ●● ●
● ●● ● ● ●
●●
●
● ●
●
●● ● ●●
●
●
● ●●● ● ●●
● ● ●
●
● ●● ●● ● ● ● ●● ●● ● ●
●
● ● ● ● ●
● ● ● ●
● ●
●
●●●●
● ●
●●●●
●●
●● ●●
● ●
●
●
●
●
●● ●

Figure 13.3: Example k-means clustering on simulated data. Panel (a) shows
data generated from five two-dimensional normal distributions with different
means. Panel (b) shows a k-means solution with k = 5, where black points
indicate the centroids, and lines show the residuals for each point. Panels (c)
and (d) are for k = 2 and k = 10 respectively.
252CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

(i.e. the sum of the squared lengths of the residual lines in Figure 13.3b-d)1 , and
add a penalty term. For the AIC, the penalty is 2mk, where m is the number
of dimensions (i.e. dependent variables) and k is the number of clusters. For
the BIC, the penalty is 0.5log(N )mk, where N is the number of data points
(observations).

For both the AIC and BIC statistics, the best model is the one that produces the
lowest score. Generally both statistics behave similarly, meaning that whichever
one you use is likely to produce the same outcome, so the choice will not matter
for most applications. For the simulated example here, both statistics actually tell
us (see Figure 13.4) that k = 4 clusters gives the most parsimonious description
of the data (despite us actually using 5 generating distributions).

50
BIC
AIC
AIC/BIC score

40

30

20
1 2 3 4 5 6 7 8 9 10
k

Figure 13.4: Figures of merit as a function of the number of clusters (k). Shaded
regions indicate 95% confidence intervals for 1000 independent data sets generated
from the same underlying distributions.

1 There are variants of AIC and BIC for several different error terms, including the residual

sums of squares and the log-likelihood. The key point though is that a penalty is added that is
dependent on the number of free parameters in the model, which here are the data dimensions
and the number of clusters.
13.3. EXAMPLE: K-MEANS CLUSTERING OF DINOSAUR SPECIES 253

13.3 Example: k-means clustering of dinosaur

species
To produce an example data set, I compiled a table of dinosaur statistics from
the excellent website ZoomDinosaurs.com (my inner 6-year-old had a great
time). The data consist of height, length and weight measurements for 61
dinosaur species, plus whether they were a carnivore or a herbivore (omnivores
were classified as carnivores). My very vague childhood dinosaur knowledge
characterised herbivores as being generally long and heavy, whereas carnivores
were taller and lighter. The data bear this out to some extent, with more blue
circles (herbivores) in Figure 13.5 appearing in the top left (heavy and long),
and black squares (carnivores) being more prevalent in the lower right (tall and
light).

3 3 3
● Herbivore
(a) Carnivore (b) (c)
2 ● 2 ● 2 ●
● ● ●
● ● ●
● ● ●
● ● ● ● ● ●
1 ● 1 ● 1 ●
Weight (log)

Weight (log)

● ● ●
● ● ●
Weight (log) ● ● ●● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ● ● ● ●
● ●
●
●
● ●
●
● ●
●

● ● ● ● ● ●
0 0 0
● ●
●
●

−1 ●
● −1 ●
● −1 ●
●

● ● ●
● ● ●

−2 ●
−2 ●
●
−2 ●

−3 −3 −3
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 −1 −0.8 −0.6 −0.4 −0.2 0 0.2
Height/Length (log) Height/Length (log) Height/Length (log)

Figure 13.5: k-means clustering of dinosaur species. Blue circles represent

herbivores, and black squares represent carnivores and omnivores. Panel (a)
shows the raw data, with silhouettes indicating some well-known species, and
panels (b) and (c) show clusterings for k = 2 and k = 3.

Since we have two types of dinosaur, the first thing we can try is setting k =
2 (see Figure 13.5b). This doesn’t do an amazing job, as there are quite a lot
of mis-classifications. In particular, there are lots of carnivores included in the
upper cluster, which should be mostly herbivores. An alternative might be k=3
(see Figure 13.5c), where we could define an intermediate cluster. Given the way
the data appear, this looks like we now have a ‘heavier carnivore’ and a ‘lighter
carnivore’ category, as well as a ‘herbivore’ category. Of course, there are still
some errors, but real data are unlikely to cluster perfectly.
If we had the length, height and weight of a newly discovered species, or
one that doesn’t appear in our original data set, we might use the cluster
arrangement to hazard a decent guess about whether they were a carnivore or
a herbivore. For example, my four year old daughter (who knows much more
about dinosaurs than I do) really likes the protoceratops, which doesn’t feature
in the data set. Apparently these weighed 85 kg and were about 0.6 m tall
and 1.8 m long. That places them firmly in the top left corner of the plot (at
254CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

x = log10 (0.6/1.8) = −0.48, y = log10 (85) = 1.93), closest to the herbivore

cluster (they were indeed herbivores).

13.4 Variants of k-means clustering

There are several variations on the k-means clustering algorithm. If our data
are skewed or have lots of outliers, we don’t necessarily need to use the mean as
the measure of central tendency. Instead we could use the median (k-medians
clustering), which is more robust to outliers and non-normal distributions.
Alternatively, for ordinal data the mode (k-modes clustering) might be a more
appropriate method. This is because the mode is the most common value in a
data set, and estimating this does not require metric data, or any assumptions
about distributions.

Another variant, called k-medioids clustering, has the constraint that the centre
of each cluster must be one of the data points, whereas in k-means clustering this
is only the case for the initial guess. This is also more robust to outliers than
the k-means algorithm because the medioid is a plausible (i.e. already observed)
data point. Finally, the spherical k-means clustering method tries to constrain
both the distance and the angle of each point relative to the cluster centroid, so
that points are evenly spaced radially. All of these variants work in a broadly
similar way, and may be more or less well-suited to a particular situation or data
type.

There are also several different algorithms for estimating the clusters. In the
standard method described at the start of the chapter, the centres of the clusters
begin as random samples from the data set, and are iteratively recalculated using
the mean of the points allocated to each cluster. This is sometimes referred to
as Lloyd’s algorithm or the Forgy method (after Lloyd (1982) and Forgy (1965)).
One modification to this algorithm, called the random partition method, is to
assign each data point to a random cluster at the start, instead of choosing k data
points to form the initial cluster centres. An alternative algorithm proposed by
Hartigan and Wong (1979) uses a function minimisation approach (see Chapter
9) to determine cluster membership.

Finally, k-means clustering is not restricted to working with two dependent

variables (i.e. in two dimensions). However, data sets with a large number of
variables are very hard to visualise. To help with this, we will next discuss a
technique called multidimensional scaling, that allows us to collapse multivariate
data into a two-dimensional representation (as we saw in our insect example).
We can combine this with k-means clustering to allow us to visualise clusters
even in complex multidimensional data sets.
13.5. THE MULTIDIMENSIONAL SCALING ALGORITHM 255

13.5 The multidimensional scaling algorithm

Multidimensional scaling is a technique that is used to reduce the dimensionality
of multivariate data sets to make them easier to visualise. It works by taking
a matrix of pairwise dissimilarities (or distances) between data points and
mapping it into an abstract space with a defined number of dimensions (usually
2). Figure 13.6a illustrates five points in a two-dimensional space that are joined
by the shortest possible straight lines between each pair of points. Figure 13.6b
summarises the distances between each pair of points, with darker greys indicating
longer distances. This representation is known as a dissimilarity matrix, because
it summarises how far apart the points are in the multidimensional space. If two
points are close together, like points 1 and 2, they are quite similar according
to the data we have. If two points are far apart, like points 2 and 4, they are
quite dissimilar. Note that the dissimilarity matrix is always calculated between
pairs of points, regardless of the dimensionality of the data set. This means that
the dissimilarity matrix will always have two dimensions, even if the underlying
data set has many more dimensions.

(a) (b)
● 5 1 1

Distance
●
4 2
● 3

3 0

● 1 4

● 2
5

1 2 3 4 5

Figure 13.6: Illustration of pairwise Euclidean distances between a set of five

points in two dimensions. The lengths of the lines in panel (a) show the Euclidean
distances, which are expressed as grey level intensity in panel (b).

The aim of multidimensional scaling is to represent the dissimilarities in a new

space with fewer dimensions than the original data set. The idea is that the
distances between the points in the lower-dimensional space should correspond
as closely as possible to the distances in the original multidimensional space.
In other words, the dissimilarity matrices in the original and lower-dimensional
spaces should be as similar as possible to each other. The eventual solution will
usually be a list of two-dimensional (x,y) coordinates mapping the points in the
256CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

original data set into the new space. This is achieved by minimising a statistic
called the strain (or in some variants the stress). The strain is a loss function
based on the Euclidean distances between points.
A good way to illustrate the results of multidimensional scaling is to use random
colour vectors. Colour is defined using mixtures of the red, green and blue pixels
on a display. We can therefore create random RGB vectors, and use MDS to
reduce from 3 to 2 dimensions for plotting. This is shown in Figure 13.7a, where
colours of a similar hue end up being grouped together. A variant in Figure
13.7b includes a fourth dimension, the alpha (transparency) setting. In this
plot the different hues still group together, but the transparency information is
clearly being factored in too, for example by placing more transparent points
nearer the lower right edge of the cloud.

RGB RGBα

(a) (b) ●

● ●●
● ●
● ● ● ●
●● ●● ●
● ●
● ● ●● ●
● ● ● ●● ● ● ● ●● ●
●● ● ● ●●● ● ● ●● ● ●
●● ● ●● ● ●● ●
●● ●● ● ●● ● ●
● ● ●● ●● ● ● ●● ● ●
●
● ● ●
● ●● ● ● ●●● ● ●● ● ● ● ● ●●● ●● ● ●
●
● ●●●●
Dimension 2

● ● ● ●●
● ●
● ●
●●● ●
● ●
●●
●●
●●●
●●●●●
●
●●
●●● ● ●●● ●
Dimension 2 ●
●
●
●
●
● ●
● ● ● ●●
● ●●●● ● ● ●● ● ● ●
● ●
● ● ● ●● ● ● ● ●● ● ●●● ●●
●
●● ● ● ●●● ● ● ● ● ●● ●● ●
● ● ● ● ● ●● ●
● ● ● ●●● ● ● ●● ● ●
● ●● ● ● ●●
●
● ●●●● ● ●
● ● ●● ●● ●●●
●●● ● ●
● ●● ●● ●●
● ●● ●● ● ●● ● ●●● ●●●
● ●● ● ●●● ●● ● ●● ●
●● ●● ●●● ● ● ●
● ●● ● ● ●●● ● ● ●●● ● ● ● ● ●● ● ●
● ● ● ●● ●● ● ●●● ●● ●●
● ●
● ● ●
●●● ● ● ● ●● ● ● ● ●
●●
● ●● ● ● ● ●●
●●●● ●●●● ● ● ●●● ● ● ●● ●●
● ●● ● ● ● ●● ● ●●●● ●
● ● ● ● ● ● ●●● ●●●● ●● ● ● ●●●●●● ●● ● ● ●●
● ●
● ● ● ● ●
● ●●●●● ● ●●● ●● ●●● ●● ●
●● ●●●● ● ● ● ●●●● ●●● ●●
● ●●
●● ●● ● ● ●●● ●● ● ●● ●●●● ● ●●●● ●●
●
●●● ● ●● ●
● ●● ●
● ● ●● ● ● ● ●● ● ● ●● ● ●● ● ● ●● ● ●
●● ●● ● ● ● ● ● ●● ●
●● ● ●●
● ●●●● ● ● ● ● ● ●●
●
●●● ● ●●
● ●● ● ●● ● ●● ● ●●● ● ●●● ● ●● ●
●
● ●● ● ● ●
●● ●●
● ●
●●●● ● ●
● ●
● ●●●●●● ● ● ● ● ●● ●●● ●● ●
● ● ●● ● ●●
●● ●●●● ● ●
● ●●● ●●●● ●● ●
●●●●
● ●● ●
●●
● ●● ●● ● ● ●
● ●
● ● ●●
● ● ●● ● ●● ●●●● ●● ●●● ●●● ●● ● ●
● ● ●
● ● ●
●● ●
●● ● ●●● ●●●● ●●●● ●●● ● ● ●● ● ● ● ● ● ●●●●●● ●●●●● ● ● ●● ●● ● ● ● ●● ●
●● ● ●●● ●●● ● ●●●● ● ● ●
●●
● ● ● ● ●
● ●● ● ●● ● ● ● ●● ●
● ●●●●● ● ● ●
● ● ● ● ● ●● ● ●● ● ●●●
● ●● ● ●● ●● ●●● ● ● ● ●●●
● ●● ●● ● ● ●
● ●
● ●● ● ●● ● ●
●●
● ●●●● ● ● ●
●
● ● ● ●
●● ● ●
● ●● ● ●● ●● ●●● ● ● ●
● ●
● ●●● ●● ● ● ● ● ●●●●●
● ●●● ● ●●
●●●●● ● ●●● ●●
●● ● ●● ● ● ●
● ●● ●●●● ● ●●●●● ● ● ● ● ●● ●● ●● ●● ● ● ● ●
● ●● ● ●●●● ● ●● ●● ● ● ●● ●● ● ● ●● ●
● ● ●
●●
● ●● ● ● ● ●● ● ● ●● ● ● ●● ● ●●
● ●
●● ●● ● ● ● ●●● ●
●● ● ●●
● ●● ● ●● ●
● ● ●●
● ●●●
●●
●●
●
●
●●
● ● ● ● ●● ● ●
● ● ● ● ●● ● ●
●● ●● ● ●●● ● ● ●● ● ● ● ● ●●● ● ●
● ●
●●●● ●● ● ● ●● ● ● ●●●● ● ●●● ● ● ●●●
●●● ● ●● ● ● ● ●● ● ●● ●
●
● ● ● ●●●
●● ●●●●●● ●●● ● ● ●
●● ●
● ●●● ●
●●●● ●● ● ●●●● ●● ● ●●● ●● ● ●● ●● ● ●● ●●● ● ● ● ●●●● ● ● ●●
●
●● ● ● ●
● ● ● ●● ● ●● ● ●●● ● ●●● ● ● ● ●
● ● ●● ● ● ●
● ● ●
●● ● ● ● ● ●● ● ● ● ●● ●●
●● ●● ● ●●● ● ● ●●●● ● ●●● ●
●● ●●● ●●
●● ● ● ● ● ● ●● ● ●● ● ● ● ● ●●
●●●●●● ●●●●● ● ●●● ●●● ●● ●●●
● ● ●● ●● ● ● ●●● ●● ● ●● ● ●● ●●● ● ● ●● ● ●●●● ●● ●●●
● ● ● ● ● ● ● ● ● ●● ● ●
●● ● ● ● ● ● ●●● ●● ● ●●
● ●●● ●●● ●
● ● ●
●●●●
● ●● ●● ● ● ●● ●
●● ● ●
● ● ●● ● ● ●● ●●● ● ●● ●● ● ● ●● ● ● ●●● ● ● ● ● ● ●
●● ● ●●● ●●●● ●
● ●● ● ● ●● ●
● ● ●● ● ● ● ● ●● ● ● ●●●● ●
● ● ● ● ● ●
●
● ●● ● ● ●● ●● ●●● ● ● ● ●● ● ●
● ● ● ●● ● ● ● ● ●●● ● ● ●
● ● ● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ● ●● ●● ● ● ●●●● ●
● ●●●●● ● ●● ●● ●● ● ●● ● ● ●
●●● ● ●● ● ● ●● ●
● ●
● ●
● ● ● ●●
● ●
● ●●● ●● ●
● ● ●●● ●● ● ● ● ● ● ●
●● ●●●
● ● ●●
● ●●●● ● ●●●
●
●
●
● ● ●● ● ●● ●● ● ●● ● ●
● ● ●● ●● ●●
●●
● ● ●● ●
●
●● ●●
● ● ●
●●● ●●● ●●● ●
● ●●●● ● ● ● ●●●
● ● ●●● ●● ●●● ● ● ● ● ●● ●● ●● ●
●
● ● ●● ● ● ●● ● ● ● ● ● ●
● ● ● ● ● ● ● ●
● ●●● ● ● ● ● ● ●●● ● ●● ●● ● ● ●● ● ●
●
● ● ● ●● ●● ●● ● ● ● ●
●● ● ● ●
● ● ●
● ●● ●
● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ●
●
● ●
● ● ●●
● ●●
●● ●
●● ●●● ●●
● ● ● ● ● ● ●
● ● ●● ● ● ● ● ● ●● ●
● ● ● ●● ●● ●● ● ●● ● ● ●● ● ● ● ●● ●
● ● ● ● ●●
● ● ● ●●● ●●● ● ●
● ●● ● ● ● ● ● ●● ●
● ● ● ●● ●●●
●
●
●
●
●
● ● ● ●
●
●● ●
●
●●
●
●
●
● ●

Dimension 1 Dimension 1

Figure 13.7: Example multidimensional scaling with random colour vectors.

Panel (a) shows a two-dimensional solution for three-dimensional RGB vectors,
and panel (b) is the solution for four-dimensional RGB-alpha vectors. The x-
and y-positions of each point are determined by multidimensional scaling, and
the colour is the original colour for each data point. In both cases, similar colours
group together.

The starting data for MDS will be an N × m matrix, where m is the number of
dependent variables (dimensions). For example:
## [,1] [,2] [,3] [,4]
## [1,] 0.692666520 0.98132464 0.08029358 0.6964390816
## [2,] 0.802897572 0.13823851 0.93906565 0.6016717958
## [3,] 0.797127023 0.88163599 0.66954995 0.6361913709
## [4,] 0.007445487 0.06651651 0.37043385 0.1724689843
## [5,] 0.621347463 0.68464959 0.11980545 0.0002071382
13.5. THE MULTIDIMENSIONAL SCALING ALGORITHM 257

## [6,] 0.318114384 0.57204209 0.33323977 0.8083065613

The output will be an N × 2 matrix, where the two dimensions are x and y
coordinates:

## [,1] [,2]
## [1,] -0.15160650 0.6517739
## [2,] 0.41646691 -0.2108153
## [3,] -0.03000302 0.3792663
## [4,] -0.05064061 -0.6095281
## [5,] 0.09436012 0.1585279
## [6,] -0.22697391 0.1237203

We can check the mapping between the original dissimilarities and the dissimi-
larities between positions in the lower dimensional space created by the MDS
algorithm (the rescaled data) using a Shepard diagram. This plots the pairwise
distances between points from the original data along the x-axis, and the pairwise
distances for the rescaled data along the y-axis. If there is no loss of information
due to the rescaling, these values should be perfectly correlated. The amount
of scatter around the diagonal is therefore an indication of how faithfully the
data have been mapped by the MDS algorithm. One can also calculate statistics,
such as Spearman’s rank correlation, between the distances in the two spaces.
Examples for the colour data are shown in Figure 13.8.

RGB RGBα

(a) (b)
Rescaled distance

Rescaled distance

Original distance Original distance

Figure 13.8: Shepard plots for rescaling the three- and four-dimensional colour
vectors. Each point represents a pairwise distance between two points. Both
panels show strong ordinal relationships between the original and rescaled values,
with the largest discrepancies in the upper right corner of each plot, representing
pairs of points that were very far apart in both spaces
258CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

13.6 Metric vs non-metric MDS

There are two main variants of multidimensional scaling, known as metric and
non-metric MDS. The metric algorithm, also referred to as principal coordinates
analysis, tries to maintain the relative distances between data points in the
rescaled space. Non-metric algorithms relax this assumption, and try to maintain
only the rank orderings of distances between points. In essence, they replace
the actual values in the distance matrix with ranks, and so work in a similar
way to other non-parametric statistical tests (e.g. the Wilcoxon signed-rank
test). This also means that non-metric algorithms can be used with ordinal
variables (variables involving categories with a clear ordering), and data that do
not meet parametric assumptions (i.e. that are not normally distributed). For
many parametric data sets, both methods will produce a similar solution and
the choice of which to use is somewhat arbitrary.

13.7 Example: multidimensional scaling of

viruses
A study by Lopes, Andrade, and Tenreiro Machado (2016) used multidimensional
scaling to represent the similarity of 22 viruses. They used quantitative data
on several measures including fatality rate, speed of spread, and incubation
period, and produced two- and three-dimensional MDS plots of virus similarity.
Because their data are provided (in Table 1 of the paper), we can attempt a
simpler version of their analysis ourselves. Figure 13.9 shows a two-dimensional
scaling of the virus data. You can see that various types of flu (SFlu = swine flu,
BFlu = bird flu) and respiratory disease (SARS, MERS) are grouped together
near the top of the plot, and some high fatality rate diseases (HIV, Ebola
and untreated rabies) are clustered towards the lower right. A Shepard plot
is shown in Figure 13.10 and indicates a good mapping between the original
and rescaled spaces. The original paper reports a more involved analysis with
multiple distance measures, and is a good example of how the technique can be
used.

13.8 Combining k-means clustering and MDS

The two techniques described in this chapter can be used together, much as
we saw in the example with insect species. For example, a multivariate data
set can be clustered using k-means clustering, and the results visualised with
multidimensional scaling. This is usually the correct order to apply the techniques,
because applying MDS before clustering will mean that some information has
been lost in the rescaling, and the clustering will be less accurate.
An example comes from a study by Coggan et al. (2019), who wanted to
create sets of stimuli with similar image properties to study how the human
brain represents objects. To do this, they took a database of 2761 object
13.8. COMBINING K-MEANS CLUSTERING AND MDS 259

Dimension 2

● SFlu ● BFlu
● Nor ● MERS
● Rhi
● ●Rot
● Mar
SARS
● Ebo ● Rab Un
● Den
● CPox ● Rab
Mea ● ●
RubSma
Mum
● ●●Chi
● Pol ● HIV Un
● HIV

● HepB

Dimension 1

Figure 13.9: Example data on virus characteristics, visualised with metric

multidimensional scaling.

images (the Bank of Standardised Stimuli, see Brodeur et al. (2010)). To

simplify the representation of these images, they converted them to greyscale,
and then applied the GIST descriptor function (Oliva and Torralba 2001), which
summarises images using the orientations and sizes of features in each image
(see also Chapter 10 for details of how such image features can be represented).
Each image was thereby reduced to a vector of 4096 numbers. However this is
too many dimensions for k-means clustering to operate on reliably given the
number of examples in the data set. So principal components analysis (a form
of factor analysis) was applied the GIST vectors to further reduce them to 20
values per image (the first 20 principal components).

The k-means clustering algorithm, with k = 10 was then applied to the matrix
of 2761 × 20 numbers. The choice of k = 10 clusters was intended to produce
a suitable number of stimulus categories for use in a neuroimaging experiment.
The 24 image examples closest to each of the cluster centroids were chosen for
use in the experiment. The distinctness of each cluster was confirmed using
multidimensional scaling to reduce the dimensionality of the image dataset from
20 dimensions to 2. It was also clear that images from individual clusters had
various properties in common - for example all being roughly circular, or oriented
in a particular direction. A summary of the image selection process is provided
in Figure 13.11 (based on Figure 2 of Coggan et al. (2019)).
260CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

●
Rescaled distance

●
●

●
●
●
●●
●●●●●
●● ●
●●
●●
●●
●●
●●●
●
●●●
●
●●
●●●●
●
●

●●●
●

●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●●
●
●
●●
●●
●●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●●●
●
●●
●
●
●●

●
●●
●●
●
●
●●
●●
●●
●
●●●
●
●●
●●
● ●

Original distance

Figure 13.10: Shepard plot for the virus data set.

13.8. COMBINING K-MEANS CLUSTERING AND MDS 261

(a) Gist PC 1−20 Clusters (b)

●● ●●●
●● ●
● ● ●● ●●
● ●
●●
● ● ●●● ● ● ●
●●
●●
● ●●●
●●
●●●
●●
●●●
●●●
● ●
●●
● ● ●●
●●●
●
●●
●● ●
●●
●
●● ● ●
●●●
●
●●●●
● ●●
2761

2761

●
●● ●●●
●●
● ●● ● ●●
● ● ● ●●●
24

●
●
●●
●● ●●
●
● ● ●
● ● ●
●● ●● ●●
● ● ●●
● ● ● ● ● ●
● ●
● ● ●●●
●●● ●●● ●● ●
●
●●
● ● ●● ●● ●●
●
● ● ●
●●
●● ● ●
●● ● ● ● ● ● ●●
●●
● ●●
● ●
●●
●●●●● ● ●
● ●●
●●●
●
●● ●●● ●
●
●
●●●
● ● ●●●
●●
●●

4096 20 20

Figure 13.11: Summary of combined k-means clustering and multidimensional

scaling, following Coggan et al. (2019). Panel (a) shows a graphical representation
(shown in red/green) of the gist vector for 2761 images, represented as 4096 values
based on their low-level image properties. This was reduced using principal
components analysis to the middle matrix (blue/yellow). k-means clustering was
used to generate ten clusters of 24 images. Five example clusters are illustrated
to the right of the panel, where the similarity of components within a cluster
is clear. Panel (b) shows a two-dimensional multidimensional scaling of the
principal components, where the clusters are illustrated by coloured points. Image
examples from each cluster are shown around the circumference, and within a
cluster have consistent image properties, such as similar shape, orientation, and
aspect ratio.
262CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

The final set of 240 images (10 categories × 24 examples) were then presented
to participants in a block design fMRI experiment. The study found that neural
responses in the ventral visual cortex (a region of the brain believed to be
specialised for detecting objects) produced distinct patterns of activity for each
cluster. This is important, because distinct patterns are usually associated with
specific categories of real-world objects (such as faces, buildings etc.), and this
in turn is interpreted as evidence that there are areas of the brain specialised
for different semantic object categories, such as faces, bodies or buildings. By
using object clusters defined entirely by their image properties (and not their
semantic properties), this study demonstrates that low level image features (such
as orientation, curvature and so on) are also important in understanding stimulus
representations in this part of the brain.

13.9 Normalizing multivariate data

Both of the techniques we have discussed in this chapter work best when the
dependent variables have consistent units and similar variances. This is because
if one variable has much larger units than the others, or a much larger variance,
it will contribute much more to estimates of the Euclidean distance between data
points and cluster centres (for k-means clustering) or between pairs of points (for
MDS) than will the other variables. In such cases, it is good practice to rescale
all of the variables, so that they have similar variances (see Section 3.7). This is
achieved by dividing each variable by its standard deviation. It is also possible
to subtract the mean from each variable, so that the values are converted to
z-scores. Normalization will generally speed up convergence of the algorithm, as
well as potentially improving the results.

13.10 Examples from the literature of k-means

clustering and MDS
As we have seen in several examples throughout the chapter, both k-means
clustering and MDS are very useful when classifying different species of organisms.
A good example from the literature is a recent study by Scott Chialvo et al.
(2018). These authors investigated varieties of tiger moth that feed on lichens,
to understand defensive chemical pathways that rely on substances ingested
from the lichens. They studied the metabolome (i.e. the chemicals inside the
organism) of different species of adult tiger moth, and measured the quantities
of chemicals called phenolics that were derived from lichen. They used k-means
clustering to identify 8 clusters of chemical profiles, and visualised the results
using nonmetric multidimensional scaling. As part of a wider analysis that also
included genetic measures, this helped them to more accurately reconstruct the
evolutionary history of lichen moths.
Multivariate methods are also useful in the commercial sector, in particular for
13.11. DOING K-MEANS CLUSTERING IN R 263

identifying groups of customers that have similar characteristics or preferences,

and in understanding how those preferences relate to different products. In the
food industry, preference mapping studies involve asking a sample of consumers
about their likes and dislikes, usually in the form of rating scale data for particular
products or product categories. Wajrock et al. (2008) evaluated several different
clustering methods, including k-means and hierarchical clustering algorithms,
for 15 preference mapping data sets. They conclude that algorithms such as
k-means clustering, that involve partitioning individuals into different categories,
outperform other approaches. In a different study investigating online shopping
behaviour, Jain and Ahuja (2014) used k-means clustering to identify four
distinct types of online shopper. These were referred to as ‘cognizant techno
strivers’, ‘conversant appraisers’, ‘moderate digital ambivalents’ and ‘techno
savvy impulsive consumers’. The rationale behind such classification is to better
target advertising at consumers more likely to be receptive to specific product
types. Multidimensional scaling is also useful in marketing and brand analysis,
for example Bijmolt, Wedel, and DeSarbo (2020) used it to model consumer
perception of different car brands, based on ratings of their perceived similarity.
In a highly competitive industry, this type of information is useful for designing
new products, as well as for marketing existing ones.

13.11 Doing k-means clustering in R

The kmeans function is built into the core stats package, and can conduct k-means
clustering with several different algorithms. It takes an N (rows, containing
observations) by m (columns, containing different variables) matrix as its input,
as well as a number to specify the value of k (the desired number of clusters).
For the example dataset from Figure 13.3, we could request a clustering with k
= 5 and inspect the output as follows:
clusters <- kmeans(dataset[,1:2],5)
clusters

## K-means clustering with 5 clusters of sizes 31, 33, 13, 27, 21

##
## Cluster means:
## X1 X2
## 1 -0.39729954 -0.03181516
## 2 0.39216792 0.06878313
## 3 0.01485053 -0.66225151
## 4 -0.03320486 0.51792763
## 5 0.18295912 -0.30327400
##
## Clustering vector:
## [1] 5 2 5 1 2 4 2 5 2 2 2 2 2 2 5 2 1 5 4 1 4 1 2 1 2 4 4 4 4 4 4 4 4 4 4 4 4
## [38] 4 4 4 4 4 1 4 4 4 2 4 4 4 2 5 5 5 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 5 5 2 2
## [75] 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 4 1 1 1 1 5 1 1 1 5 5 5 5 1 3 5 3 3 5
264CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

## [112] 5 3 5 3 5 3 3 3 5 3 3 3 3 3
##
## Within cluster sum of squares by cluster:
## [1] 2.0466979 2.2336343 0.3633141 2.1496942 1.4766339
## (between_SS / total_SS = 75.6 %)
##
## Available components:
##
## [1] "cluster" "centers" "totss" "withinss" "tot.withinss"
## [6] "betweenss" "size" "iter" "ifault"

The first line of the output tells us how many clusters we have generated, and
their sizes (i.e. how many observations are assigned to each cluster). Then,
it gives the cluster means as the x and y coordinates of the cluster centres.
Note that for data sets with more than two dependent variables, the means
will contain a value for each dependent variable. The clustering vector gives
cluster assignments to each of the individual observations from the data set. The
summed squared error for each cluster is also provided, and gives an estimate
of the residual variance within each cluster. Finally, the ratio of between and
total sums of squares is given - this is the same as the R2 value from ANOVA or
regression, and tells us the proportion of the total variance that is explained by
cluster assignment.

The output data object allows us to access all of these values, as well as incidental
information about things like the number of iterations required for the clustering
algorithm to converge. We can use this information to plot the lines between
each data point and its assigned cluster centroid as follows (see Figure 13.12 for
the output):
# set up an empty plot axis
plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=c(-1,1), ylim=c(-1,1))
axis(1, at=c(-1,1), tck=0.01, lab=F, lwd=2)
axis(2, at=c(-1,1), tck=0.01, lab=F, lwd=2)

# draw lines between each cluster centre and the assigned data point
for (n in 1:(nrow(dataset))){
lines(c(clusters$centers[clusters$cluster[n],1],dataset[n,1]),
c(clusters$centers[clusters$cluster[n],2],dataset[n,2]),
col='grey')}

# draw individual data points

points(dataset[,1],dataset[,2],pch=16,col=pal2tone[1])

# draw the cluster centroids

points(clusters$centers[,1],clusters$centers[,2],pch=16,cex=2)

Additional colours for each cluster, or for true group membership if this is known,
13.11. DOING K-MEANS CLUSTERING IN R 265

●
●
● ●
● ●
● ●
● ●
● ●
● ● ● ●
●
● ● ●
● ●
● ● ●●
● ●● ● ●
● ●●●
●
●● ● ● ●● ● ●
● ● ●● ● ●●
●
● ●●
● ● ●●
●● ●● ●
● ● ●
● ● ●
●
●
● ● ● ● ● ● ●
● ● ●
● ●● ● ●● ●● ● ●
●
●
● ●● ● ● ●
● ●● ●● ●● ● ●
●
●
●
●

● ● ●
●
● ●
●●●●
● ● ●
● ●
● ●

Figure 13.12: Output of R code demonstrating k-means clustering.

can also be used (as in Figure 13.3).

There are several optional inputs to the kmeans function. One option is to
specify the algorithm we use. The default is the Hartigan-Wong algorithm
(Hartigan and Wong 1979), which uses a function minimisation method to assign
cluster membership. Alternatives are the Lloyd (1982) and Forgy (1965) methods
described earlier in the chapter (these are identical, but both names are accepted
by the function), and the MacQueen (1967) algorithm, which starts with the
centroids as far apart as possible. There may be particular data sets that are
better suited to one or other algorithm, so it is worth experimenting with different
options. Figure 13.13 shows some examples for comparison (the differences are
quite subtle), generated by the following code:
par(mfrow=c(1,3), las=1)

algorithmlist <- c('Hartigan-Wong','Lloyd','MacQueen')

for (plt in 1:3){
clusters <- kmeans(dataset[,1:2],5,algorithm = algorithmlist[plt])
plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=c(-1,1), ylim=c(-1,1))
axis(1, at=c(-1,1), tck=0.01, lab=F, lwd=2)
axis(2, at=c(-1,1), tck=0.01, lab=F, lwd=2)
title(paste(algorithmlist[plt],'algorithm'))
for (n in 1:(nrow(dataset))){
266CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

lines(c(clusters$centers[clusters$cluster[n],1],dataset[n,1]),
c(clusters$centers[clusters$cluster[n],2],dataset[n,2]),
col='grey')}
points(dataset[,1],dataset[,2],pch=16,col=pal2tone[1])
points(clusters$centers[,1],clusters$centers[,2],pch=16,cex=2)
}

Hartigan−Wong algorithm Lloyd algorithm MacQueen algorithm

● ● ●
● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ● ● ● ●
● ● ●
● ● ●
● ● ● ● ● ● ● ● ● ● ●
● ● ● ●
● ● ● ●
● ● ● ● ●
● ● ●
● ● ● ● ● ● ● ● ●
● ●
● ● ●
● ● ●
●
●● ● ● ●● ● ● ●● ● ●
● ● ● ● ● ●
● ●●● ●● ● ●●● ●● ● ●●● ●●
●
●● ● ● ● ● ●
●● ● ● ● ● ●
●● ● ● ● ●
● ● ● ● ●● ●● ● ● ● ● ●● ●● ● ● ● ● ●● ●●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
●
● ●● ● ●
● ●●
●
●●● ●
●
●
● ●● ● ●
● ●●
●
●●● ●
●
●
● ●● ●
● ●●
● ● ●● ●
●
● ● ● ● ● ● ● ● ● ● ● ●
● ●
●
●
● ● ●●
● ●
●
●
● ● ●●
● ●
●
●
● ● ● ●●
● ● ●● ● ● ● ● ●● ● ● ● ● ●● ● ●
●
● ●● ● ● ● ●
● ●● ● ● ● ●
● ●● ● ● ●
●● ● ● ●● ● ● ●● ● ●
● ● ● ● ● ●
● ●● ● ● ●● ● ● ●● ●
●● ● ●● ● ●● ●
● ● ● ● ● ● ● ● ●
●
● ● ●
● ● ●
● ●
● ● ●
● ● ●
● ●●
● ● ● ●
● ●●
● ● ● ●
● ●●
●
●● ●● ●●

● ● ●
● ● ●
● ● ●
● ● ●

Figure 13.13: Demonstration of different k-means clustering algorithms.

It is also possible to constrain the number of iterations permitted, and to start

from multiple random centroids using the nstart option, or even to specify the
starting centroids manually. These options allow substantial control over how
the clustering proceeds.
If we need to rescale the data before conducting k-means clustering, this can be
achieved using the scale function as described in Section 3.7. Scaled data will
typically produce better clustering when the dependent variables are in different
units.

13.12 Doing multidimensional scaling in R

There are several implementations of MDS in R, as well as in other software
packages and programming languages. Here we will demonstrate the cmdscale
function in the core stats package, which implements classical (metric) scaling,
as well as the isoMDS function for non-metric multidimensional scaling in the
MASS package (Venables, Ripley, and Venables 2002).
Multidimensional scaling operates on a matrix of distances between each pair
of points in the dataset. We need to compute this first using the dist function
(also part of the core stats package). This will take an N (observations) × m
(dependent variables) matrix and calculate the Euclidean (straight line) distances
between every possible pairing of data points, producing an N × N matrix.
Note that the output therefore does not depend on m, the number of dependent
variables. This is because the distance between points is calculated as a single
13.12. DOING MULTIDIMENSIONAL SCALING IN R 267

vector in the multidimensional space, regardless of its dimensionality. The data

object colourdata contains the example data from Figure 13.7b, arranged in
a 1000 (observations) × 4 (RGBα) matrix. We can calcluate the distances as
follows:
coldist <- dist(colourdata)

The coldist data object is a special class of matrix that contains the pairwise
distances in the correct format to use for multidimensional scaling. We can pass
this matrix (but not the raw data) to the cmdscale function as follows:
scaledxy <- cmdscale(coldist,2)
head(scaledxy)

## [,1] [,2]
## [1,] -0.15160650 0.6517739
## [2,] 0.41646691 -0.2108153
## [3,] -0.03000302 0.3792663
## [4,] -0.05064061 -0.6095281
## [5,] 0.09436012 0.1585279
## [6,] -0.22697391 0.1237203

The second argument for the cmdscale function (the number 2) defines the
dimensionality of the output. So, if we wanted to produce a 3D plot of the
data points, we could change this to 3. It is not clear how one might represent
dimensions higher than 3 graphically, but in principle any number of output
dimensions is possible. Each row of the output (scaledxy) shows the x,y position
of a data point (with row number consistent with the original data matrix).

We can then plot the rescaled data. Often it is helpful to colour-code the
individual points by some meaningful category. For this example, each data
point already has an RGBα colour vector associated with it, which we can use
to produce an attractive diagram (see 13.7b) as follows:
plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=c(-1,1), ylim=c(-1,1))
axis(1, at=c(-1,1), tck=0.01, lab=F, lwd=2)
axis(2, at=c(-1,1), tck=0.01, lab=F, lwd=2)
title(xlab="Dimension 1", col.lab=rgb(0,0,0), line=1.2, cex.lab=1.5)
title(ylab="Dimension 2", col.lab=rgb(0,0,0), line=1.5, cex.lab=1.5)
title(expression(paste('RGB',alpha,sep='')))

points(scaledxy[,1],scaledxy[,2],pch=16,cex=0.5,col=
rgb(colourdata[,1],colourdata[,2],colourdata[,3],alpha=colourdata[,4]))

An alternative is to use non-metric multidimensional scaling, which has less

stringent requirements about the solution. The following code uses the isoMDS
function from the MASS package:
268CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING

library(MASS)
mdsout <- isoMDS(coldist,k=2)
scaledxy <- mdsout$points

plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=c(-1,1), ylim=c(-1,1))

axis(1, at=c(-1,1), tck=0.01, lab=F, lwd=2)
axis(2, at=c(-1,1), tck=0.01, lab=F, lwd=2)
title(xlab="Dimension 1", col.lab=rgb(0,0,0), line=1.2, cex.lab=1.5)
title(ylab="Dimension 2", col.lab=rgb(0,0,0), line=1.5, cex.lab=1.5)
title(expression(paste('RGB',alpha,sep='')))

points(scaledxy[,1],scaledxy[,2],pch=16,cex=0.5,col=
rgb(colourdata[,1],colourdata[,2],colourdata[,3],alpha=colourdata[,4]))

If you run this code, you will notice that the solution is similar to the metric
version, but the non-metric diagram has more outliers at the extremes. In both
cases, the units of the scaled solution are arbitrary for both dimensions.

13.13 Practice questions

1. Standard k-means clustering works on what kind of distance between
points?
A) Rank-order
B) Manhattan
C) Cartesian
D) Euclidean
2. The accuracy of a cluster solution is based on:
A) The distances between data points and cluster centroids
B) The distances between pairs of data points
C) The distances between pairs of centroids
D) The total variance in the data set
3. Which statistic can be used to compare different numbers of clusters, taking
into account the degrees of freedom?
A) Summed squared error
B) Akaike information criterion
C) Root mean squared error
D) R2
4. Why is it important to normalize (scale) variables that have very different
units?
A) To remove outliers
B) Distance estimates cannot be calculated when the units are different
C) Variables with larger units will contribute too much to the distance
estimates
D) Variables with smaller units will contribute too much to the distance
estimates
13.13. PRACTICE QUESTIONS 269

5. Which of the following is not a variant of k-means clustering?

A) k-medians clustering
B) k-mediods clustering
C) inverse k-means clustering
D) spherical k-means clustering
6. Multidimensional scaling is used to:
A) Represent multivariate data in a lower-dimensional space
B) Represent bivariate data in a multidimensional space
C) Normalize all dimensions in a data set so they have equal variance
D) Represent multivariate data in a higher-dimensional space
7. We can check how well the distances between points are preserved after
multidimensional scaling using:
A) The summed Euclidean distance between data points
B) A Shepard diagram and correlation coefficient
C) The change in the number of dimensions
D) A Q-Q plot and the summed squared differences
8. Metric multidimensional scaling algorithms:
A) Use ranks rather than absolute values of the dependent variables
B) Are typically used for categorical dependent variables
C) Attempt to preserve the rank ordering of distances between data
points
D) Attempt to preserve the relative distances between data points
9. For a data set with N observations and m dimensions, the distance matrix
will have size:
A) N x m
B) m x m
C) N x 2
D) N x N
10. For a data set with 5 dimensions, which of the following would not be a
valid number of dimensions for the output of multidimensional scaling?
A) 1
B) 2
C) 3
D) 6
Answers to all questions are provided in section 20.2.
270CHAPTER 13. MULTIDIMENSIONAL SCALING AND K-MEANS CLUSTERING
Chapter 14

Multivariate pattern
analysis

The final multivariate technique we will discuss in this book is called Multivariate
Pattern Analysis (MVPA). In recent years it has been widely used to analyse
MRI recordings of brain activity, where it is sometimes referred to as MultiVoxel
Pattern Analysis, though the techniques (and acronyms) are much the same.
These methods are a subset of machine learning - a family of artificial intelligence
(AI) methods that aim to train computer algorithms to perform classification tasks
on some sort of complex data. Prominent examples of machine learning include
object identification algorithms (i.e. for labelling the contents of photographs)
and dictation software that converts speech to text (and in some cases can
act on verbal instructions). Machine learning methods also have substantial
promise in the area of personalised medicine and automated diagnosis, with one
prominent example being the diagnosis of eye disease (De Fauw et al. 2018).
These techniques will become more widespread and accurate in the future, and
at the time of writing (2021) are attracting substantial media attention and
commercial investment. In such a fast-moving field, it is always worth keeping
up with new developments. However, for a more detailed discussion of the core
aspects of pattern analysis and other machine learning methods, the classic
text, Pattern recognition and machine learning by Bishop (2006), is an excellent
resource.

14.1 Why use machines?

In some instances (e.g. object identification), the task we are interested in is one
that can be easily achieved by humans with a high degree of accuracy. In these
cases, the main benefit of automation is the speed and scale offered by computers.
For example, the internet contains millions of unlabelled images, and having

271
272 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

human operators manuallxy label each one would be prohibitively expensive (not
to mention tedious). An algorithm that can automatically identify their contents
makes the images searchable using text keywords, without requiring extensive
human labour. In other situations, algorithms can be used to identify patterns
in data that would be hard for humans to spot, perhaps owing to the complexity
of the data. The great promise of this aspect of AI is that it could help improve
critical real-world problems such as disease diagnosis and risk prediction in the
insurance industry.

14.2 Predicting group membership

To introduce the basic concept of classification, we will use a minimal example.
Consider some data sampled from two groups with different means, shown in
Figure 14.1a. Most students of statistics know what to do with such data: a
t-test (as demonstrated in Chapter 4) can tell us whether the group means differ
significantly or not. But what if we wanted to ask a different question? What if
we wanted to predict which group a participant is a member of, knowing only
their score? This is the basic idea of classification, and is the central problem
that MVPA algorithms are used to solve.

30 30
(a) (b)

20 20
●●● ● ● ●●● ● ●
●● ●●
● ●● ● ●●● ●
● ●●
●●●
● ●● ●● ● ●●● ● ●●
●●●
●
●● ●●● ● ●● ●●● ●
● ●●●
●● ● ●●● ●● ●● ● ●●●
●● ● ●●● ● ● ●●
●●● ●●● ● ●●● ●●● ●
10 ● 10 ●

0 0
Group A Group B

Figure 14.1: Minimal illustration of classification. Panel (a) shows example

simulated data from two groups with differing means (black lines). Panel (b)
shows the same data replotted with a category boundary (dashed line)

To think about classification, we could replot the data as a single cloud, as shown
in Figure 14.1b (the x-position of each point is arbitrary here). A good way
to try to classify group membership is to plot a category boundary that best
separates the two groups. This is shown by the dashed line in Figure 14.1b, and
14.3. DIFFERENT TYPES OF CLASSIFIER ALGORITHM AND PATTERN ANALYSIS273

a sensible decision rule would be to say that data points below the line (most
of the grey circles) are more likely to be in group A, and those above the line
(most of the blue squares) will be in group B. Of course, this classification is not
totally accurate for the current example - there are several blue squares below
the line and grey circles above it, and these will be misclassified.
The category boundary is a basic classification algorithm, and it prompts some
observations that will generalise to more complex cases. First, we can work out
the accuracy of the classification by calculating the percentage of data points that
are correctly identified. In the example in Figure 14.1b, this is something like
90%. It is also clear that if the group means were more similar, accuracy would
reduce, and if they were more different, accuracy would increase. Additionally,
the variance (spread) of the data points is important. If the variance were
greater, classification accuracy would decrease, and if the variance were smaller,
classification accuracy would increase.
A convenient way of summarising the mean difference and variance is to use the
Cohen’s d metric introduced in Chapter 5. The d statistic is the difference in
means, divided by the standard deviation. We can calculate how classification
accuracy (for a two-category data set) changes as a function of d, as shown by
the black curve in Figure 14.2. As we might expect, increasing the separation
between the group means increases the accuracy of our classifications. So far so
good, but up until this point our examples have had only a single dependent
variable - isn’t MVPA supposed to be a multivariate technique?
The same logic as described for a single variable can easily be extended to the case
of multiple variables. If we have two variables, we can try to classify data points
using both pieces of information by placing a category boundary to separate the
two-dimensional space created by plotting the variables against each other. This
is shown in Figure 14.3a for a linear classifier, where the line separating the white
and grey areas indicates the category boundary. Adding extra informative (and
uncorrelated) variables increases accuracy (see dashed blue curve in Figure 14.2).
In principle this same trick can be applied for any number of variables. In a real
data set, some variables will be informative whereas others will not, and there
will usually be some level of covariance between different measures. However this
is not generally a problem - classifier algorithms will tend to ignore uninformative
variables, and assign more weight to measures that improve accuracy.

14.3 Different types of classifier algorithm and

pattern analysis
There are many varieties of classifier algorithm available. The straight lines in
the examples we have seen so far were determined using a popular algorithm
called a linear support vector machine. This works by maximising the distances
between the boundary line and the closest example data points from each category
(termed the support vectors). Fundamentally this is an optimization problem,
274 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

100
Percentage correct

90

80

70

60 One DV
Two DVs

50
0 1 2 3 4
Difference in means/SD

Figure 14.2: Increase in accuracy with Cohen’s d for one (black) and two (blue)
dependent variables (DVs).
14.3. DIFFERENT TYPES OF CLASSIFIER ALGORITHM AND PATTERN ANALYSIS275

30 30
(a) (b)

20 20

● ● ● ●
●● ● ●● ●
● ●
● ●● ●●●● ● ●● ●●●●
●
● ● ●● ● ●
● ● ●● ●
10 ● ● ●●●● ● ● 10 ● ● ●●●●● ●
● ●● ●●● ●●
● ● ● ●● ●●● ●●
● ●
● ● ● ●
●● ●●
●● ● ● ●● ● ●
● ● ● ●

0 0
0 10 20 30 0 10 20 30

Figure 14.3: Examples of two-dimensional classification using a linear classifier

(a) and a radial basis function classifier (b).

and works in a similar way to the nonlinear curve fitting methods described in
Chapter 9. Support vector machines can also be created with nonlinear (radial)
basis functions (the basis function is the mathematical equation that is used to
construct the boundary line). These work by enclosing an ‘island’ of values from
one category (see Figure 14.3b). These can be more efficient for some types of
data, but also sometimes suffer from problems with generalisation to new data
sets (Schwarzkopf and Rees 2011).
Other types of algorithm can involve neural networks, which are based on
interconnected multi-layer processing of the type that happens in biological
neural systems. The input layer consists of a set of detectors that respond to
particular inputs or combinations of input. Each successive layer of the network
then applies a mathematical operation to the output of the previous layer. These
operations are usually fairly basic ones, such as weighted averaging (described
in section 6.8 in a different context), choosing the largest input, or a nonlinear
transform such as squaring. The end result of the network is an output layer,
from which classification decisions can be read. Neural networks can produce
very sophisticated operations (much like the brain), though it can sometimes be
difficult to understand fully what a trained neural network is actually doing.
A particularly useful variety is the deep convolutional neural network. These are
based on the early stages of sensory processing in the brain, and involve taking
images (or other natural inputs) and passing them through a bank of filters
that pick out specific low-level features from the input. An example filter bank
is shown in Figure 14.4, involving filters of different orientations and spatial
frequencies (see also Chapter 10 for details on how filters are applied to images).
276 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

Figure 14.4: Filter bank showing filters of different orientations and spatial
frequencies.

Deep neural networks are now advanced enough to classify images into different
categories, though it can sometimes be unclear precisely which features of an
image set are being used to do this. It is also important to avoid any confounds
in the input images that might produce false levels of precision, such as the
background of an image. For example, if you wanted to train a network to classify
criminals vs non-criminals from their photographs, it would be important to
make sure that all photographs were taken under the same conditions. Otherwise
it could be the case that all criminal photographs were taken against the same
background (e.g. the height gauge traditionally shown in police mugshots) and
these extraneous features would provide the network with a spurious cue.

A rather different approach to MVPA, that has been very influential in the fMRI
literature, is to forgo classification algorithms and instead use a correlation-based
approach (Haxby et al. 2001). In correlational MVPA, the pattern of brain
activity across multiple voxels (a voxel is the volumetric version of a pixel) is
correlated between two data sets derived from the same condition, or data sets
derived from two separate conditions. The logic is that if there is a distinct and
robust pattern of activity in a region of the brain, the correlation scores will
be higher when the data comes from the same condition than when it comes
from separate conditions. There are several variants of this method depending
on whether the data sets are derived within a single individual, or averaged
across multiple participants. Although this approach sounds very different from
the classification-based methods discussed above, in direct comparisons (e.g.
14.4. SITUATIONS WITH MORE THAN TWO CATEGORIES 277

Coggan, Baker, and Andrews 2016; Isik et al. 2014; Grootswagers, Wardle, and
Carlson 2017) they behave quite similarly. A related approach is to directly
calculate the multivariate effect size (the Mahalanobis distance, see section 3.4.3)
between conditions, and use this as a measure of pattern distinctness (Allefeld
and Haynes 2014).

14.4 Situations with more than two categories

Classifiers are not limited to discriminating between two possible categories.
The same principles can be applied to problems with an arbitrary number of
categories or conditions. If this is done, it is important to be aware that the
baseline guess rate will also change. For two categories, the guess rate will be
0.5 (or 50%), because an algorithm that assigned data to one of two categories
randomly will tend to get things right half of the time (assuming a balanced
data set). For ten categories, the guess rate will be 0.1 (or 10%). The guess rate
is therefore given by 1/m, where m is the number of categories (see also section
16.6 for a related situation).
However, it is worth being aware that in situations with more than two categories,
good performance may not be equally distributed between the categories. For
example, a situation with three categories (A, B and C) could involve good
discrimination between A and B, but poor discrimination between B and C. This
is an analogous problem to interpreting a main effect in ANOVA - the main effect
itself does not tell us which pairs of conditions differ. Just as ANOVAs are often
followed up by pairwise post hoc comparisons between different conditions, it is
also possible to follow up a multi-way MVPA analysis with pairwise discrimination
between different categories.

14.5 Preparing data for analysis

Before performing classification, it is good practice to normalize one’s data.
Usually this is done by subtracting the group mean from all data points that will
be involved in classification, independently for each variable. It is also common
to scale by the variance. The idea here is to remove substantial differences in the
scale of the different variables, much like converting to z-scores. Some MVPA
software does this automatically, but it can also be done manually (see section
3.7).
Next, the data will usually be split into a training set and a test set. The classifier
is trained using the training set, and then tested on the unseen observations
in the test set to determine its accuracy. This is an important step to avoid a
‘double dipping’ confound, where a model is trained and tested on the same data.
Such confounds will tend to inflate accuracy because of overfitting. Classification
will generally be repeated many times with different allocations of observations
to training and test sets, to work out an average accuracy that should be more
278 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

robust than for a single partitioning. Some MVPA software will implement this
automatically using a technique called k-fold cross validation. The value of the k
parameter determines the number of subsets the data are split into. The model
is trained on k-1 subsets, and tested on the remaining subset. This is repeated
for all permuatations (i.e. each subset is the test set once). We can replace the k
with its value when referring to this type of analysis, e.g. 5-fold cross validation,
or 10-fold cross validation.

14.6 Assessing statistical significance

Once we have determined the classifier accuracy, it is sometimes useful to test
whether this exceeds chance level, or perhaps to compare accuracy statistically
between different classifications. In situations where we have a single measure
of classifier accuracy, the binomial test is an appropriate way to do this. The
binomial test can compare the proportion of correct classifications with the guess
rate, taking into account the number of observations. This statistic will have
greater power (see Chapter 5) the more observations are included in the test set.
An alternative is to generate 95% confidence intervals on the classifier accuracy
estimate using a resampling method (see Chapter 8). One way to do this would
be to repeat the classification using many random partitionings of the original
data into training and test sets, to build up a distribution of classifier accuracy
scores. The empirical 95% quantiles (i.e. the points at 2.5% and 97.5% on the
distribution) act a bit like a t-test - if these do not overlap the guess rate, the
accuracy can be said to differ significantly from chance.
Finally, in experimental designs involving multiple participants, and where the
classification is performed separately on each individual participant, traditional
statistical tests can be used to assess significance of the mean accuracy level
across participants. These might include parametric or non-parametric t-tests,
or analysis of variance (ANOVA). Tests can be performed on either the classifier
accuracy scores, or measures derived from these, such as d′ (see Chapter 16).
One would take the score(s) for each participant, and compare these values
to chance using a one-sample t-test, or across groups with a two-sample t-test
or ANOVA. In situations where many separate classifications are conducted
(e.g. across time, or across different locations in the brain), cluster correction
algorithms can be used to correct for multiple comparisons (see Chapter 15).

14.7 Ethical issues with machine learning and

classification
This is not intended to be a sociology text. However, machine learning techniques
carry with them several potential dangers, because they are often presented as
being fair and unbiased. After all, how could a computer algorithm be sexist
or racist? The reality is that any algorithm can be very easily biased by the
14.8. DOING MVPA IN R USING THE CARET PACKAGE 279

prejudices of those who construct them, can perpetuate systemic inequalities,

and have the capacity to be used for nefarious purposes. Two very high profile
real-world examples are gender discrimination in automated recruitment software,
and algorithms designed to classify based on sexual orientation.

In the first example, in around 2018 a large technology company scrapped its
automated recruitment system because it was shown to rate women’s CVs less
highly than men’s for software and technology-related jobs. It did this in some
surprisingly blatant ways, such as penalizing graduates of female-only universities,
and down-weighting CVs that included the word “women’s” - male candidates
would be unlikely to mention being captain of the women’s basketball team, for
example. Why did this happen? It turns out that the algorithm was trained on
historical data from two groups of candidates - those who had been hired, and
those who had not. All of those hiring decisions were made in the traditional
way, by humans with their own prejudices about what makes a good software
engineer. Far from being unbiased, the algorithm inherited and perpetuated the
prejudices of the industry it was created to serve.

The second controversial example was a study claiming that deep neural networks
could classify sexual orientation from photographs more accurately than humans
(Wang and Kosinski 2018). The authors proposed that subtle differences in
facial morphology might reveal exposure to various sex hormones in the womb,
which also influence sexual orientation in adulthood. This work was criticised
on several grounds, including that most of the photographs used to train the
algorithm were of caucasian models, and that the algorithm appeared to be
classifying photographs based on cues that were unrelated to what the authors
claimed, including makeup and the presence or absence of glasses. But the key
point is that the use of machine learning algorithms in this way is extremely
unethical. There are many societies where homosexuality is illegal, and tools
that can be used to classify sexual orientation (no matter how accurate, or using
what cues) could be used to oppress innocent people. Machine learning is a
powerful tool, but it is crucial that it is used responsibly and ethically, and
that the apparent objectvity of computer algorithms is not used to mask human
prejudice and bias.

14.8 Doing MVPA in R using the Caret package

The following examples will use a package called caret (pronounced like ‘carrot’),
which is short for Classification And REgression Training (Kuhn 2008). Caret
is a general purpose package, that provides access to a wide range of algorithms
(over 200) through a consistent interface. It is certainly not the only machine
learning package available in R, but it is one of the most flexible. There are
MVPA packages available in most other contemporary programming languages,
including built-in tools in Matlab, the PyMVPA toolbox in Python, and the
cross-platform LibSVM package.
280 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

We will use two functions from the caret package: train and predict. These do
much as you would expect from the names. The train function is used to train
a pattern classifier algorithm, and outputs a data object containing the model
specification. We can then pass this model specification into the predict function,
along with some unseen data, to get predictions out. If we know the ground
truth categories (i.e. the actual categories) for the unseen data, we can also
calculate the classifier’s accuracy. Of course caret is capable of far more than
what we are doing here, but this is a good starting point to demonstrate the
basics. There is extensive documentation available on the package’s web pages
at http://caret.r-forge.r-project.org.

14.9 Categorising cell body segmentation

The first scenario we will consider to demonstrate MVPA uses an example data
set from the caret package, originally from a paper by Hill et al. (2007). The
segmentationData data set contains data for cells that are either classed as ‘well
segmented’ or ‘poorly segmented’. There are over 2000 example cells and 58
predictor variables. The data has been pre-allocated to training and test sets,
using the Factor column segmentationData$Case. We can therefore load in the
caret package and the data set, and take a look at what we have:
library(caret) # load in the caret package
data(segmentationData) # load the data set into the Environment
dim(segmentationData) # request the size of the data frame

## [1] 2019 61
head(segmentationData[,1:6])

## Cell Case Class AngleCh1 AreaCh1 AvgIntenCh1

## 1 207827637 Test PS 143.247705 185 15.71186
## 2 207932307 Train PS 133.752037 819 31.92327
## 3 207932463 Train WS 106.646387 431 28.03883
## 4 207932470 Train PS 69.150325 298 19.45614
## 5 207932455 Test PS 2.887837 285 24.27574
## 6 207827656 Test WS 40.748298 172 325.93333
We see that the data set has 2019 rows and 61 columns. The first three columns
are the cell ID number, whether it is to be used for test or training, and whether
it is well segmented or poorly segmented. The remaining 58 columns contain
different measurements like the areas, widths and lengths of different parts of
the cell, obtained by microscope imaging. We can extract the training and test
data into separate matrices as follows:
# separate the data for the training set and the test set and convert to matrices
# only include the columns (4 to 61) that contain actual measurements
trainingdata <- as.matrix(segmentationData[which(segmentationData$Case=='Train'),4:61])
14.9. CATEGORISING CELL BODY SEGMENTATION 281

testdata <- as.matrix(segmentationData[which(segmentationData$Case=='Test'),4:61])

Each data object is now an N × 58 matrix. In the training set, N = 1009, and
in the test set N = 1010, where N corresponds to the number of cells.
Next, we can store the true categories (well segmented, WS, or poorly segmented,
PS) in separate data objects for the training and test sets. These are found in
the third column of the segmentationData data frame, that is headed Class:
traininglabels <- segmentationData[which(segmentationData$Case=='Train'),3]
testlabels <- segmentationData[which(segmentationData$Case=='Test'),3]

traininglabels[1:20] # output some example labels

## [1] PS WS PS WS PS PS PS WS WS WS WS PS WS PS PS PS PS PS PS PS
## Levels: PS WS
The labels are a factor variable (as described in section 3.9), just as one would
use to specify group membership in an ANOVA design. Now that we have our
data prepared for classification, we can train a classifier on the training data set
using the train function as follows:
svmFit <- caret::train(trainingdata, traininglabels, method = "svmLinear")
svmFit

## Support Vector Machines with Linear Kernel

##
## 1009 samples
## 58 predictor
## 2 classes: 'PS', 'WS'
##
## No pre-processing
## Resampling: Bootstrapped (25 reps)
## Summary of sample sizes: 1009, 1009, 1009, 1009, 1009, 1009, ...
## Resampling results:
##
## Accuracy Kappa
## 0.7968955 0.5614672
##
## Tuning parameter 'C' was held constant at a value of 1
Note that there are multiple functions called train in different R packages, so
here we specify the one from the caret package with the syntax caret::train. The
model object (svmFit) contains some information about what has been done, as
shown above, though we rarely need to look at this directly. We can then take
the trained model, and ask caret to predict categories for the unseen (test) data
using the predict function. The outcome (stored in a new data object called p)
is a set of category predictions (PS or WS) for each cell in the test set.
282 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

p <- predict(svmFit,newdata = testdata)

p[1:20]

## [1] PS PS WS WS PS PS PS WS PS PS WS PS PS WS PS WS PS PS WS PS
## Levels: PS WS

We can compare this to the true categories by counting up how many match the
true categories (stored in testlabels) and converting to a percentage:
numbercorrect <- sum(testlabels==p)
totalexamples <- length(testlabels)
100*(numbercorrect/totalexamples)

## [1] 79.80198

The classifier has done very well, getting about 80% of the cells in the test set
correct. We can see if this is statistically significant using a binomial test to
compare this to chance performance (0.5, or 50% correct):
binom.test(numbercorrect,totalexamples,0.5)

##
## Exact binomial test
##
## data: numbercorrect and totalexamples
## number of successes = 806, number of trials = 1010, p-value < 2.2e-16
## alternative hypothesis: true probability of success is not equal to 0.5
## 95 percent confidence interval:
## 0.7719120 0.8223772
## sample estimates:
## probability of success
## 0.7980198

The test is highly significant, which is telling us that a linear support vector
machine with the full training set does pretty well. It classifies about 80% of
the test set correctly, which is significantly above chance performance of 50%
correct. What if we used a different kernel? Switching to a radial basis function
improves things by around 1%:
svmFit <- caret::train(trainingdata, traininglabels, method = "svmRadial")
p <- predict(svmFit,newdata = testdata)
100*(sum(testlabels==p)/length(testlabels))

## [1] 80.49505

Alternatively, we can try a neural network model called a Multilayer Perceptron.

For this particular data set, the perceptron does much worse than the support
vector machines (as well as taking longer):
14.9. CATEGORISING CELL BODY SEGMENTATION 283

svmFit <- caret::train(trainingdata, traininglabels, method = "mlp")

p <- predict(svmFit,newdata = testdata)
100*(sum(testlabels==p)/length(testlabels))

## [1] 66.53465

Presumably there will be other types of data where the perceptron would be
a better choice. Finally, we can explore how accuracy increases as we include
more of the dependent variables in the classification (see Figure 14.5).
perccor <- NULL
for (n in 1:19){
trainingdata <- as.matrix(segmentationData[which(segmentationData$Case=='Train'),4:(n+4)])
testdata <- as.matrix(segmentationData[which(segmentationData$Case=='Test'),4:(n+4)])
svmFit <- caret::train(trainingdata, traininglabels, method = "svmLinear")
p <- predict(svmFit,newdata = testdata)
perccor[n] <- 100*(sum(testlabels==p)/length(testlabels))
}

plot(2:20,perccor,type='l',ylim=c(50,100),lwd=3)
points(2:20,perccor,pch=16)
100
90
80
perccor

● ● ● ● ● ●
● ● ● ●
● ● ● ●
● ● ● ●
70

●
60
50

5 10 15 20

2:20

Figure 14.5: Classification accuracy as a function of number of dependent

variables, for the cell body segmentation data set.
284 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

As more dependent variables are added, classifier accuracy improves - rapidly

at first (e.g. from 2 to 3 variables), and then more gradually, before plateauing
around 15 dependent variables.

14.10 Decoding fMRI data

The second example data set we will use to demonstrate MVPA originates from
a classic study by Haxby et al. (2001). In this study, participants viewed images
from different image categories, while their brain activity was recorded using
an MRI scanner. We will use a subset of conditions from a single participant,
consisting of 12 trials in which the stimuli were either faces, houses, or scrambled
images. We will compare responses from 30 locations (voxels) in a region of the
brain called the fusiform face area (FFA), which is thought to be specialised
for processing faces. The full data set is available on the PyMVPA website
(http://www.pymvpa.org/datadb/haxby2001.html) (and note that if you read
the original paper, you will see that Haxby et al. actually used an alternative
form of MVPA that is based on correlation).

First of all, let’s see if the FFA can tell the difference between face images and
scrambled images. This is a standard test for ‘selectivity’ of a particular category.
We can set up the trainingdata matrix to contain half of the data from each
condition to train the classifier on as follows:
# create an empty matrix
trainingdata <- matrix(0,nrow=12,ncol=30)
# copy half of the face data into the matrix
trainingdata[1:6,] <- facedata[1:6,]
# copy half of the scrambled data into the matrix
trainingdata[7:12,] <- scrambdata[1:6,]
# convert to a data frame
trainingdata <- data.frame(trainingdata)

We will also need to create numerical labels to tell the classifier which condition
each observation corresponds to. We can just use the numbers 1 and 2 for this
as we have two conditions (face and scrambled):
# create a factor with two levels and six repetitions
traininglabels <- gl(2,6)
traininglabels

## [1] 1 1 1 1 1 1 2 2 2 2 2 2
## Levels: 1 2

Then we can do the same thing with the other half of the data to create a testing
set (notice here we choose trials 7 to 12 instead of 1 to 6), for assessing the
accuracy of the trained model:
14.10. DECODING FMRI DATA 285

testdata <- matrix(0,nrow=12,ncol=30)

testdata[1:6,] <- facedata[7:12,]
testdata[7:12,] <- scrambdata[7:12,]
testdata <- data.frame(testdata)
testlabels <- gl(2,6)

As for the previous example, we next train the model using the train function
and test it on an unseen data set with the predict function. We pass in the
training data, the labels identifying the conditions, and specify the algorithm we
want to use (in this case a linear support vector machine). Then we plug the
trained model into the predict function along with the test data:
svmFit <- caret::train(trainingdata, traininglabels, method = "svmLinear")
p <- predict(svmFit,newdata = testdata)
p

## [1] 2 1 1 1 1 1 2 2 2 2 2 2
## Levels: 1 2
The predictions are stored in the data object p, and are condition labels with
values of 1 or 2. You can see that five of the 12 examples have been classified as
condition 1, and 7 have been classified as condition 2. How did this correspond
to the true values? As before, we can work out the accuracy by adding up the
number of correctly classified examples, and then converting to a percentage:
numbercorrect <- sum(testlabels==p)
totalexamples <- length(testlabels)
100*(numbercorrect/totalexamples)

## [1] 91.66667
So for this example, we can see that the algorithm has over 90% accuracy. This
strongly suggests that the brain region we are looking at responds differently to
faces than it does to scrambled images. Next, let’s see if it produces a distinct
response to pictures of houses. The following code duplicates the example above,
except that I have replaced instances of facedata with housedata.
trainingdata <- matrix(0,nrow=12,ncol=30)
trainingdata[1:6,] <- housedata[1:6,]
trainingdata[7:12,] <- scrambdata[1:6,]
trainingdata <- data.frame(trainingdata)
traininglabels <- gl(2,6)

testdata <- matrix(0,nrow=12,ncol=30)

testdata[1:6,] <- housedata[7:12,]
testdata[7:12,] <- scrambdata[7:12,]
testdata <- data.frame(testdata)
testlabels <- gl(2,6)
286 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

svmFit <- caret::train(trainingdata, traininglabels, method = "svmLinear")

p <- predict(svmFit,newdata = testdata)
100*(sum(testlabels==p)/length(testlabels))

## [1] 50
For houses, accuracy is at 50% correct, so the classifier has not exceeded chance
levels. This indicates that the FFA is not selective for images of houses. Finally,
we could do a 3-way classification between all stimulus types:
trainingdata <- matrix(0,nrow=18,ncol=30)
trainingdata[1:6,] <- facedata[1:6,]
trainingdata[7:12,] <- housedata[1:6,]
trainingdata[13:18,] <- scrambdata[1:6,]
trainingdata <- data.frame(trainingdata)
traininglabels <- gl(3,6)

testdata <- matrix(0,nrow=18,ncol=30)

testdata[1:6,] <- facedata[7:12,]
testdata[7:12,] <- housedata[7:12,]
testdata[13:18,] <- scrambdata[7:12,]
testdata <- data.frame(testdata)
testlabels <- gl(3,6)

svmFit <- caret::train(trainingdata, traininglabels, method = "svmLinear")

p <- predict(svmFit,newdata = testdata)
100*(sum(testlabels==p)/length(testlabels))

## [1] 44.44444
This time around, we have above chance decoding at 44% correct (remember
that because there are three categories, the guess rate is 1/3, or 33% correct).
This basic MVPA analysis of MRI data has gone pretty well as it has given us
quite a clear answer. As mentioned above, in a real MRI study, we would repeat
the classifications many times in a loop, randomly reshuffling the examples we
use to train and test the model, and then averaging the accuracies that are
produced (see Chapter 8 for details of resampling methods). Accuracy scores
across multiple participants can then be compared using traditional statistics
such as t-tests. For some data sets, perhaps using EEG or MEG methods, we
can repeat classification at different moments in time to see how brain signals
evolve (see Chapter 15 for further discussion of this).

14.11 Practice questions

1. For classification between four categories, what is the baseline (guess) rate?
A) 4% correct
14.11. PRACTICE QUESTIONS 287

B) 25% correct
C) 50% correct
D) 70% correct
2. As the Cohen’s d effect size between two conditions increases, classifier
accuracy should:
A) Increase
B) Decrease
C) Stay the same
D) It will depend on how many dependent variables there are
3. What will a linear classifier use to partition data into categories?
A) Any arbitrary curve
B) A straight line, plane or hyperplane
C) A radial curve
D) A sine wave
4. Instead of using a classifier algorithm, MVPA can also be conducted based
on:
A) An extremely fast supercomputer
B) Scores rounded to the nearest integer
C) Reduced data from a factor analysis
D) Correlation
5. An important step in data pre-processing before running MVPA is:
A) Subtracting the mean differences between conditions
B) Conducting univariate analyses
C) Normalization
D) Squaring all measurements
6. Neural network classifier algorithms involve at least one:
A) Spatial scale of filter
B) Simulated calcium channel
C) Hidden network layer
D) Real human neuron
7. If a classifier is trained and tested on the same data, what is the most
likely outcome?
A) Accuracy will be perfect
B) Accuracy will be inflated because of overfitting
C) Accuracy will be reduced because of overfitting
D) Accuracy will be at chance
8. A significant 3-category classification can be interpreted by:
A) Running post-hoc pairwise classifications
B) Running an Analysis of Variance (ANOVA)
C) Removing the least informative category
D) Adjusting the guess rate for a two-category classification
9. In a support vector machine, the support vectors refer to:
A) The dependent variables
B) The distance from the category boundary to the nearest points
C) The distance from the category boundary to each data point
D) The weights that each dependent variable is multiplied by
288 CHAPTER 14. MULTIVARIATE PATTERN ANALYSIS

10. Deep convolutional neural networks are inspired by the structure of:
A) The convoluted (folded) structure of the human brain
B) Complex databases of natural images
C) A bank of filters with different orientations and spatial frequencies
D) Biological sensory systems
Answers to all questions are provided in section 20.2.
Chapter 15

Correcting for multiple

comparisons

Most introductory statistics courses introduce the concept of the familywise error
rate, and correction for multiple comparisons. The idea is that the more statistical
tests you run to investigate a given hypothesis, the higher the probability that
one of them will be significant, even if there is no true effect. Traditionally,
the solution to this problem has been to correct the criterion for significance
(i.e. the α-level) to account for the number of comparisons. We will first discuss
several such methods (and their shortcomings), before introducing two newer
ideas: the false discovery rate, and cluster correction. Both approaches deal with
multiple comparisons in a principled way, whilst maintaining statistical power at
higher levels than older methods. Controlling the false discovery rate is generally
appropriate when the tests are independent, whereas cluster correction should
be used in situations where correlations are expected between adjacent levels of
an independent variable (e.g. across space or time).

15.1 The problem of multiple comparisons

All frequentist statistical tests have a built in false positive rate (or Type I
error rate), determined by the criterion for significance (the α level). In many
disciplines the widely accepted criterion is α = 0.05. This means that the false
positive rate for a single test is 5% - one in twenty statistical tests will return a
significant result even when there is no true effect. The overall false positive rate
for a family of tests - the chance that at least one test is erroneously significant -
increases the more tests we conduct. For tests on independent data sets, this
family-wise error rate (FWER) is calculated as:

F W ER = 1 − (1 − α)m (15.1)

289
290 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

where α is the criterion for significance, and m is the number of tests. This
function is plotted in Figure 15.1 for α = 0.05, and shows that the false positive
rate rises rapidly, such that with 14 tests there is a 50% chance of at least one
test being significant. With >60 tests, a false positive is virtually guaranteed.

1
Familywise error rate

0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
Number of tests (m)

Figure 15.1: Familywise error rate as a function of the number of tests (m),
assuming no true effect. The horizontal dashed line indicates the alpha level of
0.05.

The multiple comparisons problem is frequently presented as justification for

avoiding using multiple t-tests, and instead using an omnibus test such as
Analysis of Variance (ANOVA). However, ANOVA itself suffers from issues with
familywise error rates that are not widely appreciated (see Cramer et al. 2016;
Luck and Gaspelin 2017). Although ANOVA takes into account the number of
comparisons within an independent variable (i.e. the number of levels of that
variable), it does not control for familywise error across the number of main
effects and interactions. This means that the chance of a large factorial ANOVA
producing a false positive is worryingly high. For example, with 4 factorially
combined independent variables, there will be 6 (2-way) + 4 (3-way) + 1 (4-way)
= 11 interaction terms, so 15 effects in total including the main effects. The
interdependencies inherent in factorial designs mean that the curve in Figure
15.1 doesn’t directly apply to ANOVA, but it’s clear that marginally significant
effects in large factorial designs should be treated with caution.
The multiple comparisons problem is also an issue for some modern high through-
put and data mining techniques, where large amounts of data are readily available.
15.2. THE TRADITIONAL SOLUTION: BONFERRONI CORRECTION291

In genetics, genome-wide association studies involve many thousands of simulta-

neous comparisons between different SNPs (single-nucleotide polymorphisms)
and some phenotype such as disease susceptibility. In brain imaging studies, many
thousands of voxels (volumetric pixels) are recorded simultaneously, and often
compared across different experimental conditions to determine which brain re-
gions respond to a particular stimulus. Finally, the age of big data and widespread
internet availability means that researchers have access to an arbitrarily huge
number of variables, some of which will be spuriously associated by chance. There
is an excellent website (https://www.tylervigen.com/spurious-correlations) and
book (Vigen 2015) with many absurd examples, such as the correlation (from
1999 to 2009) between people drowning in swimming pools, and films released
starring Nicholas Cage (r = 0.67).
The issues with Type I errors are well-known in their respective fields, and are
potentially the cause of numerous published results of dubious provenance. Some
especially questionable research practices that take advantage of an inflated false
positive rate include p-hacking (measuring many variables, but reporting only
those that are significant), and optional stopping (continuing to collect data until
a significant result is found). This may partly explain why large-scale attempts
to replicate published findings have extremely low success rates (Open Science
Collaboration 2015).

15.2 The traditional solution: Bonferroni cor-

rection
The simplest solution to the multiple comparisons problem is to modify the α
level by dividing it by the number of tests (m):

α
ᾱ = (15.2)
m

where ᾱ is the corrected significance criterion. This is known as Bonferroni

correction, and has a reputation for being a conservative solution. It reduces the
family-wise error rate to be less than or equal to α, which can be confirmed by
calculating:

α m
F W ER = 1 − (1 − ) (15.3)
m

For large values of m, the error rate will always be slightly below α with
this formula. For example, with α = 0.05 and m = 100, the error rate is
1 − (1 − 0.05
100 )
100
= 0.0488. The corrected α-level (ᾱ) is then used to threshold
the p-values of each statistical test to determine significance.
An alternative implementation of Bonferroni correction, which is the default in
some statistical software packages such as SPSS, is to adjust the p-values instead
292 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

of the α-level. This is achieved by multiplying the p-value by m, capping the

upper limit at 1. For this approach, the adjusted p-value is then compared to
the standard (i.e. uncorrected) threshold value of α.
To illustrate the distinction between these two implementations, imagine that
we run eight tests, and one of the p-values comes out at p = 0.0043. We could
correct our α level to 0.05/8 = 0.00625, and compare the p-value to this threshold.
Since 0.0043 < 0.00625, we would conclude the test is significant. Alternatively
we can correct the p-value itself by calculating that 0.0043*8 = 0.034. Since
0.034 < 0.05, we would again conclude the test is significant. These two methods
are equivalent, but it is important to report which one you are using so that
someone reading your results understands what you have done.

15.3 A more exact solution: Sidak correction

A very similar correction that is slightly more precise, and slightly less conserva-
tive, than Bonferroni correction is the Sidak correction. This sets the α level as
the inverse of the equation for calculating familywise error. So, the adjusted α
value, ᾱ, is given by:

ᾱ = 1 − (1 − α)1/m (15.4)

Notice that this is very similar to equation (15.1), except that the exponent is
1/m instead of m. It has the effect of reversing the familywise error rate, so that
when it is plugged into the familywise error calculation, the false positive rate
remains fixed at α (i.e. 0.05), regardless of the number of tests. For our example
scenario of 8 tests, the Sidak-corrected criterion will be:

ᾱ = 1 − (1 − α)1/m = 1 − (1 − 0.05)1/8 = 0.0064.

15.4 A higher power solution: the Holm-

Bonferroni correction
Bonferroni and Sidak correction mean that each individual test must reach a
higher level of significance (i.e. a smaller p-value) before it can be considered
significant, compared with if it were run in isolation (and not part of a family
of tests). The Holm-Bonferroni method is less conservative than this (Holm
1979). It works by taking an ordered list of p-values from all tests, and adjusting
the criterion for significance sequentially for each item in the list. The test
α
with the smallest p-value will be assessed against a criterion of m , the next
α
smallest p-value against a criterion of m−1 and so on. If we define j as the ranked
α
position of a p-value, the criterion is m−(j−1) . The largest p-value is assessed
against the original uncorrected value of α, because when j = m, it follows that
15.5. ADJUSTING α REDUCES STATISTICAL POWER 293

α
m−(m−1) = α1 . This has the effect of keeping the familywise error rate at or
below α, but reducing the Type II error rate (i.e. being less likely to miss true
effects). There is an equivalent method known as the Holm-Sidak method, where
the Sidak correction is progressively applied instead of the Bonferroni correction.

15.5 Adjusting α reduces statistical power

A big problem with corrections that adjust the α level is that they reduce
statistical power. As discussed in Chapter 5, power is the ability of a statistical
test to detect a true effect. Because power depends directly on the α level, a
consequence of adjusting α is that either effect sizes or sample sizes must become
much larger to detect significant effects, compared with situations where no
correction is applied. For example, Figure 15.2 illustrates how power for an
individual test is reduced as the number of tests in the family increases, for
Bonferroni corrected α levels. Figure 15.2a shows that for a medium effect size
(d = 0.5), correcting across 10 comparisons requires an increase in sample size
from around N = 64 to N = 109 to maintain power at 0.8 (see the arrows).
Figure 15.2b shows that for a fixed sample size (N = 50), the minimum effect
size that can be detected with a power of 0.8 increases from around d = 0.57
with m = 1 to d = 0.89 with m = 100. This means that designing a study that
will involve conducting many tests will tend to require a much larger sample size
(or be limited to detecting only large effects), compared with a study involving
fewer tests.
In the context of making adjustments to the α level, it is worth some consideration
of where the criterion of α = 0.05 comes from, and whether it remains appropriate.
This threshold is often treated as an absolute limit on determining statistical
significance, however it is really only a historical convention, and differs across
research fields. The value of 0.05 was proposed by Ronald Fisher, the inventor
of ANOVA, almost 100 years ago (Fisher 1926). Fisher felt that a one-in-twenty
chance of a false positive gave a reasonable balance between identifying a real
effect and being misled by noise in the data. However he was not as dogmatic
about this value as many more recent scientists (and journals) have become -
he also suggested considering α values of 0.02 and 0.01 if the level of 0.05 ‘does
not seem high enough odds’. In recent years, there have been calls to redefine
significance at a lower value of α = 0.005 (Benjamin et al. 2017), or to provide
an explicit rationale for choosing a given value of α (Lakens et al. 2018), or
to dispense with the concept of significance entirely (Amrhein, Greenland, and
McShane 2019). Ultimately it is the job of a researcher or analyst to choose a
criterion that is convincing to both themselves and those reading their work,
and this also applies to correction for multiple comparisons. Alternatively, see
Chapter 17 for a different approach to statistical hypothesis testing.
Whatever baseline level of α we choose, one potential solution to avoiding overly
stringent multiple comparison correction is to reduce the number of tests. In
recent years, preregistering an analysis plan in which the anticipated comparisons
294 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

1 1
(a) d = 0.5 (b) N = 50
0.8 0.8

0.6 0.6
Power

Power
0.4 0.4

m=1
0.2 m=3 0.2
m = 10
m = 100
0 0
0 50 100 150 200 0 0.2 0.4 0.6 0.8 1
Sample size (N) Effect size (d)

Figure 15.2: Power curves for a single corrected test, for different sizes of test
family (m), following Bonferroni correction. Panel (a) shows how power increases
as a function of sample size, for an effect size of d = 0.5. Panel (b) shows how
power increases as a function of effect size, for a sample size of N = 50. The
dashed horizontal line in both panels shows the target of 80% power.

are specified in advance has been suggested as one way to reduce researcher
degrees of freedom. But in many situations, such as novel and exploratory work,
it is not possible to specify which conditions are expected to produce significant
results until after the data have been collected. In such cases, there are two
relatively recent developments designed to control the rate of Type I errors
but still maintain statistical power as far as possible. These are false-discovery
rate correction, and cluster correction. The former method generally makes
the assumption that the various tests are independent. The latter method is
specifically used when there are correlations between successive tests.

15.6 Controlling the false-discovery rate

The familywise error rate described above is the probability that one out of
a family of tests will be a false positive. The false-discovery rate (FDR) is
subtly different - it is the proportion of all significant results that are false
positives. A seminal paper by Benjamini and Hochberg (1995) introduced
methods for controlling the false-discovery rate at a specific level. From a
philosophical perspective, this approach accepts that some of our significant
results will be false positives (Type I errors), and tries to keep this at a known
level, so as to minimise the number of missed true effects (Type II errors). It is
particularly appropriate for research designs where significant results are followed
15.7. COMPARISON OF FDR VS FWER CORRECTION 295

up by further experiments, for example in drug development where a series of

candidates might be screened, with those producing significant results subjected
to additional tests.
The desired false-discovery rate is represented by a value called q ∗ . We will
assume, as Benjamini and Hochberg (1995) did, that q ∗ = 0.05. However for
large data sets it is possible to estimate q ∗ empirically (see Storey and Tibshirani
2003). Much like the Holm corrections, the Benjamini-Hochberg method also
takes a rank-ordered
∗
list of p-values, but this time compares each value to the
criterion ᾱ = qmj (where again, j is the position in the list, from 1 to m). The
threshold for significance across three methods is compared in Figure 15.3, for a
family of m = 100 tests at each position (j) in the rank-ordered list of p-values.
Because the Benjamini-Hochberg curve sits above the other two lines, it is by
far the least conservative of the three methods.
0.05
Significance threshold

Benjamini−Hochberg
0.005

Holm−Bonferroni

Bonferroni
0.0005

0 10 20 30 40 50 60 70 80 90 100
Ranked position of test (j)

Figure 15.3: Effect of three types of multiple comparison correction across 100
tests on the threshold for significance (y-axis) at each position in a rank-ordered
list of p-values (x-axis). Note the logarithmic scaling of the y-axis.

15.7 Comparison of FDR vs FWER correction

To see how false-discovery rate (FDR) and family-wise error rate (FWER)
correction compare, let’s imagine that we run a series of t-tests that produce the
following set of ten p-values:
296 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

Table 15.1: Example of multiple comparisons correction for a set of 10 tests. j

is the position in the rank ordered list of p-values, and the column headed p
gives the p-values. The columns headed Bonf, Holm and FDR (thresh) give the
critical (threshold) values for Bonferroni correction, Holm-Bonferroni correction,
and false-discovery rate correction.

j p Bonf (thresh) Holm (thresh) FDR (thresh)

1 0.002 0.005 0.005 0.005
2 0.003 0.005 0.006 0.010
3 0.005 0.005 0.006 0.015
4 0.006 0.005 0.007 0.020
5 0.012 0.005 0.008 0.025
6 0.022 0.005 0.010 0.030
7 0.049 0.005 0.012 0.035
8 0.088 0.005 0.017 0.040
9 0.197 0.005 0.025 0.045
10 0.754 0.005 0.050 0.050

## [1] 0.006 0.754 0.012 0.003 0.005 0.049 0.197 0.022 0.088 0.002

We can compare Bonferroni correction, Holm-Bonferroni correction, and FDR

correction for these p-values. We do this by sorting the numbers into ascending
order, as shown in Table 15.1. The final three columns of the table give the
critical threshold values for significance for Bonferroni correction (0.05/10 =
0.005 for all tests), Holm-Bonferroni correction, and FDR correction. A test is
significant for each type of correction if the p-value is lower than the threshold
value in these columns.

For Bonferroni correction, the first two p-values in the list are below the threshold,
and so only these two tests are significant - the other 8 are not. Holm-Bonferroni
correction is a little more forgiving, and finds the first four tests significant. False-
discovery rate correction is the most liberal, with tests 1-6 reaching significance.
This direct comparison shows the potential for these three methods to lead to
different answers for a given family of tests.

Alternatively, we could correct the p-values themselves, as shown in Table

15.2. Now we can compare each adjusted p-value to the uncorrected α level of
0.05. The same tests are significant as before. Note that the adjustment for
the Benjamini-Hochberg method is slightly different from the others. It is a
sequential process, where we start with the largest p-value, and work our way
through the list. Each adjusted p-value is the lesser of the previous adjusted
value, and the current p-value multiplied by an adjustment factor of mj (where
m is the number of tests, and j is the rank-ordered position). The reason we
take the lesser of the two numbers is to prevent the adjustments potentially
corrupting the rank ordering of p-values.
15.8. CLUSTER CORRECTION FOR CONTIGUOUS MEASUREMENTS297

Table 15.2: As for the previous table, but the columns headed Bonf, Holm and
FDR (corr) give the corrected p-values for Bonferroni correction, Holm-Bonferroni
correction, and false-discovery rate correction.

j p Bonf (corr) Holm (corr) FDR (corr)

1 0.002 0.02 0.020 0.015
2 0.003 0.03 0.027 0.015
3 0.005 0.05 0.040 0.015
4 0.006 0.06 0.042 0.015
5 0.012 0.12 0.072 0.024
6 0.022 0.22 0.110 0.037
7 0.049 0.49 0.196 0.070
8 0.088 0.88 0.264 0.110
9 0.197 1.00 0.394 0.219
10 0.754 1.00 0.754 0.754

Correcting the false-discovery rate has become a standard approach in many

fields, particularly genetics where data sets are large enough to estimate the false-
discovery rate directly from the data (e.g. Storey and Tibshirani 2003). There
also are variants that can cope with dependencies (i.e. correlations) between the
different tests (Benjamini and Yekutieli 2001). Overall, it is a suitable choice for
dealing with the multiple comparisons problem without having as dramatic an
impact on statistical power as afforded by other methods.

15.8 Cluster correction for contiguous measure-

ments
For many types of data acquisition, sampling occurs in either space or time, or
sometimes both. For example, cameras and microscopes create a two-dimensional
image sampled at some particular resolution (number of pixels per unit area).
Physiological signals like muscle activity, skin conductance, pupil diameter, or
electromagnetic brain activity, are sampled rapidly on the order of milliseconds.
Finally, movie cameras and MRI scanners produce two or three dimensional
images at multiple points in time.
For all of the above types of data, the sample rate is largely arbitrary, and is
often determined by hardware restrictions. Imagine you used an infra-red video
camera to record the movements of fireflies at night. Perhaps your camera returns
images that are 256x256 pixels 10 times per second, and that you perform some
pixel-wise statistics across time to determine the number of fireflies in a sequence
of images. At that resolution you would make 256 × 256 = 65536 comparisons.
But now imagine we upgraded our camera to one with a higher resolution -
1024x1024 pixels equates to over a million comparisons. Even a relatively liberal
correction for multiple comparisons that adjusts the α level will struggle to
298 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

reach significance for very many image locations because statistical power will
be massively reduced. But there is clearly something wrong here, as a better
camera should give us better results! Cluster statistics are a solution to this
issue because they group significant observations together into clusters, which
become the meaningful unit for significance testing regardless of the sampling
resolution.
It is reasonable to assume in many situations that adjacent sample points will
be correlated to some extent. For example, two adjacent pixels in an image are
likely to be more similar in luminance than two randomly selected pixels (Field
1987), and samples from successive moments in time are also likely to be highly
correlated (the extent of which can be assessed by calculating the autocorrelation
function). Sometimes, data preprocessing methods such as low-pass filtering (see
Chapter 10) or smoothing will deliberately blur adjacent samples together to
reduce noise. This is a problem for traditional multiple comparison corrections,
which generally assume that tests are independent. Again, cluster correction
avoids these issues because it takes into account correlation between adjacent
observations.
The basic idea of cluster correction is that we identify clusters - contiguous regions
of space and/or time where a test statistic is significant at some threshold level.
We then aggregate (add up) the test statistic within each cluster, and compare
these values to an empirically derived null distribution. Those clusters that fall
outside some quantile (i.e. the 95% region) of the null distribution are retained
and considered significant. Various algorithms have been developed along these
same lines, but some of them have recently been shown to suffer inflated Type
I error rates (see Eklund, Nichols, and Knutsson 2016). The method we will
discuss in detail here is a nonparametric cluster correction technique described
by Maris and Oostenveld (2007), that does not suffer these issues. We will
demonstrate its use on an example data set.

15.9 Example of cluster correction for EEG data

Our example data set is a unpublished EEG dataset for a cueing experiment,
collected in my lab during 2017. The reaction times from the same experiment
have been reported by Pirrone et al. (2018). In the experiment 38 participants
were first shown a direction cue (a face with its eyes looking either left or right).
Then a visual stimulus (a sine-wave grating) appeared on either the left or the
right side of the screen. There were 800 trials per participant, with half of
the stimuli appearing randomly on each side of the screen. Brain activity was
recorded from 64 locations on the scalp using electroencephalography (EEG),
and epoched relative to the time of stimulus onset (0 ms in all plots). Each
trial was normalized by subtracting the average of the voltage during the 200 ms
before stimulus onset. The waveforms in Figure 15.4 show group averages from
electrode P8, which is on the right side of the head near the back (see grey point
in upper left insert), adjacent to brain regions that respond to visual stimuli.
15.9. EXAMPLE OF CLUSTER CORRECTION FOR EEG DATA 299

The black trace is the averaged response to stimuli shown on the right side of
the screen, and the blue trace is the response to stimuli shown on the left side of
the screen.

4 Contralateral
Ipsilateral
Voltage (µV)

●
2

−2

−4

−200 0 200 400 600 800 1000

Time (ms)

Figure 15.4: Example EEG data from a cueing experiment. Data are averaged
across 400 trials per condition for each participant, and 38 participants. Data
were bandpass filtered from 0.01 to 30Hz, and taken from electrode P8 (grey
point in upper left insert). Shaded regions show ±1SE across participants.

There are some substantial differences between the two waveforms, which we can
summarise by subtracting the two conditions to calculate a difference waveform.
This is shown by the trace in Figure 15.5. There are differences from shortly
after stimulus onset, that persist throughout the epoch. But are these differences
statistically significant? If we conduct a series of paired t-tests to compare
the conditions, these reveal many time points where the waveforms differ, as
summarised by the dark blue lines at y = -2 in Figure 15.5. However, this
involves running 1200 t-tests (because the data are sampled at 1000 Hz), and
with α = 0.05 we would expect at least 60 of these to be significant even if there
were no true effects. A good example of a definite false positive is at the very
start of the time window (-200 ms). This is before the stimulus was presented,
so it must necessarily be a statistical artefact.

Implementing traditional family-wise error rate correction will substantially

affect our results. The Bonferroni-corrected α level is 0.05/1200 = 0.00004. If
we plot a histogram of uncorrected p-values, it shows that very few of the tests
300 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

4 Difference
p < 0.05
Difference (µV)

Bonferroni
2 FDR

−2

−4

−200 0 200 400 600 800 1000

Time (ms)

Figure 15.5: Difference waveform (ipsilateral - contralateral). Shaded regions

indicate ±1SE across participants. The horizontal lines at y = -2 and below
indicate time windows where the difference waveform differed significantly from
0 (one sample t-test), following different corrections (legend).
15.9. EXAMPLE OF CLUSTER CORRECTION FOR EEG DATA 301

remain significant with such severe correction (see Figure 15.6). The blue lines at
y = -3 in Figure 15.5 show very few significant time points. Even with the more
liberal false discovery rate correction (see light blue lines at y = -4 in Figure
15.5), we still lose quite a lot of our significant time points. Indeed, our largest
cluster splits into two, and our other clusters are reduced in size.
Bonferroni corrected

Uncorrected

0.00001 0.0001 0.001 0.01 0.1 1

p

Figure 15.6: Kernel-density histogram of uncorrected p-values (blue shading).

The solid vertical line indicates the uncorrected threshold of 0.05, and the dashed
line indicates the Bonferroni corrected threshold of 0.00004. Note the log-scaling
of the x-axis.

To perform cluster correction, we need more than just our p-values. We also
need to think about the test statistic, which in this case is a t-value, though it
could equally well be an F-ratio, correlation coefficient, or any other test statistic.
Figure 15.7 shows the trace of t-statistics as a function of time. This looks
broadly similar to the difference waveform, because the t-statistic is the mean
scaled by the standard error at each time point. The thin blue horizontal lines
indicate the critical t-value for a test with 37 degrees of freedom at α = 0.05
(which is around t = ±2). The significant clusters correspond to the time periods
when the t-statistics are outside of these bounds, as shown by the blue lines at y
= -4, and the grey shaded regions.
For the current example, we have five clusters, which have been numbered
consecutively in Figure 15.7. These range from very brief durations (i.e cluster 1)
to a very long one (cluster 4). We next calculate the summed t-value across all
time points within each cluster. This just involves adding up all of the t-values
302 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

2
t−statistic

−2

1 2 3 4 5
−4

−200 0 200 400 600 800 1000

Time (ms)

Figure 15.7: Trace of t-statistics as a function of time. The thin blue lines bound
the critical t-values, and the blue dashed line shows the critical Bonferroni-
corrected t-value. Grey shaded regions indicate periods of significant t-values,
which also correspond to the clusters at y = -4.
15.9. EXAMPLE OF CLUSTER CORRECTION FOR EEG DATA 303

in a single cluster. The summed t-statistics for our five clusters are as follows:
## [1] -7.22676 -73.03767 182.08464 2176.61865 254.14413
We select the largest of our summed t-values (ignoring the sign), which in this
case is cluster 4, with a summed t-statistic of 2177. The raw data from this
cluster will be used to generate a null distribution using a resampling technique
(see also section 8.4.2). The idea here is that we randomly reassign the condition
labels for each participant, and then at each time point within the cluster, we
repeat the t-test and recalculate the summed t-statistic. This is done at least
a thousand times (with different reshufflings of condition labels), to build a
distribution of resampled summed t-statistics. Because of the random condition
assignments, the distribution should have a mean around 0, and some spread
determined by the variability of the data. What we are doing here by randomising
the condition assignments is building up a picture of how we should expect the
data to look if there were no true difference between the groups. We will then
compare this to our observed data to see if there is evidence for real differences.

21 35 4

−1500 −1000 −500 0 500 1000 1500 2000 2500

Summed t−statistic

Figure 15.8: Null distribution of resampled summed t-values. The dashed black
lines give the upper and lower 95% confidence intervals. The numbered vertical
lines correspond to the clusters. Only cluster 4 exceeds the 95% confidence
interval, so only this cluster is considered significant.

The null distribution is shown by the blue shaded curve in Figure 15.8. The
vertical dashed lines indicate the 95% confidence intervals of the null distribution.
These are used as thresholds to compare with each individual cluster’s summed
304 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

t-value. The clusters are indicated by the shorter numbered lines. Clusters 1, 2,
3 and 5 are inside the 95% confidence intervals of the null distribution, so these
are rejected as not reaching statistical significance. Cluster 4 falls outside of the
confidence intervals, and is retained as being significant.

4 Difference
Clusters
Difference (µV)

−2

−4

−200 0 200 400 600 800 1000

Time (ms)

Figure 15.9: Difference waveform (black) showing significant clusters following

cluster correction (blue line).

The final difference waveform with the surviving cluster is shown in Figure
15.9, indicated by the horizontal blue line. Note that the smaller clusters have
been removed, including the spurious one that occurred before the stimulus
was presented. Also, the remaining cluster has not shrunk - its start and end
points are preserved. So, we would conclude that the significant difference in
activity between conditions begins around 200 ms, and extends to around 850
ms. Cluster correction takes a little getting used to, as it involves very different
processes to other multiple comparison corrections. Nevertheless, it has been
robustly validated, and achieves a good balance between maintaining the desired
familywise error rate, and avoiding Type II errors and cluster shrinkage.

15.10 Correcting for multiple comparisons in R

As described previously, the two classical approaches to multiple comparison
correction are to either correct the α level, or adjust the p-values. To make
the procedures clear, we will demonstrate both approaches manually, and also
15.10. CORRECTING FOR MULTIPLE COMPARISONS IN R 305

describe the use of a built-in R function called p.adjust, that can be used to
adjust p-values using various methods. First, let’s look at manual Bonferroni
correction. For a vector of p-values (from Tables 15.1 and 15.2, but these could
be collated across all the tests you have run for a data set), we can identify the
significant ones at α = 0.05 as follows:
pvals

## [1] 0.006 0.754 0.012 0.003 0.005 0.049 0.197 0.022 0.088 0.002
alphalevel <- 0.05
pvals < alphalevel

## [1] TRUE FALSE TRUE TRUE TRUE TRUE FALSE TRUE FALSE TRUE
With no correction, 7 values reach significance. We can Bonferroni correct our α
level by dividing it by the number of tests (which we store in a new data object
called m, defined as the length of the pvals vector):
m <- length(pvals)
alphahat <- alphalevel/m
pvals < alphahat

## [1] FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE TRUE
Only two of our p-values survive this correction. Alternatively, we can apply the
Sidak correction, which finds one additional test to be positive:
alphahat <- 1 - (1 - alphalevel)^(1/m)
pvals < alphahat

## [1] FALSE FALSE FALSE TRUE TRUE FALSE FALSE FALSE FALSE TRUE
The Holm-Bonferroni method is a sequential technique, so we need to rank order
our list of p-values first using the sort function:
pvals <- sort(pvals)
pvals

## [1] 0.002 0.003 0.005 0.006 0.012 0.022 0.049 0.088 0.197 0.754
The thresholds for significance depend on the position in the list, and can be
calculated as follows:
holm <- alphalevel/(m - ((1:m)-1))
print(holm, digits=1)

## [1] 0.005 0.006 0.006 0.007 0.008 0.010 0.013 0.017 0.025 0.050
Finally, we can compare our p-values to the thresholds as follows:
pvals < holm
306 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

## [1] TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE FALSE FALSE

The thresholds for false-discovery rates are calculated as follows:

fdr <- (alphalevel*(1:m))/m
fdr

## [1] 0.005 0.010 0.015 0.020 0.025 0.030 0.035 0.040 0.045 0.050

And again can be compared with the rank ordered p-values:

pvals < fdr

## [1] TRUE TRUE TRUE TRUE TRUE TRUE FALSE FALSE FALSE FALSE

Next, let’s look at how to correct the p-values themselves, instead of adjusting
the α level (recall that these methods produce the same outcome). If we want
to apply the false discovery rate algorithm manually, we can do so using a loop
(see section 2.10) that starts with the second to largest p-value, and works its
way down the list:
fdr <- pvals
for (i in ((m-1):1)){
fdr[i] <- min(pvals[i]*m/i, fdr[i+1])}
print(fdr, digits=2)

## [1] 0.015 0.015 0.015 0.015 0.024 0.037 0.070 0.110 0.219 0.754

However, in R, the p.adjust function will do all of our adjustments for us

automatically. We just need to give it the vector of p-values, and tell it which
correction to apply. For false-discovery rate correction, we get the same results
as above:
p.adjust(pvals,method='fdr')

## [1] 0.01500000 0.01500000 0.01500000 0.01500000 0.02400000 0.03666667

## [7] 0.07000000 0.11000000 0.21888889 0.75400000

We can also apply Bonferroni correction to adjust our p-values as follows:

p.adjust(pvals,method='bonferroni')

## [1] 0.02 0.03 0.05 0.06 0.12 0.22 0.49 0.88 1.00 1.00

Or Holm-Bonferroni correction:
p.adjust(pvals,method='holm')

## [1] 0.020 0.027 0.040 0.042 0.072 0.110 0.196 0.264 0.394 0.754

There are several additional correction methods available that you can read
about in the help file for the p.adjust function.
15.11. IMPLEMENTING CLUSTER CORRECTION IN R 307

15.11 Implementing cluster correction in R

Cluster correction is built into most neuroimaging analysis packages, including
FSL, SPM, AFNI, FieldTrip, Brainstorm and EEGlab. Several R packages also
implement some form of cluster correction algorithm, including the eyetrackingR
and permuco packages. However, such packages typically require data to be in
a specific format, and it is instructive to demonstrate how we can perform the
calculations manually so that they can be applied to different situations and
data types.
Let’s repeat the cluster correction on the ERP data from earlier in the chapter
(see Figure 15.4), which is contained in the file ERPdata.RData available from
the book’s GitHub repository. The file contains a matrix called data, which has
38 rows (participants) and 1200 columns (time points):
N <- nrow(data)
m <- ncol(data)
N

## [1] 38
m

## [1] 1200
Each data point is the voltage difference for a particular participant and time
point. We can therefore perform one-sample t-tests (comparing to 0) at each
time point, and store the results in vectors of p-values and t-statistics as follows:
allp <- NULL
allt <- NULL
for (n in 1:m){
output <- t.test(data[,n]) # run a t-test
allp[n] <- output$p.value # store the p-value
allt[n] <- output$statistic # store the t-value
}

allp[1:5]

## [1] 0.2240673 0.2230249 0.2207536 0.2171818 0.2121924

Next, we need to find all of the significant time points that are contiguous
- i.e. those that occur next to each other. There are doubtless some fancy
sophisticated ways of doing this, but we will use a loop to be as transparent as
possible. The loop checks each p-value in turn. If it is significant and we are not
currently in a cluster, we start a new one, and note the start time for this cluster.
If it is non-significant and we were in the middle of a cluster, we note the end
time for the cluster. After the loop, if we are still in a cluster, we record the end
of the cluster as the end of the data set. At the end of this section of code, we
see that five clusters have been identified, along with the start and end indices:
308 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

clusterstarts <- NULL

clusterends <- NULL
nclusters <- 0 # at the start, we have no clusters
incluster = 0 # at the start, we are not in a cluster
for (n in 1:m){ # loop through all time points
if (allp[n]<0.05){ # if this test is significant
if (incluster==0){ # and if we are not in a cluster
nclusters <- nclusters + 1 # create a new cluster
clusterstarts[nclusters] <- n # note the start time
incluster <- 1 # record that we are now in a cluster
}
}
if (allp[n]>=0.05){ # if this test is not significant
if (incluster==1){ # and if we are in a cluster
clusterends[nclusters] <- n-1 # note the end of cluster
incluster <- 0 # record that we are not in a cluster
}
}
}
# if we end the loop in a cluster, record the end of the cluster
if (incluster>0 & nclusters>0){clusterends[nclusters] <- m}

nclusters

## [1] 5
clusterstarts

## [1] 15 269 334 416 1101

clusterends

## [1] 17 298 385 1058 1189

Now that we have identified our clusters, we can add up the t-values for each
cluster. This uses the vector of t-values that we created earlier, and the start
and end indices of the clusters:
summedt <- NULL
for (n in 1:nclusters){
summedt[n] <- sum(allt[clusterstarts[n]:clusterends[n]])}

summedt

## [1] -7.22676 -73.03767 182.08464 2176.61865 254.14413

To generate the null distribution, we need to choose the largest cluster and note
its start and endpoints:
15.11. IMPLEMENTING CLUSTER CORRECTION IN R 309

biggestcluster <- which(summedt==max(summedt)) # find largest cluster

cstart <- clusterstarts[biggestcluster] # cluster start index
cend <- clusterends[biggestcluster] # cluster end index

Because we are using a one-sample t-test comparing to zero, we can permute

the group labels by randomly inverting the sign of the voltage difference for
half of our participants. This is a bit of a shortcut, and if we were doing a
between-groups comparison we would explicitly switch the group assignments
instead. The following code calculates a t-statistic manually because this is much
faster than calling the t.test function.
# vector of positive and negative weights to be permuted
signlabels <- sign(c(-(1:(N/2)),(1:(N/2))))
nullT <- NULL
for (i in 1:1000){ # repeat lots of times
tsum <- 0 # create a data object to store the summed t-value
randsigns <- sample(signlabels,replace=TRUE) # permute weights
for (n in cstart:cend){ # loop across all cluster indices
tempdata <- data[,n]*randsigns # weight the data
# calculate the t-value and add to the sum
tsum <- tsum + (mean(tempdata)/(sd(tempdata)/sqrt(N)))
}
nullT[i] <- tsum # store summed t-value in a list
}

The null distribution is now stored in the data object nullT, and contains 1000
summed t-statistics. We need to compare the summed t-value for each cluster to
this distribution (see Figure 15.8). We can do this using the quantile function to
identify the lower and upper 95% quantiles (i.e. the points between which 95%
of the values in the distribution lie). We can then compare each of our summed
t-values to see if they fall outside of these limits, and retain any that meet this
criteria:
distlims <- quantile(nullT,c(0.05/2,1-(0.05/2)))
distlims

## 2.5% 97.5%
## -988.5799 952.6620
sigclusts <- c(which(summedt<distlims[1]),which(summedt>distlims[2]))
sigclusts

## [1] 4
Finally, we can retain the start and end indices of any significant clusters (in
this case, just cluster 4):
if (length(sigclusts)>0){
clustout <- matrix(0,nrow=length(sigclusts),ncol=2)
310 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

clustout[,1] <- clusterstarts[sigclusts]

clustout[,2] <- clusterends[sigclusts]}

clustout

## [,1] [,2]
## [1,] 416 1058
So, our significant cluster starts at index 416, and ends at 1058. We can use
these to index the vector of sample times (from -199 to 1000 ms), to find the
true start and end times for the cluster (in ms):
timevals <- -199:1000
timevals[clustout]

## [1] 216 858

Finally, we can plot the results, along with an indication of the significant cluster
(see e.g. Figure 15.9):
plot(timevals,colMeans(data),type='l')
lines(timevals[clustout],c(0,0),col='blue')

We can package up all of the stages for cluster correction into a single function
(a version of this function is included in the FourierStats package that we
encountered in Chapter 11). The function expects four inputs to define the data,
the number of resamples for building the null distribution, and the α levels for
forming clusters and determining significance:
doclustcorr <- function(data,nresamples,clustformthresh,clustthresh){

clustout <- NULL

N <- nrow(data)
m <- ncol(data)

allp <- NULL

allt <- NULL
for (n in 1:m){
output <- t.test(data[,n])
allp[n] <- output$p.value
allt[n] <- output$statistic
}

clusterstarts <- NULL

clusterends <- NULL
nclusters <- 0
incluster = 0
for (n in 1:m){
if (allp[n]<clustformthresh){
15.11. IMPLEMENTING CLUSTER CORRECTION IN R 311

if (incluster==0){
nclusters <- nclusters + 1
clusterstarts[nclusters] <- n
incluster <- 1
}
}
if (allp[n]>=clustformthresh){
if (incluster==1){
clusterends[nclusters] <- n-1
incluster <- 0
}
}
}
if (incluster>0 & nclusters>0){clusterends[nclusters] <- m}

if (nclusters>0){
summedt <- NULL
for (n in 1:nclusters){
summedt[n] <- sum(allt[clusterstarts[n]:clusterends[n]])}

biggestcluster <- which(summedt==max(summedt))

cstart <- clusterstarts[biggestcluster]
cend <- clusterends[biggestcluster]
signlabels <- sign(c(-(1:(N/2)),(1:(N/2))))
nullT <- NULL
for (i in 1:nresamples){
tsum <- 0
randsigns <- sample(signlabels,replace=TRUE)
for (n in cstart:cend){
tempdata <- data[,n]*randsigns
tsum <- tsum + (mean(tempdata)/(sd(tempdata)/sqrt(N)))
}
nullT[i] <- tsum
}
distlims <- quantile(nullT,c(clustthresh/2,1-(clustthresh/2)))
sigclusts <- c(which(summedt<distlims[1]),which(summedt>distlims[2]))

if (length(sigclusts)>0){
clustout <- matrix(0,nrow=length(sigclusts),ncol=2)
clustout[,1] <- clusterstarts[sigclusts]
clustout[,2] <- clusterends[sigclusts]}
}

return(clustout)}
312 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS

We can then perform cluster correction with a single line of code:

doclustcorr(data,1000,0.05,0.05)

## [,1] [,2]
## [1,] 416 1058
The first column contains the cluster start indices, and the second contains the
cluster end indices. If there were multiple significant clusters, there would be
additional rows in the output matrix.
Cluster correction is a very flexible technique, and I hope that by including a
basic implementation, readers will see how to apply the method to their own
data analysis. Although the example here involves a signal that varies in time,
the same principle applies to other dimensions, such as spatial position. For a
two-dimensional image (or three-dimensional volume), determining the adjacency
of points in a cluster is somewhat more challenging.

15.12 Practice questions

1. What is the familywise error rate for 6 tests, with α = 0.01?
A) 0.059
B) 0.265
C) 0.002
D) 0.008
2. What is the Bonferroni corrected α value for 20 tests and an uncorrected
value of α = 0.05?
A) 0.0642
B) 0.0025
C) 0.0049
D) 1.0000
3. For Holm-Bonferroni correction of a family of 12 tests, the adjusted value
of α = 0.05 for the 9th smallest p-value is:
A) 0.025
B) 0.005
C) 0.013
D) 0.004
4. Two consequences of adjusting the α level using Sidak correction are:
A) Power is increased and the Type I error rate is increased
B) Power is reduced and the Type I error rate is reduced
C) Power is increased but the Type I error rate remains constant
D) Power is reduced but the Type I error rate remains constant
5. The false-discovery rate is defined as:
A) The proportion of significant tests that are Type I errors
B) The proportion of all tests that are Type I errors
C) The proportion of significant tests that are Type II errors
D) The proportion of all tests that are Type II errors
15.12. PRACTICE QUESTIONS 313

6. The FDR-adjusted α level for the 50th p-value in a family of 88 tests,

assuming q ∗ = 0.05 is:
A) 0.05/(88-(50-1))
B) 0.05*50/88
C) 0.05*88/50
D) 0.05/88
7. Which of the following would not be a plausible situation for using cluster
analysis to compare between two experimental conditions?
A) Continuous measures of pupil diameter from a video eyetracker
B) Opotogenetic neural activity across a 1 cm2 section of mouse cortex
C) A clinical trial, in which each participant contributes one outcome
measure
D) Differences in brain activity measured with an fMRI scan
8. Nonparametric cluster statistics can be calculated using:
A) The largest p-value within each cluster
B) The larget test statistic within each cluster
C) The summed p-value within each cluster
D) The summed test statistic within each cluster
9. An advantage of cluster correction over other methods of correcting for
multiple comparisons is:
A) Significant clusters retain their original extent
B) The largest cluster is always significant
C) The smallest clusters can always be rejected
D) Clusters can be broken into smaller sub-clusters
10. In cluster correction, a null distribution is generated by:
A) Resampling data from a random subset of observations within the
largest cluster
B) Resampling data from the largest cluster, with random condition
assignments
C) Summing the test statistics across all clusters
D) Summing the test statistics across a random selection of clusters
Answers to all questions are provided in section 20.2.
314 CHAPTER 15. CORRECTING FOR MULTIPLE COMPARISONS
Chapter 16

Signal detection theory

Signal detection theory (Green and Swets 1966; Macmillan and Creelman 1991)
is a method for formalising how humans (and other systems) make decisions
under uncertainty. By uncertainty I mean that either the information being
used to make the decision, or the decision process itself, is noisy - that is it
fluctuates over time. Everyday examples might include trying to understand
what someone is saying on a loud train, spot a faint star in the sky at night, or
work out whether you have added salt to your dinner. The theory was developed
in the 1950s, primarily to characterise the performance of radar operators, but
it has much wider applicability. It is the foundation of modern psychophysical
studies of perception and memory, and in recent years the same concepts have
been applied in artificial intelligence (machine learning) research. This chapter
will cover the basic concepts of signal detection theory, and also discuss common
experimental designs. However, we will begin with an example where we consider
how best to determine the sex of a chicken.

16.1 The curious world of chicken sexing

The global poultry industry is unimaginably huge. Current estimates at time of
writing are that there are in excess of 24 billion chickens on the planet, meaning
they outnumber humans 3:1. But there is an inherent problem in chicken farming
- half of all fertilised eggs hatch into male chicks, and these are useless for egg
and meat production. Chicken farmers are therefore highly motivated to identify
the female hatchlings as soon as possible. Although recent methods have been
developed to determine sex before the eggs hatch (based on tiny emissions of
gas from the shell), these are expensive and not in widespread use. Instead,
the standard technique for many years has been to employ highly trained, and
highly paid, expert chicken sexers.
The job of a chicken sexer is to take trays of day-old hatchlings (see Figure 16.1)

315
316 CHAPTER 16. SIGNAL DETECTION THEORY

Figure 16.1: Recently hatched chicks. Most of them turned out to be male.

and sort them by sex. They do this by inspecting the nether regions (the vent) of
the chick under a bright light for specific features that indicate sex. The precise
features are to do with the shape of a small protruberance called the bead, which
is pointier in female chicks, and rounder in male chicks (see Biederman and
Shiffrar 1987 for diagrams). This is an expert task, and it typically requires many
years of training to reach the levels of performance (>99% accuracy) expected
by the industry. Furthermore, each chick is inspected for only a few seconds, so
professional sexers will process around 1000 chicks per hour (and over a million
each year).

16.2 Hits, misses, false alarms and correct re-

jections
Chicken sexing can be considered a signal detection problem. Suppose that our
signal is whatever visual information indicates that a chick is female (i.e. the
pointiness of the bead). There are then four possible outcomes for every chick
that is inspected.
The first possibility is that we decide that a chick appears to be female, and it
turns out (when it grows up) to actually be female. In signal detection terms, we
consider this a hit. We correctly identified the signal we were looking for. The
second possibility is that we decide that a chick is not female, and it turns out
(when it grows up) to indeed be male as we predicted. This is known as a correct
rejection, because we are saying the ‘female’ signal is not there, and it isn’t. It is
very common to calculate an overall (percentage) accuracy in detection tasks,
and this is taken as the sum of hits and correct rejections, divided by the total
number of trials (chicks). However, it is also important to think about instances
when we are incorrect, as there are also two different types of these.
One way we can be wrong is by deciding that a chick is female, when it actually
16.2. HITS, MISSES, FALSE ALARMS AND CORRECT REJECTIONS 317

turns out to be male. This is referred to as a false alarm, because we have

incorrectly decided that the ‘female’ signal is present. The converse situation
is where we decide that a chick is male, but it turns out to be female. This is
known as a miss (because we have missed the ‘female’ signal).
Thinking about these four possible outcomes is important, because it gives us
information about how an individual is going about the task. For example,
imagine three student chicken sexers: Aya, Bia and Che. Each student is given a
tray of 100 chicks to sex (half of which are female), and all three have the same
overall accuracy: 80/100 chicks are correctly sexed, so their percentage accuracy
is 80%. Yet as we will see, the pattern of responses is very different across the
three students.
Aya is a good all-rounder. She correctly identifies 40 of the female chicks (hits)
and 40 of the male chicks (correct rejections). She incorrectly identifies 10 of the
female chicks as being male (misses), and 10 of the male chicks as being female
(false alarms). We can summarise her performance with the grid shown on the
left of Figure 16.2.

Assigned sex Assigned sex Assigned sex

Aya F M Total Bia F M Total Che F M Total

True sex

True sex

True sex

F 40 10 50 F 49 1 50 F 35 15 50
Hits Misses Hits Misses Hits Misses

M 10 40 50 M 19 31 50 M 5 45 50
FAs CRs FAs CRs FAs CRs

Total 50 50 80% Total 68 32 80% Total 40 60 80%

correct correct correct

Figure 16.2: Grids showing the accuracy of three student chicken sexers. See
text for details. FAs: false alarms, CRs: correct rejections.

Bia has a somewhat different strategy. She is overall more likely to assign a
chick as being female. Out of all her chicks, she assigns 68 as being female, and
the remaining 32 as being male. In other words she has a bias towards assigning
chicks as being female. This strategy works well for spotting the female chicks,
as it means means she correctly identifies nearly all of them (49 hits, 1 miss).
However, because of her bias she incorrectly assigns 19 of the male chicks to
the female category (19 false alarms) and correctly identifies the rest as male
(31 correct rejections). Bia’s performance is summarised in the middle grid of
Figure 16.2.
Che has a bias in the opposite direction to Bia. He is more likely to assign
a chick as being male (perhaps to avoid getting fined for letting male chicks
through as female). This strategy means that he misses 15 of the female chicks,
318 CHAPTER 16. SIGNAL DETECTION THEORY

incorrectly assigning them as male. However he correctly identifies most of the

male chicks. His results are shown in the final grid of Figure 16.2.
What is striking is that all three students have the same percentage accuracy,
even though both Bia and Che have substantial biases. It would be very useful
if we could quantify these biases, and if we could separate them from the true
sensitivity at doing the task. This is precisely what signal detection theory aims
to do, as we will describe in the following section.

16.3 Internal responses and decision criteria

Signal detection theory makes some assumptions about what is happening inside
the receiver, in other words the mind of the person making the decision. The first
assumption is that the sensory information used to make the decision is mapped
onto a monotonic scale termed the internal response. If we are comparing male
and female chicks, once we have learned to do the task, we expect that some
group of neurons in our brain might respond more to female chicks and less to
male chicks (on average). We could then use the magnitude of this response to
make a decision. The same idea applies if we were trying to detect a weak signal
- for example a faint light, or a quiet sound. When the signal is present, the
internal response should be larger on average than when it is absent. It is not
necessary to know precisely which population of neurons are doing this task, or
where they are in the brain.
The second assumption is that incoming information is processed imperfectly, and
is subject to noise. This noise could come from a number of sources, including
external noise associated with the stimulus itself (perhaps the chick moves whilst
it is being inspected), and internal noise from the inherent variability of biological
nervous systems. A consequence of noise is that repeated presentations of even
the same stimulus will produce different amounts of response.
We can characterise the internal response to the two stimulus categories (here
male and female) with two probability distributions (see examples in Figure 16.3).
One distribution tells us how likely it is that a male chick will produce a given
response, and the other tells us how likely it is that a female chick will produce a
given response. The mid-point of the distribution is the average internal response
for that category, and the spread is determined by the noise. Different example
stimuli (i.e. chicks, or trials in some experiment), can be thought of as drawing
samples from these underlying distributions.
Let’s start off by thinking about sexing adult birds. This should be very easy,
as the males have very prominent wattles and crow loudly, and the females
lay eggs. The distributions in Figure 16.3a are very far apart, and random
samples (individual birds) from the two distributions (points) are very easy to
categorise correctly. In this situation, we would expect perfect accuracy - all
animals would be either hits (correctly identified females) or correct rejections
(correctly identified males).
16.3. INTERNAL RESPONSES AND DECISION CRITERIA 319

Adults Eggs Chicks

(a) (b) (c)

● ●
●● ● ● ● ●
●
● ●● ● ● ● ●
● ●● ● ● ● ● ● ●
● ● ●● ● ●
● ●● ● ●
● ● ●
● ● ●● ● ● ●
● ● ● ● ● ●
● ● ●
● ●● ●● ●● ● ●● ●
●
●● ●● ● ●
●● ● ● ●
● ●
●● ● ● ●● ●● ● ●● ●
● ●● ●

Male Female Male Female Male Female

Internal response

Figure 16.3: Example distributions of internal responses (curves), with data

points sampled from each distribution (points). In panel (a), the distributions are
far apart and narrow, consistent with an easy task. In panel (b), the distributions
overlap entirely, and distinguishing male from female animals is impossible. In
panel (c), the distributions overlap somewhat, indicating that there will be some
animals that are misclassified.

Next let’s think about the opposite extreme. Imagine we have to categorise
eggs as male and female before they have hatched, just by looking at them.
Obviously this is impossible, and so the two distributions will completely overlap
(see Figure 16.3b), with any given egg being equally likely to be male or female.
In this situation, hits, misses, false alarms and correct rejections are all equally
likely (each consisting of 25% of trials), and accuracy will be at chance (50%
correct).

A more interesting situation occurs for day-old chicks, where sexing is possible
but it is challenging. The probability distributions in Figure 16.3c overlap
somewhat. This means that some individual chicks will be difficult to categorise,
and may produce misses or false alarms, but accuracy should be somewhere
above chance.

So what is the best strategy for categorising male and female chicks? The
example student chicken sexers described above illustrate that there are different
criteria we can adopt when doing this task. Recall that Bia is more likely to
assign chicks as being female, Che more likely to assign them as male, and Aya
somewhere in between. Figure 16.4 shows us the overlapping distributions in
a bit more detail with an extra feature - the vertical dotted line indicates the
criterion. Any individual observations that fall to the left of this line will be
classified as male, and any falling to the right as female. Of the chicks classified
as female, some will be hits (chicks that really are female) and some will be false
alarms (chicks that are actually male). Of the chicks classified as male, some
will be correct rejections (chicks that really are male) and some will be misses
(chicks that are actually female).
320 CHAPTER 16. SIGNAL DETECTION THEORY

Sensitivity (d')

Criterion

Correct Hits
Rejections

Misses False
Alarms
Male Female
Internal response

Figure 16.4: Internal response distributions with additional features. The vertical
dotted line indicates the criterion - samples to the left of this line will be classified
as male, and those to the right as female.
16.4. CALCULATING SENSITIVITY AND BIAS 321

A key idea in signal detection theory is that we can move our criterion around
if we want to. So, we could instruct our students that misses are a very bad
thing (because potentially productive female chicks would be missed). This
would make them shift their criterion to the left to avoid misses (and increase
hits). However the number of false alarms would also increase (and correct
rejections would decrease) - in other words, they would be like Bia, who has
a bias towards saying chicks are female. Alternatively we could instruct the
students that false alarms are a bigger problem (because it wastes money raising
male chicks), so that they shift their criterion to the right. Then false alarms
(and hits) would decrease, but misses (and correct rejections) would increase. On
the other hand, the sensitivity at the task is determined by the overlap of the two
internal response distributions, and is not affected by changing the criterion. So
this is generally assumed to be fixed for a given observer and situation (though
learning and training are possible).

16.4 Calculating sensitivity and bias

We can fully characterise the pattern of performance of our student chicken
sexers (and indeed any rater in a similar situation) by calculating two properties:
sensitivity and bias. The sensitivity is the amount of overlap between the two
distributions in Figure 16.4. This depends on both the offset of the means, and
the spread of the distributions. We can standardise this by taking the difference
between the means, and dividing by the standard deviation. This results in
a statistic called d′ (pronounced d-prime). Note that d′ is conceptually very
similar to measures of effect size such as Cohen’s d (which is calculated as the
difference between group means divided by the pooled standard deviation, see
Chapter 5). If the distributions are narrow and far apart (e.g. Figure 16.3a), this
will result in a large d′ value. If they are wide and close together (and therefore
overlap a lot), we will get a small d′ value. Because d′ is based on the underlying
internal response, it is referred to as the sensitivity index.
The other component is the bias (C ). This depends on the placement of the
criterion, and can be altered by instruction. As described in the previous section,
we can often choose a particularly strict or lax criterion to achieve a particular
outcome. Sometimes rewards or penalties can be used to shift the criterion in a
specific direction.
How can we calculate d′ and C if we just have information about hits, misses,
false alarms and correct rejections? The first thing we need to do is convert our
trial counts from each category into proportions of the total number of trials
where the signal was either present or absent (we can also think of these values
as probabilities). Let’s do this for Aya’s ratings in the first grid in Figure 16.2 -
these are shown in Figure 16.5.
Something we notice in the proportions grid is that the proportions within each
category of true sex (i.e. the rows) must sum to one. If we have a fixed number of
322 CHAPTER 16. SIGNAL DETECTION THEORY

Assigned sex

Aya F M Total
True sex

F 0.8 0.2 1
P(Hit) P(Miss)

M 0.2 0.8 1
P(FA) P(CR)

Figure 16.5: Detection performance for chick sexing, converted to proportions.

16.4. CALCULATING SENSITIVITY AND BIAS 323

chicks that are truly female, each of them must either be a hit or a miss. Similarly,
for a fixed number of truly male chicks, each of them must be classified as either
a correct rejection or a false alarm. This means that P (M iss) = 1 − P (Hit) and
also P (CR) = 1 − P (F A). Because of these interdependencies, once the data
are converted to proportions, we actually only need to know the hit rate and
the false alarm rate to fully characterise performance.

1
Cumulative probability

0.8 Hits
0.6

0.4

0.2
FA ●
0
−5 −4 −3 −2 −1 0 1 2 3 4 5
z (sd units)

Figure 16.6: Cumulative normal distribution, mapping standardized units (z-

scores) to probabilities. This function (and its inverse) are used to convert
between proportions and z-scores.

In order to calculate d′ , we next need to convert the proportions to z-scores using

an inverse cumulative normal distribution (we will demonstrate how to do this
in R later). Recall that a z-score is a standardised score that tells you how far
an observation is from the group mean, expressed in multiples of the standard
deviation. These are unitless measures, which can easily be compared across
datasets. There is a straightforward monotonic transform between probabilities
and z-score (standard deviation) units as shown in Figure 16.6. For Aya, the
probability for hits (0.8) converts to z(Hit) = 0.84 and the probability for false
alarms (0.2) converts to z(F A) = −0.84. The equation for d′ is the difference
between these values:

d′ = z(Hit) − z(F A) = 0.84 − (−0.84) = 1.68

We can also convert the performance of Bia and Che (our biased chicken sexers)
324 CHAPTER 16. SIGNAL DETECTION THEORY

to z-scores and d′ values in the same way. For Bia, her hit rate is 49/50 = 0.98,
and her false alarm rate is 19/50 = 0.38. This works out as:

d′ = z(Hit) − z(F A) = z(0.98) − z(0.38) = 2.05 − (−0.31) = 2.36

So despite being heavily biased, Bia is actually more sensitive (and therefore
potentially better at the task) than Aya. Che’s hit rate is 35/50 = 0.7 and his
false alarm rate is 5/50 = 0.1. His sensitivity can be calculated as:

d′ = z(Hit) − z(F A) = z(0.7) − z(0.1) = 0.52 − (−1.28) = 1.81

This means that Che is slightly more sensitive than Aya, but less sensitive than
Bia.
We can also quantify the bias of each rater by calculating the criterion. The
equation for the criterion is:

C = −(z(Hit) + z(F A))/2

For Aya (who has no bias), this reassuringly works out to 0:

C = −(z(Hit) + z(F A))/2 = −(0.84 + (−0.84))/2 = 0

For Bia, who has a substantial bias, this works out as:

C = −(2.05 + (−0.31))/2 = −0.87

And for Che, whose bias is a bit smaller (and in the opposite direction), this
works out as:

C = −(0.52 + (−1.28))/2 = 0.38

In many contexts, signal detection theory is used to estimate sensitivity (d′ ) for
a task, by removing the confounding effect of bias (C ). For our chicken sexing
example, we might want to decide which student to offer a job to. Although
they all have the same accuracy, Bia was the most sensitive, and so she might
be our best choice - after all, we could ask her to adjust her bias if necessary,
but her underlying ability was the strongest. Without dividing accuracy data
into sensitivity and bias, we would have wrongly judged all three students to
be equally good at the task. Having a good understanding of signal detection
theory might help to avoid analogous errors in your own research.
Estimating sensitivity and bias is important in many lab-based studies of human
perceptual and memory abilities, and more recently has been applied to assess the
ability of artificial intelligence systems to judge the content of images. Sometimes
16.5. RADIOLOGY EXAMPLE: RATING SCALES AND ROC CURVES325

an unbiased estimate of sensitivity can be of substantial practical importance:

examples include a football (soccer) referee’s ability to detect a handball, an
airport security officer’s ability to detect a weapon in a luggage x-ray, a passport
control officer’s ability to detect a fake passport, or disease diagnosis in the
context of medical screening. The following section describes an example from
the literature on cancer diagnosis.

16.5 Radiology example: rating scales and ROC

curves
Radiologists must interpret diagnostic scans (usually x-ray or CT scans) to
determine whether patients have cancer. In signal detection terms, the images
will be in two categories: cancer (signal) present and cancer (signal) absent. The
radiologists’ decisions will fall into four categories:
• cancer diagnosed when cancer present (hits)
• cancer diagnosed when cancer absent (false alarms)
• cancer not diagnosed when cancer present (misses)
• cancer not diagnosed when cancer absent (correct rejections)
Exactly the same approach can be taken to analyse these data as described
above. However in this context, the consequences of bias (or criterion) are worth
further consideration. If a radiologist is very liberal, they might diagnose cancer
even when the evidence is quite marginal. This will catch most real cancer cases,
but also result in a lot of false alarms. These will cause anxiety to patients, and
use additional resources running further tests (such as biopsies) to confirm the
diagnosis. If a radiologist is very conservative, they might diagnose cancer only
when the evidence is overwhelming. This will miss some true cancer cases, which
could be fatal if not caught early. On the other hand, there will be few false
alarms, so resources will not be wasted running additional tests.
A recent study (Evans et al. 2016) tested 49 radiologists in a series of controlled
experiments. The stimuli were mammograms (breast images) where the eventual
true diagnosis of the patient was known (i.e. the experimenters knew if the patient
really did have cancer or not based on follow up tests). Although in clinical
practice radiologists can inspect images for as long as they like, in the experiment
the images were presented very briefly for 500 ms so that all participants had the
same inspection time. Also, to provide more information about each radiologist’s
ratings, instead of a binary classification (cancer vs no cancer), participants
provided ratings on a 101-point scale, from 0 (clearly abnormal) to 100 (clearly
normal).
This finer-grained information allows a more detailed analysis because we can
place criteria at multiple locations on the rating scale. For example, if we place
a criterion at a rating of 10, and count all images with ratings below this value
as cancer cases and all those above as clear, this will produce very few hits and
326 CHAPTER 16. SIGNAL DETECTION THEORY

very few false alarms. On the other hand, if we place a criterion at a rating of
90, this will produce many hits, but also many false alarms. A useful way of
presenting such data is to plot the hit rate against the false alarm rate for a
range of criterion levels, as shown in Figure 16.7a.

1 1
(a) ● ● (b)
●
0.8 ● 0.8

●
Hit rate

Hit rate
0.6 0.6
●
0.4 0.4
● Criterion d' = 0.5
● ● Liberal d' = 1
0.2 ● Conservative 0.2 d' = 2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
False alarm rate False alarm rate

Figure 16.7: Example ROC curves. Panel (a) shows hit rates against false alarm
rates for a range of criterion levels, with sensitivity fixed at d’ = 1. Panel (b)
shows ROC curves for three different levels of d’.

This plotting convention is called a receiver operating characteristic (ROC) plot.

Note that only the criterion changes across the points - the sensitivity (d′ )
remains constant. So, the ROC curve that we produce should correspond to
a fixed value of d′ . Figure 16.7b shows ROC curves for three different levels
of d′ . Notice how the highest sensitivity (d′ = 2, dashed line) stretches closer
to the top left corner, which represents perfect performance (i.e. a hit rate of
1 and a false alarm rate of 0). The lowest sensitivity (d′ = 0.5, dotted line) is
much closer to the diagonal, which represents chance performance (hits = false
alarms). Also note that any single point on the graph corresponds to a unique
combination of hit and false alarm rates, which maps onto a single estimate of
d′ using the equations we encountered above.
In the Evans et al. (2016) study, this method was used to estimate sensitivity
for cancer diagnosis with various stimuli. When comparing breast images
containing tumours with those that did not, trained radiologists had a sensitivity
around d′ = 1.2 (recall that presentation was very brief). Most strikingly, when
comparing scans from the opposite breast of cancer patients (i.e. the one without
a tumour) to breasts from healthy controls, it was still possible to identify the
cancer patients with a sensitivity of d′ = 0.6. This suggests that global changes
in breast tissue provide a distributed signal that can potentially indicate patients
likely to have (or to develop) cancer, even in the absence of visible tumours.
16.6. REMOVING BIAS USING FORCED CHOICE PARADIGMS 327

ROC curves can be constructed in other ways besides using a rating scale (and
also by using rating scales with fewer levels), and applied in many contexts besides
cancer diagnosis. In some lab experiments, performance can be aggregated across
many participants who all naturally have different levels of bias, mapping out an
ROC curve for the population. In other studies, participants can be instructed
to change their criterion in different blocks of the experiment (Morgan et al.
2012). This can be done explicitly, by asking participants to be more liberal or
more conservative (e.g. “Only respond ‘yes’ if you are really sure”), or implicitly
by implementing a reward and penalty structure (this is popular in behavioural
economics studies). For example, rewarding hits will encourage a more liberal
criterion, whereas penalising false alarms will encourage a more conservative
criterion. The data from blocks with different instructions should map out a
single ROC curve.

16.6 Removing bias using forced choice paradigms

Signal detection theory gives us a principled way to estimate both sensitivity
and bias in a class of experimental methods referred to as yes/no paradigms. In
these experiments, a single stimulus is presented on each trial (for the examples
above, this was a single chick, or a single breast image). But there is another
class of method that aims to avoid the issue of bias completely. This is called a
forced choice paradigm. The most common version is the two-alternative forced
choice (2AFC) paradigm.

In a 2AFC experiment, two stimuli will be presented (either one after the other,
or simultaneously but in different locations). One will always come from the
null distribution (i.e. a male chick, or a healthy breast image), the other will
always come from the target distribution (i.e. a female chick, or a cancerous
breast image). The participant’s task is to indicate which of the two stimuli
contains the target. This seems like a very minor difference to yes/no designs,
but it has an important consequence. In terms of the internal response, we now
have two discrete values to compare (see Figure 16.8). The optimal strategy is
always to pick the stimulus that produces the largest internal response, as that
is more likely to be the target. There is no need for us to have a criterion, and
no opportunity for bias. Forced choice paradigms are therefore often referred to
as bias free.

Forced choice methods are not limited to having two alternatives, but the number
of alternatives will determine the baseline (chance) performance level. For 2AFC,
the guess rate is 0.5 (50% correct). For 3AFC, the guess rate is 0.33 (33%
correct). In general, the guess rate is 1/m, where m is the number of alternatives
(see also Section 14.4). Depending on the other constraints of an experiment,
there may be good reasons for choosing designs where m > 2. For example,
3AFC can often be explained to participants as an odd-one out paradigm, where
they choose the stimulus that appears different from the other two.
328 CHAPTER 16. SIGNAL DETECTION THEORY

Sensitivity (d')

Null Target
●

Internal response

Figure 16.8: Example internal response distributions (curves) and samples

(points) for a 2AFC paradigm. The black solid curve is for the null stimulus
(e.g. target absent), and the blue dashed curve is for the target stimulus (e.g. tar-
get present). There is no criterion; the optimal strategy is to select the larger of
the two responses, as this is more likely to arise from the target distribution.
16.7. MANIPULATING STIMULUS INTENSITY 329

Because bias is irrelevant in mAFC designs, there is a monotonic mapping

between accuracy and sensitivity. This means we can convert directly between
the proportion of correct trials and d′ . This is done by rescaling the accuracy,
and then mapping it to the cumulative Gaussian distribution illustrated in Figure
16.6. For 2AFC, we calculate d′ as:

√
d′ = φ−1 (Pc ) 2

where Pc is the proportion of correct responses, and φ−1 indicates the inverse of
the cumulative Gaussian function shown in Figure 16.6.
Note that it is very common to see single interval experiments described as being
2AFC when they are actually yes/no. The confusion stems from the two possible
response options that a participant is ‘forced’ to choose between. For example,
being shown a line stimulus and having to indicate if it is tilted leftwards or
rightwards. However, this is really a yes/no paradigm in disguise, because there
is only one stimulus being presented. For true 2AFC there must always be two
distinct stimuli, and the participant selects one of them. If bias is negligibly
small then accuracy and d′ are closely related even for yes/no paradigms, so this
distinction becomes largely semantic, but without estimating bias for a yes/no
task this might be difficult to argue convincingly.

16.7 Manipulating stimulus intensity

The examples we have encountered so far in this chapter have involved a fixed
signal level, and therefore a single corresponding level of sensitivity (d′ ). But it
is quite common in empirical studies to manipulate properties of the signal to
measure performance at a range of stimulus intensities. Consider an experiment
designed to estimate a detection threshold - that is, the weakest signal strength
that can be reliably detected at a given level of performance. We would typically
present a range of signal intensity levels on different trials, measuring accuracy
at each one. Performance (accuracy or d′ ) can then be plotted as a function of
signal strength. Figure 16.9 illustrates the underlying distributions of internal
response for each signal level (panel a), and the corresponding proportion correct
scores (panel b) that form a psychometric function.
In such an experiment, we are more interested in finding out the signal level that
corresponds to some particular threshold, such as 82% correct (the midpoint of
a cumulative Weibull function), or a sensitivity of d′ = 1. This is traditionally
estimated by fitting a sigmoidal function to our discrete data points (such as
the curve Figure 16.9b; see Chapter 9 for details of how nonlinear models can
be fitted to data), and interpolating to find the required performance level
(indicated by the dashed lines). Thresholds can then be used as a dependent
variable and compared across individuals, or between different experimental
conditions (e.g. for different types of stimuli). These paradigms are widely used
in psychophysical studies in both humans and animals, and form the basis of our
330 CHAPTER 16. SIGNAL DETECTION THEORY

Probability (a)

0
Internal response

1
(b) ●
●
●
Proportion correct

0.8
●

0.6
●
●
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Signal strength

Figure 16.9: Probability distributions of internal response for different signal

strengths (top panel) and proportion correct scores for a simulated 2AFC ex-
periment (bottom panel). The black distribution is the null interval in each
condition, and d’ scores for the six stimulus levels range from 0 to 2.5 in steps of
0.5. The curve in the lower panel is a fitted Weibull function.
16.8. TYPE II SIGNAL DETECTION THEORY (METACOGNITION) 331

contemporary understanding of sensation and perception (Kingdom and Prins

2010).

16.8 Type II signal detection theory (metacog-

nition)
The preceding sections describe what is known as Type I signal detection theory,
and is what most people think of when signal detection theory is mentioned -
the quantification of performance in detection tasks. However, another aspect
to the theory has gained widespread interest in recent years. This is known as
Type II signal detection theory and involves having participants make confidence
judgements about their own performance. This is a form of metacognition, or
knowledge about our own cognitive processes. Broadly speaking, the same types
of calculations can be applied to Type I and Type II measures, though the
equivalent sensitivity measure is called meta-d′ .
The additional information from Type II measures can be used to test theories
and models about the source of different effects. The most famous example is
the case of blindsight, a curious phenomenon experienced by patients with brain
damage to specific parts of their visual cortex (Weiskrantz 2007). These patients
do not experience conscious visual perception in the affected regions of space.
If you ask them what is there, they cannot tell you. However, they are able to
respond to visual information in the affected region, and their accuracy in some
tasks is above chance (though they report that they are guessing). In terms
of signal detection theory, Type I performance is above chance, but Type II
performance is not.

16.9 Practical problems with ceiling perfor-

mance
One common practical issue with calculating d′ occurs when participants get
every trial correct in one or other category of trials in a yes/no design (i.e. the
hit rate or correct rejection rate is 1), or every trial is correct in an mAFC
experiment. This can sometimes occur by chance for easy tasks with finite
numbers of trials, but it implies that d′ = ∞. This is implausible in reality, and
the true d′ value is most likely lower. Possible solutions include adding one ‘fake’
incorrect trial for each participant, or establishing a maximum possible d′ score,
such as d′ = 8.
A related issue that can interfere with fitting psychometric functions (as in
Figure 16.9b) is when performance at high stimulus intensity levels asymptotes
at a proportion correct slightly below 1. This happens because of lapse trials,
where the participant gets the trial wrong even though it is very easy, perhaps
because of inattention, tiredness, or unpredictable events such as blinks or
332 CHAPTER 16. SIGNAL DETECTION THEORY

sneezes. Lapses can skew psychometric function fits by making them much
shallower than they would usually be (to capture the reduced performance at
high stimulus intensities). This can make the threshold estimate incorrect. Many
software tools for fitting psychometric functions now give the option of setting
an upper asymptote slightly below 1 to counteract this problem (see Section 4.2
of Kingdom and Prins 2010).

16.10 Calculating d′ in R
For yes/no tasks, we can calculate z-scores from proportions directly using the
qnorm function to generate quantiles from the normal distribution. This is the
curve shown in Figure 16.6. Here are the z-scores for a range of proportion
values:
proportions <- seq(0.1,0.9,0.1)
proportions

## [1] 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
zscores <- qnorm(proportions)
round(zscores,digits=2)

## [1] -1.28 -0.84 -0.52 -0.25 0.00 0.25 0.52 0.84 1.28
We can then calculate d′ manually by subtracting z-scores for hits and false
alarms. For a hit rate of 0.95 and a false alarm rate of 0.1, d′ is calculated as
follows:
hits <- 0.95
FA <- 0.1
dprime <- qnorm(hits) - qnorm(FA)
dprime

## [1] 2.926405
Similarly, bias (C ) is calculated using the same values:
C <- -(qnorm(hits) + qnorm(FA))/2
C

## [1] -0.181651
For 2AFC, the appropriate conversion is:
dprime2afc <- qnorm(proportions)*sqrt(2)
round(dprime2afc,digits=2) # round to 2 decimal places

## [1] -1.81 -1.19 -0.74 -0.36 0.00 0.36 0.74 1.19 1.81
Notice that the proportions below chance performance (0.5) produce negative
d-prime scores. Most participants will tend to perform above chance in 2AFC
16.11. PRACTICE QUESTIONS 333

experiments, so this should occur only rarely. Because there is only one proportion
involved in this calculation,
√ we can convert in the opposite direction as well, by
scaling the d′ scores by 2 and using the pnorm function:
proportionsconverted <- pnorm(dprime2afc/sqrt(2))
proportionsconverted

## [1] 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
For mAFC tasks (where m > 2), the calculations become more complex (see
section 6.3.2 of Kingdom and Prins (2010) for a detailed explanation). However,
there is a function called dprime.mAFC available in the psyphy package. It will
convert proportion correct scores to d′ , provided m is also specified. However
it can only convert one score at a time (it does not work with vectors), so to
convert multiple values we can either use a loop, or use the sapply function (a
built in function that applies a function to all values in a vector) as follows:
library(psyphy)
dprime3afc <- sapply(proportions,dprime.mAFC,3)
dprime4afc <- sapply(proportions,dprime.mAFC,4)
dprime5afc <- sapply(proportions,dprime.mAFC,5)

If we plot the d′ scores (in Figure 16.10), you can see (from the vertical dotted
lines) that chance performance at the baseline (of 1/m) always corresponds to
d′ = 0.
These tools will allow a signal detection analysis to be applied to data from
yes/no and forced choice experiments. Similar tools exist in other programming
languages and toolboxes, for example the Palamedes toolbox provides signal
detection functions, along with tools for fitting psychometric functions, in the
Matlab environment. In R, the quickpsy package (Linares and López-Moliner
2016) can be used to fit psychometric functions. For further reading on signal
detection theory, I recommend Kingdom and Prins (2010), Macmillan and
Creelman (1991), and (for an excellent historical account) Wixted (2020).

16.11 Practice questions

1. In a yes/no task, responding ‘no’ when the signal is absent would be a:
A) Hit
B) Miss
C) False alarm
D) Correct rejection
2. In a yes/no task, responding ‘no’ when the signal is present would be a:
A) Hit
B) Miss
C) False alarm
D) Correct rejection
334 CHAPTER 16. SIGNAL DETECTION THEORY

1
d'

m=2
m=3
−1

m=4
m=5

0.2 0.4 0.6 0.8

Proportion correct

Figure 16.10: Mapping from proportion correct to d’ for different forced choice
designs.

3. In a yes/no task with 28 hits and 15 misses, the hit rate (as a proportion)
is:
A) 0.65
B) 0.35
C) 0.54
D) 0.45
4. The sensitivity index in signal detection theory is called:
A) Cohen’s d
B) The criterion
C) d′
D) z
5. For a yes/no task, d’ is calculated as:
A) z(Hit) - z(Miss)
B) z(Hit) - z(FA)
C) z(Hit) - z(CR)
D) z(FA) - z(Hit)
6. Using the graph in Figure 16.6, what is the approximate z-score for a
probability of 0.9?
A) 0.5
B) 1.3
C) 3.1
16.11. PRACTICE QUESTIONS 335

D) 4.5
7. By collecting data on a task at a range of criterion levels, we can plot:
A) A psychometric function
B) The change in d′ with bias
C) A normal distribution
D) An ROC curve
8. An experiment is described as being 2AFC if there are:
A) Two possible responses
B) Two different stimuli to choose between
C) Two trials, and the participant responds to both of them
D) Additional responses to indicate confidence
9. Use the qnorm function in R to calculate d’ for a yes/no experiment with
a hit rate of 0.95 and a correct rejection rate of 0.8.
A) d′ = 0.80
B) d′ = 2.32
C) d′ = 2.49
D) d′ = 3.28
10. Use the qnorm function in R to calculate d’ for a 2AFC experiment where
the participant got 76% of trials correct.
A) d′ = 1.00
B) d′ = 0.95
C) d′ = 0.77
D) d′ = 0.71
Answers to all questions are provided in section 20.2.
336 CHAPTER 16. SIGNAL DETECTION THEORY
Chapter 17

Bayesian statistics

This chapter will introduce an approach to statistics that is quite different from
that traditionally taught. Statistical tests that are evaluated for significance
using a p-value are known as Frequentist statistics. This category probably
includes most of the tests you are familiar with - things like t-tests, ANOVA,
correlation and regression. Bayesian methods involve quite different assumptions
and different procedures, and there are situations where they could be a sensible
choice in data analysis. But before we go into detail about how they work, it is
worth discussing some of the basic assumptions of, and problems with, frequentist
methods. It might seem counterintuitive to start a chapter on Bayesian statistics
by talking about frequentist methods, but these are issues that people who use
statistics routinely have often not thought about in depth.

17.1 The philosophy of frequentist statistics

When we run a frequentist test, we are asking whether the data we have collected
provides evidence to support our experimental hypothesis. Depending on the
study design, this will often be of the form that the group mean differs from
some stated value, or that the means of multiple groups or conditions differ from
each other. The null hypothesis is that there are no differences. If the data
provide evidence for a difference, we say that our experimental hypothesis is
supported by the data, and we reject the null hypothesis. On the other hand, if
the data do not provide evidence for a difference we have failed to reject the null
hypothesis.
Failing to reject the null is not the same as accepting the null hypothesis, and
we cannot conclude with confidence that there is no true difference. This might
be the case, but it could also be that our experiment was underpowered to detect
the effect we are interested in (termed a Type II error). This might happen
if either the true effect size or the sample size were too small (see Chapter 5).

337
338 CHAPTER 17. BAYESIAN STATISTICS

With frequentist methods, we generally cannot distinguish whether there really

is no difference, or if our study design lacks power.
The p-value is usually reported alongside a test statistic as a way of summarising
the outcome. Formally, the p-value is a statement about the probability that
we would have observed our data if the experimental hypothesis were untrue.
A small p-value suggests that it is very unlikely that we would have observed
the data were the hypothesis untrue, and so this allows us to reject the null
hypothesis. A large p-value means we have failed to reject the null but, as
mentioned above, we do not obtain an estimate of how likely the null hypothesis
is to be true.

17.2 An unfortunate consequence of the fixed

Type I error rate
A rather curious feature of frequentist methods is that the Type I error rate is
fixed, no matter the sample size. The Type I error rate is the false positive rate
- the probability of reporting an effect as significant when in fact there is no true
effect. It is determined by the α level, the threshold value below which a p-value
is considered significant. In many disciplines, this is fixed at α = 0.05, meaning
there is a one in 20 (or 1/α) chance of getting an apparently significant effect
when there is no true effect. Other disciplines use a more stringent threshold
- for example some research in high energy physics uses the 5σ limit, which
corresponds to a threshold of α = 0.0000003, and a Type I error rate of about 1
in 3.3 million.
The curious thing about the Type I error rate is that it does not change at all if
we collect more data. Whether we have ten participants or ten thousand, there
is still the same probability that we will get a false positive when there is no
true effect. The same is not true of situations where there is a real effect. Here,
our statistical power will increase with sample size, and so we will be more likely
to detect a (true) effect the more observations we make. We demonstrate this
by stochastic simulation (see Chapter 8) in Figure 17.1. Simulated experiments
with no true effect (a Cohen’s d effect size of 0) have a constant (false) positive
rate of 0.05 (black circles), whereas simulated experiments involving real effects
(blue squares, d = 0.25; white diamonds, d = 0.5) are more likely to be accurate
with larger sample sizes.
This is a rather counterintuitive feature of frequentist methods. We might
reasonably expect that as we gather more evidence, our tests should become
more accurate. That is the case in the presence of a true effect, but not in the
presence of a null effect. But this seems irrational, since in the real world two
things can be the same, and we would ideally like our statistical tests to be able
to show this.
So what has caused this curious state of affairs? Primarily it is due to a
17.2. AN UNFORTUNATE CONSEQUENCE OF THE FIXED TYPE I ERROR RATE339

1
Proportion significant

0.5

0.2 d=0.5
d=0.25
0.1 ● d=0

0.05 ● ● ● ● ● ● ● ●

0.02

0.01
4 8 16 32 64 128 256 512
Sample size (N)

Figure 17.1: Outcome of 100000 simulated experiments for different sample sizes
and effect sizes (Cohen’s d).
340 CHAPTER 17. BAYESIAN STATISTICS

philosophical position taken by many of the originators of frequentist tests.

Jacob Cohen, for example, who did seminal work on effect sizes (and from whom
Cohen’s d gets its name), took the position that every variable always influences
every other variable to some extent, even if the effect size is miniscule. Ronald
Fisher (who developed Analysis of Variance) argued that the null hypothesis
could never be proven or established, but that each experiment had the potential
to disprove the null.
If this position seems unbalanced, that’s because it is. Indeed, in recent years
it has been realised that combining Type I errors with publication bias (the
preference for both journals and researchers to publish results that are statistically
significant) has had a distorting effect on the scientific literature. Imagine that
20 laboratories across the world independently investigate an effect that does
not truly exist. Purely by chance, one laboratory is likely to find an apparently
significant result (assuming α = 0.05). If this is the only study that gets published
(because of publication bias) then a spurious effect will enter the literature. This
will be the case even if all the studies have a large sample size.

17.3 Bayesian statistics: an alternative ap-

proach
The methods we are about to introduce take a different philosophical position.
Instead of assuming that the experimental hypothesis is either true or unproven
(as in frequentist methods), Bayesian techniques try to estimate the probabilities
of both the experimental and the null hypotheses being true. In other words,
in Bayesian analysis we use our observations about the world (data) to assign
probabilities to our competing hypotheses. As we will discover in the following
sections, this avoids the problems outlined above with frequentist methods, as
well as several other shortcomings of the frequentist approach (see also Chapter
11 of Kruschke 2014; Dienes and Mclatchie 2018; Wagenmakers 2007).

17.4 The legacy of Thomas Bayes: Reverend

Thomas Bayes (probably not pictured in Figure 17.2) was an 18th century
English clergyman, who had a particular interest in mathematics. He developed
a mathematical theorem that specifies how conditional probabilities can be
combined. This is the basis of Bayesian statistics, and the reason why the
method bears his name. However much of the theoretical elaboration of the
approach was developed after his death by the French mathematician Pierre-
Simon Laplace. Whereas frequentist statistics can often be straightforwardly
conducted by hand (being based on the means and variances of the data),
Bayesian methods are often more computationally intensive. This means that,
although some aspects of Bayesian statistics predate their frequentist cousins,
they were not commonly used until computers become widely available.
17.5. BAYES’ THEOREM 341

Figure 17.2: Portrait of a Presbyterian clergyman, claimed (controversially) to

be Thomas Bayes.

The dominance of frequentist methods in scientific research for many decades

has lent Bayesian statistics a curious outsider status. Scientists who use them
often self-identify as ‘Bayesians’, and there is even a society that meets regularly
to discuss methodological development (the International Society for Bayesian
Analysis). Visits to Bayes’ grave in a London cemetery are common (indeed, I
have even been myself). Most interestingly, Bayesians are almost evangelical
about the use of their methods, and the literature contains many articles arguing
that frequentist methods are fundamentally flawed and should no longer be used.
I will leave the reader to come to their own conclusions about this debate.

17.5 Bayes’ Theorem

Bayes’ theorem is an equation that governs the combination of conditional
probabilities. A conditional probability is written P (A | B), and expresses the
probability of A given B. Everyday examples might be the probability that it
will rain given that it is February, or the probability that you have an infectious
disease given a positive test result. Where possible this book avoids equations,
but Bayes’ theorem (also called Bayes’rule) is sufficiently important that we will
discuss it explicitly. It is defined as:

P (B | A)P (A)
P (A | B) = . (17.1)
P (B)

In plain language, this equation tells us how to work out the probability of A
given B, if we already know the general probabilities of both A and B, and also
the probability of B given A. That probably doesn’t sound very plain! A more
concrete example might be more useful to introduce how this works.
Imagine that you are a general practitioner, and you want to know how likely
it is that a patient has a specific disease. If you know nothing else about that
patient, your estimate will be based on the prevalence of that disease in the
342 CHAPTER 17. BAYESIAN STATISTICS

population. Of course, in real life doctors know lots of other things about a
patient, not least whether they have any symptoms of the disease, and this
information would also factor into their diagnosis. But for the purposes of this
example, let’s imagine that the best estimate from the medical literature of the
baseline probability is P(A) = 0.01 (e.g. that one in a hundred individuals have
this disease). So your initial expectation, in the absence of any symptoms, is a
1% chance your patient has the disease, and a 99% chance they do not.
To improve your estimate, you take a blood sample from the patient and send it
off to be tested. You know that the test will correctly identify 90% of people
who have the disease; this is referred to as the sensitivity of the test, or the
true positive rate, and is conceptually similar to statistical power (see Chapter
5). This also means that the test will miss 10% of people who have the disease
- they will still be infected, but the test result will be negative, meaning they
appear to be healthy. When real clinical tests are developed, the aim is to get the
sensitivity as high as possible, though there will usually be technical limitations
that prevent perfect sensitivity.
The other aspect of test performance is what happens with uninfected patients,
who do not actually have the disease. This is summarised by the specificity, or
true negative rate of the test. This is how often the test will correctly identify
healthy patients as not having the disease - let’s say it’s 95% of the time for our
example. The remaining 5% of healthy patients will produce a false positive, and
might be incorrectly diagnosed as having the disease. Balancing the trade-off
between these two values (sensitivity and specificity) is a major challenge in test
development, as increasing sensitivity can often decrease specificity. A test that
is very sensitive but has low specificity might catch most people with the disease,
but also give false positives to lots of people who do not, perhaps resulting in
unnecessary treatment. An insensitive test that has high specificity would avoid
false positives, but might also miss many people who are truly ill.
A good way to think about the probabilities involved is to construct a table (see
Figure 17.3) showing the four possible outcomes for a hypothetical group of 1000
people tested. This table will look very familiar to those who have just read
Chapter 16 on signal detection theory, and indeed this can be thought of as a
signal detection problem (where the ‘signal’ is the presence of the disease). Out
of our 1000 people, we expect that 10 of them (1%) actually have the disease.
So the top row of the table (showing people with a true positive disease status)
adds up to 10 individuals, and the lower row (those with a true negative disease
status) is the remaining 990 individuals. The columns indicate the test results.
In the left column are the positive test results. This is 90% of the 10 people who
actually have the disease (so 9 of them), and 5% (the false positive rate) of the
people who don’t have it (~50 people). In the right column are the negative test
results. These comprise the one missed person who actually has the disease, and
940 correct rejections - people who don’t have the disease and tested negative.
Now let’s imagine we get back a positive test result for our patient. Can we
update our estimate from the baseline of 0.01 to incorporate the information
17.5. BAYES’ THEOREM 343

Test result

Positive Negative Total

True status

Positive 9 1 10
Hits Misses

Negative 50 940 990

FAs CRs

Total 59 941

Figure 17.3: Grid showing test results and true disease status for a group of
1000 individuals. FAs: false alarms; CRs: correct rejections.
344 CHAPTER 17. BAYESIAN STATISTICS

from the test, as well as the information we have about how accurate the test is?
We already know the baseline probability P(A) is 0.01. The other part of the
numerator of equation (17.1) is the probability of getting a positive test result if
we actually have the disease. This is the sensitivity value we identified above,
and so P (B | A) = 0.9.
The denominator of the equation is P(B). This is the probability overall of
getting a positive test result. We don’t know this yet, but we can work it out,
because we have enough information about the sensitivity and specificity, as
well as the baseline probability. The overall probability of getting a positive test
result will be a combination of the probability of getting a positive result if you
have the disease, and the probability of a positive result if you don’t have the
disease. Mathematically:

P (B) = P (B | A)P (A) + P (B | Ā)P (Ā). (17.2)

The first part of this equation is just the numerator term again (P (B | A)P (A)).
The second part is the probability of getting a positive test result if you don’t
have the disease. The bar over the A symbols are indicating the probability of
not having the disease. So the second section of the equation is the probability
of getting a positive result if you do not have the disease (the false positive rate),
multiplied by the probability of not having the disease in the first place. Overall,
the equation we need to calculate is:

P (B | A)P (A)
P (A | B) = , (17.3)
P (B | A)P (A) + P (B | Ā)P (Ā)

or, more specifically for this example:

P (positive | disease)P (disease)

P (disease | positive) = .
P (positive | disease)P (disease) + P (positive | healthy)P (healthy)
(17.4)
We can now plug in the probabilities from the above scenario to calculate the
posterior probability (in this instance, the probability of having the disease if
you get a positive test result). P (positive | disease) is the sensitivity, which is
90%, or 0.9. P (disease) is the prior probability, or baseline prevalence rate of
0.01 (1 in a hundred). P (healthy) is the rest of the population that don’t have
the disease, so it must be 1 − P (disease) = 0.99. Finally, P (positive | healthy)
is 1-specificity = 1-0.95 = 0.05. Completing the calculation with these numbers
gives:

0.9 × 0.01
P (A | B) = = 0.15 (17.5)
0.9 × 0.01 + (1 − 0.95) × 0.99
17.5. BAYES’ THEOREM 345

This outcome is known as the posterior probability. In this case, it is telling

us that if you get a positive result on this test, you only have a 15% chance
of actually having the disease. In some respects, that might not seem like an
especially great test, even though the sensitivity and specificity are quite high.
But note that the final probability we calculated depended not just on the
specificity and sensitivity of the test, but also on the prevalence of the disease
in the population. If any one of these factors changes, for example if the test
becomes more sensitive, or the disease becomes more or less prevalent, it will
affect our calculation. Figure 17.4 shows how different values of baseline and
sensitivity alter the posterior probability (P (A | B)). A test is most likely to
correctly identify a disease when the sensitivity and specificity are both high,
and the disease has a high prevalence in the population.
Probabilty of disease if positive

0.8

0.6

0.4

0.2 Baseline = 0.9

Baseline = 0.1
0
0 0.2 0.4 0.6 0.8 1
Sensitivity

Figure 17.4: Change in posterior probability for different values of test sensitivity
and baseline prevalence. The specificity was fixed at 0.9.

As a topical aside - as previously mentioned, this book was mostly written during
2020 when a novel coronavirus caused a deadly global pandemic. Crucial to
attempts to control the spread of the virus was the development of rapid and
accurate testing. Of course, in the early stages of the pandemic, the baseline
prevalence rate was changing by the day and in fact was acknowledged to be
inaccurate in most countries because only patients with severe symptoms were
being tested. For individual tests, sensitivity and specificity were often unclear,
and were certainly not communicated to the general public. In such situations,
assessing the likelihood of a test result being accurate is very challenging. It
346 CHAPTER 17. BAYESIAN STATISTICS

was also the case that many scientists and politicians had prior expectations
about the virus that were ultimately incorrect. It was assumed to be like a
severe flu with a relatively low fatality rate, which meant that implementing
countermeasures such as lockdowns and mass testing were delayed in many
countries. In the next section we will introduce probability distributions, which
allow us to quantify our uncertainty about a given situation and consider a range
of possibilities.

17.6 Probability distributions

The example of Bayes’ theorem above applies to point probabilities - exact
values of a probability. But typically our knowledge of probabilities is not exact,
and we would ideally like to incorporate this uncertainty into our calculations.
An important concept in Bayesian statistics that allows us to do this is the
probability distribution. These distributions specify the mean and spread of
relevant probabilities, allowing us to estimate our confidence in the final outcome.

Three critical concepts in Bayesian statistics are the prior, evidence and posterior
distributions. The prior distribution quantifies our expectations before we have
made any new observations. This is equivalent to the baseline probability of
having a disease in the above example (P (A)). In an experimental setting, it
might be generated using data (such as effect sizes) from previous studies, or
we might use a default (neutral) prior if other information is not available. The
evidence distribution (also called the likelihood function) is derived from our
new observations, which in an experimental or observational study would be the
data we collect. The posterior distribution is the outcome, and is calculated by
multiplying the prior and evidence distributions together according to Bayes’rule.
Example distributions are shown in Figure 17.5.

The prior distribution is an interesting feature of the Bayesian approach. It allows

us to incorporate our existing knowledge into the statistic we are calculating. One
criticism of Bayesian methods is that this is very subjective, and two different
researchers could easily derive different priors. However a counterargument is
that many aspects of statistics involve subjective decisions, which are often
hidden. Examples might include the choice of which test to conduct, or the
decision about when to stop collecting data, particularly if the experimenter
is aware of the results they have collected so far (known as optional stopping).
The prior is at least explicit, allowing other researchers to see if the outcomes
remain constant when the prior is changed. As we will see in the following
section, contemporary implementations of Bayesian methods typically offer a
default prior to make them more directly comparable to standard frequentist
tests. However the option remains for an experienced researcher to modify
the prior based on previous evidence. These two approaches are referred to as
subjective Bayes (when a prior is informed by earlier research) and objective
Bayes (when the prior is generic).
17.6. PROBABILITY DISTRIBUTIONS 347

Prior
Evidence
Posterior

−1 −0.5 0 0.5 1
Effect size (d)

Figure 17.5: Hypothetical probability distributions for the prior (dotted curve),
empirical evidence (dashed curve) and posterior (solid curve). Curves have been
vertically scaled for visualisation.
348 CHAPTER 17. BAYESIAN STATISTICS

17.7 The Bayesian approach to statistical infer-

ence
Now that we have introduced the idea of conditional probabilities, they provide a
convenient way to think about frequentist p-values. A p-value is the probability
that we will observe some data given the null hypothesis (i.e. assuming that
two groups really are equal): P (data | null). When the p-value is small, this
is telling us that it is very unlikely that we would have observed the data if
the null hypothesis were correct. Under these circumstances we reject the null
hypothesis, and conclude that the alternative (experimental) hypothesis is more
likely to be correct. This approach treats the data as a random variable, which
is compared to a fixed model (the null).
The Bayesian perspective is rather different. In this tradition, we treat the data
as a fixed observation about the world, and instead ask which of our hypotheses
(experimental or null) is more likely. In other words, we use Bayes’ theorem
to calculate P (experimental | data) and P (null | data). Note that here we are
no longer calculating the probability of observing the data (because we have
definitely observed it already), but instead working out the probability that a
model (or hypothesis) is true, given the data we have already observed. This is
the crucial distinction between the two traditions (frequentist and Bayesian).
Exactly how these posterior probabilities are calculated depends on the type of
data you are working with and your choice of prior. For some relatively simple
data such as proportions (e.g. coin tosses, patients infected in a vaccine trial,
correct trials, flies killed by a pesticide, etc.), the posterior probabilities can be
derived analytically using specific distributions. For a long time, until computers
were invented, this was really the only way of doing Bayesian statistics, and
explains why the approach was often ignored in favour of frequentist methods.
For more complex types of data, like continuous and categorical variables, the
mathematics required to derive the posterior becomes intractable. In those
situations, the posterior is estimated using a sophisticated stochastic method
known as Markov Chain Monte Carlo (MCMC). Very briefly, this combines
aspects of the methods discussed in Chapters 8 and 9 to randomly sample the
posterior distribution using random numbers. A detailed explanation of this
requires a whole book, so the interested reader is referred to Kruschke (2014)
and Lambert (2018).
Once the probabilities have been calculated, a useful summary statistic is to
take their ratio: P (experimental|data)
P (null|data) . This statistic is known as the Bayes factor
(Jeffreys 1961), and is often reported as the outcome measure for Bayesian tests1 .
1 Actually it’s a bit more complex than that, but explaining it this way makes for a clearer

comparison with frequentist methods. This value is technically called the posterior odds, and
is the product of the prior odds and the ratio of marginal likelihoods. In most implementations,
we set the prior odds to equal 1, because we assume that the null and alternative hypotheses
are equally likely. The posterior odds therefore depends entirely on the ratio of marginal
likelihoods, which is also called the Bayes Factor. So, in situations where we consider the
17.8. HEURISTICS FOR BAYES FACTOR SCORES 349

Table 17.1: Table of Bayes factor scores corresponding to different levels of

evidence.

Amount of evidence Supporting null hypothesis Supporting experimental hypothesis

Inconclusive 1.00 1
Some evidence 0.33 3
Strong evidence 0.10 10
Very strong evidence 0.03 30

If the experimental hypothesis is more likely, the Bayes factor score will be
>1. If the null hypothesis is more likely, the Bayes factor score will be <1.
Confusingly, some software calculates the ratio the other way up, which inverts
the interpretation of the scores. It is possible to indicate which way round the
probabilities were considered by using the subscripts BF10 and BF01 (where 1
is the experimental and 0 is the null hypothesis, and the first value corresponds
to the numerator). The Bayes factor tells us which model is best supported by
the data.

17.8 Heuristics for Bayes Factor scores

Unlike the p-values ubiquitous in frequentist statistics, Bayes factor scores do
not have a hard threshold that determines statistical ‘significance’. However,
Jeffreys (1961) proposed some heuristics that have been widely adopted, and are
given in Table 17.1. Bayes factors around 1 are inconclusive, typically as a result
of small sample size or very noisy data. As Bayes factors increase or decrease
from 1, this provides evidence in support of one or other hypothesis. Larger
values support the experimental hypothesis, with values around 3 considered
reasonably convincing evidence. These descriptors will often be referenced when
reporting the results of Bayesian statistics. See Dienes and Mclatchie (2018) for
an in-depth discussion of the Bayes factor.

17.9 Bayes factors can support the null hypoth-

esis
A useful feature of Bayes factors is that they can provide evidence in support
of the null hypothesis (see Table 17.1). To illustrate this, we can repeat our
earlier simulations of experiments with different sample and effect sizes, but this
time calculating Bayes factor scores instead of frequentist p-values. These are
shown in Figure 17.6. Bayes factors for the two conditions with positive effect
sizes (blue squares and white diamonds) increase as a function of sample size,

null and alternative hypotheses equally likely, the posterior odds and the Bayes factor are
equivalent.
350 CHAPTER 17. BAYESIAN STATISTICS

just as with their frequentist counterparts (Figure 17.1). But in the case where
the effect size is d = 0, the Bayes factors decrease as a function of sample size
(black circles). This means that some data sets will provide evidence for the null
hypothesis (Bayes factors <1/3), and this will not be conflated with inconclusive
data sets that may have insufficient sample size (Bayes factors around 1).

1000

100
Bayes Factor

10

d=0.5
1
●
●
d=0.25
● ● d=0
●
●
●
0.1 ●
●
4 8 16 32 64 128 256 512
Sample size (N)

Figure 17.6: Bayes factors for 100000 simulated experiments with different
sample sizes and effect sizes.

17.10 Bayesian implementations of different sta-

tistical tests
The general Bayesian approach has been applied to various statistical tests. In
brief, empirical observations (data) are combined with an appropriate prior to
estimate a marginal likelihood for both the null and experimental hypotheses.
The ratio of these likelihood estimates can then be used to calculate a Bayes
factor. Specific implementations vary, and these methods are still under active
development, particularly with respect to a suitable choice of prior (see Jeffreys
1961; Kass and Raftery 1995; Rouder et al. 2009; Gronau, Ly, and Wagenmakers
2019). The approach we will consider here uses a default prior, which involves as
few assumptions as possible (following the objective Bayes tradition). The idea
behind a default prior is that it covers the range of plausible non-zero effect sizes
(often using a Gaussian-like distribution), and is compared to a model where the
17.11. SO WHICH APPROACH IS THE BEST? 351

prior is zero (i.e. no effect). A detailed breakdown of the mathematics involved in

these calculations is beyond the scope of this chapter, and the interested reader
is directed to recent work on this topic (Rouder et al. 2009; Gronau, Ly, and
Wagenmakers 2019). For more complex experimental designs, estimating the
posterior distributions is typically done using the stochastic Markov Chain Monte
Carlo (MCMC) method, which means the implementations are computationally
intensive, and can take some time to calculate.

17.11 So which approach is the best?

We have seen that frequentist tests have two key shortcomings: a fixed false
positive rate (regardless of sample size), and an inability to provide evidence to
support the null hypothesis. Bayesian methods fix both of these problems, and
furthermore allow (in subjective Bayes) the incorporation of prior knowledge, so
that research can build iteratively upon previous findings. However, Bayesian
methods can be harder to implement than their frequentist counterparts, and
are less well understood by the wider research community.

Much has been written about the relative strengths of each approach, and we
will not attempt to resolve this question here. However, a practical suggestion
that may be of value is to start to report Bayes factors alongside frequentist
tests, in much the same way that effect sizes are now routinely reported in many
papers. In most cases, these will simply reinforce the frequentist outcome. But in
situations where a non-significant result is found, reporting the Bayes factor can
help to distinguish between an inconclusive result, and one in which the evidence
favours the null hypothesis. This adds substantial value to the interpretation of
null results, at very little cost.

17.12 Software for performing Bayesian analy-

ses
There are several software options for conducting Bayesian statistics. For those
more comfortable with spreadsheet-based graphical interfaces, the JASP package
is freeware, and offers both frequentist and Bayesian analyses. Within R, there
are many options depending on the level of detail required. Advanced users
can program their own bespoke hierarchical models using advanced stochastic
Markov Chain Monte Carlo (MCMC) methods, available in packages such as
rstan, rjags and brms. For those who wish to access Bayesian alternatives to
popular frequentist tests, the BayesFactor package offers a selection of functions
that are designed to mirror the generic frequentist R functions as closely as
possible. We will demonstrate some examples from this package in the following
section.
352 CHAPTER 17. BAYESIAN STATISTICS

17.13 Calculating Bayes Factors in R using the

BayesFactor package
First, we will generate some synthetic data and run a one-sample frequentist
t-test. The rationale for using random data (rather than one of the built-in
data sets in R) is that we know the ground truth population mean that we have
drawn our sample from (here a mean of 2, with a standard deviation of 3).
data <- rnorm(20, mean=2, sd=3)
t.test(data)

##
## One Sample t-test
##
## data: data
## t = 4.644, df = 19, p-value = 0.000177
## alternative hypothesis: true mean is not equal to 0
## 95 percent confidence interval:
## 1.800734 4.755699
## sample estimates:
## mean of x
## 3.278217
The effect size for the synthetic data is large, so this test gives a highly significant
result (p < 0.05). We can run a Bayesian t-test on the same data by calling the
ttestBF function from the BayesFactor package:
library(BayesFactor)
ttestBF(data)

## Bayes factor analysis

## --------------
## [1] Alt., r=0.707 : 167.5407 ±0%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS
This produces a substantial Bayes factor score of 167.5, which is given on the
third line of the output. This is consistent with the significant frequentist result,
and indicates very strong evidence for the experimental hypothesis.
The other parts of the output give us information about the prior and the
value we are comparing to. The value of r (0.707) is the prior scale parameter -
changing this as an optional input to the ttestBF function affects the behaviour
of the test. Larger values of the rscale argument might be appropriate if large
effect sizes are expected (and vice versa). The value of mu (0 for this example)
17.13. CALCULATING BAYES FACTORS IN R USING THE BAYESFACTOR PACKAGE353

is the comparison value for the test - in other words it is the prediction of the
null hypothesis, which is why it is described as being ‘against denominator’ - the
null appears on the denominator of the Bayes factor equation. The final line of
the output tells us the type of test (a one sample Bayes factor) and the type of
prior (JZS). The JZS prior (standing for Jeffreys-Zellner-Siow) is a particular
type of default prior constructed using a Cauchy distribution for the effect size
and a Jeffreys prior on the variance, as described by Rouder et al. (2009).
What about if we have a null result? To explore this, we can generate a new
data set with a true mean of zero, and run both tests again:
data <- rnorm(20, mean=0, sd=3)
t.test(data)

##
## One Sample t-test
##
## data: data
## t = -0.18955, df = 19, p-value = 0.8517
## alternative hypothesis: true mean is not equal to 0
## 95 percent confidence interval:
## -1.442209 1.202684
## sample estimates:
## mean of x
## -0.1197628
ttestBF(data)

## Bayes factor analysis

## --------------
## [1] Alt., r=0.707 : 0.2361331 ±0.02%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS
This time, there is no significant difference in the frequentist test, and the Bayes
factor score is much less than 1 (it is 0.24). This indicates support for the null
hypothesis, which of course in this case we know is true (because we drew values
from a distribution with zero mean).
Finally, how about a situation where the result is ambiguous? The following
simulation has a very small sample size (of N = 5):
data <- rnorm(5, mean=0.6, sd=1)
t.test(data)

##
## One Sample t-test
354 CHAPTER 17. BAYESIAN STATISTICS

##
## data: data
## t = 1.3714, df = 4, p-value = 0.2422
## alternative hypothesis: true mean is not equal to 0
## 95 percent confidence interval:
## -0.6562693 1.9373199
## sample estimates:
## mean of x
## 0.6405253
ttestBF(data)

## Bayes factor analysis

## --------------
## [1] Alt., r=0.707 : 0.7510559 ±0%
##
## Against denominator:
## Null, mu = 0
## ---
## Bayes factor type: BFoneSample, JZS

As you can see from the output, this produces a non-significant p-value, and
a Bayes factor score around 1 (0.75). Because the sample size is so small, we
simply don’t have enough data from our observations to conclude that either
the null or alternative hypothesis is more convincing.

If we have more than two groups to compare, we can conduct a one-way ANOVA
using the frequentist aov function, and the Bayesian anovaBF function. First,
we simulate and plot some data (see Figure 17.7:
# generate three levels of a dependent variable
dv1 <- rnorm(60, mean = 3, sd = 3)
dv2 <- rnorm(60, mean = 4, sd = 3)
dv3 <- rnorm(60, mean = 5, sd = 3)
alldata <- c(dv1,dv2,dv3) # combine together
group <- gl(3,60,labels = c("DV1", "DV2", "DV3")) # make condition labels
dataset <- data.frame(group,alldata) # combine into a data frame
plot(alldata ~ group) # plot as box plots

A frequentist ANOVA shows a significant main effect of group:

summary(aov(alldata ~ group, data = dataset))

## Df Sum Sq Mean Sq F value Pr(>F)

## group 2 127.4 63.70 7.19 0.000995 ***
## Residuals 177 1568.0 8.86
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
17.13. CALCULATING BAYES FACTORS IN R USING THE BAYESFACTOR PACKAGE355

10
alldata

5
0
−5

DV1 DV2 DV3

group

Figure 17.7: Simulated data for a one-way ANOVA.

(For those unfamiliar with R’s linear model syntax, the term alldata ~ group
is interpreted as ‘alldata is predicted by group’. In other words, alldata is the
dependent variable, and group is the independent variable - see Chapter 4). The
frequentist effect is complemented by a large Bayes factor score of 26.7:
bf <- anovaBF(alldata ~ group, data = dataset)
summary(bf)

## Bayes factor analysis

## --------------
## [1] group : 26.69677 ±0.01%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
Finally, we can calculate a Bayes Factor score for correlation and linear regression.
We first generate two weakly correlated data sets, by making the dependent
variable incorporate some proportion of the independent variable:
IV <- rnorm(60, mean = 3, sd = 1) # create independent variable
# create dependent variable with some portion of the IV
DV <- rnorm(60, mean = 4, sd = 1) + 0.5*IV
356 CHAPTER 17. BAYESIAN STATISTICS

regressiondata <- data.frame(IV,DV) # combine into a data frame

plot(IV,DV,type='p') # plot the data
cor.test(IV,DV) # run a correlation test

##
## Pearson's product-moment correlation
##
## data: IV and DV
## t = 3.6749, df = 58, p-value = 0.0005212
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## 0.2030679 0.6200815
## sample estimates:
## cor
## 0.4345837
The data show a significant positive correlation (not plotted here). There is also
a Bayesian correlation test (the correlationBF function) that can be called as
follows:
correlationBF(IV,DV)

## Bayes factor analysis

## --------------
## [1] Alt., r=0.333 : 72.60094 ±0%
##
## Against denominator:
## Null, rho = 0
## ---
## Bayes factor type: BFcorrelation, Jeffreys-beta*
Again, the Bayes factor is substantial. A frequentist linear regression can be
conducted on the same data using the lm function as follows:
rsreg <- lm(DV ~ IV, data = regressiondata)
summary(rsreg)

##
## Call:
## lm(formula = DV ~ IV, data = regressiondata)
##
## Residuals:
## Min 1Q Median 3Q Max
## -2.57064 -0.83313 0.06715 0.65323 2.38935
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 4.2038 0.4216 9.971 3.46e-14 ***
17.14. FURTHER RESOURCES ON BAYESIAN METHODS 357

## IV 0.4933 0.1342 3.675 0.000521 ***

## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.041 on 58 degrees of freedom
## Multiple R-squared: 0.1889, Adjusted R-squared: 0.1749
## F-statistic: 13.5 on 1 and 58 DF, p-value: 0.0005212
As we would expect based on the correlation results, there is a signficant effect of
the independent variable on the dependent variable, with a significant p-value in
the IV row of the coefficients table. The Bayesian equivalent is the lmBF function,
which uses the same syntax to define the relationships between variables:
bflr <- lmBF(DV ~ IV, data = regressiondata)
summary(bflr)

## Bayes factor analysis

## --------------
## [1] IV : 53.64218 ±0%
##
## Against denominator:
## Intercept only
## ---
## Bayes factor type: BFlinearModel, JZS
Again, the Bayes Factor score of 53.64 suggests strong evidence for the alternative
hypothesis, consistent with the significant frequentist test result and positive
correlation coefficient.

17.14 Further resources on Bayesian methods

The above examples give a flavour of how straightforward it can be to implement
Bayesian versions of standard statistical tests. Of course, more elaborate experi-
mental designs will require the use of other packages, and there are some very
good books available that explain how this can be done (Kruschke 2014; Lambert
2018). Bespoke Bayesian models usually require using stochastic methods to
estimate the shape of the posterior distribution, and these two books (known
as ‘the dogs book’ and ‘the chillis book’ because of the pictures on their front
covers) describe the current state of the art on how to go about this. Despite
their potential complexity, Bayesian methods have some clear advantages over
their frequentist counterparts, and are likely to become more heavily used as
they become better developed, standardised, and more widely understood.

17.15 Practice questions

1. In frequentist methods, the p-value is a statement about:
358 CHAPTER 17. BAYESIAN STATISTICS

A) The probability of the experimental hypothesis being true

B) The probability of the null hypothesis being true
C) The probability of observing the data if the null hypothesis is true
D) The probability of observing the data if the experimental hypothesis
is true
2. In frequentist methods, the false positive rate:
A) Increases with sample size
B) Decreases with sample size
C) Stays constant regardless of sample size
D) Depends on statistical power
3. In Bayesian methods, information we already know is called:
A) The posterior
B) The prior
C) The likelihood
D) The data
4. Bayes’ theorem allows us to combine:
A) Conditional probabilities
B) p-values from t-tests
C) Bayes factor scores
D) Data from experiments with different sample sizes
5. A Bayes factor denoted BF01 indicates:
A) The ratio of the experimental over the null hypothesis
B) The probability of the experimental hypothesis being true
C) The probability of the null hypothesis being true
D) The ratio of the null over the experimental hypothesis
6. A Bayes factor of BF10 = 7.5 indicates:
A) Some evidence in support of the alternative hypothesis
B) Some evidence in support of the null hypothesis
C) Strong evidence in support of the alternative hypothesis
D) Strong evidence in support of the null hypothesis
7. In a situation with an effect size of d = 0, the Bayes factor (BF10 ) score:
A) Increases with sample size
B) Decreases with sample size
C) Stays constant regardless of sample size
D) Depends on statistical power
8. In the case of a non-significant frequentist test, running a Bayesian alter-
native would help to:
A) Distinguish between the null hypothesis and an underpowered (incon-
clusive) study
B) Distinguish between the experimental and null hypotheses
C) Increase the statistical power
D) Get a significant result
9. If some data support the experimental hypothesis with a probability of
0.67, and the null with a probability of 0.02, what is the Bayes factor score
BF10 ?
A) 0.03
17.15. PRACTICE QUESTIONS 359

B) 0.01
C) 33.5
D) 0.65
10. Suppose that the probability of a man sporting a moustache is 5%. However,
during November lots more people grow moustaches for charity, and the
probability of any individual man having a moustache increases to 30%.
Use Bayes’rule to calculate the conditional probability that it is November
given that you see a man with a moustache. The probability is:
A) 0.08
B) 0.25
C) 0.35
D) 0.49
Answers to all questions are provided in section 20.2.
360 CHAPTER 17. BAYESIAN STATISTICS
Chapter 18

Plotting graphs and data

visualisation

An essential part of scientific communication is to be able to display your data

in a clear and interpretable way. Despite the importance of this skill, it is rarely
taught explicitly, and most researchers have to pick up some general heuristics
about graph plotting by osmosis. This chapter will attempt to distill some
principles of data visualisation that I have found useful, though of course such
guidelines are necessarily subjective - others may disagree on what looks good.

There are several plotting packages in R that one might choose, the most popular
being the ggplot2 package (Wickham 2016). This is considered part of the
tidyverse suite of packages created by the makers of RStudio. The ggplot2
functions use a plotting grammar, whereby a plot is constructed by adding layers
of components. This can produce some excellent results, however in my personal
experience I find that there is usually something I need to change from the
default way that ggplot2 behaves. This often proves to be very difficult indeed,
and so I have found that the built-in plotting functions from the base R graphics
package afford more flexibility. We will use these functions throughout the
chapter. Remember that the code to produce all figures in the book, including
this chapter, is available at: https://github.com/bakerdh/ARMbookOUP.

We will start by discussing some key principles of data visualisation, and con-
siderations regarding the choice of colour palettes. We will then build up a
plot with multiple components using the basic plot commmand, demonstrating
various options. Finally, we will discuss how to export graphs, and also how
to combine several graphs into a multi-part figure. These tools are sufficient
for assembling publication-quality plots entirely within R, without the need for
additional software (such as Adobe Illustrator).

361
362 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

18.1 Four principles of clear data visualisation

The purpose of plotting graphs is to communicate data in a clear and unbiased
way, so that it can be easily interpreted by your reader with a minimum of effort.
The following principles are things I have picked up informally over a period of
many years. Of course others may have a different aesthetic sense and disagree
with many of my specific choices, but hopefully this section will at least prompt
readers to consider how they construct figures, and why they make the choices
they do.

18.1.1 Different conditions should be maximally distin-

guishable
When I first started writing up the results of experiments, I had in my head
the idea that ‘proper’ science needed to involve black and white graphs with
thin lines joining some very small data points (see an example from my own
undergraduate work in Figure 18.1). I probably got this idea from reading too
many papers from the 1980s and before, when figures were created by hand
using a ruler, and colour reproduction was not available in most journals (or was
prohibitively expensive). This is no longer the case, and modern tools can make
graphs that are more visually appealing, and also clearer to the reader.
The main problems with Figure 18.1 are that the data points are very small,
and extremely hard to distinguish. Different symbols were used to represent
different conditions (circles, squares and triangles), but the reader needs to work
quite hard to realise this, and to see which symbol is which. We can replot these
data using larger symbols, thicker lines, and also giving each condition its own
colour and line style. Note that in common with many of the figures in this
book, I have been quite sparing with colour. This is partly due to using the 2
colour printing system common in publishing, but it is also good practice to
avoid over-using colour when it is not strictly necessary. However, the reader
should not be dissuaded from using colour when it is appropriate, and when it
makes a figure easier to understand.
Figure 18.2 shows the same data replotted with these changes, and it is im-
mediately much clearer which condition is which. Notice that the symbols are
now large enough to easily distinguish, and the use of colour makes linking with
the figure legend effortless and automatic. Finally, the axis labels are more
appropriate in the revised figure, avoiding block capitals, and the tick marks
point in towards the data.

18.1.2 Axes should be labelled and span an informative

range
It is critical that axis labels explain clearly what is plotted on that axis, including
appropriate units. Although it may seem obvious, the axis tick marks should
always be plotted at regular intervals in the chosen units (see Figure 18.3a) -
18.1. FOUR PRINCIPLES OF CLEAR DATA VISUALISATION 363

100 ●
●
●
●
90 ●
PROPORTION CORRECT

● ●
80
70
60
50
40
30 ● 0 Degrees
20 20 Degrees
10 40 Degrees
0
0 15 30 45 60 75 90
ELEMENT ORIENTATION (deg)

Figure 18.1: Reproduction of a rubbish graph from my own final year undergrad-
uate project. The experiment involved finding contours in a field of randomly
oriented elements. Performance is best when all elements in the contour had a
similar orientation (circles), but rapidly degrades as each element’s orientation
is jittered. Error bars give 95% confidence intervals.
364 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

100 ● ●
● ●
90 ●
● ●
Percent correct
80
70
60
50
40
30 ● 0 Degrees
20 20 Degrees
10 40 Degrees
0
0 15 30 45 60 75 90
Orientation (deg)

Figure 18.2: Replotted data from my undergraduate project, using colour, line
style and symbol size to maximally distinguish conditions.

there are some truly awful examples in the media that violate this basic rule.
For interval and ratio data, the units of measurement are meaningful, and so
these need to be represented appropriately in a plot. Some spreadsheet software
that is widely used to produce graphs has a default setting where the x-axis is
categorical even for continuous data, which can result in an unequal spacing of
values that is very misleading.

For some types of data, a logarithmic scale may be more appropriate than a
linear one. Log-scalings can be used when the data have a positive skew in linear
units, as is often the case with values involving time (such as reaction times),
and also when the data are ratios. There are several ways to plot a log axis.
Figure 18.3b shows the tick marks in linear increments, but on a log axis, so
that successive spaces become smaller (as used on old logarithmic graph paper).
An alternative is to log-transform the data, and plot this on a linear axis, where
the labels are in log units (Figure 18.3c). A disadvantage of this method is that
the log units themselves are arbitrary, and hard to relate to the original values.
The method I find most transparent is to give tick labels in the original linear
units (rather than in log units), but spaced in logarithmic steps, such as factors
of 2 or 10 (see Figure 18.3d). Conventions differ across disciplines about which
approach to use, and so it is always best to check some related papers in your
field.
 18.1. FOUR PRINCIPLES OF CLEAR DATA VISUALISATION 365

1000 1000 3.2 1600

(c) (d)
●

(a) ●
900 (b) ●

900 ● 800 ●

700 ●
●
●

● ●
●
● ●
●
●●

800
● ● ● ● ●
● ● ●
● ●●
● ● ●● ●

600
● ● ● ●
● ●● ● ● ●
Reaction time (log(ms))
●●
●● ●●●●● ●●
●
● ●● ●
●● ●
● ●●●
●
● ●●
● ● ●● ●● ● ●●
●
●●● ●
●
●
● ●

● ●●●
● ●
●
●●●
●●●●●
●●
●
●
800 ●
Reaction time (ms)

Reaction time (ms)

Reaction time (ms)

●● ● ● ● ●● ●
● ●
● ●● ● ● ●

700
●●●

500 ●● ●● ● ●●
● ●● ●●
● ● ● ● ●● ● ●● ● ●●
●●
● ● ● ● ●● ● ● ● ●●
●●
● ●● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ●
● ● ●●●●●
● ●● ● ●● ●
● ● ●●●●
●●●●● ●●
●●● ● ● ●● ● ● ●●

2.8
●●● ●● ●
●● ● ● ●● ● ● ● ●
● ●●●● ●
●●● ●
●●● ●●●●● ●●●●● ●●●● ● ●
● ● ● ●
● ●
●●● ●
●●● ● ●●●●● ●● ●●
● ● ●● ● ●●●● ●● ●
●●● ●● ●●●● ●●
● ● ●● ●● ● ● ● ● ●● ● ● ● ●●
● ●
● ●●●● ● ●
●
●● ●●●●● ● ● ●● ●● ● ● ●
● ●● ●● ● ● ● ● ●● ●● ● ● ● ●
● ● ● ●● ● ●
●●●● ● ●●● ● ●● ●● ●
● ● ● ●●●● ●● ●●●● ●● ●● ●●●● ●●
● ● ●●
●●●● ● ●
●● ● ● ●●
●●●● ●●●●●● ●● ●● ●● ● ● ●●● ● ● ● ● ●●● ● ●
● ● ●● ●● ● ●● ●● ● ●●
●●
●●● ● ● ●●● ●●●● ●
● ● ● ● ● ● ●
●●● ●●●
● ●● ● ●● ● ● ● ● ● ●● ●● ● ● ● ● ● ●● ●●

400
● ● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ●● ●● ● ● ●

600
●●●●●
● ●● ●● ●
●●● ●● ●● ●● ●●
●● ● ● ● ● ●● ●●●●●●●● ●●● ●●●● ● ●
●●●● ●● ●● ●●● ●
● ●●● ●● ● ● ●● ●● ●●● ●
● ●●● ●● ● ●
● ●● ● ●● ● ●●● ●● ● ● ●●● ● ●● ●● ● ● ●●● ● ●●
● ● ● ● ●●●● ●
●●
●
●●●●
●● ●●●
●
● ●●● ●
● ●● ●●●●
●
●●
●●●●●●
●● ●
●● ●● ●●●●●
●●● ●●● ●●
●●
●
●● ●● ●●●●●
●●● ●●● ●●
●●
● ● ●●●●● ●●●●● ●●● ● ●●●●●● ●
●●●● ● ●● ● ● ●
●● ● ●●●
● ● ●● ● ● ●
●● ● ●●●
●
● ●● ●●● ●●●●● ● ●
●●●● ● ●●●
● ● ●
● ● ●●● ● ●● ●●
● ●● ● ●● ●● ● ● ● ● ● ●● ●●● ●●
● ●● ● ●● ●● ● ● ● ● ● ●● ●●●
● ● ●●
● ● ●
●
● ●
●●●
●
● ● ●● ●● ●
●● ●●● ●● ●●
●●
●
●
●● ● ● ● ●● ●● ●●●● ●● ●●
● ● ● ● ● ●● ●● ●●●● ●● ●●
● ●
● ●● ●●
● ●●●
●● ●
●●● ●●
● ●● ●●
●●●● ●●●
●●
●● ●●●●●● ●● ●●
● ●
●●
● ●
● ●●● ●●●
●● ●●
●
●●● ● ● ● ●
●●● ●● ●●
● ●
●●
● ●
● ●●● ●●●
●● ●●
●
●●● ● ● ● ●
●●● ●●
● ● ●● ●● ● ● ●●●
●●●●● ● ● ●● ●● ●
●● ●
●● ●●●●●●
● ●●●●●● ●
● ● ●●●● ●●●● ●●●●●●●●●
● ●●●●● ●
● ● ●●●● ●●●●
● ●● ● ●
● ● ●● ● ●
● ●●●●●● ●●● ● ●
●
●●
● ●● ● ●●●● ●● ●
● ●●●●● ●●●●● ●● ●● ● ●● ● ●●●●
●
●● ●
● ●●●●● ●●●●● ●● ●● ●
●● ● ● ●● ●
●● ●● ● ● ●●● ●● ● ●
●
●●● ●● ●● ●● ●●● ●●●●● ●●
●● ●
●● ●● ●● ●●●● ●
●●
●● ● ● ●●●●● ●●
●● ●
●● ●● ●● ●
●●
●● ● ●
●
●●● ● ● ●●●
● ●
● ● ●
●●● ●● ●●●
● ●● ● ●●●●●● ●●●● ● ●● ●●● ●●●●● ● ●●●
● ●●●●● ●● ●●● ●●●●●
●●●●
● ●●●
● ●●●●●
●● ● ●
●
● ●● ● ●●●●●
● ● ●●● ●●● ● ● ● ●●●● ●
●
●●●● ● ● ●●●
●●●
● ●● ●●● ● ● ●●
●●●● ●●●●●● ● ●● ●●● ●
●●● ●●●
● ●● ●●● ● ● ●●
●●●● ●●●●●● ● ●● ●●● ●
●●●
● ● ●● ● ● ● ● ●●● ● ●●● ●●● ● ● ● ● ● ●● ●
●● ●
●●● ●
● ●● ● ● ● ● ● ●● ●
●● ●
●●● ●
● ●●

400
● ●●● ●●● ●●● ●●
●● ● ● ●● ● ●●● ●●
●● ● ● ●● ●

500
● ●● ● ● ● ● ●●● ● ●
●
●●● ●●
● ● ●● ●●●●●
●●● ● ● ● ●
●●● ●● ●●●●●●●●● ●● ●
●●● ●
●● ● ●
●●● ●● ●●●●●●●●● ●● ●
●●● ●
●●
● ● ● ●●●● ●● ● ●● ●● ●● ●●
●●●● ●● ●●● ● ●● ● ●●
●
● ●● ●● ●●
●●●● ●● ●●● ● ●● ● ●●
●
●
●● ●● ● ●● ● ● ● ● ●●
● ●● ●● ●● ●●● ●●●●●
● ●●● ● ● ●● ●●●● ●
● ●●●● ●●●●
●●●
●●●
●● ●● ●●● ● ●●● ●●●●● ●
● ●●●● ●●●●
●●●●● ●● ●●● ● ●●● ●●●●●
● ●
●●
●
● ● ●● ● ●● ●●● ●● ● ● ●● ●●● ●● ●
●●●●● ●● ●●● ●●●
●● ● ● ●● ●●● ●● ●
●●●●●

300
● ●●● ●●●
●●●●
●● ● ●● ● ● ● ●●● ● ●●● ●
●●
●
●●●●●●●● ● ●● ●● ●● ●●● ●
●●
●
●●●●●●●● ● ●● ●● ●●
●●● ●●● ●●●● ●● ●●●●●● ● ● ●● ●●● ●●●●● ●●●● ●●●●●● ●●●●●● ● ●●●
● ●● ●●● ●●●●● ●●●● ●●●●●● ●●●●●● ● ●●●
●
●●●● ● ● ●● ●●●● ●
●● ● ●● ●●
●
● ● ● ●●● ● ●●● ● ● ● ● ● ●
●●● ● ●●● ●●●● ●●
● ● ●● ●● ●
●● ●●● ●● ●●
●
● ●
●●● ● ●●● ●●●● ●●
● ● ●● ●● ●
●● ●●● ●● ●●
●
●
●●● ●
● ●● ● ●● ●●● ● ●● ● ● ●●● ● ● ●
● ●●
● ●●● ●●
●● ●
●●● ●● ●●●●
● ● ●●● ● ●●● ● ● ●●
● ●●● ●●
●● ●
●●● ●● ●●●●
● ● ●●● ● ●●● ● ●
●● ●●●●
●●
●●●
● ● ●●● ● ● ●●●● ● ●●● ●● ● ●● ● ● ●● ● ● ●● ●●●●● ●● ●
●●● ●●●● ●●● ● ●●●●● ●● ●
●●● ●●●● ●●● ●
●●
●● ● ● ●●●● ●
● ●
● ●●
● ●● ● ●
●
●
● ●
● ●● ●● ● ●
● ● ● ● ●● ●● ● ● ●●
●
●●● ●●● ●●●
●●
●● ●●●● ●●● ● ●● ●● ● ● ●●
●
●●● ●●● ●●●
●●
●● ●●●● ●●●
● ● ● ●● ●●●●●●
●● ●●●● ● ● ●● ● ●● ●● ● ●●●● ●● ● ●●●●●●● ● ●●●● ●● ●●
●● ● ●● ● ●●●● ●● ● ●●●●●●● ● ●●●● ●● ●●
●● ●
●●● ●●● ●● ● ●● ● ●●● ●●● ●●● ● ●● ● ●● ● ● ● ●● ● ● ● ●● ●●●
●
●● ● ● ● ●
●
●●●● ● ● ●● ●● ●●●
●
●● ● ● ● ●
●
●●●● ● ● ●●
●●
●●●● ● ●●● ●
●●●● ●●●●●● ● ●● ● ● ● ● ●
● ● ●● ●● ●
●●● ●● ●● ●● ● ●●● ●
● ● ●● ●● ●
●●● ●● ●● ●● ● ●●●
●●● ●● ●●●
●● ●
● ● ● ● ● ●● ●
●●● ●● ● ● ●● ●●●
● ●
● ● ●
●●● ●● ● ● ●● ●●●
● ●
● ●
● ●●
●●●
●
●
●●●
● ● ●● ●●●●●● ●
●● ●●● ● ● ● ●●●●● ●
●●
●
● ●● ● ● ●●● ● ●●
● ● ● ● ●●●●● ●
●●
●
● ●● ● ● ●●● ● ●●
● ● ● ●

400
● ● ●●●●● ●● ●● ●●
●●● ●● ● ● ● ● ● ●●● ● ●● ● ●●● ● ● ● ● ●● ●● ●●● ● ● ● ● ●● ●●
●●●●●
● ●●
●●●●●● ●●● ●●●● ● ●
●●
●
●●
● ● ● ●● ●● ●●●●● ●● ● ● ● ● ● ● ●●
●●●●● ●● ● ● ● ● ● ● ●●
●
● ●●●● ●● ●●
● ●●
● ● ●● ●●● ●● ● ●●●
● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●
●● ●●● ●●●
●●
● ● ● ●● ●●●●
●●●
●●●●●
●●
●●●
● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●●● ● ● ● ● ● ●● ● ● ● ●● ●●●● ●
●●● ●
●●●
●
●●●
● ●●●●
●●● ● ● ● ● ● ● ● ●● ●●
● ● ● ● ● ● ●● ●●
● ●
●● ●●● ●●
● ●● ●●●●● ●●●●●● ●● ●
●●●● ● ●●● ● ● ●
●
●● ●● ●
●
● ●●
● ● ● ●● ●● ●
●
● ●●
● ● ●
●
● ●●
● ●
● ●● ●● ●
●
●● ●●● ●●● ●
●● ●●● ● ●
●
● ● ● ● ● ● ●●● ● ●● ●● ● ● ● ● ●●● ● ●● ●● ●
●●
● ●●● ●
●●● ●●●
● ● ●● ●●●● ●●● ● ●●● ● ● ●● ● ● ● ● ● ●● ●● ●●● ● ● ● ● ●● ● ●● ●● ●●● ● ● ● ● ●●
●●●●● ●● ●●
●●● ●●●● ●●
●● ● ●
●● ●
● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ●
●●●●●●

2.4
●● ●
●● ● ● ● ● ●● ●
●● ●
● ● ●●●
●
● ● ●
●● ●
● ●●●● ●
●
●●●● ●● ●●
●● ● ● ● ● ● ● ● ● ●● ●● ●●
● ● ● ● ●● ● ● ● ● ● ●● ●● ●●
● ● ● ● ●●
●● ●●● ●●●● ● ●
●● ●●
●
●●●● ● ● ●● ●● ● ●
●
●
●
●● ● ●● ●●
●●●
●●
●●
● ●● ●●
●●●●●●
● ●
●●●
● ● ● ●● ● ●● ● ● ●● ● ●● ●
●● ●●●● ●● ●●● ●● ● ●●●●●●● ● ● ●● ● ● ● ● ●● ● ●

200
●

300
● ● ●● ● ●●
●● ● ● ● ●●●●
● ● ● ● ●
●●● ● ●●● ● ● ●● ●● ●●● ●
● ●● ●
●
● ●● ●
● ●● ● ●
● ● ● ●● ●● ● ● ● ● ● ●●
● ●●
●
●● ●● ● ● ●●● ● ● ●● ●●
● ● ● ● ●
● ●● ●● ● ●● ● ● ● ●
● ●● ●● ● ●●● ●●● ● ●● ●● ● ●● ● ●
● ● ● ● ●●

200
● ● ● ● ● ●
● ● ● ● ● ●● ● ●
● ●●
● ●● ●● ● ● ●●●● ● ● ●
● ●● ●
● ● ●
●●●● ● ● ● ●
●● ●

200
●
●
●

100
0 100 2 100

Figure 18.3: Examples of linear and logarithmic axes. Panel (a) shows a linear
axis for reaction time data in milliseconds (ms). Panel (b) shows a logarithmic
axis for the same data, with tick marks spaced every 100 ms, resulting in uneven
tick placement. Panel (c) shows the logarithmic transform of the data - the
units are now much harder to interpret. Finally, panel (d) shows logarithmic
scaling with tick marks spaced at equal factors, but values given in linear units.
Position on the x-axis is arbitrary.
366 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

There is also some disagreement about whether it is necessary for a linear axis
to extend to zero (especially for bar charts). Sometimes starting the axis at
an arbitrary point closer to the means can exaggerate an apparent difference
between conditions (see Figure 18.4). This is bad practice if the intention is to
mislead, and exaggerate the apparent size of an effect. Yet it is not necessarily
the case that zero should always be included. For many variables, zero is not a
plausible value for any of our observations - reaction times being a good example
again. Stretching an axis to zero arbitrarily can obscure real and meaningful
differences in the data. Instead, I favour using axes that span an informative
range, and combining this with error bars to give a clear visual indication of
which differences are meaningful, as we will discuss in the following section.
A good rule of thumb might be to have the axes extend around 2 standard
deviations above the highest point, and around 2 standard deviations below the
lowest point you are plotting. However, this heuristic also needs a good dose
of common sense - round the upper and lower values to a sensible value, such
as an integer, or a factor of 10 or 100, depending on the natural scale of your
data. You should ideally use the same range for multiple plots of the same type
of data.

10 14
(a) (b)
9 12

10
8
Height (m)

Height (m)

8
7
6
6
4

5 2

4 0

Figure 18.4: Illustration of how apparent differences between groups can be

exaggerated by starting the y-axis at a value other than zero (a), compared with
at zero (b). The data are identical in both panels.

18.1.3 A measure of variability should always be included

A third principle is that we should include some indication of the variability
of our data. This might feel like second nature to anyone with a scientific
background, but it is surprising how often graphs are shown in the media,
and even in technical reports (such as market research reports), that have no
18.1. FOUR PRINCIPLES OF CLEAR DATA VISUALISATION 367

visual indication of how precise our estimates of means or medians might be. For
example, the data shown in Figure 18.5a might suggest a large difference between
the two conditions, yet without an indication of how variable each estimate is,
we cannot tell whether the apparent differences are real, or just due to sampling
error.

The most widely used measures of variability are the standard deviation, standard
error, and confidence intervals. The standard deviation (SD, or σ) is the square
root of the variance, so it is the most direct indication of the underlying variability
in the data (see Figure 18.5b). It does not depend on the number of estimates,
meaning it should remain approximately constant as the sample size changes.
On the other hand, the standard error (SE) scales the standard deviation by the
square root of the sample size (see Figure 18.5c). This means that as we collect
more data the standard error becomes smaller, reflecting the greater precision of
our estimate of the mean (or other measure of central tendency).

Alternatively, confidence intervals (CIs) explicitly represent our confidence in the

location of the mean. Actually, standard errors are a type of confidence interval -
they are the 68% confidence intervals of the mean. But it is more typical to plot
95% confidence intervals (see Figure 18.5d), because these can act a bit like a
t-test - if the confidence intervals do not overlap with some particular point, then
the mean is likely to differ significantly from that point (at p < 0.05). Confidence
intervals can be estimated with the parametric assumption of 1.96 × SE, or by
bootstrap resampling of the data (see Chapter 8).

Something worth observing in Figure 18.5 is that the different types of error
bars can appear to indicate different amounts of variance, if we don’t take into
account what they are actually showing (remember it is the same data set in
each case). For example, standard deviations will always produce larger error
bars than standard errors, and without clarity about what is being shown a
reader might develop a spurious understanding of the variability of the data. It
turns out that even trained researchers often misinterpret what error bars are
showing, and treat standard errors and 95% confidence intervals very similarly
(Belia et al. 2005). Many people appear to use heuristics, such as whether the
error bars ‘just touch’, which are not a good indicator of statistical significance.
All of this confirms that we always need to explicitly state in the figure caption
what our error bars represent.

In terms of graphical presentation, error bar whiskers such as those shown in

Figure 18.5b-d are widely used and well understood. However in some types of
plot they overlap across conditions, which can make them hard to see properly
(see e.g. the error bars for the square and triangle at 60 degrees orientation
in Figure 18.2). An alternative visualisation is to plot a shaded region that
encompasses the error bars, but is semi-transparent. This solves the problem of
overlap, and also looks visually appealing (see Figure 18.6).
368 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

14 14
(a) (b)
12 12

10 10
Height (m)

Height (m)
8 8

6 6

4 4

2 2

0 0

14 14
(c) (d)
12 12

10 10
Height (m)

Height (m)

8 8

6 6

4 4

2 2

0 0

Figure 18.5: Illustration of different types of error bar. Panel (a) shows two
means with no indication of the variance. Panel (b) shows the same data with
standard deviations plotted. The lower row shows the same data with standard
errors (c) and 95% confidence intervals (d) plotted. In this example, the two
groups are significantly different (t=2.5, df=18, p=0.02, d=1.1).
18.1. FOUR PRINCIPLES OF CLEAR DATA VISUALISATION 369

100 ● ●
● ●
90 ●
● ●
Percent correct

80
70
60
50
40
30
20 ● 0 Degrees
20 Degrees
10
40 Degrees
0
0 15 30 45 60 75 90
Orientation (deg)

Figure 18.6: Replotted data from my undergraduate project, with shading to

represent the error region (95% confidence intervals).
370 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

18.1.4 Where possible, include data at the finest mean-

ingful level
Traditional bar plots (e.g. Figure 18.7a) can hide a multitude of sins. The
main problem is that without seeing the original data points that went into
calculating the means, it is impossible to tell whether one or more outliers might
be responsible for the results. Or, there could be unexpected patterns in the
data (such as bimodally distributed results) that are hidden by the averaging
process. A classic demonstration is Anscombe’s quartet (Anscombe 1973) - a
set of four scatterplots that have identical summary statistics, including the
correlation coefficient, but appear very differently when the individual data
points are plotted.

The most straightforward solution is to plot the individual data points that went
into calculating the averages being displayed (see Figure 18.7b). Interestingly, it
has been shown that the presence of a bar can distort the reader’s interpretation of
individual data points. Newman and Scholl (2012) found that people judge points
falling within the bar to be more likely to be part of the underlying distribution
than those falling outside of it. Since the bar itself is not informative, it can
be safely removed, and replaced by a line or point to indicate the average.
In practical terms, individual data points should not get in the way of the
representation of the mean, and so alpha transparency can be useful. Plotting
data points semi-transparently also allows us to see when multiple points with
similar values are overlapping with each other.

For some data sets, it can also be helpful to plot distributions, or kernel density
functions (smoothed histograms). There are several conventions for plotting
these. For example, with bivariate scatterplots showing correlations, it is common
to plot histograms along the margins of the plot (we will see an example of this
later in the chapter). For univariate data, the violin plot replaces bars with
mirrored distributions (Figure 18.7c). A more recent suggestion is the raincloud
plot (Allen et al. 2019), which shows a distribution with individual data points
either underneath or to one side, and a mean and error bar in between (Figure
18.7d).

For data types that involve continuous functions rather than individual points
(e.g. where a timecourse has been measured), individual functions can still be
plotted with a thin line-width, and the grand mean overlaid in a thicker line and
more salient colour (Rousselet, Foxe, and Bolam 2016). An example of this is
shown in Figure 18.8, replotting the ERP difference data described in Section
15.9. For this data set there is one clear outlier (highlighted in blue) that is
contributing only noise to the data set, and might reasonably be excluded on
various criteria. In the absence of outliers, this method of presentation allows
the reader to confirm that nothing has been hidden by the averaging process,
giving them more confidence in the results of any inferential statistics, and the
conclusions drawn from them.
 18.1. FOUR PRINCIPLES OF CLEAR DATA VISUALISATION 371

1000 1000 1000 1000

(a) (b) (c) (d)
900 900 900 900
800 800 800 800
Reaction time (ms)

Reaction time (ms)

Reaction time (ms)

Reaction time (ms)

700 700 700 700
600 600 600 600
500 ● 500 ● 500 ● 500 ●

400 400 400 400

300 300 300 300
200 200 200 200
100 100 100 100
0 0 0 0

Figure 18.7: Example visualisations that provide more information about the
distribution of data. Panel (a) shows a generic bar plot (error bars indicate
95% confidence intervals). Panel (b) shows the raw data points plotted semi-
transparently. Panel (c) shows a violin plot, illustrating the distribution of points
as a symmetrical kernel density function. Panel (d) shows a raincloud plot,
which features the distribution and the raw data.
372 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

20

15
Difference (µV)

10

−5

−10
−200 0 200 400 600 800 1000
Time (ms)

Figure 18.8: Replotted ERP difference data, showing results for individual partic-
ipants (thin traces) and the group mean (thick line). The shaded region indicates
±1SE across participants. The blue curve highlights an outlier participant that
might sensibly be excluded.
18.2. CHOOSING COLOUR PALETTES 373

18.2 Choosing colour palettes

The choice of colour palette for plotting a figure is more than a purely aesthetic
one. Two key factors that should be considered are accessibility to readers with
atypical colour vision, and the introduction of spurious features through a colour
palette that is not perceptually uniform. In this section, we discuss both issues.

18.2.1 Palettes that are safe for colourblind individuals

As most readers will be aware, some individuals are ‘colourblind’ (indeed, you
may be yourself). It is not necessarily the case that these individuals see
the world like a black and white photograph. Such severe colourblindness
(known as achromatopsia) is quite rare (around 1 in 30,000), though it can be
acquired through injury or partial asphyxiation. A more common form of genetic
colourblindness affects around 2% of men, but far fewer women because the genes
involved are linked to the X-chromosome. This is red-green colourblindness,
with two variants: deuteranopia and protanopia. Affected individuals lack a
class of cone photoreceptor (a cell that detects light) in their eye. In general
terms, these individuals cannot tell the difference between red and green, with
both appearing as a yellowish brown colour. A rarer variant of colourblindness
(tritanopia) involves similar problems with perceiving the colour blue. Finally,
many individuals (>5% of men) have anomalous colour vision, where they do
not lack classes of photoreceptor, but the photopigments are shifted in their
sensitivity, making discrimination between some colours more difficult.
These various types of colourblindness will have important consequences for how
graphs are perceived by some readers (see also Crameri, Shephard, and Heron
2020). A very simple heuristic that will aid the largest population of colourblind
readers is to avoid using red and green to distinguish two different conditions.
If red and green need to be used together, they should be paired with different
symbol types (as in Figure 18.2) and/or use different lightnesses - for example
a light red and a dark green. This will allow the brightness and shape of the
symbols to differentiate conditions even if the colour information is lost (which
also happens when using a black and white printer).
We can simulate how a colourblind individual will perceive a figure or image
through our understanding of the conditions. Websites and applications exist
that can do this for individual images, for example http://www.vischeck.com
and https://colororacle.org. For a given colour palette, the likely appearance
for colourblind individuals can also be simulated, as demonstrated in Figure
18.9 for the rainbow palette. Notice the presence of peaks and discontinuities of
luminance in the second row. For all four simulated varieties of colourblindness
there are regions of the palette that are spaced far apart but have a similar
appearance. These features will make it hard to interpret images plotted with
this palette.
Because of the different profiles of various types of colourblindness, it is challeng-
374 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

Rainbow palette

Original

Black/White

Deutan

Protan

Tritan

Figure 18.9: Simulation of colourblindness for a rainbow colour palette. The

first row shows the original palette, which transitions through the colours of the
rainbow. The second row shows a greyscale rendering, as might be perceived by
an individual with achromatopsia. The third and fourth rows show simulations
for individuals with two types of red/green colourblindness - reds, greens and
yellows all appear similar. The final row simulates tritanopia, where blue and
green cannot be distinguished.
18.2. CHOOSING COLOUR PALETTES 375

ing to produce palettes that are appropriate for everyone. Palettes such as the
rainbow palette shown in Figure 18.9 are problematic partly because they involve
many different hues. A better alternative are colour maps that pass smoothly
from one hue to another, and simultaneously change in brightness. One example,
the parula palette, is shown in Figure 18.10. Both ends of the palette are clearly
distinguishable for all four simulated varieties of colourblindness. When used to
plot real data (e.g. contour maps), the parula palette is still attractive and clear,
even when the colour information is removed (see Figure 18.11).

Parula palette

Original

Black/White

Deutan

Protan

Tritan

Figure 18.10: Simulation of colourblindness for the parula colour palette.

18.2.2 Perceptually uniform palettes

Another important consideration when choosing a colour palette is whether the
palette is perceptually uniform. Many standard colour maps vary in luminance
(brightness) across their range (see e.g. Figure 18.9, and Ware (1988)), and also
ignore features of the human visual system that mean we are less sensitive to
changes in some regions of colour space (Borland and Taylor 2007). Generating
colour maps that avoid these problems is a challenge, and some useful principles
are described by Kovesi (2015).
In general, most colour palettes involve linearly changing the values of the
red, green and blue colour channels in various combinations. But this often
produces an overall luminance profile that has unwanted peaks in different
parts of the palette, as shown by the black curve in the left panel of Figure
376 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

Figure 18.11: The Himmelblau function, plotted with two colour palettes and
their greyscale equivalents - the rainbow palette (left) and the parula palette
(right). The upper row shows the original palettes, and the lower row shows the
same values but with colour information removed (simulating achromatopsia).
Notice how the rainbow palette introduces sharp boundaries of brightness between
regions, whereas the parula palette has a smoother transition.
18.2. CHOOSING COLOUR PALETTES 377

18.12. Perceptually uniform palettes reverse these priorities, and aim to create
consistent changes in luminance (and perceived hue) by manipulating the red,
green and blue levels nonlinearly (see right panel of Figure 18.12). Notice how
the standard ramp on the left has bands of higher brightness (especially in the
reddest part of the palette), but the perceptually uniform palette has a much
smoother transition across the whole range. This is important when representing
data using a colourmap, because a nonuniform palette will introduce spurious
features that are not present in the data.

Linear ramp Perceptually uniform

1 1

0.8 R 0.8 R
Luminance

Luminance

0.6 L 0.6 L
B B
0.4 0.4

0.2 G 0.2 G

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Palette level Palette level

Figure 18.12: Profile of red (R), green (G) and blue (B) values, and overall
luminance (L) for a standard linear ramp from black through red, orange, yellow
and white (left panel), and a perceptually uniform equivalent (right panel).

To demonstrate this, Figure 18.13 shows a time-frequency plot of neural responses

to a visual stimulus, plotted using a linear ramp (left) and a perceptually uniform
palette (right). Notice how the linear ramp produces banding and distortions in
378 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

the data, making the peaks and troughs appear more salient and over-saturated.
The perceptually uniform palette produces smoother transitions between features,
and the true location of the peak in the lower left corner is clearer. For most
types of scientific image, these characteristics are to be preferred for the accurate
communication of results.

Figure 18.13: Time-frequency plots of neural responses to a visual stimulus

presented at 0 ms (indicated by the blue line). Data were recorded using MEG
and are shown for a single sensor. Left panel shows a linear colour palette, right
panel shows a perceptually uniform colour palette

As well as being used for displaying scientific image data and surfaces, perceptu-
ally uniform colour palettes can also be used for selecting colours to plot data
points and curves. We will demonstrate this in the practical sections of the
chapter.

18.3 Building plots from scratch in R

My preferred approach to creating plots in R is to create a blank axis first, and
add individual features to it. In the following section we will construct the plots
shown in Figure 18.14 one component at a time. This will illustrate how to build
and label axes, and how to add widely-used features such as symbols, points,
lines, error bars, and arbitrary shapes.
For all plots, we first need to create a new empty figure, and add appropriate
axes. This requires us to know the dimensions of the plot (i.e. the range of the
x- and y-axes), and then to call an empty plot with the axes and annotations
suppressed:
18.3. BUILDING PLOTS FROM SCRATCH IN R 379

1 1
(a) ● Condition 1 (b)
Condition 2

0.8 0.8
●
● ●
Y axis title

Y axis title
0.6 0.6
●

●
0.4 0.4
●
●
●
●
0.2 0.2 ●

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X axis title X axis title

1 1
(c) (d)
0.8 0.8
Y axis title

Y axis title

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X axis title X axis title

Figure 18.14: Example outputs from the plotting function. Panel (a) shows an
empty axis, ready for data to be plotted. Panel (b) shows an example of two
conditions plotted with lines, points, error bars and a figure legend. Panel (c)
shows an example of the polygon function with alpha transparency. Panel (d)
shows an example of kernel density functions for an x-y scatterplot.
380 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

# create an empty axis of the correct dimensions

plotlims <- c(0,1,0,1)
plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=plotlims[1:2], ylim=plotlims[3:4])

The above code will appear to do nothing at all, but it will define an empty
axis in the plot window of RStudio. We can then add axes with tick marks at
sensible intervals using the axis command, tick labels with the mtext command,
and axis labels with the title command.
ticklocs <- seq(0,1,0.2) # locations of tick marks
axis(1, at=ticklocs, tck=0.01, lab=F, lwd=2)
axis(2, at=ticklocs, tck=0.01, lab=F, lwd=2)
mtext(text = ticklocs, side = 1, at=ticklocs) # add the tick labels
# the 'line' command moves away from the axis, the 'las' command rotates to vertical
mtext(text = ticklocs, side = 2, at=ticklocs, line=0.2, las=1)
title(xlab="X axis title", col.lab=rgb(0,0,0), line=1.2, cex.lab=1.5)
title(ylab="Y axis title", col.lab=rgb(0,0,0), line=1.5, cex.lab=1.5)

The above code will produce the axes shown in Figure 18.14a. These axes are
quite minimal - it is also possible to add the off-axes (again using the axis
command, but with the side option of 3 or 4). If required, we can also force the
axes to be square by adding par(pty=“s”) before the call to the plot command.
Additional text can be added anywhere in the plot using the text function. This
is often useful for adding panel labels (a, b, etc.) for example:
text(-0.05,0.95,'(a)',cex=1.8, pos=4)

The first two inputs are the x- and y-coordinates, and the third input is the text
string we want to plot. The cex option specifies the font size, and the pos option
has several possible values that determine the position of the text relative to the
coordinates: 4 means that the string is plotted to the right of the coordinates.
A brief note about font size and style - many readers have visual impairments
that make small text particularly difficult to read. Therefore it is a good idea to
make text within your figure as large as is practical. Some people also find that
fonts without serifs (the little decorative lines added to some letters) are easier
to read. Popular sans serif fonts include Arial, Calibri and Helvetica.

18.4 Adding data to plots

Once our axes are defined, it is time to add data to the plot. The examples
here will illustrate plotting data points, error bars, lines and semi-transparent
polygons. First, let’s create some synthetic data to plot using a random number
generator as follows:
# create some synthetic data to plot as points and lines
datax <- sort(runif(10,min=0,max=1))
18.4. ADDING DATA TO PLOTS 381

datay <- sort(runif(10,min=0.2,max=0.8))

SEdata <- runif(10,min=0,max=0.1)

The generic plotting functions do not require the data to be stored in a specific
format such as a data frame. They work fine with vectors of numbers. Lines and
error bars should ideally be plotted first so that they appear behind the data
points. A line connecting the data points can be defined using the lines function:
# draw a line connecting the points
lines(datax, datay, col='cornflowerblue', lwd=3, lty=1)

We provide the x- and y-values as vectors, define the colour using the col option,
and the width using the lwd option. There are also several line styles that can
be selected with the lty option, though here we have chosen the default option
(1), which plots a continuous straight line. Other options include dashed (option
2) and dotted (option 3) lines.

Error bars are plotted using the arrows function. We need to manually add and
subtract the standard error to define the upper and lower extents of the bars as
follows:
# add lower error bars
arrows(datax, datay, x1=datax, y1=datay-SEdata, length=0.015, angle=90, lwd=2)
# add upper error bars
arrows(datax, datay, x1=datax, y1=datay+SEdata, length=0.015, angle=90, lwd=2)

Notice that we plot all of the lower error bars at once, followed by all of the
upper error bars. The arrows function allows us to enter a vector of values for
each input, just like the lines function. As the name suggests, the arrows the
function is designed to plot actual arrows, but setting the angle of the arrow to
90 degrees (as in the above code) produces error bars that are flat at the ends.

Next, we can add our data points with the points function. There are many
different point styles available that can be selected through the pch option. I
tend to favour styles 15-20 (solid shapes of a single colour) and styles 21-25
(filled shapes with an outline). Details on the available shapes are given in the
help file for the points function. For solid shapes of a single colour, the colour is
specified by the col option. For filled shapes, we specify the outline colour with
the col option, and the fill colour with the bg (background) option, allowing the
two parts of the shape to have different colours.
# draw the data points themselves
points(datax, datay, pch=21, col='black', bg='cornflowerblue', cex=1.6, lwd=3)

We can also include a legend to indicate what different symbols refer to. The
legend function requires x and y coordinates corresponding to its upper left
corner. We also provide a list of text labels for the conditions, and repeat some of
the options from the points and lines functions to specify the symbol properties.
382 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

legend(0, 1, c('Condition 1','Condition 2'), col=c('black','black'),

pt.cex=1.6, pt.bg=c('cornflowerblue','grey'),lty=1, lwd=3,
pch=21:22, pt.lwd=3, cex=1, box.lwd=2)

Putting all of the above lines of code together (with some additional data for a
second condition) gives us the plot in Figure 18.14b.

18.5 Setting alpha transparency for points and

shapes
Whenever a colour is defined in R, we also have the option of setting its trans-
parency using the alpha channel. Images on a computer are usually defined in
terms of the intensity of the red, green and blue pixels on the display. But we
can also specify a fourth value, alpha (or α), which sets transparency between 0
(fully transparent, and therefore invisible) and 1 (fully opaque). Some level of
transparency is useful for two reasons. First it allows us to see plot features that
are layered on top of each other, which allows us to estimate the density of clouds
of many small points. Second, it is a handy way of reducing the saturation of a
colour. There is evidence that highly saturated colours can introduce perceptual
distortions that affect how data are interpreted, whereas less highly saturated
colours are perceived more veridically (Cleveland and McGill 1983).
We can set the alpha level using the rgb function, which also requires that
we specify the red, green, and blue values. The following code specifies semi-
transparent blue and black colours:
colour1 <- rgb(0,0,1,alpha=0.5)
colour2 <- rgb(0,0,0,alpha=0.5)

colour1

## [1] "#0000FF80"
colour2

## [1] "#00000080"
Notice that the colours are converted to a hexadecimal representation. If we
don’t know the RGB values of the colour we want to use, but instead have the
colour name, the following helper function (addalpha) will convert to the correct
format and add the alpha level:
addalpha <- function(col, alpha=1){
apply(sapply(col, col2rgb)/255, 2, function(x) rgb(x[1], x[2], x[3],alpha=alpha))}

colour1 <- addalpha('cornflowerblue',0.5)

colour2 <- addalpha('black',0.5)
18.6. CHOOSING, CHECKING AND USING COLOUR PALETTES 383

We can then use these colour values with lines, points and polygons. Polygons
are two-dimensional shapes of arbitrary specification. For example, we could
draw two rectangles using the polygon function as follows:
polygon(c(0.2,0.8,0.8,0.2),c(0.2,0.2,0.6,0.6),col=colour1)
polygon(c(0.4,0.6,0.6,0.4),c(0.4,0.4,0.9,0.9),col=colour2)

In these two lines of code, the first two vectors of numbers are x- and y- values
that specify the corners (vertices) of the polygons. Adding the code to an empty
plot produces the image in Figure 18.14c. Of course we are not limited to only
four vertices, and a common use of the polygon function is to plot histograms
as kernel density functions. With a larger set of correlated data, we can add
histograms to the margins of a scatterplot as follows:
datax <- rnorm(1000,mean=0.5,sd=0.1)
datay <- rnorm(1000,mean=0.5,sd=0.1) + (datax-0.5)

a <- density(datax)
a$y <- 0.2*(a$y/max(a$y))
polygon(a$x, 1-a$y, col=colour1,border=NA)

a <- density(datay)
a$y <- 0.2*(a$y/max(a$y))
polygon(1-a$y, a$x, col=colour2,border=NA)

The density function generates the kernel density distribution, and the subsequent
line of code rescales the height to be between 0 and 0.2 so that it fits neatly
onto our plot. Notice that the x- and y- variables are reversed for the second
polygon so that it appears on the right side of the plot. Figure 18.14d plots
these polygons along with the data points, which are also semi-transparent.

18.6 Choosing, checking and using colour

palettes
There are many ways to define colours in R. We have already seen that we can
define a colour by its RGBα values, using a name (such as ‘red’), or using a
hexadecimal number. There are various functions for converting between these
different formats. For example, the rgb function takes values for the individual
colours (scaled from 0-1) and outputs a hexadecimal code:
rgb(1,0.5,0,alpha=0.8)

## [1] "#FF8000CC"

The col2rgb converts a colour name or hexadecimal code back to red, green and
blue values (scaled 0 - 255):
384 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

col2rgb('orange')

## [,1]
## red 255
## green 165
## blue 0
col2rgb("#FF8000CC")

## [,1]
## red 255
## green 128
## blue 0
Note that this function removes the transparency information.
To obtain a custom colour palette, we can use the colorRamp function to
transition smoothly between two or more colours, and sampling the ramp at
an arbitrary number of intermediate points. For example, the following code
produces the palette shown on the left side of Figure 18.12:
cr <- colorRamp(c("black","red","orange","yellow","white"))
thispal <- rgb(cr(seq(0, 1, length = 256)), max = 255)

The colorRamp function is unusual in that it actually returns a custom function

(named cr in the above example) rather than a standard data object, which we
can then sample at a sequence of values between 0 and 1.
There are also many built-in and default palettes that can be accessed through
the grDevices package. Built-in palettes include rainbow (sometimes called jet,
and shown in Figure 18.9), heat.colors, terrain.colors, topo.colors and cm.colors
(cyan-magenta). Further details are available in the help for the Palettes function.
Finally, a great many additional packages provide further colour palettes. The
pals package is particularly useful, as it contains a series of perceptually uniform
colour palettes, as well as tools for assessing the properties of a colour palette.
The perceptually uniform colour maps are described in the kovesi function (there
are around 50), and can be used to generate a palette with an arbitrary number
of levels, for example:
thispal <- kovesi.linear_bgyw_15_100_c67(5)
thispal

## [1] "#1B0084" "#2441D4" "#4C9A41" "#C0CC23" "#FFFFFF"

When plotting data from several conditions, setting the number of colours to
equal the number of conditions is a sensible approach. This will give a set of
hues that should distinguish the conditions, but also appear to be drawn from a
consistent set.
18.6. CHOOSING, CHECKING AND USING COLOUR PALETTES 385

The pal.test function produces a series of test images for a given palette, as
shown in Figure 18.15, which shows the output of the following expression:
pal.test(kovesi.linear_bgyw_15_100_c67)

kovesi.linear_bgyw_15_100_c67

250
200
150
100
50
0

0 50 100 150

Figure 18.15: Test image for colour palettes, generated by the pal.test function
for the Kovesi linear blue-green-yellow-white palette.

The graph in the lower right shows the red, green, blue and luminance ramps in
the same format as Figure 18.12. The upper right image shows the full palette
with a sine wave modulation along the upper edge. If the oscillations are harder
to see in some parts of the palette, this indicates perceptual non-uniformity.
The volcano images in the lower row are the same image with the colour palette
ranging from 0 - 1 and from 1 - 0. If some features are visible in only one of
these images, it again suggests problems with the palette. Since the Kovesi
palettes are all perceptually uniform, they pass most of these tests. It is worth
also viewing the output of the test function for a non-uniform palette (see Figure
18.16). Notice that the two volcanoes appear to differ in size, and that parts of
the sine-wave oscillation are not visible (particularly in the green band).
To check how a palette will appear to individuals with several types of colour-
blindness, the pal.safe function produces images like those shown in Figures 18.9
and 18.10.
Once we have selected a suitable colour palette, we can apply it to an image using
the image function. A suitable test image is the interesting two-dimensional
386 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

rainbow

250
200
150
100
50
0

0 50 100 150

Figure 18.16: Test image for colour palettes, generated by the pal.test function
for the rainbow palette. Observe in particular that the sinusoidal oscillation in
the upper right panel is very hard to see in some regions of the palette (e.g. the
green band).
18.7. SAVING GRAPHS AUTOMATICALLY 387

mathematical function known as Mishra’s Bird, defined by the equation:

2 2
f (x, y) = sin(y)e(1−cox(x)) + cos(x)e(1−sin(y)) + (x − y)2 (18.1)

The following code plots this function using the palette stored in the rpal object,
with the output shown in Figure 18.17:
x <- matrix(rep(seq(-10,0,length.out=200),200),nrow=200,ncol=200)
y <- t(matrix(rep(seq(-6.5,0,length.out=200),200),nrow=200,ncol=200))
z <- sin(y)*exp((1 - cos(x))^2) + cos(x)*exp((1 - sin(y))^2) + (x - y)^2

rpal <- kovesi.diverging_bky_60_10_c30(256)

image(x[,1], y[1,], z, col=rpal, useRaster=TRUE, axes=FALSE, ann=FALSE)

The image function requires vectors of x and y values, and a matrix of z (intensity)
values, as well as a palette specification for the col option. When it is called,
the function generates a new plot, for which axes and labels can be plotted if
required. The useRaster option tells the function to plot either as a raster image
(if TRUE) or a vector image (if FALSE). Raster images are like photographs,
where the RGB value of each pixel is stored at a particular resolution. Vector
graphics use graphical primitives such as lines, points and polygons to represent
images. For most plots (i.e. lines, points, polygons), vector graphics are preferred
because they can easily be enlarged with no loss of quality. However the vector
rendering of images such as the one shown in Figure 18.17 often look worse
than a raster version, because the individual elements of the texture surface get
separated by little white lines. It is worth trying both options to see what looks
best.

18.7 Saving graphs automatically

By default, RStudio sends all graphics to the Plots window. We can manually
export a plot from here using the ‘Export’ button at the top of the window,
which provides options to save in a number of formats. However, it is often more
efficient to automatically export the figure we have constructed. This can be
done in a variety of graphics formats. But when we choose to export a figure, it
will no longer be drawn to the plot window. It is therefore sensible to develop
the code that generates the plot first, and add in the commands to export it
later.
In computer graphics there are two main types of file format: vector and raster
files. As we mentioned in the previous section, raster images are like photographs.
They have a set resolution (i.e. number of pixels), and the colour value of each
pixel is represented explicitly as a triplet of red, green and blue values. It is
appropriate to use raster formats (such as .jpeg, .tiff, .png and .bmp files) to
store photographs, medical images, and similar types of data. However there are
two big downsides to raster formats. First, for high resolution images, the file
388 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

Figure 18.17: The Mishra Bird function, plotted using the image function and
the Kovesi diverging bky palette.
18.7. SAVING GRAPHS AUTOMATICALLY 389

sizes can become very large. Second, for a fixed resolution, if we zoom into a
raster image too far it will start to look blocky and pixelated, because we can
see each individual pixel (and sometimes compression artifacts). This can look
particularly bad for text, as shown in the example in Figure 18.18a.

Figure 18.18: Comparison of raster and vector art. Panel (a) shows a figure
legend from a graph that was saved in a raster format (jpeg) and then zoomed
in. Panel (b) shows the same thing for a vector format (pdf) - the text and
objects are sharper, with no visible pixels.

An alternative are vector file formats (including .pdf, .svg, .ps and .eps formats).
Vector images are lists of instructions about how to draw an image using a
collection of lines, shapes and text - much like the way we actually construct
graphs in R. Vector files do not have a set resolution, and if we zoom in on a
particular part the computer can redraw it with a high level of detail (see Figure
18.18b). This is the natural format for scientific figures, and most journals prefer
you to submit artwork in a vector file format. An added advantage is that vector
images typically have much smaller file sizes than their raster equivalents. It is
possible to embed raster images as part of a vector file, though these will still
have all of the limitations of the original image (e.g. fixed resolution, and large
file size).
We can export a graph as a vector pdf file (pdf stands for portable document
format) using the pdf function:
pdf('filename.pdf', bg='transparent', height = 5.5, width = 5.5)

The height and width parameters are in inches, and other background colours
can be set using the bg option if required. All plots you create will then be
exported to this file up until the following command is called:
390 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

dev.off()

The svg function exports in scalable vector graphics format in a similar way.
Other functions can export as raster graphics formats (jpeg, tiff, png and bmp
functions). For all of the raster functions the file size is set in pixels by default.
Finally, we can export in ps and eps format using the postscript function:
postscript('filename.ps', horizontal = FALSE, onefile = FALSE,
paper = 'special', height = 5.5, width = 5.5)

For this function the paper option needs to be set to ‘special’ for the height
and width options to be used. Postscript files are vector images (though they
can have raster images embedded inside of them). In the next section we will
discuss how to re-import individual graphs that have been saved in ps format,
to combine into a single figure without losing the vector information. There is
one downside to this format - postscript files cannot store alpha transparency
information. However we will suggest a workaround for this issue.

18.8 Combining multiple graphs

In most published journal articles, figures contain multiple sub-panels to show
different aspects of some data. If these components are all the same size, there
are some useful R functions and packages that will allow us to define these
individual panels within the same figure. For example, many of the multi-panel
figures in this book were constructed using the par function:
par(mfrow=c(2,2))

The above code will generate a 2 x 2 plot like Figure 18.14, with each subsequent
panel being added in sequence. The layout function is slightly more sophisticated,
and can allow for plots of different sizes. However, sometimes publication-quality
figures need more flexibility than is permitted by regular layouts. For this reason,
this section will demonstrate an alternative approach to combining plots that is
built on functions from the grImport and grid packages.
The grImport package allows us to import a postscript (.ps) file by first converting
it to a custom xml format as follows:
library(grImport)

PostScriptTrace('filename.ps')
e1 <- readPicture('filename.ps.xml')

Note that for this to work, you will need the free Ghostscript tools installed
on your system (see https://www.ghostscript.com/). The imported figure is
described as a set of lines, text and filled surfaces stored in a systematic way - in
other words it is a vector image. We can draw this figure in its entirety onto a
new plot, either in the Plots window or (more typically) a pdf file that we wish
18.8. COMBINING MULTIPLE GRAPHS 391

to export to. To do this, we use the grid.picture function from the grid package
as follows:
grid.picture(e1,x=0.5,y=0.5,width=0.5,height=1)

The advantage of this approach is that we have total control over the size and
placement of the figure panel in the new plot. By default, the image will be
centred on the x and y coordinates provided, with the width value specifying a
proportion of the width of the plot window. Somewhat unintuitively, the height
option specifies a proportion of the width, and so should generally be set to 1 to
avoid distorting the aspect ratio of the figure.

Putting everything together, we can create a plot, save it to a postscript file,

re-import it and plot it using the following code, which produces Figure 18.19.
# generate the graph and save as postscript
postscript('outputfile.ps', horizontal = FALSE, onefile = FALSE,
paper = 'special', height = 5.5, width = 5.5)

plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=plotlims[1:2], ylim=plotlims[3:4])

axis(1, at=ticklocs, tck=0.01, lab=F, lwd=2)
axis(2, at=ticklocs, tck=0.01, lab=F, lwd=2)
mtext(text = ticklocs, side = 1, at=ticklocs)
mtext(text = ticklocs, side = 2, at=ticklocs, line=0.2, las=1)
title(xlab="X axis title", col.lab=rgb(0,0,0), line=1.2, cex.lab=1.5)
title(ylab="Y axis title", col.lab=rgb(0,0,0), line=1.5, cex.lab=1.5)

datax <- rnorm(1000,mean=0.5,sd=0.1)

datay <- rnorm(1000,mean=0.5,sd=0.1) + (datax-0.5)

a <- density(datax)
a$y <- 0.2*(a$y/max(a$y))
polygon(a$x, 1-a$y, col=pal2tone[2],border=NA)

a <- density(datay)
a$y <- 0.2*(a$y/max(a$y))
polygon(1-a$y, a$x, col='black',border=NA)

points(datax,datay,col=rgb(0,0,0),pch=16,cex=0.6)

dev.off()

# import the figure

library(grImport)
PostScriptTrace('outputfile.ps')
e1 <- readPicture('outputfile.ps.xml')
392 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

# create a new plot and add the graph at several scales

plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=plotlims[1:2], ylim=plotlims[3:4])
grid.picture(e1,x=0.8,y=0.2,width=0.2,height=1)
grid.picture(e1,x=0.6,y=0.4,width=0.4,height=1)
grid.picture(e1,x=0.25,y=0.75,width=0.5,height=1)

Figure 18.19: Example of multi-panel plot with panels of different sizes and
positions.

# clean up the external files we generated

file.remove(c('outputfile.ps','outputfile.ps.xml'))

This approach to collating plots works very well, but there is one problem - the
postscript format does not support alpha transparency. This means that when
18.9. ADDING RASTER IMAGES TO A PLOT 393

we create our individual sub-plots, we cannot include any transparent features as

they will generate an error. However, we can specify transparency when the plots
are loaded back in with the readPicture function. The picture object created by
this function is a list of all of the components that make up a plot. We can loop
through this list, find the features we want to be transparent, and change their
colour settings:
for (n in 1:length(e1@paths)){ # loop through all features
temp <- class(e1@paths[n]$path)[1] # look at just one feature
# check if this feature is a filled shape
if (pmatch(temp,"PictureFill",nomatch=0)){
# set any filled shape to be semi-transparent
e1@paths[n]$path@rgb <- addalpha(e1@paths[n]$path@rgb,alpha=0.3)}}

# plot everything again

plot(x=NULL,y=NULL,axes=FALSE, ann=FALSE, xlim=plotlims[1:2], ylim=plotlims[3:4])
grid.picture(e1,x=0.8,y=0.2,width=0.2,height=1)
grid.picture(e1,x=0.6,y=0.4,width=0.4,height=1)
grid.picture(e1,x=0.25,y=0.75,width=0.5,height=1)

18.9 Adding raster images to a plot

Sometimes it is necessary to include a raster image as part of a figure, for example
to illustrate experimental stimuli, or show something like a microscope image or
Western blot. We can load images from a wide range of graphics formats using
packages such as jpeg (for .jpg files) and tiff (for .tiff files). These images are
stored as matrices in R, either with two dimensions (x and y) for a greyscale
image, or with three dimensions for a colour image (x, y and the colour channels:
red, green and blue). For example, we can load in a greyscale jpeg image of a
violin (from the Bank of Standardised Stimuli, see Brodeur et al. (2010)) using
the readJPEG function as follows:
library(jpeg)
violin <- readJPEG('images/BOSSr/violin.jpg')
dim(violin)

## [1] 256 256

We see using the dim function that the image is 256 x 256 pixels (I have resized
it - the original database is at 2000 x 2000 pixels). This image can then be
inserted into a plot using the rasterImage function from the graphics package as
follows:
rasterImage(violin,0.2,0.2,0.5,0.5)

This will plot the image at the x- and y-coordinates given by the first two
numbers for the bottom left corner, to the x- and y-coordinates given by the
394 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

Figure 18.20: Example of multi-panel plot with transparency.

18.10. PRACTICE QUESTIONS 395

last two numbers for the top right corner. Note that graphs in which the axes
have different limits or units will require careful thought about how to set these
coordinates to preserve the correct aspect ratio of the image. For example, if
the x-axis runs from 0 to 1, but the y-axis spans from 0 to 2, the extent of a
square image in the y-direction must be twice that in the x-direction.

It is also possible to rotate the image if required using the additional angle
option, specified in degrees. Rotation is about the lower left corner of the image
(not the centre), so this will also shift the centre of the image, meaning that the
coordinates often need to be adjusted for appropriate placement. A good way to
think about this is that the second pair of coordinates are really specifying the
size of the image (rather than its top right corner). For example, consider the
code:
rasterImage(violin,0.8,0.8,1.1,1.1,angle=180)

This actually plots the image between x=0.5 and x=0.8, because the rotation
through 180 degrees about the lower left hand corner means that the top right
corner has moved: Figure 18.21 shows the violin image plotted at both rotations.
One practical consideration when plotting raster images is that they will cover
up anything plotted underneath them. It is often pragmatic to plot the images
first, so that any points, lines and other features of the figure appear on top of
the image.

The methods described in this chapter can be combined to produce publication-

quality plots entirely within R. I have used this approach for a number of papers,
and example scripts are available for several of them. For example, scripts to
produce a dozen figures from D. Baker et al. (2021) are available on the Open
Science Foundation website at: https://osf.io/ebhnk/. I hope that the principles
for clear data visualisation and choosing colour palettes described in this chapter
will help you to produce more informative, accessible and beautiful figures.

18.10 Practice questions

1. Plotting an axis in log units can be problematic because:
A) The units are harder to interpret
B) The tick marks are spaced irregularly
C) Skewed distributions appear more normal
D) The base of the logarithm affects the appearance
2. Starting an axis at a number other than zero is sometimes (unethically)
used to:
A) Distort the distribution of the data
B) Allow logarithmic units to be used
C) Minimise the apparent differences between conditions
D) Exaggerate the apparent differences between conditions
3. Standard errors will always be:
396 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Figure 18.21: Demonstration of using the rasterImage function to add a square

image of a violin to a plot at two rotations.
18.10. PRACTICE QUESTIONS 397

A) Smaller than standard deviations but larger than 95% confidence

intervals
B) Smaller than standard deviations and smaller than 95% confidence
intervals
C) Larger than standard deviations but smaller than 95% confidence
intervals
D) Larger than standard deviations and larger than 95% confidence
intervals
4. Adding individual data points to a plot:
A) Allows us to check visually for outliers
B) Makes error bars unnecessary
C) Makes the means unnecessary
D) Is not possible for continuous data like timecourses
5. A person who cannot discriminate between green and blue, but can dis-
criminate between red and green is most likely to have:
A) Achromatopsia
B) Deuteranopia
C) Protanopia
D) Tritanopia
6. The parula palette:
A) Changes smoothly in hue but has luminance peaks
B) Has plateaus of hue but changes smoothly in luminance
C) Changes smoothly in hue and luminance
D) Has peaks of luminance and plateaus of hue
7. Perceptually uniform palettes:
A) Aim to introduce spurious features that are not present in the data
B) Are created by linear changes in the values of the red, green and blue
channels
C) Aim to produce smooth changes in perceived luminance and hue
D) Look identical even to individuals who are colourblind
8. Transparency can be set through a channel referred to as:
A) gamma
B) alpha
C) beta
D) hexadecimal
9. A vector image:
A) Stores the value of every pixel
B) Stores a description of lines, text and shapes that make up a figure
C) Has a fixed resolution (number of pixels)
D) Will tend to be larger than a high quality raster image
10. Violin plots show:
A) The raw data points
B) The raw data and a distribution
C) Two mirrored distributions
D) The mean and an error bar only
398 CHAPTER 18. PLOTTING GRAPHS AND DATA VISUALISATION

Answers to all questions are provided in section 20.2.

Chapter 19

Reproducible data analysis

As we have mentioned several times throughout the book, many researchers

have become more aware in recent years that there are substantial issues with
contemporary scientific practice. In particular, modern research can be extremely
complex, involving analysis pipelines with many stages to go from the raw data
to the graphs and summary statistics presented in a final publication. The
reader of a paper must then take it on trust that the authors have conducted
the analysis correctly and competently, and that the reporting is accurate and
unbiased.
In the interests of transparency, many advocates of open science have begun to
make raw data and analysis scripts publicly available, with the intention that
others could reproduce their analysis independently (note that reproducibility is
distinct from replication, which is where other researchers run an experiment
themselves to see if they get the same result). This may seem like a scary
prospect - most scientists are not trained programmers, and are often concerned
that there might be errors in their code that would invalidate some analysis, or
even that the code looks messy and inelegant, even if it works.
However, these worries aside, sharing data and code is a critical step in improving
scientific practice. It gives the readers and reviewers of a paper confidence that
the analysis is as described in the paper, and that the data exist and have not
been manipulated inappropriately. It should ideally be possible to independently
reproduce the full analysis pipeline, by downloading and running it yourself
to confirm the findings. Grant agencies and journals increasingly require such
materials to be made available, as they realise the importance of openness for
increasing the impact of research, and improving the quality and societal impact
of science.
Open data and code allow other researchers to build on your work, perhaps
reanalysing the data for another purpose, or using parts of the code in their own
analysis. Existing data are valuable for conducting meta-analyses (see Chapter

399
400 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

6), and to inform power analyses in planning new studies (see Chapter 5). Open
practices should generally act to speed up scientific progress, and even suggest
new projects that were not previously viable. Over the past few years I have
used open data in my own research, and also had others use data that I have
made available. Both experiences have been very positive, and I have no regrets
about sharing data publicly. In particular I have been contacted by researchers
who lack laboratory facilities to collect their own data, but are able to make
progress and test hypotheses by re-using existing data sets. Such secondary data
analysis is also likely to increase citations to the original work.
The aim of this chapter is to discuss some details associated with making data
and scripts open and accessible. Some of these points will transcend specific
implementation issues, but we will also discuss how to interface with platforms
that are widely used today, in particular GitHub and the Open Science Framework
(OSF). In general, making code and data available will improve the integrity
and progress of science, and is worthwhile for researchers in all disciplines and
at all career stages.

19.1 Version control of analysis scripts

Computer scientists and professional programmers have developed robust sys-
tems for version control of code. The most widely used is a system called git.
Essentially the point of git is to keep track of every time you change some code.
This means you can roll back any changes that didn’t work, and for collaborative
work you can see who wrote each piece of code. A detailed explanation of how
to use git is beyond the scope of this book, and many other (free) resources are
available (see e.g. Chacon and Straub (2014), and Happy Git With R). However
we will briefly cover integration of RStudio and online git repositories such as
GitHub because they are an important resource for conducting open, reproducible
analyses.
The first stage is to install git on your computer. It is a standalone set of tools
that can be accessed through the command line, as well as integrating with
RStudio (and most other contemporary programming environments). Git can be
obtained in several different ways, the simplest is probably to download it from
the git website at: https://git-scm.com/downloads. You will then need to tell
RStudio where your local git installation lives, which you can do through the
Git/SVN section of the Preferences dialog window in RStudio. The location of
git will be different on different platforms (in Windows it is usually in Program
Files, and on the Mac it should be somewhere like /usr/local/git/bin/git), but
some searching online will help you to find it on your own machine.
If the project you want to work on already exists as a GitHub repository, you
can access it by cloning the repository to your own computer (if you want to
create a totally new project, the easiest way to do this is through the GitHub
website). Cloning creates a local copy of all of the files in the repository. I have
19.1. VERSION CONTROL OF ANALYSIS SCRIPTS 401

set up a public dummy example to demonstrate how this works. Within RStudio,
choose New Project. . . from the File menu. Select Version Control, followed by
Git. You will then see a dialog box asking for details of the repository. The one
I’ve created is at:
https://github.com/bakerdh/gitdemo
You should enter this URL into the dialog box. You will also need to choose a
name for the project, and a directory to save it to on your computer. Then, when
you click Create Project, all of the files from the repository will automatically be
downloaded and stored as a local copy on your computer.
At the moment the project really just has a single file called gitdemo.R, which
doesn’t do very much (it prints out a message to the console). You can make
changes to this file on your computer, and then upload the changes to the
repository. There are three stages to this process, which are all accessed through
the Git tab in the top right corner of the main RStudio window (where the
Environment usually appears).
First you can click the Diff button to stage the changes - this shows you a
colour-coded visual representation of everything that you have changed from the
latest version you downloaded, in a separate window headed Review Changes.
If you are happy with the changes you can Commit them. This means you are
saying that you are ready to upload the changes to the online repository so that
others can see them. You should also add a commit message, where you can
briefly detail the reasons for the changes. Finally, you can push the changes to
the repository using the upward pointing green arrow - this will send everything
you have altered to the GitHub server and incorporate it into the main project.
If you are working on a group project, you can also periodically pull changes
made by your collaborators using the downward pointing blue arrow. This will
update your local copy of the project. But this is where things can get tricky -
if multiple people are working on the same files, there might be conflicts, where
two people have made incompatible changes. There are tools within git to cope
with this, including making separate branches of the code that you can work on
individually, before merging them back with the master branch. Dedicated git
references go into more detail about how this works (Chacon and Straub 2014).
GitHub repositories are designed to contain computer code and other small
files, so it is not generally a good idea to store large data files in them (the
amount of storage space is usually limited for this reason). Happily, as we will
outline in Section 19.6, other repositories exist that can store your data, and
you can include code to automatically download the data to your computer.
Furthermore, the OSF website integrates well with GitHub, in that you can
associate a GitHub repository with an OSF project. The files contained in the
repository will appear as if they are files within the OSF project, meaning that
data and code are accessible from a single location.
There are two additional advantages to maintaining a GitHub repository for your
402 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

programming projects. First, your code exists independently of your computer,

and is automatically backed up every time you push your changes to the repository.
This means that if your computer breaks or gets lost or stolen, you will not
have lost your work (I have been extremely paranoid about accidentally deleting
computer code ever since I was ten years old and tried to write a BASIC program
to remove backup files, but accidentally deleted everything on the disc). The
other advantage is that, if your repository is public, other people can see how
great you are at programming. Coding is now an essential skill in many areas of
science, and potential employers (particularly for those at an early career stage)
will often want to see evidence of your programming abilities. Giving them a
link to a well-organised GitHub page might be the thing that lands you a great
job!

19.2 Writing code for others to read

As well as sharing code, it is well worth bearing a few things in mind to help
others to read and understand your code. This may seem obvious, but it is
surprising how rarely people do this (myself included). In addition to helping
others to understand your code, these suggestions might also help future versions
of yourself (by which I mean you, but in the future; I’m not suggesting you’re
going to be cloned) to remember exactly what your past code is doing. Coding
decisions that made sense when you were immersed in a project can often be
impenetrable a decade later.
First of all, it is important to give your data objects, functions and files meaningful
names. When you are in the process of writing code, it is easy to come up with
arbitrary names for data objects (a is a perennial favourite, as is data, and I
am aware that I have used both liberally throughout this book). But a much
clearer approach is to have data object names briefly describe what they contain:
rawdata and cleandata is much better than data1 and data2. It can also be
useful to come up with a naming convention that distinguishes data objects
from functions. For example, I sometimes begin function names with d_ (d for
Daniel), especially when they refer to generic functions from a ‘lab toolbox’ that
will be used in multiple scripts.
For longer object names, many programmers favour camel case, where the first
letters of each word are capitalised. So rawdata would become RawData and
cleandata would be CleanData, which can make the object names easier to parse
and avoid some ambiguities. A variant of camel case is to keep the first letter
of the object name in lower case (e.g. rawData) - this is called either lower
camel case or Dromedary case (a dromedary is a type of camel with one hump).
Remember that R is case sensitive, so the object names RawData and rawdata
refer to different data objects - it would be very confusing to use both of these
in the same script!
A second recommendation is to provide clear and detailed comments about what
19.2. WRITING CODE FOR OTHERS TO READ 403

your code is doing. In R, the hash symbol (#) and anything following it are
ignored by R, allowing you to include comments as follows:
randomSum <- 0 # initialise a data object to store a number in

# loop ten times

for (n in 1:10){

# add a random number to our data object each time around the loop
randomSum <- randomSum + rnorm(1)

You can use multiple hash symbols to introduce hierarchical structure into your
code, for example by using three hash symbols to indicate a new section of code.
I am pretty bad at commenting my own code, so this is something I need to work
on myself! But it is definitely helpful for anyone trying to understand what you
have done, or attempting to modify your code for their own needs. Beginning
your script with some comment lines giving an overview of what it aims to do is
good practice, especially in the common situation where you might have many
versions of the same piece of code which are all slightly different.
Finally, there is a balance to be struck between writing dense code with multiple
nested functions on a single line, and code that is more spread out and easier
to read. For example, the following section of code contains multiple nested
commands and a logic statement, and is therefore rather difficult to parse:
if (t.test(rnorm(100,mean=2,sd=2))$p.value<0.05){print('Test is significant')}

On the other hand, we can rewrite (or re-factor) the code by splitting things up
and de-nesting the function calls as follows:
a <- rnorm(100, mean=2, sd=2)
tout <- t.test(a)
if (tout$p.value<0.05){
print('Test is significant')}

(In both cases, the code generates some random numbers, runs a t-test on them,
and spits out some text if it is significant). Which of these approaches you
take will depend on what sort of code you are trying to write. If your aim is
to use as few lines of code as possible, and perhaps save on memory allocation
for additional data objects, the first approach might occasionally have some
advantages. If the aim is to be understood by others, the second approach is
much clearer.
If you write dense code without comment lines and meaningful object names, it
will seem like a chore to go back and add them later. Indeed, many programmers
will be reticent to do this, in case it stops the code from working! However, if
you bear these suggestions in mind when you are developing the code, then you
404 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

can embed readability into your code as you create it. This often encourages
you to think more deeply about the programming decisions you are making,
and about the structure of your code, which means you are more likely to write
better, more legible code.

19.3 Using R Markdown to combine writing

and analysis
A very useful tool that is worth knowing about is R Markdown, which is part
of the RStudio ecosystem. The idea of a markdown script is that you can
construct a document that contains both written text and computer code. As
well as formatting both types of information into a single output (such as a pdf,
Word document or webpage), the code is also executed, which means you can
perform data analysis and create figures entirely from within your document.
This makes it possible to produce a complete piece of research, including all
analyses and calculations, within a single file. This has huge advantages for
reproducibility, as your readers can download both the finished file and the
underlying markdown code. Indeed, the first draft of this book was written in
markdown. Additional features include being able to incorporate code in other
programming languages besides R (including Python, Julia and C++), and good
typesetting of equations (based on LaTeX syntax). It is not an exaggeration
to say that markdown has changed the way that I (and many others) produce
written outputs. Some scientific journals are also investigating how to embed
working code into published articles to create computationally reproducible
articles, so this approach is likely to become more widespread in the future. A
good reference to learn more is R Markdown: The Definitive Guide (Xie, Allaire,
and Grolemund 2019).

19.4 Automated package installation

One of the big strengths of R is the way that packages are handled (see Section
2.8). This is much more straightforward than in any other programming language
I have used. However there is a minor issue when writing code that will be used
by other people, or even just run on more than one of your own computers. The
first time you need to use a package on a given system, you need to install it,
for example:
install.packages('zip')

But once the package has been installed, repeating this line of code is unnecessary.
In fact it will actually re-install the same package(s), which can substantially
slow down your script, especially if there are many packages to install. As an
alternative, the following code will check the currently installed packages, and
only install those that are missing:
19.5. IDENTIFYING A ROBUST DATA REPOSITORY 405

packagelist <- c('zip') # a list of packages we want to install

# work out which packages are not currently installed
missingpackages <- packagelist[!packagelist %in% installed.packages()[,1]]
# install the missing packages only
if (length(missingpackages)>0){install.packages(missingpackages)}

We can do a similar trick for activating the packages, so this is only done for
packages that are not currently active:
# work out which packages are not currently activated
toactivate <- packagelist[which(!packagelist %in% (.packages()))]
# silently activate the inactive packages using the library function
invisible(lapply(toactivate,library,character.only=TRUE))

These five lines of code will avoid repeatedly installing and activating the same
packages, and is a reasonably compact code snippet to include at the start of a
script that will make your code more effortlessly accessible to others.

19.5 Identifying a robust data repository

It is worth noting that until relatively recently (i.e. since the turn of the mil-
lennium), suitable infrastructure to facilitate data sharing was not generally
available. If researchers wished to share their data, they would either indicate
in a paper that they would provide it upon request (usually by email), or they
might post the data files to a personal website. There are obvious problems with
both of these solutions - email addresses change, as do websites, and both would
be lost in the event of the author’s death.
In recent years, more reliable platforms for sharing data have been developed.
There are two main varieties of these. In many subject areas, there are repositories
that are intended for sharing particular types of data. For example, the websites
Neurovault and OpenNeuro store neuroimaging data, the Crystallography Open
Database is a well-organised collection of crystal structure data, and the Electron
Microscopy Public Image Archive (EMPIAR) shares electron microscopy images.
The other option is to use a more generic data archive website. Again there are
several of these, such as Figshare, DataDryad and the Open Science Framework
(OSF), all of which will archive data from any research field, and make it publicly
available. A key feature to look for is that the data are given a stable DOI (a
digital online identifier - basically a number that will always link to that data
set), as this can be included in any publications associated with the data set,
and also allow the data set to be cited directly. It is also worth checking that the
platform has contingency plans for the future that will ensure the longevity of
the data. For example, the OSF website has a preservation fund that is sufficient
to maintain hosting of their archive for over 50 years.
OSF is currently my preferred repository, because it is free, integrates well
406 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

with other systems (such as GitHub, which we discussed in Section 19.1), and
has many additional features such as hosting of preregistration documents and
preprints. This means it can be used as a repository for an entire project, hosting
the data files, analysis scripts, and other information. An ideal to strive for is
that another researcher could download all of your materials, and fully reproduce
your entire analysis pipeline, through from processing the raw data to creating
the figures and statistics reported in a paper. The structure of the OSF website
makes this goal achievable. An example repository from a recent paper of mine
(D. Baker et al. 2021) is available at: https://osf.io/ebhnk/.

19.6 Automatically uploading to and download-

ing from OSF
One of the main reasons that I am a particular fan of the OSF repository is
that they have provided an R package to automate uploading and downloading
data from the site. For uploading data, this can save a lot of time with large
data sets, and is far quicker than the standard drag-and-drop web interface. For
downloading data, it raises the possibility of other users requiring only a single
analysis script, which would then automatically download and analyse all of the
data associated with a project. Both of these are big advantages for creating
efficient and reproducible workflows.

The osfr package is hosted on GitHub, rather than the CRAN repository. We
need to use the remotes package to download and install it as follows:
install.packages('remotes')
library(remotes)
remotes::install_github("centerforopenscience/osfr")
library(osfr)

I have created a dummy OSF project to demonstrate how uploading and down-
loading works. However, in order to upload data (to a repository you own) you
would first need to authorise R to interact with your OSF account using an
access token. The instructions for how to create a token are in the settings
section of the OSF website. Once you have a token, this is entered as follows:
osf_auth(token = 'MY_TOKEN_HERE')

Of course your token will only allow you to upload to your own projects, and
importantly it should not be shared with anyone else. This means you need to
remember to remove the token if you are posting your analysis code somewhere
publicly (just as I have done above).

The dummy project is located at: https://osf.io/thm3j/. We can retrieve details

of the project ‘node’ by entering the unique 5-character part of the URL (thm3j
for this example) into the osf_retrieve_node function as follows:
19.6. AUTOMATICALLY UPLOADING TO AND DOWNLOADING FROM OSF407

osfproject <- osf_retrieve_node("thm3j")

osfproject

## # A tibble: 1 x 3
## name id meta
## <chr> <chr> <list>
## 1 Upload & Download examples thm3j <named list [3]>
The osfproject data object then contains metadata about the project, and we
can use it to list the files contained in the project with the osf_ls_files function:
filelist <- osf_ls_files(osfproject)
filelist

## # A tibble: 5 x 3
## name id meta
## <chr> <chr> <list>
## 1 File1.csv 5ef0d99265982801b4cf0c9f <named list [3]>
## 2 File2.csv 5ef0d99665982801abcf2ad5 <named list [3]>
## 3 File3.csv 5ef0d99a145b1a01cc52cc8f <named list [3]>
## 4 File4.csv 5ef0d99e65982801a8cf2b7b <named list [3]>
## 5 File5.csv 5ef0d9a2145b1a01cb52c874 <named list [3]>
From this table, we see that the project contains five csv files. I originally
uploaded these files using the following code (which of course could also be done
in a loop):
osf_upload(osfproject,'data/File1.csv')
osf_upload(osfproject,'data/File2.csv')
osf_upload(osfproject,'data/File3.csv')
osf_upload(osfproject,'data/File4.csv')
osf_upload(osfproject,'data/File5.csv')

The OSF website has automatically generated ID numbers for each file. We
can then download the files to the R working directory using the osf_download
function:
osf_download(filelist) # download all files in the list

Note that we are downloading all the files at once here, which might be a
dangerous strategy with large data sets! If we want to download only a single
file, we can index different rows of the filelist data object as follows:
osf_download(filelist[3,]) # download the third file in the list only

The osf_download function takes additional arguments that can be used to

specify the destination path to save the files to on your local computer.
If we want to find a particular file, we can use R’s native string matching
functions as follows:
408 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

a <- which(filelist[,1]=='File3.csv')
osf_download(filelist[a,])

The OSF repository allows additional nodes (also called components) to be

nested inside the main one, in much the same way that computer file systems
can have multiple subdirectories. A list of sub-nodes within a given node can be
generated using the osf_ls_nodes function, which is called in a similar way to
the osf_ls_files function described above. The table returned by this function is
a list of additional nodes, which can also be indexed using either of these listing
functions, and the enclosed files uploaded and downloaded. The package also
contains functions to generate components and folders (osf_create_component
and osf_mkdir). By combining these functions, we can generate online data
structures of arbitrary complexity entirely from within an R script.

19.7 Future proofing with open data formats

One common issue with data sharing is that many data acquisition devices have
a proprietary native file format. Sometimes this requires special software to
access the data, and this software may not be free to obtain. This can result
in a barrier to data sharing, as not everyone who wants to access the data
will have the resources necessary to obtain the software. It is also an issue
for future-proofing data access. Proprietary data formats tied to a particular
company may no longer exist in 20 years time, especially if the company has
ceased trading, or have replaced their software with something else that is not
backwards-compatible.
A good example from my own research area is electroencephalography (EEG)
data. There are currently dozens of EEG systems available, and each of them
has a different proprietary format for saving the resulting data. For example
the EEGlab analysis software (Delorme and Makeig 2004) includes extensions
for importing over 30 different file formats. If I were to make available the raw
EEG data files from an experiment, it might become impossible (or at least very
difficult) to open them in the future.
There are two main solutions to this problem. One option is to convert the data
to a standard format that is widely used in the field. For MRI data, the NIFTI
(Neuroimaging Informatics Technology Initiative) file format is a great example
- it is an open standard that has been widely adopted and is compatible with
all current analysis software. Crucially, the file format is described in detail
and is actively maintained by a working group that are not tied to a particular
commercial organisation. Similarly, the Brain Imaging Data Structure (BIDS)
is a standard for organising and describing neuroimaging data from multiple
modalities (Gorgolewski et al. 2016). In genetics and bioinformatics, the FASTA
format is widely used to store DNA data and has similar advantages.
The other alternative is to convert your data into a generic data format that is
19.8. PROVIDING INFORMATIVE META-DATA 409

not specific to a particular scientific field or data type. The most basic example is
to use a text file format, such as the csv (comma-separated values) format. Each
individal value in a csv file is separated by a comma, with new lines indicated by
a carriage return. This format can store both numbers and text, and the data
can be easily read by any modern programming language, as well as spreadsheet
packages such as Microsoft Excel, Google Sheets, or OpenOffice Calc. Finally,
column headings can be included as the first row of data.
A shortcoming of text files is that they are not very efficient for storing large
amounts of data. If storage space is an issue, lossless compression algorithms
(such as zip and gzip) can be used to reduce file size. Alternatively, generic
formats for structuring and storing large amounts of data also exist. One
example is the HDF5 (Hierarchical Data Format 5) specification, which can store
arbitrarily large data structures. This is an open source format, developed by
a not-for-profit group dedicated to ensuring accessibility of data stored in the
format. The hdf5r package contains tools for reading and writing in this format
within R, and there are similar libraries for other programming languages.
For several reasons, data formats associated with a specific programming language
or software package are not typically very open. For example, the widely-used
Matlab language is owned by a company (The MathWorks Inc.), who also own
the specification for the .mat file format. Although toolboxes currently exist
to import such data files into other programming environments (such as the
R.matlab toolbox in R), there is no guarantee that this will always be the case.
Similar arguments apply to Microsoft’s xls and xlsx spreadsheet formats. Overall,
making your data available in an open format is a better choice for ensuring it
is accessible to others both now and in the future.

19.8 Providing informative meta-data

In addition to being able to access and open data files, it is important for
potential users to understand what they contain. So, including useful metadata
is critical. Sometimes this can be incorporated into the data files themselves, for
example in the HDF5 format. But it can also make sense to include a separate
metadata file, describing the data format, the experimental conditions and so
on. At its simplest, this could just be a plain text file (like a readme file) that
explains the structure of the data.
I have recently begun sharing EEG data as compressed csv text files, and include
a metadata file that describes the full data set, the experimental conditions and
how these map to trigger codes, the participants, and the equipment parameters
(e.g. electrode locations). An example of the first few rows and columns might
look like this:
hdata <- read.csv('data/headerfile.csv')

head(hdata[,1:4],n=12)
410 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS

## Details1 Details2 Trigger Description

## 1 SSVEP masking experiment 2014 10 Target 0 Mask 0
## 2 Sample rate 1000Hz 20 Target 2 Mask 0
## 3 Stimulus duration 11s 30 Target 4 Mask 0
## 4 Target frequency 7Hz 40 Target 8 Mask 0
## 5 Mask frequency 5Hz 50 Target 16 Mask 0
## 6 Repetitions per condition 8 60 Target 32 Mask 0
## 7 Trials per block 40 70 Target 64 Mask 0
## 8 Total participants 100 80 Target 0 Mask 32
## 9 Number of electrodes 64 90 Target 2 Mask 32
## 10 Montage layout 5 percent 100 Target 4 Mask 32
## 11 EEG system ANT Neuroscan 110 Target 8 Mask 32
## 12 Original file format ANT EEprobe .cnt files 120 Target 16 Mask 32

Some of this information could be read in by a script, and used to automate

parts of the analysis. The aim of including metadata should be that another
researcher could use it, along with the description of the experiment in an
associated publication, in order to work out how to re-analyse your data. Two
added bonuses: if you manage to do this you are much less likely to be contacted
by other researchers asking for help and explanations, and also your future self
will have a much easier job if you ever need to go back to look at your own data
again.

19.9 Practice questions

1. Which of the following is not an advantage of sharing analysis code?
A) Other researchers can see what you did without having to contact
you directly
B) If you share code, it allows you to access other people’s shared code
C) If there are mistakes in the code, other researchers might find them
for you
D) Future work can more easily build on the work you have done
2. One problem with analysis packages that involve graphical interfaces
instead of writing code is that:
A) It is often hard to reproduce exactly the options that were selected
for the analysis
B) There is only one possible way to analyse a given data set
C) The calculations are less accurate than they would be in a program-
ming language
D) There is no way to save the results of the analysis
3. An important feature to look for in a data repository is:
A) That it is maintained by a commercial company
B) That it stores only data from the particular type of experiment you
have run
C) That they will remove your data if you die
19.9. PRACTICE QUESTIONS 411

D) A stable digital online identifier is created for each data set

4. An OSF personal access token will allow you to use R to:
A) Upload to your own projects
B) Download from public projects
C) Upload to public projects
D) Access all projects on the site
5. Which of the following are both open data formats?
A) HDF5 and xlsx
B) csv and .mat
C) csv and HDF5
D) .mat and xlsx
6. When using git, a pull request is where you:
A) Upload your changes to the repository
B) Download other programmers’ changes from the repository
C) Create a new branch of the code
D) Integrate any changes into the main branch
7. In git, the order of the three stages to upload some changes is:
A) Clone, Push and Pull
B) Pull, Push and Commit
C) Commit, Pull and Stage
D) Stage, Commit and Push
8. Which of the following data object names is in lower camel case?
A) AllData
B) allData
C) alldata
D) Alldata
9. In R, comments are indicated by:
A) #
B) %
C) ’ ’ (the comment is enclosed in quote marks)
D) ∗
10. What is the difference between reproducibility and replication?
A) Replication is where other researchers try and produce the same result
as you but using different methods, whereas with reproducibility they
use the same methods
B) Reproducibility is where other researchers try and produce the same
result as you but using different methods, whereas with replication
they use the same methods
C) Reproducibility means that others can recreate your analysis pipeline
(using your data), whereas replication is where they try to repeat
your experiment
D) Replication means that others can recreate your analysis pipeline
(using your data), whereas reproducibility is where they try to repeat
your experiment
Answers to all questions are provided in section 20.2.
412 CHAPTER 19. REPRODUCIBLE DATA ANALYSIS
Chapter 20

Endnotes

20.1 A parting note

I hope that you have found at least some parts of this book useful. As discussed
in Chapter 1, the aim was to provide enough information to grasp and implement
the basics of several techniques. Hopefully this will serve as a starting point for
your own further reading and investigation of the methods. Inevitably there
are many things I have not included here, and I welcome any feedback and
suggestions for new material in future editions. Remember that the CRAN
repository contains many thousands of packages to investigate, and that the
code to produce all figures in this book is available in a GitHub repository at:
https://github.com/bakerdh/ARMbookOUP
As your abilities using R increase, you will certainly run into problems, many of
which will take time and patience to solve. As mentioned in Chapter 1, the R
community is generally very supportive and inclusive, and there is a wealth of
useful information available online. This includes discussion boards, blog posts
and tutorials created by other users, and formal groups like RLadies and the
local R user groups coordinated by the R consortium. You could also set up
an R interest group in your own town, workplace, or University department to
build a peer group of R users who can help each other develop their skills.
In several places I have talked about the Open Science movement, for example
discussing statistical power in Chapter 5, publication bias in Chapter 6, and
reproducibility and data sharing in Chapter 19. My hope is that open research
practices will continue to become embedded in the way science is conducted across
all fields and disciplines. Future readers of this book might find it endearingly
quaint that I have talked explicitly about things like sharing data and code,
if these have become second nature to all scientists. However at the time of
writing, open practices are only just gaining traction, and are still applied to
only a minority of projects.

413
414 CHAPTER 20. ENDNOTES

Ultimately the scientific enterprise is about finding out how the world works.
But real data are always noisy, and rarely lead to a clear interpretation in their
raw state. To this end, the analysis methods described in this book are tools
that help us to test hypotheses and better understand empirical data. Yet the
choice of statistical tool, and the way we interpret its outcome, will often be
somewhat subjective. My aim is that by understanding the fundamentals of a
range of techniques, researchers can make more informed decisions about how to
conduct their analyses. Mastery of advanced methods allows for creative flair in
the presentation of results, and in some situations may lead to new knowledge
that would be missed by more basic analyses. Although it may not feel like
it to a beginner, it can be very satisfying to immerse oneself in data analysis
(“wallowing in your data” as a former colleague of mine used to say). I hope that
this book helps others on their way to experiencing data analysis as a pleasure
rather than a chore, and leads to many new and exciting discoveries.

20.2 Answers to practice questions

Chapter 2:
1. B - R is based on the S language, which was first released in 1976.
2. A - the Environment tab is in the upper right corner, and contains variables
currently held in memory.
3. B - although scripts are useful, the syntax of different programming lan-
guages means that you usually cannot run a script in a language other
than the one it was written for.
4. D - variable names cannot begin with a number, so 9thNumber is not
permitted.
5. C - matrices and data frames have both rows and columns, much like an
Excel spreadsheet.
6. B - square brackets are used to index variables, which distinguishes them
from function calls (which use normal brackets).
7. D - the packages installed on your computer are visible in the Packages
tab in the lower right panel of RStudio.
8. B - the loop will repeat five times, with the variable n having the values
31, 32, 33, 34 and 35.
9. D - the loop will repeat three times, so the end value of a will be 0 + 4
+ 4 + 4 = 12. This means that the first and fourth if statements will be
true: a>10 and a==12. So the code will print out Bananas, followed by
Oranges.
10. D - the transpose function in R is called t. You can work this out by trying
to call the help for each of the four options, or by searching online.
20.2. ANSWERS TO PRACTICE QUESTIONS 415

Chapter 3:
1. B - The data layouts are called wide and long - see the chapter for examples.
2. C - The error bars on a boxplot usually show the inner fence. Data
points falling outside of the whiskers are classified as outliers using Tukey’s
method.
3. D - the interquartile range gives the points between which 50% of the data
points lie. This is important for some methods of identifying outliers, but
it is not used to replace outlying values.
4. D - the Mahalanobis distance is suitable for identifying outliers with
multivariate data. The other three methods are used for univariate data.
5. C - the inner fence is 1.5 times the interquartile range beyond Q1 and Q3,
which works out as 2.698 standard deviations from the mean.
6. A - a 20% trimmed mean would exclude the highest 20% and lowest 20%
of values, leaving only the central 60%.
7. B - the Kolmogorov-Smirnov test is a test of the normality assumption,
rather than an alternative to parametric methods. The other responses
are all plausible alternatives to using parametric tests.
8. C - the scale function performs normalization of the mean and standard
deviation. If we want to avoid subtracting the mean, we set the ‘centre’
argument to FALSE, as in the third answer.
9. D - all of the previous options are used to assess deviations from normality.
The q-q plot shows the quantiles of the data and reference distributions,
and the two tests give a quantitive comparison.
10. A - logarithmic transforms will squish large values closer together, so the
skewed tail of the distribution will shrink.
Chapter 4:
1. A - Gossett was employed by the Guinness corporation in Dublin as their
head brewer, to analyse data on batches of beer
2. B - the model formula uses the tilde (~) to indicate relationships, and is
always of the form DV ~ IV. Since we want to know how age predicts brain
volume, we need brainvolume to come first in the formula.
3. C - you can confirm that the other three options are not R functions by
trying to call the help function for them - they do not exist.
4. D - the residuals are the left over variance that cannot be explained by the
model. A good way to think about this is the error between the model’s
predictions and the data points.
5. C - a null regression model is flat, so it has a slope of zero. See Figure 4.2
for an example.
416 CHAPTER 20. ENDNOTES

6. D - the slope of a fitted regression line will depend on the data we are
fitting, and the strength of any relationship between the two variables.
7. B - the degrees of freedom are always one less than the number of groups.
A good way to think about this is the number of straight lines required to
join successive pairs of points (see e.g. Figure 4.5).
8. C - factors are categorical variables, used to define groups e.g. in ANOVA.
9. B - we cannot use the t.test function because we have more than two
levels, but we could use either the aov or lm functions, as illustrated in
the chapter.
10. A - the asterisk symbol is used to indicate factorial combination, and
to generate interaction terms. A plus symbol would request additive
combination, as in multiple regression.
Chapter 5:
1. B - the power is the proportion of significant tests, so 2000/10000 = 0.2.
2. C - Cohen’s d is the difference in means divided by the standard deviation,
so (15-12)/20 = 0.15.
3. C - as we collect more data, our precision in estimating the true effect size
is less subject to noise from random sampling error.
4. D - power is the probability of producing a significant effect, so a low
powered study is unlikely to do this
5. A - when power is low, only studies with large effect sizes will be significant
(see Figure 5.2), which overestimates the true effect
6. D - enter pwr.t.test(d = 0.8, sig.level = 0.05, power = 0.8,
type=‘one.sample’). The answer is 14.3, which rounds up to 15.
7. A - enter pwr.r.test(r=0.3, n=24, sig.level=0.05). The answer is 0.30.
8. B - enter pwr.anova.test(f=0.33,k=8,n=30,sig.level=0.01). The answer is
0.91.
9. C - enter pwr.chisq.test(N = 12, df = 10, power = 0.5, sig.level=0.05).
The answer is 0.88.
10. A - enter pwr.f2.test(u=2, v=12, sig.level=0.05, power=0.8). The answer
is 0.83.
Chapter 6:
1. D - enter mes(0.34,0.3,0.1,0.1,24,24), the value of d is 0.4.
2. B - enter res(0.7,n=10), the value of d is 1.96.
3. A - enter pes(0.01,30,30), the value of the odds ratio is 3.48.
4. D - enter fes(13.6,17,17), the value of g is 1.24.
20.2. ANSWERS TO PRACTICE QUESTIONS 417

5. C - enter propes(0.7,0.6,3,3), the value of the log-OR is 0.44.

6. C - enter meta.summaries(c(0.1,0.6,-0.2,0.9,1.1),c(0.2,0.3,0.1,0.4,0.5),method=‘random’),
the summary effect size is 0.371.
7. B - enter meta.summaries(c(0.1,0.6,-0.2,0.9,1.1),c(0.2,0.3,0.1,0.4,0.5),method=‘fixed’),
the summary effect size is -0.00552.
8. A - assign the output of the meta.summaries function to a variable
(e.g. called metaoutput). Then enter: metaplot(effectsizes,standarderrors,summn=metaoutput$summary,sumse=m
metaoutput$se.summaryˆ-2,xlab=‘Effect size (d)’). The diamond overlaps
the vertical dashed line, so there is no true effect.
9. A - small sample studies have lower statistical power, so in general they
are less likely to produce significant effects.
10. D - the odds for the treatment group is 10/(500-10) = 0.02040816. The
odds for the control group is 5/(400-5) = 0.01265823. The ratio of the two
numbers is 1.61.
Chapter 7:
1. B - fixed effects is the term that encompasses traditional independent
variables.
2. C - random effects involve group differences on the dependent variable, so
it is
3. D - mixed effects models are no more guaranteed to produce significant
results than any other technique, as this depends upon the data.
4. A - mixed effects models are able to estimate model parameters (e.g. slope
and intercept terms), despite some missing data.
5. A - the term defining the random effects (in the brackets) features a 0 for
the intercept, meaning that only the slopes are random.
6. B - the term defining the random effects (in the brackets) features both a
1 requesting random intercepts, and the IV term is included, specifying
random slopes.
7. C - the quantile-quantile (QQ) plot is used to check the residuals. It plots
the data quantiles against the model prediction quantiles.
8. A - the conditional R2 value gives the overall proportion of the variance
explained by the model, including both fixed and random effects.
9. B - the AIC score incorporates the inverse likelihood and also the number
of free parameters as a penalty term. Smaller scores indicate better fits.
FOr the other three statistics, larger values indicate better fits.
10. D - celebrities will have different amounts of fame and recognisability, so
image should be a random effect. Although country is a grouping variable
418 CHAPTER 20. ENDNOTES

that might be treated as a random effect, it has only two levels and is
likely to be the independent variable of interest.
Chapter 8:
1. D - pseudo-random numbers are generated by an algorithm. The computer’s
clock is entirely predictable, but can be used to seed a random number
generator. The other two options are inherently random.
2. A - regardless of the shape of the individual distributions, their sum should
always be normal (see the example in Figure 8.3).
3. B - the seed is the name for the value used by the algorithm to generate a
sequence of random numbers.
4. C - bootstrapping is primarily used to estimate confidence intervals, though
this can be applied to any statistic including the mean, median or t-statistic.
5. C - it is very difficult to distinguish between truly random and pseudo-
random numbers, and it is unlikely that this could be done with stochastic
simulations.
6. A - enter median(rgamma(100000,shape=2,scale=2)). The output will be
approximately 3.35.
7. D - we use the quantile function to request points from an empirical
distribution that correspond to particular probabilities.
8. B - the 95% confidence intervals are taken at 2.5% and 97.5% on a distri-
bution, which means that 95% of the values lie between those points.
9. A - this is a permutation of the original set, because each of the five
numbers appears once. The other three examples include either duplicate
numbers (implying resampling with replacement), or numbers that are not
in the original set.
10. C - enter hist(rpois(10000,lambda=2)). Poisson distributions have positive
skew as they are bounded at 0.
Chapter 9:
1. C - there is always one more dimension than the number of free parameters
(which represents the error between model and data).
2. B - for a three dimensional space, the x- and y-coordinates represent the
parameter values, and the height represents the error.
3. A - circulation is not something associated with simplex algorithms, but
the other three options are geometric operations that change the shape
and location of the simplex.
4. D - local minima are low regions of the error space that are not as low as
the global minimum.
20.2. ANSWERS TO PRACTICE QUESTIONS 419

5. A - because the model code needs to be repeated many times, making it

more efficient can speed up the fitting process.
6. C - since negative rates are not possible, this would be a sensible constraint.
7. D - nelder_mead can minimise arbitrary functions that we define, but the
function must return the error between the model and data.
8. B - the AIC combines the error of the fit with a penalty term that is
a multiple of the number of free parameters, so it takes the degrees of
freedom into account when comparing models.
9. B - a small RMS error indicates a good fit. The number of free parameters
will depend on a balance between model simplicity and accuracy, so there
are no firm rules about how many parameters there should be.
10. A - functional models try to say something about the processes involved,
as compared with descriptive models that aim only to summarise the data.
Chapter 10:
1. C - the inverse transform converts from the Fourier spectrum back into
the temporal/spatial domain.
2. A - the Nyquist limit is the highest frequency that can be represented, and
is always half of the sample rate.
3. C - the signal duration determines the frequency resolution, with the
minimum frequency step being determined by 1/duration. For example, a
1 second signal has a frequency resolution of 1Hz.
4. D - phase is an angular term measured in degrees (or alternatively, radians).
5. B - low pass filters allow through low frequencies, but block high frequencies.
6. C - these arguments calculate the Fourier transform, and then take the
absolute values, which discards the phase information.
7. D - convolving two signals in the temporal domain is the same as taking
the Fourier transform of both signals and multiplying them together.
8. A - the convention is to plot the DC component in the centre, with spatial
frequency increasing as a function of distance from the centre, so the
highest frequencies appear in the corners of the spectrum.
9. B - the phase information is extracted from the complex spectrum using
the angle function.
10. B - assuming that the filter has already been Fourier transformed, this line
of code calculates the Fourier transform, applies the filer, then transforms
back and takes the real values of the output.
Chapter 11:
420 CHAPTER 20. ENDNOTES

1. D - multivariate tests involve multiple dependent variables, regardless of

the number of independent variables.
2. A - Hotelling’s T 2 does not make assumptions about the relative variances
of the DVs, so their units do not have to be the same. However there
should be at least two dependent variables, otherwise you could just use a
standard t-test.
3. B - the T 2 statistic includes terms for the means and variances of each
dependent variable, as well as the covariance between them.
4. B - the F-distribution is used to determine significance, following a trans-
form of the test statistic based on the sample size and number of dependent
variables.
5. C - both N and m are important in calculating the degrees of freedom,
which are used to assess statistical significance.
6. A - if the variances are not equal or the variables are correlated, the Type
I error rate (false positive rate) will be inflated.
7. B - the Eigenvalues are the lengths, the Eigenvectors are the lines them-
selves, and the condition index is the square root of the ratio of Eigenvalues.
8. C - the pairwise Mahalanobis distance takes into account the covariance
matrices of both groups, whereas the standard Mahalanobis distance just
considers one group’s covariance when comparing it to a point.
9. D - the two-sample Hotelling’s T 2 test requires pooling the covariances
across both samples, taking into account the sample size (see eqn. (11.4)).
10. A - surprisingly, the T 2 statistic discounts any correlation between the
dependent variables, so it has no effect on statistical power.
Chapter 12:
1. C - latent variables are by definition internal variables that are unobservable.
However we can infer their existence from other observations.
2. B - goodness of fit is calculated by comparing the model covariance matrix
with the empirical covariance matrix.
3. B - measured variables are always squares or rectangles, whereas latent
variables are ovals or circles.
4. D - double headed arrows indicate the residual variance that the other
model components cannot explain.
5. C - using the equation N = p(p+1)/2 for p = 5 gives 5*(5+1)/2 = 15.
6. A - correlations are indicated by the double tilde, whereas a latent variable
is =~.
20.2. ANSWERS TO PRACTICE QUESTIONS 421

7. A - the number of data points is N = p(p+1)/2, which for p = 4 gives

4*(4+1)/2 = 10. The number of free parameters is the 4 measured variables
plus the latent variable, plus the two covariances, giving 7 free parameters.
This means that the model is over-identified (more data points than
parameters).
8. C - the root mean square error tends to zero as the model fit improves,
because it is calculated by taking the square root of the mean squared
error between the model and the empirical data. A perfect fit would have
no residual error, so the RMSEA would be zero.
9. D - the Wald test indicates which parameters can be removed.
10. A - this is a rule of thumb, but in general correlation coefficients are hard
to estimate accurately with N<200.
Chapter 13:
1. D - Euclidean distances are the shortest straight lines between the points.
2. A - the residual lines are used to calculate the goodness of fit, as shown in
Figure 13.2.
3. B - the Akaike information criterion (AIC) includes a penalty term for the
number of free parameters (clusters).
4. C - rescaling ensures that each dependent variable is approximately equal
in its contribution to the distances between points.
5. C - I made up inverse k-means clustering, so as far as I know it isn’t a real
thing!
6. A - MDS is a visualisation technique that collapses multidimensional data
to 2 or 3 dimensions for plotting.
7. B - Shepard diagrams plot the pairwise distances in the original space
against those in the rescaled space. If the values are highly correlated this
suggests that the algorithm has preserved the ordering of distances.
8. D - metric algorithms care about the distances between points, whereas
non-metric algorithms just work on the rank ordered distances.
9. D - the distance matrix does not depend on the number of dimensions (m),
just the number of observations (N ). There will be an entry in the matrix
for each pairwise combination of points, so it will have size N x N.
10. D - although technically possible, it does not make sense to produce a
solution with more dimensions than in the original data set.
Chapter 14:
1. B - the guess rate is 1/n, which for 4 categories is 1/4 or 25%.
422 CHAPTER 20. ENDNOTES

2. A - Cohen’s d is a measure of effect size calculated as the difference

between means scaled by the standard deviation. As it increases, the group
differences get bigger and so classification becomes easier (and therefore
more accurate).
3. B - linear classifiers use straight lines to partition data, but with many
dimensions this will become a hypoerplane.
4. D - there are versions of MVPA that use correlation instead of classification
to identify consistent patterns in data.
5. C - normalization involves subtracting the mean and scaling by the vari-
ances. This will often speed up classification and make it more accurate.
6. C - neural networks have hidden (sometimes called ‘deep’) layers that
convert the inputs to outputs using various mathematical operations.
7. B - overfitting inflates accuracy, but it will not necessarily be perfect
because of overlap between the data across categories (see Figure 14.1).
8. A - as with ANOVA, we can follow up the omnibus test (involving all 3
categories) using pairwise classifications between different groups.
9. B - support vectors are calculated using the closest few points to the
decision boundary, but not those that are futher away.
10. D - the structure of biological visual and auditory systems are the inspira-
tion for deep neural networks. Note that this structure does often involve
banks of filters, but these are not the primary motivation.
Chapter 15:
1. A - the familywise error rate is calculated as 1 − (1 − α)m , which here is
1 − (1 − 0.01)6 = 0.059.
α 0.05
2. B - this is calculated as m = 20 = 0.0025.
α 0.05
3. C - this is calculated as m−(k−1) = 12−(9−1) = 0.0125.

4. D - adjusting the alpha level reduces power. However Sidak correction

fixes the Type I error rate so that it does not change.
5. A - the FDR is the proportion of discoveries that are false. Since Type I
errors are false positives, and only significant tests are counted as discoveries,
option A is the best description.
q∗ k 0.05×50
6. B - this is calculated as m = 88 , which is option B.
7. C - cluster correction makes sense when there are some correlations between
conditions, so is only well-suited to repeated measures type designs. If only
one outcome measure is contributed by each participant, cluster correction
would probably not be suitable.
20.2. ANSWERS TO PRACTICE QUESTIONS 423

8. D - the summed test-statistic (i.e. t-value) is used to calculate cluster

statistics.
9. A - the cluster correction method avoids cluster shrinkage due to adjusting
the α level.
10. B - null distributions are generated by randomly permuting the conditions
for each participant. This is done many times for the largest cluster,
to build an empirical null distribution that can be used to determine
significance for each cluster.
Chapter 16:
1. D - if the signal is absent and we respond ‘no’, we have ‘correctly rejected’
the possibility that the signal was present.
2. B - if the signal is present and we respond ‘no’, we have missed the signal,
so this is called a Miss.
3. A - this is calculated as hits/(hits+misses) = 28/(28+15) = 0.65.
4. C - although the statistic is closely related to Cohen’s d, and is calculated
using z-scores, the sensitivity index is known as d′ .
5. B - d′ is the difference between z-scores for hits and false alarms.
6. B - find the value of 0.9 on the y-axis of the graph, and trace along until
this intercepts the curve. Then trace down to find the appropriate value
on the x-axis - it is between 1 and 2, so the closest answer is x=1.3.
7. D - the ROC curve plots hits against false alarms for a range of criteria.
8. B - for a design to be true 2AFC there must always be two stimuli to
choose between. Yes/no paradigms have two responses, but only a single
stimulus.
9. C - to calculate d′ we need the hit rate (0.95) and the false alarm rate.
The false alarm rate is 1 - correct rejections, so it is 1 - 0.8=0.2. Then
enter qnorm(0.95) - qnorm(0.2) to get the answer 2.49.
10. A - enter qnorm(0.76) * sqrt(2) to get the answer 0.9988627, which rounds
to 1.
Chapter 17:
1. C - a significant p-value means we are very unlikely to have observed the
data if the null hypothesis were true.
2. C - frequentist methods have a fixed Type I error rate, determined by the
α level (often 0.05).
3. B - the prior describes our existing knowledge about a situations.
4. A - Bayes theorem describes how conditional probabilities are combined
and calculated, for example P (A|B) - the probability of A given B.
424 CHAPTER 20. ENDNOTES

5. D - in the subscript 0 indicates the null hypothesis, and 1 indicates the

alternative hypothesis. So BF01 is the ratio of probabilities of null over
experimental.
6. A - from Jeffrey’s (1961) approximation, Bayes Factor (BF10 ) values above
3 indicate some evidence in support of the alternative hypothesis (see Table
17.1).
7. B - Bayes factors decrease as a function of sample size when there is no
true effect (see Figure 17.6).
8. A - Bayesian tests can provide evidence in support of the null hypothesis,
when an experiment has appropriate power.
9. C - calculate as 0.67/0.02, the ratio is 33.5.
P (B|A)P (A)
10. D - from equation (17.1): P (A | B) = P (B) , which here will be
P (moustache|N ov)P (N ov)
P (N ov | moustache) = P (moustache) . Plugging in the numbers
30
0.3× 365
P (N ov | moustache) = 0.05 = 0.49.
Chapter 18:
1. A - log units can be quite hard to interpret. The tick marks will usually be
in equal log steps, and only positively skewed distributions appear more
normal. The choice of base number does not affect the appearance, but
does affect the units.
2. D - starting an axis closer to the smallest value can make the differences
between conditions appear larger without changing the data.
3. B - standard
p errors are smaller than standard deviations because SE =
SD/ (N ). Standard errors will be smaller than 95% confidence intervals
because CI = 1.96 × SE.
4. A - individual data helps to reassure the reader that there are no problems
with outliers, or when there are the outliers become clear and explicit.
5. D - tritanopia occurs when the blue photoreceptors are absent, making it
hard to distinguish between blue and green hues.
6. C - the parula palette is designed to change smoothly in both hue and
luminance.
7. C - the idea of a perceptually uniform palette is that the perceived colour
changes smoothly, even if this involves nonlinear changes to the output of
the three colour channels.
8. B - the alpha channel codes transparency information.
9. B - vector files contain a description of the components that make up a
figure. They are typically smaller than their raster equivalents, and do not
have a fixed resolution or an explicit representation of pixels.
20.3. ALPHABETICAL LIST OF KEY R PACKAGES USED IN THIS BOOK425

10. C - violin plots show a kernel density function mirrored about its midpoint.
They are often said to look rather suggestive (see https://xkcd.com/1967/).
Chapter 19:
1. B - if code is shared publicly, it can be accessed by anyone, so there is
no need to have shared your own code. Note that answer C might seem
like a disadvantage in some ways, but really it is a good thing to correct
mistakes, even if it requires corrections to a publication.
2. A - with some complex analyses, graphical interfaces often have so many
options that it can be difficult to precisely reproduce the original choices
(even if you are the one who made them!).
3. D - digital online identifiers are important to ensure that it will always be
possible to find the data even if website URLs change in the future.
4. A - the PAT gives R the permission to access your own OSF projects
so that you can upload and download data. Public projects are already
accessible without requiring the PAT, and there is no permission setting
that will allow you to access other private projects belonging to other users.
5. C - HDF5 and csv files are both open data formats, whereas .mat and xlsx
are proprietary.
6. B - A pull request downloads the current version, including any changes
made by other programmers working on the project.
7. D - we first stage the changes to check differences from the main branch,
then we commit those changes, and finally push them to the repository.
8. B - in lower camel case the first letter is always lower case, but subsequent
words start with a captial letter.
9. A - the hash symbol (#) is used to indicate comments in R (other pro-
gramming languages use different characters, for example Matlab uses a %
symbol).
10. C - reproducibility is about repeating an analysis on a given data set,
whereas replication is about repeating an experiment.

20.3 Alphabetical list of key R packages used in

this book
BayesFactor - package for running Bayesian versions of various statistical tests,
used in Chapter 17. Key functions include ttestBF, anovaBF and lmBF.
caret - package containing functions for Classification And REgression Training,
used in Chapter 14. Key functions include train and predict.
426 CHAPTER 20. ENDNOTES

compute.es - tools for calculating and converting effect sizes, used in Chapter 6.
Key functions include des, mes, pes and tes.
grImport - tools to import and manipulate vector graphics files, used in Chapter
18. The key functions are PostScriptTrace, readPicture and grid.picture.
FourierStats - functions to conduct Hotelling’s T 2 and Tcirc
2
tests, used in Chapter
11. Key functions include tsqh.test, tsqc.test, CI.test and pairwisemahal.
jpeg - package for loading in JPEG images, used in Chapters 10 and 18. The
package contains two functions, readJPEG and writeJPEG.
lavaan - Latent Variable Analysis package for running structural equation
models, used in Chapter 12. The key functions are cfa, lavTestScore and
lavTestWald.
MAd - meta-analysis with mean differences, including tools for calculating effect
sizes, used in Chapter 6. Key functions include r_to_d, t_to_d and or_to_d.
MASS - helper package from the textbook Modern Applied Statistics with S
(Venables, Ripley, and Venables 2002). Contains the isoMDS function for non-
metric multidimensional scaling used in Chapter 13, as well as the Shepard
function to produce Shepard plots.
pals - tools for creating and evaluating colourmaps and palettes for plotting,
used in Chapter 18. The kovesi palettes are very useful, the pal.safe function
simulates colourblindness, and the pal.test function can be used to evaluate a
palette.
pracma - package containing many practical mathematical functions. This
includes the nelder_mead function to implement the Nelder and Mead (1965)
downhill simplex algorithm, described in Chapter 9.
PRISMAstatement - functions for creating PRISMA diagrams for use in meta-
analysis, used in Chapter 6. The prisma_graph function generates a diagram.
psyphy - functions for analysing psychophysical data, including d′ calculation
with the dprime.mAFC function, as used in Chapter 16.
pwr - package of power analysis functions used in Chapter 5. Key functions
include pwr.t.test, pwr.r.test and pwr.anova.test.
quickpsy - package for fitting psychometric functions, used in Chapter 16. The
main function is also called quickpsy.
remotes - tools for installing packages from repositories such as GitHub, as
demonstrated in Chapter 19. The key function is install_github.
rmeta - meta anlaysis tools used in Chapter 6. Key functions include
meta.summaries, metaplot and funnelplot.
semPlot - package used to produce graphical representations of structural equa-
tion models, used in Chapter 12. The key function is semPaths.
20.3. ALPHABETICAL LIST OF KEY R PACKAGES USED IN THIS BOOK427

signal - package for signal processing, the fir1 function was used to construct
filters in Chapter 10.
428 CHAPTER 20. ENDNOTES
Chapter 21

References

Akaike, H. 1974. “A New Look at the Statistical Model Identification.” IEEE

Transaction on Automatic Control 19: 716–23.
Allefeld, Carsten, and John-Dylan Haynes. 2014. “Searchlight-Based Multi-Voxel
Pattern Analysis of fMRI by Cross-Validated Manova.” Neuroimage 89 (April):
345–57. https://doi.org/10.1016/j.neuroimage.2013.11.043.
Allen, Micah, Davide Poggiali, Kirstie Whitaker, Tom Rhys Marshall, and
Rogier A Kievit. 2019. “Raincloud Plots: A Multi-Platform Tool for Robust
Data Visualization.” Wellcome Open Res 4: 63. https://doi.org/10.12688/
wellcomeopenres.15191.1.
Allison, Paul D. 2003. “Missing Data Techniques for Structural Equation
Modeling.” J Abnorm Psychol 112 (4): 545–57. https://doi.org/10.1037/0021-
843X.112.4.545.
Amrhein, Valentin, Sander Greenland, and Blake McShane. 2019. “Scientists
Rise up Against Statistical Significance.” Nature 567 (7748): 305–7. https:
//doi.org/10.1038/d41586-019-00857-9.
Anscombe, F. J. 1973. “Graphs in Statistical Analysis.” The American Statisti-
cian 27 (1). Taylor & Francis: 17–21. https://doi.org/10.1080/00031305.1973.
10478966.
Baayen, R.H., D.J. Davidson, and D.M. Bates. 2008. “Mixed-Effects Modeling
with Crossed Random Effects for Subjects and Items.” Journal of Memory and
Language 59 (4): 390–412. https://doi.org/https://doi.org/10.1016/j.jml.2007.
12.005.
Baker, D.H. 2021. “Statistical Analysis of Periodic Data in Neuroscience.”
Neurons, Behavior, Data Analysis and Theory. http://arxiv.org/abs/2101.04408.
Baker, D.H., F.A. Lygo, T.S. Meese, and M.A. Georgeson. 2018. “Binocular

429
430 CHAPTER 21. REFERENCES

Summation Revisited: Beyond Sqrt(2).” Psychological Bulletin 144 (11): 1186–99.

https://doi.org/10.1037/bul0000163.
Baker, D.H., and B. Richard. 2019. “Dynamic Properties of Internal Noise
Probed by Modulating Binocular Rivalry.” PLoS Comput Biol 15 (6): e1007071.
https://doi.org/10.1371/journal.pcbi.1007071.
Baker, D.H., G. Vilidaite, F.A. Lygo, A.K. Smith, T.R. Flack, A.D. Gouws, and
T.J. Andrews. 2021. “Power Contours: Optimising Sample Size and Precision
in Experimental Psychology and Human Neuroscience.” Psychological Methods
26 (3): 295–314. https://doi.org/10.1037/met0000337.
Bates, D.M., Martin Mächler, Ben Bolker, and Steve Walker. 2015. “Fitting
Linear Mixed-Effects Models Using Lme4.” Journal of Statistical Software 67 (1).
Foundation for Open Access Statistic. https://doi.org/10.18637/jss.v067.i01.
Begley, C Glenn, and Lee M Ellis. 2012. “Drug Development: Raise Standards
for Preclinical Cancer Research.” Nature 483 (7391): 531–3. https://doi.org/10.
1038/483531a.
Belia, Sarah, Fiona Fidler, Jennifer Williams, and Geoff Cumming. 2005.
“Researchers Misunderstand Confidence Intervals and Standard Error Bars.”
Psychol Methods 10 (4): 389–96. https://doi.org/10.1037/1082-989X.10.4.389.
Benedek, Mathias, Barbara Wilfling, Reingard Lukas-Wolfbauer, Björn H Katzur,
and Christian Kaernbach. 2010. “Objective and Continuous Measurement of
Piloerection.” Psychophysiology 47 (5): 989–93. https://doi.org/10.1111/j.1469-
8986.2010.01003.x.
Benjamin, Daniel J., James O. Berger, Magnus Johannesson, Brian A. Nosek,
E.-J. Wagenmakers, Richard Berk, Kenneth A. Bollen, et al. 2017. “Redefine
Statistical Significance.” Nature Human Behaviour 2 (1). Springer Science;
Business Media LLC: 6–10. https://doi.org/10.1038/s41562-017-0189-z.
Benjamini, Y, and Y Hochberg. 1995. “Controlling the False Discovery Rate:
A Practical and Powerful Approach to Multiple Testing.” Journal of the Royal
Statistical Society. Series B (Methodological) 57 (1): 289–300. http://www.jstor.
org/stable/2346101.
Benjamini, Yoav, and Daniel Yekutieli. 2001. “The Control of the False Discovery
Rate in Multiple Testing Under Dependency.” The Annals of Statistics 29 (4).
Institute of Mathematical Statistics: 1165–88. https://doi.org/10.1214/aos/
1013699998.
Biederman, I, and M.M. Shiffrar. 1987. “Sexing Day-Old Chicks: A Case Study
and Expert Systems Analysis of a Difficult Perceptual-Learning Task.” Journal
of Experimental Psychology: Learning, Memory and Cognition 13 (4): 640–45.
Bijmolt, Tammo H.A., Michel Wedel, and Wayne S. DeSarbo. 2020. “Adaptive
Multidimensional Scaling: Brand Positioning Based on Decision Sets and Dissim-
ilarity Judgments.” Customer Needs and Solutions. Springer Science; Business
 431

Media LLC. https://doi.org/10.1007/s40547-020-00112-7.

Bishop, Christopher M. 2006. Pattern Recognition and Machine Learning. New
York: Springer. http://www.loc.gov/catdir/enhancements/fy0818/2006922522-
d.html.
Bland, M. 2000. An Introduction to Medical Statistics. Oxford University Press.
Blohm, Gunnar, Konrad P Kording, and Paul R Schrater. 2020. “A How-
to-Model Guide for Neuroscience.” eNeuro 7 (1). https://doi.org/10.1523/
ENEURO.0352-19.2019.
Bollen, Kenneth A., James B. Kirby, Patrick J. Curran, Pamela M. Paxton,
and Feinian Chen. 2007. “Latent Variable Models Under Misspecification:
Two-Stage Least Squares (2SLS) and Maximum Likelihood (Ml) Estimators.”
Sociological Methods & Research 36 (1). SAGE Publications: 48–86. https:
//doi.org/10.1177/0049124107301947.
Borland, David, and M Russell Taylor 2nd. 2007. “Rainbow Color Map (Still)
Considered Harmful.” IEEE Comput Graph Appl 27 (2): 14–17. https://doi.org/
10.1109/mcg.2007.323435.
Brodeur, Mathieu B, Emmanuelle Dionne-Dostie, Tina Montreuil, and Martin
Lepage. 2010. “The Bank of Standardized Stimuli (Boss), a New Set of 480
Normative Photos of Objects to Be Used as Visual Stimuli in Cognitive Research.”
PLoS One 5 (5): e10773. https://doi.org/10.1371/journal.pone.0010773.
Button, K.S., J.P.A. Ioannidis, C. Mokrysz, B.A. Nosek, J. Flint, E.S.J. Robinson,
and M.R. Munafò. 2013. “Power Failure: Why Small Sample Size Undermines
the Reliability of Neuroscience.” Nature Reviews Neuroscience 14 (5): 365–76.
https://doi.org/10.1038/nrn3475.
Bürkner, Paul-Christian. 2017. “brms: An R Package for Bayesian Multilevel
Models Using Stan.” Journal of Statistical Software 80 (1). https://doi.org/10.
18637/jss.v080.i01.
———. 2018. “Advanced Bayesian Multilevel Modeling with the R Package brms.”
The R Journal 10 (1): 395–411. https://doi.org/10.32614/RJ-2018-017.
Cangur, Sengul, and Ilker Ercan. 2015. “Comparison of Model Fit Indices Used
in Structural Equation Modeling Under Multivariate Normality.” Journal of
Modern Applied Statistical Methods 14 (1): 152–67. https://doi.org/10.22237/
jmasm/1430453580.
Chacon, S., and B. Straub. 2014. Pro Git. 2nd ed. Apress. https://git-
scm.com/book/en/v2.
Cham, Heining, Evgeniya Reshetnyak, Barry Rosenfeld, and William Breitbart.
2017. “Full Information Maximum Likelihood Estimation for Latent Variable
Interactions with Incomplete Indicators.” Multivariate Behav Res 52 (1): 12–30.
https://doi.org/10.1080/00273171.2016.1245600.
432 CHAPTER 21. REFERENCES

Chauvenet, W. 1863. A Manual of Spherical and Practical Astronomy Vii.

Dover.
Cleveland, William S., and Robert McGill. 1983. “A Color-Caused Optical
Illusion on a Statistical Graph.” The American Statistician 37 (2). [American
Statistical Association, Taylor & Francis, Ltd.]: 101–5. http://www.jstor.org/
stable/2685868.
Coggan, David D, Daniel H Baker, and Timothy J Andrews. 2016. “The Role of
Visual and Semantic Properties in the Emergence of Category-Specific Patterns
of Neural Response in the Human Brain.” eNeuro 3 (4). https://doi.org/10.
1523/ENEURO.0158-16.2016.
Coggan, David D, Afrodite Giannakopoulou, Sanah Ali, Burcu Goz, David M
Watson, Tom Hartley, Daniel H Baker, and Timothy J Andrews. 2019. “A Data-
Driven Approach to Stimulus Selection Reveals an Image-Based Representation
of Objects in High-Level Visual Areas.” Hum Brain Mapp 40 (16): 4716–31.
https://doi.org/10.1002/hbm.24732.
Cohen, J. 1988. Statistical Power Analysis for the Behavioral Sciences. Lawrence
Erlbaum, Hillsdale, New Jersey.
Colegrave, N., and G. D. Ruxton. 2020. Power Analysis: An Introduction for
the Life Sciences. Oxford Biology Primers. Oxford University Press.
Cramer, Angélique O J, Don van Ravenzwaaij, Dora Matzke, Helen Stein-
groever, Ruud Wetzels, Raoul P P P Grasman, Lourens J Waldorp, and
Eric-Jan Wagenmakers. 2016. “Hidden Multiplicity in Exploratory Multi-
way Anova: Prevalence and Remedies.” Psychon Bull Rev 23 (2): 640–7.
https://doi.org/10.3758/s13423-015-0913-5.
Crameri, Fabio, Grace E. Shephard, and Philip J. Heron. 2020. “The Misuse of
Colour in Science Communication.” Nature Communications 11 (1). Springer
Science; Business Media LLC. https://doi.org/10.1038/s41467-020-19160-7.
Crowley, P, I Chalmers, and M J Keirse. 1990. “The Effects of Corticosteroid
Administration Before Preterm Delivery: An Overview of the Evidence from
Controlled Trials.” Br J Obstet Gynaecol 97 (1): 11–25. https://doi.org/10.
1111/j.1471-0528.1990.tb01711.x.
Curry, Haskell B. 1944. “The Method of Steepest Descent for Non-Linear
Minimization Problems.” Quarterly of Applied Mathematics 2 (3). American
Mathematical Society (AMS): 258–61. https://doi.org/10.1090/qam/10667.
De Fauw, Jeffrey, Joseph R Ledsam, Bernardino Romera-Paredes, Stanislav
Nikolov, Nenad Tomasev, Sam Blackwell, Harry Askham, et al. 2018. “Clinically
Applicable Deep Learning for Diagnosis and Referral in Retinal Disease.” Nat
Med 24 (9): 1342–50. https://doi.org/10.1038/s41591-018-0107-6.
Delorme, Arnaud, and Scott Makeig. 2004. “EEGLAB: An Open Source
Toolbox for Analysis of Single-Trial EEG Dynamics Including Independent
 433

Component Analysis.” J Neurosci Methods 134 (1): 9–21. https://doi.org/10.

1016/j.jneumeth.2003.10.009.
Dienes, Zoltan, and Neil Mclatchie. 2018. “Four Reasons to Prefer Bayesian
Analyses over Significance Testing.” Psychon Bull Rev 25 (1): 207–18. https:
//doi.org/10.3758/s13423-017-1266-z.
Dong, Ensheng, Hongru Du, and Lauren Gardner. 2020. “An Interactive Web-
Based Dashboard to Track Covid-19 in Real Time.” Lancet Infect Dis 20 (5):
533–34. https://doi.org/10.1016/S1473-3099(20)30120-1.
Efron, Bradley, and Robert Tibshirani. 1993. An Introduction to the Boot-
strap. Vol. 57. New York: Chapman & Hall. http://www.loc.gov/catdir/
enhancements/fy0730/93004489-d.html.
Eklund, Anders, Thomas E Nichols, and Hans Knutsson. 2016. “Cluster Failure:
Why fMRI Inferences for Spatial Extent Have Inflated False-Positive Rates.”
Proc Natl Acad Sci U S A 113 (28): 7900–7905. https://doi.org/10.1073/pnas.
1602413113.
Evans, Karla K, Tamara Miner Haygood, Julie Cooper, Anne-Marie Culpan, and
Jeremy M Wolfe. 2016. “A Half-Second Glimpse Often Lets Radiologists Identify
Breast Cancer Cases Even When Viewing the Mammogram of the Opposite
Breast.” Proc Natl Acad Sci U S A 113 (37): 10292–7. https://doi.org/10.1073/
pnas.1606187113.
Fahlman, A, M Brodsky, R Wells, K McHugh, J Allen, A Barleycorn, J C
Sweeney, D Fauquier, and M Moore. 2018. “Field Energetics and Lung Function
in Wild Bottlenose Dolphins, Tursiops Truncatus, in Sarasota Bay Florida.” R
Soc Open Sci 5 (1): 171280. https://doi.org/10.1098/rsos.171280.
Faul, F., E. Erdfelder, A. Buchner, and A-G. Lang. 2009. “Statistical Power
Analyses Using G*Power 3.1: Tests for Correlation and Regression Analyses.”
Behav Res Methods 41 (4): 1149–60. https://doi.org/10.3758/BRM.41.4.1149.
Faul, F., E. Erdfelder, A-G. Lang, and A. Buchner. 2007. “G*Power 3: A
Flexible Statistical Power Analysis Program for the Social, Behavioral, and
Biomedical Sciences.” Behav Res Methods 39 (2): 175–91. https://doi.org/10.
3758/BF03193146.
Fieberg, John R, Kelsey Vitense, and Douglas H Johnson. 2020. “Resampling-
Based Methods for Biologists.” PeerJ 8: e9089. https://doi.org/10.7717/peerj.
9089.
Field, A., J. Miles, and Z. Field. 2012. Discovering Statistics Using R. Sage.
Field, Andy P, and Raphael Gillett. 2010. “How to Do a Meta-Analysis.”
Br J Math Stat Psychol 63 (Pt 3): 665–94. https://doi.org/10.1348/
000711010X502733.
Field, D J. 1987. “Relations Between the Statistics of Natural Images and
the Response Properties of Cortical Cells.” J Opt Soc Am A 4 (12): 2379–94.
434 CHAPTER 21. REFERENCES

https://doi.org/10.1364/josaa.4.002379.
Fisher, Ronald A. 1926. “The Arrangement of Field Experiments.” Journal
of the Ministry of Agriculture 33. Ministry of Agriculture; Fisheries: 503–15.
https://doi.org/10.23637/ROTHAMSTED.8V61Q.
Forgy, E.W. 1965. “Cluster Analysis of Multivariate Data: Efficiency Vs Inter-
pretability of Classifications.” Biometrics 21: 768–69.
Fox, J., and S. Weisberg. 2018. An R Companion to Applied Regression. 3rd ed.
SAGE Publications Inc.
Friston, Karl. 2012. “Ten Ironic Rules for Non-Statistical Reviewers.” Neuroim-
age 61 (4): 1300–1310. https://doi.org/10.1016/j.neuroimage.2012.04.018.
Friston, Karl J, Thomas Parr, Peter Zeidman, Adeel Razi, Guillaume Flandin,
Jean Daunizeau, Ollie J Hulme, et al. 2020. “Dynamic Causal Modelling of
Covid-19.” Wellcome Open Res 5: 89. https://doi.org/10.12688/wellcomeopenres.
15881.2.
Galecki, A., and T. Burzykowski. 2013. Linear Mixed-Effects Models Using R.
Springer-Verlag New York.
Glasziou, P P, and D E Mackerras. 1993. “Vitamin a Supplementation in
Infectious Diseases: A Meta-Analysis.” BMJ 306 (6874): 366–70. https://doi.
org/10.1136/bmj.306.6874.366.
Gorgolewski, Krzysztof J, Tibor Auer, Vince D Calhoun, R Cameron Craddock,
Samir Das, Eugene P Duff, Guillaume Flandin, et al. 2016. “The Brain
Imaging Data Structure, a Format for Organizing and Describing Outputs of
Neuroimaging Experiments.” Sci Data 3: 160044. https://doi.org/10.1038/sdata.
2016.44.
Green, D M, and John A Swets. 1966. Signal Detection Theory and Psy-
chophysics. Wiley.
Gronau, Quentin F., Alexander Ly, and Eric-Jan Wagenmakers. 2019. “Informed
Bayesian T-Tests.” The American Statistician 0 (0). Taylor & Francis: 1–14.
https://doi.org/10.1080/00031305.2018.1562983.
Grootswagers, Tijl, Susan G Wardle, and Thomas A Carlson. 2017. “Decoding
Dynamic Brain Patterns from Evoked Responses: A Tutorial on Multivariate
Pattern Analysis Applied to Time Series Neuroimaging Data.” J Cogn Neurosci
29 (4): 677–97. https://doi.org/10.1162/jocn_a_01068.
Harbord, Roger M, Matthias Egger, and Jonathan A C Sterne. 2006. “A Modified
Test for Small-Study Effects in Meta-Analyses of Controlled Trials with Binary
Endpoints.” Stat Med 25 (20): 3443–57. https://doi.org/10.1002/sim.2380.
Hartigan, J. A., and M. A. Wong. 1979. “Algorithm as 136: A K-Means
Clustering Algorithm.” Journal of the Royal Statistical Society. Series C
 435

(Applied Statistics) 28 (1). [Wiley, Royal Statistical Society]: 100–108. http:

//www.jstor.org/stable/2346830.
Haxby, J V, M I Gobbini, M L Furey, A Ishai, J L Schouten, and P Pietrini.
2001. “Distributed and Overlapping Representations of Faces and Objects in
Ventral Temporal Cortex.” Science 293 (5539): 2425–30. https://doi.org/10.
1126/science.1063736.
Higgins, J.P.T., and S. Green. 2011. Cochrane Handbook for Systematic Reviews
of Interventions Version 5.1.0. The Cochrane Collaboration. http://handbook.
cochrane.org.
Hill, Andrew A, Peter LaPan, Yizheng Li, and Steve Haney. 2007. “Impact of
Image Segmentation on High-Content Screening Data Quality for Sk-Br-3 Cells.”
BMC Bioinformatics 8: 340. https://doi.org/10.1186/1471-2105-8-340.
Hoenig, J.M., and D.M. Heisey. 2001. “The Abuse of Power.” The American
Statistician 55 (1): 19–24. https://doi.org/10.1198/000313001300339897.
Holland, John H. 1992. Adaptation in Natural and Artificial Systems: An Intro-
ductory Analysis with Applications to Biology, Control, and Artificial Intelligence.
1st MIT Press ed. Cambridge, Mass.: MIT Press.
Holm, S. 1979. “A Simple Sequentially Rejective Multiple Test Procedure.”
Scandinavian Journal of Statistics 6 (2). [Board of the Foundation of the
Scandinavian Journal of Statistics, Wiley]: 65–70. http://www.jstor.org/stable/
4615733.
Holzinger, K, and F Swineford. 1939. A Study in Factor Analysis: The Stability
of a Bifactor Solution. Supplementary Educational Monograph 48. University
of Chicago Press.
Hotelling, H. 1931. “The Generalization of Student’s Ratio.” The Annals of
Mathematical Statistics 2 (3). Institute of Mathematical Statistics: 360–78.
http://www.jstor.org/stable/2957535.
Huber, P.J. 2004. Robust Statistics. Wiley.
Ioannidis, J.P.A. 2008. “Why Most Discovered True Associations Are Inflated.”
Epidemiology 19 (5): 640–8. https://doi.org/10.1097/EDE.0b013e31818131e7.
Isik, Leyla, Ethan M Meyers, Joel Z Leibo, and Tomaso Poggio. 2014. “The
Dynamics of Invariant Object Recognition in the Human Visual System.” J
Neurophysiol 111 (1): 91–102. https://doi.org/10.1152/jn.00394.2013.
Jain, Neha, and Vandana Ahuja. 2014. “Segmenting Online Consumers Using
K-Means Cluster Analysis.” International Journal of Logistics Economics and
Globalisation 6 (2). Inderscience Publishers: 161. https://doi.org/10.1504/ijleg.
2014.068274.
Jeffreys, H. 1961. Theory of Probability. 3rd ed. Oxford University Press.
436 CHAPTER 21. REFERENCES

Karaboga, D. 2005. “An Idea Based on Honey Bee Swarm for Numberical
Optimization.” Technical Report TR06. Ericyes University.
Kass, R.E., and A.E. Raftery. 1995. “Bayes Factors.” Journal of the American
Statistical Association 90 (430): 773–95.
Kennedy, J., and R. Eberhart. 1995. “Particle Swarm Optimization.” Proceedings
of ICNN’95 4: 1942–8. https://doi.org/10.1109/ICNN.1995.488968.
Kingdom, F.A.A., and N. Prins. 2010. Psychophysics: A Practical Introduction.
Elsevier.
Kline, Rex B. 2015. Principles and Practice of Structural Equation Modelling.
4th ed. Guilford, New York.
Kolmogorov, A.N. 1992. Selected Works Ii: Probability Theory and Mathematical
Statistics. Edited by A.N. Shiryaev. Vol. 26. Springer Netherlands.
Kovesi, P. 2015. “Good Colour Maps: How to Design Them.” arXiv, no.
1509:03700. https://arxiv.org/abs/1509.03700.
Kruschke, John K. 2014. Doing Bayesian data analysis: a tutorial with R, JAGS,
and Stan. 2nd ed. Elsevier, Academic Press.
Kuhn, Max. 2008. “Building Predictive Models inRUsing thecaretPackage.”
Journal of Statistical Software 28 (5). Foundation for Open Access Statistic.
https://doi.org/10.18637/jss.v028.i05.
Kuznetsova, Alexandra, Per B. Brockhoff, and Rune H. B. Christensen. 2017.
“LmerTest Package: Tests in Linear Mixed Effects Models.” Journal of Statistical
Software 82 (13). Foundation for Open Access Statistic. https://doi.org/10.
18637/jss.v082.i13.
Lakens, Daniel, Federico G. Adolfi, Casper J. Albers, Farid Anvari, Matthew A.
J. Apps, Shlomo E. Argamon, Thom Baguley, et al. 2018. “Justify Your Alpha.”
Nature Human Behaviour 2 (3). Springer Science; Business Media LLC: 168–71.
https://doi.org/10.1038/s41562-018-0311-x.
Lambert, Ben. 2018. A Student’s Guide to Bayesian Statistics. SAGE Publica-
tions.
Lenhard, W., and A. Lenhard. 2016. Calculation of Effect Sizes. Dettelbach
(Germany): Psychometrica. https://doi.org/10.13140/RG.2.2.17823.92329.
Lilja, David. 2016. Linear Regression Using R: An Introduction to Data Modeling.
University of Minnesota Libraries Publishing. https://doi.org/10.24926/8668/
1301.
Linares, Daniel, and Joan López-Moliner. 2016. “quickpsy: An R Package to
Fit Psychometric Functions for Multiple Groups.” The R Journal 8 (1): 122–31.
https://doi.org/10.32614/RJ-2016-008.
 437

Liu, Yiwen, Robert Carmer, Gaonan Zhang, Prahatha Venkatraman, Skye

Ashton Brown, Chi-Pui Pang, Mingzhi Zhang, Ping Ma, and Yuk Fai Leung.
2015. “Statistical Analysis of Zebrafish Locomotor Response.” PLoS One 10
(10): e0139521. https://doi.org/10.1371/journal.pone.0139521.
Lloyd, S. 1982. “Least Squares Quantization in Pcm.” IEEE Transactions on
Information Theory 28 (2): 129–37. https://doi.org/10.1109/TIT.1982.1056489.
Lopes, António M, José P Andrade, and J A Tenreiro Machado. 2016. “Mul-
tidimensional Scaling Analysis of Virus Diseases.” Comput Methods Programs
Biomed 131 (July): 97–110. https://doi.org/10.1016/j.cmpb.2016.03.029.
Luck, Steven J, and Nicholas Gaspelin. 2017. “How to Get Statistically Signifi-
cant Effects in Any Erp Experiment (and Why You Shouldn’t).” Psychophysiology
54 (1): 146–57. https://doi.org/10.1111/psyp.12639.
Macmillan, N.A., and C.D. Creelman. 1991. Detection Theory: A User’s Guide.
Cambridge University Press.
MacQueen, J. 1967. “Proceedings of the Fifth Berkeley Symposium on Mathe-
matical Statistics and Probability.” In, edited by L.M. Le Cam and J Neyman,
281–97. Berkeley, CA: University of California Press.
Mahalanobis, P.C. 1936. “On the Generalised Distance in Statistics.” Proceedings
of the National Academy of Sciences of India 2: 49–55.
Maris, Eric, and Robert Oostenveld. 2007. “Nonparametric Statistical Testing
of EEG- and MEG-Data.” J Neurosci Methods 164 (1): 177–90. https://doi.org/
10.1016/j.jneumeth.2007.03.024.
Matsumoto, Makoto, and Takuji Nishimura. 1998. “Mersenne Twister: A 623-
Dimensionally Equidistributed Uniform Pseudo-Random Number Generator.”
ACM Transactions on Modeling and Computer Simulation 8 (1). Association for
Computing Machinery (ACM): 3–30. https://doi.org/10.1145/272991.272995.
Meteyard, Lotte, and Robert A.I. Davies. 2020. “Best Practice Guidance for
Linear Mixed-Effects Models in Psychological Science.” Journal of Memory and
Language 112: 104092. https://doi.org/10.1016/j.jml.2020.104092.
Miranda-Mendizábal, A, P Castellvı’, O Parés-Badell, J Almenara, I Alonso, M
J Blasco, A Cebrià, et al. 2017. “Sexual Orientation and Suicidal Behaviour in
Adolescents and Young Adults: Systematic Review and Meta-Analysis.” Br J
Psychiatry 211 (2): 77–87. https://doi.org/10.1192/bjp.bp.116.196345.
Moher, David, Alessandro Liberati, Jennifer Tetzlaff, and Douglas G. Altman.
2009. “Preferred Reporting Items for Systematic Reviews and Meta-Analyses:
The Prisma Statement.” PLoS Medicine 6 (7). Public Library of Science (PLoS):
e1000097. https://doi.org/10.1371/journal.pmed.1000097.
Morgan, Michael, Barbara Dillenburger, Sabine Raphael, and Joshua A Solomon.
2012. “Observers Can Voluntarily Shift Their Psychometric Functions Without
438 CHAPTER 21. REFERENCES

Losing Sensitivity.” Atten Percept Psychophys 74 (1): 185–93. https://doi.org/

10.3758/s13414-011-0222-7.
Navarro, D. 2019. Learning Statistics with R. https://learningstatisticswithr.com/.
Nelder, J. A., and R. Mead. 1965. “A Simplex Method for Function Minimization.”
The Computer Journal 7 (4): 308–13. https://doi.org/10.1093/comjnl/7.4.308.
Newman, George E, and Brian J Scholl. 2012. “Bar Graphs Depicting Averages
Are Perceptually Misinterpreted: The Within-the-Bar Bias.” Psychon Bull Rev
19 (4): 601–7. https://doi.org/10.3758/s13423-012-0247-5.
Oliva, Aude, and Antonio Torralba. 2001. “Modeling the Shape of the Scene:
A Holistic Representation of the Spatial Envelope.” International Journal
of Computer Vision 42 (3). Springer Science; Business Media LLC: 145–75.
https://doi.org/10.1023/a:1011139631724.
Open Science Collaboration. 2015. “Estimating the Reproducibility of Psycho-
logical Science.” Science 349 (6251): aac4716. https://doi.org/10.1126/science.
aac4716.
Pincus, Martin. 1970. “A Monte Carlo Method for the Approximate Solution of
Certain Types of Constrained Optimization Problems.” Operations Research 18
(6). Institute for Operations Research; the Management Sciences (INFORMS):
1225–8. https://doi.org/10.1287/opre.18.6.1225.
Pirrone, Angelo, Wen Wen, Sheng Li, Daniel H Baker, and Elizabeth Milne.
2018. “Autistic Traits in the Neurotypical Population Do Not Predict Increased
Response Conservativeness in Perceptual Decision Making.” Perception 47 (10-
11): 1081–96. https://doi.org/10.1177/0301006618802689.
Press, W.H., S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. 1986. Numer-
ical Recipes: The Art of Scientific Computing. Cambridge University Press.
R Core Team. 2013. “R: A Language and Environment for Statistical Computing.”
R Foundation for Statistical Computing, Vienna, Austria. http://www.R-project.
org/.
Rietveld, Toni, and Roeland van Hout. 2007. “Analysis of Variance for Repeated
Measures Designs with Word Materials as a Nested Random or Fixed Factor.”
Behavior Research Methods 39 (4). Springer Science; Business Media LLC:
735–47. https://doi.org/10.3758/bf03192964.
Rosenthal, R. 1979. “The File Drawer Problem and Tolerance for Null Results.”
Psychological Bulletin 86 (3): 638–41.
Rosseel, Yves. 2012. “Lavaan: An R Package for Structural Equation Modeling.”
Journal of Statistical Software 48 (2). Foundation for Open Access Statistic.
https://doi.org/10.18637/jss.v048.i02.
Rouder, Jeffrey N, Paul L Speckman, Dongchu Sun, Richard D Morey, and
Geoffrey Iverson. 2009. “Bayesian T Tests for Accepting and Rejecting the Null
 439

Hypothesis.” Psychon Bull Rev 16 (2): 225–37. https://doi.org/10.3758/PBR.

16.2.225.
Rousselet, Guillaume A, John J Foxe, and J Paul Bolam. 2016. “A Few Simple
Steps to Improve the Description of Group Results in Neuroscience.” Eur J
Neurosci 44 (9): 2647–51. https://doi.org/10.1111/ejn.13400.
Satorra, A., and W. E. Saris. 1985. “Power of the Likelihood Ratio Test in
Covariance Structure Analysis.” Psychometrika 50 (1): 83–90.
Sánchez-Meca, Julio, and Fulgencio Marin-Martinez. 2008. “Confidence Intervals
for the Overall Effect Size in Random-Effects Meta-Analysis.” Psychol Methods
13 (1): 31–48. https://doi.org/10.1037/1082-989X.13.1.31.
Schwarz, Gideon. 1978. “Estimating the Dimension of a Model.” The Annals of
Statistics 6 (2). Institute of Mathematical Statistics: 461–64. https://doi.org/
10.1214/aos/1176344136.
Schwarzkopf, D S, and G Rees. 2011. “Pattern Classification Using Functional
Magnetic Resonance Imaging.” Wiley Interdiscip Rev Cogn Sci 2 (5): 568–79.
https://doi.org/10.1002/wcs.141.
Scott Chialvo, Clare H, Pablo Chialvo, Jeffrey D Holland, Timothy J Anderson,
Jesse W Breinholt, Akito Y Kawahara, Xin Zhou, Shanlin Liu, and Jennifer
M Zaspel. 2018. “A Phylogenomic Analysis of Lichen-Feeding Tiger Moths
Uncovers Evolutionary Origins of Host Chemical Sequestration.” Mol Phylogenet
Evol 121 (April): 23–34. https://doi.org/10.1016/j.ympev.2017.12.015.
Shanks, David R, Miguel A Vadillo, Benjamin Riedel, Ashley Clymo, Sinita
Govind, Nisha Hickin, Amanda J F Tamman, and Lara M C Puhlmann. 2015.
“Romance, Risk, and Replication: Can Consumer Choices and Risk-Taking Be
Primed by Mating Motives?” J Exp Psychol Gen 144 (6): e142–58. https:
//doi.org/10.1037/xge0000116.
Shapiro, S. S., and M. B. Wilk. 1965. “An analysis of variance test for normality
(complete samples).” Biometrika 52 (3-4): 591–611. https://doi.org/10.1093/
biomet/52.3-4.591.
Smirnov, N. 1948. “Table for Estimating the Goodness of Fit of Empirical
Distributions.” The Annals of Mathematical Statistics 19 (2). Institute of
Mathematical Statistics: 279–81. https://doi.org/10.1214/aoms/1177730256.
Spearman, C. 1904. “"General Intelligence," Objectively Determined and Mea-
sured.” The American Journal of Psychology 15 (2). University of Illinois Press:
201–92. http://www.jstor.org/stable/1412107.
Storey, John D, and Robert Tibshirani. 2003. “Statistical Significance for
Genomewide Studies.” Proc Natl Acad Sci U S A 100 (16): 9440–5. https:
//doi.org/10.1073/pnas.1530509100.
Stroop, J.R. 1935. “Studies of Interference in Serial Verbal Reactions.” Journal
of Experimental Psychology 18 (6): 643–62. https://doi.org/10.1037/h0054651.
440 CHAPTER 21. REFERENCES

Thompson, R. 1985. “A Note on Restricted Maximum Likelihood Estimation

with an Alternative Outlier Model.” Journal of the Royal Statistical Society.
Series B (Methodological) 47. [Royal Statistical Society, Wiley]: 53–55. http:
//www.jstor.org/stable/2345543.
Tukey, J.W. 1977. Exploratory Data Analysis. Addison-Wesley.
Venables, W. N, Brian D. Ripley, and W. N Venables. 2002. Modern Applied
Statistics with S. 4th ed. New York: Springer. http://www.loc.gov/catdir/toc/
fy042/2002022925.html.
Victor, J D, and J Mast. 1991. “A New Statistic for Steady-State Evoked
Potentials.” Electroencephalogr Clin Neurophysiol 78 (5): 378–88. https://doi.
org/10.1016/0013-4694(91)90099-p.
Vigen, T. 2015. Spurious Correlations. Hachette Books.
Vilidaite, G., Anthony M Norcia, Ryan J H West, Christopher J H Elliott,
Francesca Pei, Alex R Wade, and Daniel H Baker. 2018. “Autism Sensory
Dysfunction in an Evolutionarily Conserved System.” Proc Biol Sci 285 (1893):
20182255. https://doi.org/10.1098/rspb.2018.2255.
Wagenmakers, Eric-Jan. 2007. “A Practical Solution to the Pervasive Problems
of P Values.” Psychon Bull Rev 14 (5): 779–804. https://doi.org/10.3758/
bf03194105.
Wajrock, Sophie, Nicolas Antille, Andreas Rytz, Nicolas Pineau, and Corinne
Hager. 2008. “Partitioning Methods Outperform Hierarchical Methods for
Clustering Consumers in Preference Mapping.” Food Quality and Preference 19
(7): 662–69. https://doi.org/https://doi.org/10.1016/j.foodqual.2008.06.002.
Walters, Charlotte L., Robin Freeman, Alanna Collen, Christian Dietz, M. Brock
Fenton, Gareth Jones, Martin K. Obrist, et al. 2012. “A Continental-Scale
Tool for Acoustic Identification of European Bats.” Edited by JeroenEditor
Minderman. Journal of Applied Ecology 49 (5). Wiley: 1064–74. https://doi.
org/10.1111/j.1365-2664.2012.02182.x.
Wang, Y. Andre, and Mijke Rhemtulla. 2021. “Power Analysis for Parameter Es-
timation in Structural Equation Modeling: A Discussion and Tutorial.” Advances
in Methods and Practices in Psychological Science 4 (1): 2515245920918253.
https://doi.org/10.1177/2515245920918253.
Wang, Yilun, and Michal Kosinski. 2018. “Deep Neural Networks Are More
Accurate Than Humans at Detecting Sexual Orientation from Facial Images.” J
Pers Soc Psychol 114 (2): 246–57. https://doi.org/10.1037/pspa0000098.
Ware, C. 1988. “Color Sequences for Univariate Maps: Theory, Experiments
and Principles.” IEEE Computer Graphics and Applications 8 (5): 41–49.
https://doi.org/10.1109/38.7760.
Weiskrantz, Larry. 2007. “Blindsight.” Scholarpedia 2 (4). Scholarpedia: 3047.
https://doi.org/10.4249/scholarpedia.3047.
 441

Weisstein, N. 1980. “The Joy of Fourier Analysis.” In Visual Coding and

Adaptability, edited by S.C. Harris, 365–80. Hillsdale, N.J.: Lawrence Erlbaum.
Westfall, Jacob, David A Kenny, and Charles M Judd. 2014. “Statistical
Power and Optimal Design in Experiments in Which Samples of Participants
Respond to Samples of Stimuli.” J Exp Psychol Gen 143 (5): 2020–45. https:
//doi.org/10.1037/xge0000014.
Wickham, H. 2016. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag
New York.
Wixted, John T. 2020. “The Forgotten History of Signal Detection Theory.” J Exp
Psychol Learn Mem Cogn 46 (2): 201–33. https://doi.org/10.1037/xlm0000732.
Wolf, Erika J, Kelly M Harrington, Shaunna L Clark, and Mark W Miller. 2013.
“Sample Size Requirements for Structural Equation Models: An Evaluation
of Power, Bias, and Solution Propriety.” Educ Psychol Meas 76 (6): 913–34.
https://doi.org/10.1177/0013164413495237.
Xie, Y., J.J. Allaire, and G. Grolemund. 2019. R Markdown: The Definitive
Guide. Chapman & Hall.
Yap, B. W., and C. H. Sim. 2011. “Comparisons of Various Types of Normality
Tests.” Journal of Statistical Computation and Simulation 81 (12). Informa UK
Limited: 2141–55. https://doi.org/10.1080/00949655.2010.520163.
Yarkoni, Tal, Russell A Poldrack, Thomas E Nichols, David C Van Essen, and
Tor D Wager. 2011. “Large-Scale Automated Synthesis of Human Functional
Neuroimaging Data.” Nat Methods 8 (8): 665–70. https://doi.org/10.1038/
nmeth.1635.
Yen, Tian-Ming. 2016. “Culm Height Development, Biomass Accumulation and
Carbon Storage in an Initial Growth Stage for a Fast-Growing Moso Bamboo
(Phyllostachy Pubescens).” Bot Stud 57 (1): 10. https://doi.org/10.1186/s40529-
016-0126-x.