Analysing data using linear models

Stéphanie M. van den Berg

Fifth edition (R)

(January 4, 2021)
Copyright © 2018, 2021 by Stéphanie M. van den Berg
University of Twente
Department of Learning, Data Analytics, and Technology
Licensed under Creative Commons, see https://creativecommons.org/licenses/
For source code and updates: github.com/pingapang/book
Email: [email protected]
cbna
First edition: October 2018
Second edition: November 2018
Third edition: January 2019
Fourth edition: November 2019
Fifth edition: January 2021
Preface

This book is for bachelor students in social, behavioural and management sciences
that want to learn how to analyse their data, with the specific aim to answer
research questions. The book has a practical take on data analysis: how to do
it, how to interpret the results, and how to report the results. All techniques
are presented within the framework of linear models: this includes simple and
multiple regression models, linear mixed models and generalised linear models.
This approach is illustrated using R.

iii
Contents

1 Variables, variation and co-variation 1

1.1 Units, variables, and the data matrix . . . . . . . . . . . . . . . . 1
1.2 Data matrices in R . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Multiple observations: wide format and long format data matrices 3
1.4 Wide and long format in R . . . . . . . . . . . . . . . . . . . . . 6
1.4.1 From wide to long . . . . . . . . . . . . . . . . . . . . . . 6
1.4.2 From long to wide . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Measurement level . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5.1 Numeric variables . . . . . . . . . . . . . . . . . . . . . . 9
1.5.2 Ordinal variables . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.3 Categorical variables . . . . . . . . . . . . . . . . . . . . . 11
1.5.4 Treatment of variables in data analysis . . . . . . . . . . . 12
1.6 Measurement level in R . . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Frequency tables, frequency plots and histograms . . . . . . . . . 15
1.8 Frequencies, proportions and cumulative frequencies and proportions 17
1.9 Frequencies and proportions in R . . . . . . . . . . . . . . . . . . 18
1.10 Quartiles, quantiles and percentiles . . . . . . . . . . . . . . . . . 20
1.11 Quantiles in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.12 Measures of central tendency . . . . . . . . . . . . . . . . . . . . 23
1.12.1 The mean . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.12.2 The median . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.12.3 The mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.13 Relationship between measures of tendency and measurement level 26
1.14 Measures of central tendency in R . . . . . . . . . . . . . . . . . 27
1.15 Measures of variation . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.15.1 Range and interquartile range . . . . . . . . . . . . . . . . 28
1.15.2 Sum of squares . . . . . . . . . . . . . . . . . . . . . . . . 29
1.15.3 Variance and standard deviation . . . . . . . . . . . . . . 30
1.16 Variance, standard deviation, and standardisation in R . . . . . . 32
1.17 Density plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.18 Density plots in R . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.19 The normal distribution . . . . . . . . . . . . . . . . . . . . . . . 35
1.20 Obtaining quantiles of the normal distribution using R . . . . . . 40
1.21 Visualising numeric variables: the box plot . . . . . . . . . . . . 40

v
 1.22 Box plots in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.23 Visualising categorical variables . . . . . . . . . . . . . . . . . . . 41
1.24 Visualising categorical and ordinal variables in R . . . . . . . . . 45
1.25 Visualising co-varying variables . . . . . . . . . . . . . . . . . . . 46
1.25.1 Categorical by categorical: cross-table . . . . . . . . . . . 46
1.25.2 Categorical by numerical: box plot . . . . . . . . . . . . . 47
1.25.3 Numeric by numeric: scatter plot . . . . . . . . . . . . . . 48
1.26 Visualising two variables using R . . . . . . . . . . . . . . . . . . 50
1.27 Overview of the book . . . . . . . . . . . . . . . . . . . . . . . . . 51

2 Inference about a mean 53

2.1 The problem of inference . . . . . . . . . . . . . . . . . . . . . . . 53
2.2 Sampling distribution of mean and variance . . . . . . . . . . . . 55
2.3 The effect of sample size . . . . . . . . . . . . . . . . . . . . . . . 57
2.4 The standard error . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.5 Confidence intervals . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.6 The t-statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.7 Interpreting confidence intervals . . . . . . . . . . . . . . . . . . . 70
2.8 t-distributions and degrees of freedom . . . . . . . . . . . . . . . 71
2.9 Constructing confidence intervals . . . . . . . . . . . . . . . . . . 74
2.10 Obtaining a confidence interval for a population mean in R . . . 75
2.11 Null-hypothesis testing . . . . . . . . . . . . . . . . . . . . . . . . 76
2.12 Null-hypothesis testing with t-values . . . . . . . . . . . . . . . . 79
2.13 The p-value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.14 One-sided versus two-sided testing . . . . . . . . . . . . . . . . . 86
2.15 One-tailed testing applied to LH levels . . . . . . . . . . . . . . . 88
2.16 Type I and type II errors . . . . . . . . . . . . . . . . . . . . . . 91

3 Inference about a proportion 97

3.1 Sampling distribution of the sample proportion . . . . . . . . . . 97
3.2 The binomial distribution . . . . . . . . . . . . . . . . . . . . . . 98
3.3 Confidence intervals . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.4 Null-hypothesis concerning a proportion . . . . . . . . . . . . . . 102
3.5 Inference on proportions using R . . . . . . . . . . . . . . . . . . 103

4 Linear modelling: introduction 107

4.1 Dependent and independent variables . . . . . . . . . . . . . . . 107
4.2 Linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.3 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.4 Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.5 Least squares regression lines . . . . . . . . . . . . . . . . . . . . 115
4.6 Linear models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.7 Finding the OLS intercept and slope using R . . . . . . . . . . . 120
4.8 Pearson correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.9 Covariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.10 Numerical example of covariance, correlation and least square slope127
 4.11 Correlation, covariance and slopes in R . . . . . . . . . . . . . . . 128
4.12 Explained and unexplained variance . . . . . . . . . . . . . . . . 131
4.13 More than one predictor . . . . . . . . . . . . . . . . . . . . . . . 132
4.14 R-squared . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.15 Multiple regression in R . . . . . . . . . . . . . . . . . . . . . . . 135
4.16 Multicollinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.17 Simpson’s paradox . . . . . . . . . . . . . . . . . . . . . . . . . . 140

5 Inference for linear models 145

5.1 Population data and sample data . . . . . . . . . . . . . . . . . . 146
5.2 Random sampling and the standard error . . . . . . . . . . . . . 147
5.2.1 Standard error and sample size . . . . . . . . . . . . . . . 150
5.2.2 From sample slope to population slope . . . . . . . . . . . 150
5.3 t-distribution for the model coefficients . . . . . . . . . . . . . . . 153
5.4 Confidence intervals for the slope . . . . . . . . . . . . . . . . . . 154
5.5 Residual degrees of freedom in linear models . . . . . . . . . . . . 157
5.6 Null-hypothesis testing with linear models . . . . . . . . . . . . . 160
5.7 p-values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.8 Hypothesis testing . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.9 Inference for linear models in R . . . . . . . . . . . . . . . . . . . 167
5.10 Type I and Type II errors in decision making . . . . . . . . . . . 169
5.11 Statistical power . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.12 Power analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.13 Criticism on null-hypothesis testing and p-values . . . . . . . . . 177
5.14 Relationship between p-values and confidence intervals . . . . . . 181
5.15 The intercept only model . . . . . . . . . . . . . . . . . . . . . . 182

6 Categorical predictor variables 185

6.1 Dummy coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.2 Using regression to describe group means . . . . . . . . . . . . . 187
6.3 Making inferences about differences in group means . . . . . . . 191
6.4 Regression analysis using a dummy variable in R . . . . . . . . . 191
6.5 Two independent variables: one dummy and one numeric variable 194
6.6 Dummy coding for more than two groups . . . . . . . . . . . . . 197
6.7 Analysing categorical predictor variables in R . . . . . . . . . . . 198
6.7.1 Creating your own dummy variables . . . . . . . . . . . . 198
6.7.2 Let R create dummy variables automatically . . . . . . . 199
6.7.3 Interpreting the regression table . . . . . . . . . . . . . . 200
6.8 Analysis of Variance . . . . . . . . . . . . . . . . . . . . . . . . . 201
6.9 The logic of the F -statistic . . . . . . . . . . . . . . . . . . . . . 205
6.10 Small ANOVA example . . . . . . . . . . . . . . . . . . . . . . . 207
6.11 Reporting ANOVA . . . . . . . . . . . . . . . . . . . . . . . . . . 210
6.12 Relationship between F - and t-distributions . . . . . . . . . . . . 211
7 Assumptions of linear models 215
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
7.2 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
7.3 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.4 Equal variances . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
7.5 Residuals normally distributed . . . . . . . . . . . . . . . . . . . 231
7.6 General approach to testing assumptions . . . . . . . . . . . . . . 233
7.7 Checking assumptions in R . . . . . . . . . . . . . . . . . . . . . 234

8 When assumptions are not met: non-parametric alternatives 239

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
8.2 Spearman’s ρ (rho) . . . . . . . . . . . . . . . . . . . . . . . . . . 243
8.3 Spearman’s rho in R . . . . . . . . . . . . . . . . . . . . . . . . . 246
8.4 Kendall’s rank-order correlation coefficient τ . . . . . . . . . . . 246
8.5 Kendall’s τ in R . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.6 Kruskal-Wallis test for group comparisons . . . . . . . . . . . . . 250
8.7 Kruskal-Wallis test in R . . . . . . . . . . . . . . . . . . . . . . . 251

9 Moderation: testing interaction effects 253

9.1 Interaction with one numeric and one dichotomous variable . . . 253
9.2 Interaction effect with a dummy variable in R . . . . . . . . . . . 256
9.3 Interaction effects with a categorical variable in R . . . . . . . . 261
9.4 Interaction between two dichotomous variables in R . . . . . . . 264
9.5 Moderation involving two numeric variables in R . . . . . . . . . 268

10 Linear mixed modelling: introduction 273

10.1 Fixed effects and random effects . . . . . . . . . . . . . . . . . . 273
10.2 Pre-post intervention designs . . . . . . . . . . . . . . . . . . . . 277
10.3 Parameter estimation in linear mixed models . . . . . . . . . . . 286

11 Linear mixed models for more than two measurements 289

11.1 Pre-mid-post intervention designs . . . . . . . . . . . . . . . . . . 289
11.2 Pre-mid-post intervention design: linear effects . . . . . . . . . . 293
11.3 Linear mixed models and interaction effects . . . . . . . . . . . . 297
11.4 Mixed designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
11.5 Mixed design with a linear effect . . . . . . . . . . . . . . . . . . 305

12 Non-parametric alternatives for linear mixed models 311

12.1 Checking assumptions . . . . . . . . . . . . . . . . . . . . . . . . 311
12.2 Friedman’s test for k measures . . . . . . . . . . . . . . . . . . . 314
12.3 How to perform Friedman’s test in R . . . . . . . . . . . . . . . . 319
12.4 Wilcoxon’s signed ranks test for 2 measures . . . . . . . . . . . . 320
12.5 How to perform Wilcoxon’s signed ranks test in R . . . . . . . . 322
12.6 Ties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
13 Generalised linear models: logistic regression 327
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
13.2 Logistic regression . . . . . . . . . . . . . . . . . . . . . . . . . . 331
13.2.1 Bernoulli distribution . . . . . . . . . . . . . . . . . . . . 331
13.2.2 Odds and logodds . . . . . . . . . . . . . . . . . . . . . . 334
13.2.3 Logistic link function . . . . . . . . . . . . . . . . . . . . . 337
13.3 Logistic regression in R . . . . . . . . . . . . . . . . . . . . . . . 340

14 Generalized linear models for count data: Poisson regression 343

14.1 Poisson regression . . . . . . . . . . . . . . . . . . . . . . . . . . 343
14.2 Poisson regression in SPSS . . . . . . . . . . . . . . . . . . . . . . 348
14.3 Interaction effects in Poisson models . . . . . . . . . . . . . . . . 350
14.4 Cross-tabulation and the Pearson chi-square statistic . . . . . . . 354
14.5 Poisson regression or logistic regression? . . . . . . . . . . . . . . 357

15 Introduction to big data analytics 361

15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
15.1.1 Model selection . . . . . . . . . . . . . . . . . . . . . . . . 365
15.1.2 Cross-validation . . . . . . . . . . . . . . . . . . . . . . . 366
15.1.3 The p > n problem . . . . . . . . . . . . . . . . . . . . . . 366
15.1.4 Steps in big data analytics . . . . . . . . . . . . . . . . . . 367

Appendices 371

A Cumulative probabilities for the standard normal distribution 373

B Critical values for the t-distribution 377

Chapter 1

Variables, variation and

co-variation

1.1 Units, variables, and the data matrix

Data is the plural of datum, and datum is the Latin translation of ’given’. That
the world is round, is a given. That you are reading these lines, is a given, and
that my dog’s name is Philip, is a given. Sometimes we have a bunch of given
facts (data), for example the names of all students in a school, and their marks
for a particular course. We could put these data in a table, like the one in Table
1.1. There we see information (’facts’) about seven students. And of these seven
students we know two things: their name and their grade. You see that the data
are put in a matrix with seven (horizontal) rows and two (vertical) columns.
Each row stands for one student, and each column stands for one property.
In data analysis, we nearly always put data in such a matrix format. In
general, we put the objects of our study in rows, and their properties in columns.
The objects of our study we call units, and the properties we call variables.

Table 1.1: Data matrix with 7 units and 2 variables.

name grade
Mark Zimmerman 5
Daisy Doe 8
Mohammed Solmaz 5
Monique Gambin 9
Inga Svensson 10
Piet van der Keuken 2
Floor de Vries 6

Let’s look at the first column in Table 1.1. We see that it regards the variable
name. We call the property name a variable, because it varies across our units
(the students): in this case, every unit has a different value for the variable

1
name. In sum, a variable is a property of units that shows different values for
different units.
The second column represents the variable grade. Grade is here a variable,
because it takes different values for different students. Note that both Mark
Zimmerman and Mohammed Solmaz have the same value for this variable.
What we see in Table 1.1 is called a data matrix : it is a matrix (a collection
of rows and columns) that contains information on units (in the rows) in the
form of variables (in the columns).
A unit is something we’d like to say something about. For example, I might
want to say something about students and how they score on a course. In that
case, students are my units of analysis.
If my interest is in schools, the data matrix in Table 1.2 might be useful,
which shows a different row for each school with a couple of variables. Here
again, we see a variable for grade on a course, but now averaged per school. In
this case, school is my unit of analysis.

Table 1.2: Data matrix on schools.

school number students grade average teacher
1 5 6.1 Alice Monroe
2 8 5.9 Daphne Stuart
3 5 6.9 Stephanie Morrison
4 9 5.9 Clark Davies
5 10 6.4 David Sanchez Gomez
6 2 6.1 Metin Demirci
7 6 5.2 Frederika Karlsson
8 9 6.8 Advika Agrawal

1.2 Data matrices in R

In R, data matrices are called data frames. A data frame consists of different
vectors, one vector for each variable, and each vector contains values. Each
vector/variable is stored as a column in a data frame. In the tidyverse version
of R that we use in this book, we work with a particular form of a data frame:
a tibble. Below we see some R code that creates a tibble: we first load the
tidyverse package, then we create the vectors studentID, course, grade, and
shirtsize, and then combine these 4 vectors into a tibble.

library(tidyverse)
studentID <- seq(4132211, 4132215)
course <- c("Chemistry", "Physics", "Math", "Math", "Chemistry")
grade <- c(4, 6, 3, 6, 8)
shirtsize <- c("medium", "small", "large", "medium", "small")
tibble(studentID, course, shirtsize, grade)

2
## # A tibble: 5 x 4
## studentID course shirtsize grade
## <int> <chr> <chr> <dbl>
## 1 4132211 Chemistry medium 4
## 2 4132212 Physics small 6
## 3 4132213 Math large 3
## 4 4132214 Math medium 6
## 5 4132215 Chemistry small 8

From the output, you see that the tibble has dimensions 5 × 4: that means it
has 5 rows (units) and 4 columns (variables). Under the variable names, it can
be seen how the data are stored. The variable studentID is stored as a numeric
variable, more specifically as an integer (<int>). The course variable is stored
as a character variable (<chr>), because the values consist of text. The same is
true for shirtsize. The last variable, grade, is stored as <dbl> which stands
for ’double’. Whether a numeric variable is stored as integer or double depends
on the amount of computer memory that is allocated to a variable. Double
variables have a decimal part (e.g., 2.0), integers don’t (e.g., 2).

1.3 Multiple observations: wide format and long

format data matrices
In many instances, units of analysis are observed more than once. This means
that we have more than one observation for the same variable for the same unit
of analysis. Storing this information in the rows and columns of a data matrix
can be done in two ways: using wide format or using long format. We first look
at wide format, and then see that generally, long format is to be preferred.
Suppose we measure depression levels in four men four times during cognitive
behavioural therapy. Sometimes you see data presented in the way of Table
1.3, where there are four separate variables for depression level, one for each
measurement: depression 1, depression 2, depression 3, and depression 4.

Table 1.3: Data matrix with depression levels in wide format.

client depression 1 depression 2 depression 3 depression 4
1 5 6 9 3
2 9 5 8 7
3 9 0 9 3
4 9 2 8 6

This way of representing data on a variable that was measured more than
once is called wide format. We call it wide because we simply add columns when
we have more measurements, which increases the width of the data matrix. Each
new observation of the same variable on the same unit of analysis leads to a
new column in the data matrix.

3
 Table 1.4: Data matrix with depression levels in long format.
client time depression
1 1 5
1 2 6
1 3 9
1 4 3
2 1 9
2 2 5
2 3 8
2 4 7
3 1 9
3 2 0
3 3 9
3 4 3
4 1 9
4 2 2
4 3 8
4 4 6

Note that this is only one way of looking at this problem of measuring
depression four times. Here, you can say that there are really four depression
variables: there is depression measured at time point 1, there is depression
measured at time point 2, and so on, and these four variables vary only across
units of analysis. This way of thinking leads to a wide format representation.
An alternative way of looking at this problem of measuring depression four
times, is that depression is really only one variable and that it varies across
units of analysis (some people are more depressed than others) and that it also
varies across time (at times you feel more depressed than at other times).
Therefore, instead of adding columns, we could simply stick to one variable
and only add rows. That way, the data matrix becomes longer, which is
the reason that we call that format long format. Table 1.4 shows the same
information from Table 1.3, but now in long format. Instead of four different
variables, we have only one variable for depression level, and one extra variable
time that indicates to which time point a particular depression measure refers
to. Thus, both Tables 1.3 and 1.4 tell us that the second depression measure
for client number 3 was 0.
Now let’s look at a slightly more complex example, where the advantage
of long format becomes clear. Suppose the depression measures were taken on
different days for different clients. Client 1 was measured on Monday, Tuesday,
Wednesday and Thursday, while client 2 was measured on Thursday, Friday,
Saturday and Sunday. If we would put that information into a wide format
table, it would look like Figure 1.5, with missing values for measures on Monday
thru Wednesday for client 2, and missing values for measures on Friday thru
Sunday for patient 1.

4
 Table 1.5: Data matrix with depression levels in wide format.
client Monday Tuesday Wednesday Thursday Friday Saturday Sunday
1 5 6 9 3
2 9 5 8 7

Table 1.6 shows the same data in long format. The data frame is considerably
smaller. Imagine that we would also have weather data for the days these
patients were measured: whether it was cloudy or sunny, whether it rained or
not, and what the maximum temperature was. In long format, storing that
information is easy, see Table 1.7. Try and see if you can think of a way to store
that information in a wide table!

Table 1.6: Data matrix with depression levels in long format.

client depression day
1 5 Monday
1 6 Tuesday
1 9 Wednesday
1 3 Thursday
2 9 Thursday
2 5 Friday
2 8 Saturday
2 7 Sunday

Table 1.7: Data matrix with depression levels in wide format, including data on
the time of measurement.
client depression day maxtemp rain
1 5 Monday 23 rain
1 6 Tuesday 24 no rain
1 9 Wednesday 23 rain
1 3 Thursday 25 no rain
2 9 Thursday 25 no rain
2 5 Friday 22 no rain
2 8 Saturday 21 rain
2 7 Sunday 22 no rain

Thus, storing data in long format is often more efficient in terms of storage of
information. Another reason for preferring long format over wide format is the
most practical one for data analysis: when analysing data using linear models,
software packages require your data to be in long format. In this book, all the
analyses with linear models require your data to be in long format. However,
we will also come across some analyses apart from linear models that require
your data to be in wide format. If your data happen to be in the wrong format,

5
rearrange your data first. Of course you should never do this by hand as this
will lead to typing errors and would take too much time. Statistical software
packages have helpful tools for rearranging your data from wide format to long
format, and vice versa.

1.4 Wide and long format in R

Making a data matrix longer or wider can be done with the functions pivot longer()
and pivot wider(), respectively. These functions are part of the tidyr package,
and available when you load the tidyverse collection of packages.

library(tidyverse)

1.4.1 From wide to long

The relig income dataset stores counts based on a survey which (among other
things) asked people about their religion and annual income:

relig_income

## # A tibble: 18 x 11
## religion ‘<$10k‘ ‘$10-20k‘ ‘$20-30k‘ ‘$30-40k‘ ‘$40-50k‘ ‘$50-75k‘ ‘$75-100k‘
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 Agnostic 27 34 60 81 76 137 122
## 2 Atheist 12 27 37 52 35 70 73
## 3 Buddhist 27 21 30 34 33 58 62
## 4 Catholic 418 617 732 670 638 1116 949
## 5 Don’t k~ 15 14 15 11 10 35 21
## 6 Evangel~ 575 869 1064 982 881 1486 949
## 7 Hindu 1 9 7 9 11 34 47
## 8 Histori~ 228 244 236 238 197 223 131
## 9 Jehovah~ 20 27 24 24 21 30 15
## 10 Jewish 19 19 25 25 30 95 69
## 11 Mainlin~ 289 495 619 655 651 1107 939
## 12 Mormon 29 40 48 51 56 112 85
## 13 Muslim 6 7 9 10 9 23 16
## 14 Orthodox 13 17 23 32 32 47 38
## 15 Other C~ 9 7 11 13 13 14 18
## 16 Other F~ 20 33 40 46 49 63 46
## 17 Other W~ 5 2 3 4 2 7 3
## 18 Unaffil~ 217 299 374 365 341 528 407
## # ... with 3 more variables: ‘$100-150k‘ <dbl>, ‘>150k‘ <dbl>, ‘Don’t
## # know/refused‘ <dbl>

This dataset contains three variables:

6
 1. religion, stored in the rows,

2. income, spread across the column names, and

3. count, stored in the cell values.

To put the values that we see in the columns into one single column, we use
pivot longer():

relig_income %>%
pivot_longer(cols = -religion, # columns that need to be restructured
names_to = "income", # name of new variable with old column names
values_to = "count") # name of new variable with values

## # A tibble: 180 x 3
## religion income count
## <chr> <chr> <dbl>
## 1 Agnostic <$10k 27
## 2 Agnostic $10-20k 34
## 3 Agnostic $20-30k 60
## 4 Agnostic $30-40k 81
## 5 Agnostic $40-50k 76
## 6 Agnostic $50-75k 137
## 7 Agnostic $75-100k 122
## 8 Agnostic $100-150k 109
## 9 Agnostic >150k 84
## 10 Agnostic Don’t know/refused 96
## # ... with 170 more rows

ˆ The cols argument describes which columns need to be reshaped. In this

case, it is every column except religion.

ˆ The names to argument gives the name of the variable that will be created
using the column names, i.e. income.

ˆ The values to argument gives the name of the variable that will be
created from the data stored in the cells, i.e. count.

1.4.2 From long to wide

The us rent income dataset contains information about median income and
rent for each state in the US for 2017 (from the American Community Survey,
retrieved with the tidycensus package).

us_rent_income

7
## # A tibble: 104 x 5
## GEOID NAME variable estimate moe
## <chr> <chr> <chr> <dbl> <dbl>
## 1 01 Alabama income 24476 136
## 2 01 Alabama rent 747 3
## 3 02 Alaska income 32940 508
## 4 02 Alaska rent 1200 13
## 5 04 Arizona income 27517 148
## 6 04 Arizona rent 972 4
## 7 05 Arkansas income 23789 165
## 8 05 Arkansas rent 709 5
## 9 06 California income 29454 109
## 10 06 California rent 1358 3
## # ... with 94 more rows

Here both estimate and moe are variables (column names), so we can supply
them to the function argument values from to make new variables:

us_rent_income %>%
pivot_wider(names_from = variable,
values_from = c(estimate, moe))

## # A tibble: 52 x 6
## GEOID NAME estimate_income estimate_rent moe_income moe_rent
## <chr> <chr> <dbl> <dbl> <dbl> <dbl>
## 1 01 Alabama 24476 747 136 3
## 2 02 Alaska 32940 1200 508 13
## 3 04 Arizona 27517 972 148 4
## 4 05 Arkansas 23789 709 165 5
## 5 06 California 29454 1358 109 3
## 6 08 Colorado 32401 1125 109 5
## 7 09 Connecticut 35326 1123 195 5
## 8 10 Delaware 31560 1076 247 10
## 9 11 District of Columbia 43198 1424 681 17
## 10 12 Florida 25952 1077 70 3
## # ... with 42 more rows

ˆ The names from argument gives the name of the variable that will be used
for the new column names, i.e. variable

ˆ The values from argument gives the name(s) of the variable(s) that store
the value that you wish to see spread out across several columns. Here we
have two such variables, i.e. moe and estimate

For more examples, see the vignette on pivoting.

8
vignette("pivot")

1.5 Measurement level

Data analysis is about variables and the relationships among them. In essence,
data analysis is about describing how different values in one variable go together
with different values in one or more other variables (co-variation). For example,
if we have the variable age with values ’young’ and ’old’, and the variable
happiness with values ’happy’ and ’unhappy’, we’d like to know whether ’happy’
mostly comes together with either ’young’ or ’old’. Therefore, data analysis is
about variation and co-variation in variables.
Linear models are important tools when describing co-varying variables.
When we want to use linear models, we need to distinguish between different
kinds of variables. One important distinction is about the measurement level of
the variable: numeric, ordinal or categorical.

1.5.1 Numeric variables

Numeric variables have values that describe a measurable quantity as a number,
like ’how many’ or ’how much’. A numeric variable can be a count variable,
for instance the number of children in a classroom. A count variable can
only consist of discrete, natural, positive numbers: 0, 1, 2, 3, etcetera. But
a numeric variable can also be a continuous variable. Continuous variables can
take any value from the set of real numbers, for instance values like -200.765,
-9.78, -2, 0.001, 4, and 7.8. The number of decimals can be as large as the
instrument of measurement allows. Examples of continuous variables include
height, time, age, blood pressure and temperature. Note that in all these
examples, quantities (age, height, temperature) are expressed as the number
of a particular measurement unit (years, inches, degrees).
Whether a numeric variable is a count variable or a continuous variable, it
is always expressing a quantity, and therefore numeric variables can be called
quantitative variables.
For numeric variables, there is a further distinction between interval variables
and ratio variables. The distinction is rather technical. The difference between
interval and ratio variables is that for ratio variables, the ratio between two
measurement values is meaningful, and for interval variables it is not. An
example of a ratio variable is height. You could measure height in two persons
where one measures 1 meter and the other measures 2 meters. It is then
meaningful to say that the second person is twice as tall as the first person.
This is meaningful, because had we chosen a different measurement unit, the
ratio would be the same. For instance, suppose we express the heights of the two
persons in inches, we would get 39.37 and 78.74 inches respectively. The ratio
remains 2: namely 78.74/39.37. The same ratio would hold for measurements
in feet, miles, millimetres or even light years. Thus, whatever the unit of

9
measurement you use, the ratio of height for these individuals would always
be 2. Therefore, if we have a variable that measures height in meters, we are
dealing with a ratio variable.
Now let’s look at an example of an interval variable. Suppose we measure
the temperature in two classrooms: one is 10 degrees Celsius and the other is
20 degrees Celsius. The ratio of these two temperatures is 20/10 = 2, but does
that ratio convey meaningful information? Could we state for example that the
second classroom is twice as warm as the first classroom? The answer is no,
and the reason is simple: had we expressed temperature in Fahrenheit, we would
have gotten a very different ratio. Temperatures of 10 and 20 degrees Celsius
correspond to 50 and 68 degrees Fahrenheit, respectively. These Fahrenheit
temperatures have a ratio of 68/50=1.36. Based on the Fahrenheit metric, the
second classroom would now be 1.36 times warmer than the first classroom. We
therefore say that the ratio does not have a meaningful interpretation, since the
ratio depends on the metric system that you use (Fahrenheit or Celsius). It
would be strange to say that there is twice more warmth in classroom B than
in classroom A, but only if you measure temperature in Celsius, not when you
measure it in Fahrenheit!
The reason why the ratios depend on the metric system, is because both
the Celsius and Fahrenheit metrics have arbitrary zero-points. In the Celsius
metric, 0 degrees does not mean that there is no warmth, nor is that implied in
the Fahrenheit metric. In both metrics, a value of 0 is still warmer than a value
of -1.
Contrasting this to the example of height: a height of 0 is indeed the absence
of height, as you would not even be able to see a person with a height of 0,
whatever metric you would use. Thus, the difference between ratio and interval
variables is that ratio variables have a meaningful zero point where zero indicates
the absence of the quantity that is being measured. This meaningful zero-point
makes it possible to make meaningful statements about ratios (e.g., 4 is twice
as much as 2) which gives ratio variables their name.
What ratio and interval variables have in common is that they are both
numeric variables, expressing quantities in terms of units of measurements. This
implies that the distance between 1 and 2 is the same as the distances between
3 and 4, 4 and 5, etcetera. This distinguishes them from ordinal variables.

1.5.2 Ordinal variables

Ordinal variables are also about quantities. However, the important difference
with numeric variables is that ordinal variables are not measured in units. An
example would be a variable that would quantify size, by stating whether a
T-shirt is small, medium or large. Yes, there is a quantity here, size, but there
is no unit to state exactly how much of that quantity is present in that T-shirt.
Even though ordinal variables are not measured in specific units, you can
still have a meaningful order in the values of the variable. For instance, we know
that a large T-shirt is larger than a medium T-shirt, and a medium T-shirt is
larger than a small T-shirt.

10
 Similar for age, we could code a number of people as young, middle-aged
or old, but on the basis of such a variable we could not state by how much
two individuals differ in age. As opposed to numeric variables that are often
continuous, ordinal variables are usually discrete: there isn’t an infinite number
of levels of the variable. If we have sizes small, medium and large, there are no
meaningful other values in between these values.
Ordinal variables often involve subjective measurements. One example would
be having people rank five films by preference from one to five. A different
example would be having people assess pain: ”On a scale from 1 to 10, how bad
is the pain?”
Ordinal variables often look numeric. For example, you may have large,
medium and small T-shirts, but these values may end up in your data matrix as
’3’, ’2’ and ’1’, respectively. However, note that with a truly numeric variable
there should be a unit of measurement involved (3 of what? 2 of what?), and
that numeric implies that the distance between 3 and 2 is equal to the distance
between 2 and 1. Here you would not have that information: you only know
that a large T-shirt (coded as ’3’) is larger than a medium T-shirt (coded as
’2’), but how large that difference is, and whether that difference is that same
as the difference between a medium T-shirt (’2’) is larger than a small T-shirt
(’1’), you do not know. Therefore, even though we see numbers in our data
matrix, the variable is called an ordinal variable.

1.5.3 Categorical variables

Categorical variables are not about quantity at all. Categorical variables are
about quality. They have values that describe ’what type’ or ’which category’ a
unit of belongs to. For example, a school could either be publicly funded or not,
or a person could either have the Swedish nationality or not. A variable that
indicates such a dichotomy between publicly funded ’yes’ or ’no’, or Swedish
nationality ’yes’ or ’no’, is called a dichotomous variable, and is a subtype of a
categorical variable. The other subtype of a categorical variable is a nominal
variable. Nominal comes from the Latin nomen, which means name. When you
name the nationality of a person, you have a nominal variable. Table 1.8 shows
an example of both a dichotomous variable (Swedish) that always has only two
different values, and a nominal variable (Nationality), that can have as many
different values as you want (usually more than two).
Another example of a nominal variable could be the answer to the question:
”name the colours of a number of pencils”. Nothing quantitative could be
stated about a bunch of pencils that are only assessed regarding their colour.
In addition, there is usually no logical order in the values of such variables,
something that we do see with ordinal variables.

11
 Table 1.8: Nationalities.
ID Swedish Nationality
1 Yes Swedish
2 Yes Swedish
3 No Angolan
4 No Norwegian
5 Yes Swedish
6 Yes Swedish
7 No Danish
8 No Unknown

1.5.4 Treatment of variables in data analysis

For data analysis with linear models, you have to decide for each variable
whether you want to treat it as numeric or as categorical.1 The easiest choice is
for numeric variables: numeric variables should always be treated as numeric.
Categorical data should always be treated as categorical. However, the
problem with categorical variables is that they often look like numeric variables.
For example, take the categorical variable country. In your data file, this
variable could be coded with strings like ”Netherlands”, ”Belgium”, ”Luxembourg”,
etc. But the variable could also be coded with numbers: 1, 2 and 3. In
a codebook that belongs to a data file, it could be stated that 1 stands for
”Netherlands”, 2 for ”Belgium”, and 3 for ”Luxembourg” (these are the value
labels), but still in your data matrix your variable would look numeric. You
then have to make sure that, even though the variable looks numeric, it should
be interpreted as a categorical variable and therefore be treated like a categorical
variable.
The most difficult problem involves ordinal variables: in linear models you
can either treat them as numeric variables or as categorical variables. The choice
is usually based on common sense and whether the results are meaningful. For
instance, if you have an ordinal variable with 7 levels, like a Likert scale, the
variable is often coded with numbers 1 through 7, with value labels 1=”completely
disagree”, 2=”mostly disagree”, 3=”somewhat disagree”, 4=”ambivalent”, 5=”somewhat
agree”, 6=”mostly agree”, and 7=”completely agree”. In this example, you
could choose to treat this variable as a categorical variable, recognising that
this is not a numeric variable as there is no measurement unit. However, if
you feel this is awkward, you could choose to treat the variable as numeric, but
be aware that this implies that you feel that the difference between 1 and 2 is
the same as the difference between 2 and 3. In general, with ordinal data like
Likert scales or sizes like, Small, Medium and Large, one generally chooses to
use categorical treatment for low numbers of categories, say 3 or 4 categories,
and numerical treatment for variables with many categories, say 5 or more.
However, this should not be used as a rule of thumb: first think about the
1 In data analysis, it is possible to treat variables as ordinal, but only in more advanced

models and methods than treated in this book.

12
meaning of your variable and the objective of your data analysis project, and
only then take the most reasonable choice. Often, you can start with numerical
treatment, and if the analysis shows peculiar results2 , you can choose categorical
treatment in secondary analyses.
In the coming chapters, we will come back to the important distinction
between categorical and numerical treatment (mostly in Chapter 6). For now,
remember that numeric variables are always treated as numeric variables, categorical
variables are always treated as categorical variables (even when they appear
numeric), and that for ordinal variables you have to think before you act.

1.6 Measurement level in R

In a previous section we saw the creation of a data frame. Let’s store the
resulting data frame as an object called course results.

studentID <- seq(4132211, 4132215)

course <- c("Chemistry", "Physics", "Math", "Math", "Chemistry")
grade <- c(4, 6, 3, 6, 8)
shirtsize <- c("medium", "small", "large", "medium", "small")
course_results <- tibble(studentID, course, shirtsize, grade)
course_results

## # A tibble: 5 x 4
## studentID course shirtsize grade
## <int> <chr> <chr> <dbl>
## 1 4132211 Chemistry medium 4
## 2 4132212 Physics small 6
## 3 4132213 Math large 3
## 4 4132214 Math medium 6
## 5 4132215 Chemistry small 8

We see that the variable studentID is stored as integer. That means that the
values are stored as numeric values. However, the values are quite meaningless,
they are only used to identify persons. If we want to treat this variable as a
categorical variable in data analysis, it is necessary to change this variable into
a factor variable. We can do this by typing:

course_results$studentID <-
course_results$studentID %>%
factor()

When we look at this variable after the transformation, we see that this new
categorical variable has 5 different categories (levels).
2 For instance, you may find that the assumptions of your linear model are not met, see

Chapter 7.

13
course_results$studentID

## [1] 4132211 4132212 4132213 4132214 4132215

## Levels: 4132211 4132212 4132213 4132214 4132215

When we look at the variable course, we see that it is stored as a character

variable. If we want R to treat it as a categorical variable in data analysis, we
can also transform this variable into a factor variable. We could use the same
code as above, or we could use the function mutate().

course_results <- course_results %>%

mutate(course = factor(course))

The shirtsize variable is stored as character, but we tell R that this is an

ordinal variable. For this we need to turn it into a factor variable, indicating
that there is an order in the values, where small is the lowest quantity, and large
the highest quantity.

course_results <- course_results %>%

mutate(shirtsize = factor(shirtsize,
levels = c("small", "medium", "large"),
ordered = TRUE)
)
course_results$shirtsize

## [1] medium small large medium small

## Levels: small < medium < large

The last variable grade is stored as double. Variables of this type will be
treated as numeric in data analyses. If we’re fine with that for this variable, we
leave it as it is. If we want the variable to be treated as ordinal, then we need
the same type of factor transformation as for shirtsize. For now, we leave it as
it is. The resulting data frame then looks like this:

course_results

## # A tibble: 5 x 4
## studentID course shirtsize grade
## <fct> <fct> <ord> <dbl>
## 1 4132211 Chemistry medium 4
## 2 4132212 Physics small 6
## 3 4132213 Math large 3
## 4 4132214 Math medium 6
## 5 4132215 Chemistry small 8

Now both studentID and course are stored as factors and will be treated
as categorical. Variable shirtsize is stored as an ordinal factor and will be

14
treated accordingly. Variable grade is still stored as double and will therefore
be treated as numeric.

1.7 Frequency tables, frequency plots and histograms

Variables have different values. For example, age is a (numeric, ratio) variable:
lots of people have different ages. Suppose we have an imaginary town with 1000
children. For each age measured in years, we can count the number of children
who have that particular age. The results of the counting are in Table 1.9.
The number of observed children with a certain age, say 8 years, is called the
frequency of age 8. The table is therefore called a frequency table. Generally in
a frequency table, values that are not observed are omitted (i.e., the frequency
of children with age 16 is 0).

Table 1.9: Frequency table for age, with proportions and cumulative
proportions.
age frequency proportion cum frequency cum proportion
0 2 0.002 2 0.002
1 7 0.007 9 0.009
2 20 0.020 29 0.029
3 50 0.050 79 0.079
4 105 0.105 184 0.184
5 113 0.113 297 0.297
6 159 0.159 456 0.456
7 150 0.150 606 0.606
8 124 0.124 730 0.730
9 108 0.108 838 0.838
10 70 0.070 908 0.908
11 34 0.034 942 0.942
12 32 0.032 974 0.974
13 14 0.014 988 0.988
14 9 0.009 997 0.997
15 2 0.002 999 0.999
17 1 0.001 1000 1.000

The data in the frequency table can also be represented using a frequency
plot. Figure 1.1 gives the same information, not in a table but in a graphical
way. On the horizontal axis we see several possible values for age in years,
and on the vertical axis we see the number of children (the count) that were
observed for each particular age. Both the frequency table and the frequency
plot tell us something about the distribution of age in this imaginary town with
1000 children. For example, both tell us that the oldest child is 17 years old.
Furthermore, we see that there are quite a lot of children with ages between 5
and 8, but not so many children with ages below 3 or above 14. The advantage

15
 160
150
140
130
120
110
100
90
count

80
70
60
50
40
30
20
10
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
age in years

Figure 1.1: A frequency plot

of the table over the graph is that we can get the exact number of children of a
particular age very easily. But on the other hand, the graph makes it easier to
get a quick idea about the shape of the distribution, which is hard to make out
from the table.

Instead of frequency plots, one often sees histograms. Histograms contain

the same information as frequency plots, except that groups of values are taken
together. Such a group of values is called a bin. Figure 1.2 shows the same
age data, but uses only 9 bins: for the first bin, we take values of age 0 and 1
together, for the second bin we take ages 2 and 3 together, etcetera, until we
take ages 16 and 17 together for the last bin. For each bin, we compute how
often we observe the ages in that bin.

Histograms are very convenient for continuous data, for instance if we have
values like 3.473, 2.154, etcetera. Or, more generally, for variables with values
that have very low frequencies. Suppose that we had measured age not in years
but in days. Then we could have had a data set of 1000 children where each and
every child had a unique value for age. In that case, the length of the frequency
table would be 1000 rows (each value observed only once) and the frequency
plot would be very flat. By using age measured in years, what we have actually
done is putting all children with an age less than 365 days into the first bin (age
0 years) and the children with an age of at least 365 but less than 730 days into
the second bin (age 1 year). And so on. Thus, if you happen to have data with
many many values with very low frequencies, consider binning the data, and
using a histogram to visualise the distribution of your numeric variable.

16
 320
300
280
260
240
220
200
180
count

160
140
120
100
80
60
40
20
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
age in years

Figure 1.2: A histogram

1.8 Frequencies, proportions and cumulative frequencies

and proportions

When we have the frequency for each observed age, we can calculate the relative
frequency or proportion of children that have that particular age. For example,
when we look again at the frequencies in Table 1.9 we see that there are two
children who have age 0. Given that there are in total 1000 children, we know
2
that the proportion of people with age 0 equals 1000 = 0.002. Thus, the
proportion is calculated by taking the frequency and dividing it by the total
number.
We can also compute cumulative frequencies. You get cumulative frequencies
by accumulating (summing) frequencies. For instance, the cumulative frequency
for the age of 3, is the frequency for age 3 plus all frequencies for younger ages.
Thus, the cumulative frequency of age 3 equals 50 + 20 (for age 2) + 7 (for age
1) + 2 (for age 0) = 79. The cumulative frequencies for all ages are presented
in Table 1.9.
We can also compute cumulative proportions: if we take for each age the
proportion of people who have that age or less, we get the fifth column in Table
1.9. For example, for age 2, we see that there are 20 children with an age of 2.
This corresponds to a proportion of 0.020 of all children. Furthermore, there
are 9 children who have an even younger age. The proportion of children with
an age of 1 equals 0.007, and the proportion of children with an age of 0 equals
0.002. Therefore, the proportion of all children with an age of 2 or less equals
0.020 + 0.007 + 0.002 = 0.029, which is called the cumulative proportion for the
age of 2.

17
1.9 Frequencies and proportions in R
The mtcars data set contains information about a number of cars: miles per
gallon (mpg), number of cylinders (cyl), etcetera.

mtcars

## mpg cyl disp hp drat wt qsec vs am gear carb

## Mazda RX4 21.0 6 160.0 110 3.90 2.620 16.46 0 1 4 4
## Mazda RX4 Wag 21.0 6 160.0 110 3.90 2.875 17.02 0 1 4 4
## Datsun 710 22.8 4 108.0 93 3.85 2.320 18.61 1 1 4 1
## Hornet 4 Drive 21.4 6 258.0 110 3.08 3.215 19.44 1 0 3 1
## Hornet Sportabout 18.7 8 360.0 175 3.15 3.440 17.02 0 0 3 2
## Valiant 18.1 6 225.0 105 2.76 3.460 20.22 1 0 3 1
## Duster 360 14.3 8 360.0 245 3.21 3.570 15.84 0 0 3 4
## Merc 240D 24.4 4 146.7 62 3.69 3.190 20.00 1 0 4 2
## Merc 230 22.8 4 140.8 95 3.92 3.150 22.90 1 0 4 2
## Merc 280 19.2 6 167.6 123 3.92 3.440 18.30 1 0 4 4
## Merc 280C 17.8 6 167.6 123 3.92 3.440 18.90 1 0 4 4
## Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3
## Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3
## Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3
## Cadillac Fleetwood 10.4 8 472.0 205 2.93 5.250 17.98 0 0 3 4
## Lincoln Continental 10.4 8 460.0 215 3.00 5.424 17.82 0 0 3 4
## Chrysler Imperial 14.7 8 440.0 230 3.23 5.345 17.42 0 0 3 4
## Fiat 128 32.4 4 78.7 66 4.08 2.200 19.47 1 1 4 1
## Honda Civic 30.4 4 75.7 52 4.93 1.615 18.52 1 1 4 2
## Toyota Corolla 33.9 4 71.1 65 4.22 1.835 19.90 1 1 4 1
## Toyota Corona 21.5 4 120.1 97 3.70 2.465 20.01 1 0 3 1
## Dodge Challenger 15.5 8 318.0 150 2.76 3.520 16.87 0 0 3 2
## AMC Javelin 15.2 8 304.0 150 3.15 3.435 17.30 0 0 3 2
## Camaro Z28 13.3 8 350.0 245 3.73 3.840 15.41 0 0 3 4
## Pontiac Firebird 19.2 8 400.0 175 3.08 3.845 17.05 0 0 3 2
## Fiat X1-9 27.3 4 79.0 66 4.08 1.935 18.90 1 1 4 1
## Porsche 914-2 26.0 4 120.3 91 4.43 2.140 16.70 0 1 5 2
## Lotus Europa 30.4 4 95.1 113 3.77 1.513 16.90 1 1 5 2
## Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4
## Ferrari Dino 19.7 6 145.0 175 3.62 2.770 15.50 0 1 5 6
## Maserati Bora 15.0 8 301.0 335 3.54 3.570 14.60 0 1 5 8
## Volvo 142E 21.4 4 121.0 109 4.11 2.780 18.60 1 1 4 2

The object is a data frame. We can turn it into a tibble as follows:

mtcars <- mtcars %>% as_tibble()

The function as tibble() is available when you load the tidyverse package.
From now on, we assume that you load the tidyverse package at the start of

18
every R session.
If we want to know how many cars belong to which category of number of
cylinders, we can use the function count():

mtcars %>%
count(cyl)

## # A tibble: 3 x 2
## cyl n
## <dbl> <int>
## 1 4 11
## 2 6 7
## 3 8 14

The new variable n is the frequency. We see that the value 4 occurs 11 times,
the value 6 occurs 7 times and the value 8 occurs 14 times. Thus, in this data
set there are 11 cars with 4 cylinders, 7 cars with 6 cylinders, and 14 cars with
8 cylinders.
We obtain proportions when we divide the frequencies by the total number
of cars (the sum of all the values in the n variable):

mtcars %>%
count(cyl) %>%
mutate(proportion = n/sum(n))

## # A tibble: 3 x 3
## cyl n proportion
## <dbl> <int> <dbl>
## 1 4 11 0.344
## 2 6 7 0.219
## 3 8 14 0.438

Cumulative frequencies and cumulative proportions can be obtained using

the cumsum() function:

mtcars %>%
count(cyl) %>%
mutate(proportion = n/sum(n)) %>%
mutate(cumfreq = cumsum(n),
cumprop = cumsum(proportion))

## # A tibble: 3 x 5
## cyl n proportion cumfreq cumprop
## <dbl> <int> <dbl> <int> <dbl>
## 1 4 11 0.344 11 0.344
## 2 6 7 0.219 18 0.562
## 3 8 14 0.438 32 1

19
 A frequency plot can be made using ggplot combined with geom line():

mtcars %>%
count(cyl) %>%
mutate(proportion = n/sum(n)) %>%
ggplot(aes(x = cyl, y = n)) +
geom_line()

A histogram of the mpg variable can be made using geom histogram():

mtcars %>%
ggplot(aes(x = mpg)) +
geom_histogram(breaks = seq(5, 40, 5))

It is wise to play around with the number of bins that you’d like to make,
or with the boundaries of the bins. Here we choose boundaries 5, 10, 15, . . . , 40.

1.10 Quartiles, quantiles and percentiles

Suppose we want to split the group of 1000 children into 4 equally-sized subgroups,
with the 25% youngest children in the first group, the 25% oldest children in the
last group, and the remaining 50% of the children in two equally sized middle
groups. What ages should we then use to divide the groups? First, we can
order the 1000 children on the basis of their age: the youngest first, and the
oldest last. We could then use the concept of quartiles (from quarter, a fourth)
to divide the group in four. In order to break up all ages into 4 subgroups, we
need 3 points to make the division, and these three points are called quartiles.
The first quartile is the value below which 25% of the observations fall, the
second quartile is the value below which 50% of the observations fall, and the
third quartile is the value below which 75% of the observations fall.3
Let’s first look at a smaller but similar problem. For example, suppose your
observed values are 10, 5, 6, 21, 11, 1, 7, 9. You first order them from low to
high so that you obtain 1, 5, 6, 7, 9, 10, 11, 21. You have 8 values, so the first
25% of your values are the first two. The highest value of these two equals 5,
and this we define as our first quartile.4 We find the second quartile by looking
at the values of the first 50% of the observations, so 4 values. The first 4 values
are 1, 5, 6, and 7. The last of these is 7, so that is our second quartile. The
first 75% of the observations are 1, 5 ,6 ,7 , 9, and 10. The value last in line is
10, so our fourth quartile is 10.
3 The fourth quartile would be the value below which all values are, so that would be the

largest value in the row (the age of the last child in the row).
4 Note that we could also choose to use 6, because 1 and 5 are lower than 6. Don’t worry,

the method that we show here to compute quartiles is only one way of doing it. In your life,
you might stumble upon alternative ways to determine quartiles. These are just arbitrary
agreements made by human beings. They can result in different outcomes when you have
small data sets, but usually not when you have large data sets.

20
 1.00

0.75
cum.proportion

0.50

0.25

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
value

Figure 1.3: Cumulative proportions.

The quartiles as defined here can also be found graphically, using cumulative
proportions. Figure 1.3 shows for each observed value the cumulative proportion.
It also shows where the cumulative proportions are equal to 0.25, 0.50 and 0.75.
We see that the 0.25 line intersects the other line at the value of 5. This is the
first quartile. The 0.50 line intersects the other line at a value of 7, and the 0.75
line intersects at a value of 10. The three percentiles are therefore 5, 7 and 10.
If you have a large data set, the graphical way is far easier than doing it by
hand. If we plot the cumulative proportions for the ages of the 1000 children,
we obtain Figure 1.4. We see a nice S-shaped curve. We also see that the
three horizontal quartile lines no longer intersect the curve at specific values, so
what do we do? By eye-balling we can find that the first quartile is somewhere
between 4 and 5. But which value should we give to the quartile? If we look at
the cumulative proportion for an age of 4, we see that its value is slightly below
the 0.25 point. Thus, the proportion of children with age 4 or younger is lower
than 0.25. This means that the child that happens to be the 250th cannot be 4
years old. If we look at the cumulative proportion of age 5, we see that its value
is slightly above 0.25. This means that the proportion of children that is 5 years
old or younger is slightly more than 0.25. Therefore, of the the total of 1000
children, the 250th child must have age 5. Thus, by definition, the first quantile
is 5. The second quartile is somewhere between 6 an 7, so by using the same
reasoning as for the first quartile we know that 50% of the youngest children is
7 years old or younger. The third quartile is somewhere between 8 and 9 and
this tells us that the youngest 75% of the children is age 9 or younger. Thus,
we can call 5, 7 and 9 our three quartiles.
Alternatively, we could also use the frequency table (Table 1.9). First, if we
want to have 25% of the children that are the youngest, and we know that we
have 1000 children in total, we should have 0.25 × 1000 = 250 children in the

21
 1.00

cum.proportion 0.75

0.50

0.25

0.00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
age

Figure 1.4: Cumulative proportions.

first group. So if were to put all the children in a row, ordered from youngest
to oldest, we want to know the age of the 250th child.
In order to find the age of this 250th child, and we look at Table 1.9, we see
that 29.7% of the children have an age of 5 or less (297 children), and 18.4% of
the children have an age of 4 or less (184 children). This tells us that, since 250
comes after 184, the 250th child must be older than 4, and because 250 comes
before 297, it must be younger than or equal to 5, hence the child is 5 years old.
Furthermore, if we want to find a cut-off age for the oldest 25%, we see from
the table, that 83.8% of the children (838 children) have an age of 9 or less, and
73.0% of the children (730) have an age of 8 or less. Therefore, the age of the
750th child (when ordered from youngest to oldest) must be 9.
What we just did for quartiles, (i.e. 0.25, 0.50, 0.75) we can do for any
proportion between 0 and 1. We then no longer call them quartiles, but quantiles.
A quantile is the value below which a given proportion of observations in a group
of observations fall. From this table it is easy to see that a proportion of 0.606
of the children have an age of 7 or less. Thus, the 0.606 quantile is 7. One
often also sees percentiles. Percentiles are very much like quantiles, except that
they refer to percentages rather than proportions. Thus, the 20th percentile is
the same as the 0.20 quantile. And the 0.81 quantile is the same as the 81st
percentile.
The reason that quartiles, quantiles and percentiles are important is that
they are very short ways of saying something about a distribution. Remember
that the best way to represent a distribution is either a frequency table or a
frequency plot. However, since they can take up quite a lot of space sometimes,
one needs other ways to briefly summarise a distribution. Saying that ”the
third quartile is 454” is a condensed way of saying that ”75% of the values is
either 454 or lower”. In the next sections, we look at other ways of summarising

22
information about distributions.
Another way in which quantiles and percentiles are used is to say something
about individuals, relative to a group. Suppose a student has done a test and
she comes home saying she scored in the 76th percentile of her class. What does
that mean? Well, you don’t know her score exactly, but you do know that of
her classmates, 76 percent had the same score or lower. That means she did
pretty well, compared to the others, since only 24 percent had a higher score.

1.11 Quantiles in R
Obtaining quartiles, quantiles and percentiles can be done with the quantile()
function:

quantile(mtcars$mpg,
probs = c(0.25, 0.50, 0.75, 0.90))

## 25% 50% 75% 90%

## 15.425 19.200 22.800 30.090

1.12 Measures of central tendency

The mean, the median and the mode are three different measures that say
something about the central tendency of a distribution. If you have a series of
values: around which value do they tend to cluster?

1.12.1 The mean

Suppose we have the values 1, 2 and 3, then we compute the mean by first
adding these numbers and then divide them by the number of values we have.
In this case we have three values, so the mean is equal to (1 + 2 + 3)/3 = 2.
In statistical formulas, the mean of a variable is indicated by a bar above that
variable. So if our values of variable Y are 1, 2 and 3, then we denote the mean
by Y (pronounced as ’y-bar’). When taking the sum of a set of values, statistical
formulas show the summation sign Σ (the Greek letter sigma). So we often see
the following formula for the mean of a set of n values for variable Y 5 :

Σni=1 Yi
Y = (1.1)
n
In words, in order to compute Y , we take every value for variable Y from
i = 1 to i = n and sum them, and the result is divided by n. Suppose we have
5 Variables are symbolised by capitals, e.g., Y . Specific values of a variable are indicated

in lowercase, e.g., y.

23
variable Y with the values 6, -3, and 21, then the mean of Y equals:

Σi Yi Y1 + Y2 + Y3 6 + (−3) + 21 24
Y = = = = =8 (1.2)
n n 3 3

1.12.2 The median

The mean is only one of the measures of central tendency. An alternative
measure of central tendency is the median. The median is nothing but the
middle value of an ordered series. Suppose we have the values 45, 567, and 23.
Then what value lies in the middle when ordered? Let’s first order them from
small to large to get a better look. We then get 23, 45 and 567. Then it’s easy
to see that the value in the middle is 45.
Suppose we have the values 45, 45, 45, 65, and 23. What is the middle value
when ordered? We first order them again and see what value is in the middle:
23, 45, 45, 45 and 65. Obviously now 45 is the median. You can also see that
half of the values is equal or smaller than this value, and half of the values is
equal or larger than this value. The median therefore is the same as the second
quartile.
What if we have two values in the middle? Suppose we have the values 46,
56, 45 and 34. If we order them we get 34, 45, 46 and 56. Now there are two
values in the middle: 45 and 46. In that case, we take the mean of these two
middle values, so the median is 45.5.
When do you use a median and when do you use a mean? For numeric
variables that have a more or less symmetric distribution (i.e., a frequency plot
that is more or less symmetric), the mean is most often used. Actually, for
distributions that are more or less symmetric the mean and median are very
similar. For numeric variables that do not have a symmetric distribution, it is
usually more informative to use the median. An example of such a situation
is income. Figure 1.5 shows a typical distribution of yearly income. The
distribution is highly asymmetric, it is severely skewed to the right. The bulk of
the values are between 20,000 and 40,000, with only a very few extreme values
on the high end. Even though there are only a few people with a very high
income, the few high values have a huge effect on the mean.
The mean of the distribution turns out to be 23604. The largest value in the
distribution is an income of 75051. Imagine what would happen to the mean
and the median if we would change only this one value, that is, the highest
observed income. Which would be most affected, do you think: the mean or the
median?
Well, if we would change this value into 85051, you see an immediate impact
on the mean: the mean is then 23614. This means that the mean is very sensitive
to extreme values. One single change in a data set can have a huge effect on
the mean. The median on the other hand is much more stable. The median
remains unaffected by changes in the extremes. This because it only looks at
the middle value. The middle value is unaffected by a change in the extreme
values, as long as the order of the values remains the same and the middle value

24
 150
count

100

50

0
0 20000 40000 60000
income

Figure 1.5: Distribution of yearly income.

remains the same.

This can be seen even more clearly by looking at the example in Table 1.10.
There we have three values, X1, X2 and X3, for which we compute both the
mean and the median. First, suppose we have the values 4, 5, and 8 (like in the
first row of Table 1.10). Obviously, the median is 5. Next, instead of 4, 5 and 8,
we could have values 4, 5 and 80, or 4, 5 and 800, or 4, 5 and 8000. Regardless,
the middle value of this series remains 5. In contrast, the mean would be very
much affected by having either an 8, an 80, an 800 or an 8000 in the series. In
sum: the median is a more stable measure of central tendency than the mean.

Table 1.10: Four series of values and their respective medians and means.
X1 X2 X3 median mean
4 5 8 5 5.7
4 5 80 5 29.7
4 5 800 5 269.7
4 5 8000 5 2669.7

1.12.3 The mode

A third measure of central tendency is the mode. The mode is defined as the
value that we see most frequently in a series of values. For example, if we have
the series 4, 7, 5, 5, 6, 6, 6, 4, then the value observed most often is 6 (three
times). Modes are easily inferred from frequency tables: the value with the
largest frequency is the mode. They are also easily inferred from frequency
plots: the value on the horizontal axis for which we see the highest count (on
the vertical axis).

25
 The mode can also be determined for categorical variables. If we have the
observed values ’Dutch’, ’Danish’, ’Dutch’, and ’Chinese’, the mode is ’Dutch’
because that is the value that is observed most often.
If we look back at the distribution in Figure 1.5, we see that the peak of
the distribution is around the value of 19,000. However, whether this is the
mode, we cannot say. Because income is a more or less continuous variable,
every value observed in the Figure occurs only once: there is no value of income
with a frequency more than 1. So technically, there is no mode. However, if
we split the values into 20 bins, like we did for the histogram in Figure 1.5, we
see that the fifth bin has the highest frequency. In this bin there are values
between 17000 and 21000, so our mode could be around there. If we really want
a specific value, we could decide to take the average value in the fifth bin. There
are many other statistical tricks to find a value for the mode, where technically
there is none. The point is that for the mode, we’re looking for the value or the
range of values that are most frequent. Graphically, it is the value under the
peak of the distribution. Similar to the median, the mode is also quite stable: it
is not affected by extreme values and is therefore to be preferred over the mean
in the case of asymmetric distributions.

1.13 Relationship between measures of tendency

and measurement level
There is a close relationship between measures of tendency and measurement
level. For numeric variables, all three measures of tendency are meaningful.
Suppose you have the numeric variable age measured in years, with the values
56, 68, 68, 99 and 100. Then it is meaningful to say that the average age is 78.2
years, that the median age is 68 years, and that the mode is 68 years.
For ordinal variables, it is quite different. Suppose you have 5 T-shirts, with
the following sizes: M, S, M, L, XL. Then what is the average size? There are
no numeric values here to put in the algebraic formula. But we can determine
the median: if we order the values from small to large we get the set S, M, M,
L, XL and we see that the middle value is M. So M is our median in this case.
6
The other meaningful measure of tendency for ordinal variables is the mode.
For categorical variables, both the mean and the median are pointless to
report. Suppose we have the nominal variable Study Programme with observed
values ”Medicine”, ”Engineering”, ”Engineering”, ”Mathematics”, and ”Biology”.
It would be impossible to derive a numerical mean, nor would it be possible to
determine the middle value to determine the median, as there is no logical
or natural order.7 It is meaningful though to report a mode. It would be
meaningful to state that the study programme mentioned most often in the
news is ”Psychology”, or that the most popular study programme in India is
6 However, suppose that our collection of T-shirts had the following sizes: S, M, L, L. Then

there would be no single middle value in we would have to average the M and L values, which
would be impossible!
7 Unless you see one? But then it would not be a categorical value but an ordinal variable.

26
”Engineering”. Thus, for categorical variables, both dichotomous and nominal
variables, only the mode is a meaningful measure of central tendency.
As stated earlier, the appearance of a variable in a data matrix can be
quite misleading. Categorical variables and ordinal variables can often look like
numeric variables, which makes it very tempting to compute means and medians
where they are completely meaningless. Take a look at Table 1.11. It is entirely
possible to compute the average University, Size, or Programme, but it would
be utterly senseless to report these values.
It is entirely possible to compute the median University, Size, or Programme,
but it is only meaningful to report the median for the variable Size, as Size is
an ordinal variable. Reporting that the median size is equal to 2 is saying that
about half of the study programmes is of medium size or small, and about half
of the study programmes is of medium size or large.
It is entirely possible to compute the mode for the variables University, Size,
or Programme, and it is always meaningful to report them. It is meaningful to
say that in your data there is no University that is observed more than others.
It is meaningful to report that most study programmes are of medium size, and
that most study programmes are study programme number 2 (don’t forget to
look up and write down which study programme that actually is!).

Table 1.11: Study programmes and their relative sizes (1=small, 2=medium,
3=large) for six different universities.
University Size Programme
1 1 2
2 3 2
3 2 3
4 2 3
5 3 4
6 2 1

1.14 Measures of central tendency in R

The mean and median for numeric variables can be obtained as follows:

mtcars %>%
summarise(mean_cyl = mean(cyl),
median_cyl = median(cyl))

## # A tibble: 1 x 2
## mean_cyl median_cyl
## <dbl> <dbl>
## 1 6.19 6

27
 R does not have an in-built function to calculate modes. So we create our
own function getmode(). This function takes a vector as input and gives the
mode value as output.

getmode <- function(variable){

unique_values <- unique(variable)
unique_values[
match(variable, unique_values) %>%
tabulate() %>%
which.max()
]
}

mtcars %>%
summarise(mode_cyl = getmode(cyl))

## # A tibble: 1 x 1
## mode_cyl
## <dbl>
## 1 8

1.15 Measures of variation

Above we saw that we can summarise distributions by measures of central
tendency. Here we discuss how we can summarise distributions of numeric
variables by a measure that describes their variation. Variables show variation,
by definition, but how much variation do they actually show?
Suppose we measure the height of 3 children, and their heights (in cms) are
120, 120 and 120. There is no variation in height: all heights are the same.
There are no differences. Then the average height is 120, the median height is
120, and the mode is 120. The variation is 0: non-existing, absent.
Now suppose their heights are 120, 120, 135. Now there are differences: one
child is taller than the other two, who have the same height. There is some
variation now. We know how to quantify the mean, which is 125, we know how
to quantify the median, which is 120, and we know how to quantify the mode,
which is also 120. But how do we quantify the variation? Is there a lot of
variation, or just a little, and how do we measure it?

1.15.1 Range and interquartile range

One thing you could think of is measuring the distance or difference between
the lowest value and the highest value. We call this the range. The lowest
value is 120, and the highest value is 135, so the range of the data is equal to
135 − 120 = 15. As another example, suppose we have the values 20, 20, 21,

28
20, 19, 20 and 454. Then the range is equal to 454 − 19 = 435. That’s a large
range, for a series of values that for the most part hardly differ from each other.
Instead of measuring the distance from the lowest to the highest value, we
could also measure the distance between the first and the third quartile: how
much does the third quartile deviate from the first quartile? This distance or
deviation is called the interquartile range (IQR) or the interquartile distance.
Suppose that we have a large number of systolic blood pressure measurements,
where 25% are 120 or lower, and 75% are 147 or lower, then the interquartile
range is equal to 147 − 120 = 27.
Thus, we can measure variation using the range or the interquartile range.
A third measure for variation is variance, and variance is based on the sum of
squares.

1.15.2 Sum of squares

What we call a sum of squares is actually a sum of squared deviations. But
deviations from what? We could for instance be interested in how much the
values 120, 120, 135 vary around the mean of these values. The mean of these
three values equals 125. The first value differs 120 − 125 = −5, the second value
also differs 120 − 125 = −5, and the third value differs 135 − 125 = 10.
Whenever we look at deviations from the mean, some deviations are positive
and some deviations will be negative (except when there is no variation). If we
want to measure variation, it should not matter whether deviations are positive
or negative: any deviation should add to the total variation in a positive way.
Moreover, if we would add up all deviations from the mean, we would always
end up with 0, as you can see in our example. Adding up -5, -5 and +10 would
lead to a sum of 0. This would mean no variation. However, as you can see,
there is variation. So that is why we would better make all deviations positive,
and this can be done by taking the square of the deviations, since a negative
number squared is always positive. So for our three values 120, 120 and 135,
we get the deviations -5, -5 and +10, and if we square these deviations, we get
25, 25 and 100. If we add these three squares, we obtain the sum 150. This is
a sum of squared differences, or sum of squares.
In most cases, the sum of squares (SS) refers to the sum of squared deviations
from the mean. In brief, suppose you have n values of a variable Y , you first
take the mean of those values (this is Y ), you subtract this mean from each of
these n values (Yi − Y ), then you take the squares of these deviations, (Yi − Y )2 ,
and then add them up (take the sum of these squared deviations, Σ(Yi − Y )2 .
In formula form, this process looks like:

SS = Σni (Yi − Y )2 (1.3)

As an example, suppose you have the values 10, 11 and 12, then the mean
is 11. Then the deviations from the mean are -1, 0 and +1. If you square them
you get (−1)2 = 1, 02 = 0 and (+1)2 = 1, and if you sum these three values,
you get SS = 1 + 0 + 1 = 2. In formula form:

29
 SS = (Y1 − Y )2 + (Y2 − Y )2 + (Y3 − Y )2 (1.4)
= (10 − 11)2 + (11 − 11)2 + (12 − 11)2 = (−1)2 + 02 + 12 = 2

Now let’s use some values that are more different from each other, but with
the same mean. Suppose you have the values 9, 11 and 13. The average value
is still 11, but the deviations from the mean are larger. The deviations from 11
are -2, 0 and +2. Taking the squares, you get (−2)2 = 4, 02 = 0 and (+2)2 = 4
and if you add them you get SS = 4 + 0 + 4 = 8.

SS = (Y1 − Y )2 + (Y2 − Y )2 + (Y3 − Y )2 (1.5)

2 2 2 2 2 2
= (9 − 11) + (11 − 11) + (13 − 11) = (−2) + 0 + 2 = 8

Thus, the more the values differ from each other, the larger the deviations
from the mean. And the larger the deviations from the mean, the larger the
sum of squares. The sum of squares is therefore a nice measure of how much
values differ from each other.

1.15.3 Variance and standard deviation

The sum of squares can be seen as a measure of total variation: all (squared)
deviations from a certain value are added up. This means that the more data
values you have, the larger the sum of squares. Often-times, you are not
interested in the total variation, but you’re interested in the average variation.
Suppose we have the values 10, 11 and 24. The mean is then 45/3 = 15. We
have two values that are smaller than the mean and one value that is larger than
the mean, so two negative deviations and one positive deviation. Squaring them
makes them all positive. The squared deviations are 25, 16, and 81. The third
value has a huge squared deviation (81) compared to the other two values. If
we take the average squared deviation, we get (25 + 16 + 81)/3 ≈ 40.67. So the
average squared deviation is equal to 40.67. This value is called the variance. So
the variance of a bunch of values is nothing but the SS divided by the number
of values, n. The variance is the average squared deviation from the mean. The
symbol used for the variance is usually σ 2 (pronounced as ’sigma squared’).8

SS Σi (Yi − Y )2
Var(Y ) = = (1.6)
n n
8 Online Σ (Y −Y )2
you will often find the formula i n−1 i
. The difference is that here we are
talking about the definition of the variance of an observed variable Y , and that elsewhere one
talks about trying to figure out what the variance might be of all values of Y when we only see
a small portion of the values of Y . When we use all values of Y , we talk about the population
variance, denoted by σ 2 . When we only see a small part of the values of Y , we talk about a
sample of Y -values. We will come back to the distinction between population variance and
sample variance and why they differ in Chapter 2.

30
 As an example, suppose you have the values 10, 11 and 12, then the average
value is 11. Then the deviations are -1, 0 and 1. If you square them you get
(−1)2 = 1, 02 = 0 and 12 = 1, and if you add these three values, you get
SS = 1 + 0 + 1 = 2. If you divide this by 3, you get the variance: 32 . Put
differently, if the squared deviations are 1, 0 and 1, then the average squared
deviation (i.e., the variance) is 1+0+1
3 = 23 .
As another example, suppose you have the values 8, 10, 10 and 12, then
the average value is 10. Then the deviations from 10 are -2, 0, 0 and +2.
Taking the squares, you get 4, 0, 0 and 4 and if you add them you get SS = 8.
To get the variance, you divide this by 4: 8/4 = 2. Put differently, if the
squared deviations are 4, 0, 0 and 4, then the average squared deviation (i.e.,
the variance) is 4+0+0+4
4 = 2.
Often we also see another measure of variation: the standard deviation. The
standard deviation is the square root of the variance and is therefore denoted
as σ:
s
√ p Σi (Yi − Y )2
σ= σ2 = Var(Y ) = (1.7)
n

The standard deviation is often used to indicate how deviant a particular

value is from the rest of the values. Take for instance an IQ score of 105. Is that
a high IQ score or a low IQ score? Well, if someone tells you that the average
person has an IQ score of 100, you know that a score of 105 is above average.
However, still you do not know whether it is much higher than average, or just
slightly higher than average. Suppose I tell you that the standard deviation of IQ
scores is 15, then you know that a score of 105 is a third of a standard deviation
above the mean. Therefore, in order to know how deviant a particular value is
relative to a the rest of the values, one needs both a measure of central tendency
and a measure of variation. In psychological testing, IQ testing for instance, one
usually uses the mean and the standard deviation to express someone’s score
as the number of standard deviations above or below the average score. This
process of counting the number of standard deviations is called standardisation.
If we go back to the IQ score of 105, and if we want to standardise the score
in terms of standard deviations from the mean, we saw that a score of 105 was
a third of a standard deviation above the mean, so + 31 . As another example,
suppose the mean is 100 and we observe an IQ score of 80, we see that we
are 20 points below the average of 100. This is equal to 20/15 = 4/3 standard
deviations below the average, so our standardised measure equals −4/3 (note the
negative sign: it indicates we are below the mean). In general, a standardised
score can be computed by subtracting the mean and dividing the result by the
standard deviation. A standardised score for a particular value of Y , Y = y, is
usually denoted by the z-score:

y−Y
z= (1.8)
σ

31
1.16 Variance, standard deviation, and standardisation
in R
The functions var() and sd() calculate the variance and standard deviation
for a variable, respectively.

mtcars %>%
summarise(var_mpg = var(mpg),
std_mpg = sd(mpg))

## # A tibble: 1 x 2
## var_mpg std_mpg
## <dbl> <dbl>
## 1 36.3 6.03
2
q
2
i −Y ) i −Y )
However, these functions use the formulas Σi (Yn−1 and Σi (Yn−1 , respectively.
We will discuss this further in Chapter 2. If you want to use the formula
Σi (Yi −Y )2
n , you need to write your own function that computes the sum of squares
(SS) and divides by n:

var_n <- function(variable){

SS <- (variable - mean(variable))**2 %>%
sum()
return(SS/length(variable)) # dividing by N
}

mtcars %>%
summarise(var_mpg = var_n(mpg),
std_mpg = sqrt(var_n(mpg))) # taking the square root

## # A tibble: 1 x 2
## var_mpg std_mpg
## <dbl> <dbl>
## 1 35.2 5.93

Note that you get different results. For large data sets (large n), the differences
will be negligible.
Standardised measures can be obtained using the scale() function:

mtcars %>%
mutate(z_mpg = scale(mpg)) %>%
select(mpg, z_mpg)

## # A tibble: 32 x 2
## mpg z_mpg[,1]
## <dbl> <dbl>

32
## 1 21 0.151
## 2 21 0.151
## 3 22.8 0.450
## 4 21.4 0.217
## 5 18.7 -0.231
## 6 18.1 -0.330
## 7 14.3 -0.961
## 8 24.4 0.715
## 9 22.8 0.450
## 10 19.2 -0.148
## # ... with 22 more rows

1.17 Density plots

Earlier in this chapter we saw that when we have a number of values for a
numeric variable, frequency tables and frequency plots fully describe all values
of the variable that are observed. A histogram is a helpful tool to visualise
the distribution of a variable when there are so many different values that a
frequency table would be too long and a frequency plot would become too
cluttered.
A histogram can then be used to give a quick graphical overview of the
distribution. The bin width is usually chosen rather arbitrarily. Figure 1.6
shows a histogram of one million values of a numeric variable, say yearly wage
for an administrative clerk. Figure 1.7 shows a histogram for the exact same
data, but now using a much smaller bin size. You see that when you have a lot
of values, a million in this case, you can choose a very small bin size, and in
some cases this can result in a very clear shape of the distribution.
The shape of the distribution that we discern in Figure 1.7 can be represented
by a density plot. Density plots are an elegant representation of how the
frequency of certain values are distributed across a continuum. They are particularly
suited for large amounts of non-discrete (continuous) values, typically more
than 1000. Figure 1.8 shows a density plot of the one million wages. They
more or less ’smooth’ the histogram: drawing a smooth line connecting the dots
of the histogram in Figure 1.7 while looking through your eyelashes. On the
vertical axis, we no longer see ’count’ or ’frequency’, but ’density’. The quantity
density is defined such that the area under the curve equals 1. Density plots
are particularly suited for large data sets, where one is no longer interested in
the particular counts, but more interested in relative frequencies: how often are
certain values observed, relative to other values. From this density plot, it is
very clear that, relatively speaking, there are more values around 30,000 than
around 27,500 or 32,500.

33
 400000

300000
count

200000

100000

0
24000 27000 30000 33000 36000
wage

Figure 1.6: A histogram of wages with bin size 1000.

4000

3000
count

2000

1000

0
25000 27500 30000 32500 35000
wage

Figure 1.7: A histogram of wages with bin size 10.

34
 0.0004

0.0003
density

0.0002

0.0001

0.0000
25000 27500 30000 32500 35000
wage

Figure 1.8: A density plot of the wage variable.

1.18 Density plots in R

Density plots can be obtained using geom density():

mtcars %>%
ggplot(aes(x = mpg)) +
geom_density()

1.19 The normal distribution

Sometimes distributions of observed variables bear close resemblance to theoretical
distributions. For instance, Figure 1.8 bears close resemblance to the theoretical
normal distribution with mean 30,000 and standard deviation 1000. This theoretical
shape can be described with the mathematical function

1 (x−30000)2
−
f (x) = √ e 2×10002 (1.9)
2π10002
which you are allowed to forget immediately. It is only to illustrate that
distributions observed in the wild (empirical distributions) sometimes resemble
mathematical functions (theoretical distributions).
The density function of that distribution is plotted in Figure 1.9. Because
of its bell-shaped form, the normal distribution is sometimes informally called
’the bell curve’. The histogram in Figure 1.8 and the normal density function
in Figure 1.9 look so similar, they are practically indistinguishable.
Mathematicians have discovered many interesting things about the normal
distribution. If the distribution of a variable closely resembles the normal

35
 0.0004

0.0003

inflexion point inflexion point

density

0.0002 68 percent

0.0001

2.5 percent 2.5 percent

0.0000
25000 30000 35000
wage

Figure 1.9: The theoretical normal distribution with mean 30,000 and standard
deviation 1000.

distribution, you can infer many things. One thing we know about the normal
distribution is that the mean, mode and median are always the same. Another
thing we know from theory is that the inflexion points9 are one standard deviation
away from the mean. Figure 1.9 shows the two inflexion points. From theory we
also know that if a variable has a normal distribution, 68% of the observed values
lies between these two inflexion points. We also know that 5% of the observed
values lie more than 1.96 standard deviations away from the mean (2.5% on both
sides, see Figure 1.9). Theorists have constructed tables that make it easy to
see what proportion of values lies more than 1, 1.1, 1.2 . . . , 3.8, 3.9, . . . standard
deviations away from the mean. These tables are easy to find online or in books,
and these are fully integrated into statistical software like SPSS and R. Because
all these percentages are known for the number of standard deviations, it is
easier to talk about the standard normal distribution.
In such tables online or in books, you find information only about this
standard normal distribution. The standard normal distribution is a normal
distribution where all values have been standardised (see Section 1.15.3). When
values have been standardised, they automatically have a mean of 0 and a
standard deviation of 1. As we saw in Section 1.15.3, such standardised values
are obtained if you subtract the mean score from each value, and divide the
result by the standard deviation. A standardised value is usually denoted as
a z-score. Thus in formula form, a value Y = y is standardised by using the
following equation:

y−Y
z= (1.10)
σ
9 The inflexion point is where concave turns into convex, and vice versa. Mathematically,

the inflexion point can be found by equating the second derivative of a function to 0.

36
 Table 1.12: Standardising scores.
Y mean Y minus mean Z
7.2 10.4 -3.2 -0.7
8.8 10.4 -1.5 -0.3
17.8 10.4 7.4 1.6
10.4 10.4 -0.0 -0.0
10.6 10.4 0.3 0.1
18.6 10.4 8.2 1.7
12.3 10.4 1.9 0.4
3.7 10.4 -6.7 -1.4
6.6 10.4 -3.8 -0.8
7.8 10.4 -2.6 -0.5

Table 1.12 shows an example set of values for Y that are standardised. The
mean of the Y -values turns out to be 10.38, and the standard deviation 4.77.
By subtracting the mean, we ensure that the average z-score becomes 0, and
by subsequently dividing by the standard deviation, we make sure that the
standard deviation of the z-scores becomes 1.
This standardisation makes it much easier to look up certain facts about
the normal distribution. For instance, if we go back to the normally distributed
wage values, we see that the average is 30,000, and the standard deviation is
1,000. Thus, if we take all wages, subtract 30,000 and divide by 1,000, we
get standardised wages with mean 0 and standard deviation 1. The result is
shown in Figure 1.10. We know that the inflexion points lie at one standard
deviation below and above the mean. The mean is 30,000, and the standard
deviation equals 1,000, so the inflexion points are at 30000 − 1000 = 29000 and
30000 + 1000 = 31000. Thus we know that 68% of the wages are between 29,000
and 31,000.
How do we know that 68% of the observations lie between the two inflexion
points? Similar to proportions and cumulative proportions, we can plot the
cumulative normal distribution. Figure 1.11 shows the cumulative proportions
curve for the normal distribution. Note that we no longer see dots because the
variable Z is continuous.
We know that the two inflexion points lie one standard deviation below and
above the mean. Thus, if we look at a z-value of 1, we see that the cumulative
probability equals about 0.84. This means that 84% of the z-values are lower
than 1. If we look at a z-value of -1, we see that the cumulative probability
equals about 0.16. This means that 16% of the z-values are lower than -1.
Therefore, if we want to know what percentage of the z-values lie between -1
and 1, we can calculate this by subtracting 0.16 from 0.84, which equals 0.68,
which corresponds to 68%.
All quantiles for the standard normal distribution can be looked up online10
or in Appendix A, but also using R. Table 1.13 gives a short list of quantiles.
10 See for example www.normaltable.com or www.mathsisfun.com/data/standard-normal-

distribution-table.html

37
 0.4

0.3
density

0.2 68 percent

0.1

2.5 percent 2.5 percent

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
Z

Figure 1.10: The standard normal distribution.

1.00

0.84
Cumulative proportion

0.75

0.50

0.25

0.16

0.00
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5
Z

Figure 1.11: The cumulative standard normal distribution.

38
From this table, you see that 1% of the z-values is lower than -2.33, and that
25% of the z-values is lower than -0.67. We also see that half of all the z-values
is lower than 0.00 and that 10% of the z-values is larger than 1.28, and that the
1% largest values are higher than 2.33.
Although tables are readily found online, it’s helpful to memorise the so-
called 68 – 95 – 99.7 rule, also called the empirical rule. It says that 68%
of normally distributed values are at most 1 standard deviation away from the
mean, 95% of the values are at most 2 standard deviations away (more precisely,
1.96), and 99.7% of the values are at most 3 standard deviations away. In other
words, 68% of standardised values are between -1 and +1, 95% of standardised
values are between -2 and +2 (-1.96 and +1.96), and 99.7% of standardised
values are between -3 and +3.

Table 1.13: Some quantiles for the standard normal distribution.

Z cum proportion
-2.33 0.01
-1.28 0.10
-0.67 0.25
0.00 0.50
0.67 0.75
1.28 0.90
2.33 0.99

Thus, if we return to our wages with mean 30,000 and standard deviation
1,000, we know from Table 1.13 that 99% of the wages are below 30000 + 2.33
times the standard deviation = 30000 + 2.33 × 1000=32330.
Returning back to the IQ example of Section 1.15.3. Suppose we have IQ
scores that are normally distributed with a mean of 100 and a standard deviation
of 15. What IQ score would be the 90th percentile? From Table 1.13 we see
that the 90th percentile is a z-value of 1.28. Thus, the 90th percentile for our IQ
scores lies 1.28 standard deviations above the mean (above because the z-value
is positive). The mean is 100 so we have to look at 1.28 standard deviations
above that. The standard deviation equals 15, so we have to look at an IQ score
of 100 + 1.28 × 15, which equals 119.2. This tells us that 90% of the IQ scores
are equal to or lower than 119.2.
As a last example, suppose we have a personality test that measures extraversion.
If we know that test scores are normally distributed with a mean of 18 and a
standard deviation of 2, what would be the 0.10 quantile? From Table 1.13
we see that the 0.10 quantile is a z-value of -1.28. This tells us that the 0.10
quantile for the personality scores lies at 1.28 standard deviations below the
mean. The mean is 18, so the 0.10 quantile for the personality scores lies at
1.28 standard deviations below 18. The standard deviation is 2, so this amounts
to 18 − 1.28 × 2 = 15.44. This tells us that 10% of the scores on this test are
15.44 or lower.
Such handy tables are also available for other theoretical distributions. Theoretical

39
distributions are at the core of many data analysis techniques, including linear
models. In this book, apart from the normal distribution, we will also encounter
other theoretical distributions: the t-distribution (Chapter 2), the F -distribution
(Chapter 6), the chi-square distribution (Chapters 2, 8, 12, 13 and 14) and the
Poisson distribution (14).

1.20 Obtaining quantiles of the normal distribution

using R
Quantiles of a normal distribution with a certain mean and standard deviation
(sd) can be obtained using the qnorm() function:

qnorm(c(0.05, 0.50, 0.95), mean = 100, sd = 15)

## [1] 75.3272 100.0000 124.6728

This means that if you have a normal distribution with mean 100 and
standard deviation 15, 5% of the values are 75.3272 or less, 50% of the values
are 100 or less, and 95% of the values are 124.6728 or less.
If you want to know the cumulative proportion for a certain value of a
variable that is normally distributed, you can use pnorm():

pnorm(-1, mean = 0, sd = 1)

## [1] 0.1586553

So 15.86% of the values from a standard normal distribution (mean 0,

standard deviation 1), are -1 or less.

1.21 Visualising numeric variables: the box plot

We started this chapter with variables that can be stored in a data matrix. With
a variable with a large number of values on a large number of units of analysis,
it is hard to get a an intuitive feel for the data. Making a frequency table
is one way of summarising a variable, computing measures of central tendency
and variation is another way. Visualisation is probably the best way of getting a
quick and dirty feel for the information contained in a large data matrix. Earlier
in this chapter we came across frequency plots, histograms, and density plots to
visualise the distribution of a single variable. A fourth plot for a single variable
that we discuss in this book is the box plot.
A box plot gives a quick overview of the distribution of a numeric variable
in terms of its quartiles. Figure 1.12 gives an example of a box plot of (part of)
the wage data. The white box represents the interquartile range. The top of the
white box equals the third quartile, and the bottom of the white box equals the
first quartile. Therefore, we know that half of the workers have a wage between

40
29,400 and 30,800 The horizontal black line within the white box represents the
second quartile (the median), so half of the workers earn less than 30,100.
A box plot also shows whiskers: two vertical lines sprouting from the white
box. There are several ways to draw these two whiskers. One way is to draw
the top whisker to the largest value (the maximum) and the bottom whisker to
the smallest value (the minimum). Another way, used in Figure 1.12, is to have
the upper whisker extend from the third quartile to the observed value equal
to at most 1.5 times the interquartile range away from the median, and the
lower whisker extend from the first quartile to the value at most 1.5 times the
interquartile range below the median (the interquartile range is of course the
height of the white box). The dots are outlying values, or simply called outliers:
values that are even further away from the median. This is displayed in Figure
1.12. There you see first and third quartiles of 29,400 and 30,800, respectively, so
an interquartile range (IQR) of 30800 − 29400 = 1400. Multiplying this IQR by
1.5 we get 1.5 × 1400 = 2100. The whiskers therefore extend to 29400 − 2100 =
27300 and 30800 + 2100 = 32900.
Thus, the box plot is a quick way of visualising in what range the middle
half of the values are (the range in the white box), where most of the values are
(the range of the white box plus the whiskers), and where the extreme values
are (the outliers, individually plotted as dots). Note that the white box always
contains 50% of the values. The whiskers are only extensions of the box by a
factor of 1.5. In many cases you see that they contain most of the values, but
sometimes they miss a lot of values. You will see that when you notice a lot of
outliers.

1.22 Box plots in R

A box plot can be made using geom boxplot():

mtcars %>%
ggplot(aes(x = "", y = mpg)) +
geom_boxplot() +
xlab("")

1.23 Visualising categorical variables

The histogram, the density plot and the box plot can be used for numeric
variables, but also for ordinal variables that you’d like to treat numerically. For
categorical variables and ordinal variables that can’t be treated numerically, we
need other types of plots.
For example, suppose we are in a lecture hall with 456 students and we
count the number of Dutch, German, Belgian, Indian, Chinese and Indonesian
students. We could summarise the results in a frequency table (see Table 1.14),
but a bar chart shows the distribution in a more dramatic way, see Figure 1.13.

41
 33000
32800
32600
32400
32200
32000
31800
31600
31400
31200
31000
30800
30600
30400
30200
wage

30000
29800
29600
29400
29200
29000
28800
28600
28400
28200
28000
27800
27600
27400
27200
27000

Figure 1.12: A boxplot of the wages earned by a sample of 150 administrative

clerks

42
 Table 1.14: A frequency table of nationalities.
nationality n
Chinese 10
Dutch 145
German 284
Indian 7
Indonesian 10

200
count

100

0
Chinese Dutch German Indian Indonesian
nationality

Figure 1.13: A bar chart of the observed nationalities in a lecture hall.

Sometimes, counts of values of a categorical variable are displayed as a pie

chart, see Figure 1.14. Pie charts are however best avoided. First, because
compared to bar charts, they show no information about the actual counts; you
only observe relative sizes of the counts. Second, it is very hard to see from a
pie chart what the exact proportions are. For example, from the bar chart in
Figure 1.13 it is easily seen that the ratio German students to Dutch students
is about 2 to 1. Research shows that this ratio cannot be read with the same
precision from the pie chart in Figure 1.14. In sum, pie charts are best replaced
by bar charts.

Ordinal variables are often visualised using bar charts. Figure 1.15 shows
the variation of the answers to a Likert questionnaire item, where Nairobi
inhabitants are asked ”To what degree do you agree with the statement that
the climate in Iceland is agreeable?”. With ordinal variables, make sure that
the labels are in the natural order.

43
 Nationality
Chinese
Dutch
German
Indian
Indonesian

Figure 1.14: A pie chart of nationalities.

75
count

50

25

0
Co

So

Co
So
mp

me

mp
me
Dis
let

wh

Ag

let
wh
ely

ag

at

ely
ree
at
re

d
dis

isa

ag

ag
e
ag

ree
gre

ree
ree

The climate in Iceland is agreeable

Figure 1.15: Opinions on the climate in Iceland.

44
1.24 Visualising categorical and ordinal variables
in R

If a categorical variable is stored as numeric, turn it into a factor first. Then R

will treat it as categorical. A bar plot with the frequencies on the y-axis can be
made with geom bar():

mtcars %>%
mutate(cyl = factor(cyl, ordered = TRUE)) %>%
ggplot(aes(x = cyl)) +
geom_bar()

10
count

0
4 6 8
cyl

If you really want a pie chart, then do:

mtcars %>%
count(cyl) %>%
mutate(proportion = n/sum(n)) %>%
ggplot(aes(x = "",
y = proportion,
fill = factor(cyl))) +
geom_col(width = 1) +
coord_polar(theta = "y") +
xlab("") +
ylab("") +
theme_void() +
scale_fill_brewer(palette = "Blues") +
labs(fill = "Cylinders")

45
 Cylinders
4
6
8

1.25 Visualising co-varying variables

1.25.1 Categorical by categorical: cross-table
Variables are properties that vary: from person to person, or from location to
location, or from time to time, or from object to object. Sometimes when you
have two variables, you see that they co-vary: when one variable changes, the
other variable changes too. For example, suppose I have 20 pencils. These
pencils may vary in colour: twelve of them are red, and eight of them are blue.
Therefore, colour is a variable with values ”red” and ”blue”. The twenty pencils
also vary in length: four are unused and therefore still long, and sixteen of them
have been used many times so that they are short. Therefore, length is also a
variable, with values ”long” and ”short”. Note that these variables have been
measured using the same pencils. In theory I could have long blue pencils, long
red pencils, short blue pencils and short red pencils. Let’s look at the pencils
that I have: for each combination of length and colour, I count the number
of pencils. The result I put in Table 1.15.

Table 1.15: Cross-tabulation of colour and length for twenty pencils.

blue red
long 4 0
short 8 8

Such a table is called a cross-table. For every combination of two variables, I

see the number of objects (units of analysis) that have that combination. From
the table we see that there is not a single pencil that is both red and long (count
is 0). At the same time you see that all long pencils are blue. A cross-table is
therefore a nice way to show how two variables co-vary. From this particular
table for instance, you can easily see that once you know that a pencil is long,
you automatically know it is blue.

46
 Cross-tables are a nice visualisation of how two categorical variables co-vary.
But what if one of the two variables is not a categorical variable?

1.25.2 Categorical by numerical: box plot

Suppose instead of determining length by values ”short” and ”long”, we could
measure the exact length of the pencils in centimetres. The results are displayed
in Table 1.16. We see that the table is much larger than Table 1.15. We also
see quite a few cells with zeros. In most cases, for every particular combination
of length and colour we only see a count of 1 pencil. In general, you see that
when one of the variables is numeric, the cross-table becomes very large and in
addition it becomes sparse, that is, with many zeros. With such a large and
sparse table, it is hard to get a quick impression of how two variables co-vary.

Table 1.16: Cross-tabulation of colour and length for twenty pencils.

blue red
2 0 1
2.7 1 0
3.3 1 0
3.4 0 1
3.5 0 1
3.6 1 0
4.1 1 1
4.4 1 1
4.5 1 1
4.7 0 1
5.2 1 0
5.7 1 0
5.8 0 1
9 4 0

The alternative for two variables where one is categorical and the other one
is numeric, is to create a box plot. Figure 1.16 shows a box plot of the pencil
data. A box plot gives a quick overview of the distribution of the pencils: one
distribution of the blue pencils, and one distribution of the red pencils. Let’s
have a look at the distribution of the blue pencils on the left side of the plot.
The white box represents the interquartile range (IQR), so that we know that
half of the blue pencils have a length between 4 and 9. The horizontal black
line within the white box represents the median (the middle value), so half of
the blue pencils are smaller than 4.85. The vertical lines are called whiskers.
These typically indicate where the data points are that lie at most 1.5 times
the IQR away from the median. For the blue pencils, we see no whisker on top
of the white box. That means that there are no data points that lie more than
1.5 times the IQR above the median of 4.85 (here the IQR equals 5.03). We see
a whisker on the bottom of the white box, to the lowest observed value of 2.7.

47
 8

6
length

blue red
colour

Figure 1.16: A boxplot of the pencil data.

This value is less than 1.5 times 5.03 = 7.545 away from the median of 4.85 so
it is included in the whisker. It is the lowest observed value for the blue pencils
so the whisker ends there.
From a box plot like this it is easy to spot differences in the distribution of a
quantitative measure for different levels of a qualitative measure. From Figure
1.16 we easily spot that the red pencils (varying between 2 and 6 cm) tend to
be shorter than the blue pencils (varying between 3 and 9 cm). Thus, in these
pencils, length and colour tend to co-vary: red pencils are often short and
blue pencils are often long.

1.25.3 Numeric by numeric: scatter plot

Suppose we also measure the weight of my pencils in grams. Table 1.17 shows
the cross-tabulation of length and weight. This is a very sparse table (i.e.,
with lots of zeros), which makes it very hard to see any systematic co-variation
in weight and length. Figure 1.17 shows a box plot of weight and length.
Also this plot seems a bit strange, because for every observed weight value under
4 grams, there is only one observation, so that only the median can be plotted.
Therefore, in cases where we have two numeric variables, we generally use
a scatter plot. Figure 1.18 shows a scatter plot of weight by length. Now,
the relationship between weight and length is easily understood: it appears
there is a linear relationship between weight and length. For every increase
in weight, there is also an increase in length. The relationship is called linear
because we could summarise the relationship by drawing a straight line through
the dots. This line is shown in Figure 1.19.
You see that by visualising two variables, important patterns may emerge

48
 8

6
length

2
3.3 3.4 3.5 3.6 3.7 4
weight

Figure 1.17: A boxplot of the pencil data.

6
length

2
3.4 3.6 3.8 4.0
weight

Figure 1.18: A scatterplot of length and weight.

49
Table 1.17: Cross-tabulation of length (rows) and weight (columns) for twenty
pencils.
3.3 3.4 3.5 3.6 3.7 4
2 1 0 0 0 0 0
2.7 0 1 0 0 0 0
3.3 0 1 0 0 0 0
3.4 0 1 0 0 0 0
3.5 0 0 1 0 0 0
3.6 0 0 1 0 0 0
4.1 0 0 2 0 0 0
4.4 0 0 2 0 0 0
4.5 0 0 2 0 0 0
4.7 0 0 0 1 0 0
5.2 0 0 0 1 0 0
5.7 0 0 0 0 1 0
5.8 0 0 0 0 1 0
9 0 0 0 0 0 4

that you can easily overlook when only looking at the values. Cross-tables, box
plots and scatter plots are powerful tools to find regularities but also oddities
in your data that you’d otherwise miss. Some such patterns can be summarised
by straight lines, as we see in Figure 1.19. The remainder of this book focuses
on how we can use straight lines to summarise data, but also how to make
predictions for data that we have not seen yet.

1.26 Visualising two variables using R

A scatter plot for two numeric variables can be made using geom point():

mtcars %>%
ggplot(aes(x = wt, y = mpg)) +
geom_point()

A box plot for one categorical and one numeric variable can be made using
geom boxplot():

mtcars %>%
mutate(cyl = factor(cyl)) %>%
ggplot(aes(x = cyl, y = mpg)) +
geom_boxplot()

A cross table for two categorical variables can be made using table():

50
 8

6
length

2
3.4 3.6 3.8 4.0
weight

Figure 1.19: A scatterplot of length and weight, with a straight line that
summarises the relationship.

table(mtcars$cyl, mtcars$gear)

##
## 3 4 5
## 4 1 8 2
## 6 2 4 1
## 8 12 0 2

Note that the number of cylinders (first-named variable) is in the rows (here
4, 6 and 8 cylinders), and the number of gears (second-named variable) is in the
columns (3, 4, and 5 gears).

1.27 Overview of the book

Chapter 2 will introduce the problem of inference: if you only have a small
selection of data points, what can they tell us about the rest of the data? We
will use the example of a mean computed using a small number of numerical
data points and try to figure out what the mean is likely to be if we would
have all the data points. Chapter 3 discusses the same problem but then for a
proportion.
Chapter 4 will show how we can use a straight line to summarise the relationship
between two numeric variables (simple regression), where one variable is the
outcome variable, and the other variable is a predictor variable, that predicts
the value on the outcome variable. Such a straight line is a simple form of a
linear model. We also describe how we can use straight lines (linear models)
to summarise relationships between one outcome variable and more than two

51
numeric predictor variables (multiple regression). In Chapter 5 we will discuss
how you can draw conclusions about linear models for data that you have not
seen. For example, in the previous section we described the relationship between
weight and length of twenty pencils. The question that you may have is whether
this linear relationship also holds for all pencils of the same make, that is,
whether the same linear model holds for both the observed twenty pencils and
the total collection of pencils.
In Chapter 6 we will show how we can use straight lines to summarise
relationships with predictor variables that we want to treat as categorical.
Chapter 7 discusses when it is appropriate to use linear models to summarise
your data, and when it is not. It introduces methods that enable you to decide
whether to trust a linear model or not. Chapter 8 then discusses alternative
methods that you can use when linear models are not appropriate.
Chapter 9 focuses on moderation: how one predictor variable can affect the
effect that a second predictor variable has on the outcome variable.
Chapter ?? shows how you can make elaborate statements about differences
between groups of observations, in case one of the predictor variables is a
categorical variable.
Chapters 10 and 11 show how to deal with variables that are measured more
than once in the same unit of analysis (the same participant, the same pencil,
the same school, etc.). For example, you may measure the weight of a pencil
before and after you have made a drawing with it. Models that we use for
such data are called linear mixed models. Similar to linear models, linear mixed
models are not always appropriate for some data sets. Therefore, Chapter 12
discusses alternative methods to study variables that are repeatedly measured
in the same research unit.
Chapters 13 and 14 discuss generalised linear models. These are models
where the outcome variable is not numeric and continuous. Chapter 13 discusses
a method that is appropriate when the outcome variable has only two values, say
”yes” and ”no”, or ”pass” and ”fail”. Chapter 14 discusses a method that can
be used when the outcome variable is a count variable and therefore discrete,
for example the number of children in a classroom, or the number of harvested
zucchini from one plant.
Chapter 15 discusses relatively new statistical methodology that is needed
when you have a lot of variables. In such cases, traditional inferential data
analysis as discussed in the previous chapters often fails.

52
Chapter 2

Inference about a mean

2.1 The problem of inference

The human body is heavily controlled by hormones. One of the hormones
involved in a healthy reproductive system is luteinising hormone (LH). This
hormone is present in both females and males, but with different roles. In
females, a sudden rise in LH levels triggers ovulation (the release of an egg
from an ovary). We have a data set on luteinising hormone (LH) levels in one
anonymous female. The data are given in Figure 2.1. In this data set, we have
48 measures, taken at 10-minute intervals. We see that LH levels show quite
some variation over time. Suppose we want to know the mean level of luteinising
hormone level in this woman, how could we do that?
The easiest way is to compute the mean of all the values that we see in this
graph. If we do that here, we get the value 2.4. That value is displayed as the
red line in Figure 2.1. However, is that really the mean of the hormone levels
during that time period? The problem is that we only have 48 measures; we do
not have information about the hormone levels in between measurements. We
see some very large differences between two consecutive measures, which makes
the level of hormone look quite unstable. We lack information about hormone
levels in between measurements because we do not have data on that. We only
have information about hormone levels at the times where we have observed
data. For the other times, we have unobserved or missing data.
Suppose that instead of the mean of the observed hormone levels, we want
to know the mean of all hormone levels during this time period: not only those
that are measured at 10-minute intervals, but also those that are not measured
(unobserved/missing).
You could imagine that if we would measure LH not every 10 minutes, but
every 5 minutes, we would have more data, and the mean of those measurements
would probably be somewhat different than 2.4. Similarly, if we would take
measurements every minute, we again would obtain a different mean. Suppose
we want to know what the true mean is: the mean that we would get if we

53
Luteinising hormone (LH) level IU/L
3.5
3.3
3.1
2.9
2.7
2.5
2.3
2.1
1.9
1.7
1.5
1.3
1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839404142434445464748
measure

Figure 2.1: Luteinising hormone levels measured in one female, 48 measures

taken at 10-minute intervals.

would measure LH continuously, that is, an infinite number of measurements.

Unfortunately we only have these 48 measures to go on. We would like to
infer from these 48 measures, what the mean is of LH level had we measured
continuously.
This is the problem of inference: how to infer something about complete
data, when you only see a small subset of the data. The problem of statistical
inference is when you want to say something about an imagined complete data
set, the population, when you only observe a relatively small portion of the data,
the sample.
In order to show you how to do that, we do a thought experiment. Imagine
a huge data set on African elephants where we measured the height of each
elephant currently living (today around 415,000 individuals). Let’s imagine
that for this huge data set, the mean and the variance are computed: a mean
of 3.25 m and a variance of 0.14 (recall, from Chapter 1, that the variance is
a measure of spread, based on the sums of squared differences between values
and the mean). We call this data set of all African elephants currently living
the population of African elephants.
Now that we know that the actual mean equals 3.25 and the actual variance
equals 0.14, what happens if we only observe 10 of these 415,000 elephants? In
our thought experiment we randomly pick 10 elephants. Random means that
every living elephant has an equal chance of being picked. This random sample
of 10 elephants is then used to compute a mean and a variance. Imagine that we
do this exercise a lot of times: every time we pick a new random sample of 10
elephants, and you can imagine that each time we get slightly different values
for our mean, but also for our variance. This is illustrated in Table 2.1, where
we show the data from 5 different samples (in different columns), together with
5 different means and 5 different variances.
What we see from this table is that the 5 sample means vary around the
population mean of 3.25, and that the 5 variances vary around the population

54
variance of 0.14. We see that therefore the mean based on only 10 elephants
gives a rough approximation of the mean of all elephants: the sample mean
gives a rough approximation of the population mean. Sometimes it is too low,
sometimes it is too high. The same is true for the variance: the variance based
on only 10 elephants is a rough approximation, or estimate, of the variance of
all elephants: sometimes it is too low, sometimes it is too high.

Table 2.1: Imaginary data on elephant height when 5 random samples (columns)
of 10 elephants (rows) are drawn from the population data.
1 2 3 4 5
1 3.77 2.52 3.26 3.61 3.16
2 3.61 3.41 3.09 3.33 2.74
3 3.12 2.91 3.14 3.22 3.91
4 2.95 3.20 2.85 3.40 3.60
5 2.53 3.45 2.69 3.20 3.19
6 3.12 3.11 3.45 2.31 2.94
7 3.31 3.22 2.98 3.65 4.39
8 2.59 3.76 2.81 2.20 3.24
9 2.91 3.44 3.63 3.12 3.21
10 3.36 2.84 4.15 2.73 2.75
mean 3.13 3.19 3.20 3.08 3.31
variance 0.14 0.12 0.18 0.23 0.24

2.2 Sampling distribution of mean and variance

How high and how low the sample mean can be, is seen in Figure 2.2. There you
see a histogram of all sample means when you draw 10,000 different samples of
each consisting of 10 elephants and for each sample compute the mean. This
distribution is a sampling distribution. More specifically, it is the sampling
distribution of the sample mean.
The red vertical line indicates the mean of the population data, that is, the
mean of 3.25 (the population mean). The blue line indicates the mean of all
these sample means together (the mean of the sample means). You see that
these lines practically overlap.
What this sampling distribution tells you, is that if you randomly pick 10
elephants from a population, measure their heights, and compute the mean, this
mean is on average a good estimate (approximation) of the mean height in the
population. The mean height in the population is 3.25, and when you look at the
sample means in Figure 2.2, they are generally very close to this value of 3.25.
Another thing you may notice from Figure 2.2 is that the sampling distribution
of the sample mean looks symmetrical and resembles a normal distribution.
Now let’s look at the sampling distribution of the sample variance. Thus,
every time we randomly pick 10 elephants, we not only compute the mean but

55
 1000
count

500

0
2.75 3.00 3.25 3.50 3.75
sample mean

Figure 2.2: A histogram of 10,000 sample means when the sample size equals
10.

750
count

500

250

0
0.0 0.1 0.2 0.3
sample variance

Figure 2.3: A histogram of 10,000 sample variances when the sample size equals
10. The red line indicates the population variance. The blue line indicates the
mean of all variances observed in the 10,000 samples.

56
also the variance. Figure 2.3 shows the sampling distribution. The red line
shows the variance of the height in the population, and the blue line shows the
mean variance observed in the 10,000 samples. Clearly, the red and blue line do
not overlap: the mean variance in the samples is slightly lower than the actual
variance in the population. We say that the sample variance underestimates the
population variance a bit. Sometimes we get a sample variance that is lower
than the population value, sometimes we get a value that is higher than the
population value, but on average we are on the low side.

Overview

ˆ population: all values, both observed and unobserved

ˆ population mean: the mean of all values (observed and unobserved

values)
ˆ sample: a limited number of observed values

ˆ sample size: the number of observed values

ˆ sample mean: the mean of the values in the sample

ˆ random sample: values that you observe when you randomly pick a
subset of the population

ˆ random: each value in the population has an equal probability of being

observed
ˆ sampling distribution of the sample mean: the distribution of means
that you get when you randomly pick new samples from a population and
for each sample compute the mean
ˆ sampling distribution of the sample variance: the distribution of
variances that you get when you randomly pick new samples from a
population and for each sample compute the variance

2.3 The effect of sample size

What we have seen so far is that when the population mean is 3.25 m and
we observe only 10 elephants, we may get a value for the sample mean of
somewhere around 3.25, but on average, we’re safe to say that the sample mean
is a good approximation for the population mean. In statistics, we call the
sample mean an unbiased estimator of the population mean, as the expected
value (the average value we get when we take a lot of samples) is equal to the
population value.
Unfortunately the same could not be said for the variance: the sample
variance is not an unbiased estimator for the population variance. We saw

57
that on average, the values for the variances are too low.
Another thing we saw was that the distribution of the sample means looked
symmetrical and close to normal. If we look at the sampling distribution of
the sample variance, this was less symmetrical, see Figure 2.3. It actually has
the shape of a so-called χ2 -(pronounced ’chi-square’) distribution, which will be
discussed in Chapters 8, 12, 13 and 14. Let’s see what happens when we do not
take samples with 10 elephants each time, but 100 elephants.

Stop and think: What will happen to the sampling distributions of the mean
and the variance? For instance, in what way will Figure 2.2 change when we
use 100 elephants instead of 10?

Figure 2.4 shows the sampling distribution of the sample mean. Again the
distribution looks normal, again the blue and red lines overlap. The only
difference with Figure 2.2 is the spread of the distribution: the values of the
sample means are now much closer to the population value of 3.25 than with
a sample size of 10. That means that if you use 100 elephants instead of 10
elephants to estimate the population mean, on average you get much closer to
the true value!

Now stop for a moment and think: is it logical that the sample means are
much closer to the population mean when you have 100 instead of 10 elephants?

Yes, of course it is, with 100 elephants you have much more information about
elephant heights than with 10 elephants. And if you have more information,
you can make a better approximation (estimation) of the population mean.
Figure 2.5 shows the sampling distribution of the sample variance. Compared
to a sample size of 10, the shape of the distribution now looks more symmetrical
and closer to normal. Second, similar to the distribution of the means, there is
much less variation in values: all values are now closer to the true value of 0.14.
And not only that: it also seems that the bias is less, in that the blue and the
red lines are closer to each other.
Here we see three phenomena. The first is that if you have a statistic like
a mean or a variance and you compute that statistic on the basis of randomly
picked sample data, the distribution of that statistic (i.e., the sampling distribution)
will generally look like a normal distribution if sample size is large enough.
It can actually be proven that the distribution of the mean will become a
normal distribution if sample size becomes large enough. This phenomenon is
known as the Central Limit Theorem. It is true for any population, no matter
what distribution it has.1 Thus, this means that height in elephants itself does
not have to be normally distributed, but the sampling distribution of the sample
mean will be normal for large sample sizes (e.g., 100 elephants).
The second phenomenon is that the sample mean is an unbiased estimator
1 This is true except for the case that you have fewer than 3 data points and for a few

special cases, that you don’t need to know about in this book.

58
of the population mean, but that the variance of the sample data is not an
unbiased estimator of the population variance. Let’s denote the variance of the
sample data as S 2 . Remember from Chapter 1 that the formula for the variance
is

Σ(yi − ȳ)2
S 2 = Var(Y ) = (2.1)
n

We saw that the bias was large for small sample size and small for larger
sample size. So somehow we need to correct for sample size. It turns out that
n
the correction is a multiplication with n−1 :

n
s2 = S2 (2.2)
n−1

where s2 is the corrected estimator of population variance, S 2 is the variance

observed in the sample, and n is sample size. When we rewrite this formula and
cancel out n, we get a more direct way to compute s2 :

Σ(yi − ȳ)2
s2 = (2.3)
n−1

Thus, if we are interested to know the variance or the standard deviation in

the population, and we only have sample data, it is better to take the sums of
squares and divide by n − 1, and not by n.

2
c2 = s2 = Σ(yi − ȳ)
σ (2.4)
n−1

where σc2 (pronounced ’sigma-squared hat’) signifies the estimator of the

population variance (the little hat stands for estimator or estimated value).
The third phenomenon is that if sample size increases, the variability of the
sample statistic gets smaller and smaller: the values of the sample means and
the sample variances get closer to their respective population values. We will
delve deeper into this phenomenon in the next section.

59
 4000

3000
count

2000

1000

0
2.75 3.00 3.25 3.50 3.75
means

Figure 2.4: A histogram of 10,000 sample means when the sample size equals
100.

2500

2000

1500
count

1000

500

0
0.0 0.1 0.2 0.3
var

Figure 2.5: A histogram of 10,000 sample variances when the sample size equals
100.

60
Overview
ˆ Central Limit Theorem: says that the sampling distribution of the
sample mean will be normally distributed for infinitely large sample sizes.
ˆ estimator: a quantity that you compute based on sample data, that
you hope says something about a quantity in the population data. For
instance, you can use the sample mean and hope that it is close to the
population mean. You use the sample mean as an approximation of the
population mean.
ˆ estimate: the actual value that you get when computing an estimator.
For instance, we can use the sample mean as the estimator of the
population mean. The formula for the sample mean is Σy n so this formula
i

is our estimator. Based on a sample of 10 values, you might get a sample

mean of 3.5. Then 3.5 is the estimate for the population mean.
ˆ unbiased estimator: an estimator that has the population value as
expected value (the mean that you get when averaging over many
samples). For example, the sample mean is an unbiased estimator for
the population mean because if you draw an infinite number of samples,
the mean of the sample means will be equal to the population mean.
ˆ biased estimator: an estimator that does not have the population value
as expected value. For example, the variance calculated using a sample
is a biased estimator for the population variance because if you draw an
infinite number of samples, the mean of the variances will not be equal to
the population variance.
ˆ S 2 : the variance of the values in the sample, computed by taking the sums
of squares and divide by sample size n.
ˆ s2 : an unbiased estimator for the population variance, often confusingly
called the ’sample variance’, computed by taking the sums of squares and
divide by n − 1.

2.4 The standard error

In Chapter 1 we saw that a measure for spread and variability was the variance.
In the previous section we saw that with sample size 100, the variability of the
sample mean was much lower than with sample size 10. Let’s look at this more
closely.
When we look at the sampling distribution in Figure 2.2 with sample size
10, we see that the means lie between 2.8 and 3.71. If we compute the standard
deviation of the sample means, we obtain a value of 0.118. This standard
deviation of the sample means is technically called the standard error, in this

61
case the standard error of the mean. It is a measure of how uncertain we are
about a population mean when we only have sample data to go on. Think about
this: why would we associate a large standard error with very little certainty?
In this case we have only 10 data points for each sample, and it turns out that
the standard error of the mean is a function of both the sample size n and the
population variance σ 2 .

r
σ2
σȳ = (2.5)
n

Here, the
q population variance equals 0.14 and sample size equals 10, so the
σȳ equals 0.14
10 = 0.118, close to our observed value. If we fill in the formula
for a sample size of 100, we obtain a value of 0.037. This is a much smaller
value for the spread and this is indeed observed in Figure 2.4. Figure 2.6 shows
the standard error of the mean for all sample sizes between 1 and 200.
In sum, the standard error of the mean is the standard deviation of the
sample means, and serves as a measure of the uncertainty about the population
mean. The larger the sample size, the smaller the standard error, the closer a
sample mean is expected to be around the population mean, the more certain
we can be about the population mean.
Similar to the standard error of the mean, we can compute the standard
error of the variance. This is more complicated – especially if the population
distribution is not normal – and we do not treat it here. Software can do
the computations for you, and later in this book you will see examples of the
standard error of the variance.
Summarising the above: when we have a population mean, we usually see
that the sample mean is close to it, especially for large sample sizes. If you do
not understand this yet, go back before you continue reading.
The larger the sample size, the closer the sample means are to the population
means. If you turn this around, if you don’t know the population mean, you
can use a large sample size, calculate the sample mean, and then you have a
fairly good estimate for the population. This is useful for our problem of the
LH levels, where we have 48 measures. The mean of the 48 measurements could
be a good approximation of the mean LH level in general.
As an indication of how close you are to the population mean, the standard
error can be used. The standard error of the mean is the standard deviation of
the sampling distribution of the sample mean. The smaller the standard error,
the more confident you can be that your sample mean is close to the population
mean. In the next section, we look at this more closely. If we use our sample
mean as our best guess for the population mean, what would be a sensible range
of other possible values for the population mean, given the standard error?

62
 0.20
standard error of the mean

0.15

0.10

0.05

0 50 100 150 200

sample size

Figure 2.6: Relationship between sample size and the standard error of the
mean, when the population variance equals 0.14.

Overview
ˆ standard error of the mean: the standard deviation of the distribution
of sample means (the sampling distribution of the sample mean). Says
something about how spread out the values of the sample means are. It
can be used to quantify the uncertainty about the population mean when
we only have the sample mean to go on.
ˆ standard error of the variance: the standard deviation of the sampling
distribution of the sample variance. Says something about how spread
out the values of the sample variances are. It can be used to quantify the
uncertainty about the population variance when we only have the variance
of the sample values to go on.

2.5 Confidence intervals

If we take a sample mean as our best guess of the population mean, we know
that we are probably a little bit off. If we have a large standard error we know
that the population mean could be very different from our best guess, and if we
have a small standard error we know that the true population mean is pretty
close to our best guess, but could we quantify this in a better way? Could we
give a range of plausible values for the population mean?
In order to do that, let’s go back to the elephants: the true population mean
is 3.25 m with variance 0.14. What would possible values of sample means look
like if sample size is 4? Of course it would look like the sampling distribution
of the sample mean with a sample size of 4. Its mean would be the population
mean of 3.25 and its standard deviation would be equivalent to the standard

63
error,qcomputed as a function of the population variance and sample size, in our
case 0.14 4 = 0.19. Now imagine that for a bunch of samples we compute the
sample means. We know that the means for large sample sizes will look more
or less like a normal distribution, but how about for a small sample size like
n = 4? If it would look like a normal distribution too, then we could use the
knowledge about the standard normal distribution to say something about the
distribution of the sample means.
For the moment, let’s assume the sample size is not 4, but 4000. From
the Central Limit Theorem we know that the distribution of sample means is
almost identical to a normal distribution, so let’s assume it is normal. From
the normal distribution, we know that 68% of the observations lies between 1
standard deviation below and 1 standard deviation above the mean (see Section
1.19 and Figure 1.9). If we would therefore standardise our sample means, we
could say something about their distribution given the standard error, since the
standard error is the standard deviation of the sampling distribution. Thus, if
the sampling distribution looks normal, then we know that 68% of the sample
means lies between one standard error below the population mean and one
standard error above the population mean.
So suppose we take a large number of samples from the population, compute
means and variances for each sample, so that we can compute standardised
scores. Remember from Chapter 1 that a standardised score is obtained by
subtracting an observed score from the mean and divide by the standard deviation:
y − ȳ
zy = (2.6)
sdy
If we apply standardisation of the sample means, we get the following: for
a given sample mean ȳ we subtract the population mean µ and divide by the
standard deviation of the sample means (the standard error):
ȳ − µ
zȳ = (2.7)
σȳ
If we then have a bunch of standardised sample means, their distribution
should have a standard normal distribution with mean 0 and variance 1. We
know that for this standard normal distribution, 68% of the values lie between
-1 and +1, meaning that 68% of the values in a non-standardised situation lie
between -1 and +1 standard deviations from the mean (see Section 1.19). That
implies that 68% of the sample means lie between -1 and +1 standard deviations
(standard errors!) from the population mean. Thus, 68% of the sample means
lie between −1×σȳ and +1×σ
q ȳ from the population mean µ. If we have sample
0.14
size 4000, σȳ is equal to 4000 = 0.0059161 and µ = 3.25, so that 68% of the
sample means lie between 3.2440839 and 3.2559161.
This means that we also know that 100 − 68 = 32% of the sample means
lie farther away from the mean: that it occurs in only 32% of the samples that
a sample mean is smaller than 3.2440839 and larger than 3.2559161. Taking
this a bit further, since we know that 95% of the values in a standard normal

64
distribution lie between -1.96 and +1.96 (see Section 1.19), we know that it
happensqin only 5% of the samples that the sample mean isqsmaller than 3.25 −
0.14 0.14
1.96 × 4000 = 3.2384045 or larger than 3.25 + 1.96 × 4000 = 3.2615955.
Another way of putting this is that
q it happens in only 95% of the samples that
0.14
a sample mean is at most 1.96 × 4000 away from the population mean 3.25.
This distance of 1.96 times the standard error is called the margin of error
(MoE). Here we focus on the margin of error that is based on 95% observations
of the observations seen in the normal distribution:

M oE0.95 = z0.95 × σȳ = 1.96 × σȳ (2.8)

where z0.95 is the standardised value z for which holds that 95% of the values
are between µ − z and µ + z (i.e., 1.96).
Knowing the population mean, we know that it is very improbable (5%) that
a sample mean is farther away from the population mean than this margin of
error. The next step is tricky, so pay close attention. If we know the population
mean, we can construct an interval based on the margin of error for where we
expect sample means to lie. In the above case, knowing that the population
mean is 3.25, and we use an MoE based on 95%, we expect that 95% of the
sample means will lie between 3.25 − MoE and 3.25 + MoE.
But what if we don’t know the population mean, but do know the sample
mean? We could use the same interval but centred around the sample mean
instead of the population mean. Thus, we have a 95% interval if we take the
sample mean as the centre and the MoE around it. Suppose that we randomly
draw 4000 elephants and we obtain a sample mean of ȳ = 3.26, then we construct
the 95% interval as running from ȳ − M oE = 3.26 − M oE to ȳ + M oE =
3.26 + M oE. The margin of error is based on the standard error, which is in
turn dependent on the population variance. If we don’t know that, we have to
2
estimate it from the sample. So suppose we find a sample q variance s = 0.15,
0.15
we get the 95% interval from ȳ − M oE = 3.26 − 1.96 × 4000 to ȳ + M oE =
q
0.15
3.26 + 1.96 × 4000 .
Such an interval, centred around the sample mean, is called a confidence
interval. Because it is based on 95% of the sampling distribution (centred around
the population mean) it is called a 95% confidence interval.
One way of thinking about this interval is that it represents 95% of the
sample means had the population mean been equal to the sample mean. For
example, a 95% interval around the sample mean of 3.26 represents 95% of
the sample means that you would get if you would take many random samples
from a population distribution with mean 3.26: the middle 95% of the sampling
distribution for a population mean of 3.26.

A 95% confidence interval contains 95% of the sample means had the population
mean been equal to the sample mean. Its construction is based on the estimated
sampling distribution of the sample mean.

65
 60

40
density

20

MoE MoE

0
3.22 3.23 3.24 3.25 3.26 3.27 3.28 3.29
Height in m

Figure 2.7: Illustration of the construction of a 95% confidence interval. Suppose

we find a sample mean of 3.26 and a sample variance of 0.15, with N = 4000.
The black curve represents the sampling distribution if the population mean
would be 3.26 and a variance of 0.15. In reality, we don’t know the population
mean, it could be 3.25 or any other value. The sampling distribution for 3.25
is shown by the blue curve. Whatever the case, the length of an interval that
contains 95% of the sample means is always the same: twice the margin of error.
This interval centred around the sample mean, is called the 95% confidence
interval.

The idea is illustrated in Figure 2.7. There you see two sampling distributions:
one for if the population mean is 3.25 (blue) and one for if the population mean
is 3.26 (black). Both are normal distributions because sample size is large, and
both have the same standard error that can be estimated using the sample
variance. Whatever the true population mean, we can estimate the margin of
error that goes with 95% of the sampling distribution. We can then construct
an interval that stretches the length of about twice (i.e., 1.96) the margin of
error around any value. We can do that for the real population mean (in blue),
but the problem that we face in practice is that we don’t know the population
mean. We do know the sample mean, and if we centre the interval around that
value, we get what is called the 95% confidence interval. We see that it ranges
from 3.248 to 3.272. This we can use as a range of plausible values for the
unknown population mean. With some level of ’confidence’ we can say that the
population mean is somewhere in this interval.
Note that when we say: the 95% confidence interval runs from 3.248 to
3.272, we cannot say, we are 95% sure that the population mean is in there.
’Confidence’ is not the same as probability. We’ll talk about this in a later
section. First, we look at the situation where sample size is small so that we
cannot use the Central Limit Theorem.

66
2.6 The t-statistic
In the previous section, we constructed a 95% confidence interval based on the
standard normal distribution. We know from the standard normal distribution
that 95% of the values are between -1.96 and +1.96. We used the standard
normal distribution because the sampling distribution will look normal if sample
size is large. We took the example of a sample size of 4000, and then this
approach works fine, but remember that the actual sample size was 4. What if
sample size is not large? Let’s see what the sampling distribution looks like in
that case.
Remember from the previous section that we standardised the sample means.

ȳ − µ
zȳ =
σȳ

and that zȳ has a standard normal distribution. But, this only works if we
have a good estimate of σȳ , the standard error. If sample size is limited, our
estimate is not perfect. You can probably imagine that if you take one sample
of 4 randomly
q selected elephants, you get one value for the estimated standard
2
error ( sn ), and if you take another sample of 4 elephants, you get a slightly
different value for the estimated standard error. Because we do not always have
a good estimate for σȳ , the standardisation becomes a bit more tricky. Let’s
call the standardised sample mean t instead of z:

ȳi − µ
tȳi = q 2
si
n

Thus, a standardised sample mean for sample i, will be constructed using an

estimate for the standard error by computing the sample variance s2 for sample
i.
If you standardise every sample mean, each time using a slighly different
standard deviation, and you plot a histogram of the t-values, you do not get a
standard normal distribution, but a slightly different one.
In summary: if you know the standard error (because you know the population
variance), the standardised sample means will show a normal distribution. If
you don’t know the standard error, you have to estimate it based on the sample
variance. If sample size is really large, you can estimate the population variance
pretty well, and the sample variances will be very similar to each other. In that
case, the sampling distribution will look very much like a normal distribution.
But if sample size is relatively small, each sample will show a different sample
variance, resulting in different standard error estimates. If you standardise each
sample mean with a different standard error, the sampling distribution will
not look normal. This distribution is called a t-distribution. The difference
between this distribution and the standard normal distribution is shown in
Figure 2.8. The blue curve is the standard normal distribution, the red curve is

67
 0.4

0.3

Distribution
density

0.2 normal
t

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5

Figure 2.8: Distribution of t with sample size 4, compared with the standard
normal distribution.

ȳ i −µ
the distribution we get if we have sample size 4 and we compute tȳi = r for
s2
i
n
many different samples.
When you compare the two distributions, you see that compared to the
normal curve, there are fewer observations around 0 for the t-distribution: the
density around 0 is lower for the red curve than for the blue curve. That’s
because there are more observations far away from 0: in the tails of the distributions,
you see a higher density for the red curve (t) than for the blue curve (normal).
They call this phenomenon ’heavy-tailed’: relatively more observations in the
tails than around the mean.
That the t-distribution is heavy-tailed has important implications. From
the standard normal distribution, we know that 5% of the observations lie more
than 1.96 away from the mean. But since there are relatively more observations
in the tails of the t-distribution, 5% of the values lie farther away from the
mean than 1.96. This is illustrated in Figure 2.9. If we want to construct a 95%
confidence interval, we can therefore no longer use the 1.96 value.
With this t-distribution, 95% of the observations lie between -3.18 and +3.18.
Of course, that is in the standardised situation. If we move back to our scale
of elephant heights with a sample mean of 3.26, we have to transform this back
to elephant heights. So -3.18q times the standard error away from the mean of
0.15
3.26, is equal to 3.26 − 3.18 × = 2.6441956, and +3.18 times the standard
4
q
error away from the mean of 3.26, is equal to 3.26 + 3.18 × 0.15 4 = 3.8758044.
So the 95% interval runs from 2.64 to 3.88. This interval is called the 95%
confidence interval, because 95% of the sample means will lie in this interval, if
the population mean would be 3.26.
Notice that the interval includes the population mean of 3.25. If we would
interpret this interval around 3.26 as containing plausible values for the population
mean, we see that in this case, this is a fair conclusion, because the true value

68
 0.4

0.3

Distribution
density

0.2 normal
t

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5

Figure 2.9: Distribution of t with sample size 4, compared with the standard
normal distribution. Shaded areas represent 2.5% of the respective distribution.

3.25 lies within this interval.

2.7 Interpreting confidence intervals

The interpretation of confidence intervals is very difficult, and it often goes
wrong, even in many textbooks on the matter.
One thing that should be very clear is that a confidence interval is constructed
as if you know the population mean and variance, which, of course, you don’t.
We assume that the population is a certain value, say q µ = m0 , we assume
2
that the standard error of the mean is equal to σȳ = sn , and we know that
if we would look at many many samples and compute standardised sample
means, their distribution would be a t-distribution. Based on that t-distribution,
we know in which interval 95% of the standardised sample means would lie
and we use that to compute the margin of error and to construct an interval
around the sample mean that we actually obtain. A lot of this reasoning is
imagination: imagining that you know the population mean and that you have
a good estimate for the population variance. Then you imagine what sample
means would be reasonable to find. But of course, it’s in fact the opposite: you
only know the sample mean and sample variance and you want to know what
are plausible values for the population mean.
You have to bear this reversal in mind when interpreting the 95% confidence
interval around a sample mean. Many people state the following: with 95%
probability, the 95% confidence interval contains the population mean. This
is wrong. It is actually the opposite: the 95% interval around the population
mean contains 95% of the sample means.
If you know the population mean µ, then 95% of the confidence intervals that
you construct around the sample means that you get from random sampling will
contain the mean µ. This is illustrated in Figure 2.10. Suppose we take µ = 3.25.

69
Then if we imagine that we take 100 random samples from this population
distribution, we can calculate 100 sample means and 100 sample variances. If
we then construct 100 confidence intervals around these 100 sample means, we
obtain the confidence intervals displayed in Figure 2.10. We see that 95 of these
intervals contain the value 3.25, and 5 of them don’t: only in samples 1, 15, 20,
28 and 36, the interval does not contain 3.25.
It can be mathematically shown that given a certain population mean, when
taking many, many samples and constructing 95% confidence intervals, you can
expect 95% of them will contain that population mean. That does not mean
however that given a sample mean with a certain 95% interval, that interval
contains the population mean with a probability of 95%. It only means that were
this procedure of constructing confidence intervals to be repeated on numerous
samples, the fraction of calculated confidence intervals that contain the true
population mean would tend toward 95%. If you only do it once (you obtain a
sample mean and you calculate the 95% confidence interval) it either contains
the population mean or it doesn’t: you cannot calculate a probability for this.
In the statistical framework that we use in this book, one can only say something
about the probability of data occurring given some population values:

Given that the population value is 3.25, and if you take many, many independent
samples from the population, you can expect that 95% of the confidence intervals
constructed based on resulting sample means will contain that population value
of 3.25.

Using this insight, we therefore conclude that the fact we see the value of
3.25 in our 95% confidence interval around 2.9, gives us some reason to believe
(’confidence’) that 3.25 could also be a plausible candidate for the population
mean.
Summarising, if we find a sample mean of say 2.9, we know that 2.9 is a
reasonable guess for the population mean (it’s an unbiased estimator). Moreover,
if we construct a 95% confidence interval around this sample mean, this interval
contains other plausable candidates for the population mean. However, it might
be possible that the true population mean is not included.

2.8 t-distributions and degrees of freedom

The standardised deviation of a sample mean from a hypothesised population
mean has a t-distribution. This happens when the population variance is not
known, and we therefore have to estimate the standard error based on the
sample variance. Because of this uncertainty about the population variance and
consequently the standard error, the standardised score does not have a normal
distribution but a t-distribution.
In the previous section we saw the distribution for the case that we had a
sample size of 4. With such a small sample size, we have a very inaccurate
estimate of the population variance. The sample variance s2 will be very

70
 100
99
98
97
96
95
94
93
92
91
90
89
88
87
86
85
84
83
82
81
80
79
78
77
76
75
74
73
72
71
70
69
68
67
66
65
64
63
62
61
60
59
58
57
56
55
54
53
sample

52
51
50
49
48
47
46
45
44
43
42
41
40
39
38
37
36
35
34
33
32
31
30
29
28
27
26
25
24
23
22
21
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1

2.6 2.7 2.8 2.9 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8
x

Figure 2.10: Confidence intervals

71
different for every new sample of size 4. But if sample size increases, our
estimates for the population variance will become more precise, and they will
show less variability. This results in the sampling distribution to become less
heavy-tailed, until it closely resembles the normal distribution for very large
sample sizes.
This means that the shape of the sampling distribution is a t-distribution but
that the shape of this t-distribution depends on sample size. More precisely, the
shape of the t-distribution depends on its so-called degrees of freedom (explained
below). Degrees of freedom are directly linked to the sample size. Degrees of
freedom can be as small as 1, very large like 250, or infinitely larger. The t-
distribution with a very large number like 2500, is practically indistinguishable
from a normal distribution. However for a relatively low number of degrees
of freedom, the shape is very different: relatively more observations are in the
tails of the distribution and less so in the middle, compared to the normal
distribution, see Figure 2.11.
The shape of the t-distribution is determined by its degrees of freedom: the
higher the degrees of freedom, the more it resembles the normal distribution.
So which t-distribution do we have to use when we are dealing with sample
means and we want to infer something about the population mean, and what
are degrees of freedom? As stated already above, the degrees of freedom is
directly related to sample size: sample size determines the degrees of freedom
of the t-distribution that we need. Degrees of freedom stands for the amount
of information that we have and of course that depends on how many data
values we have. In its most general case, the number of degrees of freedom is
the number of values in the final calculation of a statistic that are free to vary.
More specifically in our case, the degrees of freedom for a statistic like t are
equal to the number of independent scores that go into the estimate, minus
the number of parameters used as intermediate steps in the estimation of the
parameter itself.
In the example above we had information about 4 elephants (4 values), so
our information content is 4. However, remember that when we construct our
t-value, we have to first compute the sample mean in order to compute the
sample variance s2 . But, suppose you know the sample mean, you don’t have to
know all the 4 values anymore. Suppose the heights of the first three elephants
are 3.24, 3.25 and 3.26, and someone computes the mean of all four elephants as
3.25, then you automatically know that the fourth elephant has a height of 3.25
(why?). Thus, once you know the mean of n elephants, you can give imaginary
values for the heights of only n − 1 elephants, because given the other heights
and the mean, it is already determined.
The same is true for the t-statistic: once you know 3 elephant heights and
statistic t, then you know the height of the fourth elephant automatically.
Because we assume the mean in our computation of s2 (we fix it) we loose one
information point, leaving 3. The shape of the standardised scores of fictitious
new samples then looks like a t-distribution with 3 degrees of freedom.
Generally, if we have a sample size of n and the population variance is
unknown, the shape of the standardised sample means (i.e., t-scores) of fictitious

72
 0.4

0.3 Distribution
normal
density

0.2 t, df = 1
t, df = 3
t, df = 9
0.1

0.0
−5.0 −2.5 0.0 2.5 5.0

Figure 2.11: Difference in the shapes of the standard normal distribution and
t-distributions with 1, 3 and 9 degrees of freedom.

new samples is that of a t-distribution with n − 1 degrees of freedom.

2.9 Constructing confidence intervals

In previous sections we discussed the 95% confidence interval, because it is the
most widely used interval. But other intervals are also seen, for instance 99%
confidence intervals or 90% confidence intervals. A 99% confidence interval
is wider than a 95% confidence interval, which in turn is wider than a 90%
confidence interval. The width of the confidence interval also depends on the
sample size. Here we show how to construct 90%, 99% and other intervals, for
different sample sizes.
As we discussed for the 95% interval above, we looked at the t-distribution
of 3 degrees of freedom because we had a sample size of 4 elephants. Suppose
we have a sample size of 200, then we would have to look at a t-distribution of
200−1 = 199 degrees of freedom. Table 2.2 shows information about a couple of
t-distributions with different degrees of freedom. In the first column, cumulative
probabilities are given, and the next column gives the respective quantiles. For
instance, the column ’norm’ shows that a cumulative proportion of 0.025 is
associated with a quantile of -1.96 for the standard normal distribution. This
means that for the normal distribution, 2.5% of the observations are smaller
than -1.96. In the same column we see that the quantile 1.96 is associated with
a cumulative probability of 0.975. This means that 97.5% of the observations in
a normal distribution are smaller than 1.96. This implies that 100% - 97.5% =
2.5% of the observations are larger than 1.96. Thus, if 2.5% of the observations
are larger than 1.96 and 2.5% of the observations are smaller than -1.96, then
5% of the observations are outside the interval (-1.96, 1.96), and 95% are inside
this interval.
From Table 2.2, we see that for such a 95% interval, we have to use the

73
values -1.96 and 1.96 for the normal distribution, but for the t-distribution we
have to use other values, depending on the degrees of freedom. We see that for 3
degrees of freedom, we have to use the values -3.18 and 3.18, and for 199 degrees
of freedom the values -1.97 and +1.97. This means that for a t-distribution with
3 degrees of freedom, 95% of the observations lie in the interval from -3.18 to
3.18. Similarly, for a t-distribution with 199 degrees of freedom, the values for
cumulative probabilities 0.025 and 0.975 are -1.97 and 1.97 respectively, so we
can conclude that 95% of the observations lie in the interval from -1.97 to 1.97.

Now instead of looking at 95% intervals for the t-distribution, let’s try to
construct a 90% confidence interval around an observed sample mean. With
a 90% confidence interval, 10% lies outside the interval. We can divide that
equally to 5% on the low side and 5% on the high side. We therefore have to
look at cumulative probabilities 0.05 and 0.95 in Table 2.2. The corresponding
quantiles for the normal distribution are -1.64 and 1.64, so we can say that for
the normal distribution, 90% of the values lie in the interval (-1.64, 1.64). For
a t-distribution with 9 degrees of freedom, we see that the corresponding values
are -1.83 and 1.83. Thus we conclude that with a t-distribution with 9 degrees
of freedom, 90% of the observed values lie in the interval (-1.83, 1.83).

However, now note that we are not interested in the values of the t-distribution,
but in likely values for the population mean. The standard normal and the
t-distribution are standardised distributions. In order to get values for the
confidence interval around the sample mean, we have to unstandardise the
values. The value of 1.83 above means ”1.83 standard errors away from the mean
(the sample mean)”. So suppose we find a sample mean of 3, with a standard
error of 0.5, then we say that a 90% confidence interval for the population mean
runs from 3 − 1.83 × 0.5 to 3 + 1.83 × 0.5, so from 2.085 to 3.915.

Follow these steps to compute a x% confidence interval:

74
Constructing confidence intervals
1. Compute the sample mean ȳ.
(yi −ȳ)
2. Estimate the population variance s2 = Σi n−1 .
q
2
3. Estimate the standard error σ̂ȳ = sn .

4. Compute degrees of freedom as n − 1.

5. Look up t 1−x . Take the t-distribution with the right number of degrees
2
of freedom and look for the critical t-value for the confidence interval: if
x is the confidence level you want, then look for quantile 1−x
2 . Then take
its absolute value. That’s your t 1−x .
2

6. Compute margin of error (MoE) as MoE = t 1−x × σ̂ȳ .

2

7. Subtract and sum the sample mean with the margin of error: (ȳ−MoE, ȳ+
MoE).

Note that for a large number of degrees of freedom, the values are very close
to those of the standard normal.

Table 2.2: Quantiles for the standard normal and several t-distributions.
probs norm t199 t99 t9 t5 t3
0.0005 -3.29 -3.34 -3.39 -4.78 -6.87 -12.92
0.0010 -3.09 -3.13 -3.17 -4.30 -5.89 -10.21
0.0050 -2.58 -2.60 -2.63 -3.25 -4.03 -5.84
0.0100 -2.33 -2.35 -2.36 -2.82 -3.36 -4.54
0.0250 -1.96 -1.97 -1.98 -2.26 -2.57 -3.18
0.0500 -1.64 -1.65 -1.66 -1.83 -2.02 -2.35
0.1000 -1.28 -1.29 -1.29 -1.38 -1.48 -1.64
0.9000 1.28 1.29 1.29 1.38 1.48 1.64
0.9500 1.64 1.65 1.66 1.83 2.02 2.35
0.9750 1.96 1.97 1.98 2.26 2.57 3.18
0.9900 2.33 2.35 2.36 2.82 3.36 4.54
0.9950 2.58 2.60 2.63 3.25 4.03 5.84
0.9990 3.09 3.13 3.17 4.30 5.89 10.21
0.9995 3.29 3.34 3.39 4.78 6.87 12.92

75
2.10 Obtaining a confidence interval for a population
mean in R
Suppose we have values on miles per gallon (mpg) in a sample of cars, and we
wish to construct a 99% confidence interval for the population mean. We can
do that in the following manner. We take all the mpg values from the mtcars
data set, and set our confidence level to 0.99 in the following manner:

t.test(mtcars$mpg, conf.level = 0.99)$conf.int

## [1] 17.16706 23.01419

## attr(,"conf.level")
## [1] 0.99

It shows that the 99% confidence interval runs from 17.2 to 23.0. The
t.test() function does more than simply constructing confidence intervals.
That is the topic of the next section.

2.11 Null-hypothesis testing

Suppose a professor of biology claims, based on years of measuring the height
of elephants in Tanzania, that the mean height of elephants in Tanzania is
3.38 m. Suppose that you come up with data on a relatively small number of
South-African elephants and the professor would like to know whether the two
groups of elephants have the same population mean. Do both the Tanzanian
and South-African populations have the same mean of 3.38, or is there perhaps
a difference in the means? A difference in means could indicate that there are
genetic differences between the two elephant populations. The professor would
like to base her conclusion on your sample of data, and you assume that the
professor is right in that the population mean of Tanzanian elephants is 3.38 m.
One way of addressing a question like this is to look at the confidence interval
for the South-African mean. Suppose you construct a 95% confidence interval.
Based on a sample mean of 3.27, a sample variance s2 of 0.14 and a sample
size of 40, you calculate that the interval runs from 3.15 to 3.39. Based on that
interval, you can conclude that 3.38 is a reasonable value for the population
mean, and that it could well be that the both the Tanzanian and South-African
populations have the same mean height of 3.38 m.
However, as we have seen in the previous section, there are many confidence
intervals that we could compute. If instead of the 95% confidence interval, we
would compute a 90% confidence interval, we would end up with an interval
that runs from 3.17 to 3.37. In that case, the Tanzanian population mean is
no longer included in the confidence interval for the South-African population
mean, and we’d have to conclude that the populations have different means.
What interval to choose? Especially if you have questions like ”Do the
two populations have the same mean” and you want to have a clear yes or no

76
answer, then null-hypothesis testing might be a solution. With null-hypothesis
testing, a null-hypothesis is stated, after which you decide based on sample data
whether or not the evidence is strong enough to reject that null-hypothesis. In
our example, the null-hypothesis is that the South-African mean has the value
3.38 (the Tanzanian mean). We write that as follows:

H0 : µSA = 3.38 (2.9)

We then look at the data on South-African elephants that could give us
evidence that is either in line with this hypothesis or not. If it is not, we say
that we reject the null-hypothesis.
The objective of null-hypothesis testing is that we either reject the null-
hypothesis, or not. This is done using the data from a sample. In the null-
hypothesis procedure, we simply assume that the null-hypothesis is true, and
compare the sample data with data that would result if the null-hypothesis were
true.
So, let’s assume the null-hypothesis is true. In our case that means that
the mean height of all South-African elephants is equal to that of all Tanzanian
elephants, namely 3.38 m. Next, we compare our actual observed data with
data that would theoretically result from a population mean of 3.38. What
would sample data theoretically look like if the population mean is 3.38? In
the previous sections, we learned what possible sample means would look like.
Thus, let’s focus on the sample mean.2
Based on what we learned about the sampling distribution of the sample
mean, we know that possible values for the sample mean come from this distribution.
It is more or less a normal distribution with mean 3.38, but what the variance
is (the standard error), we don’t know. We’d have to take a guess, based on
the sample data that we have. Based on the sample data, we could compute q
s 2
the sample variance s2 , and then estimate the standard error as σ̂ȳ = n.
However, as we saw earlier, because we have to estimate the standard error, the
sample means are no longer normally distributed, but t-distributed.
Suppose we observe 40 South-African elephants, and we obtain a sample
mean of 3.27 and a sample variance s2 of 0.14. The hypothesised population
mean is 3.38. We know that the sampling distribution is a t-distribution because
we do not know the population variance. To know the shape of the sampling
distribution, we need three things: the mean of the sampling distribution (assuming
the population mean is 3.38), the standard deviation (or variance) of the distribution,
and the exact shape of the t-distribution (the degrees of freedom). The mean
is easy: that is equal to the hypothesised population mean of 3.38 (why?). The
standard deviation (standard error) is more difficult, but we can use the qsample
s2
data to estimate it. We compute it using the sample variance: σ̂ȳ = n =
2 The sample mean is called a sufficient statistic for the population mean. That means, if

you want to know something about the population mean, the only information you need to
get from the sample data is the mean of the sample values. Knowing the exact values does
not give you extra information: the sample mean suffices. The proof for this is beyond this
book.

77
0.059. And the last bit is easy: the degrees of freedom is simply sample size
minus 1: 40 − 1 = 39.
We plot this sampling distribution of the sample mean in Figure 2.12. This
figure tells us that if the null-hypothesis is really true and that the South-African
mean height is 3.38, and we would take many different random samples of 40
elephants, we would see only sample means between 3.20 and 3.35. Other values
are in fact possible, but very unlikely. But how likely is our observed sample
mean of 3.27: do we feel that it is a likely value to find if the population mean
is 3.38, or is it rather unlikely?
What do you think? Think this over for a bit before you continue to read.
In fact, every unique value for a sample mean is rather unlikely. If the
population mean is 3.38, it will be very improbable that you will find a sample
mean of exactly 3.38, because by sheer chance it could also be 3.39, or 3.40 or
3.37. But relatively speaking, those values are all more likely to find than more
deviant values. The density curve tells you that values around 3.38 are more
likely than values around 3.27 or 3.50, because the density is higher around the
value of 3.38 than around those other values.
What to do?
The solution is to define regions for sample means where we think the sample
mean is no longer probable under the null-hypothesis, and a region where it is
probable enough to believe that the null-hypothesis could be true.
For example, we could define an acceptance region where 95% of the sample
means would fall if the null-hypothesis is true, and a rejection region where only
5% of the sample means would fall if the null-hypothesis is true. Let’s put the
rejection region in the tails of the distribution, where the most extreme values
can be found (farthest away from the mean). We put half of the rejection region
in the left tail and half of it in the right tail of the distribution, so that we have
two regions that each covers 2.5% of the sampling distribution. These regions
are displayed in Figure 2.13. The red ones are the rejection regions, and the
green one is the acceptance region (covering 95% of the area).
Why 5%, why not 10% or 1%? Good question. It is just something that is
accepted in a certain group of scientists. In the social and behavioural sciences,
researchers feel that 5% is a small enough chance. In contrast, in quantum
mechanics, researchers feel that 0.000057% is a small enough chance. Both
values are completely arbitrary. We’ll dive deeper into this arbitrary chance
level in a later section. For now, we continue to use 5%.
From Figure 2.13 we see that the sample mean that we found for your 40
South-African elephants (3.27) does not lie in the red rejection region. We
see that 3.27 lies well within the green section where we decide that sample
means are likely to occur when the population is 3.38. Because this is likely, we
think that the null-hypothesis is plausible: if the population mean is 3.28, it is
plausible to expect a sample mean of 3.27, because in 95% of random samples
we would see a sample mean between 3.255 and 3.500. The value 3.27 is a
very reasonable value and we therefore do not reject the null-hypothesis. We
conclude therefore that it could well be that both Tanzanian and South-African
elephants have the same average height of 3.38, that is, we do not have any

78
 0.4

0.3
density

0.2

0.1

0.0 3.27 3.38

3.175 3.200 3.225 3.250 3.275 3.300 3.325 3.350 3.375 3.400 3.425 3.450 3.475 3.500 3.525 3.550 3.575
sample mean

Figure 2.12: The sampling distribution under the null-hypothesis that the
South-African population mean is 3.38. The blue line represents the sample
mean for our observed sample mean of 3.27.

evidence that the population mean is not 3.38.

This is the core of null-hypothesis testing for a population mean: 1) you
determine a null-hypothesis that states that the population mean has a certain
value, 2) you figure out what kind of sample means you would get if the
population mean would have that value, 3) you check whether the sample mean
that you actually have is far enough from the population mean to say that it
is unlikely enough to result from the hypothesised population mean. If that is
the case, then you reject the null-hypothesis, meaning you don’t believe in it.
If it is likely to result from the hypothesised population, you do not reject the
null-hypothesis: there is no reason to suspect that the null-hypothesis is false.

2.12 Null-hypothesis testing with t-values

In the above example, we looked explicitly at the sampling distribution for a
hypothesised value for the population mean. By determining what the distribution
would look like (determining the mean, standard error and degrees of freedom),
we could see whether a certain sample mean would give enough evidence to
reject the null-hypothesis.
In this section we will show how to do this hypothesis testing more easily
by first standardising the problem. The trick is that we do not have to make a
picture of the sampling distribution every time we want to do a null-hypothesis
test. We simply know that its shape is that of a t-distribution with degrees of
freedom equal to n − 1. t-distributions are standardised distributions, always
with a mean of 0. They are the distribution of standardised t-statistics, where a
sample mean is standardised by subtracting the population mean and dividing
the result by the standard error.
Let’s do this standardisation for our observed sample mean of 3.27. With a

79
 0.4

0.3
density

0.2

0.1

2.5 % 2.5 %
0.0
3.27 3.38
3.175 3.200 3.225 3.250 3.275 3.300 3.325 3.350 3.375 3.400 3.425 3.450 3.475 3.500 3.525 3.550 3.575
sample mean

Figure 2.13: The sampling distribution under the null-hypothesis that the
South-African population mean is 3.38. The red area represents the range of
values for which the null-hypothesis is rejected (rejection region), the green
area represents the range of values for which the null-hypothesis is not rejected
(acceptance region).

q q
s2 0.14
population mean of 3.38 and a standard error of σ̂ȳ = n = 40 = 0.059,
we obtain:

3.27 − 3.38
t= = −1.864 (2.10)
0.059
We can then look at a t-distribution of 40 − 1 = 39 degrees of freedom to see
how likely it is that we find such a t-score if the null-hypothesis is true. The t-
distribution with 39 degrees of freedom is depicted in Figure 2.14. Again we see
the population mean represented, now standardised to a t-score of 0 (why?), and
the observed sample mean, now standardised to a t-score of -1.864. As you can
see, this graph gives you the same information as the sampling distribution in
Figure 2.13. The advantage of using standardisation and using the t-distribution
is that we can now easily determine whether or not an observed sample mean
is somewhere in the red zone or in the green zone, without making a picture.
We have to find the point in the t-distribution where the red and green zones
meet. These points in the graph are called critical values. From Figure 2.14 we
can see that these critical values are around -2 and 2. But where exactly? This
information can be looked up in the t-tables that were discussed earlier in this
chapter. We plot such a table again in Table 2.3. A larger version is given in
Appendix B.
In such a table, you can look up the 2.5th percentile. That is, the value for

80
which 2.5% of the t-distribution is equal or smaller. Because we are dealing with
a t-distribution with 39 degrees of freedom, we look in the column t39, and then
in the row with cumulative probability 0.025 (equal to 2.5%), we see a value of
-2.02. This is the critical value for the lower tail of the t-distribution. To find
the critical value for the upper tail of the distribution, we have to know how
much of the distribution is lower than the critical value. We know that 2.5% is
higher, so it must be the case that the rest of the distribution, 100−2.5 = 97.5%
is lower than that value. This is the same as a probability of 0.975. If we look
for the critical value in the table, we see that it is 2.02. Of course this is the
opposite of the other critical value, because the t-distribution is symmetrical.
Now that we know that the critical values are -2.02 and +2.02, we know
that for our standardised t-score of -1.864 we are still in the green area, so we
do not reject the null-hypothesis. We don’t need to draw the distribution any
more. For any value, we can directly compare it to the critical values. And not
only for this example of 40 elephants and a sample mean of 3.27, but for any
combination.
Suppose for example that we would have had a sample size of 10 elephants,
and we would have found a sample mean of 3.28 with a slightly different sample
variance, s2 = 0.15. If we want to test the null-hypothesis again that the
population mean is 3.38 based on these results, we would have to do the following
steps:

Null-hypothesis testing
q
s2
1. Estimate the standard error σ̂ȳ = n.

ȳ−µ
2. Calculate the t-statistic t = σ̂ȳ , µ is the population mean under the
null-hypothesis.
3. Determine the degrees of freedom, n − 1.

4. Determine the critical values for lower and upper tail of the appropriate
t-distribution, using Appendix B.
5. If the t-statistic is between the two critical values, then we’re in the green,
we still believe the null-hypothesis is plausible.

6. If the t-statistic is not between the two critical values, we are in the red
zone and we reject the null-hypothesis.

So let’s do this for our hypothetical result:

q
0.15
1. Estimate the standard error: 10 = 0.1224745

3.28−3.38
2. Calculate the t-statistic: t = 0.1224745 = −0.8164966
3. Determine the degrees of freedom: sample size minus 1 equals 9

81
 0.4

0.3
density

0.2

0.1

0.0 2.5 % 2.5 %

−1.864
−2 0 2
t

Figure 2.14: A t-distribution with 39 degrees of freedom to test the null-

hypothesis that the South-African population mean is 3.38. The blue line
represents the T -score for our observed sample mean of 3.27.

4. In Table 2.3 we look for the row with probability 0.025 and the column
for t9. We see a value of -2.26. The other critical value then must be 2.26.

5. The t-statistic of -0.8164966 lies between these two critical values, so these
sample data do not lead to a rejection of the null-hypothesis that the
population mean is 3.38. In other words, these data from 10 elephants do
not give us reason to doubt that the population mean is 3.38.

Table 2.3: Quantiles for the standard normal and several t-distributions.
probs norm t199 t99 t47 t39 t9 t5 t3
0.0005 -3.29 -3.34 -3.39 -3.51 -3.56 -4.78 -6.87 -12.92
0.0010 -3.09 -3.13 -3.17 -3.27 -3.31 -4.30 -5.89 -10.21
0.0050 -2.58 -2.60 -2.63 -2.68 -2.71 -3.25 -4.03 -5.84
0.0100 -2.33 -2.35 -2.36 -2.41 -2.43 -2.82 -3.36 -4.54
0.0250 -1.96 -1.97 -1.98 -2.01 -2.02 -2.26 -2.57 -3.18
0.0500 -1.64 -1.65 -1.66 -1.68 -1.68 -1.83 -2.02 -2.35
0.1000 -1.28 -1.29 -1.29 -1.30 -1.30 -1.38 -1.48 -1.64
0.9000 1.28 1.29 1.29 1.30 1.30 1.38 1.48 1.64
0.9500 1.64 1.65 1.66 1.68 1.68 1.83 2.02 2.35
0.9750 1.96 1.97 1.98 2.01 2.02 2.26 2.57 3.18
0.9900 2.33 2.35 2.36 2.41 2.43 2.82 3.36 4.54
0.9950 2.58 2.60 2.63 2.68 2.71 3.25 4.03 5.84
0.9990 3.09 3.13 3.17 3.27 3.31 4.30 5.89 10.21
0.9995 3.29 3.34 3.39 3.51 3.56 4.78 6.87 12.92

82
 0.4

0.3
density

0.2

0.1

0.0
0.217 0.566 0.217
t = −0.82
−2 0 2
t

Figure 2.15: Illustration of what a p-value is. The total blue area represents the
probability that under the null-hypothesis, you find a more extreme value than
the t-score or its opposite. The blue area covers a proportion of .217 + .217 =
0.434 of the t-distribution. This amounts to a p-value of .434.

2.13 The p-value

What we saw in the previous section was the classical null-hypothesis testing
procedure: calculating a t-statistic and determine whether or not this t-score is
in the red zone or green zone, by comparing them to critical values. In the old
days, this was done by hand: the calculation of t and looking up the critical
values in tables published in books.
These days we have the computer do the work for us. If you have a data
set, a program can calculate the t-score for you. However, when you look at the
output, you actually never see whether this t-score leads to a rejection of the
null-hypothesis or not. The only thing that a computer prints out is the t-score,
the degrees of freedom, and a so-called p-value. In this section we explain what
a p-value is and how you can use it for null-hypothesis testing.
Let’s go back to our example in the previous section, where we found a
sample mean height of 3.28 with only 10 elephants. We computed the t-score
and obtained -0.82. We illustrate this result in Figure 2.15 where the red line
indicates the t-score. By comparing this t-value with the critical values, we
could decide that we do not reject the null-hypothesis. However, if you would
do this calculation with a computer program like R, we would get the following
result:
t = -0.82 , df = 9 , p - value = 0.434
Figure 2.15 shows what this p-value of 0.434 means. The green area in the
middle represents the probability that a t-score lies between -0.82 and +0.82.
That probability is shown in the figure as 0.567, so 56.7%. The left blue region
represents the probability that if the null-hypothesis is true, the t-score will be
less than -0.82. That probability is 0.217, so 21.7%. Because of symmetry, the

83
probability that the t-score is more than 0.82 is also 0.217. The blue regions
together therefore represent the probability that you find a t-score of less than
-0.82 or more than 0.82, and that probability equals 0.217 + 0.217 = 0.434.
Therefore, the probability that you find a t-value of ±0.82 or more extreme
equals 0.434. This probability is called the p-value.
Why is this value useful?
Let’s imagine that we find a t-score of exactly equal to one of the critical
values. The critical value for a sample size of 10 animals related to a cumulative
proportion of 0.025 equals -2.26 (see Table 2.3). Based on this table, we know
that the probability of a t-value of -2.26 or lower equals 0.025. Because of
symmetry, we also know that the probability of a t-value of -2.26 or higher also
equals 0.025. This brings us to the conclusion that the probability of a t-score
of ±2.26 or more extreme, is equal to 0.025 + 0.025 = 0.05 = 5%. Thus, when
the t-score is equal to the critical value, then the p-value is equal to 5%. You
can imagine that if the t-score becomes more extreme than the critical value
the p-value will become less than 5%, and if the t-score becomes less extreme
(closer to 0), the p-value becomes larger.
In the previous section, we said that if a t-score is more extreme than one of
the critical values (when it doesn’t have a value between them) then we reject
the null-hypothesis. Thus, a p-value of 5% or less means that we have a t-score
more extreme than the critical values, which in turn means we have to reject the
null-hypothesis. Thus, based on the computer output, we see that the p-value
is larger than 0.05, so we do not reject the null-hypothesis.

Overview
ˆ critical value: the minimum (or maximum) value that a t-score should
have to be in the red zone (the rejection region). If a t-value is more
extreme than a critical value, then the null-hypothesis is rejected. The
red zone is often chosen such that a t-score will be in that zone 5% of the
time, assuming that the null-hypothesis is true.
ˆ p-value: indicates the probability of finding a t-value equal or more
extreme than the one found, assuming that the null-hypothesis is true.
Often a p-value of 5% or smaller is used to support the conclusion that
the null-hypothesis is not tenable. This is equivalent to a rejection region
of 5% when using critical values.

Let’s apply this null-hypothesis testing to our luteinising hormone (LH) data.
Based on the medical literature, we know that LH levels for women in their
child-bearing years vary between 0.61 and 56.6 IU/L. Values vary during the
menstrual period. If values are lower than normal, this can be an indication that
the woman suffers from malnutrition, anorexia, stress or a pituitary disorder.
If the values are higher, this is an indication that the woman has gone through
menopause.

84
 We’re going to use the LH data presented earlier in this chapter to make a
decision whether the woman has a healthy range of values for a woman in her
child-bearing years by testing the null-hypothesis that the mean LH level in this
woman is the same as the mean of LH levels in healthy non-menopausal women.
First we specify the null-hypothesis. Suppose we know that the mean LH
level in this woman should be equal to 2.54, given her age and given the timing
of her menstrual cycle. Thus our null-hypothesis is that the mean LH in our
particular woman is equal to 2.54:

H0 : µ = 2.54 (2.11)
Next, we look at our sample mean and see whether this is a likely or unlikely
value to find under this null-hypothesis. The sample mean is 2.40. To know
whether this is a likely value to find, we have to know the standard error
of the sampling distribution, and we can estimate this by using the sample
variance. The sample variance s2 = 0.3042553
q q and we had 48 measures, so we
s2 0.3042553
estimate the standard error as σˆȳ = n = 48 = 0.08. We then apply
standardisation to get a t-value:
2.40 − 2.54
t= = −1.75 (2.12)
0.08
Next, we look up in a table whether this t-value is extreme enough to be
considered unlikely under the null-hypothesis. In Table 2.3, we see that for 47
degrees of freedom, the critical value for the 0.025 quantile equals -2.01. For the
0.975 quantile it is 2.01. Our observed t-value of -1.75 lies within this range.
This means that a sample mean of 2.40 is likely to be found when the population
mean is 2.54, so we do not reject the null-hypothesis. We conclude that the LH
levels are healthy for a woman her age.
We can do the null-hypothesis testing also with a computer. Let’s analyse
the data in R and do the computations with the following code. First we load
the LH data:

data(lh)

Next, we test whether the population mean could be 2.54:

t.test(lh, mu = 2.54)

##
## One Sample t-test
##
## data: lh
## t = -1.7584, df = 47, p-value = 0.08518
## alternative hypothesis: true mean is not equal to 2.54
## 95 percent confidence interval:
## 2.239834 2.560166

85
## sample estimates:
## mean of x
## 2.4

In the output we see that the t-value is equal to -1.7584, similar to our -1.75.
We see that the number of degrees of freedom is 47 (n − 1) and that the p-value
equals 0.08518. This p-value is larger than 0.05, so we do not reject the null-
hypothesis that the mean LH level in this woman equals 2.54. Her LH level is
healthy.

2.14 One-sided versus two-sided testing

In the previous section, we tested a null-hypothesis in order to find evidence
that an observed sample mean was either too large or too small to result from
random sampling. For example, in the previous section we saw that the observed
LH levels were not too low and we did not reject the null-hypothesis. But had
the LH levels been too high or too low, then we would have rejected the null-
hypothesis.
In the reasoning that we followed, there were actually two hypotheses: the
null-hypothesis that the population mean was exactly 2.54, and the alternative
hypothesis that the population is not exactly 2.54:

H0 : µ = 2.54 (2.13)
HA : µ 6= 2.54 (2.14)
This kind of null-hypothesis testing is called two-sided or two-tailed testing:
we look at two critical values, and if the computed t-score is outside this
range (i.e., somewhere in the two tails of the distribution), we reject the null-
hypothesis.
The alternative to two-sided testing is one-sided or one-tailed testing. Sometimes
before an analysis you already have an idea of what direction the data will
go. For instance, imagine a zoo where they have held elephants for years.
These elephants always were of Tanzanian origin, with a mean height of 3.38.
Lately however, the manager observes that the opening that connects the indoor
housing with the outdoor housing gets increasingly damaged. Since the zoo
recently acquired 4 new elephants of South-African origin, the manager wonders
whether South-African elephants are on average taller than the Tanzanian elephants.
To figure out whether South-African elephants are on average taller than the
Tanzanian average of 3.38 or not, the manager decides to apply null-hypothesis
testing. She has two hypotheses: null-hypothesis H0 and alternative hypothesis
HA :

H0 : µSA = 3.38 (2.15)

HA : µSA > 3.38 (2.16)

86
 This set of hypotheses leaves out one option: the South-African mean might
be lower than the Tanzanian one. Therefore, one often writes the set of hypotheses
like this:

H0 : µSA ≤ 3.38 (2.17)

HA : µSA > 3.38 (2.18)

She next tests the null-hypothesis, more specifically the one where µSA =
3.38. From the damaged doorway she expects the sample mean to be higher
than 3.38, but is it high enough to serve as evidence that the population mean
is also higher than 3.38? She decides that when the sample mean is in the
rejection zone in the right tail of the sampling distribution, then she will decide
that the null-hypothesis is not true, but that the alternative hypothesis must
be true. This is illustrated in Figure 2.16.
It shows the sampling distribution if we happen to have 4 new South-African
elephants, with a sample mean of 3.45 and a standard error of 0.059. In red,
we see the rejection region: if the sample mean happens to be in that zone we
decide to reject the null-hypothesis. Similar to two-tailed testing, we decide
that an area of 5% is small enough to suggest that the null-hypothesis is not
true. Note that in two-tailed testing, this area of 5% was divided equally into
the upper tail and the lower tail of the distribution, but with one-tailed testing
we put it all in the tail where we expect to find the sample mean based on a
theory or a hunch.
In this sampling distribution, based on 3 degrees of freedom, we see that the
sample mean is not in the red zone – the rejection region – therefore we do not
reject the null-hypothesis. We conclude that based on this random sample of 4
elephants, there is no evidence to suggest that South-African elephants are on
average taller than Tanzanian elephants.
The same procedure can be done with standardisation. We compute the
t-statistic as

3.45 − 3.38
t= = 1.19 (2.19)
0.059
In Table 2.3 we have to look up where the red zone starts: that is for the 0.95
quantile, because below that value lies 95% (green zone) and above it 5% (the
red zone). We see that the 95th percentile for a t-distribution with 3 degrees
of freedom is equal to 2. Our t-value 1.19 is less than that, so that we do not
reject the null-hypothesis.
A third way is to compute a one-tailed p-value. This is illustrated in Figure
2.17. The one-tailed p-value for a t-statistic of 1.19 and 3 degrees of freedom
turns out to be 0.16. That is the proportion of the t-distribution that is blue.
That means that if the null-hypothesis is true, you will find a t-value of 1.19 or
larger in 16% of the cases. Because this proportion is more than 5%, we do not
reject the null-hypothesis.

87
 0.3
density

0.2

0.1

0.0 5%
3.38 3.45
3.175 3.200 3.225 3.250 3.275 3.300 3.325 3.350 3.375 3.400 3.425 3.450 3.475 3.500 3.525 3.550 3.575
sample mean

Figure 2.16: The sampling distribution under the null-hypothesis that the
South-African population mean is 3.38. In one-tailed testing, the rejection
area is located in only one of the tails. The red area represents the range
of values for which the null-hypothesis is rejected (rejection region), the green
area represents the range of values for which the null-hypothesis is not rejected
(acceptance region).

0.3
density

0.2

0.1

0.84 0.16
0.0
t = 1.19
−2.5 0.0 2.5
t

Figure 2.17: The sampling distribution under the null-hypothesis that the
South-African population mean is 3.38. For one-tailed testing, the rejection
area is located in only one of the tails. The green area represents the probability
of seeing a t-value smaller than 1.19, the blue are represents the probability of
seeing a t-value larger than 1.19. The latter probability is the p-value.

88
 0.3
density

0.2

0.1

0.05 0.95
0.0
Critical t = −1.68
−2.5 0.0 2.5
t

Figure 2.18: One-tailed decision process for deciding whether the average LH
level in a woman is too low.

2.15 One-tailed testing applied to LH levels

As we have seen, LH levels that are too high are indicative of menopause, a
normal transition for women. However, LH levels that are too low are indicative
of an illness or malnutrition. In that case, it is important that the source of this
malnutrition or the specific illness is diagnosed. You could therefore say that if
LH levels are too low, a red flag should be put up, whereas if the LH levels are
normal or higher, then there is usually no reason to worry.
LH levels can therefore be used to construct a diagnostic red flag decision
system. If normal or high, then nothing happens, if too low, then something
should be done. We could formulate these two alternative states of reality as
two hypotheses:

H0 : µLH ≥ 2.54 (2.20)

HA : µLH < 2.54 (2.21)

We decide beforehand that if a t-value is too far out in the left tail of the
distribution, the LH levels are too low. We again use 5% of the area of the
t-distribution. This decision process is illustrated in Figure 2.18 where we see a
critical t-value of -1.68 when we we have 47 degrees of freedom (see Table 2.3).
We calculate our t-value and find -1.75, see section 2.13. We see that this
t-value is smaller than the critical value -1.68, so it is in the red rejection area.
This is the area that we use for the rejection of the null-hypothesis, so based on
these data we decide that the mean LH level in this woman is abnormally low.
Importantly, note that when we applied two-tailed hypothesis testing, we
decided to not reject the null-hypothesis, whereas here with one-tailed testing,
we decide to reject the null-hypothesis. All based on the same data, and the same
null-hypothesis. The difference lies in the choice of the alternative hypothesis.

89
When doing one-tailed testing, we put all of the critical region in only one tail
of the t-distribution. This way, it becomes easier to reject a null-hypothesis, if
the mean LH level is indeed lower than normal. However, it could also be easier
to make a mistake: if the mean LH level is in fact normal, we could make a
mistake in thinking that the sample mean is deviant, where it is actually not.
Making mistakes in inference is the topic of the next section.
It is generally advised to use two-tailed testing rather than one-tailed testing.
The reason is that in hypothesis testing, it is always the null-hypothesis that
is being used as the starting point: what would the sample means (or their
standardised versions: t-scores) look like if the null-hypothesis is true? Based
on a certain null-hypothesis, say population mean µ equals 2.54, sample means
could be as likely higher or lower than the population mean (since the sampling
distribution is symmetrical). Even if you suspect that µ is actually lower, based
on a very good theory, you would help yourself too much to falsify the null-
hypothesis by putting the rejection area only in the left tail of the distribution.
And what do you actually do if you find a sample mean that is in the far end
of the right tail? Do you still accept the null-hypothesis? That would not make
much sense. It is therefore better to just stick to the null-hypothesis, and see
whether the sample mean is far enough removed to reject the null-hypothesis.
If the sample mean is in the anticipated tail of the distribution, that supports
the theory you had, and if the sample mean is in the opposite tail, it does not
support the theory you had.
Compare one-tailed and two-tailed testing in R using the LH data. By
default, R applies two-tailed testing. R gives the following output:

t.test(lh, mu = 2.54)

##
## One Sample t-test
##
## data: lh
## t = -1.7584, df = 47, p-value = 0.08518
## alternative hypothesis: true mean is not equal to 2.54
## 95 percent confidence interval:
## 2.239834 2.560166
## sample estimates:
## mean of x
## 2.4

If you want one-tailed testing, where you expect that the mean LH level is
lower than 2.54, you do that in the following manner3 :

t.test(lh, mu = 2.54, alternative = "less")

##
3 If you expect that the LH level will higher than 2.54, you use ”greater” instead of ”less”.

90
## One Sample t-test
##
## data: lh
## t = -1.7584, df = 47, p-value = 0.04259
## alternative hypothesis: true mean is less than 2.54
## 95 percent confidence interval:
## -Inf 2.533589
## sample estimates:
## mean of x
## 2.4

When you compare the p-values, you see that the p-value using one-tailed
testing is half the size of the p-value using two-tailed testing (0.04 vs 0.08).
Based on the previous sections, you should know why the p-value is halved!
In the second output, using a critical p-value of 5% you would reject the null-
hypothesis, whereas in the first output, you would not reject the null-hypothesis.
Using one-tailed testing could lead to a big mistake: thinking that the sample
mean is deviant enough to reject the null-hypothesis, while the null-hypothesis
is actually true. We delve deeper into such mistakes in the next section.

2.16 Type I and type II errors

In the preceding sections, we have used the value of 5% a lot of times. We
deemed that this was a fairly low probability, that allows us to take the decision
to reject the null-hypothesis. We looked at the distribution of sample means,
given that there was a certain population mean, and we looked at how often we
can expect a sample mean that is smaller or larger than certain critical values.
These critical values were based on 5% of the area of the sampling distribution.
With two-tailed testing, this 5% was divided over the two extreme tails of the
sampling distribution, and with one-tailed testing, this 5% rejection area was
put in the tail end where we expected the population to be according to the
alternative hypothesis (based on theory).
In this null-hypothesis testing procedure there is always the risk that we
take the wrong decision. Let’s return to our elephant example where we had
the null-hypothesis that the population mean for South-African elephants equals
3.38. The alternative two-sided hypothesis was that the population mean was
not equal to 3.38. After calculating the standard error, we calculated the t-
score. We said that we reject the null-hypothesis when the obtained t-score was
somewhere in the extreme ends of the tail: more specifically, in the rejection
area that made up 5% of the area of the t-distribution. That means that if the
null-hypothesis is true, there is a 5% probability that we find such a t-score. In
that case we reject the null-hypothesis. But that could be the wrong decision:
if the null-hypothesis is true it will happen in 5% of the cases that a t-score
will be in the 5% rejection region. We then reject the null-hypothesis while it
is actually true! Such a mistake is called a Type I error. In this case, type I

91
error rate is 5%. It is a conditional probability. Conditional probabilities are
probabilities that start from some given information. In this case, the given
information is that the null-hypothesis is true: given that the null-hypothesis is
true, it is the probability that we reject the null-hypothesis. Because we do not
like to make mistakes, we want to have the probability of a mistake as low as
possible.
In the social and behavioural sciences, one thinks that a probability of 5%
is low enough to take the risk of making the wrong decision. As stated earlier,
in quantum mechanics one is even more careful, using a probability of 0.000057
%. So why don’t we also use a much lower probability of making a type I error?
The answer is that we do not want to make another type of mistake: a type II
error. A type II error is the mistake that we make when we do not reject the
null-hypothesis, while it is not true. Taking the example of the elephants again,
suppose that the population mean is not equal to 3.38, but the t-score is not in
the rejection area, so we believe that the population mean is 3.38. This is then
the wrong decision. The type of mistake we then make is a type II error.
Let’s take this example further. Suppose we have a two-tailed decision
process, where we compare two hypotheses about South-African elephants:
either their mean height is equal to 3.38 (H0 ), or it is not (HA ). We compute
the t-statistic and determine the critical values based on 5% area in the tails of
the t-distribution. This means that we allow ourselves to make a mistake in 5%
of the cases: the probability that we find a t-score in one of the 2.5% tails equals
2.5% + 2.5% = 5%. This is the probability of a type I error. Note that we chose
this value deliberately. This 5% we call α (’alpha’): it is the relative frequency
we allow ourselves to make a type I error. We say then that our α is fixed to
0.05, or 5%. This means that if the null-hypothesis is true, the probability that
the t-statistic will be in in the tails will be 5%.
Then what is the probability of a type II error? A type II error is based on
the premise that the alternative hypothesis is true. That alternative hypothesis
states that the population mean is not equal to 3.38. Given that, what is the
probability that we do not reject the null-hypothesis?
This is impossible to compute, because the alternative hypothesis is very
vaguely stated: it could be anything, as long as it is not 3.38. Let’s make it a
bit easier and state that the alternative hypothesis states that the population
mean equals 3.42.

H0 : µSA = 3.38 (2.22)

HA : µSA = 3.42 (2.23)

If the population mean height is equal to 3.42, what would sample means
look like? That’s easy, that is the sampling distribution of the sample mean.
The mean of that sampling distribution would be 3.42. This is illustrated in
Figure 2.19. The left curve is the sampling distribution for a population mean
of 3.38. The red area represents the probability of a type I error. The right
curve is the sampling distribution for a population mean of 3.42. The blue area

92
 0.4

0.3
density

0.2

0.1

2.5 % 2.5 %
0.0

3.18 3.23 3.28 3.33 3.38 3.43 3.48 3.53 3.58

sample mean

Figure 2.19: Two sampling distributions, one for a population mean of 3.38
(null-hypothesis) and one for a population mean of 3.42 (alternative hypothesis).
The red areas represent the probability of a type I error, the dark green area
the probability of a type II error. The blue areas represent the probability of
making the (correct) decision that the null-hypothesis is not true when it is
indeed not true.

93
 0.4

0.3
density

0.2

0.1

0.5 % 0.5 %
0.0

3.18 3.23 3.28 3.33 3.38 3.43 3.48 3.53 3.58

sample mean

Figure 2.20: Two sampling distributions, one for a population mean of 3.38
(null-hypothesis) and one for a population mean of 3.42 (alternative hypothesis).
The red areas represent the probability of a type I error, the dark green area
the probability of a type II error. The blue area represents the probability of
making the (correct) decision that the null-hypothesis is not true when it is
indeed not true.

represents the probability of rejecting the null-hypothesis. This is because if the

sample mean is smaller than 3.260336 or larger than 3.499664, the sample mean
is in the rejection area of the null-hypothesis testing and the null-hypothesis
will therefore be rejected. The probability of this happening given that the
alternative hypothesis is true (HA = 3.42), is represented by the area under
that curve: the blue area. If we determine the two blue areas in Figure 2.19,
we end up with 0.004 + 0.097 = 0.101. This is the probability of rejecting the
null-hypothesis while it is not true, so this is no mistake at all. We would make
a mistake when the alternative hypothesis is true, and we would not reject the
null-hypothesis. This is represented by the dark green area. That area is equal
to 1 minus the blue area: 1 − 0.101 = 0.899.
When we have to make a definite decision about a population mean, the
null-hypothesis framework can be used for that. Usually we don’t want to make
type I error mistakes, so we pick a low probability like 5% for the tails of the
sampling distribution under the H0 . This value is called α: if the null-hypothesis
is true, we don’t want to reject it, so we allow this to happen in only 5% of the
cases. One chooses α before collecting the data. You have to be careful with this
choice of α though because it directly affects the probability of making a type
II error. This probability is denoted by β (’beta’): how often does it happen
that if the alternative hypothesis is true, we do not reject the null-hypothesis.
This relationship is illustrated in Figure 2.20. There, an α of 1% is chosen, using

94
both tails (a two-tailed null-hypothesis test). You immediately see that the blue
areas have also become smaller, and that by consequence the dark green area
becomes larger: the probability of a type II error.
Thus, the α should be chosen wisely: if it is too large, you run a high risk of
a type I error. But if it is too low, you run a high risk of a type II error. Let’s
think about this in the context of our luteinising hormone problem.
We saw that if the LH level is not normal, this is an indication of malnutrition
or a disease and the patient should have further checks to see what the problem
is. But if the LH level is normal or above, there is no disease and no further
checks are required. Again we take the null-hypothesis that the mean LH level
in this woman equals 2.54. What would be a type I error this case, and what
would be type II error?
The type I error is the mistake of rejecting the null-hypothesis while it is in
fact true. Thus, the woman’s mean LH level is 2.54, but by coincidence, the
mean of the 48 measurements that we have turns out to be in the rejection area
of the sampling distribution. If this happens we make the mistake that we do
a lot of tests with this woman to find out what’s wrong with her, while in fact
she is perfectly healthy! How bad would such a mistake be? It would certainly
lead to extra costs, but also a lot of the woman’s time. She would also probably
start worrying that something is wrong with her. So we definitely don’t want
this to happen. We can minimize the risk of a type I error by choosing a low α.
The type II error is the mistake of not rejecting the null-hypothesis while it
is in fact not true. Thus, the woman’s mean LH level is lower than 2.54, but by
coincidence, the sample mean of the 48 measurements that we have turns out
to be in the acceptance area of the sampling distribution. This means that the
woman’s LH level does not seem to be abnormal, and the woman is sent home.
How bad would such a mistake be? Well, pretty bad because the woman’s
hormone level is not normal, but everybody thinks that she is OK. She could
be very ill but nothing is found in further tests, because there are no further
tests. So we definitely don’t want this to happen. We can minimize the risk of
a type II error by choosing a higher α.
So here we have a conflict, and we have to make a balanced choice for α:
too low we run the risk of type II errors, too high we run the risk of a type I
error. Then you have to decide what is worse: a type I mistake or a type II
mistake. In this case, you could say that sending the woman home while she is
ill, is worse than spending money on tests that are actually not needed. Then
you would choose a rather high α, say 10%. That means that if you have several
women who are in fact healthy, 10% of them would receive extra testing. This
is a fairly high percentage, but you are more sure that women with an illness
will be detected and receive proper care.
But if you think it is most important that you don’t spend too much money
and that you don’t want women to start worrying when it is not needed, you
can pick a low α like 1%: then when you have a lot of healthy women, only 1%
of them will receive unnecessary testing.

95
Overview
ˆ Type I error: the mistake of rejecting the null-hypothesis, while it is
true
ˆ Type II error: the mistake of not rejecting the null-hypothesis, while it
is not true
ˆ α: the relative frequency we allow ourselves to make a type I error

ˆ β: the relative frequency of making a type II error

96
Chapter 3

Inference about a
proportion

3.1 Sampling distribution of the sample proportion

So far, we focused on inference about a population mean: starting from a sample
mean, what can we infer about the population mean? However, there are also
other sample statistics we could focus on. We briefly touched on the variance in
the sample and what it tells us about the population variance. In this section,
we focus on inference regarding a proportion.
Let’s go back to the example of the elephants in the zoo, and that the
manager saw a damaged doorway. This is most likely caused by elephants that
are taller than a certain height, making their heads bump the doorway when
moving from one space to the other. Let’s suppose the height of the doorway
is 3.40 m and that the manager observes that of the 4 elephants in the zoo, 3
bump their head when passing the doorway. Suppose that the 4 elephants are
randomly sampled from the entire population of elephants worldwide. What
could we say based on these observations about the proportion of elephants
worldwide that are taller than 3.40 m?
Let’s again start from the population. Let’s do the thought experiment
that the population proportion of elephants taller than 3.40 m equals 0.6: 60%
of all the elephants in the world are taller than 3.40 m. Let’s randomly pick
4 elephants from this population. We might get 2 tall elephants and 2 less
tall elephants. This means we get a sample proportion of 24 = 0.5. If we
do this sampling a lot of times, we obtain the sampling distribution of the
sample proportion. It is shown in Figure 3.1. It is a discrete (non-continuous)
distribution that is clearly not a normal distribution. But, as we know from the
Central Limit Theorem (Chapter 2), it will become a normal distribution when
sample size increases.
Actually, the sampling distribution that we see in Figure 3.1 is based on
the binomial distribution. Using the binomial distribution, we can calculate the

97
 0.3

probability

0.2

0.1

0.0
0 0.25 0.5 0.75 1
sample proportion

Figure 3.1: Sampling distribution of the sample proportion, when the population
proportion is 0.60

probabilities of getting various sample proportions in a straightforward manner,

without relying on the normal distribution.

3.2 The binomial distribution

The binomial distribution gives us the probability of obtaining a certain number
of elements, given how many elements there are in total and the population
probability. In our case, the binomial distribution gives us the probability of
having exactly 2 elephants taller than 3.40 m, given that there are 4 elephants
in our sample and the population proportion equals 0.6. Let’s go through the
reasoning step by step.
The proportion of tall elephants in the population is p = 0.6. The sample
size equals n = 4. Let’s begin with randomly picking the first elephant: what’s
the probability that we select an elephant that is taller than 3.40 m? Well, that
probability is equal to the proportion of 0.6. Next, what is the probability that
the second elephant is taller than 3.40? Again, this is equal to 0.6.
Now something more complicated: what is the probability that both the first
and the second elephant are taller than 3.40? This is equal to 0.6 × 0.6 = 0.36.
What is the probability that all 4 elephants are taller than 3.40 m? That is
equal to 0.6 × 0.6 × 0.6 × 0.6 = 0.604 = 0.1296. The probability that all 4
elephants are shorter than 3.40 m is equal to (1 − 0.6)4 = 0.44 = 0.0256.
The probability for a mix of 2 tall elephants and 2 shorter elephants is more
difficult to compute. You might remember from high school that it involves
combinations. For example, the probability that the first 2 elephants are taller
than 3.40, and the last 2 elephants shorter, is equal to 0.62 ×(1−0.6)2 = 0.0576,
but there are many other ways in which we can find 2 tall elephants and 2 shorter

98
elephants when we randomly and sequentially pick 4 elephants. There are in
fact 6 different ways of randomly selecting 4 elephants where only 2 are tall.
When we use A to denote a tall elephant and B to denote a short elephant,
the 6 possible combinations of having two As and two Bs are in fact: AABB,
BBAA, ABAB, BABA, ABBA, and BAAB.
This number of combinations is calculated using the binomial coefficient:
 
4 4!
= =6 (3.1)
2 2!2!

This number 42 (’four choose two’) is called the binomial coefficient. It can

be calculated using factorials: the exclamation mark ! stands for factorial. For
instance, 5! (’five factorial’) means 5 × 4 × 3 × 2 × 1.
In its general form, the binomial coefficient looks like:
 
n n!
= (3.2)
r r!(n − r)!
So suppose sample size n is equal to 4 and r equal to 2 (the number of tall
elephants in the sample), we get:
 
4 4! 4! 4×3×2×1
= = = =6 (3.3)
2 2!(n − r)! 2!2! 2×1×2×1

Going back to the elephant example, there are 42 = 6 possible ways of

getting 2 tall elephants and 2 short elephants when we sequentially pick 4

elephants. Each of these possibilities has a probability of 0.62 × (1 − 0.6)2 =
0.0576. This is explained in Table 3.1. For instance, the probability of getting
the ordering ABAB, is equal to the multiplication of the respective probabilities:
0.6×0.4×0.6×0.4. In the table you can see that the probability for any ordering
is always 0.0576. Since any ordering will qualify as obtaining 2 tall elephants
from a total of 4, we can sum these probabilities: the probability of getting the
ordering AABB or BBAA or ABAB or BABA or ABBA or BAAB, is equal to
0.0576 + 0.0576 + 0.0576 + 0.0576 + 0.0576 + 0.0576 = 6 × 0.0576 = 0.3456.
Here
6 is the number of combinations, calculated as the binomial coefficient 42 . We
could therefore in general compute the probability of having 2 tall elephants in
a sample of 4 as

 
4
p(#A = 2|n = 4, p = 0.6) = ×0.62 ×(1−0.6)2 = 6×0.0576 = 0.3456 (3.4)
2

The probability of ending up with 2 tall elephants in a sample of 4 elephants,

in any order, and where the proportion of tall elephants in the population is
0.6, is therefore equal to 0.3456.
In the more general case, if you have a population with a proportion p of As,
a sample size of n, and you want to know the probability of finding r instances

99
 0.3

probability

0.2

0.1

0.0
0 1 2 3 4
r

Figure 3.2: Binomial distribution with N = 4 and p = 0.60.

of A in your sample, it can be computed with the formula

 
n
p(#A = r|n, p) = × pr × (1 − p)(n−r) (3.5)
r
For example, the probability
of obtaining 3 tall elephants when the total
number of elephants is 4, is 43 × 0.63 × (1 − 0.6)1 = 4 × 0.216 × 0.4 = 0.3456.
When we calculate the probabilities of finding 0, 1, 2, 3, or 4 tall elephants
in sample of 4 when the population proportion is 0.6, we obtain the binomial
distribution that is plotted in Figure 3.2. It is exactly the same as the sampling
distribution in Figure 3.1, except that we plot the number of tall elephants in
the sample on the horizontal axis, instead of the proportion. This means that
we can use the binomial distribution to describe the sampling distribution of
the sample proportion. To get the proportions, we simply divide the number of
tall elephants in our sample by the total number of elephants (n) and we get
Figure 3.1.

Table 3.1: Four possible ways of selecting 2 tall elephants (A) and 2 short
elephants (B), together with the probability for each selection.
ordering computation of probability probability
AABB 0.6 x 0.6 x 0.4 x 0.4 0.0576
ABAB 0.6 x 0.4 x 0.6 x 0.4 0.0576
ABBA 0.6 x 0.4 x 0.4 x 0.6 0.0576
BAAB 0.4 x 0.6 x 0.6 x 0.4 0.0576
BABA 0.4 x 0.6 x 0.4 x 0.6 0.0576
BBAA 0.4 x 0.4 x 0.6 x 0.6 0.0576

100
Overview
ˆ sampling distribution of the sample proportion: the distribution
of proportions that you get when you randomly pick new samples from a
population and for each sample compute the proportion.

ˆ binomial distribution: a discrete distribution showing the probabilities

of finding a certain number of successes (r), given sample size n and
population proportion p.
ˆ binominal coefficient: a coefficient used to calculate binomial
probabilities. It represents the number of ways in which you can find
r instances in a sample of size n. It is calculated as nr = r!(n−r)!
n!


.

3.3 Confidence intervals

Based on what we know about the binomial distribution, we can perform inference
on proportions. In Chapter 2 we saw that inference is very much based on the
standard error (i.e., the standard deviation of the sampling distribution). We
know from theory that the variance of the binomial distribution can be easily
calculated as n×p×(1−p). Because we want to have the variance in proportions
rather than in numbers, we have to divide this variance by n to get the variance
of proportions: n×p×(1−p)
n = p×(1−p). Next, because the variance of a sampling
distribution gets smaller with increasing n, we divide by n again, in a similar
way as we did for the sampling distribution of the sample mean in Chapter 2.
Taking the square root of this variance gives us the standard deviation of the
sampling distribution (i.e., the standard error):
r
p(1 − p)
σp̂ = (3.6)
n
This standard error makes it easy to construct confidence intervals. We know
from the Central Limit Theorem that if n becomes infinitely large, the sampling
distribution will become normal. When n = 50, the sampling distribution is
already close to normal, as is shown in Figure 3.3. This fact, together with the
standard error makes it easy to construct approximate confidence intervals.
Suppose that we had 50 elephants in our zoo, and the manager observed that
42 of them bump their head against the doorway. That is a sample proportion of
42
50 = 0.84. When we want to have a range of plausible values for the population
proportion, we can construct a 95% confidence interval around this sample
proportion. Because we know that for the standard normal distribution, 95% of
the observations are between -1.96 and +1.96, we constructq the 95% confidence
interval by multiplying 1.96 with the standard error, σp̂ = p(1−p)
n .
However, since we do not know the population proportion p, we have to
estimate it. From theory, we know that an unbiased estimator for the population

101
 0.12

0.09
probability

0.06

0.03

0.00
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
proportion

Figure 3.3: Sampling distribution with N = 50 and p = 0.60.

proportion is the sample proportion: p̂ = 42 50 = 0.84. Our estimate for the

q
p̂(1−p̂)
standard error is then n = 0.0518459.
If we use that value, we get the interval from 0.84 − 1.96 × 0.0518459 to
0.84 + 1.96 × 0.0518459: thus, our 95% confidence interval for the population
proportion runs from 0.738382 to 0.941618.

3.4 Null-hypothesis concerning a proportion

Suppose that a researcher has measured all Tanzanian elephants and noted that
a proportion of 0.60 was taller than 3.40 m. Suppose also that the manager in
the zoo finds that 42 out of the 50 elephants bump their head and are therefore
taller than 3.40. How can we know that the elephants could be a representative
sample of Tanzanian elephants?
To answer this question with a yes or a no, we could apply the logic of null-
hypothesis testing. Let the null-hypothesis be that the population proportion
is equal to 0.60, and the alternative hypothesis that it is not equal to 0.60.

H0 : p = 0.60 (3.7)
HA : p 6= 0.60 (3.8)

Is the proportion of 0.84 that we observe in the sample (the zoo) a probable
value to find if the proportion of all Tanzanian is equal to 0.60? If this is the
case, we do not reject the null-hypothesis, and believe that the zoo data could
have been randomly selected from the Tanzanian population and are therefore
representative. However, if the proportion of 0.84 is very unprobable given that

102
the population proportion is 0.60, we reject the null-hypothesis and believe that
the data are not representative.
With null-hypothesis testing we always have to fix our α first: the probability
with which we are willing to accept a type I error. We feel it is really important
that the sample is representative of the population, so we definitely do not want
to make the mistake that we think the sample is representative (not rejecting
the null-hypothesis) while it isn’t (HA is true). This would be a type II error
(check this for yourself!). If we want to minimise the probability of a type II
error (β), we have to pick a relatively high α (see Chapter 2), so let’s choose
our α = .10.
Next, we have to choose a test statistic and determine critical values for it
that go with an α of .10. Because we have a relatively large sample size of 50,
we assume that the sampling distribution for a proportion of 0.60 is normal.
From the standard normal distribution, we know that 90% (1 − α!) of the
values lie between −1.6448536 and 1.6448536 (see Table 2.3). If we therefore
standardise our proportion, we have a measure that should show a standard
normal distribution:

ps − p0
zp = (3.9)
sd
where zp is the z-score for a proportion, ps is the sample proportion, p0 is
the population proportion assuming H0 , and sd is the standard deviation of the
sampling distribution, which is the standard error. Note that we should take
the standard error that we get when the null-hypothesis is true. We then get

0.84 − 0.6 0.24 0.24

zp = =q = = 3.4641016 (3.10)
se p0 (1−p0 ) 0.069282
n

90% of the values in any normal distribution lie between ±1.64 standard
deviations away from the mean (see Table 2.3). Here we see a z-score that
exceeds these critical values, and we therefore reject the null-hypothesis. We
conclude that the proportion of tall elephants observed in the sample is larger
than to be expected under the assumption that the population proportion is
0.6. We decide that the zoo data are not representative of the population data.
The decision process is illustrated in Figure 3.4.

3.5 Inference on proportions using R

Using the normal distribution is a nice trick when you have to do the calculations
by hand. However, this approach is of course only valid when you have large
sample sizes, so that you know that the shape of the normal distribution is a
good approximation of the binomial distribution. In contrast, using the binomial
distribution always gives you the most exact answers. However it can be very
tiresome to do all the computations by hand. In this section we discuss how to
let R do the calculations for you.

103
 0.4

0.3
density

0.2

0.1

0.0 5% 5%
3.464
−2 0 2
Z

Figure 3.4: A normal distribution to test the null-hypothesis that the population
proportion is 0.6. The blue line represents the z-score for our observed sample
proportion of 0.84.

Suppose we have a sample of 50 elephants, and we see that 42 of them bump

their head against the doorway. What can we say about the population: what
proportion of elephants in the entire population will bump their heads? In R,
we use the binom.test() function to do inference on proportions. This function
does all the calculations using the binomial distribution, so that the results are
always trustworthy, even for small sample sizes. We state the number of oberved
elephants that bump their head (x = 42), the sample size (n = 50), the kind
of confidence interval (95%: conf.level = 0.95) and the proportion that we
want to use for the null-hypothesis (p = 0.6):

binom.test(x = 42, n = 50, conf.level = 0.95, p = 0.6)

##
## Exact binomial test
##
## data: 42 and 50
## number of successes = 42, number of trials = 50, p-value = 0.0004116
## alternative hypothesis: true probability of success is not equal to 0.6
## 95 percent confidence interval:
## 0.7088737 0.9282992
## sample estimates:
## probability of success
## 0.84

The output shows the sample proportion: the probability of success is 0.84.
42
This is of course 50 . If we want to know what the population proportion is, we
look at the 95% confidence interval that runs from 0.7088737 to 0.9282992. If
you want to test the null-hypothesis that the population proportion is equal to

104
0.60, then we see that the p-value for that test is 0.0004116.
As said, the binomial test also works fine for small sample sizes. Let’s go
back to the very first example of this chapter: the zoo manager sees that of the
4 elephants they have, 3 bump their head and are therefore taller than 3.40 m.
What does that tell us about the proportion of elephants worldwide that are
taller than 3.40 m? If we assume that the 4 zoo elephants were randomly selected
from the entire population of elephants, we can use the binomial distribution.
In this case we type in R:

binom.test(x = 3, n = 4)

##
## Exact binomial test
##
## data: 3 and 4
## number of successes = 3, number of trials = 4, p-value = 0.625
## alternative hypothesis: true probability of success is not equal to 0.5
## 95 percent confidence interval:
## 0.1941204 0.9936905
## sample estimates:
## probability of success
## 0.75

By default, binom.test() yields 95% confidence intervals, as can be seen in

the output.1 We see that the confidence interval for the population proportion
runs from 0.1941204 to 0.9936905. Thus, based on this sample proportion of
0.75, we can see with some degree of confidence that the population proportion
is somewhere between 0.19 and 0.99. That’s of course not very informative,
which makes sense considering we only observe 4 elephants.

1 Note in the output that by default, binom.test() chooses the null-hypothesis that the

population proportion is 0.5.

105
106
Chapter 4

Linear modelling:
introduction

4.1 Dependent and independent variables

In the previous two chapters we discussed single variables. In Chapter 2 we
discussed a numeric variable that had a certain mean, for instance we talked
about the height of elephants. In Chapter 3 we talked about a dichotomous
categorical variable: elephants being taller than 3.40 m or not, with a certain
proportion of tall elephants. This chapter deals with the relationship between
two variables, more specifically the relationship between two numeric variables.
In Chapter 1 we discussed the distinction between numeric, ordinal and
categorical variables. In linear modelling, there is also another important distinction
between variables: dependent and independent variables. Dependency of a
variable is not really a property of a variable but it is the result of the data
analyst’s choice. Let’s first think about relationships between two variables.
Determining whether a variable is to be treated as independent or not, is often
either a case of logic or a case of theory. When studying the relationship between
the height of a mother and that of her child, the more logical it would be to
see the height of the child as dependent on the height of the mother. This
is because we assume that the genes are transferred from the mother to the
child. The mother comes first, and the height of the child is partly the result of
the mother’s genes that were transmitted during fertilisation. The height of a
child depends in part on the height of the mother. The variable that measures
the result is usually taken as the dependent variable. The theoretical cause or
antecedent is usually taken as the independent variable.
The dependent variable is often called the response variable. An independent
variable is often called a predictor variable or simply predictor. Independent
variables are also often called explanatory variables. We can explain a very tall
child by the genes that it got from its very tall mother. The height of a child
is then the response variable, and the height of the mother is the explanatory

107
variable. We can also predict the adult height of a child from the height of the
mother.
The dependent variable is usually the most central variable. It is the variable
that we’d like to understand better, or perhaps predict. The independent
variable is usually an explanatory variable: it explains why some people have
high values for the dependent variable and other people have low values. For
instance, we’d like to know why some people are healthier than others. Health
may then be our dependent variable. An explanatory variable might be age
(older people tend to be less healthy), or perhaps occupation (being a dive
instructor induces more health problems than being a university professor).
Sometimes we’re interested to see whether we can predict a variable. For
example, we might want to predict longevity. Age at death would then be our
dependent variable and our independent (predictor) variables might concern
lifestyle and genetic make-up.
Thus, we often see four types of relations:

ˆ Variable A affects/influences another variable B.

ˆ Variable A causes variable B.

ˆ Variable A explains variable B.

ˆ Variable A predicts variable B.

In all these four cases, variable A is the independent variable and variable
B is the dependent variable.
Note that in general, dependent variables can be either numeric, ordinal, or
categorical. Also independent variables can be numeric, ordinal, or categorical.

4.2 Linear equations

From secondary education you might remember linear equations. Suppose you
have two quantities, X and Y , and there is a straight line that describes best
their relationship. An example is given in Figure 4.1. We see that for every
value of X, there is only one value of Y . Moreover, the larger the value of X,
the larger the value of Y . If we look more closely, we see that for each increase
of 1 unit in X, there is an increase of 2 units in Y . For instance, if X = 1, we
see a Y -value of 2, and if X = 2 we see a Y -value of 4. So if we move from
X = 1 to X = 2 (a step of one on the X-axis), we move from 2 to 4 on the
Y -axis, which is an increase of 2 units. This increase of 2 units for every step
of 1 unit in X is the same for all values of X and Y . For instance, if we move
from 2 to 3 on the X-axis, we go from 4 to 6 on the Y -axis: an increase of again
2 units. This constant increase is typical for linear relationships. The increase
in Y for every unit increase in X is called the slope of a straight line. In this
figure, the slope is equal to 2.
The slope is one important characteristic of a straight line. The second
important property of a straight line is the intercept. The intercept is the value

108
 6

3
Y

−1

−2

−1 0 1 2 3 4 5
X

Figure 4.1: Straight line with intercept 0 and slope 2.

of Y , when X = 0. In Figure 4.1 we see that when X = 0, Y is 0, too. Therefore

the intercept of this straight line is 0.
With the intercept and the slope, we completely describe this straight line:
no other information is necessary. Such a straight line describes a linear relationship
between X and Y . The linear relationship can be formalised using a linear
equation. The general form of a linear equation for two variables X and Y is
the following:

Y = intercept + slope × X (4.1)

For the linear relationship between X and Y in Figure 4.1 the linear equation
is therefore

Y = 0 + 2X (4.2)

which can be simplified to

Y = 2X (4.3)

With this equation, we can find the Y -value for all values of X. For instance,
if we want to know the Y -value for X = 3.14, then using the linear equation we
know that Y = 2×3.14 = 6.28. If we want to know the Y -value for X = 49876.6,
we use the equation to obtain Y = 2 × 49876.6 = 99753.2. In short, the linear
equation is very helpful to quickly say what the Y -value is on the basis of the
X-value, even if we don’t have a graph of the relationship or if the graph does
not extent to certain X-values.

109
 1

0
Y

−1

−2

−1 0 1 2 3 4 5 6
X

Figure 4.2: Straight line with intercept -2 and slope 0.5.

In the linear equation, we call Y the dependent variable, and X the independent
variable. This is because the equation helps us determine or predict our value
of Y on the basis of what we know about the value of X. When we graph the
line that the equation represents, such as in Figure 4.1, the common way is to
put the dependent variable on the vertical axis, and the independent variable
on the horizontal axis.
Figure 4.2 shows a different linear relationship between X and Y . First we
look at the slope: we see that for every unit increase in X (from 1 to 2, or
from 4 to 5) we see an increase of 0.5 in Y . Therefore the slope is equal to 0.5.
Second, we look at the intercept: we see that when X = 0, Y has the value -2.
So the intercept is -2. Again, we can describe the linear relationship by a linear
equation, which is now:

Y = −2 + 0.5X (4.4)

Linear relationships can also be negative, see Figure 4.3. There, we see that
if we move from 0 to 1, we see a decrease of 2 in Y (we move from Y = −2 to
Y = −4), so −2 is our slope value. Because the slope is negative, we call the
relationship between the two variables negative. Further, when X = 0, we see
a Y -value of -2, and that is our intercept. The linear equation is therefore:

Y = −2 − 2X (4.5)

110
 0

−1

−2

−3

−4

−5
Y

−6

−7

−8

−9

−10

−11

−12

−1 0 1 2 3 4 5
X

Figure 4.3: Straight line with intercept -2 and slope -2.

Overview
ˆ dependent variable: the variable that we want to describe, understand,
predict or explain. Usually denoted as Y .

ˆ independent variable: the variable that we use in order to understand,

predict or explain something. Usually denoted as X.
ˆ linear relationship: two variables are said to be linearly related if their
relationship can be described by a linear equation with an intercept and
a slope.
ˆ intercept: the value for Y (dependent variable) if X = 0 (independent
variable).
ˆ slope: the change in Y when we increase X by 1 unit.

4.3 Linear regression

In the previous section, we saw perfect linear relationships between quantities
X and Y : for each X-value there was only one Y -value, and the values are all
described by a straight line. Such relationships we hope to see in physics, but
mostly see only in mathematics.
In social sciences we hardly ever see such perfectly linear relationships between
quantities (variables). For instance, let us plot the relationship between yearly

111
 1100

1000

900
Euros yearly spent on holidays
800

700

600

500

400

300

200

100

0 10 20 30 40 50 60 70 80 90 100
Yearly income in k Euros

Figure 4.4: Data on holiday spending.

income and the amount of Euros spent on holidays. Yearly income is measured in
thousands of Euros (k Euros), and money yearly spent on holidays is measured
in Euros. Let us regard money spent on holidays as our dependent variable
and yearly income as our independent variable (we assume money needs to be
saved before it can be spent). We therefore plot yearly income on the X-axis
(horizontal axis) and holiday spendings on the Y -axis (vertical axis). Let’s
imagine we find the data from 100 women between 30 and 40 years of age that
are plotted in Figure 4.4.
In the scatter plot, we see that one woman has a yearly income of 100,000
Euros, and that she spends almost 1100 Euros per year on holidays. We also
see a couple of women who earn less, between 10,000 and 20,000 Euros a year,
and they spend between 200 and 300 Euros per year on holiday.
The data obviously do not form a straight line. However, we tend to think
that the relationship between yearly income and holiday spending is more or
less linear: there is a general linear trend such that for every increase of 10,000
Euros in yearly income, there is an increase of about 100 Euros.
Let’s plot such a straight line that represents that general trend, with a slope
of 100 straight through the data points. The result is seen in Figure 4.5. We
see that the line with a slope of 100 is a nice approximation of the relationship
between yearly income and holiday spendings. We also see that the intercept of
the line is 100.
Given the intercept and slope, the linear equation for the straight line
approximating the relationship is

HolidaySpendings = 100 + 100 × YearlyIncome (4.6)

112
 1100

1000

900 Sandra Schmidt

Euros yearly spent on holidays

800

700

600

500

400

300

200

100

0 10 20 30 40 50 60 70 80 90 100
Yearly income in k Euros

Figure 4.5: Data on holiday spending with an added straight line.

In summary, data on two variables may not show a perfect linear relationship,
but in many cases, a perfect straight line can be a very reasonable approximation
of the data. Another word for a reasonable approximation of the data is a
prediction model. Finding such a straight line to approximate the data points
is called linear regression. In this chapter we will see what method we can
use to find a straight line. In linear regression we describe the behaviour of
the dependent variable (the Y -variable on the vertical axis) on the basis of
the independent variable (the X-value on the horizontal axis) using a linear
equation. We say that we regress variable Y on variable X.

4.4 Residuals
Even though a straight line can be a good approximation of a data set consisting
of two variables, it is hardly ever perfect: there are always discrepancies between
what the straight line describes and what the data actually tell us.
For instance, in Figure 4.5, we see a woman, Sandra Schmidt, who makes 69
k Euros a year and who spends 809 Euros on holidays. According to the linear
equation that describes the straight line, a woman that earns 69 k Euros a year
would spend 100 + 100 × 69 = 786 Euros on holidays. The discrepancy between
the actual amount spent and the amount prescribed by the linear equation
equals 809 − 786 = 23 Euros. This difference is rather small and the same holds
for all the other women in this data set. Such discrepancies between the actual
amount spent and the amount as prescribed or predicted by the straight line
are called residuals or errors. The residual (or error) is the difference between
a certain data point (the actual value) and what the linear equation predicts.

113
 2500

2000

1500
Y

1000

500

0 10 20 30 40 50 60 70 80 90 100
X

Figure 4.6: Data on variables X and Y with an added straight line.

Let us look at another fictitious data set where the residuals (errors) are a
bit larger. Figure 4.6 shows the relationship between variables X and Y . The
dots are the actual data points and the blue straight line is an approximation
of the actual relationship. The residuals are also visualised: sometimes the
observed Y -value is greater than the predicted Y -value (dots above the line)
and sometimes the observed Y -value is smaller than the predicted Y -value (dots
below the line). If we denote the ith predicted Y -value (predicted by the blue
line) as Ybi (pronounced as ’y-hat-i’), then we can define the residual or error as
the discrepancy between the observed Yi and the predicted Ybi :

ei = Yi − Ybi (4.7)

where ei stands for the error (residual) for the ith data point .
If we compute residual ei for all Y -values in the data set, we can plot them
using a histogram, as displayed in Figure 4.7. We see that the residuals are on
average 0, and that the histogram resembles the shape of a normal distribution.
We see that most of the residuals are around 0, and that means that most of
the values Y -values are close to the line (where the predicted values are). We
also see some large residuals but that there are not so many of these. Observing
a more or less normal distribution of residuals happens often in research. Here,
the residuals show a normal distribution with mean 0 and variance of 13336
(i.e., a standard deviation of 115).

114
 40

30
count

20

10

−400 −350 −300 −250 −200 −150 −100 −50 0 50 100 150 200 250 300 350 400
residuals

Figure 4.7: Histogram of the residuals (errors).

4.5 Least squares regression lines

You may ask yourself how to draw a straight line through the data points: How
do you decide on the exact slope and the exact intercept? And what if you don’t
want to draw the data points and the straight line by hand? That can be quite
cumbersome if you have more than 2000 data points to plot!
First, because we are lazy, we always use a computer to draw the data
points and the line, that we call a regression line. Second, since we could draw
many different straight lines through a scatter of points, we need a criterion to
determine a nice combination of intercept and slope. With such a criterion we
can then let the computer determine the regression line with its equation for us.
The criterion that we use in this chapter is called Least Squares, or Ordinary
Least Squares (OLS). To explain the Least Squares principle, look again at
Figure 4.6 where we see both small and large residuals. About half of them
are positive (above the blue line) and half of them are negative (below the blue
line).
The most reasonable idea is to draw a straight line that is more or less in the
middle of the Y -values, in other words, with about half of the residuals positive
and about half of them negative. Or perhaps we could say that on average, the
residuals should be 0. A third way of saying the same thing is that the sum of
the residuals should be equal to 0.
However, the criterion that all residuals should sum to 0 is not sufficient.
In Figure 4.8 we see a straight line with a slope of 0 where the residuals sum
to 0. However, this regression line does not make intuitive sense: it does not
describe the structure in the data very well. Moreover, we see that the residuals
are generally much larger than in Figure 4.6.

115
 2500

2000

1500
Y

1000

500

0 10 20 30 40 50 60 70 80 90 100
X

Figure 4.8: Data on variables X and Y with an added straight line. The sum
of the residuals equals 0.

We therefore need a second criterion to find a nice straight line. We want

the residuals to sum to 0, but also want the residuals to be as small as possible:
the discrepancies between what the linear equation predicts (the Yb -values) and
the actual Y -values should be as small as possible.
So now we have two criteria: we want the sum of the residuals to be 0 (about
half of them negative, half of them positive), and we want the residuals to be
as small as possible. We can achieve both of these when we use as our criterion
the idea that the sum of the squared residuals be as small as possible. Recall
from Chapter 1 that the sum of the squared deviations from the mean is closely
related to the variance. So if the sum of the squared residuals is as small as
possible, we know that the variance of the residuals is as small as possible. Thus,
as our criterion we can use the regression line for which the sum of the squared
differences between predicted and observed Y -values is as small as possible.
Figure 4.9 shows three different regression lines for the same data set. Figure
4.10 shows the respective distributions of the residuals. For the first line, we see
that the residuals sum to 0, for the residuals are on average 0 (the red vertical
line). However, we see quite large residuals. The residuals for the second line
are smaller: we see very small positive residuals, but the negative residuals are
still quite large. We also see that the residuals do not sum to 0. For the third
line, we see both criteria optimised: the sum of the residuals is zero and the
residuals are all very small. We see that for regression line 3, the sum of squared
residuals is at its minimum value. It can also be mathematically shown that if
we minimise the sum of squared differences between the predicted and observed
Y -values, they automatically show a mean of 0, satisfying the first criterion.

116
 1 2 3

2500

2000

1500
Y

1000

500

0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

X

Figure 4.9: Three times the same data set, but with different regression lines.

1 2 3
60

SSR: 18813520 SSR: 7711466 SSR: 2653808

40
count

20

−500 0 500 −500 0 500 −500 0 500

residual

Figure 4.10: Histogram of the residuals (errors) for three different regression
lines, and the respective sums of squared residuals (SSR).

117
 In summary, when we want to have a straight line that describes our data
best (i.e., the regression line), we’d like a line such that the residuals are on
average 0 (i.e, sum to 0), and where we see the smallest residuals possible. We
reach these criteria when we use the line in such a way that we have the lowest
value for the sum of the squared residuals possible. This line is therefore called
the least squares or OLS regression line.
There are generally two ways of finding the intercept and the slope values
that satisfy the Least Squares principle.

1. Numerical search Try some reasonable combinations of values for the

intercept and slope, and for each combination, calculate the sum of the
squared residuals. For the combination that shows the lowest value, try to
tweak the values of the intercept and slope a bit to find even lower values
for the sum of the squared residuals. Use some stopping rule otherwise
you keep looking forever.

2. Analytical approach For problems that are not too complex, like this
linear regression problem, there are simple mathematical equations to find
the combination of intercept and slope that gives the lowest sum of squared
residuals.

Using the analytical approach, it can be shown that the Least Squares slope
can be found by solving:

P
(Xi − X)(Yi − Y )
slope = P (4.8)
(Xi − X)2

and the Least Squares intercept can be found by:

intercept = Y − slope × X (4.9)

where X and Y are the means of the independent Xi and dependent Yi

observations, respectively.
In daily life, we do not compute this by hand but let computers do it for us,
with software like for instance R.

118
Overview
ˆ residual: the difference between a certain data point (the actual value)
and what the linear equation predicts.
ˆ linear regression: When we want to describe the behaviour of the
dependent variable (the Y -variable on the vertical axis) on the basis of the
independent variable (the X-value on the horizontal axis) by a straight
line, linear regression is the process of finding such a straight line.
ˆ Least Squares principle: In order to find the best regression line, you
need a criterion. The Least Squares principle is such a criterion and
specifies that the sum of the squares of the residuals should be as small
as possible.

4.6 Linear models

By performing a regression analysis of Y on X, we try to predict the Y -value
from a given X on the basis of a linear equation. We try to find an intercept
and a slope for that linear equation such that our prediction is ’best’. We define
’best’ as the linear equation for which we see the lowest possible value for the
sum of the squared residuals (least squares principle).
Thus, the prediction for the ith value of Y (Ybi ) can be computed by the
linear equation

Ybi = b0 + b1 Xi (4.10)
where we use b0 to denote the intercept, b1 to denote the slope and Xi as
the ith value of X.
In reality, the predicted values for Y always deviate from the observed values
of Y : there is practically always an error e that is the difference between Ybi and
Yi . Thus we have for the observed values of Y

Yi = Ybi + ei = b0 + b1 Xi + ei (4.11)
Typically, we assume that the residuals e have a normal distribution with a
mean of 0 and a variance that is often unknown but that we denote by σe2 . Such
a normal distribution is denoted by N (0, σe2 ). Taking the linear equation and
the normally distributed residuals together, we have a model for the variables
X and Y .

Yi = b0 + b1 Xi + ei (4.12)
ei ∼ N (0, σe2 ) (4.13)

119
 A model is a specification of how a set of variables relate to each other. Note
that the model for the residuals, the normal distribution, is an essential part
of the model. The linear equation only gives you predictions of the dependent
variable, not the variable itself. Together, the linear equation and the distribution
of the residuals give a full description of how the dependent variable depends on
the independent variable.
A model may be an adequate description of how variables relate to each
other or it may not, that is for the data analyst to decide. If it is an adequate
description, it may be used to predict yet unseen data on variable Y (because
we can’t see into the future), or it may be used to draw some inferences on data
that can’t be seen, perhaps because of limitations in data collection. Remember
Chapter 2 where we made a distinction between sample data and population
data. We could use the linear equation that we obtain using a sample of data
to make predictions for data in the population. We delve deeper into that issue
in Chapter 5.
The model that we see in Equations 4.12 and 4.13 is a very simple form
of the linear model. The linear model that we see here is generally known
as the simple regression model : the simple regression model is a linear model
for one numeric dependent variable, an intercept, a slope for only one (hence
’simple’) numeric independent variable, and normally distributed residuals. In
the remainder of this book, we will see a great variety of linear models: with one
or more independent variables, with numeric or with categorical independent
variables, and with numeric or categorical dependent variables. All these models
can be seen as extensions of this simple regression model. What they all have
in common is that they aim to predict one dependent variable from one or more
independent variables using a linear equation.

4.7 Finding the OLS intercept and slope using

R
Figure 4.11 shows a data set on the relationship between the number of cylinders
(cyl) and miles per gallon (mpg) in 1 cars. The blue line is the least squares
regression line. The coefficients for this line can be found with R using the
following code:
model < - mtcars % >%
lm ( mpg ∼ cyl , data = .)
model

In the syntax we first indicate that we start from the mtcars data set. Next,
we use the lm() function to indicate that we want to apply the linear model
to these data. Next, we say that we want to model the variable mpg. The ∼
(’tilde’) sign means ”is modelled by” or ”is predicted by”, and next we plug in
the independent variable cyl. Thus, this code says we want to model the mpg
variable by the cyl variable, or predict mpg scores by cyl. Next, because we

120
 35

30

25
mpg

20

15

10
4 5 6 7 8
cyl

Figure 4.11: Data set on number of cylinders (cyl) and miles per gallon (mpg)
in 32 cars.

already indicated we use the mtcars data set, the data argument for the lm()
function should be left empty. Finally, we store the results in the object model.
In the last line of code we indicate that we want to see the results, that we
stored in model.

model

##
## Call:
## lm(formula = mpg ~ cyl, data = .)
##
## Coefficients:
## (Intercept) cyl
## 37.885 -2.876

The output above shows us a repetition of the lm() analysis, and then two
coefficients. These are the regression coefficients that we wanted: the first is
the intercept, and the second is the slope. These coefficients are the parameters
of the regression model. Parameters are parts of a model that can vary from
data set to data set, but that are not variables (variables vary within a data set,
parameters do not). Here we use the linear model from Equations 4.12 and 4.13
where b0 , b1 and σe2 are parameters since they are different for different data
sets.
The output does not look very pretty. Using the broom package, we can get
the same information about the analysis, and more:

121
library(broom)
model <- mtcars %>%
lm(mpg ~ cyl, data = .)
model %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 37.9 2.07 18.3 8.37e-18
## 2 cyl -2.88 0.322 -8.92 6.11e-10

R then shows two rows of values, one for the intercept and one for the slope
parameter for cyl. For now, we only look at the first two columns. In these
columns we find the least squares values for these parameters for this data set
on 32 cars that we are analysing here.
In the second column, called estimate, we see that the intercept parameter
has the value 37.9 (when rounded to 1 decimal) and the slope has the value
-2.88. Thus, with this output, the linear equation for the regression equation
can be filled in:

mpg = 37.9 − 2.88 × cyl + e (4.14)

With this equation we can predict values for mpg for number of cylinders
that are not even in the data set displayed in Figure 4.11. For instance, that
plot does not show a car with 2 cylinders, but on the basis of the linear equation,
the best bet would be that such a car would run 37.9 − 2.88 × 2 = 32.14 miles
per gallon.
The OLS linear model parameters are in the estimate column of the R
output, but there are also a number of other columns: standard error, statistic
(t), and p-value, terms that we encountered earlier in Chapter 2. These columns
will be discussed further in Chapter 5.

4.8 Pearson correlation

For any set of two numeric variables, we can determine the least squares regression
line. However, it depends on the data set how well that regression line describes
the data. Figure 4.12 shows two different data sets on variables X and Y . Both
plots also show the least squares regression line, and they both turn out to be
exactly the same: Y = 100 + 10X.
We see that the regression line describes data set A very well (left panel):
the observed dots are very close to the line, which means that the residuals are
very small. The regression line does a worse job for data set B (right panel)
since there are quite large discrepancies between the observed Y -values and
the predicted Y -values. Put differently, the regression equation can be used

122
 A B

2000

1000
Y

−1000

0 25 50 75 100 0 25 50 75 100
X

Figure 4.12: Two data sets with the same regression line.

to predict Y -values in data set A very well, almost without error, whereas the
regression line cannot be used to predict Y -values in data set B very precisely.
The regression line is also the least squares regression line for data set B, so any
improvement by choosing another slope or intercept is not possible.
Francis Galton was the first to think about how to quantify this difference in
the ability of a regression line to predict the dependent variable. Karl Pearson
later worked on this measure and therefore it came to be called Pearson’s
correlation coefficient. It is a standardised measure, so that it can be used
to compare different data sets.
In order to get to Pearson’s correlation coefficient, you first need to standardise
both the independent variable, X, and the dependent variable, Y . You standardise
scores by taking their values, subtract the mean from them, and divide by the
standard deviation (see Chapter 1). So, in order to obtain a standardised value
for X = x we compute zX ,

x−X
zX = (4.15)
σX
and in order to obtain a standardised value for Y = y we compute zY ,

y−Y
zY = . (4.16)
σY
Let’s do this both for data set A and data set B, and plot the standardised
scores, see Figure 4.13. If we then plot the least squares regression lines for
the standardised values, we obtain different equations. For both data sets, the
intercept is 0 because by standardising the scores, the means become 0. But

123
 A B

0
Z_Y

−2

−1 0 1 −1 0 1
Z_X

Figure 4.13: Two data sets, with different regression lines after standardisation.

the slopes are different: in data set A, the slope is 0.997 and in data set B, the
slope is 0.376.

ZY = 0 + 0.997 × ZX = 0.997 × ZX (4.17)

ZY = 0 + 0.376 × ZX = 0.376 × ZX (4.18)

These two slopes, the slope for the regression of standardized Y -values on
standardized X-values, are the correlation coefficients for data sets A and B,
respectively. For obvious reasons, the correlation is sometimes also referred to
as the standardised slope coefficient or standardised regression coefficient.
Correlation stands for the co-relation between two variables. It tells you
how well one variable can be predicted from the other. The correlation is bi-
directional: the correlation between Y and X is the same as the correlation
between X and Y . For instance in Figure 4.13, if we would have put the ZX -
variable on the ZY -axis, and the ZY -variable on the ZX -axis, the slopes would be
exactly the same. This is true because the variances of the Y - and X-variables
are equal after standardisation (both variances equal to 1).
Since a slope can be negative, a correlation can be negative too. Furthermore,
a correlation is always between -1 and 1. Look at Figure 4.13: the correlation
between X and Y is 0.997. The dots are almost on a straight line. If the dots
would all be exactly on the straight line, the correlation would be 1.
Figure 4.14 shows a number of scatter plots of X and Y with different
correlations. Note that if dots are very close to the regression line, the correlation
can still be close to 0: if the slope is 0 (bottom-left panel), then one variable
cannot be predicted from the other variable, hence the correlation is 0, too.

124
 1 1
1 2

500

−500

−1000 0.54 −0.99

2 2
Y

1 2

500

−500

−1000 −0.04 −0.42

90 100 110 90 100 110

X

Figure 4.14: Various plots showing different correlations between variables X

and Y.

125
 In summary, the correlation coefficient indicates how well one variable can
be predicted from the other variable. It is the slope of the regression line if both
variables are standardised. If prediction is not possible (when the regression
slope is 0), the correlation is 0, too. If the prediction is perfect, without errors
(no residuals) and with a slope unequal to 0, then the correlation is either -1
or +1, depending on the sign of the slope. The correlation coefficient between
variables X and Y is usually denoted by rXY for the sample correlation and
ρXY (pronounced ’rho’) for the population correlation.

4.9 Covariance
The correlation ρXY as defined above is a standardised measure for how much
two variables co-relate. It is standardised in such a way that it can never be
outside the (-1, 1) interval. This standardisation happened through the division
of X and Y -values by their respective standard deviation. There exists also an
unstandardised measure for how much two variables co-relate: the covariance.
The correlation ρXY is the slope when X and Y each have variance 1. When
you multiply correlation ρXY by a quantity indicating the variation of the two
variables, you get the covariance. This quantity is the product of the two
respective standard deviations.
The covariance between variables X and Y , denoted by σXY , can be computed
as:

σXY = ρXY × σX × σY (4.19)

For example, if the variance of X equals 49 and the variance of Y equals
25, then the respective standard deviations are 7 and 5. If the correlation
between X and Y equals 0.5, then the covariance between X and Y is equal to
0.5 × 7 × 5 = 17.5.
Similar to the correlation, the covariance of two variables indicates by how
much they co-vary. For instance, if the variance of X is 3 and the variance of
Y is 5, then a covariance of 2 indicates that X and Y co-vary: if X increases
by a certain amount, Y also increases. If you want to know how many standard
deviations Y increases if X increases with one standard deviation, you can turn
the covariance into a correlation by dividing the covariance by the respective
standard deviations.
σXY 2
ρXY = = √ √ = 0.52 (4.20)
σX σY 3 5
Similar to correlations and slopes, covariances can also be negative.
Instead of computing the covariance on the basis of the correlation, you
can also compute the covariance using the data directly. The formula for the
covariance is

P
(Xi − X)(Yi − Y )
σXY = (4.21)
n

126
 so it is the mean of the squared cross-products of two variables.1 Note that
the numerator bears close resemblance to the numerator of the equation that
we use to find the least squares slope, see Equation 4.8. This is not strange
since both the slope and the covariance say something about the relationship
between two variables. Also note that in the equation that we use to find the
least squares slope the denominator bears close relationship to the formula for
P 2
2
the variance, since σX = (Xin−X) (see Chapter 1). We could therefore rewrite
Equation 4.8 that finds the least squares or OLS slope as:

P
(Xi − X)(Yi − Y )
slopeOLS = P (4.22)
(Xi − X)2
σXY × n
= 2 ×n
σX
σXY
= 2
σX
This shows how all three quantities slope, correlation and covariance say
something about the linear relationship between two variables. The slope says
how much the dependent variable increases if the independent variable increases
by 1, the correlation says how much of a standard deviation the dependent
variable increases if the independent variable increases by one standard deviation
(alternatively: the slope after standardisation), and the covariance is the mean
cross-product of two variables (alternatively: the unstandardised correlation).

4.10 Numerical example of covariance, correlation

and least square slope

Table 4.1: Computing cross-products for the covariance of two variables.

X Y X - meanX Y - meanY Crossproduct
-1 2 -0.60 2.20 -1.32
0 -1 0.40 -0.80 -0.32
1 -2 1.40 -1.80 -2.52
-2 1 -1.60 1.20 -1.92
0 -1 0.40 -0.80 -0.32

Table 4.1 shows a small data set on two variables X and Y with 5 observations.
The mean value of X is -0.4 and the mean value of Y is -0.2. If we subtract
1 Again, similar to what was said about the formula for the variance of a variable, on-line
P
(X −X)(Y −Y )
i i
you will often find the formula n−1
. The difference is that here we are talking
about the definition of the covariance of two observed variables, and that elsewhere one talks
about trying to estimate the covariance between two variables in the population. Similar
to the variance, the covariance in a sample is a biased estimator of the covariance in the
population. To remedy this bias, we divide the cross-products not by n but by n − 1

127
 2

Y 1

−1
−1.23

−2
−2 −1 0 1 2
X

Figure 4.15: Data example and the regression line.

the respective mean from each observed value and multiply, we get a column of
cross-products. For example, take the first row: X − X = −1 − (−0.4) = −0.6
and Y − Y = 2 − (−0.2) = 2.20. If we multiply these numbers we get the
cross-product −0.6 × 2.20 = −1.32. If we compute all cross-products and sum
them, we get -6.40. Dividing this by the number of observations (5), yields the
covariance: -1.28.
If we compute the variances of X and Y (see Chapter 1), we obtain 1.04 and
2.16, respectively. Taking the square roots we obtain the standard deviations:
1.0198039 and 1.4696938. Now we can calculate the correlation on the basis of
−1.28
the covariance as ρXY = σσXXYσY = 1.0198039×1.4696938 = −0.85.
We can also calculate the least squares slope as σσXY
2 = −1.28
1.04 = −1.23.
X
The original data are plotted in Figure 4.15 together with the regression
line. The standardised data and the corresponding regression line are plotted
in Figure 4.16. Note that the slopes are different, and that the slope of the
regression line for the standardised data is equal to the correlation.

4.11 Correlation, covariance and slopes in R

Let’s use the mtcars dataframe and compute the correlation between the number
of cylinders (cyl) and miles per gallon (mpg). We do that with the function
cor():

mtcars %>%
select(cyl, mpg) %>%
cor()

## cyl mpg

128
 2

1
Z_Y

−0.85

−1

−2
−2 −1 0 1 2
Z_X

Figure 4.16: Data example (standardised values) and the regression line.

## cyl 1.000000 -0.852162

## mpg -0.852162 1.000000

In the output we see a correlation matrix. On the diagonal are the correlations
of cyl and mpg with themselves, which are perfect (a correlation of 1). On the
off-diagonal, we see that the correlation between cyl and mpg equals -0.852162.
This is a strong negative correlation, which means that generally, the more
cylinders a car has, the lower the mileage. We can also compute the covariance,
with the function cov:

mtcars %>%
select(cyl, mpg) %>%
cov()

## cyl mpg
## cyl 3.189516 -9.172379
## mpg -9.172379 36.324103

On the off-diagonal we see that the covariance between cyl and mpg equals
-9.172379. On the diagonal we see the variances of cyl and mpg. Note that
R uses the formula with n − 1 in the denominator. If we want R to compute
the (co-)variance using n in the denominator, we have to write an alternative
function ourselves:

cov_alt <- function(x, y){

X <- (x - mean(x)) # deviations from mean x
Y <- (y - mean(y)) # deviations from mean y

129
 XY <- X %*% Y # multiply each X with each Y and sum them
return(XY / length(x)) # divide by n
}
cov_alt(mtcars$cyl, mtcars$mpg)

## [,1]
## [1,] -8.885742

To determine the least squares slope for the regression line of mpg on cyl,
we divide the covariance by the variance of cyl (Equation 5.1):

cov(mtcars$cyl, mtcars$mpg) / var(mtcars$cyl)

## [1] -2.87579

Note that both cov() and var() use n − 1. Since this cancels out if we do
the division, it doesn’t matter whether we use n or n − 1.
If we first standardise the data with the function scale() and then compute
the least squares slope, we get

z_mpg <- mtcars$mpg %>% scale() # standardise mpg

z_cyl <- mtcars$cyl %>% scale() # standardise cyl

cov(z_mpg, z_cyl) / var(z_cyl)

## [,1]
## [1,] -0.852162

cor(z_mpg, z_cyl)

## [,1]
## [1,] -0.852162

cov(z_mpg, z_cyl)

## [,1]
## [1,] -0.852162

We see from the output that the slope coefficient for the standardised situation
is equal to both the correlation and the covariance of the standardised values.
The data and the least squares regression line can be plotted using geom smooth():

mtcars %>%
ggplot(aes(x = cyl, y = mpg)) +
geom_point() +
geom_smooth(method = "lm", se = F)

130
 35

30

25
mpg

20

15

10
4 5 6 7 8
cyl

4.12 Explained and unexplained variance

So far in this chapter, we have seen relationships between two variables: one
dependent variable and one independent variable. The dependent variable we
usually denote as Y , and the independent variable we denote by X. The
relationship was modelled by a linear equation: an equation with an intercept
b0 and a slope parameter b1 :

Y = b0 + b1 X (4.23)
Further, we argued that in most cases, the relationship between X and Y
cannot be completely described by a straight line. Not all of the variation in Y
can be explained by the variation in X. Therefore, we have residuals e, defined
as the difference between the observed Y -value and the Y -value that is predicted
by the straight line, (denoted by Yb ):

e = Y − Yb (4.24)
Therefore, the relationship between X and Y is denoted by a regression
equation, where the relationship is approached by a linear equation, plus a
residual part e:

Y = b0 + b1 X + e (4.25)
The linear equation gives us only the predicted Y -value, Yb :

Yb = b0 + b1 X (4.26)
We’ve also seen that the residual e is assumed to have a normal distribution,
with mean 0 and variance σe2 :

e ∼ N (0, σe2 ) (4.27)

131
 Remember that linear models are used to explain (or predict) the variation
in Y : why are there both high values and low values for Y ? Where does the
variance in Y come from? Well, the linear model tells us that the variation is
in part explained by the variation in X. If b1 is positive, we predict a relatively
high value for Y for a high value of X, and we predict a relatively low value for
Y if we have a low value for X. If b1 is negative, it is of course in the opposite
direction. Thus, the variance in Y is in part explained by the variance in X,
and the rest of the variance can only be explained by the residuals e.

Var(Y ) = Var(Yb ) + Var(e) = Var(b0 + b1 X) + σe2 (4.28)

Because the residuals do not explain anything (we don’t know where these
residuals come from), we say that the explained variance of Y is only that part
of the variance that is explained by independent variable X: Var(b0 +b1 X). The
unexplained variance of Y is the variance of the residuals, σe2 . The explained
variance is often denoted by a ratio: the explained variance divided by the total
variance of Y :

Var(b0 + b1 X) Var(b0 + b1 X)
Varexplained = = (4.29)
Var(Y ) Var(b0 + b1 X) + σe2
From this equation we see that if the variance of the residuals is large, then
the explained variance is small. If the variance of the residuals is small, the
variance explained is large.

4.13 More than one predictor

In regression analysis, and in linear models in general, we try to make the
explained variance as large as possible. In other words, we try to minimise the
residual variance, σe2 . One way to do that is to use more than one independent
variable. If not all of the variance in Y is explained by X, then why not include
multiple independent variables?
Let’s use an example with data on the weight of books, the size of books
(area), and the volume of books. These data are available in R, and we will
show how to perform the followint analyses in a later section. Let’s try first to
predict the weight of a book, weight, on the basis of the volume of the book,
volume. Suppose we find the following regression equation and a value for σe2 :

weight = 107.7 + 0.71 × volume + e (4.30)

e ∼ N (0, 15362) (4.31)

In the data set, we see that the variance of the weight, Var(weight) is equal
to 72274. Since we also know the variance of the residuals, we can solve for the
variance explained by volume:

132
 Var(weight) = 72274 = Var(107.7 + 0.7 × volume) + 15362
Var(107.7 + 0.7 × volume) = 72274 − 15362 = 56912
So the proportion of explained variance is equal to 56912
72274 = 0.7874478. This
is quite a high proportion: nearly all of the variation in the weight of books is
explained by the variation in volume.

But let’s see if we can explain even more variance if we add an extra independent
variable. Suppose we know the area of each book. We expect that books with
a large surface area weigh more. Our linear equation then looks like this:

weight = 22.4 + 0.71 × volume + 0.5 × area + e (4.32)

e ∼ N (0, 6031) (4.33)
How much of the variance in weight does this equation explain? The amount
of explained variance equals the variance of weight minus the residual variance:
72274 − 6031 = 66243. The proportion of explained variance is then equal to
66243
72274 = 0.9165537. So the proportion of explained variance has increased!
Note that the variance of the residuals has decreased; this is the main reason
why the proportion of explained variance has increased. By adding the extra
independent variable, we can explain some of the variance that without this
variable could not be explained! In summary, by adding independent variables
to a regression equation, we can explain more of the variance of the dependent
variable. A regression analysis with more than one independent variable we
call multiple regression. Regression with only one independent variable is called
simple regression.

4.14 R-squared
With regression analysis, we try to explain the variance of the dependent variable.
With multiple regression, we use more than one independent variable to try to
explain this variance. In regression analysis, we use the term R-squared to
refer to the proportion of explained variance, usually denoted with the symbol
R2 . The unexplained variance is of course the variance of the residuals, Var(e),
usually denoted as σe2 . So suppose the variance of dependent variable Y equals
200, and the residual variance in a regression equation equals say 80, then R2
or the proportion of explained variance is (200 − 80)/200 = 0.60.

R2 = σexplained
2
/σY2 = (σY2 − σunexplained
2
)/σY2 = (σY2 − σe2 )/σY2 (4.34)
This is the definition of R-squared at the population level, where we know
the exact values of the variances. However, we do not know these variances,
since we only have a sample of all values.

133
 We know from Chapter 2 that we can take estimators of the variances σY2
and σe2 . We should not use the variance of Y observed in the sample, but the
unbiased estimator of the variance of Y in the population

2 Σi (Yi − Y )2
σc
Y = (4.35)
n−1
where n is sample size (see Section 2.3).
For σe2 we take the unbiased estimator of the variance of the residuals e in
the population

2 2
c2 = Σi (ei − ē) = Σi ei
σ (4.36)
e
n−1 n−1
Here we do not have to subtract the mean from the residuals, because the
mean is 0 by definition.
If we plug these estimators into Equation 4.34, we get

Σ(Yi −Y )2 Σe2
2 c2
Y − σe
σc n−1 − n−1i
c2
R = =
2 Σ(Yi −Y )2
σc
Y n−1
Σ(Yi − Y )2 − Σe2i Σ(Yi − Y )2 Σe2i
= 2
= −
Σ(Yi − Y ) Σ(Yi − Y )2 Σ(Yi − Y )2
SSR
= 1− (4.37)
SST
where SSR refers to the sum of the squared residuals (errors)2 , and SST
refers to the total sum of squares (the sum of the squared deviations from the
mean for variable Y ).
As we saw in Section 4.5, in a regression analysis, the intercept and slope
parameters are found by minimising the sum of squares of the residuals, SSR.
Since the variance of the residuals is based on this sum of squares, in any
regression analysis, the variance of the residuals is always as small as possible.
The values of the parameters for which the SSR (and by consequence the
variance) is smallest, are the least squares regression parameters. And if the
variance of the residuals is always minimised in a regression analysis, the explained
variance is always maximised!
Because in any least squares regression analysis based on a sample of data,
the explained variance is always maximised, we may overestimate the variance
explained in the population data. In regression analysis, we therefore very often
use an adjusted R-squared that takes this possible overestimation (inflation)
into account. The adjustment is based on the number of independent variables
and sample size.
2 In the literature and online, sometimes you see SSR and sometimes you see SSE, both

referring to the sum of the squared residuals

134
 The formula is

2 n−1
Radj = 1 − (1 − R2 )
n−p−1
where n is sample size and p is the number of independent variables. For
example, if R2 equals 0.10 and we have a sample size of 100, and 2 independent
100−1
variables, the adjusted R2 is equal to 1 − (1 − 0.10) 100−2−1 = 1 − (0.90) 99
97 =
0.08. Thus, the estimated proportion of variance explained at population level,
corrected for inflation, equals 0.08. Because R2 is inflated, the adjusted R2 is
never larger than the unadjusted R-squared.
2
Radj ≤ R2

4.15 Multiple regression in R

Let’s use the book data and run a multiple regression in R. The data set is
called allbacks and is available in the R package DAAG (you may need to install
that package first). The syntax looks very similar to simple regression, except
that we now specify two independent variables, volume and area, instead of
one. We combine these two independent variables using the +-sign.
library ( DAAG )
library ( broom )
model < - allbacks % >%
lm ( weight ∼ volume + area , data = .)
model % >%
tidy ()
Below we see the output:

library(DAAG)
library(broom)
model <- allbacks %>%
lm(weight ~ volume + area, data = .)
model %>%
tidy()

## # A tibble: 3 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 22.4 58.4 0.384 0.708
## 2 volume 0.708 0.0611 11.6 0.0000000707
## 3 area 0.468 0.102 4.59 0.000616

There we see an intercept, a slope parameter for volume and a slope parameter
for area. Remember from Section 4.2 that the intercept is the predicted value

135
when the independent variable has value 0. This extends to multiple regression:
the intercept is the predicted value when the independent variables all have
value 0. Thus, the output tells us that the predicted weight of a book that has
a volume of 0 and an area of 0, is 22.4. The slopes tell us that for every unit
increase in volume, the predicted weight increases by 0.708, and for every unit
increase in area, the predicted weight increases by 0.468.
So the linear model looks like:

weight = 22.4 + 0.708 × volume + 0.468 × area + e (4.38)

Thus, the predicted weight of a book that has a volume of 10 and an area
of 5, the expected weight is equal to 22.4 + 0.708 × 10 + 0.468 × 5 = 31.82.
In R, the R-squared and the adjusted R-squared can be obtained by first
making a summary of the results, and then accessing these statistics directly.
sum < - model % >% summary ()
sum $ r . squared
sum $ adj . r . squared

sum$r.squared

## [1] 0.9284738

sum$adj.r.squared

## [1] 0.9165527

The output tells you that the R-squared equals 0.93 and the adjusted R-
squared 0.92. The variance of the residuals can also be found in the summary
object:

sum$sigma^2

## [1] 6031.052

4.16 Multicollinearity
In general, if you add independent variables to a regression equation, the proportion
explained variance, R2 , increases. Suppose you have the following three regression
equations:

weight = b0 + b1 × volume + e (4.39)

weight = b0 + b1 × area + e (4.40)
weight = b0 + b1 × volume + b2 × area + e (4.41)

136
 If we carry out these three analyses, we obtain an R2 of 0.8026346 if we only
use volume as predictor, and an R2 of 0.1268163 if we only use area as predictor.
So perhaps you’d think that if we take both volume and area as predictors in
the model, we would get an R2 of 0.8026346 + 0.1268163 = 0.9294509. However,
if we carry out the multiple regression with volume and area, we obtain an R2
of 0.9284738, which is slightly less! This is not a rounding error, but results
from the fact that there is a correlation between the volume of a book and the
area of a book. Here it is a tiny correlation of 0.002, but nevertheless it affects
the proportion of variance explained when you use both these variables.
Let’s look at what happens when independent variables are strongly correlated.
Table 4.2 shows measurements on a breed of seals (only measurements on the
first 6 seals are shown). These data are in the dataframe cfseals in the package
DAAG. Often, the age of an animal is gauged from its weight: we assume that
heavier seals are older than lighter seals. If we carry out a simple regression of
age on weight, we get the output

library(DAAG)
data(cfseal) # available in package DAAG
out1 <- cfseal %>%
lm(age ~ weight , data = .)
out1 %>% tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 11.4 4.70 2.44 2.15e- 2
## 2 weight 0.817 0.0716 11.4 4.88e-12

var(cfseal$age) # total variance of age

## [1] 1090.855

summary(out1)$sigma^2 # variance of residuals

## [1] 200.0776

resulting in the equation:

age = 11.4 + 0.82 × weight + e (4.42)

e ∼ N (0, 200) (4.43)

From the data we calculate the variance of age, and we find that it is
1090.8551724. The variance of the residuals is 200, so that the proportion of
explained variance is (1090.8551724 − 200)/1090.8551724 = 0.8166576.
Since we also have data on the weight of the heart alone, we could try to
predict the age from the weight of the heart. Then we get output

137
 Table 4.2: Part of Cape Fur Seal Data.
age weight heart
33.00 27.50 127.70
10.00 24.30 93.20
10.00 22.00 84.50
10.00 18.50 85.40
12.00 28.00 182.00
18.00 23.80 130.00

out2 <- cfseal %>%

lm(age ~ heart , data = .)
out2 %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 20.6 5.21 3.95 0.000481
## 2 heart 0.113 0.0130 8.66 0.00000000209

sum2 <- out2 %>%

summary()
sum2$sigma^2 # variance of residuals

## [1] 307.1985

that leads to the equation:

age = 20.6 + 0.11 × heart + e (4.44)

e ∼ N (0, 307) (4.45)

Here the variance of the residuals is 307, so the proportion of explained

variance is (1090.8551724 − 370)/1090.8551724 = 0.6608166.
Now let’s see what happens if we include both total weight and weight of
the heart into the linear model. This results in the following output

out3 <- cfseal %>%

lm(age ~ heart + weight , data = .)
out3 %>% tidy()

## # A tibble: 3 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 10.3 4.99 2.06 0.0487

138
## 2 heart -0.0269 0.0373 -0.723 0.476
## 3 weight 0.993 0.254 3.91 0.000567

sum3 <- out3 %>% summary()

sum3$sigma^2 # variance of residuals

## [1] 203.55

with model equation:

age = 10.3 − 0.03 × heart + 0.99 × weight + e (4.46)

e ∼ N (0, 204) (4.47)

Here we see that the regression parameter for weight has increased from 0.82
to 0.99. At the same time, the regression parameter for heart has decreased,
has even become negative, from 0.11 to -0.03. From this equation we see that
there is a strong relationship between the total weight and the age of a seal, but
on top of that, for every unit increase in the weight of the heart, there is a very
small decrease in the expected age. The slope for heart has become practically
negligible, so we could say that on top of the effect of total weight, there is no
remaining relationship between the weight of the heart and age. In other words,
once we can use the total weight of a seal, there is no more information coming
from the weight of the heart.
This is because the total weight of a seal and the weight of its heart are
strongly correlated: heavy seals generally have heavy hearts. Here the correlation
turns out to be 0.96, almost perfect! This means that if you know the total
weight of a seal, you practically know the weight of its heart. This is logical of
course, since the total weight is a composite of all the weights of all the parts
of the animal: the total weight variable includes the weight of the heart.
Here we have seen, that if we use multiple regression, we should be aware
of how strongly the independent variables are correlated. Highly correlated
predictor variables do not add extra predictive power. Worse: they can cause
problems in obtaining regression parameters because it becomes hard to tell
which variable is more important: if they are strongly correlated (positive or
negative), then they measure almost the same thing!
When two predictor variables are perfectly correlated, either 1 or -1, regression
is no longer possible, the software stops and you get a warning. We call such
a situation multicollinearity. But also if the correlation is close to 1 or -1, you
should be very careful interpreting the regression parameters. If this happens,
try to find out what variables are highly correlated, and select the variable that
makes most sense.
In our seal data, there is a very high correlation between the variables heart
and weight that can cause computational and interpretation problems. It makes
more sense to use only the total weight variable, since when seals get older, all
their organs and limbs grow larger, not just their heart.

139
Salary 90000

60000

30000

0 2 4 6
Neuroticism

Figure 4.17: Simulated HR data set.

4.17 Simpson’s paradox

With multiple regression, you may uncover very surprising relationships between
two variables, that can never be found using simple regression. Here’s an
example from Paul van der Laken3 , who simulated a data set on the topic
of Human Resources (HR).
Assume you run a company with 1000 employees and you have asked all of
them to fill out a Big Five personality survey. Per individual, you therefore
have a score depicting their personality characteristic Neuroticism, which can
run from 0 (not at all neurotic) to 7 (very neurotic). Now you are interested
in the extent to which this Neuroticism of employees relates to their salary
(measured in Euros per year).
We carry out a simple regression, with salary as our dependent variable and
Neuroticism as our independent variable. We then find the following regression
equation:

salary = 45543 + 4912 × Neuroticism + e (4.48)

Figure 4.17 shows the data and the regression line. From this visualisation
it looks like Neuroticism relates positively to their yearly salary: more neurotic
people earn more salary than less neurotic people. More precisely, we see in the
equation that for every unit increase on the Neuroticism scale, the predicted
salary increases with 4912 Euros a year.
3 https://paulvanderlaken.com/2017/09/27/simpsons-paradox-two-hr-examples-with-r-

code/

140
 Next we run a multiple regression analysis. We suspect that one other very
important predictor for how much people earn is their educational background.
The Education variable has three levels: 0, 1 and 2. If we include both
Education and Neuroticism as independent variables and run the analysis,
we obtain the following regression equation:

salary = 50935 − 3176 × Neuroticism + 20979 × Education + e (4.49)

Note that we now find a negative slope parameter for the effect of Neuroticism!
This implies there is a relationship in the data where neurotic employees earn
less than their less neurotic colleagues! How can we reconcile this seeming
paradox? Which result should we trust: the one from the simple regression, or
the one from the multiple regression?
The answer is: neither. Or better: both! Both analyses give us different
information.
Let’s look at the last equation more closely. Suppose we make a prediction
for a person with a low educational background (Education = 0). Then the
equation tells us that the expected salary of a person with a neuroticism score
of 0 is around 50935, and of a person with a neuroticism score of 1 is around
47759. That’s an increase of -3176, which is the slope for Neuroticism in the
multiple regression. So for employees with low education, the more neurotic
employees earn less! If we do the same exercise for average education and high
education employees, we find exactly the same pattern: for each unit increase
in neuroticism, the predicted yearly salary drops by 3176 Euros.
It is true that in this company, the more neurotic persons generally earn
a higher salary. But if we take into account educational background, the
relationship flips around. This can be seen from Figure 4.18: looking only
at the people with a low educational background (Education = 0, the red
data points), then the more neurotic people earn less than their less neurotic
colleagues with a similar educational background. And the same is true for
people with an average education (Education = 1, the green data points) and
a high education (Education = 2, the blue data points). Only when you put
all employees together in one group, you see a positive relationship between
Neuroticism and salary.
Simpson’s paradox tells us that we should always be careful when interpreting
positive and negative correlations between two variables: what might be true
at the total group level, might not be true at the level of smaller subgroups.
Multiple linear regression helps us investigate correlations more deeply and
uncover exciting relationships between multiple variables.
Simpson’s paradox helps us in interpreting the slope coefficients in multiple
regression. In simple regression, when we only have one independent variable,
we saw that the slope for an independent variable A is the increase in the
dependent variable if we increase variable A by one unit. In multiple regression,
we have multiple independent variables, say A, B and C. The interpretation for
the slope coefficient for variable A is then the increase in the dependent variable

141
 Education
0
1
90000
2
Salary

60000

30000

0 2 4 6
Neuroticism

Figure 4.18: Same HR data, now with markers for different education levels.

if we increase variable A by one unit, with the other independent variables B

and C held constant. For example, the slope for variable A is the increase when
we take particular values for variables B and C, say B = 5 and C = 7.
Multiple regression therefore plays an important part in studying causation.
Suppose that a researcher finds in South-African beach data that on days with
high ice cream sales there are also more shark attacks. Might this indicate
that there is a causal relationship between ice cream sales and shark attacks?
Might bellies full of ice cream be more attractive to sharks? Or when there
are many shark attacks, might people prefer eating ice cream over swimming?
Alternatively, there might be a third variable that explains both the shark
attacks and the ice cream sales: temperature! Sharks attack during the summer
when temperature is high, and that’s also the time people eat more ice cream.
There is no causal relationship, since if you only look at data from sunny summer
days (holding temperature constant), you don’t see a relationship between shark
attacks and ice cream sales (just many shark attacks and high ice cream sales).
And if you only look at cold wintry days, you also see no relationship (no
shark attacks and no ice cream sales). But if you take all days into account,
you see a relationship between shark attacks and ice cream sales. Because this
correlation is non-causal and explained by the third variable temperature, we
call this correlation a spurious correlation.
This spurious correlation is plotted in Figure 4.19. If you look at all the
data points at once, you see a steep slope in the least squares regression line for
shark attacks and ice cream sales. However, if you hold temperature constant
by looking at only the light blue data points (high temperatures), there is no
linear relationship. Neither is there a linear relationship when you only look at

142
Number of ice creams sold per day

1000
temperature
30

20
500
10

10 20 30 40 50
Number of shark attacks per day

Figure 4.19: A spurious correlation between the number of shark attacks and
ice cream sales.

the dark blue data points (low temperatures).

143
144
Chapter 5

Inference for linear models

In Chapter 4 on regression we saw how a linear equation can describe a data

set: the linear equation describes the behaviour of one variable, the dependent
variable, on the basis of one or more other variables, the independent variable(s).
Sometimes we are indeed interested in the relationship between two variables in
one given data set. For instance, a primary school teacher wants to know how
well the exam grades in her class of last year predict how well the same students
do on another exam a year later.
But very often, researchers are not interested in the relationships between
variables in one data set on one specific group of people, but interested in the
relationship between variables in general, not limited to only the observed data.
For example, a researcher would like to know what the relationship is between
the temperature in a brewery and the amount of beer that goes into the beer
bottles. In order to study the effect of temperature on volume, the researcher
measures the volume of beer in a limited collection of 200 bottles under standard
conditions of 20 degrees Celsius and determines from log files the temperature
in the factory during production for each measured bottle. The linear equation
might be volume = 32.35 − 0.1207 × temp + e, see Figure 5.1. Thus, for every
unit increase in degrees Celsius, say from 20 to 21 degrees, the volume of beer
that is measured increases by -0.1207 centilitres, or put differently, the volume
of beer decreases by 0.1207.
But the researcher is not at all interested in these 200 bottles specifically:
the question is what would the linear equation be if the researcher had used
information about all bottles produced in the same factory? In other words,
we may know about the linear relationship between temperature and volume
in a sample of bottles, but we might really be interested to know what the
relationship would look like had we been able to measure the volume in the
population of all bottles.
In this chapter we will see how to do inference in the case of a linear
model. Many important concepts that we already saw in earlier chapters will
be mentioned again. Some repetition of those rather difficult concepts will be
helpful, especially when now discussed within the context of linear models.

145
Volume in centilitres at 20 degrees Celsius
32

30

28

17 18 19 20 21 22
Temperature in degrees Celsius during production

Figure 5.1: The relationship between temperature and volume in a sample of

200 bottles.

5.1 Population data and sample data

In the beer bottle example above, the volume of beer was measured in a total
of 200 bottles. Let’s do a thought experiment, similar to the one in Chapter
2. Suppose we could have access to volume data about all bottles of beer on
all days on which the factory was operating, including information about the
temperature for each day of production. Suppose that the total number of
bottles produced is 80,000 bottles. When we plot the volume of each bottle
against the temperature of the factory we get the scatter plot in Figure 5.2.
In our thought experiment, we could determine the regression equation using
all bottles that were produced: all 80,000 of them. We then find the blue
regression line displayed in Figure 5.2. Its equation is volume = 29.98 + 0.001 ×
temp + e. Thus, for every unit increase in temperature, the volume increases
by 0.001 centilitres. Thus, the slope is slightly positive in the population, but
negative in the sample of 200 bottles.
In the data example above, data were only collected on 200 bottles. These
bottles were randomly selected1 : there were many more bottles but we could
measure only a limited number of them. This explains why the regression
equation based on the sample differed from the regression equation based on
all bottles: we only see part of the data.
Here we see a discrepancy between the regression equation based on the
sample, and the regression equation based on the population. We have a slope
of 0.001 in the population, and we have a slope of -0.1207 in the sample. Also
the intercepts differ. To distinguish between the coefficients of the population
and coefficients of the sample, a population coefficient is often denoted by the
1 Random selection means that each of the 80,000 bottles had an equal probability to end

up in this sample of 200 bottles.

146
 34
Volume in centilitres at 20 degrees Celsius

32

30

28

26

17 18 19 20 21 22
Temperature in degrees Celsius during production

Figure 5.2: The relationship between temperature and volume in all 80,000
bottles.

Greek letter β and a sample coefficient by the Roman letter b.

P opulation : volume = β0 + β1 × temp = 29.98 + 0.001 × temp

Sample : volume = b0 + b1 × temp = 32.35 − 0.1207 × temp

The discrepancy between the two equations is simply the result of chance:
had we selected another sample of 200 bottles, we probably would have found
a different sample equation with a different slope and a different intercept.
The intercept and slope based on sample data are the result of chance and
therefore different from sample to sample. The population intercept and slope
(the true ones) are fixed, but unknown. If we want to know something about
the population intercept and slope, we only have the sample equation to go on.
Our best guess for the population equation is the sample equation; the unbiased
estimator for a regression coefficient in the population is the sample coefficient.
But how certain can we be about how close the sample intercept and slope are
to the population intercept and slope?

5.2 Random sampling and the standard error

In order to know how close the intercept and slope in a sample are to their
values in the population, we do another thought experiment. Let’s see what
happens if we take more than one random sample of 200 bottles.
We put the 200 bottles that we selected earlier back into the population and
we again blindly pick a new collection of 200 bottles. We then measure for each
bottle the volume of beer it contains and we determine the temperature in the
factory on the day of its production. We then apply a regression analysis and

147
determine the intercept and the slope. Next, we put these bottles back into
the population, draw a second random sample of 200 bottles and calculate the
intercept and slope again.
You can probably imagine that if we repeat this procedure of randomly
picking 200 bottles from a large population of 80,000, each time we find a
different intercept and a different slope. Let’s carry out this procedure 100
times by a computer. Table 5.1 shows the first 10 regression equations, each
based on a random sample of 200 bottles. If we then plot the histograms of
all 100 sample intercepts and sample slopes we get Figure 5.3. Remember from
Chapters 2 and 3 that these are called sampling distributions. Here we look at
the sampling distributions of the intercept and the slope.
The sampling distributions in Figure 5.3 show a large variation in the intercepts,
and a smaller variation in the slopes (i.e., all values very close to another).

Table 5.1: Ten different sample equations based on ten different random samples
from the population of bottles.
sample equation
1 volume = 28.87 + 0.06 x temperature + e
2 volume = 30.84 – 0.05 x temperature + e
3 volume = 31.05 – 0.06 x temperature + e
4 volume = 31.67 – 0.09 x temperature + e
5 volume = 30.59 – 0.03 x temperature + e
6 volume = 29.53 + 0.02 x temperature + e
7 volume = 28.36 + 0.08 x temperature + e
8 volume = 27.78 + 0.11 x temperature + e
9 volume = 28.29 + 0.09 x temperature + e
10 volume = 30.75 – 0.03 x temperature + e

For now, let’s focus on the slope. We do that because we are mostly
interested in the linear relationship between volume and temperature. However,
everything that follows also applies to the intercept. In Figure 5.4 we see the
histogram of the slopes if we carry out the random sampling 1000 times. We
see that on average, the sample slope is around 0.001, which is the population
slope (the slope if we analyse all bottles). But there is variation around that
mean of 0.001: the standard deviation of all 1000 sample slopes turns out to be
0.08.
Remember from Chapter 2 that the standard deviation of the sampling
distribution is called the standard error. The standard error for the sampling
distribution of the sample slope represents the uncertainty about the population
slope. If the standard error is large, it means that if we would draw many
different random samples from the same population data, we would get very
different sample slopes. If the standard error is small, it means that if we would
draw many different random samples from the same population data, we would
get sample slopes that are very close to one another, and very close to the

148
 intercept slope

40

30
count

20

10

0
0 10 20 30 0 10 20 30

Figure 5.3: Distribution of the 100 sample intercepts and 100 sample slope.

60

40
count

20

0
−0.2 0.0 0.2
sample.slope

Figure 5.4: Distribution of 1000 sample slopes.

149
population slope.2

5.2.1 Standard error and sample size

Similar to the sample mean, the standard error for a sample slope depends on
the sample size: how many bottles there are in each random sample. The larger
the sample size, the smaller the standard error, the more certain we are about
the population slope. In the above example, the sample size is 200 bottles.
The left panel of Figure 5.6 shows the distribution of the sample slope where
the sample size is 2. You see that for quite a number of samples, the slope is
larger than 10, even if the population slope is 0.001. But when you increase the
number of bottles per sample to 20 (in the right panel), you are less dependent
on chance observations. With large sample sizes, your results from a regression
analysis become less dependent on chance, become more stable, and therefore
more reliable.
In Figure 5.6 we see the sampling distributions of the sample slope where
the sample size is either 2 (left panel) or 20 (right panel). We see quite a lot
of variation in sample slopes with sample size equal to 2, and considerably less
variation in sample slopes if sample size is 20. This shows that the larger the
sample size, the smaller the standard error, the larger the certainty about the
population slope. The dependence of a sample slope on chance and sample size
is also illustrated in Figure 5.5.

5.2.2 From sample slope to population slope

In the previous section we saw that if we have a small standard error, we can
be relatively certain that our sample slope is close to the population slope.
We did a thought experiment where we knew everything about the population
intercept and slope, and we drew many samples from this population. In reality,
we don’t know anything about the population: we only have one sample of data.
So suppose we draw a sample of 200 bottles from an unknown population of
bottles, and we find a slope of 1, we have to look at the standard error to know
how close that sample slope is to the population slope.
For example, suppose we find a sample slope of 1 and the standard error is
equal to 0.1. Then we know that the population slope is more likely to be in
the neighbourhood of values like 0.9, 1.0, or 1.1 than in the neighbourhood of
10 or -10 (we know that when using the empirical rule, see Chap. 1).
Now suppose we find a sample slope of 1 and the standard error is equal
to 10. Then we know that the sample slope is more likely to be somewhere
in the neighbourhood of values like -9, 1 or 11, than around values in the
neighbourhood of -100 or +100. However, values like -9, 1 and 11 are quite
far apart, so actually we have no idea what the population slope is; we don’t
even know whether the population slope is positive or negative! The standard
error is simply too large.
2 Because sample slopes cluster around the population slope, the sample slope is very close

to the population slope when the standard error is small.

150
 32

31
volume

30

29

28
18 19 20 21
temperature

Figure 5.5: The averaging effect of increasing sample size. The scatter plot
shows the relationship between temperature and volume for a random sample
of 20 bottles (the dots); the first two bottles in the sample are marked in red.
The red line would be the sample slope based on these first two bottles, the blue
line is the sample slope based on all 20 bottles, and the black line represents the
population slope, based on all 80,000 bottles. This illustrates that the larger
the sample size, the closer the sample regression line is expected to be to the
population regression line.

sample size 2 sample size 20

600

400
count

200

0
−20 −10 0 10 20 −20 −10 0 10 20
sample slope

Figure 5.6: Distribution of the sample slope when sample size is 2 (left panel)
and when sample size is 20 (right panel).

151
 As we have seen, the standard error depends very much on sample size.
Apart from sample size, the standard error for a slope also depends on the
variance of the independent variable, the variance of the dependent variable,
and the correlations between the independent variable and other independent
variables in the equation. We will not bore you with the complicated formula for
the standard error for regression coefficients in the case of multiple regression
3
. But here is the formula for the standard error for the slope coefficient if you
have only one predictor variable X:

s
s2R
σbb1 =
s2X × (n − 1)
v
u s
Σi (Yi −Ybi )2
Σi (Yi − Ybi )2
u
u n−2
= t = (5.1)
Σi (Xi −X)2 (n − 2)Σi (Xi − X)2
n−1 × (n − 1)

where b1 is the slope coefficient in the sample, n is sample size, s2R is the
sample variance of the residuals, and s2X the sample variance of independent
variable X. From the formula, you can see that the standard error σbb1 becomes
smaller when sample size n becomes larger.
It’s not very useful to memorise this formula; you’d better let R do the
calculations for you. But an interesting part of the formula is the nominator:
SSR
n−2 . This is the sum of the squared residuals, divided by n − 2. Remember
from Chapter 1 that the definition of the variance is the sum of squares divided
by the number of values. Thus it looks like we are looking at the variance of
the residuals. Remember from Chapter 2 that when we want to estimate a
population variance, a biased estimator is the variance in the sample. In order
to get an unbiased estimate of the variance, we have to divide by n−1 instead of
n. This was because when computing the sum of squares, we assume we know
the mean. Here we are computing the variance of the residuals, but it’s actually
an unbiased estimator of the variance in the population, because we divide by
n − 2: when we compute the residuals, we assume we know the intercept and
the slope. We assume two parameters, so we divide by n − 2. Thus, when we
have a linear model with 2 parameters (intercept and slope), we have to divide
the sum of squared residuals by n − 2 in order to obtain an unbiased estimator
of the variance of the residuals in the population.
From the equation, we see that the standard error becomes larger when there
is a large variation in the residuals, it becomes smaller when there is a large
variation in predictor variable X, and it becomes smaller with large sample size
n.

3 See https://www3.nd.edu/ rwilliam/stats1/x91.pdf for the formula. In this pdf, ’IV’

means independent variable

152
5.3 t-distribution for the model coefficients
When we look at the sample distribution of the sample slope, for instance
in Figure 5.4, we notice that the distribution looks very much like a normal
distribution. From the Central Limit Theorem, we know that the sampling
distribution will become very close to normal for large sample sizes. Using this
sampling distribution for the slope we could compute confidence intervals and
do null-hypothesis testing, similar to what we did in Chapters 2 and 3.
For large sample sizes, we could assume the normal distribution, and when
we standardise the slope coefficient, we can look up in tables such as in Appendix
A the critical value for a particular confidence interval. For instance, 200 bottles
is a large sample size. When we standardise the sample slope – let’s assume
we find a slope of 0.05 –, we need to use the values -1.96 and +1.96 to obtain
a 95% confidence interval around 0.05. The margin of error (MoE) is then
1.96 times the standard error. Suppose that the standard error is 0.10. The
MoE is then equal to 1.96 × 0.10 = 0.196. The 95% interval then runs from
0.05 − 0.196 = −0.146 to 0.05 + 0.196 = 0.246.
However, this approach does not work for small sample sizes. Again this can
be seen when we standardise the sampling distribution. When we standardise
the slope for each sample, we subtract the sample slope from the population
slope β1 , and have to divide each time by the standard error (the standard
deviation). But when we do that

b1 − β 1 b1 − β1
t= =r (5.2)
σc
bb1 s2R
s2X ×(n−1)

we immediately see the problem that when we only have sample data, we
have to estimate the standard error. In each sample, we get a slightly different
estimated standard error, because each time, the variation in the residuals (s2R )
is a little bit different, and also the variation in the predictor variable (s2X ). If
sample size is large, this is not so bad: we then can get very good estimates of the
standard error so there is little variation across samples. But when sample size
is small, both s2R and s2X are different from sample to sample (due to chance),
and the estimate of the standard error will therefore also vary a lot. The result
is that the distribution of the standardised t-value from Equation 5.2 will only
be close to normal for large sample size, but will have a t-distribution in general.
Because the standard error is based on the variance of the residuals, and
because the variance of the residuals can only be computed if you assume a
certain intercept and a certain slope, the degrees of freedom will be n − 2.
Let’s go back to the example of the beer bottles. In our first random sample
of 200 bottles, we found a sample slope of -0.121. We also happened to know
the population slope, which was 0.001. From our computer experiment, we saw
that the standard deviation of the sample slopes with sample size 200 was equal
to 0.08. Thus, if we fill in the formula for the standardised slope t, we get for

153
this particular sample

−0.1207 − 0.001
t= = −1.52 (5.3)
0.08
In this section, when discussing t-statistics, we assumed we knew the population
slope β, that is, the slope of the linear equation based on all 80,000 bottles. In
reality, we never know the population slope: the whole reason to look at the
sample slope is to have an idea about the population slope. Let’s look at the
confidence interval for slopes.

5.4 Confidence intervals for the slope

Since we don’t know the actual value of the population slope β1 , we could ask
the personnel in the beer factory what they think is a likely value for the slope.
Suppose Mark says he believes that a slope of 0.1 could be true. Well, let’s
find out whether that is a reasonable guess, given that the sample slope is -
0.121. Now we assume that the population slope β1 is 0.1, and we compute the
t-statistic for our sample slope -0.121:

−0.121 − 0.1
t= = −2.7 (5.4)
0.08
Thus, we compute how many standard errors the sample value is away from
the hypothesised population value 0.1. If the population value is indeed 0.1,
how likely is it that we find a sample slope of -0.121?
From the t-distribution, we know that such a t-value is very unlikely: the
probability of finding a sample slope -2.7 standard deviations or more away
from a population slope of 0.1 is less than 0.0075341. How do we know that?
Well, the t-statistic is -2.7 and the degrees of freedom is 200 − 2 = 198. The
cumulative proportion of a t-value can be looked up in R:

pt(-2.7, df = 198)

## [1] 0.003767051

That means that a proportion of 0.0037671 of all values in the t-distribution

with 198 degrees of freedom are lower than -2.7. Because the t-distribution
is symmetric, we then also know that 0.0037671 of all values are larger than
2.7. If we add up these two numbers, we know that 0.0075341 of all values
in a t-distribution are less than -2.7 or more than 2.7. That means that if the
population slope is 0.1, we only find a sample slope of ±−0.121 or more extreme
with a probability of 0.0075341. That’s very unlikely.
Because we know that such a t-value of ± − 2.7 or more extreme is unlikely,
we know that a sample slope of -0.1206874 is unlikely if the population slope is
equal to 0.1. Therefore, we feel 0.1 is not a realistic value for the population
slope.

154
 Now let’s ask Martha. She thinks a reasonable value for the population slope
is 0, as she doesn’t believe there is a linear relationship between temperature
and volume. She suspects that the fact that we found a sample slope that was
not 0 was a pure coincidence. Based on that hypothesis, we compute t again
and find:

−0.121 − 0
t= = −1.5 (5.5)
0.08
In other words, if we believe Martha, our sample slope is only about 1
standard deviation away from her hypothesised value. That’s not a very bad
idea, since from the t-distribution we know that the probability of finding a
value more than 1.5 standard deviations away from the mean (above or below)
is 13.35%. You can see that by asking R:

pt(-1.5, df = 198) * 2

## [1] 0.1352072

Thirteen percent, that’s about 1 in 7 or 8 times. That’s not so improbable.

In other words, if the population slope is truly 0, then our sample slope of
−0.121 is quite a reasonable finding. If we reverse this line of reasoning: if our
sample slope is −0.121, with a standard error of 0.08, then a population slope
of 0 is quite a reasonable guess! It is reasonable, since the difference between
the sample slope and the hypothesised value is only 1.5 standard errors.
So when do we no longer feel that a person’s guess of the population slope
is reasonable? Perhaps if the probability of finding a sample slope of at least a
certain size given a hypothesised population slope is so small that we no longer
believe that the hypothesised value is reasonable. We might for example choose
a small probability like 1%. We know from the t-distribution with 198 degrees
of freedom that 1% of the values lie at least 2.6 standard deviations above and
below the mean.

qt(0.005, df = 198)

## [1] -2.600887

This is shown in Figure 5.7. So if our sample slope is more than 2.6 standard
errors away from the hypothesised population slope, then that population slope
is not a reasonable guess. In other words, if the distance between the sample
slope and the hypothesised population slope is more than 2.6 standard errors,
then the hypothesised population slope is no longer reasonable.
This implies that any value closer than 2.6 standard errors from the sample
slope is a collection of reasonable values for the population slope.
Thus, in our example of the 200 bottles with a sample slope of −0.121 and a
standard error of 0.08, the interval from −0.121−2.6×0.08 to −0.121+2.6×0.08
contains reasonable values for the population slope. If we do the calculations,

155
 0.4

density 0.3

0.2

0.1

0.5 percent 0.5 percent

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
t

Figure 5.7: The t-distribution with 198 degrees of freedom.

we get the interval from −0.33 to 0.09. If we would have to guess the value for
the population slope, our guess would be that it would lie somewhere between
between -0.33 and 0.09, if we feel that 1% is a small enough probability.
In data analysis, such an interval that contains reasonable values for the
population value, if we only know the sample value, is called a confidence
interval, as we know from Chapter 2. Here we’ve chosen to use 2.6 standard
errors as our cut-off point, because we felt that 1% would be a small enough
probability to dismiss the real population value as a reasonable candidate (type
I error rate). Such a confidence interval based on this 1% cut-off point is called
a 99% confidence interval.
Particularly in social and behavioural sciences, one also sees 95% confidence
intervals. The critical t-value for a type I error rate of 0.05 and 198 degrees of
freedom is 1.97.

qt(0.975, df = 198)

## [1] 1.972017

Thus, 5% of the observations lie more than 1.97 standard deviations away
from the mean, so that the 95% confidence interval is constructed by subtracting/adding
1.97 standard errors from/to the sample slope. Thus, in the case of our bottle
sample, the 95% confidence interval for the population slope is from −0.121 −
1.97 × 0.08 to −0.121 + 1.97 × 0.08, so reasonable values for the population slope
are those values between −0.28 and 0.04. Luckily, this corresponds to the truth,
because we happen to know that the population slope is equal to 0.001. In real
life, we don’t know the population slope and of course it might happen that the
true population value is not in the 95% confidence interval. If you want to make
the likelihood of this being the case smaller, then you can use a 99%, a 99.9%

156
or an even larger confidence interval.

5.5 Residual degrees of freedom in linear models

What does the term, ”degrees of freedom” mean? In Chapter 2 we discussed
degrees of freedom in the context of doing inference about a population mean.
We saw that degrees of freedom referred to the number of values in the final
calculation of a statistic that are free to vary. More specifically, the degrees of
freedom for a statistic like t are equal to the number of independent scores that
go into the estimate, minus the number of parameters used as intermediate steps
in the estimation of the parameter itself. There, we computed a t-statistic for the
sample mean. Because in the computation of the t-statistic for a sample mean,
we divide by the standard error for the mean, and that this in turn requires
assuming a certain value for the mean, we had n − 1 degrees of freedom.
Here, we are talking about t-statistics for linear models. In the case of a
simple regression model, we only have one intercept and one slope. In order to
compute a t-value, for the slope for example, we have to estimate the standard
error as well. In Equation 5.1 we see that in order to estimate the standard
error, we need to compute the residuals ei = Yi − Ŷi . But you can only compute
residuals if you have predictions for the dependent variable, Ŷi , and for that you
need an intercept and a slope coefficient. Thus, we need to assume we know
two parameters, in order to calculate a t-value. With sample means we only
assumed we knew the mean, and therefore had n − 1 degrees of freedom. In
case of a linear model where we assume one intercept and one slope, we have
n − 2 degrees of freedom. For the same reason, if we have a linear model with
one intercept and two slopes (multiple regression with two predictors), we have
n − 3 degrees of freedom. In general then, if we have a linear model with K
independent variables, we have n − K − 1 degrees of freedom associated with
our t-statistic.
To convince you of this, we illustrate the idea of degrees of freedom in a
numerical example. Suppose that we have a sample with four Y values: 2, 6,
5, 2. There are four separate pieces of information here. There is no particular
connection between these values. They are free to take any values, in principle.
We could say that there are “four degrees of freedom” associated with this
sample of data.
Now, suppose that I tell you that three of the values in the sample are 2, 6,
and 2; and I also tell you that the sample mean is 3.75. You can immediately
deduce that the remaining value has to be 5. Were it any other value, the mean
would not be 3.75.

ΣYi 2 + 6 + Y3 + 2 10 + Y3
Y = = = = 3.75
n 4 4
10 + Y3 = 4 × 3.75 = 15
Y3 = 15 − 10 = 5

157
 6

4 Y = 3.75 + e
Y

0
1 2 3 4
X

Figure 5.8: Illustration of residual degrees of freedom in a linear model, in case

there is no slope and the intercept equals 3.75.

Once I tell you that the sample mean is 3.75, I am effectively introducing a
constraint. The value of the unknown sample value is implicitly being determined
from the other three values plus the constraint. That is, once the constraint is
introduced, there are only three logically independent pieces of information in
the sample. That is to say, there are only three ”degrees of freedom”, once the
sample mean is revealed.
Let’s carry this example to regression analysis. Suppose I have four observations
of variables X and Y , where the values for X are 1, 2, 3 and 4. Each value of
Y = y is one piece of information. These Y -values could be anything, so we
say that we have 4 degrees of freedom. Now suppose I use a linear model for
these data points, and suppose I only use an intercept. Let the intercept be 3.75
so that we have Y = 3.75 + e. Now the first bit of information for X = 1, Y
could be anything, say 2. The second and third bits of information for X = 2
and X = 4 could also be anything, say 6 and 2. Figure 5.8 shows these bits of
information as dots in a scatter plot. Since we know that the intercept is equal
to 3.75, with no slope (slope=0), we can also draw the regression line.
Before we continue, you must know that if we talk about degrees of freedom
in regression analysis, we generally talk about residual degrees of freedom. We
therefore look at residuals. If we compute the residuals, we have residuals -1.75,
2.25 and -1.75 for these data points. When we sum them, we get -1.25. Since
we know that all residuals should sum to 0 in a regression analysis (see Chapter
4), we can derive the fourth residual to be +1.25, since only then the residuals
sum to 0. Therefore, the Y -value for the fourth data point (for X = 3) has to
be 5, since then the residual is equal to 5 − 3.75 = 1.25.
In short, when we use a linear model with only an intercept, the degrees of
freedom is equal to the number of data points (combinations of X and Y ) minus

158
 12

8
Y

4 Y = 3 + 1X + e

0
1 2 3 4
X

Figure 5.9: Illustration of residual degrees of freedom, in case of a linear model

with both intercept and slope for four data points (black line). The blue line is
the regression line only using the three known data points.

1, or in short notation: n − 1, where n stands for sample size.

Now let’s look at the situation where we use a linear model with both an
intercept and a slope: suppose the intercept is equal to 3 and the slope is equal
to 1: Y = 3 + 1X + e. Then suppose we have the same X-values as the example
above: 1, 2 and 4. When we give these X-values corresponding Y -values, 2, 6,
and 2, we get the plot in Figure 5.9.
The black line is the regression line that we get when we analyse the complete
data set of four points, Y = 3 + 1X. The blue line is the regression line based
on only the three visible data points. Now the question is, is it possible for a
fourth data point with X = 3, to think of a Y -value such that the regression
line based on these four data points is equal to Y = 3 + 1X? In other words,
can we choose a Y -value such that the blue line exactly overlaps with the black
line?
Figure 5.10 shows a number of possibilities for the value of Y if X = 3.
It can be seen, that it is impossible to pick a value for Y3 such that we get a
regression equation Y = 3 + 1X. The blue line and green line intersect the black
line at X = 1, but they have slopes that are less steep than the black line. If
you use lower values for Y such as 9 (red line) or higher values like 15 (purple
line), the regression lines still do not overlap. It turns out to be impossible to
choose a value for Y3 in such a way that the regression line matches Y = 3 + 1X.
So, with sample size n = 4, we can never freely choose 3 residuals in order to
satisfy the constraint that a particular regression equation holds for all 4 data
points. We have less than 3 degrees of freedom because it is impossible to think
of a fitting fourth value. It turns out, that in this case we can only choose 2
residuals freely, and the remaining residuals are then already determined. To

159
 15

10 line
1
2
Y

3
5 4

0
1 2 3 4
X

Figure 5.10: Different regression lines for different values of Y if X = 3.

prove this requires matrix algebra, but you can see it when you try it yourself.
The gist of it is that if you have a regression equation with both an intercept
and a slope, the degrees of freedom is equal to the number of data points (sample
size) minus 2: n − 2. Generalising this to linear models with K predictors:
n − K − 1.
Generally, these degrees of freedom based on the number of residuals that
could be freely chosen, given the constraints of the model, are termed residual
degrees of freedom. When using regression models, one usually only reports
these residual degrees of freedom. Later on in this book, we will see instances
where one also should use model degrees of freedom. For now, it suffices to know
what is meant by residual degrees of freedom.

5.6 Null-hypothesis testing with linear models

Often, data analysis is about finding an answer to the question whether there
is a relationship between two variables. In most cases, the question pertains to
the population: is there a relationship between variable Y and variable X in
the population? In many cases, one looks for a linear relationship between two
variables.
One common method to answer this question is to analyse a sample of data,
apply a linear model, and look at the slope. However, one then knows the slope
in the sample, but not the slope in the population. We have seen that the slope
in the sample can be very different from the slope in the population. Suppose
we find a slope of 1: does that mean there is a slope in the population or that
there is no slope in the population?
In inferential data analysis, one often works with two hypotheses: the null-
hypothesis and the alternative hypothesis. The null-hypothesis states that the

160
 300

200
count

100

0
−1.0 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
sample.slope

Figure 5.11: Distribution of the sample slope when the population slope is 0
and sample size equals 40.

population slope is equal to 0 and the alternative hypothesis states that there
is a slope that is different from 0. Remember that if the population slope is
equal to 0, that is saying that there is no linear relationship between X and
Y (that is, you cannot predict one variable on the basis of the other variable).
Therefore, the null-hypothesis states there is no linear relationship between X
and Y in the population. If there is a slope, whether positive or negative, is
the same as saying there is a linear relationship, so the alternative hypothesis
states that there is a linear relationship between X and Y in the population.
In formula form, we have

H0 : βslope = 0 (5.6)
HA : βslope 6= 0 (5.7)

The population slope, βslope , is either 0 or it is not. Our data analysis is

then aimed at determining which of these two hypotheses is true. Key is that
we do a thought experiment on the null-hypothesis: we wonder what would
happen if the population slope would be really 0. In our imagination we draw
many samples of a certain size, say 40 data points, and then determine the
slope for each sample. Earlier we learned that the many sample slopes would
form a histogram in the shape of a t-distribution with n − 2 = 38 degrees of
freedom. For example, suppose we would draw 1000 samples of size 40, then
the histogram of the 1000 slopes would look like depicted Figure 5.11.
From this histogram we see that all observed sample slopes are well between
-0.8 and 0.8. This gives us the information we need. Of course, we have only
one sample of data, and we don’t know anything about the population data.
But we do know that if the population slope is equal to 0, then it is very unlikely

161
to find a sample slope of say 1 or -1. Thus, with our sample slope of 1, we know
that this finding is very unlikely if we hold the null-hypothesis to be true. In
other words, if the population slope is equal to 0, it would be quite improbable
to find a sample slope of 1 or larger. Therefore, we regard the null-hypothesis to
be false, since it does not provide a good explanation of why we found a sample
slope of 1. In that case, we say that we reject the null-hypothesis.

5.7 p-values
A p-value is a probability. It represents the probability of observing certain
events, given that the null-hypothesis is true.
In the previous section we saw that if the population slope is 0, and we drew
1000 samples of size 40, we did not observe a sample slope of 1 or larger. In
other words, the frequency of observing a slope of 1 or larger was 0. If we would
draw more samples, we theoretically could observe a sample slope of 1, but the
probability that that happens for any new sample we can estimate at less than
1 in a 1000, so less than 0.001: p < 0.001.
This estimate of the p-value was based on 1000 randomly drawn samples
of size 40 and then looking at the frequency of certain values in that data set.
But there is a short-cut, for we know that the distribution of sample slopes
has a t-distribution if we standardise the sample slopes. Therefore we do not
have to take 1000 samples and estimate probabilities, but we can look at the
t-distribution directly, using tables online or in statistical packages.
Figure 5.12 shows the t-distribution that is the theoretical distribution corresponding
to the histogram in Figure 5.11. If the standard error is equal to 0.19, and the
hypothetical population slope is 0, then the t-statistic associated with a slope of
1−0
1 is equal to t = 0.19 = 5.26. With this value, we can look up in the tables, how
often such a value of 5.26 or larger occurs in a t-distribution with 38 degrees of
freedom. In the tables or using R, we find that the probability that this occurs
is 0.00000294.

1 - pt(5.26, df = 38)

## [1] 0.000002939069

So, the fact that the t-statistic has a t-distribution gives us the opportunity
to exactly determine certain probabilities, including the p-value.
Now let’s suppose we have only one sample of 40 bottles, and we find a slope
of 0.1 with a standard error of 0.19. Then this value of 0.1 is (0.1−0)/0.19 = 0.53
standard errors away from 0. Thus, the t-statistic is 0.53. We then look at the
t-distribution with 38 degrees of freedom, and see that such a t-value of 0.53 is
not very strange: it lies well within the middle 95% of the t-distribution (see
Figure 5.12).
Let’s determine the p-value again for this slope of 0.1: we determine the
probability that we obtain such a t-value of 0.53 or larger. Figure 5.13 shows

162
 0.4

0.3
density

0.2

0.1

0.0
−8 −7 −6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8
t

Figure 5.12: The histogram of 1000 sample slopes and its corresponding
theoretical t-distribution with 38 degrees of freedom. The vertical line represents
the t-value of 5.26.

the area under the curve for values of t that are larger than 0.53. This area
under the curve can be seen as a probability. The total area under the curve of
the t-distribution amounts to 1. If we know the area of the shaded part of the
total area, we can compute the probability of finding t-values larger than 0.53.
In tables online, in Appendix B, or available in statistical packages, we can
look up how large this area is. It turns out to be 0.3.

1 - pt(0.53, df = 38)

## [1] 0.2995977

So, if the population slope is equal to 0 and we draw an infinite number of

samples of size 40 and compute the sample slopes, then 30% of them will be
larger than our sample slope of 0.1. The proportion of the shaded area is what
we call a one-sided p-value. We call it one-sided, because we only look at one
side of the t-distribution: we only look at values that are larger than our t-value
of 0.53.
We conclude that a slope value of 0.1 is not that strange to find if the
population slope is 0. By the same token, it would also have been probable to
find a slope of -0.1, corresponding to a t-value of -0.53. Since the t-distribution
is symmetrical, the probability of finding a t-value of less than -0.53 is depicted
in Figure 5.14, and of course this probability is also 0.3.

pt(-0.53, df = 38)

## [1] 0.2995977

163
 0.4

0.3

0.2
y

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
t

Figure 5.13: Probability of a t-value larger than 0.53.

0.4

0.3

0.2
y

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
t

Figure 5.14: Probability of finding a t-value smaller than -0.53.

164
 0.4

0.3

0.2
y

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
t

Figure 5.15: The blue vertical line represents a t-value of 0.53. The shaded area
represents the two-sided p-value: the probability of obtaining a t-value smaller
than -0.53 or larger than 0.53.

Remember that the null-hypothesis is that the population slope is 0, and the
alternative hypothesis is that the population slope is not 0. We should therefore
conclude that if we find a very large positive or negative slope, large in the sense
of the number of standard errors away from 0, that the null-hypothesis is unlikely
to be true. Therefore, if we find a slope of 0.1 or -0.1, then we should determine
the probability of finding a t-value that is larger than 0.53 or smaller than -0.53.
This probability is depicted in Figure 5.15 and is equal to twice the one-side
p-value, 2 × 0.2995977 = 0.5991953.
This probability is called the two-sided p-value. This is the one that should
be used, since the alternative hypothesis is also two-sided: the population slope
can be positive or negative. The question now is: is a sample slope of 0.1 enough
evidence to reject the null-hypothesis? To determine that, we determine how
many standard errors away from 0 the sample slope is and we look up in tables
how often that happens. Thus in our case, we found a slope that is 0.53 standard
errors away from 0 and the tables told us that the probability of finding a slope
that is at least 0.53 standard errors away from 0 (positive or negative) is equal
to 0.5991953. We find this probability rather large, so we decide that we do not
reject the null-hypothesis.

5.8 Hypothesis testing

In the previous section, we found a one-sided p-value of 0.00000294 for a sample
slope of 1 and more or less concluded that this probability was rather small.
The two-sided p-value would be twice this value, so 0.00000588, which is still
very small. Next, we determined the p-value associated with a slope of 0.1 and

165
found a p-value of 0.60. This probability was rather large, and we decided to
not reject the null-hypothesis. In other words, the probability was so large that
we thought that the hypothesis that the population slope is 0 should not be
rejected based on our findings.

When should we think the p-value is small enough to conclude that the
null-hypothesis can be rejected? When can we conclude that the hypothesis
that the population slope is 0 is not supported by our sample data? This was
a question posed to the founding father of statistical hypothesis testing, Sir
Ronald Fischer. In his book Statistical Methods for Research Workers (1925),
Fisher proposed a probability of 5%. He advocated 5% as a standard level for
concluding that there is evidence against the null-hypothesis. However, he did
not see it as an absolute rule: ”If P is between .1 and .9 there is certainly no
reason to suspect the hypothesis tested. If it is below .02 it is strongly indicated
that the hypothesis fails to account for the whole of the facts. We shall not often
be astray if we draw a conventional line at .05...”. So Fisher saw the p-value
as an informal index to be used as a measure of discrepancy between the data
and the null-hypothesis: The null-hypothesis is never proved, but is possibly
disproved.

Later, Jerzy Neyman and Egon Pearson saw the p-value as an instrument
in decision making: is the null-hypothesis true, or is the alternative hypothesis
true? You either reject the null-hypothesis or you don’t, there is nothing in
between. A slightly milder view is that you either decide that there is enough
empirical evidence to reject the null-hypothesis, or there is not enough empirical
evidence to reject the null-hypothesis (not necessarily accepting H0 as true!).
This view to data-analysis is rather popular in the social and behavioural
sciences, but also in particle physics. In order to make such black-and-white
decisions, you decide before-hand, that is, before collecting data, what level of
significance you choose for your p-value to decide whether to reject the null-
hypothesis. For example, as your significance level, you might want to choose
1%. Let’s call this chosen significance level α. Then you collect your data, you
apply your linear model to the data, and find that the p-value associated with
the slope equals p. If this p is smaller than or equal to α, you reject the null-
hypothesis, and if p is larger than α then you do not reject the null-hypothesis.
A slope with a p ≤ α is said to be significant, and a slope with a p > α is
said to be non-significant. If the sample slope is significant, then one should
reject the null-hypothesis and say there is a slope in the population different
from zero. If the sample slope is not significant, then one should not reject the
null-hypothesis and say there is no slope in the population (i.e., the slope is 0).
Alternatively, one could say there is no empirical evidence for the existence of a
slope (this leaves the possibility that there is a slope in the population but that
our method of research failed to find evidence for it).

166
5.9 Inference for linear models in R
So far, we have focused on standard errors and confidence intervals for the slope
parameter in simple regression, that is, a linear model where there is only one
independent variable. However, the same logic can be applied to the intercept
parameter, and to other slope variables in case you have multiple independent
variables in your model (multiple regression).
For instance, suppose we are interested in the knowledge university students
have of mathematics. We start measuring their knowledge at time 0, when the
students start doing a bachelor programme in mathematics. At time 1 (after 1
year) and at time 2 (after two years), we also perform measures. Our dependent
variable is mathematical knowledge, a measure with possible values between 200
and 700. The independent variables are time (the time of measurement) and
distance: the distance in kilometers between university and their home. There
are two research questions. First question is about the level of knowledge when
students enter the bachelor programme, and the second question is how much
knowledge is acquired in one year of study. The linear model is as follows:

knowledge = b0 + b1 time + b2 distance + e (5.8)

e ∼ N (0, σ 2 )

Table 5.2: Regression table as obtained from R, with knowledge predicted by

time and distance.
term estimate std.error statistic p.value
(Intercept) 299.35 7.60 39.40 0.00
time 18.13 2.60 6.96 0.00
distance -0.76 0.67 -1.15 0.25

The first question could be answered by estimating the intercept b0 : that

is the level of knowledge we expect for a student at time 0 and with a home
0 kilometres from the university. The second question could be answered by
estimating the slope coefficient for time: the expected increase in knowledge
per year. In Chapter 4 we saw how to estimate the regression parameters in R.
We saw that we then get a regression table. For our mathematical knowledge
example, we could obtain the regression table, displayed in Table 5.2. We
discussed the first column with the regression parameters in Chapter 4. We see
that the intercept is estimated at 299.35, and the slopes for time and distance
are 18.13 and -0.76, respectively. So we can fill in the equation:

knowledge = 299.35 + 18.13time − 0.76distance + e (5.9)

Let’s look at the other columns in the regression table. In the second column
we see the standard errors for each parameter. The third column gives statistics;

167
these are the t-statistics for the null-hypotheses that the respective parameters
in the population are 0. For instance, the first statistic has the value 39.40. It
belongs to the intercept. If the null-hypothesis is that the population intercept
is 0 (β0 = 0), then the t-statistic is computed as

b0 − β 0 299.35 − 0 299.35
t= = = = 39.40 (5.10)
σβ̂ 7.60 7.60
You see that the t-statistic in the regression table is simply the regression
parameter divided by its standard error. This is also true for the slope parameters.
For instance, the t-statistic of 6.96 for time is simply the regression coefficient
18.13 divided by the standard error 2.60:

b1 − β 0 18.13 − 0 18.13
t= = = = 6.96 (5.11)
σβ̂ 7.60 2.60
The last column gives the two-sided p-values for the respective null-hypotheses.
For instance, the p-value of 0.00 for the intercept says that the probability of
finding an intercept of 299.35 or larger (plus or minus), under the assumption
that the population intercept is 0, is very small (less than 0.01).
If you want to have confidence intervals for the intercept and the slope for
time, you can use the information in the table to construct them yourself. For
instance, according to the table, the standard error for the intercept equals
7.60. Suppose the sample size equals 90 students, then you know that you have
n − K − 1 = 90 − 2 − 1 = 87 degrees of freedom. The critical value for a t-
statistic with 84 degrees of freedom for a 95% confidence interval can be looked
up in Appendix B. It must be somewhere between 1.98 and 2.00, so let’s use
1.99. The 95% interval for the intercept then runs between 299.35 − 1.99 × 7.60
and 299.35 − 1.99 × 7.60, so the expected level of knowledge at the start of
the bachelor programme for students living close to or on campus is somewhere
between from 284.23 to 314.47.
To show you how this can all be done using R, we have a look at the R dataset
called ”freeny” on quarterly revenues. We would like to predict the variable
market.potential by the predictors price.index and income.level. Apart
from the tidyverse package, we also need the broom package for the tidy()
function. When we run the following code, we obtain a regression table.

library(broom)
data("freeny")
out <- freeny %>%
lm(market.potential ~ price.index + income.level, data = .)
out %>%
tidy()

## # A tibble: 3 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>

168
## 1 (Intercept) 13.3 0.291 45.6 1.86e-33
## 2 price.index -0.309 0.0263 -11.8 6.92e-14
## 3 income.level 0.196 0.0291 6.74 7.20e- 8

We can have R compute the respective confidence intervals by indicating

that we want intervals of a certain confidence level, say 99%:

out <- freeny %>%

lm(market.potential ~ price.index + income.level, data = .)
out %>%
tidy(conf.int = TRUE, conf.level = 0.99)

## # A tibble: 3 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 13.3 0.291 45.6 1.86e-33 12.5 14.1
## 2 price.index -0.309 0.0263 -11.8 6.92e-14 -0.381 -0.238
## 3 income.level 0.196 0.0291 6.74 7.20e- 8 0.117 0.276

In the last two columns we see for example that the 99% confidence interval
for the price.index slope runs from -0.363 to -0.256.

5.10 Type I and Type II errors in decision making

Since data analysis is about probabilities, there is always a chance that you
make the wrong decision: you can wrongfully reject the null-hypothesis, or
you can wrongfully fail to reject the null-hypothesis. Pearson and Neyman
distinguished between two kinds of error: one could reject the null-hypothesis
while it is actually true (error of the first kind, or type I error) and one could
accept the null-hypothesis while it is not true (error of the second kind, or type
II error). We already discussed these types of error in Chapter 2. Table 5.3
gives an overview.

Table 5.3: Four different scenarios for hypothesis tests.

Test conclusion
do not reject H0 reject H0
H0 true OK Type I Error
Truth
HA true Type II Error OK

To illustrate the difference between type I and type II errors, let’s recall
the famous fable by Aesop about the boy who cried wolf. The tale concerns a
shepherd boy who repeatedly tricks other people into thinking a wolf is attacking
his flock of sheep. The first time he cries ”There is a wolf!”, the men working in
an adjoining field come to help him. But when they repeatedly find there is no

169
wolf to be seen, they realise they are being fooled by the boy. One day, when
a wolf does appear and the boy again calls for help, the men believe that it is
another false alarm and the sheep are eaten by the wolf.
In this fable, we can think of the null-hypothesis as the hypothesis that there
is no wolf. The alternative hypothesis is that there is a wolf. Now, when the boy
cries wolf the first time, there is in fact no wolf. The men from the adjoining
field make a type I error: they think there is a wolf while there isn’t. Later,
when they are fed up with the annoying shepherd boy, they don’t react when
the boy cries ”There is a wolf!”. Now they make a type II error: they think
there is no wolf, while there actually is a wolf. See Table 5.4 for the overview.

Men in the field

Think there is no wolf Think there is a wolf
There is no wolf OK waste of time and energy
Truth
There is a wolf devoured sheep OK

Table 5.4: Four different scenarios for wolves and men working in the field.

Let’s now discuss these errors in the context of linear models. Suppose you
want to determine the slope for the effect of age on height in children. Let the
slope now stand for the wolf: either there is no slope (no wolf, H0 ) or there is a
slope (wolf, HA ). The null-hypothesis is that the slope is 0 in the population of
all children (a slope of 0 means there is no slope) and the alternative hypothesis
that the slope is not 0, so there is a slope. You might study a sample of children
and you might find a certain slope. You might decide that if the p-value is below
a critical value you conclude that the null-hypothesis is not true. Suppose you
think a probability of 10% is small enough to reject the null-hypothesis as true.
In other words, if p ≤ 0.10 then we no longer think 0 is a reasonable value for
the population slope. In this case, we have fixed our α or type I error rate to be
α = 0.10. This means that if we study a random sample of children, we look at
the slope and find a p-value of 0.11, then we do not reject the null-hypothesis.
If we find a p-value of 0.10 or less, then we reject the null-hypothesis.
Note that the probability of a type I error is the same as our α for the
significance level. Suppose we set our α = 0.05. Then for any p-value equal or
smaller than 0.05, we reject the null-hypothesis. Suppose the null-hypothesis is
true, how often do we then find a p-value smaller than 0.05? We find a p-value
smaller than 0.05 if we find a t-value that is above a certain threshold. For
instance, for the t-distribution with 198 degrees of freedom, the critical value is
±1.97, because only in 5% of the cases we find a t-value of ±1.97 or more if the
null-hypothesis is true! Thus, if the null-hypothesis is true, we see a t-value of at
least ±1.97 in 5% of the cases. Therefore, we see a significant p-value in 5% of
the cases if the null-hypothesis is true. This is exactly the definition of a Type
I error: the probability that we reject the null-hypothesis (finding a significant
p-value), given that the null-hypothesis is true. So we call our α-value the type
I error rate.
Suppose 100 researchers are studying a particular slope. Unbeknownst to

170
them, the population slope is exactly 0. They each draw a random sample from
the population and test whether their sample slope is significantly different from
0. Suppose they all use different sample sizes, but they all use the same α of
0.05. Then we can expect that about 5 researchers will reject the null-hypothesis
(finding a p-value less than or smaller than 0.05) and about 95 will not reject
the null-hypothesis (finding a p-value of more than 0.05).
Fixing the type I error rate should always be done before data collection.
How willing are you to take a risk of a type I error? You are free to make a
choice about α, as long as you do it before looking at the data, and report what
value you used.
If α represents the probability of making a type I error, then we can use β to
represent the opposite: the probability of not rejecting the null-hypothesis while
it is not true (type II error, thinking there is no wolf while there is). However,
setting the β-value prior to data collection is a bit trickier than choosing your α.
It is not possible to compute the probability that we find a non-significant effect
(p > α), given that the alternative hypothesis is true, because the alternative
hypothesis is only saying that the slope is not equal to 0. In order to compute
β, we need to think first of a reasonable size of the slope that we expect. For
example, suppose we believe that a slope of 1 is quite reasonable, given what
we know about growth in children. Let that be our alternative hypothesis:

H0 : β 1 = 0
HA : β1 = 1

Next, we determine the distribution of sample slopes under the assumption

that the population slope is 1. We know that this distribution has a mean of 1
and a standard deviation equal to the standard error. We also know it has the
shape of a t-distribution. Let sample size be equal to 102 and the standard error
2. If we standardise the slopes by dividing by the standard error, we get the two
t-distributions in Figure 5.16: one distribution of t-values if the population slope
is 0 (centred around t = 0), and one distribution of t-values if the population
slope is 1 (centred around t = 1/2 = 0.5).
Let’s fix α to 10%. The shaded areas represent the area where p ≤ α: for
all values of t smaller than −1.6859545 and larger than 1.6859545, we reject the
null-hypothesis. The probability that this happens, if the null-hypothesis is true,
is equal to α, which is 0.10 in this example. The probability that this happens
if the alternative hypothesis is true (i.e., population slope is 1), is depicted in
Figure 5.17.
The shaded area in Figure 5.17 turns out to be 0.1415543. This represents
the probability that we find a significant effect, if the population slope is 1.
This is actually 1 minus the probability of finding a non-significant effect, if
the population slope is 1, which is defined as β. Therefore, the shaded area in
Figure 5.17 represents 1 − β: the probability of finding a significant p-value, if
the population slope is 1. In this example, 1 − β is equal to 0.1415543, so β is
equal to 1 − 0.1415543 = 0.8584457.

171
 0.4

0.3 null hypothesis alternative hypothesis

density

0.2

0.1

0.0
−1.66 1.66
t

Figure 5.16: Different t-distributions of the sample slope if the population slope
equals 0 (left curve in blue), and if the population slope equals 1 (right curve
in red). Blue area depicts the probability that we find a p-value value smaller
than 0.10 if the population slope is 0 (α).

0.4

0.3 null hypothesis alternative hypothesis

density

0.2

0.1

0.0
−1.66 1.66
t

Figure 5.17: Different t-distributions of the sample slope if the population slope
equals 0 (left curve in blue), and if the population slope equals 1 (right curve in
red). Shaded area depicts the probability that we find a p-value value smaller
than 0.10 if the population slope is 1 (1 − β).

172
 In sum, in this example with an α of 0.10 and assuming a population slope
of 1, we find that the probability of a type II error is 0.86: if there is a slope of
1, then we have an 86% chance of wrongly concluding that the slope is 0.
Type I and II error rates α and β are closely related. If we feel that a
significance level of α = 0.10 is too high, we could choose a level of 0.01. This
ensures that we are less likely to reject the null-hypothesis when it is true. The
critical value for our t-statistic is then equal to ±2.6258905, see Figure 5.18. In
Figure 5.19 we see that if we change α, we also get a different value for 1 − β,
in this case 0.0196567.
Table 5.5 gives an overview of how α and β are related to type I and type
II error rates. If a p-value for a statistical test is equal to or smaller than a
pre-chosen significance level α, the probability of a type I error equals α. The
probability of a type II error rate is equal to β.

Statistical outcome
p>α p≤α
H0 1−α α
Truth
HA β 1−β

Table 5.5: The probabilities of a statistical outcome under the null-hypothesis

and the alternative hypothesis.

Thus, if we use smaller values for α, we get smaller values for 1−β, so we get
larger values for β. This means that if we lower the probability of rejecting the
null-hypothesis given that it is true (type I error) by choosing a lower value for α,
we inadvertently increase the probability of failing to reject the null-hypothesis
given that it is not true (type II error).
Think again about the problem of the sheep and the wolf. Instead of the
boy, the men could choose to put a very nervous person on watch, someone
very scared of wolves. With the faintest hint of a wolf’s presence, the man will
call out ”Wolf!”. However, this will lead to many false alarms (type I errors),
but the men will be very sure that when there actually is a wolf, they will be
warned. Alternatively, they could choose to put a man on watch that is very
laid back, very relaxed, but perhaps prone to nod off. This will lower the risk of
false alarms immensely (no more type I errors) but it will dramatically increase
the risk of a type II error!
One should therefore always strike a balance between the two types of errors.
One should consider how bad it is to think that the slope is not 0 while it is,
and how bad it is to think that the slope is 0, while it is not. If you feel that
the first mistake is worse than the second one, then make sure α is really small,
and if you feel that the second mistake is worse, then make α not too small.
Another option, and a better one, to avoid type II errors, is to increase sample
size, as we will see in the next section.

173
 0.4

0.3 null hypothesis alternative hypothesis

density

0.2

0.1

0.0
−2.625891 2.625891
t

Figure 5.18: Different t-distributions of the sample slope if the population slope
equals 0 (left curve), and if the population slope equals 1 (right curve). Blue
area depicts the probability that we find a p-value value smaller than 0.01 if the
population slope is 0.

0.4

0.3 null hypothesis alternative hypothesis

density

0.2

0.1

0.0
−2.625891 2.625891
t

Figure 5.19: Different t-distributions of the sample slope if the population slope
equals 0 (left curve in blue), and if the population slope equals 1 (right curve in
red). Red area depicts the probability that we find a p-value value smaller than
0.01 if the population slope is 1: 1 − β.

174
5.11 Statistical power
Null-hypothesis testing only involves the null-hypothesis: we look at the sample
slope, compute the t-statistic and then see how often such a t-value and larger
values occur given that the population slope is 0. Then we look at the p-
value and if that p-value is smaller than or equal to α, we reject the null-
hypothesis. Therefore, null-hypothesis testing does not involve testing the
alternative hypothesis. We can decide what value we choose for our α, but
not our β. The β is dependent on what the actual population slope is, and we
simply don’t know that.
As stated in the previous section, we can compute β only if we have a more
specific idea of an alternative value for the population slope. We saw that we
needed to think of a reasonable value for the population slope that we might
be interested in. Suppose we have the intuition that a slope of 1 could well
be the case. Then, we would like to find a p-value of less than α if indeed the
slope were 1. We hope that the probability that this happens is very high: the
conditional probability that we find a t-value large enough to reject the null-
hypothesis, given that the population slope is 1. This probability is actually
the complement of β, 1 − β: the probability that we reject the null-hypothesis,
given that the alternative hypothesis is true. This 1 − β is often called the
statistical power of a null-hypothesis test. When we think again about the boy
who cried wolf: the power is the probability that the men think there is a wolf
if there is indeed a wolf. The power of a test should always be high: if there is
a population slope that is not 0, then of course you would like to detect it by
finding a significant t-value!
In order to get a large value for 1 − β, we should have large t-values in our
data-analysis. There are two ways in which we can increase the value of the
t-statistic. Since with null-hypothesis testing t = b−0 b
σb̂ = σb̂ , we can get large
values for t if 1) we have a small standard error, σb̂ , or 2) if we have a large
value for b.
Let’s first look at the first option: a small standard error. We get a small
standard error if we have a large sample size, see Section 5.2.1. If we go back to
the example of the previous section where we had a sample size of 102 children
and our alternative hypothesis was that the population slope was 1, we found
that the t-distribution for the alternative hypothesis was centred around 0.5,
because the standard error was 2. Suppose that we would increase sample size
to 1200 children, then our standard error might be 0.2. Then our t-distribution
for the alternative hypothesis is centred at 5. This is shown in Figure 5.20.
We see from the shaded area that if the population slope is really 1, there is
a very high chance that the t-value for the sample slope will be larger than 2.58,
the cut-off point for an α of 0.01 and 1198 degrees of freedom. The probability
of rejecting the null-hypothesis while it is not true, is therefore very large. This
is our 1 − β and we call this the power of the null-hypothesis test. We see that
with increasing sample size, the power to find a significant t-value increases too.
Now let us look at the second option, a large value of b. Sample slope b1

175
 0.4

density 0.3 null hypothesis alternative hypothesis

0.2

0.1

0.0
−2.579939 2.579939
t

Figure 5.20: Different t-distributions of the sample slope if the population slope
equals 0 (left curve in blue), and if the population slope equals 1 (right curve
in red). Now for a larger sample size. Shaded area depicts the probability that
we find a p-value value smaller than 0.01 if the population slope is 1.

depends of course on the population slope β1 . The power becomes larger when
the population slope is further away from zero. If the population slope were
10, and we only had a sample of 102 children (resulting in a standard error of
2), the t-distribution for the alternative hypothesis that the population slope
is centred around σb = 10/2 = 5, resulting in the same plot as in Figure 5.20,
b̂
with a large value for 1 − β. Unfortunately, the population slope is beyond our
control: the population slope is a given fact that we cannot change. The only
thing we can change most of the times is sample size.
In sum: the statistical power of a test is the probability that the null-
hypothesis is rejected, given that it is not true. This probability is equal to
1 − β. The statistical power of a test increases with sample size, and depends
on the actual population slope. The further away the population slope is from 0
(positive or negative), the larger the statistical power. Earlier we also saw that
1 − β increases with increasing α: the larger α, the higher the power.

5.12 Power analysis

Because of these relationships between statistical power, α, sample size n, and
the actual population slope β1 , we can compute the statistical power for any
combination thereof.
If you really care about the quality of your research, you carry out a power
analysis prior to collecting data. With such an analysis you can find out how
large your sample size should be. You can find many tools online that can help
you with that.

176
 Suppose you want to minimise the probability of a type I error, so you choose
an α = 0.01. Next, you think of what kind of population slope you would like
to find, if it indeed has that value. You could perhaps base this expectation
on earlier research. Suppose that you feel that if the population slope is 0.15,
you would really like to find a significant t-value so that you can reject the
null-hypothesis. Next, you have to specify how badly you want to reject the
null-hypothesis if indeed the population slope is 0.15. If the population slope
is really 0.15, then you would like to have a high probability to find a t-value
large enough to reject the null-hypothesis. This is of course the power of the
test, 1 − β. Let’s say you want to have a power of 0.90. Now you have enough
information to calculate how large your sample size should be.
Let’s look at G*power4 , an application that can be downloaded from the
web. If we start the app, we can ask for the sample size required for a slope of
0.15, an α of 0.01, and a power (1 − β) of 0.90. Let the standard deviation of
our dependent variable (Y ) be 3 and the standard deviation of our independent
variable (X) be 2. These numbers you can guess, preferably based on some other
data that were collected earlier. Then we get the input as displayed in Figure
5.21. Note that you should use two-sided p-values, so tails = two. From the
output we see that the required sample size is 1477 children.

5.13 Criticism on null-hypothesis testing and p-

values
The practice of null-hypothesis significance testing (NHST) is widespread. However,
from the beginning it has received much criticism. One of the first to criticise
the approach was the inventor of the p-value, Sir Ronald Fisher himself. Fisher
explicitly contrasted the use of the p-value for statistical inference in science
with the Pearson-Neyman approach, which he termed ”Acceptance Procedures”.
Whereas in the Pearson-Neyman approach the only relevance of the p-value
is whether it is smaller or larger than the fixed significance level α, Fisher
emphasised that the exact p-value should be reported to indicate the strength
of evidence against the null-hypothesis. He emphasised that no single p-value
can refute a hypothesis, since chance always allows for type I and type II
errors. Conclusions can and will be revised with further experimentation; science
requires more than one study to reach solid conclusions. Decision procedures
with clear-cut decisions based on one study alone hamper science and lead to
tunnel-vision.
Apart from these science-theoretical considerations of the NHST, there are
also practical reasons why pure NHST should be avoided. In at least a number
of research fields, the p-value has become more than just the criterion for
finding an effect or not: it has become the criterion of whether the research is
publishable or not. Editors and reviewers of scientific journals have increasingly
interpreted a study with a significant effect to be more interesting than a study
4 http://www.gpower.hhu.de/

177
 Figure 5.21: G*power output for a simple regression analysis.

with a non-significant effect. For that reason, in scientific journals you will find
mostly studies reported with a significant effect. This has led to the file-drawer
problem: the literature reports significant effects for a particular phenomenon,
but there can be many unpublished studies with non-significant effects for the
same phenomenon. These unpublished studies remain unseen in file-drawers
(or these days on hard-drives). So based on the literature there might seem to
exist a particular phenomenon, but if you would put all the results together,
including the unpublished studies, the effect might disappear completely.
Remember that if the null-hypothesis is true and everyone uses an α of 0.05,
then out of 100 studies of the same phenomenon, only 5 studies will be significant
and are likely to be published. The remaining 95 studies with insignificant effects
are more likely to remain invisible.
As a result of this bias in publication, scientists who want to publish their
results are tempted to fiddle around a bit more with their data in order to get
a significant result. Or, if they obtain a p-value of 0.07, they decide to increase

178
their sample size, and perhaps stop as soon as the p-value is 0.05 or less. This
horrible malpractice is called p-hacking and is extremely harmful to science. As
we saw earlier, if you want to find an effect and not miss it, you should carry
out a power analysis before you collect the data and make sure that your sample
size is large enough to obtain the power you want to have. Increasing sample
size after you have found a non-significant effect increases your type I error rate
dramatically: if you stop collecting data until you find a significant p-value, the
type I error rate is equal to 1!
There have been wide discussions the last few years about the use and
interpretation of p-values. In a formal statement, the American Statistical
Association published six principles that should be well understood by anyone,
including you, who uses them.
The six principles are:

1. p-values can indicate how incompatible the data are with a specified
statistical model (usually the null-hypothesis).

2. p-values do not measure the probability that the studied hypothesis is

true, or the probability that the data were produced by random chance
alone. Instead, they measure how likely it is to find a sample slope of at
least the size that you found, given that the population slope is 0.

3. Scientific conclusions and business or policy decisions should not be based

only on whether a p-value passes a specific threshold. For instance, also
look at the size of the effect: is the slope large enough to make policy
changes worth the effort? Have other studies found effects of similar sizes?

4. Proper inference requires full reporting and transparency. Always report

your sample slope, the standard error, the t-statistic, the degrees of freedom,
and the p-value. Only report about null-hypotheses that your study was
designed to test.

5. A p-value or statistical significance does not measure the size of an effect

or the importance of a result. (See principle 1)

6. By itself, a p-value does not provide a good measure of evidence regarding

a model or hypothesis. At least as important is the design of the study.

These six principles are further explained in the statement online5 . The
bottom line is, p-values have worth but only when used and interpreted in a
proper way. Some disagree. The philosopher of science William Rozeboom
once called NHST ”surely the most bone-headedly misguided procedure ever
institutionalized in the rote training of science students.” The scientific journal
Basic and Applied Social Psychology even banned NHST altogether: t-values
and p-values are not allowed if you want to publish your research in that journal.
5 https://amstat.tandfonline.com/doi/abs/10.1080/00031305.2016.1154108

179
 Most researchers now realise that reporting confidence intervals is often a
lot more meaningful than reporting whether a p-value is significant or not. A p-
value only says something about evidence against the hypothesis that the slope
is 0. In contrast, a confidence interval gives a whole range of reasonable values
for the population slope. If 0 lies within the confidence interval, then 0 is a
reasonable value; if it is not, then 0 is not a reasonable value so that we can
reject the null-hypothesis.
Using confidence intervals also counters one fundamental problem of null-
hypotheses: nobody believes in them! Remember that the null-hypothesis states
that a particular effect (a slope) is exactly 0: not 0.0000001, not -0.000201, but
exactly 0.000000000000000000000.
Sometimes a null-hypothesis doesn’t make sense at all. Suppose we are
interested to know what the relationship is between age and height in children.
Nobody believes that the population slope coefficient for the regression of height
on age is 0. Why then test this hypothesis? More interesting would be to know
how large the population slope is. A confidence interval would then be much
more informative than a simple rejection of the null-hypothesis.
In some cases, a null-hypothesis can be slightly more meaningful: suppose
you are interested in the effect of cognitive behavioural therapy on depression.
You hope that the number of therapy sessions has a negative effect on the
severity of the depression, but it is entirely possible that the effect is very close to
non-existing. Of course you can only look at a sample of patients and determine
the sample slope. But think now about the population slope: think about all
patients in the world with depression that theoretically could partake in the
research. Some of them have 0 sessions, some have 1 session, and so on. Now
imagine that there are 1 million of such people. How likely is it that in the
population, the slope for the regression is exactly 0? Not 0.00000001, not -
0.0000000002, but exactly 0.0000000000. Of course, this is extremely unlikely.
The really interesting question in such research is whether there is a meaningful
effect of therapy. For instance, an effect of at least half a point decrease on the
Hamilton depression scale for 5 sessions. That would translate to a slope of
−0.5
5 = −0.1. Also in this case, a confidence interval for the effect of therapy on
depression would be more helpful than a simple p-value. A confidence interval
of -2.30 to -0.01 says that a small population effect of -0.01 might be there,
but that an effect of -0.0001 or 0.0000 is rather unlikely. It also states that
a meaningful effect of at least -0.1 is likely. You can then conclude that the
therapy is helpful. The p-value less than α only tells you that a value of exactly
0.0000 is not realistic, but who cares.
So, instead of asking research questions like ”Is there a linear relationship
between x and y?” you might ask: ”How large is the linear effect of x on y?”
Instead of a question like ”Is there an effect of the intervention?” it might be
more interesting to ask: ”How large is the effect of the intervention?”
Summarising, remember the following principles when doing your own research
or evaluating the research done by others:

ˆ Inference about a population slope or intercept can be made on the basis

180
 of sample data, but only in probabilistic terms. This means that a simple
statement like ”the value of the population slope is definitely not zero”
cannot be made. Only statements like ”A population slope of 0 is not very
likely given the sample data” can be made.

ˆ Science is cumulative. No study is definitive. Effects should be replicated

by independent researchers.

ˆ Always report your regression slope or intercept, with the standard error
and the sample size. Based on these, the t-statistics can be computed with
the degrees of freedom. Then if several other researchers have done the
same type of research, the results can be combined in a so-called meta-
analysis, so that a stronger statement about the population can be made,
based on a larger total sample size. The standard error and sample size
moreover allow for the construction of confidence intervals. But better is
to report confidence intervals yourself.

5.14 Relationship between p-values and confidence

intervals
In previous sections we stated that if the value 0 lies within a confidence interval,
it is a reasonable value for the population slope. If 0 is not within the interval,
0 is not a reasonable value for the population slope, so we have to reject the
null-hypothesis. Here we will elaborate a little on this theme.
Both the confidence interval and the p-value are based on the same t-
distribution. Suppose we set our α to 0.05, and our sample size is 102. This
means that if we find a p-value p ≤ 0.05 we reject the null-hypothesis that the
slope is 0. The p-value depends on how many standard deviations our sample
slope deviates from 0. We calculate this by computing a standardised slope. For
example, for a sample slope of 1 and a standard error of 0.5, our standardised
slope is t = (1 − 0)/0.5 = 2. In other words, our sample slope of 1 is 2 standard
errors away from 0. From t-tables, we know that with 100 degrees of freedom,
the 2.5th and 97.5th percentiles are -1.98 and 1.98, respectively (see Appendix
B). Therefore, the p-value depends on the size of the t-statistic. If it is equal to
-1.98 or 1.98, the p-value is exactly 0.05. If the t-statistic is smaller than -1.98
or larger than 1.98, the p-value is smaller than 0.05.
The values -1.98 and 1.98 are also used for the construction of the 95%
confidence interval. The lower bound lies at 1.98 times the standard error below
the sample slope, and the upper bound lies at 1.98 times above the sample slope.
Therefore, if 0 lies more than 1.98 standard errors away from the mean, it lies
outside the confidence interval. But if 0 lies more than 1.98 standard errors
away from the mean, this implies that the sample slope lies more than -1.98
standard errors away from 0, which corresponds to a t-statistic of more than
±1.98. Thus, if 0 is not within the 95% confidence interval, we know that the
p-value is smaller than 0.05.

181
 Using the same reasoning as above, we also know that if 0 is not within the
99% confidence interval, we know that the p-value is smaller than 0.01, and if 0
is not within the 99.9% confidence interval, we know that the p-value is smaller
than 0.001, etcetera.
A 95% confidence interval can therefore also be seen as the range of possible
values for the null-hypothesis that cannot be rejected with an α of 5%. By the
same token, a 99% confidence interval can be seen as the range of possible values
for the null-hypothesis that cannot be rejected with an α of 1%, etcetera.

5.15 The intercept only model

So far in this chapter, we have only discussed inference regarding the linear
model with both an intercept and one or more slope parameters.

Y = b0 + b1 X1 + b2 X2 + · · · + e (5.12)
Remember that in Chapter 2 we discussed inference regarding only a mean.
Here we show that inference regarding the mean can also be done within the
linear model framework. In Chapter 2 we wanted to get a confidence interval
for the mean luteinising hormone (LH) for a woman. We had 48 measures
(n q= 48) and the sample mean was 2.4. We computed the standard error
2
s
as n = 0.0796, so that we could construct a confidence interval using a t-
distribution of 48 − 1 = 47 degrees of freedom. In Chapter 2 we saw that we
can compute a 95% confidence interval for a population mean as

t.test(lh, conf.level = 0.95)$conf.int

## [1] 2.239834 2.560166

## attr(,"conf.level")
## [1] 0.95

Here we show that the same inference can be done with a very simple version
of the linear model: an intercept-only model. An intercept-only model has only
an intercept and no slopes.

Y = b0 + e (5.13)
e ∼ N (0, σ 2 )

We briefly discussed such a model when we discussed degrees of freedom.

This model says that the predicted/expected Y -value for any observation, is
equal to b0 , with residuals e that are normally distributed. On average they are
0, and that implies that their sum is equal to 0.
We know that when we have a bunch of Y -values, and we compute the mean,
the deviations between the Y -values and the mean also sum to 0. As a very
simple example, if we observe the Y -values 4, 5 and 6, the mean is 5. When we

182
take the deviations between this mean of 5 and the Y -values, we get -1, 0 and
1. And these sum to 0. This is true for any set of Y -values. Thus, we could use
the mean of Y as our estimate for b0 , since then the deviations with the mean
(i.e., the residuals) sum to 0.
Earlier we said that the unbiased estimator of the population mean is the
sample mean. Therefore, our b0 parameter represents the unbiased estimator of
the population mean of Y . Let’s see if this works by fitting this model in R. In
R, an intercept is indicated by a 1:

library(broom)
data(lh)
out <- lh %>%
lm(lh ~ 1, data = .)
out %>%
tidy(conf.int = TRUE)

## # A tibble: 1 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 2.4 0.0796 30.1 2.14e-32 2.24 2.56

In the output we only see an intercept. It is equal to 2.4, which is also the
mean of LH as we saw earlier. The standard error is also exactly the same as
we computed by hand (0.0796), as is the 95% confidence interval. We get the
same results, because in both cases, we use exactly the same standard error and
the same t-distribution with n − K − 1 = 48 − 0 − 1 = 47 degrees of freedom
(K equals 0, the number of independent variables).
In summary, inference about the mean of a set of values can be done using
an intercept-only linear model.

183
184
Chapter 6

Categorical predictor
variables

6.1 Dummy coding

As we have seen in Chapter 1, there are largely two different types of variables:
numeric variables and categorical variables. Numeric variables say something
about how much of an attribute is in an object: for instance height (measured
in inches) or heat (measured in degrees Celsius). Categorical variables say
something about the quality of an attribute: for instance colour (red, green,
yellow) or type of seating (aisle seat, window seat). We have also seen a
third type of variable: ordinal variables. Ordinal variables are somewhat in
the middle between numeric variables and categorical variables: they are about
quantitative differences between objects (e.g., size) but the values are sharp
disjoint categories (small, medium, large), and the values are not expressed
using units of measurements.
In the chapters on simple and multiple regression we saw that both the
dependent and the independent variables were all numeric. The linear model
used in regression analysis always involves a numeric dependent variable. However,
in such analyses it is possible to use categorical independent variables. In this
chapter we explain how to do that and how to interpret the results.
The basic trick that we need is dummy coding. Dummy coding involves
making one or more new variables, that reflects the categorisation seen with
a categorical variable. First we focus on categorical variables with only two
categories (dichotomous variables). Later in this chapter, we will explain what
to do with categorical variables with more than two categories (nominal variables).
Imagine we study bus companies and there are two different types of seating
in buses: aisle seats and window seats. Suppose we ask 5 people, who have
travelled from Amsterdam to Paris by bus during the last 12 months, whether
they had an aisle seat or a window seat during their last trip, and how much
they paid for the trip. Suppose we have the variables person, seat and price.

185
Table 6.1 shows the anonymised data. There we see the dichotomous variable
seat with values ’aisle’ and ’window’.

Table 6.1: Bus trips to Paris.

person seat price
001 aisle 57.00
002 aisle 59.00
003 window 68.00
004 window 60.00
005 aisle 61.00

With dummy coding, we make a new variable that only has values 0 and 1,
that conveys the same information as the seat variable. The resulting variable
is called a dummy variable. Let’s call this dummy variable window and give it
the value 1 for all persons that travelled in a window seat. We give the value 0
for all persons that travelled in an aisle seat. We can also call the new variable
window a boolean variable with TRUE and FALSE, since in computer science,
TRUE is coded by a 1 and FALSE by a 0. Another name that is sometimes
used is an indicator variable. Whatever you want to call it, the data matrix
including the new variable is displayed in Table 6.2.

Table 6.2: Bus trips to Paris.

person seat window price
001 aisle 0 57.00
002 aisle 0 59.00
003 window 1 68.00
004 window 1 60.00
005 aisle 0 61.00

What we have done now is coding the old categorical variable seat into a
variable window with values 0 and 1 that looks numeric. Let’s see what happens
if we use a linear model for the variables price (dependent variable) and window
(independent variable). The linear model is:

price = b0 + b1 window + e (6.1)

e ∼ N (0, σe2 ) (6.2)

Let’s use the bus trip data and determine the least squares regression line.
We find the following linear equation:

d = 59 + 5 × window
price (6.3)
If the variable window has the value 1, then the expected or predicted price
of the bus ticket is, according to this equation, 59 + 5 × 1 = 64. What does

186
 68
67
66
65
64
63
price

62
61
60
59
58
57
0.00 0.25 0.50 0.75 1.00
window

Figure 6.1: Relationship between dummy variable window and price.

this mean? Well, all persons who had a window seat also had a value of 1
for the window variable. Therefore the expected price of a window seat equals
64. By the same token, the expected price of an aisle seat (window = 0) is
59 + 5 × 0 = 59, since all those with an aisle seat scored 0 on the window
variable.
You see that by coding a categorical variable into a numeric dummy variable,
we can describe the ’linear’ relationship between the type of seat and the price
of the ticket. Figure 6.1 shows the relationship between the numeric variable
window and the numeric variable price.
Note that the blue regression line goes straight through the mean of the prices
for window seats (window = 1) and the mean of the prices for aisle seats (window
= 0). In other words, the linear model with the dummy variable actually models
the group means of people with window seats and people with aisle seats.
Figure 6.2 shows the same regression line but now for the original variable
seat. Although the analysis was based on the dummy variable window, it is
more readable for others to show the original categorical variable seat.

6.2 Using regression to describe group means

In the previous section we saw that if we replace a categorical variable with
a numeric dummy variable with values 0 and 1, we can use a linear model
to describe the relationship between a categorical independent variable and a
numeric dependent variable. We also saw that if we take the least squares
regression line, this line goes straight through the averages, the group means.
The line goes straight through the group means because then the sum of the
squared residuals is at its smallest value (the least squares principle). Have
a look at the bus trip data again in Figure 6.1 and see if you can derive the

187
 68
67
66
65
64
63
price

62
61
60
59
58
57
aisle window
seat

Figure 6.2: Relationship between type of seat and price.

residuals and the squared residuals. These are displayed in Table 6.3.

Table 6.3: Bus trip to Paris data, together with residuals and squared residuals
from the least squares regression line.
person seat window price e e squared
001 aisle 0.00 57.00 -2.00 4.00
002 aisle 0.00 59.00 0.00 0.00
003 window 1.00 68.00 4.00 16.00
004 window 1.00 60.00 -4.00 16.00
005 aisle 0.00 61.00 2.00 4.00

If we take the sum of the squared residuals we obtain 40. Now if we use
a slightly different slope, so that we no longer go straight through the average
prices for aisle and window seats (see Figure 6.3) and we compute the predicted
values, the residuals and the squared residuals (see Table 6.4), we obtain a
higher sum: 40.05.
Only the least squares regression line goes through the observed average
prices of aisle seats and window seats. Thus, we can use the least squares
regression equation to describe observed group means for categorical variables.
Conversely, when you know the group means, it is very easy to draw the
regression line: the intercept is then the mean for the category coded as 0, and
the slope is equal to the mean of the category coded as 1 minus the mean of the
category coded as 0 (i.e., the intercept). Check Figure 6.1 to verify this yourself.
But we can also show this for a new data set.
We look at results from an experiment to compare yields (as measured by
dried weight of plants) obtained under a control and two different treatment
conditions. Let’s plot the data first, where we only compare the two experimental

188
 68
67
66
65
64
63
price

62
61
60
59
58
57
aisle window
seat

Figure 6.3: Relation between type of seat and price, with the regression line
being not quite the least squares line.

Table 6.4: Bus trips to Paris, together with residuals and squared residuals from
a suboptimal regression line.
person seat window price wrongpredict e e squared
001 aisle 0.00 57.00 59.10 -2.10 4.41
002 aisle 0.00 59.00 59.10 -0.10 0.01
003 window 1.00 68.00 63.90 4.10 16.81
004 window 1.00 60.00 63.90 -3.90 15.21
005 aisle 0.00 61.00 59.10 1.90 3.61

conditions (see Figure 6.4).

With treatment 1, the average yield turns out to be 4.661, and with treatment
2, the average yield is 5.526. Suppose we make a new dummy variable treatment2
that is 0 for treatment 1, and 1 for treatment 2. Then we have the linear
equation:

d = b0 + b1 × treatment2
weight (6.4)

If we fill in the dummy variable and the expected weights (the means!), then
we have the linear equations:

4.661 = b0 + b1 × 0 = b0 (6.5)
5.526 = b0 + b1 × 1 = b0 + b1 (6.6)

So from this, we know that intercept b0 = 4.661, and if we fill that in for the

189
 6.0

5.5
weight

5.0

4.5

4.0

3.5
trt1 trt2
group

Figure 6.4: Data on yield under two experimental conditions: treatment 1 and
treatment 2.

second equation above, we get the slope:

b1 = 5.526 − b0 = 5.526 − 4.661 = 0.865. (6.7)

Thus, we get the linear equation

d = 4.661 + 0.865 × treatment

weight (6.8)

Since this regression line goes straight through the average yield for each
treatment, we know that this is the least squares regression equation. We could
have obtained the exact same result with a regression analysis using statistical
software. But this was not necessary: because we knew the group means, we
could find the intercept and the slope ourselves by doing the math.
The interesting thing about a dummy variable is that the slope of the
regression line is exactly equal to the differences between the two averages.
If we look at Equation 6.8, we see that the slope coefficient is 0.865 and this is
exactly equal to the difference in mean weight for treatment 1 and treatment
2. Thus, the slope coefficient for a dummy variable indicates how much the
average of the treatment that is coded as 1 differs from the treatment that is
coded as 0. Here the slope is positive so that we know that the treatment coded
as 1 (trt2), leads to a higher average yield than the treatment coded as 0 (trt1).
This makes it possible to draw inferences about differences in group means.

190
6.3 Making inferences about differences in group
means
In the previous section we saw that the slope in a dummy regression is equal to
the difference in group means. Suppose researchers are interested in the effects
of different treatments on yield. They’d like to know what the difference is in
yield between treatments 1 and 2, using a limited sample of 20 data points.
Based on this sample, they’d like to generalise to the population of all yields
based on treatments 1 and 2. They adopt a type I error rate of α = 0.05.

Table 6.5: Yield by treatment.

term estimate std.error statistic p.value conf.low conf.high
(Intercept) 4.66 0.20 22.94 0.00 4.23 5.09
grouptrt2 0.87 0.29 3.01 0.01 0.26 1.47

The researchers analyse the data and they find the regression table as displayed
in Table 6.5. The 95% confidence interval for the slope is from 0.26 to 1.47.
This means that reasonable values for the population difference between the
two treatments on yield lie within this interval. All these values are positive, so
we reasonably believe that treatment 2 leads to a higher yield than treatment 1.
We know that it is treatment 2 that leads to a higher yield, because the slope
in the regression equation refers to a variable grouptrt2 (see Table 6.5). Thus,
a dummy variable has been created, grouptrt2, where trt2 has been coded as
1 (and trt1 consequently coded as 0). In the next section, we will see how to do
this yourself.
If the researchers had been interested in testing a null-hypothesis about the
differences in mean yield between treatment 1 and 2, they could also use the
95% confidence interval for the slope. As it does not contain 0, we can reject
the null-hypothesis that there is no difference in group means at an α of 5%.
The exact p-value can be read from Table 6.5 and is equal to 0.01.
Thus, based on this regression analysis the researchers can write in a report
that there is a significant difference between the yield after treatment 1 and the
yield after treatment 2, t(18) = 3.01, p = 0.01. Treatment 2 leads to a yield of
about 0.87 (SE = 0.29) more than treatment 1 (95% CI: 0.26 – 1.47).

6.4 Regression analysis using a dummy variable

in R
When your independent variable is a categorical variable, the code that you use
in R is the same as with a numeric independent variable. For instance, if you
want to predict yield from the treatment group, you could run the following R
code:
data (" Plantgrowth ")

191
PlantGrowth % >%
filter ( group != " ctrl ") % >%
lm ( weight ∼ group , .) % >%
tidy ()
In this code, we take the PlantGrowth data frame that is available in R, we
omit the data points from the control group (because we are only interested in
the two treatment groups), and we model weight as a function of group. What
then happens depends on the data type of group. Let’s take a quick look at
the variables:

PlantGrowth %>%
select(weight, group) %>%
str()

## ’data.frame’: 30 obs. of 2 variables:

## $ weight: num 4.17 5.58 5.18 6.11 4.5 4.61 5.17 4.53 5.33 5.14 ...
## $ group : Factor w/ 3 levels "ctrl","trt1",..: 1 1 1 1 1 1 1 1 1 1 ...

We see that the dependent variable weight is of type numeric (num), and that
the independent variable group is of type factor. If the independent variable
is of type factor, R will automatically make a dummy variable for the factor
variable. This will not happen if the independent variable is of type numeric.
So here group is a factor variable. Below we see the regression table that
results from the linear model analysis.

data("PlantGrowth")
out <- PlantGrowth %>%
filter(group != "ctrl") %>%
lm(weight ~ group, .)
out %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 4.66 0.203 22.9 8.93e-15
## 2 grouptrt2 0.865 0.287 3.01 7.52e- 3

We no longer see the group variable, but we see a new variable called
grouptrt2. Apparently, this new variable was created by R to deal with the
group variable being a factor variable. The slope value of 0.865 now refers to
the effect of treatment 2, that is, treatment 1 is the reference category and the
value 0.865 is the added effect of treatment 2 on the yield. We should therefore
interpret these results as that in the sample data, the mean of the treatment 2
data points was 0.865 higher than the mean of the treatment 1 data points.

192
 Here, R automatically picked the treatment 1 group as the reference group.
In case you want to have treatment 2 as the reference group, you could make your
own dummy variable. For instance, make your own dummy variable grouptrt1
in the following way and check whether it is indeed stored as numeric in R:

PlantGrowth <- PlantGrowth %>%

mutate(grouptrt1 = ifelse(group == "trt1", 1, 0))

PlantGrowth %>%
select(weight, group, grouptrt1) %>%
str()

## ’data.frame’: 30 obs. of 3 variables:

## $ weight : num 4.17 5.58 5.18 6.11 4.5 4.61 5.17 4.53 5.33 5.14 ...
## $ group : Factor w/ 3 levels "ctrl","trt1",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ grouptrt1: num 0 0 0 0 0 0 0 0 0 0 ...

Next, you can run a linear model with the grouptrt1 dummy variable that
you created yourself:

out <- PlantGrowth %>%

filter(group != "ctrl") %>%
lm(weight ~ grouptrt1, .)
out %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 5.53 0.203 27.2 4.52e-16
## 2 grouptrt1 -0.865 0.287 -3.01 7.52e- 3

The results now show the effect of treatment 1, with treatment 2 being the
reference category. Of course the effect of -0.865 is now the opposite of the effect
that we saw earlier (+0.865), when the reference category was treatment 2. The
intercept has also changed, as the intercept is now the expected weight for the
other treatment group. In other words, the reference group has now changed:
the intercept is equal to the expected weight of the treatment 1 group.
In general, store variables that are essentially categorical as factor variables
in R. For instance, you could have a variable group that has two values 1 and
2 and that is stored as numeric. It would make more sense to first turn this
variable into a factor variable, before using this variable as a predictor in a
linear model. You could turn the variable into a factor only for the analysis and
leaving the data frame unchanged, like this:
model < - dataset % >%
lm ( y ∼ factor ( group ) , data = .)

193
 or change the data frame before the analysis
dataset < - dataset % >%
mutate ( group = factor ( group ))

model < - dataset % >%

lm ( y ∼ group , data = .)
R will always choose the category with the lowest internal integer value as
the reference category. The internal values are chosen alphabetically. If that
however makes the output hard to interpret, think of the best way to code your
own dummy variable. For experimental designs, it makes sense to code control
conditions as 0, and experimental conditions as 1 (you’re often interested in
the effect of the experimental condition compared to the control condition, so
the control condition is the reference group). For social surveys, if you want to
compare how a social minority group scores relative to a social majority group,
it makes sense to code the minority group as 1 and the social majority as 0.
In educational studies, it makes sense to code an old teaching method as 0 and
a new method as 1. In all of these cases, the slope can then be interpreted as
the difference of the experimental procedure/new method/minority compared
to the reference group, and the intercept can be interpreted as the mean of the
reference group.

6.5 Two independent variables: one dummy and

one numeric variable
In Chapter 4 we saw that we can have more than one predictor variable in a
linear model. If we have two or more numeric variables, one usually talks about
multiple regression models. When we have a categorical variable that we treat
as a numeric dummy variable, then we can therefore also have linear models
with both a categorical variable and a numeric variable.
Let’s return to the bus trip to Paris data. Suppose that one can choose the
amount of leg room. There are seats with 60, 70 or 80 centimetres of leg room.
You might expect that seats with more leg room are more expensive. To find
out whether this is the case, you analyse the sample data from the 5 travellers.
Leg room varies between 60 and 80 centimetres. Since you already know that
the price of seats also depends on the type of seat (aisle or window), you analyse
the data with the following linear model:

price = b0 + b1 window + b2 legroom + e (6.9)

2
e ∼ N (0, σ )

That is, your independent variables are the dummy variable window (1
coding for window seat, 0 coding for aisle seat) and the numeric variable legroom.

194
Table 6.6: Regression table for the regression of price on the dummy variable
window and the numeric variable legroom.
term estimate std.error statistic p.value
(Intercept) 41.50 9.71 4.27 0.05
window 5.00 2.50 2.00 0.18
leg room 0.25 0.14 1.83 0.21

When we look at the output, we see the regression table in Table 6.6. When
we fill in the coefficients, we obtain the following linear equation:

d = 41.5 + 5 × window + 0.25 × legroom

price (6.10)
From Chapter 4 we know how to interpret the coefficients. The slope
coefficient for window should be interpreted as ”the increase in price if we change
seat from aisle to window, given a certain amount of legroom”. That is, holding
legroom constant, for instance at 60 centimetres, the difference between an
aisle and a window seat equals 5 Euros. And of course, this difference is also
5 Euros when legroom equals 70 centimetres, and also when legroom equals 80
centimetres. Along the same vein, the slope coefficient for legroom should be
interpreted as ”the increase in price for every unit increase in legroom, given a
certain type of seat”. Thus, for an aisle seat, you pay 0.25 Euros more for every
extra centimetre. And this is also true for window seats: every extra centimetre
for your legs costs you 0.25 Euros.
The intercept of 41.5 means that the model predicts that you pay 41.5 if
you happen to have 0 centrimeters of leg room and an aisle seat (the reference
category). Of course, a seat with 0 leg room does not exist, but it is simply
what the model predicts, based on the data that are observed. The data and
these predictions are visualised in Figure 6.5.
In order to visualise the relationship between the three variables price,
legroom and window, we plotted legroom on the horizontal axis, and used
different colours for the variable window. There are a couple of things you
should notice in this figure.
The first you should notice is that there are now two regression lines, one
for window seats and one for aisle seats. This is so because the model makes
different predictions for window and aisle seats. If we take Equation 6.10 and
make predictions for aisle seats, we fill in window = 0 and we get the following
linear equation:

d = 41.5 + 5 × 0 + 0.25 × legroom = 41.5 + 0.25 × legroom

price (6.11)
Thus, the regression line for aisle seats has an intercept of 41.5 and a slope
of 0.25. Now let’s fill in the equation for window seats, that is, window = 1.
Then we obtain

d = 41.5 + 5 × 1 + 0.25 × legroom = 46.5 + 0.25 × legroom

price (6.12)

195
 68

64
window
price

0
1

60

56
60 65 70 75 80
Leg room in centimetres

Figure 6.5: The bus trip to Paris data, with the predictions from a linear model
with legroom and window as independent variables.

That is, the regression line for window seats has an intercept that is different:
it is equal to the original intercept plus the slope of the window variable, 41.5 +
5 = 46.5. On the other hand, the slope for legroom is unchanged. With the
same slope for legroom, the two regression lines are therefore parallel.

The second you should notice from Figure 6.5 is that these two regression
lines are not the least squares regression lines for window and aisle seats respectively.
For instance, the regression line for window seats (the top one) should be more
positive in order to minimise the difference between the data points and the
regression line (the residuals). On the other hand, the regression line for aisle
seats (the bottom one) should be less steep in order to have smaller residuals.
Why is this so? Shouldn’t the regression lines minimise the residuals?

Yes they should! But there is a problem, because the model also implies, as
we saw above, that the lines are parallel. Whatever we choose for values for the
multiple regression equation, the regresion lines for aisle and window seats will
always be parallel. And under that constraint, the current parameter values
give the lowest possible value for the sum of the squared residuals, that is, the
the sum of the squared residuals for both regression lines taken together. The
aisle seat regression line should be less steep, and the window seat regression
line should be steeper to have a better fit with the data, but taken together, the
estimated slope of 0.25 gives the lowest overall sum of squared residuals.

It is possible though to have linear models where the lines are not parallel.
This will be discussed in Chapter 9.

196
 Table 6.7: Height across three different countries.
ID Country height
001 A 120
002 A 160
003 B 121
004 B 125
005 C 140
... ... ...

Table 6.8: Height across three different countries with dummy variables.
ID Country height countryA countryB
001 A 120 1 0
002 A 160 1 0
003 B 121 0 1
004 B 125 0 1
005 C 140 0 0
... ... ... ... ...

6.6 Dummy coding for more than two groups

In the previous sections we saw how to code a categorical variable with 2
categories (a dichotomous variable) into 1 dummy variable. In this section,
we learn how to code a categorical variable with 3 categories into 2 dummy
variables, and to code a categorical variable with 4 categories into 3 dummy
variables, etcetera. That is, how to code nominal variables into sets of dummy
variables.
Take for instance a variable Country, where in your data set, there are three
different values for this variable, for instance, Norway, Sweden and Finland,
or Zimbabwe, Congo and South-Africa. Let’s call these countries A, B and C.
Table 6.7 shows a data example where we see height measurements on people
from three different countries.
We can code this Country variable with three categories into two dummy
variables in the following way. First, we create a variable countryA. This is a
dummy variable, or indicator variable, that indicates whether a person comes
from country A or not. Those persons that do are coded 1, and those that do
not are coded 0. Next, we create a dummy variable countryB that indicates
whether or not a person comes from country B. Again, persons that do are coded
1, and those that do not are coded 0. The resulting variables are displayed in
Table 6.8
Note that we have now for every value of Country (A, B, or C) a unique
combination of the variables countryA and countryB. All persons from country
A have a 1 for countryA and a 0 for countryB; all those from country B have a
0 for countryA and a 1 for countryB, and all those from country C have a 0 for

197
countryA and a 0 for countryB. Therefore a third dummy variable countryC
is not necessary (i.e., is redundant): the two dummy variables give us all the
country information we need.
Remember that with two categories, you only need one dummy variable,
where one category gets 1s and another category gets 0s. In this way both
categories are uniquely identified. Here with three categories we also have unique
codes for every category. Similarly, if you have 4 categories, you can code this
with 3 dummy variables. In general, when you have a variable with K categories,
you can code them with K − 1 dummy variables.

6.7 Analysing categorical predictor variables in

R
R contains a data set (PlantGrowth) on yield on a sample of thirty observations
(n = 30), under three different conditions. We already saw part of the data in
Figure 6.4. The complete data consists of weight in three different groups:
treatment 1 (trt1), treatment 2 (trt2) and a control group (ctrl). Now we
want to model weight as a function of the categorical variable group using
a linear model. Again, we discuss two options how to do that in R. First by
creating your own dummy variables, second by letting R create dummy variables
automatically.

6.7.1 Creating your own dummy variables

Suppose you want to compare treatments 1 and 2 to your control condition.
Then it makes most sense to create two dummy variables, that code for the
treatment 1 group and the treatment 2 group, respectively. Thus, we create a
new variable treatment 1 and code every specimen that belongs to treatment
1 as 1, and all other specimens as 0. Next, we create a new variable treatment 2
and code every specimen that belongs to treatment 2 as 1, and all other specimens
as 0. The code is as follows:

PlantGrowth <- PlantGrowth %>%

mutate(treatment_1 = ifelse(group == "trt1", 1, 0),
treatment_2 = ifelse(group == "trt2", 1, 0))

Next, we use these new dummy variables in a multiple regression analysis:

PlantGrowth %>%
lm(weight ~ treatment_1 + treatment_2, data = .) %>%
tidy()

## # A tibble: 3 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>

198
## 1 (Intercept) 5.03 0.197 25.5 1.94e-20
## 2 treatment_1 -0.371 0.279 -1.33 1.94e- 1
## 3 treatment_2 0.494 0.279 1.77 8.77e- 2

You now have two numeric variables that you use in an ordinary multiple
regression analysis. We see the effects (the ’slopes’) of the two dummy variables.
Based on these slopes and the intercept, we can construct the linear equation
for the relationship between treatment and weight (yield):

d = 5.03 − 0.37 × treatment 1 + 0.49 × treatment 2

weight (6.13)

Based on this we can make predictions for the mean height in the control
group, the treatment 1 group and the treatment 2 group.
Control group specimens score 0 on variable treatment 1 and 0 on variable
treatment 2. Therefore, their predicted weight equals:

5.03 − 0.37 × 0 + 0.49 × 0 = 5.03

.
Thus, the expected weight in the control group is equal to the intercept, as
we used the control group as the reference group.
Specimens in the treatment 1 group score 1 on the treatment 1 variable but
0 on the treatment 2 variable. Therefore, their predicted weight equals:

5.03 − 0.37 × 1 + 0.49 × 0 = 5.03 − 0.37 = 4.66

.
Specimens in the treatment 2 group score 0 on the treatment 1 variable but
1 on the treatment 2 variable. Therefore, their predicted weight equals:

5.03 − 0.37 × 0 + 0.49 × 1 = 5.03 + 0.49 = 5.52

.

6.7.2 Let R create dummy variables automatically

The alternative is that R creates dummy variables automatically. The upside
is that it is less work for you, the downside is that the first category (in terms
of internally numbered category) always ends up as the reference category. The
only thing that is required is that the independent variable in question is stored
as a factor variable. Thus, in this case we want R to make dummy variables for
our categorical variable group. But first, let’s check whether the group variable
is indeed a factor variable. If we look at the structure of it, we can see the types
of variables.

199
PlantGrowth %>%
str()

## ’data.frame’: 30 obs. of 5 variables:

## $ weight : num 4.17 5.58 5.18 6.11 4.5 4.61 5.17 4.53 5.33 5.14 ...
## $ group : Factor w/ 3 levels "ctrl","trt1",..: 1 1 1 1 1 1 1 1 1 1 ...
## $ grouptrt1 : num 0 0 0 0 0 0 0 0 0 0 ...
## $ treatment_1: num 0 0 0 0 0 0 0 0 0 0 ...
## $ treatment_2: num 0 0 0 0 0 0 0 0 0 0 ...

Yes, the group variable is a factor (Factor). If we use a factor in a linear

model, R will automatically code this variable into a set of dummy variables.

## lm(weight ~ group, data = .)

out %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 5.53 0.203 27.2 4.52e-16
## 2 grouptrt1 -0.865 0.287 -3.01 7.52e- 3

The regression table now looks slightly different: all values are the same as
in the analysis where we created our own dummies, except that R has used
different names for the dummy variables it created. The parameter values are
exactly the same as in the previous analysis, because in both analyses the control
condition is the reference category. This control condition is chosen by R as the
reference category, because alphabetically, ctrl comes before trt1 and trt2.
Thus, the new variable grouptrt1 codes 1s for all observations where group
= trt1 and 0s otherwise, and the new variable grouptrt2 codes 1s for all
observations where group = trt2 and 0s otherwise.

6.7.3 Interpreting the regression table

The intercept is the expected mean for the reference group, that is, the group
for which there is no dummy variable. Each slope is the difference between the
group to which the slope belongs to and the reference group. For instance, in
the yield example with manual coding of the dummy variables, the slope for the
treatment 1 dummy variable is the estimated population difference between
the mean for treatment 1 minus the mean for control condition. Similarly, the
slope for the treatment 2 dummy variable is actually the estimated population
difference between the mean for treatment 2 minus the mean for the control
condition. The confidence intervals that belong to these parameter effects are
also to be seen in this light: they are intervals for probable values for the
population difference between the respective group means and the mean of the

200
reference group. The t-values and p-values are related to null-hypothesis tests
regarding these differences to be 0 in the population.
As an example, suppose that we want to estimate the difference in mean
weight between plants from the treatment 1 group relative to the control group.
From the output, we see that our best guess for this difference (the least square
estimate) equals -0.37, where the yield is less with treatment 1 than under
control conditions. The standard error for this difference equals 0.28. So a
rough indication for the 95% confidence interval would be from −0.37 − 2 × 0.28
to −0.37 + 2 × 0.28, that is, from −0.93 to 0.19. Therefore, we infer that in the
population, our best guess for the difference is somewhere between −0.93 and
0.19.
If we would want to, we could perform three null-hypothesis tests based on
this output: 1) whether the population intercept equals 0, that is, whether the
population mean of the control group equals 0; 2) whether the slope of the
treatment 1 dummy variable equals 0, that is, whether the difference between
the population means of treatment 1 group and the control group is 0; and 3)
whether the slope of the treatment 2 group dummy variable equals 0, that is,
whether the difference between the population means of the treatment 2 group
and the control group is 0.
Obviously, the first hypothesis is not very interesting: we’re not interested to
know whether the average weight in the control group equals 0. But the other
two null-hypotheses could be interesting in some scenarios. What is missing from
the table is a test for the null-hypothesis that the means of the two treatments
conditions are equal. This could be solved by manually creating two other
dummy variables, where either treatment 1 or 2 is the reference group, or by
looking at tricks in Chapter ??. But what is also missing is a test for the null-
hypothesis that all three population means are equal. In order to do that, we
first need to explain analysis of variance.

6.8 Analysis of Variance

Since we know that applying a linear model to a categorical independent variable
is the same as modelling group means, we can test the null-hypothesis that all
group means are equal in the population. Let µt1 be the mean yield in the
population of the treatment 1 group, µt2 be the mean yield in the treatment 2
group, and µc be the mean yield in the control group. Then we can specify the
null-hypothesis using symbols in the following way:

H0 : µt1 = µt2 = µc (6.14)

If all group means are equal in the population, then all population slopes
would be 0. We want to test this null-hypothesis with a linear model in R. We
then have only one independent variable, group, and if we let R do the dummy
coding for us, R can give us an Analysis of Variance. We do that in the following
way:

201
out <- PlantGrowth %>%
lm(weight ~ group, data = .)
out %>%
anova() %>%
tidy()

## # A tibble: 2 x 6
## term df sumsq meansq statistic p.value
## <chr> <int> <dbl> <dbl> <dbl> <dbl>
## 1 group 2 3.77 1.88 4.85 0.0159
## 2 Residuals 27 10.5 0.389 NA NA

We don’t see a regression table, but output based on a so-called Analysis Of

VAriance, or ANOVA for short. This table is usually called an ANOVA table.
The function anova() is used after fitting a linear model with lm(). ANOVA
is in fact an alternative way of presenting the results of a linear model.
In the output, you see a column statistic, with the value 4.85 for the group
variable. It looks similar to the column with the t-statistic in a regression table,
but it isn’t. The statistic is an F -statistic.
The F -statistic is constructed on the basis of Sum of Squares (SS, sumsq in
the R table). Sums of squares we already encountered in Chapter 1, where they
form the basis of variances and standard deviations. We also saw sums of squares
in Chapter 4 where the sum of squared residuals (SSR) was minimised to get the
least squares estimator of regression coefficients. In Chapter 4 we also saw that
the sum of squared residuals (SSR) and the total sum of squares (SST) were
used to compute the R-squared and the adjusted R-squared. Actually, the sum
of squares that we see in the ANOVA table here in the row named Residuals
is exactly the SSR: the sum of the squared residuals. Here we see that the sum
of the squared residuals equals 10.5.
In the ANOVA table, we also see degrees of freedom (df). The degrees of
freedom in the row named Residuals are the residual degrees of freedom that we
already use when doing linear regression (Chapter 5). Here we see the residual
degrees of freedom equals 27. This is so because we have 30 data points, and for
a linear model the number of degrees of freedom is n − K − 1 = 30 − 2 − 1 = 27,
with K being the number of independent variables (two dummy variables).
Continuing, we see Mean Squares (meansq). These numbers are nothing
but the sum of squares (sumsq) divided by the respective degrees of freedom
(df). For instance, in the row for Residuals, the Sum of Squares equals 10.5,
the degrees of freedom equals 27, and the Mean square equals 10.5 27 = 0.389.
Then there is a column with F -values (statistic). F -values are test
statistics and are used in the same way as t-statistics. Under the null-hypothesis
they have a known distribution, and if they have very extreme values, you can
reject the null-hypothesis. Whether or not the null-hypothesis can be rejected,
depends on your pre-set α level and whether the p-value, reported in the last
column (p.value) is equal or smaller than your α level.
The F -value is computed based on the Mean squares values (meansq). Let’s

202
look at the F -value for the group variable. The F -value equals 4.85. This is
the ratio of the mean square of the group effect, which is 1.88, and the mean
square of the residuals (error), which is 0.389. Thus, the F -value for country is
1.88
computed as 0.389 = 4.85. Under the null-hypothesis that all three population
means are equal, this ratio is around 1. Why this is so, we will explain later.
Here we see that the F -value based on these sample data is larger than 1. But
is it large enough to reject the null-hypothesis? That depends on the degrees of
freedom. The F -value is based on two mean squares, and these in turn are based
on two separate numbers of degrees of freedom. The one for the effect of country
was 2 (3 countries so 2 degrees of freedom), and the one for the residual mean
square was 27 (27 residual degrees of freedom). We therefore have to look up in
a table whether an F -value of 4.85 is significant at 2 and 27 degrees of freedom
for a specific α. Such a table is displayed in Table 6.9. It shows critical values if
your α is 0.05. In the columns we look up our model degrees of freedom. Model
degrees of freedom is computed based on the number of independent variables.
Here we have a categorical variable group. But because this categorical variable
is represented in the analysis as two dummy variables, the number of variables
is actually 2. The model degrees of freedom is therefore 2.
In the rows of Table 6.9 we look up our residual degrees of freedom: 27.
For 2 and 27 degrees of freedom we find a critical F -value of 3.35. It means
that if we have an α of 0.05, an F -value of 3.35 or larger is large enough to
reject the null-hypothesis. Here we found an F -value of 4.85, so we reject the
null-hypothesis that the three population means are equal. Therefore, the mean
weight is not the same in the three experimental conditions.

Table 6.9: Critical values for the F -value if α = 0.05, for different model degrees
of freedom (columns) and error degrees of freedom (rows).
1 2 3 4 5 10 25 50
5 6.61 5.79 5.41 5.19 5.05 4.74 4.52 4.44
6 5.99 5.14 4.76 4.53 4.39 4.06 3.83 3.75
10 4.96 4.10 3.71 3.48 3.33 2.98 2.73 2.64
27 4.21 3.35 2.96 2.73 2.57 2.20 1.92 1.81
50 4.03 3.18 2.79 2.56 2.40 2.03 1.73 1.60
100 3.94 3.09 2.70 2.46 2.31 1.93 1.62 1.48

Note that our null-hypothesis that all group means are equal in the population
cannot be answered based on a regression table. If the population means are all
equal, then the slope parameters should consequently be 0 in the population.
Let’s have a look again at the regression table, plotting also the 95% confidence
intervals:

out <- PlantGrowth %>%

lm(weight ~ group, data = .)
out %>%
tidy(conf.int = TRUE)

203
## # A tibble: 3 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 5.03 0.197 25.5 1.94e-20 4.63 5.44
## 2 grouptrt1 -0.371 0.279 -1.33 1.94e- 1 -0.943 0.201
## 3 grouptrt2 0.494 0.279 1.77 8.77e- 2 -0.0780 1.07

Looking at the 95% confidence intervals for grouptrt1 and grouptrt2, we

see that 0 is a reasonable value for the difference between the control group (the
reference category) and treatment 1, and that 0 is also a reasonable value for the
difference between the control group and treatment 2. But what does that tell us
about the difference between the two treatment groups? How can we rigorously
test the null-hypothesis, with one clear statistic, that all three group means are
the same? In the regression table we have two t-statistics and two p-values, one
for the difference between control and treatment 1 (t = −1.33, p = 0.194) and
one of the difference between treatment 2 and the control group (t = 1.77, p =
0.0.0877), but we actually need one p-value for the null-hypothesis of three equal
means.
The ANOVA table looks slightly different: instead of two separate effects for
two dummy variables, we only see one row for the original categorical variable
group. And in the column df (degrees of freedom): instead of 1 degree of
freedom for a specific dichotomous dummy variable, we see 2 degrees of freedom
for the nominal group variable. So this suggests that the effects of the two
dummy variables are now combined into one effect, with one particular F -
statistic, and one p-value that is also different from those of the two separate
dummy variables. This is actually the p-value associated with the test of the
null-hypothesis that all 3 means are equal:

H0 : µt1 = µt2 = µc (6.15)

This hypothesis test is very different from the t-tests in the regression table.
The t-test for the grouptrt1 effect specifically tests whether the average weight
in treatment 1 group is different from the average weight in the control group
(the reference country). The t-test for the grouptrt2 effect specifically tests
whether the average weight in the treatment 2 group is different from the average
weight in the control group (the reference country). Since these two hypotheses
do not refer to our original research question regarding overall differences across
all three groups, we do not report these t-tests, but we report the overall F -test
from the ANOVA table.
In general, the rule is that if you have a specific research question that
addresses a particular null-hypothesis, you only report the statistical results
regarding that null-hypothesis. All other p-values that your software happens
to show in its output should be ignored. We will come back to this issue in
Chapter ??.

204
6.9 The logic of the F -statistic
As stated earlier, the ANOVA is an alternative way of representing the linear
model. Suppose we have a dependent variable Y , and three groups: A, B and
C. In the usual linear model, we have an intercept b0 , and we use two dummy
variables. Suppose we use C as our reference group, then we need two dummy
variables for groups A and B. We could model the data then using the following
equation, with normally distributed errors:

Y = b0 + b1 dummyA + b2 dummyB + e (6.16)

2
e ∼ N (0, σ ) (6.17)

This is the linear model as we know it. The linear equation has three
unknown parameters that need to be estimated: one intercept and two dummy
effects. The dummy effects are the differences between the means of groups A
and B relative to reference group C.
Alternatively, we could represent the same data as follows:

Y = b1 dummyA + b2 dummyB + b3 dummyC + e (6.18)

2
e ∼ N (0, σ ) (6.19)

That is, instead of estimating one intercept and two dummy effects, we
simply estimate the three population means directly! We leave out the intercept,
and we estimate three population means.
Next, we focus on the variance of the dependent variable, Y in this case,
that is split up into two parts: one part that is explained by the independent
variable (groups in this case) and one part that is not explained (cf. Chapters
4). The unexplained part is easiest of course: that is simply the part shown by
the residuals, hence σ 2 .
The logic of the F -statistic is entirely based on this σ 2 . As stated earlier,
under the null-hypothesis the F -statistic should have a value of around 1. This
is because F is a ratio and under the null-hypothesis, the numerator and the
denominator of this ratio should be more or less equal. This is so because
both the numerator and the denominator are estimators of σ 2 . Under the null-
hypothesis, these estimators should result in more or less the same numbers,
and then the ratio is more or less 1. If the null-hypothesis is not true, then the
numerator becomes larger than the denominator and hence the F -value becomes
larger than 1.
In the previous section we saw that the numerator of the F -statistic was
computed by taking the sum of squares of the group variable and dividing it
by the degrees of freedom. What is actually being done is the following: If the
null-hypothesis is really true, then the three population means are equal, and
you simply have three independent samples from the same population. Each
sample mean shows simply a slight deviation from the population mean.

205
 This variance of sample means should remind us of something. If we go
back to Chapter 2, we saw there that if we have a population with mean µ and
variance σ 2 , and if we draw many many random samples of size n and compute
sample means for each sample, their distribution is a sampling distribution (Fig.
2.2). We also saw in Chapter 2 that on average the sampling distribution will
show a mean that is the same as the population mean: the sample mean is an
unbiased estimator of the population mean. And, important for ANOVA, the
standard deviation of q the sampling distribution, known as the standard error,
2
will be equal to σY = sn (Chapter 2). If we take the square, we see that the
variance of the sample means is equal to

s2
σY2 = (6.20)
n
where s2 is the unbiased estimator of the variance of Y .
In the ANOVA model above, we have three group means. Now, suppose we
have an alternative model, under the null-hypothesis, that there is really only
one population mean µ, and that the observed different group means in groups
A, B and C are only the result of chance (random sampling). Then the variance
of the group means is nothing but the square of the standard error, and the
number of observations per group is the sample size n. If that is the case, then
we can flip the equation of the standard error around and say:

c2 = σ 2 × n = SS × n
σ (6.21)
Y 2
or in words: our estimate of the total variance of Y in the population is the
estimated variance of the group means in the population times the number of
observations per group.
So the numerator is one estimator of the variance of the residuals. For
that estimator we only used information about the group means: we looked
at variation between groups. Now let’s look at the denominator. For that
estimator we use information from the raw data and how they deviate from
the sample group means, that is we look at within-group variation. Similar to
regression analysis, for each observed value, we compute the difference between
the observed value and the group mean. We then compute the sum of squared
residuals, SSR. If we want the variance, we need to divide this by sample size,
n. However, if we want to estimate the variance in the population, we need to
divide by a corrected n. In Chapter 2 we saw that if we wanted to estimate a
variance in the population on the basis of one sample with one sample mean,
SS
we used s2 = n−1 . The n − 1 was in fact due to the loss of 1 degree of freedom
because by computing the sample variance, we used the sample mean, which was
only an estimate of the population mean. Here, because we have three groups,
we need to estimate three population means, and the degrees of freedom is
therefore n − 3. The estimated variance in the population that is not explained
by the independent variable is therefore SSR/(n − 3).

206
 4

^
Y
Y

^
Y
Y Y Y
2

^
Y
1

A B C
group

Figure 6.6: Illustration of ANOVA using a very small data set. In grey the raw
data, in black the overall sample mean, and in red the sample group means.

Thus, if we want to estimate σ 2 related to the model, we can either do that

by looking at the model residuals, computing the sum of squared residuals and
dividing it by the degrees of freedom (in this case SSR/(n − 3)), but we can
also do it by looking at the variation of the group means and multiplying it by
the group size. Now, the method of looking at the residuals will generally yield
a good estimate of σ 2 , whether the null-hypothesis is true or not. However,
only if the null-hypothesis is true, the method of looking at the variation of
group means will yield a good estimate of σ 2 . Only if the null-hypothesis is
true, both estimates will be more or less the same. But if the null-hypothesis is
not true, if the means are really different in the population, then the method of
looking at the variation of group means will yield an estimate of σ 2 that is larger
than an estimate based on residuals. Then, if you compute a ratio, the ratio
will become larger than 1. Therefore, an F -statistic larger than 1 is evidence
that the population means might not be equal. How much larger than 1 an
F -value should be to regard it as evidence against H0 depends on your level of
significance α and the degrees of freedom.

6.10 Small ANOVA example

To illustrate the idea of ANOVA and the computation of the F -statistic, let’s
assume we have a very small data set, involving height data from three countries
A, B and C. From each country, we only have 3 data points. The data are
plotted in Figure 6.6. In grey, we see the raw data values for variable Y . In
group A, we see the values 4, 2 and 1; in group B, we see the values 4, 2 and

207
2, and in group C, we see the values 2, 1 and 1. When we sum all these values
and divide by 9, we get the overall mean (the grand mean), which is equal to
Y = 2.1111111, denoted in black in Figure 6.6. In red, we see the sample group
means. For group A, that is equal to (4+2+1)/3 = 2.3333333, for group B this is
(4+2+2)/3 = 2.6666667, and for group C this equals (2+1+1)/3 = 1.3333333.
Thus, our ANOVA model for these data is the following:

Y = b1 dummyA + b2 dummyB + b3 dummyC + e (6.22)

2
e ∼ N (0, σ )

Our OLS estimates for the parameters are the sample means, so that we
have the linear equation

Yb = 2.3333333dummyA + 2.6666667dummyB + 1.3333333dummyC (6.23)

Based on this linear equation we can determine the predicted values for each
data point. Table 6.10 shows the Y -values, the group variable, the dummy
variables from the ANOVA model equation (Equation 6.22) and the predicted
values. We see that the predicted value for each observed value is equal to the
sample group mean.

Table 6.10: Small data example for illustrating ANOVA and the F -statistic.
Y group dummy A dummy B dummy C predicted residual
1 A 1 0 0 2.33 -1.33
2 A 1 0 0 2.33 -0.33
4 A 1 0 0 2.33 1.67
2 B 0 1 0 2.67 -0.67
2 B 0 1 0 2.67 -0.67
4 B 0 1 0 2.67 1.33
2 C 0 0 1 1.33 0.67
1 C 0 0 1 1.33 -0.33
1 C 0 0 1 1.33 -0.33

Using these predicted values, we can compute the residuals, also displayed
in Table 6.10, and these help us to compute the first estimate of σ 2 , the one
based on residuals, namely the SSR divided by the degrees of freedom. If we
square the residuals in Table 6.10 and sum them, we obtain SSR = 8. To obtain
the Mean Squared Error (MSE or meansq for Residuals), we divide the SSR by
the degrees of freedom. Because the linear model with 2 dummy variables has
n − K − 1 = 9 − 2 − 1 = 6 residuals degrees of freedom (see Chapter 5), we also
have only 6 residual degrees of freedom. Thus we get M SE = 8/6 = 1.3333333.
We can see these numbers in the bottom row in the ANOVA table, displayed in
Table 6.11.

208
 For our second estimate of σ 2 , the one based on the group means, we look
at the squared deviations of the group means from the overall mean (the grand
mean). We saw that the grand mean equals 2.11. The sample mean for group
A was 2.3333333, so the squared deviation equals 0.0493828. The sample mean
for group B was 2.6666667, so the squared deviation equals 0.3086421. Lastly,
the sample mean for group C was 1.3333333, so the squared deviation equals
0.3086421. Adding these squared deviations gives a sum of squares of 0.962963.
To obtain an unbiased estimate for the population variance of these means,
we have to divide this sum of squares by the number of groups minus 1 (model
degrees of freedom), thus we get 0.962963/2 = 0.6604939. This we must multiply
by the sample size per group to obtain an estimate of σ 2 (see Equation 6.21),
thus we obtain 1.4444444.

Table 6.11: ANOVA table for small data example.

term df sumsq meansq statistic p.value
group 2 2.89 1.44 1.08 0.40
Residuals 6 8.00 1.33

Table 6.12: Small data example for illustrating ANOVA and the F -statistic.
Y group predicted grand mean deviation sq deviation
1 A 2.33 2.11 0.22 0.05
2 A 2.33 2.11 0.22 0.05
4 A 2.33 2.11 0.22 0.05
2 B 2.67 2.11 0.56 0.31
2 B 2.67 2.11 0.56 0.31
4 B 2.67 2.11 0.56 0.31
2 C 1.33 2.11 -0.78 0.60
1 C 1.33 2.11 -0.78 0.60
1 C 1.33 2.11 -0.78 0.60

Obtaining the estimate of σ 2 based on the group means can also be illustrated
using Table 6.12. There again we see the raw data values for variable Y , the
predicted values (the group means), but now also the grand mean, the deviations
of the sample means from the grand mean, and their squared values. If we simply
add the squared deviations, we no longer have to multiply by sample size. Thus
we have as the sum of squares 2.8888889. Then we only have to divide by the
number of groups minus 1, so we have 2.8888889/2 = 1.4444444. This sum of
squares, the degrees of freedom of 2, and the resulting MS can also be seen in
the ANOVA table in Table 6.11.
Hence we have two estimates of σ 2 , the one called the Mean Squared Error
(MSE) that is based on the residuals (sometimes also called the MS within or
MSW), and the other one called the Mean Squared Between groups (MSB), that
is based on the sum of squares of group mean differences. For the F -statistic, we

209
use the MS Between (MSB) as the numerator and the MSE as the denominator,
M SBgroup 1.4444444
F = = = 1.0833333 (6.24)
M SE 1.3333333
We see that the F -statistic is larger than 1. That means that the estimate
for σ 2 , M SBgroup , based on the sample means is larger than the estimate based
on the residuals, M SE. This could indicate that the null-hypothesis, that the
three population means are equal, is not true. However, is the F -value really
large enough to justify such a conclusion?
To answer that question, we need to know what values the F -statistic would
take for various data sets if the null-hypothesis were true (the sampling distribution
of F ). If for each data set we have three groups, each consisting of three observed
values, then we have 2 degrees of freedom for the group effect, and 6 residual
degrees of freedom. Table 6.9 shows critical values if we want to use an α of
0.05. If we look up the column with a 2 (for the number of model degrees of
freedom) and the row with a 6 (for the residual degrees of freedom), we find
a critical F -value of 5.14. This means that if the null-hypothesis is true and
we repeatedly take random samples, we find an F -value equal to or larger than
5.14 only 5% of the time. If we want to reject the null-hypothesis, therefore, at
an alpha of 5%, the F -value has to be equal or larger than 5.14. Here we found
an F -value of only 1.0833333, which is much smaller, so we cannot reject the
null-hypothesis that the means are equal.
For illustration, Figure 6.7 shows the distribution of the F -statistic with 2
and 6 degrees of freedom under the null-hypothesis. The figure shows it happens
quite a lot under the null-hypothesis that the F -statistic is equal to 1.0833333
or larger.

6.11 Reporting ANOVA

In all cases where you have a categorical predictor variable with more than
two categories, and where the null-hypothesis is about the equality of all group
means, you have to use the factor variable in R associated with the original
nominal variable. That is, don’t make dummy variables yourself, but let R do it
for you. You then always report the corresponding F -statistic from the ANOVA
table.
For this particular example, you report the results of the analysis of variance
in the following way:

”The null-hypothesis that all 3 population means are equal was

tested with an analysis of variance. The results showed that the null-
hypothesis cannot be rejected, F (2, 6) = 1.08, M SE = 1.33, p =
0.40”.

Always check the degrees of freedom for your F -statistic carefully. The first
number refers to the degrees of freedom for the Mean Square Between: this is
the number of groups minus 1 (K − 1). This is equal to the number of dummy

210
 1.00

0.75
density

0.50

0.25

5%
0.00

0 1 2 3 4 5 6
F

Figure 6.7: Density plot of the F -distribution with 2 and 6 degrees of freedom.
In blue the observed F -statistic in the small data example, in red the critical
value for an α of 0.05. The blackened area under the curve is 5%.

variables are used in the linear model. This is also called the model degrees of
freedom. The second number refers to the residual degrees of freedom: this is
n − K − 1 as we saw Chapter 5, where K is the number of dummy variables.
In this ANOVA model you have 9 data points and you have 2 dummy variables
for the three groups. So your residual degrees of freedom is 9 − 2 − 1 = 6.
This residual degrees of freedom is equal to that of the t-statistic for multiple
regression.

6.12 Relationship between F - and t-distributions

The t-distribution and the F -distribution have much in common. Here we
will illustrate this. Suppose that we test the null-hypothesis that a certain
population slope is 0. We perform a regression analysis and obtain a t-statistic
of -2.40. Suppose our sample size was 42, so that our residual degrees of freedom
equals 42 − 2 = 40. Figure 6.8 shows the theoretical t-distribution with 40
degrees of freedom. It also shows our value of -2.40. The shaded area represents
the values for t that would be significant at an α = 0.05.
Now look closely at Figure 6.8. The density says something about the
probability of drawing certain values. Imagine that you randomly pick numbers
from this t-distribution. The density plot tells you that values around zero are
more probable than values around 2 or -2, and that values around 2 or -2 are
more probable than values around 3 or -3. Imagine that you pick a million
values for t, randomly from this t-distribution. Then imagine that you take the

211
 0.4

density 0.3

0.2

0.1

0.0
−5 −4 −3 −2 −1 0 1 2 3 4 5
t

Figure 6.8: The vertical line represents a t-value of -2.40. The shaded area
represents the extreme 5% of the possible t-values

square of each value (thus, suppose as the first 3 randomly drawn t-values you
get -3.12, 0.14, and -1.6, you then square these numbers to get the numbers 9.73,
0.02, and 2.79). If you then make a density plot of these one million squared
numbers, you get the density plot in Figure 6.9. It turns out that this density
is an F -distribution with 1 model degrees of freedom and 40 residual degrees of
freedom.
If we also square the observed test statistic t-value of -2.40, we obtain an
F -value of 5.76. From online tables, we know that, with 1 model degrees of
freedom and 40 residual degrees of freedom, the proportion of F -values larger
than 5.76 equals 0.02. The proportion of t-values, with 40 (residual) degrees of
freedom, larger than 2.40 or smaller than -2.40 is also 0.02. Thus, the two-sided
p-value associated with a certain t-value, is equal to the p-value associated with
an F -value that is the square of the t-value.

F (1, x) = t2 (x) (6.25)

This means that if you see a t-statistic of say -2.40 reported with a residual
degrees of freedom of 40, t(40) = −2.40, you can equally report this as an
F (1, 40) = (−2.40)2 = 5.76. Similarly, if you see a reported F -value of
F (1, 67) = 49, you could without problems turn this into a t(67) = 7. Note
however that this is only the case if the model degrees of freedom of the F -
statistic is equal to 1. This means you cannot do this if you are comparing more
than two groups means.

212
 0.75
density

0.50

0.25

0.00
0 1 2 3 4 5 6 7 8
F

Figure 6.9: The F -distribution with 1 model degrees of freedom and 40 error
degrees of freedom. The shaded area is the upper 5% of the distribution. The
vertical line represents the square of -2.40: 5.76

213
214
Chapter 7

Assumptions of linear
models

7.1 Introduction
Linear models are models. A model describes the relationship between two or
more variables. A good model gives a valid summary of what the relationship
between the variables looks like. Let’s look at a very simple example of two
variables: height and weight. In a sample of 100 children from a distant country,
we find 100 combinations of height in centimetres and weight in kilograms that
are depicted in the scatter plot in Figure 7.1.
We’d like to find a linear model for these data, so we determine the least
squares regression line. We also determine the standard deviation of the residuals
so that we have the following statistical model:

weight = −104.83 + 1.04 × height + e (7.1)

e ∼ N (0, σ = 4.04) (7.2)

This model, defined above, is depicted in Figure 7.2. The blue line is the
regression line, and the dots are the result of simulating (inventing) independent
normal residuals with standard deviation 4.04. The figure shows how the data
would like according to the model.
The actual data, displayed in Figure 7.1 might have arisen from this model
in Figure 7.2. The data is only different from the simulated data because of the
randomness of the residuals.
A model should be a good model for two reasons. First, a good model is a
summary of the data. Instead of describing all 100 data points on the children,
we could summarise these data with the linear equation of the regression line and
the standard deviation (or variance) of the residuals. The second reason is that
you would like to infer something about the relationship between height and

215
 60

40
weight

20

0
110 120 130 140 150
height

Figure 7.1: Data set on height and weight in 100 children.

60

40
weight

20

0
110 120 130 140 150
height

Figure 7.2: Data set on height and weight in 100 children and the least squares
regression line.

216
weight in all children from that distant country. It turns out that the standard
error, and hence the confidence intervals and hypothesis testing, are only valid
if the model describes the data well. This means that if the model is not a good
description of your sample data, then you draw the wrong conclusions about
the population.
For a linear model to be a good model, there are four conditions that need
to be fulfilled.

1. linearity The relationship between the variables can be described by a

linear equation (also called additivity)

2. independence The residuals are independent of each other

3. equal variance The residuals have equal variance (also called homoskedasticity)

4. normality The distribution of the residuals is normal

If these conditions (often called assumptions) are not met, the inference
with the computed standard error is invalid. That is, if the assumptions are
not met, the standard error should not be trusted, or should be computed using
alternative methods.
Below we will discuss these four assumptions briefly. For each assumption,
we will show that the assumption can be checked by looking at the residuals. We
will see that if the residuals do not look right, one or more of the assumptions
are violated. But what does it mean that the residuals ’look right’ ?
Well, the linear model says that the residuals have a normal distribution. So
for the height and weight data, let’s apply regression, compute the residuals for
all 100 children, and plot their distribution with a histogram, see Figure 7.3.
The histogram shows a bell-shaped distribution with one peak that is more or
less symmetric. The symmetry is not perfect, but you can well imagine that if
we had measured more children, the distribution could more and more resemble
a normal distribution.
Another thing the model implies is that the residuals are random: they are
random draws from a normal distribution. This means, if we would plot the
residuals, we should see no systematic pattern in the residuals. The scatter
plot in Figure 7.4 plots the residuals in the order in which they appear in the
data set. The figure seems to suggest a random scatter of dots, without any
kind of system or logic. We could also plot the residuals as a function of the
predicted height (the dependent variable). This is the most usual way to check
for any systemetic pattern. Figure 7.5 shows there is no systematic relationship
between the predicted height of a child and the residual.
When it looks like this, it shows that the residuals are randomly scattered
around the regression line (the predicted heights). Taken together, Figures 7.3,
7.4 and 7.5 suggest that the assumptions of the linear model are met.
Let’s have a look at the same kinds of residual plots when each of the
assumptions of the linear model are violated.

217
 20

15
count

10

0
−10 −5 0 5 10
residuals

Figure 7.3: Histogram of the residuals after regressing weight on height.

10

5
residuals

−5

0 25 50 75 100
observation

Figure 7.4: Residual plot after regressing weight on height.

218
 5
residual

−5

110 120 130 140 150

predicted height

Figure 7.5: Residuals as a function of height.

7.2 Independence
The assumption of independence is about the way in which observations are
similar and dissimilar from each other. Take for instance the following regression
equation for children’s height predicted by their age:

height = 100 + 5 × age + e (7.3)

This regression equation predicts that a child of age 5 has a height of 125 and
a child of age 10 has a height of 150. In fact, all children of age 5 have the same
predicted height of 125 and all children of age 10 have the same predicted height
of 150. Of course, in reality, children of the same age will have very different
heights: they differ. According to the above regression equation, children are
similar in height because they have the same age, but they differ because of
the random term e that has a normal distribution: predictor age makes them
similar, residual e makes them dissimilar. Now, if this is all there is, then this is
a good model. But let’s suppose that we’re studying height in an international
group of 50 Ethiopian children and 50 Vietnamese children. Their heights are
plotted in Figure 7.6.
From this graph, we see that heights are similar because of age: older children
are taller than younger children. But we see that children are also similar
because of their national background: Ethiopian children are systematically
taller than Vietnamese children, irrespective of age. So here we see that a
simple regression of height on age is not a good model. We see that, when we
estimate the simple regression on age and look at the residuals in Figure 7.7.
As our model predicts random residuals, we expect a random scatter of
residuals. However, what we see here is a systematic order in the residuals: they

219
 160

150

country
height

Ethiopian
140
Vietnamese

130

120
4 6 8 10 12
age

Figure 7.6: Data on age and height in children from two countries.

2
residual

−2

0 25 50 75 100
child

Figure 7.7: Residual plot after regressing height on age.

220
 2
residual

−2

Ethiopian Vietnamese
country

Figure 7.8: Residual plot after regressing height on age.

tend to be positive for the first 50 children and negative for the last 50 children.
These turn out to be the Ethiopian and the Vietnamese children, respectively.
This systematic order in the residuals is a violation of independence: the residuals
should be random, and they are not. The residuals are dependent on country:
positive for Ethiopians, negative for Vietnamese children. We see that clearly
when we plot the residuals as a function of country, in Figure 7.8.
Thus, there is more than just age that makes children similar. That means
that the model is not a good model: if there is more than just age that makes
children more alike, then that should be incorporated into our model. If we use
multiple regression, including both age and country, and we do the analysis,
then we get the following regression equation:

d = 102.641 + 5.017 × age − 1.712 × countryViet

height (7.4)

When we now plot the residuals we see that there is no longer a clear country
difference, see Figure 7.9.
Another typical example of non-random scatter of residuals is shown in
Figure 7.10. They come from an analysis of reaction times, done on 10 students
where we also measured their IQ. Each student was measured on 10 trials. We
predicted reaction time on the basis of student’s IQ using a simple regression
analysis. The residuals are clearly not random, and if we look more closely, we
see some clustering if we give different colours for the data from the different
students, see Figure 7.11.
We see the same information if we draw a boxplot, see Figure 7.12. We
see that residuals that are close together come from the same student. So,
reaction time are not only similar because of IQ, but also because they come
from the same student: clearly something other than IQ also explains why

221
 3

1
res

−1

−2
Ethiopian Vietnamese
country

Figure 7.9: Residual plot after regressing height on age and country.

0
residual

−5

0 25 50 75 100
trial

Figure 7.10: Residual plot after regressing reaction time on IQ.

222
 5
student
1
2
3
0
4
residual

5
6
7
−5 8
9
10

0 25 50 75 100
trial

Figure 7.11: Residual plot after regressing reaction time on IQ, with separate
colours for each student.

0
residual

−5

1 2 3 4 5 6 7 8 9 10
student

Figure 7.12: Box plot after regressing reaction time on IQ.

223
reaction times are different across individuals. The residuals in this analysis
based on IQ are not independent: they are dependent on the student. This
may be because of a number of factors: dexterity, left-handedness, practice,
age, motivation, tiredness, or any combination of such factors. You may or may
not have information about these factors. If you do, you can add them to your
model and see if they explain variance and check if the residuals become more
randomly distributed. But if you don’t have any extra information, or if do
you but the residuals remain clustered, you might either consider adding the
categorical variable student to the model or use linear mixed models, discussed
in Chapter 10.
The assumption of independence is the most important assumption in linear
models. Just a small amount of dependence among the observations causes your
actual standard error to be much larger than reported by your software. For
example, you may think that a confidence interval is [0.1, 0.2], so you reject the
null-hypothesis, but in reality the standard error is much larger, with a much
wider interval, say [-0.1, 0.4] so that in reality you are not allowed to reject
the null-hypothesis. The reason that this happens can be explained when we
look again at Figure 7.11. Objectively, there are 100 observations, and this is
fed into the software: n = 100. This sample size is then used to compute the
standard error (see Chapter 5). However, because the reaction times from the
same student are so much alike, effectively the number of observations is much
smaller. The reaction times from one student are in fact so much alike, you could
almost say that there are only 10 different reaction times, one for each student,
with only slight deviations within each student. Therefore, the real number of
observations is somewhere between 10 and 100, and thus the reported standard
error is underestimated when there is dependence in your residuals (standard
errors are inversely related to sample size, see Chapter 5).

7.3 Linearity
The assumption of linearity is often also referred to as the assumption of additivity.
Contrary to intuition, the assumption is not that the relationship between
variables should be linear. The assumption is that there is linearity or additivity
in the parameters. That is, the effects of the variables in the model should add
up.
Suppose we gather data on height and fear of snakes in 100 children from a
different distant country. Figure 7.13 plots these two variables, together with
the least squares regression line.
Figure 7.14 shows a pattern in the residuals: the positive residuals seem to
be smaller than the negative residuals. We also clearly see a problem when we
plot residuals against the predicted fear (see Fig. 7.15). The same problem is
reflected in the histogram in Figure 7.16, that does not look symmetric at all.
What might be the problem?
Take another look at the data in Figure 7.13. We see that for small heights,
the data points are all below the regression line, and the same pattern we see

224
 2000
fear

1000

110 120 130 140 150

height

Figure 7.13: Least squares regression line for fear of snakes on height in 100
children.

100

0
residuals

−100

−200

−300

0 25 50 75 100
observation

Figure 7.14: Residual plot after regressing fear of snakes on height.

225
 100

0
residual

−100

−200

−300

1000 2000
predicted fear

Figure 7.15: Residual plot after regressing fear of snakes on height.

10
count

0
−300 −200 −100 0 100
residual

Figure 7.16: Histogram of the residuals after regressing fear of snakes on height.

226
 2000
Fear for snakes

1500

1000

500

110 120 130 140 150

height

Figure 7.17: Observed and predicted fear based on a linear model with height
and height squared

for large heights. For average heights, we see on the contrary all data points
above the regression line. Somehow the data points do not suggest a completely
linear relationship, but a curved one.
This problem of model misfit could be solved by not only using height as
the predictor variable, but also the square of height, that is, height2 . For
each observed height we compute the square. This new variable, let’s call it
height2, we add to our regression model. The least squares regression equation
then becomes:

d = −2000 + 100 × height − 0.56 × height2

fear (7.5)

If we then plot the data and the regression line, we get Figure 7.17. There
we see that the regression line goes straight through the points. Note that the
regression line when plotted against height is non-linear, but equation 7.5 itself
is linear, that is, there are only two effects added up, one from variable height
and one from variable height2. We also see from the histogram (Figure 7.18)
and the residuals plot (Figure 7.19) that the residuals are randomly drawn from
a normal distribution and are not related to predicted fear. Thus, our additive
model (our linear model) with effects of height and height squared results in a
nice-fitting model with random normally scattered residuals.
In sum, the relationship between two variables need not be linear in order for
a linear model to be appropriate. A transformation of an independent variable,
such as taking a square, can result in normally randomly scattered residuals.
The linearity assumption is that the effects of a number of variables (transformed
or untransformed) add up and lead to a model with normally and independently,
randomly scattered residuals.

227
 8

6
count

0
−20 −10 0 10 20
residuals

Figure 7.18: Histogram of the residuals of the fear of snakes data with height
squared introduced into the linear model.

20

10
residual

−10

−20

500 1000 1500 2000

predicted fear

Figure 7.19: Residuals plot of the fear of snakes data with height squared
introduced into the linear model.

228
 9
Reaction time in seconds

20 40 60 80
Age in years

Figure 7.20: Least squares regression line for reaction time on age in 100 adults.

7.4 Equal variances

Suppose we measure reaction times in both young and older adults. Older
persons tend to have longer reaction times than young adults. Figure 7.20
shows a data set on 100 persons. Figure 7.21 shows the residuals as a function
of age, and shows something remarkable: it seems that the residuals are much
more varied for older people than for young people. There is more variance
at older ages than at younger ages. This is a violation of the equal variance
assumption. Remember that a linear model goes with a normal distribution for
the residuals with a certain variance. In a linear model, there is only mention
of one variance of the residuals σ 2 , not several!
The equal variance assumption is an important one: if the data show that
the variance is different for different subgroups of individuals in the data set,
then the standard errors of the regression coefficients cannot be trusted.
We often see an equal variance violation in reaction times. An often used
strategy of getting rid of such a problem is to work not with the reaction time,
but the logarithm of the reaction time. Figure 7.22 shows the data with the
computed logarithms of reaction time, and Figure 7.23 shows the residuals plot.
You can see that the log-transformation of the reaction times resulted in a much
better model.
Note that the assumption is not about the variance in the sample data, but
about the residuals in the population data. It might well be that there are slight
differences in the sample data of the older people than in the sample data of
the younger people. These could well be due to chance. The important thing to
know is that the assumption of equal variance is that in the population of older
adults, the variation in residuals is the same as the variation in residuals in the
population of younger adults.
The equal variance assumption is often referred to as the homogeneity of

229
 7.5

5.0
residuals

2.5

0.0

−2.5
20 40 60 80
Age in years

Figure 7.21: Residual plot after regressing reaction time on age.

2
Logarithm of reaction time

20 40 60 80
Age in years

Figure 7.22: Least squares regression line for log reaction time on age in 100
adults.

230
 1.5

1.0

0.5
residual

0.0

−0.5

−1.0

0.25 0.50 0.75 1.00 1.25

predicted log reaction time

Figure 7.23: Residual plot after regressing log reaction time on age.

variance assumption or homoscedasticity. It is the assumption that variance is

homogeneous (of equal size) across all levels and subgroups of the independent
variables in the population. The computation of the standard error is highly
dependent on the size of the variance of the residuals. If the size of this variance
differs across levels and subgroups of the data, the standard error also varies and
the confidence intervals cannot be easily determined. This in turn has an effect
on the computation of p-values, and therefore inference. Having no homogeneity
of variance therefore leads to wrong inference, with inflated or deflated type I
and type II error rates.
The inflation or deflation of type I and type II error rates are limited in the
case that group sizes are more or less equal. For example, suppose you have an
age variable with about an equal number of older persons and younger persons,
but unequal variances of the residuals. In that case you should not worry
too much about the precision of your p-values and your confidence intervals:
they are more or less correct. However, if you have more than 1.5 times more
elderly in your sample than youngsters (or vice versa), with unequal variances
of the residuals, then you should worry. Briefly: if the greater error variance is
associated with the greater group size, then the reported p-value is too small,
and if the greater error variance is associated with the smaller group size, then
the reported p-value is too large. If the p-value is around your pre-chosen α-level
and you’re unsure whether to reject or not to reject your null-hypothesis, look
for more robust methods of computing standard errors.

7.5 Residuals normally distributed

As we’ve already seen, the assumption of the linear model is that the residuals
are normally distributed. Let’s look at the reaction time data again and see

231
 9
count

0
−2.5 0.0 2.5 5.0 7.5
residuals

Figure 7.24: Histogram of the residuals after a regression of reaction time on

age.

what the histogram of the residuals looks like if we use reaction time as our
dependent variable. Figure 7.24 shows that in that case the distribution is not
symmetric: it is clearly skewed.
After a logarithmic transformation of the reaction times, we get the histogram
in Figure 7.25, which looks more symmetric.
Remember that if your sample size is of limited size, a distribution will
never look completely normal, even if it is sampled from a normal distribution.
It should however be likely to be sampled from a population of data that seems
normal. That means that the histogram should not be too skewed, or too
peaked, or have two peaks far apart. Only if you have a lot of observations, say
1000, you can reasonably say something about the shape of the distribution.
If you have categorical independent variables in your linear model, it is best
to look at the various subgroups separately and look at the histogram of the
residuals: the residuals e are defined as residuals given the rest of the linear
model. For instance, if there is a model for height, and country is the only
predictor in the model, all individuals from the same country are given the same
expected height based on the model. They only differ from each other because of
the normally distributed random residuals. Therefore look at the residuals for all
individuals from one particular country to see whether the residuals are indeed
normally distributed. Then do this for all countries separately. Think about
it: the residuals might look non-normal from country A, and non-normal from
country B, but put together, they might look very normal! This is illustrated
in Figure 7.26. Therefore, when checking for the assumption of normality, do
this for every subgroup separately.
It should be noted that the assumption of normally distributed residuals as
checked with a histogram is the least important assumption. Even when the

232
 15
count

10

0
−1 0 1
residuals

Figure 7.25: Histogram of the residuals after a regression of log reaction time
on age.

distribution is skewed, your standard errors are more or less correct. Only in
severe cases, like with the residuals in Figure 7.24, the standard errors start to
be somewhat incorrect.

7.6 General approach to testing assumptions

It is generally advised to always check the residuals. All four assumptions
mentioned above can be checked by looking at the residuals. We advise to do
this with three types of plots.
The first is the histogram of the residuals: this shows whether the residuals
are more or less normally distributed. The histogram should show a more or
less symmetric distribution. If the plot does not look asymmetric at all, try to
find a transformation of the dependent variable that makes the residuals more
normal. An example of this is to log-transform reaction times.
The second type of plot that you should look at is a plot where the residuals
are on the y-axis and the predicted values for the dependent variable (Yb ) is on
the x-axis. Such a plot can reveal systematic deviation from normality, but also
non-equal variance.
The third type of plot that you should study is one where the residuals are
on the vertical axis and one of the predictor variables is on the horizontal axis.
In this plot, you can spot violations of the equal variance assumption. You can
also use such a plot for candidate predictor variables that are not in your model
yet. If you notice a pattern, this is indicative of dependence, which means that
this variable should probably be included in your model.

233
 400

300
country
count

A
200
B

100

0
−10 0 10
residual

Figure 7.26: Two distributions might be very non-normal, but when taken
together, might look normal nevertheless. Normality should therefore always
be checked for each subgroup separately.

7.7 Checking assumptions in R

In this section we show the general code for making residual plots in R. We will
look at how to make the three types of plots of the residuals to check the four
assumptions.
When you run a linear model with the lm() function, you can use the package
modelr to easily obtain the residuals and predicted values that you need for your
plots. Let’s use the mpg data to illustrate the general approach. This data set
contains data on 234 cars. First we model the number of city miles per gallon
(cty) as a function of the number of cylinders (cyl).

out <- mpg %>%

lm(cty ~ cyl, data = .)

Next, we use the function add residuals from the modelr package to add
residuals to the data set and plot a histogram.

library(modelr)
mpg %>%
add_residuals(out) %>%
ggplot(aes(x = resid)) +
geom_histogram()

234
 30

20
count

10

0
−5 0 5 10 15
resid

As stated earlier, it’s even better to do this for the different subgroups
separately:

mpg %>%
add_residuals(out) %>%
ggplot(aes(x = resid)) +
geom_histogram() +
facet_wrap(. ~ cyl)

4 5
20

15

10

0
count

6 8
20

15

10

0
−5 0 5 10 15 −5 0 5 10 15
resid

For the second type of plot, we use two functions from the modelr package
to add predicted values and residuals to the data set, and use these to make a
residual plot:

mpg %>%
add_residuals(out) %>%
add_predictions(out) %>%
ggplot(aes(x = pred, y = resid)) +

235
 geom_point()

15

10
resid

−5

12 14 16 18 20
pred

When there are few values for the predictions, or when you have a categorical
predictor, it’s better to make a boxplot:

mpg %>%
add_residuals(out) %>%
add_predictions(out) %>%
ggplot(aes(x = factor(pred), y = resid)) +
geom_boxplot()

15

10
resid

−5

12.3665640272324 16.6225317099353 18.7505155512867 20.8784993926382

factor(pred)

For the third type of plot, we put the predictor on the x-axis and the residual
on the y-axis.

mpg %>%
add_residuals(out) %>%

236
 ggplot(aes(x = cyl, y = resid)) +
geom_point()

15

10
resid

−5

4 5 6 7 8
cyl

Again, with categorical variables or variables with very few categories, it is

sometimes clearer to use a boxplot:

mpg %>%
add_residuals(out) %>%
ggplot(aes(x = factor(cyl), y = resid)) +
geom_boxplot()

15

10
resid

−5

4 5 6 8
factor(cyl)

To check for independence you can also put variables on the x-axis that are
not in the model yet, for example the type of the car (class):

mpg %>%
add_residuals(out) %>%

237
ggplot(aes(x = class, y = resid)) +
geom_boxplot()

15

10
resid

−5

2seater compact midsize minivan pickup subcompact suv

class

238
Chapter 8

When assumptions are not

met: non-parametric
alternatives

8.1 Introduction
Linear models do not apply to every data set. As discussed in Chapter 7,
sometimes the assumptions of linear models are not met. One of the assumptions
is linearity or additivity. Additivity requires that one unit change in variable
X leads to the same amount of change in Y , no matter what value X has.
For bivariate relationships this leads to a linear shape. But sometimes you can
only expect that Y will change in the same direction, but you don’t believe
that this amount is the same for all values of X. This is the case for example
with an ordinal dependent variable. Suppose we wish to model the relationship
between the age of a mother and an aggression score in her 7-year-old child.
Suppose aggression is measured on a three-point ordinal scale: ’not aggressive’,
’sometimes aggressive’, ’often aggressive’. Since we do not know the quantitative
differences between these three levels, there are many graphs we could draw for
a given data set.
Suppose we have the data set given in Table 8.1. If we want to make a scatter
plot, we could arbitrarily choose the values 1, 2, and 3 for the three categories,
respectively. We would then get the plot in Figure 8.1. But since the aggression
data are ordinal, we could also choose the arbitrary numeric values 0, 2, and 3,
which would yield the plot in Figure 8.2.
As you can see from the least squares regression lines in Figures 8.1 and 8.2,
when we change the way in which we code the ordinal variable into a numeric
one, we also see the best fitting regression line changing. This does not mean
though, that ordinal data cannot be modelled linearly. Look at the example data
in Table 8.2 where aggression is measured with a 7-point scale. Plotting these

239
 4

3
Aggression

0
30.0 30.5 31.0 31.5 32.0 32.5
AgeMother

Figure 8.1: Regression of the child’s aggression score (1,2,3) on the mother’s
age.

3
Aggression2

0
30.0 30.5 31.0 31.5 32.0 32.5
AgeMother

Figure 8.2: Regression of the child’s aggression score (0,2,3) on the mother’s
age.

240
 Table 8.1: Aggression in children and age of the mother.
AgeMother Aggression
32.00 Sometimes aggressive
31.00 Often aggressive
32.00 Often aggressive
30.00 Not aggressive
31.00 Sometimes aggressive
30.00 Sometimes aggressive
31.00 Not aggressive
31.00 Often aggressive
31.00 Not aggressive
30.00 Sometimes aggressive
32.00 Often aggressive
32.00 Often aggressive
31.00 Sometimes aggressive
30.00 Sometimes aggressive
31.00 Not aggressive

data in Figure 8.3 using the values 1 through 7, we see a nice linear relationship.
So even when the values 1 thru 7 are arbitrarily chosen, a linear model can be
a good model for a given data set with one or more ordinal variables. Whether
the interpretation makes sense is however up to the researcher.
So with ordinal data, always check that your data indeed conform to a linear
model, but realise at the same time that you’re assuming a quantitative and
additive relationship between the variables that may or may not make sense. If
you believe that a quantitative analysis is meaningless then you may consider a
non-parametric analysis that we discuss in this chapter.
Another instance where we favour a non-parametric analysis over a linear
model one, is when the assumption of normally distributed residuals is not
tenable. For instance, look again at Figure 8.1 where we regressed aggression
in the child on the age of its mother. Figure 8.4 shows a histogram of the
residuals. Because of the limited number of possible values in the dependent
variable (1, 2 and 3), the number of possible values for the residuals is also
very restricted, which leads to a very discrete distribution. The histogram looks
therefore far removed from a continuous symmetric, bell-shaped distribution,
which is a violation of the normality assumption.
Every time we see a distribution of residuals that is either very skew, or has
very few different values, we should consider a non-parametric analysis. Note
that the shape of the distribution of the residuals is directly related to what scale
values we choose for the ordinal categories. By changing the values we change
the regression line, and that directly affects the relative sizes of the residuals.
First, we will discuss a non-parametric alternative for two numeric variables.
We will start with Spearman’s ρ (rho, pronouned ’row’), also called Spearman’s
rank-order correlation coefficient rs . Next we will discuss an alternative to rs ,

241
 8

6
Aggression

0
27.5 30.0 32.5 35.0 37.5
AgeMother

Figure 8.3: Regression of the child’s aggression 1 thru 7 Likert score on the
mother’s age.

4
count

0
−1.0 −0.5 0.0 0.5 1.0
residual

Figure 8.4: Histogram of the residuals after the regression of a child’s aggression
score on the mother’s age.

242
Table 8.2: Aggression in children on a 7-point Likert scale and age of the mother.
AgeMother Aggression
35.0 6
32.0 4
35.2 6
36.0 5
32.9 3
29.9 1
32.3 4
32.2 2
34.2 4
30.5 2
31.6 3
30.5 2
31.7 3
31.4 3
37.5 7

Kendall’s τ (tau, pronounced ’taw’). After that, we will discuss the combination
of numeric and categorical variables, when comparing groups.

8.2 Spearman’s ρ (rho)

Suppose we have 10 students and we ask their teachers to rate them on their
performance. One teacher rates them on geography and the other teacher rates
them on history. We only ask them to give rankings: indicate the brightest
student with a 1 and the dullest student with a 10. Then we might see the
data set in Table 8.3. We see that student 9 is the brightest student in both
geography and history, and student 7 is the dullest student in both subjects.

Table 8.3: Student rankings on geography and history.

student rank.geography rank.history
1 5 4
2 4 5
3 6 7
4 7 8
5 8 6
6 9 9
7 10 10
8 2 3
9 1 1
10 3 2

243
 Now we acknowledge the ordinal nature of the data by only having rankings:
a person with rank 1 is brighter than a person with rank 2, but we do not how
large the difference in brightness really is. Now we want to establish to what
extent there is a relationship between rankings on geography and the rankings
on history: the higher the ranking on geography, the higher the ranking on
history?
By eye-balling the data, we see that the brightest student in geography is also
the brightest student in history (rank 1). We also see that the dullest student
in history is also the dullest student in geography (rank 10). Furthermore, we
see relatively small differences between the rankings on the two subjects: high
rankings on geography seem to go together with high rankings on history. Let’s
look at these differences between rankings more closely by computing them, see
Table 8.4.

Table 8.4: Student rankings on geography and history.

student rank.geography rank.history difference
1 5 4 -1
2 4 5 1
3 6 7 1
4 7 8 1
5 8 6 -2
6 9 9 0
7 10 10 0
8 2 3 1
9 1 1 0
10 3 2 -1

So theoretically the difference could be as large as 9, but here we see a

biggest difference of -2. When all differences are small, this says something
about how the two rankings overlap: they are related. We could compute an
average difference: the average difference is the sum of these differences, divided
by 10, so we get 0. This is because we have both plus and minus values. It would
be better to take the square of the differences, so that we would get positive
values, see Table 8.5.
Now we can compute the average squared difference, which is equal to 10/10
= 1. Generally, the smaller this value, the closer the rankings of the two teachers
are together, and the more correlation there is between the two subjects.
A clever mathematician like Spearman showed that is even better to use a
somewhat different measure for a correlation between ranks. He showed that it
is wiser to compute the following statistic, where d is the difference in rank and
d2 is the squared difference (and n is sample size):

6 d2
P
rs = 1 − 3 (8.1)
n −n

244
 Table 8.5: Student rankings on geography and history.
rank.geography rank.history difference squared.difference
5 4 -1 1
4 5 1 1
6 7 1 1
7 8 1 1
8 6 -2 4
9 9 0 0
10 10 0 0
2 3 1 1
1 1 0 0
3 2 -1 1

because then you get a value between -1 and 1, just like a Pearson correlation,
where a value close to 1 describes a high positive correlation (high rank on one
variable goes together with a high rank on the other variable) and a value close
to -1 describes a negative correlation (a high rank on one variable goes together
with a low rank on the other variable). So in our case the sum of the squared
differences is equal to 10, and n is the number of students, so we get:

6 × 10 60
rs = 1 − =1− = 0.94 (8.2)
103 − 10 990
This is called the Spearman rank-order correlation coefficient rs , or Spearman’s
rho (the Greek letter ρ). It can be used for any two variables of which at least
one is ordinal. The trick is to convert the scale values into ranks, and then
apply the formula above. For instance, if we have the variable Grade with the
following values (C, B, D, A, F), we convert them into rankings by saying the A
is the highest value (1), B is the second highest value (2), C is the third highest
value (3), D is the fourth highest value (4) and F is the fifth highest value (5).
So transformed into ranks we get (3, 2, 4, 1, 5). Similarly, we could turn numeric
variables into ranks. Table 8.6 shows how the variables grade, shoesize and
height are transformed into their respective ranked versions. Note that the
ranking is alphanumerically by default: the first alphanumeric value gets rank
1. You could also do the ranking in the opposite direction, if that makes more
sense.

Table 8.6: Ordinal and numeric variables and their ranked transformations.
student grade rank.grade shoesize rank.shoesize height rank.height
1 A 1 6 1 1.70 1
2 D 4 8 3 1.82 2
3 C 3 9 4 1.92 4
4 B 2 7 2 1.88 3

245
8.3 Spearman’s rho in R
When we let R compute rs for us, it automatically ranks the data for us. Let’s
look at the mpg data on 234 cars from the ggplot2 package again. Suppose we
want to treat the variables cyl (the number of cylinders) and year (year of the
model) as ordinal variables, and we want to look whether the ranking on the
cyl variable is related to the ranking on the year variable. We use the function
rcorr() from the Hmisc package to compute Pearson’s rho:

library(Hmisc)
rcorr(mpg$cty, mpg$year, type = "spearman")

## x y
## x 1.00 -0.01
## y -0.01 1.00
##
## n= 234
##
##
## P
## x y
## x 0.9169
## y 0.9169

In the output you will see a correlation matrix very similar the one for a
Pearson correlation. Spearman’s rho is equal -0.01. You will also see whether
the correlation is significantly different from 0, indicated by a p-value. If the
p-value is very small, you may conclude that on the basis of these data, the
correlation in the population is not equal to 0, ergo, in the population there is a
relationship between the year a car was produced and the number of cylinders.

8.4 Kendall’s rank-order correlation coefficient

τ
If you want to study the relationship between two variables, of which at least
one is ordinal, you can either use Spearman’s rs or Kendall’s τ (tau, pronounced
’taw’ as in ’law’). However, if you have three variables, and you want to know
whether there is a relationship between variables A and B, over and above the
effect of variable C, you can use an extension of Kendall’s τ . Note that this is
very similar to the idea of multiple regression: a coefficient for variable X1 in
multiple regression with two predictors is the effect of X1 on Y over and above
the effect of X2 on Y . The logic of Kendall’s τ is also based on rank orderings,
but it involves a different computation. Let’s look at the student data again
with the teachers’ rankings of ten students on two subjects in Table 8.4.

246
Table 8.7: Student rankings on geography and history, now ordered according
to the ranking for geography.
student rank.geography rank.history
9 1 1
8 2 3
10 3 2
2 4 5
1 5 4
3 6 7
4 7 8
5 8 6
6 9 9
7 10 10

From this table we see that the history teacher disagrees with the geography
teacher that student 8 is brighter than student 10. She also disagrees with her
colleague that student 1 is brighter than student 2. If we do this for all possible
pairs of students, we can count the number of times that they agree and we can
count the number of times they disagree. The total number of possible pairs is
equal to 10


2 = n(n−1)/2 = 90/2 = 45 (see Chapter 3). This is a rather tedious
job to do, but it can be made simpler if we reshuffle the data a bit. We put
the students in a new order, such that the brightest student in geography comes
first, and the dullest last. This also changes the order in the variable history. We
then get the data in Table 8.7. We see that the geography teacher believes that
student 9 outperforms all 9 other students. On this, the history teacher agrees,
as she also ranks student 9 first. This gives us 9 agreements. Moving down
the list, we see that the geography teacher believes student 8 outperforms 8
other students. However, we see that the history teacher believes student 8 only
outperforms 7 other students. This results in 7 agreements and 1 disagreement.
So now in total we have 9 + 7 = 16 agreements and 1 disagreements. If we
go down the whole list in the same way, we will find that there are in total 41
agreements and 4 disagreements.
The computation is rather tedious. There is a trick to do it faster. Now
focus on Table 8.7 but start in the column of the history teacher. Start at the
top row and count the number of rows beneath it with a rank higher than the
rank in the first row. The rank in the first row is 1, and all other ranks beneath
it are higher, so the number of ranks is 9. We plug that value in the last column
in Table 8.8. Next we move to row 2. The rank is 3. We count the number of
rows below row 2 with a rank higher than 3. Rank 2 is lower, so we are left
with 7 rows and we again plug 7 in the last column of Table 8.8. Then we move
on to row 3, with rank 2. There are 7 rows left, and all of them have a higher
rank. So the number is 7. Then we move on to row 4. It has rank 5. Of the
6 rows below it, only 5 have a higher rank. Next, row 5 shows rank 4. Of the
5 rows below it, all 5 show a higher rank. Row 6 shows rank 7. Of the 4 rows

247
Table 8.8: Student rankings on geography and history, now ordered according
to the ranking for geography, with number of agreements.
student rank.geography rank.history number
9 1 1 9
8 2 3 7
10 3 2 7
2 4 5 5
1 5 4 5
3 6 7 3
4 7 8 2
5 8 6 2
6 9 9 1
7 10 10 0

below it, only 3 show a higher rank. Row 7 shows rank 8. Of the three rows
below it, only 2 show a higher rank. Row 8 shows rank 6. Both rows below it
show a higher rank. And row 9 shows rank 9, and the row below it shows a
higher rank so that is 1. Finally, when we add up the values in the last column
in Table 8.8, we find 41. This is the number of agreements. The number of
disagreements can be found by reasoning that the total number of pairs equals

the number of pairs that can be formed using a total number of 10 objects: 102
= 10(10 − 1)/2 = 45. In this case we have 45 possible pairs. Of these there are
41 agreements, so there must be 45 − 41 = 4 disagreements. We can then fill in
the formula to compute Kendall’s τ :

agreements − disagreements 37
τ= = = 0.82 (8.3)
totalnumberof pairs 45

This τ -statistic varies between -1 and 1 and can therefore be seen as a

non-parametric analogue of a Pearson correlation. Here, the teachers more
often agree than disagree, and therefore the correlation is positive. A negative
correlation means that the teachers more often disagree than agree on the
relative brightness of their students.
As said, the advantage of Kendall’s τ over Spearman’s rs is that Kendall’s τ
can be extended to cover the case that you wish to establish the strength of the
relationships of two variables A and B, over and above the relationship with C.
The next section shows how to do that in R.

8.5 Kendall’s τ in R
Let’s again use the mpg data on 234 cars. We can compute Kendall’s τ for the
variables cyl and year using the Kendall package:

248
library(Kendall)
Kendall(mpg$cyl, mpg$year)

## tau = 0.112, 2-sided pvalue =0.068798

As said, Kendall’s τ can also be used if you want to control for a third
variable (or even more variables). This can be done with the ppcor package.
Because this package has its own function select(), you need to be explicit
about which function from which package you want to use. Here you want to
use the select() function from the dplyr package (part of the tidyverse suite
of packages).

library(ppcor)
mpg %>%
dplyr::select(cyl, year, cty) %>%
pcor(method = "kendall")

## $estimate
## cyl year cty
## cyl 1.0000000 0.1642373 -0.7599993
## year 0.1642373 1.0000000 0.1210952
## cty -0.7599993 0.1210952 1.0000000
##
## $p.value
## cyl
## cyl 0.000000000000000000000000000000000000000000000000000000000000000000000000
## year 0.000189654827628346264994235736978112072392832487821578979492187500000000
## cty 0.000000000000000000000000000000000000000000000000000000000000000000770657
## year
## cyl 0.0001896548
## year 0.0000000000
## cty 0.0059236967
## cty
## cyl 0.000000000000000000000000000000000000000000000000000000000000000000770657
## year 0.005923696652933938683327497187747212592512369155883789062500000000000000
## cty 0.000000000000000000000000000000000000000000000000000000000000000000000000
##
## $statistic
## cyl year cty
## cyl 0.000000 3.732412 -17.271535
## year 3.732412 0.000000 2.751975
## cty -17.271535 2.751975 0.000000
##
## $n
## [1] 234
##

249
## $gp
## [1] 1
##
## $method
## [1] "kendall"

In the output, we see that the Kendall correlation between cyl and year,
controlled for cty, equals 0.16, with an associated p-value of 0.00019.

8.6 Kruskal-Wallis test for group comparisons

Now that we have discussed relationships between ordinal and numeric variables,
let’s have a look at the case where we also have categorical variables.
Suppose we have three groups of students that go on a field trip together:
mathematicians, psychologists and engineers. Each can pick a rain coat, with
five possible sizes: ’extra small’, ’small’, ’medium’, ’large’ or ’extra large’. We
want to know if preferred size is different in the three populations, so that
teachers can be better prepared in the future. Now we have information about
size, but this knowledge is not numeric: we do not know the difference in size
between ’medium’ and ’large’, only that ’large’ is larger than ’medium.’ We have
ordinal data, so computing a mean is impossible here. Even if we would assign
values like 1= ’extra small’, 2=’small’, 3= ’medium’, etcetera, the mean would
be rather meaningless as these values are arbitrary. So instead of focussing on
means, we can focus on medians: the middle values. For instance, the median
value for our sample of mathematicians could be ’medium’, for our sample of
psychologists ’small’, and for our sample of engineers ’large.’ Our question might
then be whether the median values in the three populations are really different.
This can be assessed using the Kruskal-Wallis test. Similar to Spearman’s
rs and Kendall’s τ , the data are transformed into ranks. This is done for all
data at once, so for all students together.
For example, if we had the data in Table 8.6, we could transform the variable
size into ranks, from smallest to largest. Student 1 has size ’extra small’ so he
or she gets the rank 1. Next, both student 4 and student 6 have size ’small’, so
they should get ranks 2 and 3. However, because there is no reason to prefer
one student over the other, we give them both the mean of ranks 2 and 3, so
they both get the rank 2.5. Next in line is student 3 with size ’medium’ and
(s)he gets rank 4. Next in line is student 5 with size ’large’ (rank 5) and last in
line is student 2 with size ’extra large’ (rank 6).
Next, we could compute the average rank per group. The group with the
smallest sizes would have the lowest average rank, etcetera. Under the null-
hypothesis, if the distribution of size were the same in all three groups, the
average ranks would be about the same. If the average rank is very different
across groups, this is an indication that size is not distributed equally among
the three groups. In order to have a proper statistical test, a rather complex
formula is used to compute the so-called KW -statistic, see Castellan & Siegel

250
 Table 8.9: Field trip data.
student group size rank
001 math extra small 1
002 math extra large 6
003 psych medium 4
004 psych small 2.5
005 engineer large 5
006 math small 2.5

35

30

25
cty

20

15

10

2seater compact midsize minivan pickup subcompact suv

class

Figure 8.5: Distributions of city mileage (city) as a function of car type (class).

(1988), that you don’t need to know. The distribution of this KW -statistic
under the null-hypothesis is known, so we know what extreme values are, and
consequently can compute p-values. This tedious computation can be done in
R.

8.7 Kruskal-Wallis test in R

Let’s look at the mpg data again. It contains data on cars from 7 different
types. Suppose we want to know whether the different types show difference
in the median city miles per gallon. The medians are plotted in Figure 8.5: it
seems that indeed the distributions of cty are very different for different car
types. But these differences could be due to sampling: maybe by accident, the
pick-up trucks in this sample happened to have a relatively low mileage, and
that the differences in the population of all cars are non-existing. To test this
null-hypothesis, we run the following R code:

251
mpg %>%
kruskal.test(cty ~ class, data = .)

##
## Kruskal-Wallis rank sum test
##
## data: cty by class
## Kruskal-Wallis chi-squared = 149.53, df = 6, p-value <
## 0.00000000000000022

The output gives us the ability to give the following report:

”The null-hypothesis that city miles per gallon is distributed equally

for all types of cars was tested using a Kruskal-Wallis test with an α
of 0.05. Results showed that the null-hypothesis could be rejected,
X 2 = 149.53, df = 6, p < 0.0001.”

252
Chapter 9

Moderation: testing
interaction effects

9.1 Interaction with one numeric and one dichotomous

variable
Suppose there is a linear relationship between age (in years) and vocabulary
(the number of words one knows): the older you get, the more words you know.
Suppose we have the following linear regression equation for this relationship:

d = 205 + 500 × age

vocab (9.1)

According to this equation, the expected number of words for a newborn

baby (age = 0) equals 205. This may sound silly, but suppose this model is a
very good prediction model for vocabulary size in children between 2 and 5 years
of age. Then this equation tells us that the expected increase in vocabulary size
is 500 words per year.
This model is meant for everybody in the Netherlands. But suppose that
one researcher expects that the increase in words is much faster in children
from high socio-economic status (SES) families than in children from low SES
families. He believes that vocabulary will be larger in higher SES children than
in low SES children. In other words, he expects an effect of SES, over and above
the effect of age:

d = b0 + b1 × age + b2 × SES
vocab (9.2)

This main effect of SES is yet unknown and denoted by b2 . Note that this
linear equation is an example of multiple regression.
Let’s use some numerical example. Suppose age is coded in years, and SES
is dummy coded, with a 1 for high SES and a 0 for low SES. Let b2 , the effect

253
 3000

Vocabulary #words
2000 SES=1

SES=0

1000

0
0 1 2 3 4 5
Age in years

Figure 9.1: Two regression lines: one for low SES children and one for high SES
children.

of SES over and above age, be 10. Then we can write out the linear equation
for low SES and high SES separately.

lowSES : vocab
d = 200 + 500 × age + 10 × 0 (9.3)
= 200 + 500 × age (9.4)
highSES : vocab
d = 200 + 500 × age + 10 × 1 (9.5)
= (200 + 10) + 500 × age (9.6)
= 210 + 500 × age (9.7)
Figure 9.1 depicts the two regression lines for the high and low SES children
separately. We see that the effect of SES involves a change in the intercept:
the intercept equals 200 for low SES children and the intercept for high SES
children equals 210. The difference in intercept is indicated by the coefficient
for SES. Note that the two regression lines are parallel: for every age, the
difference between the two lines is equal to 10. For every age therefore, the
predicted number of words is 10 words more for high SES children than for low
SES children.
So far, this is an example of multiple regression that we already saw in
Chapter 4. But suppose that such a model does not describe the data that we
actually have, or does not make the right predictions based on on our theories.
Suppose our researcher also expects that the yearly increase in vocabulary is a
bit lower than 500 words in low SES families, and a little bit higher than 500
words in high SES families. In other words, he believes that SES might moderate
(affect or change) the slope coefficient for age. Let’s call the slope coefficient
in this case b1 . In the above equation this slope parameter is equal to 500, but
let’s now let itself have a linear relationship with SES:

254
 b1 = a + b3 × SES (9.8)

In words: the slope coefficient for the regression of vocab on age, is itself
linearly related to SES: we predict the slope on the basis of SES. We model
that by including a slope b3 , but also an intercept a. Now we have two linear
equations for the relationship between vocab, age and SES:

vocab
d = b0 + b1 × age + b2 × SES (9.9)
b1 = a + b3 × SES (9.10)

We can rewrite this by plugging the second equation into the first one
(substitution):

d = b0 + (a + b3 × SES) × age + b2 × SES

vocab (9.11)

Multiplying this out gets us:

d = b0 + a × age + b3 × SES × age + b2 × SES

vocab (9.12)

If we rearrange the terms a bit, we get:

d = b0 + a × age + b2 × SES + b3 × SES × age

vocab (9.13)

Now this very much looks like a regression equation with one intercept and
three slope coefficients: one for age (a), one for SES (b2 ) and one for SES × age
(b3 ).
We might want to change the label a into b1 to get a more familiar looking
form:

d = b0 + b1 × age + b2 × SES + b3 × SES × age

vocab (9.14)

So the first slope coefficient is the increase in vocabulary for every year that
age increases (b1 ), the second slope coefficient is the increase in vocabulary for
an increase of 1 on the SES variable (b2 ), and the third slope coefficient is the
increase in vocabulary for every increase of 1 on the product of SES and age
(b3 ).
What does this mean exactly?
Suppose we find the following parameter values for the regression equation:

d = 200 + 450 × age + 125 × SES + 100 × SES × age

vocab (9.15)

255
 If we code low SES children as SES = 0, and high SES children as SES = 1,
we can write the above equation into two regression equations, one for low SES
children (SES = 0) and one for high SES children (SES = 1):

lowSES : vocab
d = 200 + 450 × age (9.16)
highSES : vocab
d = 200 + 450 × age + 125 + 100 × age (9.17)
= (200 + 125) + (450 + 100) × age
= 325 + 550 × age

Then for low SES children, the intercept is 200 and the regression slope for
age is 450, so they learn 450 words per year. For high SES children, we see
the same intercept of 200, with an extra 125 (this is the main effect of SES).
So effectively their intercept is now 325. For the regression slope, we now have
450 × age + 100 × age which is of course equal to 550 × age. So we see that the
high SES group has both a different intercept, and a different slope: the increase
in vocabulary is 550 per year: somewhat steeper than in low SES children. So
yes, the researcher was right: vocabulary increase per year is faster in high SES
children than in low SES children.
These two different regression lines are depicted in Figure 9.2. It can be
clearly seen that the lines have two different intercepts and two different slopes.
That they have two different slopes can be seen from the fact that the lines are
not parallel. One has a slope of 450 words per year and the other has a slope of
550 words per year. This difference in slope of 100 is exactly the size of the slope
coefficient pertaining to the product SES × age, b3 . Thus, the interpretation of
the regression coefficient for a product of two variables is that it represents the
difference in slope.
The observation that the slope coefficient is different for different groups
is called an interaction effect, or interaction for short. Other words for this
phenomenon are modification and moderation. In this case, SES is called the
modifier variable: it modifies the relationship between age on vocabulary. Note
however that you could also interpret age as the modifier variable: the effect of
SES is larger for older children than for younger children. In the plot you see
that the difference between vocabulary for high and low SES children of age 6
is larger than it is for children of age 2.

9.2 Interaction effect with a dummy variable in

R
Let’s look at some example output for an R data set where we have a categorical
variable that is not dummy-coded yet. The data set is on chicks and their weight
during the first days of their lives. Weight is measured in grams. The chicks
were given one of four different diets. Here we use only the data from chicks on
two different diets 1 and 2. We select only the Diet 1 and 2 data. We store the

256
 4000

3000
Vocabulary #words

SES=1
2000

1000

SES=0

0
0 2 4 6 8
Age in years

Figure 9.2: Two regression lines for the relationship between age and vocab,
one for low SES children (SES = 0) and one for high SES children (SES = 1).

Diet 1 and 2 data under the name chick data. When we have a quick look at
the data with glimpse(), we see that Diet is a factor (<fct>).

chick_data <- ChickWeight %>%

filter(Diet == 1 | Diet == 2)

chick_data %>%
glimpse()

## Rows: 340
## Columns: 4
## $ weight <dbl> 42, 51, 59, 64, 76, 93, 106, 125, 149, 171, 199, 205, 40, 49...
## $ Time <dbl> 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 21, 0, 2, 4, 6, 8, 10...
## $ Chick <ord> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, ...
## $ Diet <fct> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...

The general regression of weight on Time is shown in Figure 9.3. This

regression line for the entire sample of chicks has a slope of around 8 grams
per day. Now we want to know whether this slope is the same for chicks in the
Diet 1 and Diet 2 groups, in other words, do chicks grow as fast with Diet 1
as with Diet 2? We might expect that Diet moderates the effect of Time on
weight. We use the following code to study this Diet × Time interaction effect,
by having R automatically create a dummy variable for the factor Diet. In the
model we specify that we want a main effect of Time, a main effect of Diet, and
an interaction effect of Time by Diet:

257
 300

Diet
200
weight

1
2

100

0 5 10 15 20
Time

Figure 9.3: The relationship between Time and weight in all chicks with either
Diet 1 or Diet 2.

out <- chick_data %>%

lm(weight ~ Time + Diet + Time:Diet, data = .)
out %>%
tidy(conf.int = TRUE)

## # A tibble: 4 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 30.9 4.50 6.88 2.95e-11 22.1 39.8
## 2 Time 6.84 0.361 19.0 4.89e-55 6.13 7.55
## 3 Diet2 -2.30 7.69 -0.299 7.65e- 1 -17.4 12.8
## 4 Time:Diet2 1.77 0.605 2.92 3.73e- 3 0.577 2.96

In the regression table, we see the effect of the numeric Time variable, which
has a slope of 6.84. For every increase of 1 in Time, there is a corresponding
expected increase of 6.84 grams in weight. Next, we see that R created a dummy
variable Diet2. That means this dummy codes 1 for Diet 2 and 0 for Diet 1.
From the output we see that if a chick gets Diet 2, its weight is -2.3 grams
heavier (that means, Diet 2 results in a lower weight).
Next, R created a dummy variable Time:Diet2, by multiplying the variables
Time and Diet2. Results show that this interaction effect is 1.77.
These results can be plugged into the following regression equation:

d = 30.93 + 6.84 × Time − 2.3 × Diet2 + 1.77 × Time × Diet2 (9.18)

weight
If we fill in 1s for the Diet2 dummy variable, we get the equation for chicks
with Diet 2:

258
 300

Diet
200
weight

1
2

100

0 5 10 15 20
Time

Figure 9.4: The relationship between Time and weight in chicks, separately for
Diet 1 and Diet 2.

weight
d = 30.93 + 6.84 × Time − 2.3 × 1 + 1.77 × Time × 1
= 28.63 + 8.61 × Time (9.19)

If we fill in 0s for the Diet2 dummy variable, we get the equation for chicks
with Diet 1:

weight
d = 30.93 + 6.84 × Time (9.20)

When comparing these two regression lines for chicks with Diet 1 and Diet
2, we see that the slope for Time is 1.77 steeper for Diet 2 chicks than for Diet
1 chicks. In this particular random sample of chicks, the chicks on Diet 1 grow
6.84 grams per day (on average), but chicks on Diet 2 grow 6.84 + 1.77 = 8.61
grams per day (on average).
We visualised these results in Figure 9.4. There we see two regression lines:
one for the red data points (chicks on Diet 1) and one for the blue data points
(chicks on Diet 2). These two regression lines are the same as those regression
lines we found when filling in either 1s and 0s in the general linear model. Note
that the lines are not parallel, like in Chapter 6. Each regression line is the least
squares regression line for the subsample of chicks on a particular diet.
We see that the difference in slope is 1.77 grams per day. This is what we
observe in this particular sample of chicks. However, what does that tell us
about the difference in slope for chicks in general, that is, the population of all
chicks? For that, we need to look at the confidence interval. In the regression
table above, we also see the 95% confidence intervals for all model parameters.
The 95% confidence interval for the Time × Diet2 interaction effect is (0.58,

259
2.96). That means that plausible values for this interaction effect are those
values between 0.58 and 2.96.
It is also possible to do null-hypothesis testing for interaction effects. One
could test whether this difference of 1.77 is possible if the value in the entire
population of chicks equals 0 ? In other words, is the value of 1.77 significantly
different from 0?
The null-hypothesis is

H0 : βTime×Diet2 = 0 (9.21)

The regression table shows that the null-hypothesis for the interaction effect
has a t-value of t = 2.92, with a p-value of 3.73 × 10−3 = 0.00373. For research
reports one always also reports the degrees of freedom for a statistical test. The
(residual) degrees of freedom can be found in R by typing

out$df.residual

## [1] 336

We can report that

”we reject the null-hypothesis and conclude that there is evidence

that the Time × Diet2 interaction effect is not 0, t(336) = 2.92, p =
0.004.”

Summarising, in this section, we established that Diet moderates the effect

of Time on weight: we found a significant diet by time interaction effect. The
difference in growth rate is 1.77 grams per day, with a 95% confidence interval
from 0.58 to 2.96. In more natural English: diet has an effect on the growth
rate in chicks.
In this section we discussed the situation that regression slopes might be
different in two groups: the regression slope might be steeper in one group
than in the other group. So suppose that we had a numerical predictor X for
a numerical dependent variable Y , we said that a particular dummy variable
Z moderated the effect of X on Y . This moderation was quantified by an
interaction effect. So suppose we have the following linear equation:

Y = b0 + b1 × X + b2 × Z + b3 × X × Z + e

Then, we call b0 the intercept, b1 the main effect of X, b2 the main effect of
Z, and b3 the interaction effect of X and Z (alternatively, the X by Z interaction
effect).

260
9.3 Interaction effects with a categorical variable
in R
In the previous section, we looked at the difference in slopes between two groups.
But what we can do for two groups, we can do for multiple groups. The data
set on chicks contains data on chicks with 4 different diets. When we perform
the same analysis using all data in ChickWeight, we obtain the regression table

out <- ChickWeight %>%

lm(weight ~ Time + Diet + Time:Diet, data = .)
out %>%
tidy(conf.int = TRUE)

## # A tibble: 8 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 30.9 4.25 7.28 1.09e-12 22.6 39.3
## 2 Time 6.84 0.341 20.1 3.31e-68 6.17 7.51
## 3 Diet2 -2.30 7.27 -0.316 7.52e- 1 -16.6 12.0
## 4 Diet3 -12.7 7.27 -1.74 8.15e- 2 -27.0 1.59
## 5 Diet4 -0.139 7.29 -0.0191 9.85e- 1 -14.5 14.2
## 6 Time:Diet2 1.77 0.572 3.09 2.09e- 3 0.645 2.89
## 7 Time:Diet3 4.58 0.572 8.01 6.33e-15 3.46 5.70
## 8 Time:Diet4 2.87 0.578 4.97 8.92e- 7 1.74 4.01

The regression table for four diets is substantially larger than for two diets.
It contains one slope parameter for the numeric variable Time, three different
slopes for the factor variable Diet and three different interaction effects for the
Time by Diet interaction.
The full linear model equation is

d = 30.93 + 6.84 × Time − 2.3 × Diet2 − 12.68 × Diet3 − 0.14 × Diet4

weight
+1.77 × Time × Diet2 + 4.58 × Time × Diet3 + 2.87 × Time × Diet4
(9.22)

You see that R created dummy variables for Diet 2, Diet 3 and Diet 4. We
can use this equation to construct a separate linear model for the Diet 1 data.
Chicks with Diet 1 have 0s for the dummy variables Diet2, Diet3 and Diet4.
If we fill in these 0s, we obtain

d = 30.93 + 6.84 × Time

weight (9.23)

For the chicks on Diet 2, we have 1s for the dummy variable Diet2 and 0s
for the other dummy variables. Hence we have

261
 d = 30.93 + 6.84 × Time − 2.3 × 1 + 1.77 × Time × 1
weight
= 30.93 + 6.84 × Time − 2.3 + 1.77 × Time
= (30.93 − 2.3) + (6.84 + 1.77) × Time
= 28.63 + 8.61 × Time (9.24)

Here we see exactly the same equation for Diet 2 as in the previous section
where we only analysed two diet groups. The difference between the two slopes
in the Diet 1 and Diet 2 groups is again 1.77. The only difference for this
interaction effect is the standard error, and therefore the confidence interval is
also slightly different. We will come back to this issue in Chapter ??.
For the chicks on Diet 3, we have 1s for the dummy variable Diet3 and 0s
for the other dummy variables. The regression equation is then

d = 30.93 + 6.84 × Time − 12.68 × 1 + 4.58 × Time × 1

weight
= (30.93 − 12.68) + (6.84 + 4.58) × Time
= 18.25 + 11.42 × Time (9.25)

We see that the intercept is again different than for the Diet 1 chicks. We
also see that the slope is different: it is now 4.58 steeper than for the Diet 1
chicks. This difference in slopes is exactly equal to the Time by Diet3 interaction
effect. This is also what we saw in the Diet 2 group. Therefore, we can say
that an interaction effect for a specific diet group says something about how
much steeper the slope is in that group, compared to the reference group. The
reference group is the group for which all the dummy variables are 0. Here, that
is the Diet 1 group.
Based on that knowledge, we can expect that the slope in the Diet 4 group
is equal to the slope in the reference group (6.84) plus the Time by Diet4
interaction effect, 2.87, so 9.71.
We can do the same for the intercept in the Diet 4 group. The intercept is
equal to the intercept in the reference group (30.93) plus the main effect of the
Diet4 dummy variable, -0.14, which is 30.79.
The linear equation is then for the Diet 4 chicks:

d = 30.79 + 9.71 × Time

weight (9.26)
The four regression lines are displayed in Figure 9.5. The steepest regression
line is the one for the Diet 3 chicks: they are the fastest growing chicks. The
slowest growing chicks are those on Diet 1. The confidence intervals in the
regression table tell us that the difference between the growth rate with Diet 4
compared to Diet 1 is somewhere between 1.74 and 4.01 grams per day.
Suppose we want to test the null hypothesis that all four slopes are the same.
This implies that the Time by Diet interaction effects are all equal to 0. We

262
 300

Diet
1
weight

200 2
3
4

100

0 5 10 15 20
Time

Figure 9.5: Four different regression lines for the four different diet groups.

can test this null hypothesis

H0 : βTime×Diet2 = βTime×Diet2 = βTime×Diet4 = 0 (9.27)

by running an ANOVA. That is, we apply the anova() function to the results
of an lm() analysis:

out_anova <- out %>%

anova()
out_anova %>%
tidy()

## # A tibble: 4 x 6
## term df sumsq meansq statistic p.value
## <chr> <int> <dbl> <dbl> <dbl> <dbl>
## 1 Time 1 2042344. 2042344. 1760. 2.11e-176
## 2 Diet 3 129876. 43292. 37.3 5.07e- 22
## 3 Time:Diet 3 80804. 26935. 23.2 3.47e- 14
## 4 Residuals 570 661532. 1161. NA NA

In the output we see a Time by Diet interaction effect with 3 degrees of

freedom. That term refers to the null-hypothesis that all three interaction effects
are equal to 0. The F -statistic associated with that null-hypothesis equals 23.2.
The residual degrees of freedom equals 570, so that we can report:

”The slopes for the four different diets were significantly different
from each other, F (3, 570) = 23.2, M SE = 26935, p < 0.001.”

263
9.4 Interaction between two dichotomous variables
in R
In the previous section we discussed the situation that regression slopes might
be different in two four groups. In Chapter 6 we learned that we could also
look at slopes for dummy variables. The slope is then equal to the difference in
group means, that is, the slope is the increase in the group mean of one group
compared to the reference group.
Now we discuss the situation where we have two dummy variables, and want
to do inference on their interaction. Does one dummy variable moderate the
effect of the other dummy variable?
Let’s have a look at a data set on penguins. It can be found in the palmerpenguins
package.

# install.packages("palmerpenguins")
library(palmerpenguins)
penguins %>%
str ()

## tibble [344 x 8] (S3: tbl_df/tbl/data.frame)

## $ species : Factor w/ 3 levels "Adelie","Chinstrap",..: 1 1 1 1 1 1 1 1 1
## $ island : Factor w/ 3 levels "Biscoe","Dream",..: 3 3 3 3 3 3 3 3 3 3 ..
## $ bill_length_mm : num [1:344] 39.1 39.5 40.3 NA 36.7 39.3 38.9 39.2 34.1 42 ...
## $ bill_depth_mm : num [1:344] 18.7 17.4 18 NA 19.3 20.6 17.8 19.6 18.1 20.2 ...
## $ flipper_length_mm: int [1:344] 181 186 195 NA 193 190 181 195 193 190 ...
## $ body_mass_g : int [1:344] 3750 3800 3250 NA 3450 3650 3625 4675 3475 4250 ..
## $ sex : Factor w/ 2 levels "female","male": 2 1 1 NA 1 2 1 2 NA NA ...
## $ year : int [1:344] 2007 2007 2007 2007 2007 2007 2007 2007 2007 2007

We see there is a species factor with three levels, and a sex factor with two
levels. Let’s select only the Adelie and Chinstrap species.

penguin_data <- penguins %>%

filter(species %in% c("Adelie", "Chinstrap"))

Suppose that we are interested in differences in flipper length across species.

We then could run a linear model, with flipper length mm as the dependent
variable, and species as independent variable.

out <- penguin_data %>%

lm(flipper_length_mm ~ species, data = .)
out %>%
tidy(conf.int = TRUE)

## # A tibble: 2 x 7
## term estimate std.error statistic p.value conf.low conf.high

264
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 190. 0.548 347. 8.31e-300 189. 191.
## 2 speciesChinstrap 5.87 0.983 5.97 9.38e- 9 3.93 7.81

The output shows that in this sample, the Chinstrap penguins have on
average larger flippers than Adelie penguins. The confidence intervals tell us
that this difference in flipper length is somewhere between 3.93 and 7.81. But
suppose that this is not what we want to know. The real question might be
whether this difference is different for male and female penguins. Maybe there
is a larger difference in flipper length in females than in males?
This difference or change in one variable (flipper length mm) as a function
of another variable (sex) should remind us of moderation: maybe sex moderates
the effect of species on flipper length.
In order to study such moderation, we have to analyse the sex by species
interaction effect. By now you should know how to do that in R:

out <- penguin_data %>%

lm(flipper_length_mm ~ species + sex + species:sex, data = .)
out %>%
tidy(conf.int = TRUE)

## # A tibble: 4 x 7
## term estimate std.error statistic p.value conf.low conf.high
## <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 188. 0.707 266. 2.15e-267 186. 189.
## 2 speciesChinstrap 3.94 1.25 3.14 1.92e- 3 1.47 6.41
## 3 sexmale 4.62 1.00 4.62 6.76e- 6 2.65 6.59
## 4 speciesChinstrap:se~ 3.56 1.77 2.01 4.60e- 2 0.0638 7.06

In the output we see an intercept of 188. Next, we see an effect of a dummy

variable, coding 1s for Chinstrap penguins (speciesChinstrap). We also see
an effect of a dummy variable coding 1s for male penguins (sexmale). Then,
we see an interaction effect of these two dummy effects. That means that this
dummy variable codes 1s for the specific combination of Chinstrap penguins
that are male (speciesChinstrap:sexmale).

dlength = 188 + 3.94 × speciesChinstrap + 4.62 × sexmale + 3.56 × speciesChinstrap × sexmale

flipper (9.28)

From this we can make the following predictions. The predicted flipper
length for female Adelie penguins is

188 + 3.94 × 0 + 4.62 × 0 + 3.56 × 0 × 0 = 188 (9.29)

The predicted flipper length for male Adelie penguins is

265
 188 + 3.94 × 0 + 4.62 × 1 + 3.56 × 0 × 1
= 188 + 4.62 = 192.62
(9.30)

The predicted flipper length for female Chinstrap penguins is

188 + 3.94 × 1 + 4.62 × 0 + 3.56 × 0 × 0

= 188 + 3.94 = 191.94 (9.31)

and the predicted flipper length for male Chinstrap penguins is

188 + 3.94 × 1 + 4.62 × 1 + 3.56 × 1 × 1

= 188 + 3.94 + 4.62 + 3.56
= 200.12 (9.32)

These predicted flipper length for each male/species combination are actually
the group means. It is generally best to plot these means with a means and
errors plot. For that we first need to compute means by R. With left join()
we add these means to the data set. These diamant-shaped means (shape =
18) are plotted with intervals that are twice (mult = 2) the standard error of
those means (geom = "errorbar").

left_join(penguin_data, # adding group means to the data set

penguin_data %>%
group_by(species, sex) %>%
summarise(mean = mean(flipper_length_mm))
) %>%
ggplot(aes(x = sex, y = flipper_length_mm, colour = species)) +
geom_jitter(position = position_jitterdodge(), # the raw data
shape = 1,
alpha = 0.6) +
geom_point(aes(y = mean), # the groups means
position = position_jitterdodge(jitter.width = 0),
shape = 18,
size = 5) +
stat_summary(fun.data = mean_se, # computing errorbars
fun.args = list(mult = 2),
geom = "errorbar",
width = 0.2,
position = position_jitterdodge(jitter.width = 0),
size = 1) +
scale_color_brewer(palette = "Set1")

266
 210
flipper_length_mm

200

species
Adelie
190
Chinstrap

180

170
female male NA
sex

This plot shows also the data on penguins with unknown sex (sex = NA). If
we leave these out, we get

left_join(penguin_data, # adding group means to the data set

penguin_data %>%
group_by(species, sex) %>%
summarise(mean = mean(flipper_length_mm))
) %>%
filter(!is.na(sex)) %>%
ggplot(aes(x = sex, y = flipper_length_mm, colour = species)) +
geom_jitter(position = position_jitterdodge(), # the raw data
shape = 1,
alpha = 0.6) +
geom_point(aes(y = mean), # the groups means
position = position_jitterdodge(jitter.width = 0),
shape = 18,
size = 5) +
stat_summary(fun.data = mean_se, # computing errorbars
fun.args = list(mult = 2),
geom = "errorbar",
width = 0.2,
position = position_jitterdodge(jitter.width = 0),
size = 1) +
scale_color_brewer(palette = "Set1")

267
 210

flipper_length_mm
200

species
Adelie
190
Chinstrap

180

170
female male
sex

Comparing the Adelie and the Chinstrap data, we see that for both males
and females, the Adelie penguins have smaller flippers than the Chinstrap
penguins. Comparing males and females, we see that the males have generally
larger flippers than females. More interestingly in relation to this chapter, the
means in the females are farther apart than the means in the males. Thus, in
females the effect of species is larger than in males. This is the interaction effect,
and this difference in the difference in means is equal to 3.56 in this data set.
With a confidence level of 95% we can say that this difference in the effect of
species is probably somewhere between 0.06 and 7.06 mm in the population of
all penguins.

9.5 Moderation involving two numeric variables

in R
In all previous examples, we saw at least one categorical variable. We saw that
for different levels of a dummy variable, the slope of another variable varied.
We also saw that for different levels of a dummy variable, the effect of another
dummy variable varied. In this section, we look at how the slope of a numeric
variable can vary, as a function of the level of another numeric variable.
As an example data set, we look at the mpg data frame, available in the
ggplot2 package. It contains data on 234 cars. Let’s analyse the dependent
variable cty (city miles per gallon) as a function of the numeric variables cyl
(number of cylinders) and displ (engine displacement). First we plot the
relationship between engine displacement and city miles per gallon. We use
colours, based on the number of cylinders. We see that there is in general a
negative slope: the higher the displacement value, the lower the city miles per
gallon.

mpg %>%
ggplot(aes(x = displ, y = cty)) +

268
 geom_point(aes(colour = cyl)) +
geom_smooth(method = "lm", se = F)

30
cyl
8

7
cty

20 6

10

2 3 4 5 6 7
displ

When we run separate linear models for the different number of cylinders,
we get

mpg %>%
ggplot(aes(x = displ, y = cty, colour = cyl, group = cyl)) +
geom_point() +
geom_smooth(method = "lm", se = F)

35

30
cyl
8
25
7
cty

6
20
5

15 4

10

2 3 4 5 6 7
displ

We see that the slope is different, depending on the number of cylinders: the
more cylinders, the less negative is the slope: very negative for cars with low
number of cylinders, and slightly positive for cars with high number of cilinders.
In other words, the slope increases in value with increasing number of cylinders.
If we want to quantify this interaction effect, we need to run a linear model with
an interaction effect.

269
out <- mpg %>%
lm(cty ~ displ + cyl + displ:cyl, data = .)
out %>%
tidy()

## # A tibble: 4 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 38.6 2.03 19.0 3.69e-49
## 2 displ -5.29 0.825 -6.40 8.45e-10
## 3 cyl -2.70 0.376 -7.20 8.73e-12
## 4 displ:cyl 0.559 0.104 5.38 1.85e- 7

We see that the displ by cyl interaction effect is 0.559. It means that the
slope of displ changes by 0.559 for every unit increase in cyl.
For example, when we look at the predicted city miles per gallon with cyl
= 2, we get the following model equation:

cd
ty = 0.6 − 5.285 × displ − 2.704cyl + 0.559 × displ × cyl
cty
d = 0.6 − 5.285 × displ − 2.704 × 2 + 0.559 × displ × 2
cd
ty = 0.6 − 5.285 × displ − 5.408 + 1.118 × displ
cd
ty = (0.6 − 5.408) + (1.118 − 5.285) × displ
cd
ty = −4.808 − 4.167 × displ (9.33)

If we increase the number of cylinders from 2 to 3, we obtain the equation:

ty = 0.6 − 5.285 × displ − 2.704 × 3 + 0.559 × displ × 3

cd
ty = −7.512 − 3.608 × displ
cd (9.34)

We see a different intercept and a different slope. The difference in the slope
between 3 and 2 cylinders equals 0.559, which is exactly the interaction effect.
If you do the same exercise with 4 and 5 cylinders, or 6 and 7 cylinders, you
will always see this difference again. This parameter for the interaction effect
just says that the best prediction for the change in slope when increasing the
number of cylinders with 1, is 0.559. We can plot the predictions from this
model in the following way:

library(modelr)
mpg %>%
add_predictions(out) %>%
ggplot(aes(x = displ, y = cty, colour = cyl)) +
geom_point() +
geom_line(aes(y = pred, group = cyl))

270
 35

30
cyl
8
25
7
cty

6
20
5

15 4

10

2 3 4 5 6 7
displ

If we compare these predicted regression lines with those in the previous

figure

mpg %>%
add_predictions(out) %>%
ggplot(aes(x = displ, y = cty, group = cyl)) +
geom_point() +
geom_line(aes(y = pred), colour = "black") +
geom_smooth(method = "lm", se = F)

35

30

25
cty

20

15

10

2 3 4 5 6 7
displ

we see that they are a little bit different. That is because in the model we
treat cyl as numeric: for every increase of 1 in cyl, the slope changes by a
fixed amount. When you treat cyl as categorical, then you estimate the slope
separately for all different levels. You would then see multiple parameters for
the interaction effect:

271
out <- mpg %>%
mutate(cyl = factor(cyl)) %>%
lm(cty ~ displ + cyl + displ:cyl, data = .)
out %>%
tidy()

## # A tibble: 8 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 33.8 1.70 19.9 1.50e-51
## 2 displ -5.96 0.785 -7.60 7.94e-13
## 3 cyl5 1.60 1.17 1.37 1.72e- 1
## 4 cyl6 -11.6 2.50 -4.65 5.52e- 6
## 5 cyl8 -23.4 2.89 -8.11 3.12e-14
## 6 displ:cyl5 NA NA NA NA
## 7 displ:cyl6 4.21 0.948 4.44 1.38e- 5
## 8 displ:cyl8 6.39 0.906 7.06 2.05e-11

When cyl is turned into a factor, you see that cars with 4 cylinders are taken
as the reference category, and there are effects of having 5, 6, or 8 cylinders.
We see the same for the interaction effects: there is a reference category with 4
cylinders, where the slope of displ equals -5.96. Cars with 6 and 8 cylinders
have different slopes: the one for 6 cylinders is 5.96 + 4.21 and the one for 8
cylinders is 5.96 + 6.39. The slope for cars with 5 cylinders can’t be separately
estimated because there is no variation in displ in the mpg data set.
You see that you get different results, depending on whether you treat a
variable as numeric or as categorical. Treated as numeric, you end up with a
simpeler model with fewer parameters, and therefore a larger number of degrees
of freedom. What to choose depends on the research question and the amount
of data. In general, a model should be not too complex when you have relatively
few data points. Whether the model is appropriate for your data can be checked
by looking at the residuals and checking the assumptions.

272
Chapter 10

Linear mixed modelling:

introduction

10.1 Fixed effects and random effects

In the simplest form of linear modelling, we have one dependent numeric variable,
one intercept and one or more independent variables. Let’s look at a simple
regression equation where dependent variable Y is predicted by an intercept b0
and a linear effect of independent variable X with regression slope parameter
b1 , and an error term e, where we assume that the error term e comes from a
normal distribution.

Y = b0 + b1 X + e (10.1)
2
e ∼ N (0, σ ) (10.2)

Using this model, we know that for a person with a value of 5 for X, we
expect Y to be equal to b0 + b1 × 5. As another example, if Y is someone’s IQ
score, X is someone’s brain size in cubic millilitres, b0 is equal to 70, and b1 is
equal to 0.1, we expect on the basis of this model that a person with a brain
size of 1500 cubic millimetres has an IQ score of 70+0.01×1500, which equals 85.

Now, for any model the predicted values usually are not the same as the observed
values. If the model predicts on the basis of my brain size that my IQ is 140,
my true IQ might be in fact 130. This discrepancy is termed the residual:
the observed Y , minus the predicted Y , or Yb , so in this case the residual is
Y − Yb = 130 − 140 = −10.

Here we have the model for the relationship between IQ and brain size.

273
 IQ = 70 + 0.1 × Brain size + e (10.3)
2
e ∼ N (0, σ ) (10.4)

Note that in this model, the values of 70 and 0.1 are fixed, that is, we use the
same intercept and the same slope for everyone. You use these values for any
person, for Henry, Jake, Liz, and Margaret. We therefore call these effects of
intercept and slope fixed effects, as they are all the same for all units of analysis.
In contrast, we call the e term, the random error term or the residual in the
regression, a random effect. This is because the error term is different for every
unit. We don’t know the specific values of these random errors or residuals for
every person, but nevertheless, we assume that they come from a distribution,
in this case a normal distribution with mean 0 and an unknown variance. This
unknown variance is given the symbol σ 2 .
Here are a few more examples.

1. Suppose we study a number of schools, and for every school we use a simple
linear regression equation to predict the number of students (dependent
variable) on the basis of the number of teachers (independent variable).
For every unit of analysis (in this case: school), the intercept and the
regression slope are the same (fixed effects), but the residuals are different
(random effect).
2. Suppose we study reaction times, and for every measure of reaction time
– a trial – we use a simple linear regression equation to predict reaction
time in milliseconds on the basis of the characteristics of the stimulus.
Here, the unit of analysis is trial, and for every trial, the intercept and the
regression slope are the same (fixed effects), but the residuals are different
(random effect).
3. Suppose we study a number of students, and for every student we use
a simple linear regression equation to predict the math test score on the
basis of the number of hours of study the student puts in. Here, the unit of
analysis is student, and for every student, the intercept and the regression
slope are the same (fixed effects), but the residuals are different (random
effect).

Let’s focus for now on the last example. What happens when we have a lot
of data on students, but the students come from different schools? Suppose we
want to predict average grade for every student, on the basis of the number of
hours of study the student puts in. We again could use a simple linear regression
equation.

Y = b0 + b1 hourswork + e (10.5)
e ∼ N (0, σ 2 ) (10.6)

274
 That would be fine if all schools would be all very similar. But suppose
that some schools have a lot of high scoring students, and some schools have a
lot of low scoring students? Then school itself would also be a very important
predictor, apart from the number of hours of study. One could say that the data
are clustered : math test scores coming from the same school are more similar
than math test scores coming from different schools. When we do not take this
into account, the residuals will not show independence (see Chapter 7 on the
assumptions of linear models).
One thing we could therefore do to remedy this is to include school as a
categorical predictor. We would then have to code this school variable into a
number of dummy variables. The first dummy variable called school1 would
indicate whether students are in the first school (school1 = 1) or not (school1
= 0). The second dummy variable school2 would indicate whether students are
in the second school (school2 = 1) or not (school2 = 0), etcetera. You can
then add these dummy variables to the regression equation like this:

Y = b0 + b1 hourswork + b2 school1 + b3 school2 + b4 school3 + · · · + e

e ∼ N (0, σ 2 )

In the output we would find a large number of effects, one for each dummy
variable. For example, if the students came from 100 different schools, you
would get 99 fixed effects for the 99 dummy variables. However, one could
wonder whether this is very useful. As stated earlier, fixed effects are called
fixed because they are the same for every unit of research, in this case every
student. But working with 99 dummy variables, where students mostly score 0,
this seems very much over the top. In fact, we’re not even interested in these
99 effects. We’re interested in the relationship between test score and hours
of work, meanwhile taking into account that there are test score differences
across schools. The dummy variables are only there to account for differences
across schools; the prediction for one school is a little bit higher or lower than
for another school, depending on how well students generally perform in each
school.
We could therefore try an alternative model, where we treat the school effect
as random: we assume that every school has a different average test score, and
that these averages are normally distributed. We call these average test score
deviations school effects:

Y = b0 + b1 hourswork + schooleffect + e (10.7)

schooleffect ∼ N (0, σs2 ) (10.8)
e ∼ N (0, σe2 ) (10.9)

So in this equation, the intercept is fixed, that is, the intercept is the same
for all observed test scores. The regression coefficient b1 for the effect of hours
of work is also fixed. But the schooleffect is random, since it is different for

275
every school. The residual e is also random, being different for every student.
It could also be written like this:

Y = (b0 + schooleffect) + b1 hourswork + e (10.10)

schooleffect ∼ N (0, σs2 ) (10.11)
e ∼ N (0, σe2 ) (10.12)

This representation emphasizes that for every school, the intercept is a little
bit different: for school A the intercept might be b0 + 2, and for school B the
intercept might be b0 − 3.
So, equation 10.10 states that every observed test score is

1. partly influenced by an intercept that is random, with a certain average

b0 and variance σs2 , that is dependent on which school students are in,

2. partly influenced by the number of hours of work, an effect that is the

same no matter what school a student is in (fixed), and

3. partly influenced by unknown factors, indicated by a random residual e

coming from a normal distribution with variance σe2 .

To put it more formally: test score Yij , that is, the test score from student j
in school i, is the sum of an effect of the school b0 + schooleffecti (the average
test score in school i), plus an effect of hours of work, b1 × hourswork, and an
unknown residual eij (a specific residual for the test score for student j in school
i).

Yij = b0 + schooleffecti + b1 hourswork + eij (10.13)

schooleffecti ∼ N (0, σs2 ) (10.14)
eij ∼ N (0, σe2 ) (10.15)

So in addition to the assumption of residuals that have a normal distribution

with mean 0 and variance σe2 , we also have an assumption that the school
averages have a normal distribution, in this case with mean b0 and variance σs2 .

Let’s go back to the example of reaction times. Suppose in an experiment

we measure reaction time in a large number of trials. We want to know whether
the size of the stimulus (large or small) has an effect on reaction time. Let’s
also suppose that we carry out this experiment with 20 participants, where every
participant is measured during 100 trials: 50 large stimuli and 50 small stimuli,
in random order. Now probably, some participants show generally very fast
responses, and some participants show generally very slow responses. In other
words, the average reaction time for the 100 trials may vary from participant to
participant. This means that we can use participant as an important predictor

276
of reaction times. To take this into account we can use the following linear
equation:

Yij = b0 + speedi + b1 size + eij (10.16)

speedi ∼ N (0, σs2 ) (10.17)
eij ∼ N (0, σe2 ), (10.18)

where Yij , is the reaction time j from participant i, (b0 +speedi ) is a random
intercept representing the average speed for each participant i (where b0 is the
overall average across all participants and speedi the random deviation for each
and every participant), b1 is the fixed effect of the size of the stimulus. Unknown
residual eij is a specific residual for the reaction time for trial j of participant i.
The reason for introducing random effects is that when your observed data
are clustered, for instance student scores clustered within schools, or trial response
times are clustered within participants, you violate the assumption of independence:
two reaction times from the same person are more similar than two reaction
times from different persons. Two test scores from students from the same
school may be more similar than two scores from students in different schools
(see Chapter 7). When this is the case, when data are clustered, it is very
important to take this into account. When the assumption of independence is
violated, you are making wrong inference if you use an ordinary linear model,
the so-called general linear model (GLM). With clustered data, it is therefore
necessary to work with an extension of the general linear model or GLM, the
linear mixed model. The above models for students’ test scores across different
schools and reaction times across different participants, are examples of linear
mixed models. The term mixed comes from the fact that the models contain a
mix of both fixed and random effects. GLMs only contain fixed effects, apart
from the random residual.
If you have clustered data, you should take this clustering into account, either
by using the grouping variable as a categorical predictor or by using a random
factor in a linear mixed model. As a rule of thumb: if you have fewer than
10 groups, consider a fixed categorical factor; if you have 10 or more groups,
consider a random factor. Two other rules you can follow are: Use a random
factor if the assumption of normally distributed group differences is tenable.
Use a fixed categorical factor if you are actually interested in the size of group
differences.
Below, we will start with a very simple example of a linear mixed model,
one that we use for a simple pre-post intervention design.

10.2 Pre-post intervention designs

Imagine a study where we hope to show that aspirin helps reduce headache. We
ask 100 patients to rate the severity of their headache before they use aspirin
(on a scale from 1 - 100), and to rate the severity again 3 hours after taking 500

277
Table 10.1: Headache measurements in NY Times readers suffering from
headaches.
patient pre post
001 55 45
002 63 50
003 66 56
004 50 37
005 63 50
... ... ...

55

50
post

45

40

50 55 60 65
pre

Figure 10.1: Scatterplot of pre and post headache levels.

mg of aspirin. These patients are randomly selected among people who read
the NY Times and suffer from regular headaches. So here we have clustered
data: we have 100 patients, and for each patient we have two scores, one before
(pre) and one after (post) the intervention of taking aspirin. Of course, overall
headache severity levels tend to vary from person to person, so we might have
to take into account that some patients have a higher average level of pain than
other patients.
The data could be represented in different ways, but suppose we have the
data matrix in Table 10.1 (showing only the first five patients). What we observe
in that table is that the severity seems generally lower after the intervention
than before the intervention. But you may also notice that the severity of
the headache also varies across patients: some have generally high scores (for
instance patient 003), and some have generally low scores (for example patient
001). Therefore, the headache scores seem to be clustered, violating the assumption
of independence. We can quantify this clustering by computing a correlation
between the pre-intervention scores and the post-intervention scores. We can
also visualise this clustering by a scatter plot, see Figure 10.1. Here it appears
that there is a strong positive correlation, indicating that the higher the pain
score before the intervention, the higher the pain score after the intervention.

278
 There is an alternative way of representing the same data. Let’s look at
the same data in a new format in Table 10.2. In Chapter 1 we saw that this
representation is called long format.

Table 10.2: Headache severity measures in long format.

patient measure headache
1 1 55
1 2 45
2 1 63
2 2 50
3 1 66
3 2 56
4 1 50
4 2 37
5 1 63
5 2 50

By representing the data in long format we acknowledge that there is really

only one dependent measure: headache severity. The other two variables indicate
that this variable varies across both patients and time point (pre intervention
and post intervention). There is therefore a variable measure that indicates
whether the headache severity was measured pre intervention (measure = 1) or
post intervention (measure = 2).
Here we might consider applying a simple linear regression model, using
headache as the dependent variable and measure (1st or 2nd) as a categorical
predictor. However, since we know that there is a correlation between the pre
and post severity measures, we know that measures also systematically vary
across patients: some score high on average and some score low on average.
The assumption of independence is therefore not tenable. Thus we have to run
a linear model, including not only an effect of measure but also an effect of
patient. We then have to decide between fixed effects or random effects for
these variables.
Let’s first look at the variable measure. Since we are really interested in the
effect of the intervention, that is, we want to know how large the effect of aspirin
is, we use a fixed effect for the time effect (the variable measure). Moreover,
the variable measure has only two levels, which is another reason to opt for a
fixed effect.
Then for the patient effect, we see that we have 100 patients. Are we really
interested by how much the average pain level in say patient 23 differs from
the average pain level in say patient 45? No, not really. We only want to
acknowledge that the individual differences exist, we want to take them into
account in our model so that our inference regarding the confidence intervals
and hypothesis testing is correct. We therefore prefer to assume random effects,
assuming they are normally distributed. We therefore end up with a fixed effect
for measure and a random effect for patient resulting in the following model:

279
 Yij = b0 + patienti + b1 measure2 + eij (10.19)
patienti ∼ N (0, σp2 ) (10.20)
eij ∼ N (0, σe2 ) (10.21)

where Yij is the jth headache severity score (first or second) for patient i,
(b0 + patienti ) is the average amount of headache before aspirin that patient
i deviates from the mean, measure2 is a dummy variable, and b1 is the effect
of the intervention (by how much the severity changes from pre to post). We
assume that the average pain level for each patient shows a normal distribution
with average b0 and variance σp2 . And of course we assume that the residuals
show a normal distribution.

An analysis with this model can be done with the R package lme4. That
package contains the function lmer() that works more or less the same as the
lm() function that we saw earlier, but requires the addition of at least one
random variable. Below, we run a linear mixed model, with dependent variable
headache, a regular fixed effect for the categorical variable measure, and a
random effect for the categorical variable patient.

library(lme4)
out <- data %>%
lmer(headache ~ measure + (1|patient), data = .)
summary(out)

## Linear mixed model fit by REML [’lmerMod’]

## Formula: headache ~ measure + (1 | patient)
## Data: .
##
## REML criterion at convergence: 1189.8
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.36239 -0.42430 -0.04973 0.49702 2.17096
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 27.144 5.210
## Residual 8.278 2.877
## Number of obs: 200, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error t value
## (Intercept) 59.6800 0.5952 100.28
## measure2 -10.3600 0.4069 -25.46

280
##
## Correlation of Fixed Effects:
## (Intr)
## measure2 -0.342

In the output we see the results. We’re mainly focused on the fixed effect of
the intervention: does aspirin reduce headache? Where it says ’Fixed effects:’
in the output, we see the linear model coefficients, with an intercept of around
59 and a negative effect of the intervention dummy variable measure2, around
−10. We see that the dummy variable was coded 1 for the second measure
(after taking aspirin). So, for our dependent variable headache, we see that
the expected headache severity for the observations with a 0 for the dummy
variable measure2 (that is, measure 1, which is before taking aspirin), is equal
to 59 − (10) × 0 = 59.
Similarly, we see that the expected headache severity for the observations
with a 1 for the dummy variable measure2 (that is, after taking aspirin),
is equal to 59 − (10) × 1 = 49 − 10 = 49. So, expected pain severity is
10 points lower after the intervention than before the intervention. Whether
this difference is significant is indicated by a t-statistic. We see here that the
average headache severity after taking an aspirin is significantly different from
the average headache severity before taking an aspirin, t = −25.46. However,
note that we do not see a p-value. That’s because the degrees of freedom are not
clear, because we also have a random variable in the model. The determination
of the degrees of freedom in a linear mixed model is a complicated matter, with
different choices, the discussion of which is beyond the scope of this book.
However, we do know that for whatever the degrees of freedom really are, a t-
statistic of -25.46 will always be in the far tail of the t-distribution, see Appendix
B. So we have good reason to conclude that we have sufficient evidence to reject
the null-hypothesis that headache levels before and after aspirin intake are the
same.
If we do not want to test a null-hypothesis, but want to construct a confidence
interval, we run into the same problem that we do not know what t-distribution
to use. Therefore we do not know with what value the standard error of 0.40
should be multiplied to compute the margin of error. We could however use
a coarse rule of thump and say that the critical t-value for a 95% confidence
interval is more or less equal to 2 (see Appendix B). If we do that, then we get the
confidence level for the effect of aspirin: between −10.36 − 2 × 0.4069 = −11.17
and −10.36 + 2 × 0.4069 = −9.55.
Taking into account the direction of the effect and the confidence interval for
this effect, we might therefore carefully conclude that aspirin reduces headache
in the population of NY Times readers with headache problems, where the
reduction is around 10 points on a 1...100 scale (95% CI: 9.55 – 11.17).

Now let’s look at the output regarding the random effect of patient more
closely. The model assumed that the individual differences in headache severity
in the 100 patients came from a normal distribution. How large are these

281
 0.08

0.06
density

0.04

0.02

0.00
0 25 50 75 100
headache

Figure 10.2: Distribution of headache scores before taking aspirin, according to

the linear mixed model.

individual differences actually? This can be gleaned from the ’Random effects:’
part of the R output. The ’intercept’ (i.e., the patient effect in our model)
seems
√ to vary with a variance of 27, which is equivalent to a standard deviation
of 27 which is around 5.2. What does that mean exactly? Well let’s look at
the equation again and fill in the numbers:

Yij = b0 + patienti + b1 measure2 + eij (10.22)

Yij = 59 + patienti − 10measure2 + eij (10.23)
patienti ∼ N (0, 27) (10.24)
eij ∼ N (0, 8) (10.25)

Since R used the the headache level before the intervention as the reference
category, we conclude that the average pain level before taking aspirin is 59.
However, not everybody’s pain level before taking aspirin is 59: people show
variance (variation). The pain level before aspirin varies with a variance of 27,
which is equivalent to a standard deviation of around 5.2. Figure 10.2 shows
how much this variance actually is. It depicts a normal distribution with a mean
of 59 and a standard deviation of 5.2.
So before taking aspirin, most patients show headache levels roughly between
50 and 70. More specifically, if we would take the middle 95% by using plus or
minus twice the standard deviation, we can estimate that 95% of the patients
shows levels between 59 − 2 × 5.2 = 48.6 and 59 + 2 × 5.2 = 69.4.

Now let’s look at the levels after taking aspirin. The average headache level
is equal to 59 − 10 = 49. So 95% of the patients shows headache levels between
49 − 2 × 5.2 = 38.6 and 49 + 2 × 5.2 = 59.4 before taking aspirin.

282
 0.08

0.06 3 hrs after aspirin before aspirin

density

0.04

0.02

0.00
0 25 50 75 100
headache

Figure 10.3: Distribution of headache scores before and after taking aspirin,
according to the linear mixed model.

Together these results are visualised in Figure 10.3. In this plot you see there
is variability in headache levels before taking aspirin, and there is variation in
headache levels after taking aspirin. We also see that these distributions have the
same spread (variance): in the model we assume that the variability in headache
before aspirin is equal to the variability after aspirin (homoscedasticity). The
distributions are equal, except for a horizontal shift: the distribution for headache
after aspirin is the same as the distribution before aspirin, except for a shift to
the left of about 10 points. This is of course the effect of aspirin in the model,
the b1 parameter in our model above.
The fact that the two distributions before and after aspirin show the same
spread (variance) was an inherent assumption in our model: we only have one
random effect for patient in our output with one variance (σp2 ). If the assumption
of equal variance (homoscedasticity) is not tenable, then one should consider
other linear mixed models. But this is beyond the scope of this book. The
assumption can be checked by plotting the residuals, using different colours for
residuals from before taking aspirin and for residuals from after taking aspirin.

library(modelr)
data %>%
add_residuals(out) %>%
add_predictions(out) %>%
ggplot(aes(x = pred, y = resid, colour = measure)) +
geom_point() +
xlab("Predicted headache") +
ylab("Residual") +

283
 scale_colour_brewer(palette = "Set1") # use nice colours

measure
Residual

0 1
2

−4

40 50 60 70
Predicted headache

Alternatively one could create a box plot.

data %>%
add_residuals(out) %>%
ggplot(aes(x = measure, y = resid)) +
geom_boxplot() +
xlab("Measure") +
ylab("Residual")

4
Residual

−4

1 2
Measure

The plots show that the variation in the residuals is about the same for pre
and post aspirin headache levels (box plot) and for all predicted headache levels
(scatter plot). This satisfies the assumption of homogeneity of variance.
If all assumptions are satisfied, you are at liberty to make inferences regarding
the model parameters. We saw that the effect of aspirin was estimated at
about 10 points with a rough 95% confidence interval, that was based on the

284
rule of thumb of 2 standard errors around the estimate. For people that are
uncomfortable with such quick and dirty estimates, it is also possible to use
Satterthwaite’s approximation of the degrees of freedom. You can obtain these,
once you load the package lmerTest. After loading, your lmer() analysis will
yield (estimated) degrees of freedom and the p-values associated with the t-
statistics and those degrees of freedom.

library(lmerTest)

##
## Attaching package: ’lmerTest’
## The following object is masked from ’package:lme4’:
##
## lmer
## The following object is masked from ’package:stats’:
##
## step
out <- data %>%
lmer(headache ~ measure + (1|patient), data = .)
summary(out)

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

## lmerModLmerTest]
## Formula: headache ~ measure + (1 | patient)
## Data: .
##
## REML criterion at convergence: 1189.8
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.36239 -0.42430 -0.04973 0.49702 2.17096
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 27.144 5.210
## Residual 8.278 2.877
## Number of obs: 200, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error df t value Pr(>|t|)
## (Intercept) 59.6800 0.5952 124.7464 100.28 <0.0000000000000002 ***
## measure2 -10.3600 0.4069 99.0000 -25.46 <0.0000000000000002 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##

285
## Correlation of Fixed Effects:
## (Intr)
## measure2 -0.342

The degrees of freedom for effect of measure is 99. If we look in Appendix

B or check using R, we see that the critical t-value with 99 degrees of freedom
for a 95% confidence interval is equal to 1.98. That’s very close to our rough
estimate of 2.

qt(0.975, df = 99)

## [1] 1.984217

If we calculate the 95% confidence interval using the new critical value, we
obtain 59 − 1.98 × 5.2 = 48.7 and 59 + 1.98 × 5.2 = 69.3. That’s only a minor
difference with what we saw before.

10.3 Parameter estimation in linear mixed models

Remember that when we discussed linear regression in Chapter 4, the regression
coefficient was estimated using the principle of ordinary least squares (OLS). We
also saw that in order to obtain the estimates, we simply could fill in formulas
(the analytical approach). Unfortunately, the estimation of parameters in linear
mixed models is not that easy. There are two possibilities, both dependent on
a numerical search algorithm, explained in Chapter 4. The default in R is a
method called REML, that stands for Restricted Maximum Likelihood. The
other option is ML, standing for Maximum Likelihood. This book will not go
into the intricacies of these methods, but we can give some general rule of when
to use which method. Whenever the main interest of a study is on the variances
of the random effects, REML is the way to go. For instance, suppose that in the
aspirin study, the most important part of the analysis is to say something about
the individual differences in headache, and you want to estimate the variance of
these individual differences in the population. On the other hand, if the main
research question is about the fixed effects, the way to go is to use ML. For
instance, in the aspirin study, main interest is in the effect that aspirin has on
patients in the population. In that case, you want to deviate from the default
option. That can be done in the following way.

model <- data %>%

lmer(headache ~ measure + (1|patient), REML = FALSE, data = .)
summary(model)

## Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite’s

## method [lmerModLmerTest]
## Formula: headache ~ measure + (1 | patient)

286
## Data: .
##
## AIC BIC logLik deviance df.resid
## 1198.5 1211.7 -595.3 1190.5 196
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.37429 -0.42643 -0.04998 0.49953 2.18189
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 26.872 5.184
## Residual 8.195 2.863
## Number of obs: 200, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error df t value Pr(>|t|)
## (Intercept) 59.6800 0.5922 126.0065 100.78 <0.0000000000000002 ***
## measure2 -10.3600 0.4049 100.0000 -25.59 <0.0000000000000002 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Correlation of Fixed Effects:
## (Intr)
## measure2 -0.342

In the output you can see ”Linear mixed model fit by maximum likelihood”.
In this case, this is the output that we would be most interested in: we were
interested in the effect of aspirin on headache, and this was modelled as a
fixed effect of variable measure on the dependent variable. We only included
the random patient variable in order to deal with the dependency in the data.
We were never interested in the size of σp2 . My guess is that in the behavioural,
management and social sciences, interest is practically always in the fixed effects.
Therefore it is generally advised to add REML = FALSE when you use the lmer()
function.

287
288
Chapter 11

Linear mixed models for

more than two
measurements

11.1 Pre-mid-post intervention designs

In many intervention studies, one has more than two measurement moments.
Let’s go back to the example of the effect of aspirin on headache in Chapter
10. Suppose you’d like to know whether there is not only a short-term effect
of aspirin, but also a long-term effect. Imagine that the study on headache
among NY Times readers was extended by asking patients not only to rate
their headache before aspirin and 3 hours after intake, but also 24 hours after
intake. In this case our data could look like as presented in Table 11.1.

Table 11.1: Headache measures in NY Times readers.

patient pre post1 post2
1 52 45 47
2 59 50 55
3 65 56 58
4 51 37 42
5 62 50 55
6 61 53 57
7 56 44 55
8 62 48 53
9 56 48 49
10 58 45 44

So for each patient we have three measures: pre, post1 and post2. To see
if there is some clustering, it is no longer possible to study this by computing

289
a single correlation. We could however compute 3 different correlations: pre-
post1, pre-post2, and post1-post2, but this is rather tedious, and moreover
does not give us a single measure of the extent of clustering of the data. But
there is an alternative: one could compute not a Pearson correlation, but an
intraclass correlation (ICC). To do this, we need to bring the data again into
long format, as opposed to wide format, see Chapter 1. This is done in Table
11.2.

Table 11.2: Headache measures in NY Times readers in long format.

patient measure headache
1 pre 52
1 post1 45
1 post2 47
2 pre 59
2 post1 50
2 post2 55
3 pre 65
3 post1 56
3 post2 58
4 pre 51

Next, we can perform an analysis with the lmer() function from the lme4
package.

library(lme4)
model1 <- datalong %>%
lmer(headache ~ measure + (1|patient), data = ., REML = FALSE)
model1

## Linear mixed model fit by maximum likelihood [’lmerModLmerTest’]

## Formula: headache ~ measure + (1 | patient)
## Data: .
## AIC BIC logLik deviance df.resid
## 1740.998 1759.517 -865.499 1730.998 295
## Random effects:
## Groups Name Std.Dev.
## patient (Intercept) 5.290
## Residual 2.908
## Number of obs: 300, groups: patient, 100
## Fixed Effects:
## (Intercept) measurepost2 measurepre
## 49.32 2.36 9.85

In the output we see the fixed effects of two automatically created dummy
variables measurepost2 and measurepre, and the intercept. We also see the

290
standard deviations of the random effects: the standard deviation of the residuals
and the standard deviation of the random effects for the patients.
From this output, we can plug in the values into the equation:

headacheij = 49.32 + patienti + 2.36 measurepost2 + 9.85 measurepre + eij

patienti ∼ N (0, σp = 5.29)
eij ∼ N (0, σe = 2.91)

Based on this equation, the expected headache severity score in the population
three hours after aspirin intake is 49.32 (the first post measure is the reference
group). Dummy variable measurepost2 is coded 1 for the measurements 24
hours after aspirin intake. Therefore, the expected headache score 24 hours after
aspirin intake is equal to 49.32 + 2.36 = 51.68. Dummy variable measurepre
was coded 1 for the measurements before aspirin intake. Therefore, the expected
headache before aspirin intake is equal to 49.32 + 9.85 = 59.17. In sum, in this
sample we see that the average headache level decreases directly after aspirin
intake from 59.17 to 49.32, but then increases again to 51.68.

There is quite some variation in individual headache levels: the variance is

equal to 5.292 = 27.98, since the standard deviation (its square root) is equal
to 5.29. Therefore, if we look at roughly 95% of the sample, we see that
prior to taking aspirin, the scores vary between 59.17 − 2 × 5.29 = 48.59 and
59.17 + 2 × 5.29 = 69.75. For the short-term effect of aspirin after 3 hours, we
see that roughly 95% of the scores lie between 49.32 − 2 × 5.29 = 38.74 and
49.32 + 2 × 5.29 = 59.90. The normal distributions, predicted by this model,
are depicted in Figure 11.1.
So, are these distributions significantly different, in other words, do the
means differ significantly before aspirin, 3 hrs after aspirin and 24 hrs after
aspirin?
To answer a question about the equality of three means, we need an analysis
of variance (ANOVA).

model1 %>% anova()

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

## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## measure 5289.7 2644.9 2 200 312.71 < 0.00000000000000022 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

The F -statistic for the equality of the group means is quite large (remember
that the expected value is always 1 if the null-hypothesis is true, see Chapter 6).
We see no degrees of freedom. These need to be estimated, for example using
Satterthwaite’s method, provided by the lmerTest package:

291
 0.06 3 hrs after aspirin before aspirin
density 24 hrs after aspirin

0.04

0.02

0.00
0 25 50 75 100
headache

Figure 11.1: Distributions of the three headache levels before aspirin intake,
3 hours after intake and 24 hours after intake, according to the linear mixed
model.

library(lmerTest)
model2 <- datalong %>%
lmer(headache ~ measure + (1|patient),
data = .,
REML = FALSE)
model2 %>% anova()

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

## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## measure 5289.7 2644.9 2 200 312.71 < 0.00000000000000022 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

Using Satterthwaite’s method, we see that the error degrees of freedom is

about 200. The model degrees of freedom is of course 2, because we needed
two dummy variables for the three measurements. The p-value associated with
the F -value of 309.58 at 2 and 200 degrees of freedom is less than 0.05, so we
can reject the null-hypothesis. Thus we report that aspirin has an effect on
headache levels in NY Times readers, F (2, 198) = 312.71, p < 0.001.
If one has specific hypotheses regarding short-term and long-term effects,
one could perform a planned contrast analysis (see Chapter ??), comparing the
first measure with the second measure, and the first measure with the third
measure. If one is just interested in whether aspirin has an effect on headache,
then the overall F -test should suffice. If apart from this general effect one wishes
to explore whether there are significant differences between the three groups of
data, without any prior research hypothesis about this, then one could perform

292
a post hoc analysis of the three means. See Chapter ?? on how to perform
planned comparisons and post hoc tests.

Now recall that we mentioned an intraclass correlation, or ICC. An intraclass

correlation indicates how much clustering there is within the groups, in this case,
clustering of headache scores within individual NY Times readers. How much
are the three scores alike that come from the same patient? This correlation
can be computed using the following formula:

2
σpatient
ICC = 2 (11.1)
σpatient + σe2

Here, the variance of the patient random effects is equal to 5.292 = 27.98,
and the variance of the residuals e is equal to 2.912 = 8.47, so the intraclass
correlation for the headache severity scores is equal to
27.98
ICC = = 0.77 (11.2)
28.3 + 8.47
As this correlation is substantially higher than 0, we conclude there is quite
a lot of clustering. Therefore it’s a good thing that we used random effects for
the individual differences in headache scores among NY Times readers. Had
this correlation been 0 or very close to 0, however, then it would not have
mattered to include these random effects. In that case, we might as well use an
ordinary linear model, using the lm() function. Note from the formula that the
correlation becomes 0 when the variance of the random effects for patients is 0.
It approaches 0 as the random effects for patients grows small relative to the
residual variance. It approaches 1 as the random effects for patients grows large
relative to the residual variance. Because variance cannot be negative, ICCs
always have values between 0 and 1.

11.2 Pre-mid-post intervention design: linear effects

In the previous section, we’ve looked at categorical variables: measure (pre-
intervention, 3 hours after, and 24 hours after). We can use the same type of
analysis for numerical variables. In fact, we could have used a linear effect for
time in the headache example: using time of measurement as a variable. Let’s
look at the headache data again. But now we’ve created a new variable time
that is based on the variable measure: all first measurements are coded as time
= 0, all second measurements after 3 hours are coded as time = 3, and all third
measurements after 24 hours are coded as time = 24. The data are presented
in Table 11.3.
Instead of using a categorical intervention variable, with three levels, we
now use a numeric variable, time, indicating the number of hours that have
elapsed after aspirin intake. At point 0 hours, we measure headache severity,

293
Table 11.3: Headache measures in NY Times readers in long format with a new
variable time.
patient measure headache time
1 pre 52 0
1 post1 45 3
1 post2 47 24
2 pre 59 0
2 post1 50 3
2 post2 55 24
3 pre 65 0
3 post1 56 3
3 post2 58 24
4 pre 51 0

and patients take an aspirin. Next we measure headache after 3 hours and 24
hours. Above, we wanted to know if there were differences in average headache
between before intake and 3 hrs and 24 hrs after intake. Another question we
might ask ourselves: is there a linear reduction in headache severity after taking
aspirin?
For this we can do a linear regression type of analysis. We want to take into
account individual differences in headache severity levels among patients, so we
perform an lmer() analysis, using the following code, replacing the categorical
variable measure by numerical variable time:

model3 <- datalongnew %>%

lmer(headache ~ time + (1|patient), data = ., REML = FALSE)
model3

## Linear mixed model fit by maximum likelihood [’lmerModLmerTest’]

## Formula: headache ~ time + (1 | patient)
## Data: .
## AIC BIC logLik deviance df.resid
## 1997.2264 2012.0416 -994.6132 1989.2264 296
## Random effects:
## Groups Name Std.Dev.
## patient (Intercept) 4.533
## Residual 5.546
## Number of obs: 300, groups: patient, 100
## Fixed Effects:
## (Intercept) time
## 54.7913 -0.1557

In the output we see that the model for our data is equivalent to

294
 80

70
headache score

60

50

40

30
0 5 10 15 20 25
time after aspirin intake

Figure 11.2: Headache levels before aspirin intake, 3 hours after intake and 24
hours after intake.

headacheij = 54.79 + patienti − 0.1557 × time + eij (11.3)

patienti ∼ N (0, σp = 4.53) (11.4)
eij ∼ N (0, σe = 5.55) (11.5)

This model predicts that at time 0, the average headache severity score
equals 54.79, and that for every hour after intake, the headache level drops by
0.1557 points. So it predicts for example that after 10 hours, the headache has
dropped 1.557 points to 53.23.

Is this a good model for the data? Probably not. Look at the variance of the
residuals: with a standard deviation of 5.55 it is now a lot bigger than in the
previous analysis with the same data (see previous section). Larger variance of
residuals means that the model explains the data worse: predictions are worse,
so the residuals increase in size.

That the model is not appropriate for this data set is also obvious when we
plot the data, focusing on the relationship between time and headache levels,
see Figure 11.2.
The line shown is the fitted line based on the output. It can be seen that the
prediction for time = 0 is systematically too low, for time = 3 systematically
too high, and for time = 24 again too low. So for this particular data set on
headache, it would be better to use a categorical predictor for the effect of time
on headache, like we did in the previous section.

295
 80

70
headache score

60

50

40

30
0 1 2 3
time after aspirin intake

Figure 11.3: Alternative headache levels before aspirin intake, 3 hours after
intake and 24 hours after intake.

As an example of a data set where a linear effect would have been appropriate,
imagine that we measured headache 0 hours, 2 hours and 3 hours after aspirin
intake (and not after 24 hours). Suppose these data would look like those in
Figure 11.3. There we see a gradual increase of headache levels right after
aspirin intake. Here, a numeric treatment of the time variable would be quite
appropriate.
Suppose we would then see the following output.

model4 <- datalongnew2 %>%

lmer(headache ~ time + (1|patient), data = ., REML = FALSE)
model4 %>% summary()

## Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite’s

## method [lmerModLmerTest]
## Formula: headache ~ time + (1 | patient)
## Data: .
##
## AIC BIC logLik deviance df.resid
## 1745.4 1760.2 -868.7 1737.4 296
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.78967 -0.55395 0.05095 0.56015 2.30971
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 27.889 5.281

296
## Residual 8.732 2.955
## Number of obs: 300, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error df t value Pr(>|t|)
## (Intercept) 58.9721 0.6000 134.6807 98.29 <0.0000000000000002 ***
## time -3.3493 0.1368 200.0000 -24.48 <0.0000000000000002 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Correlation of Fixed Effects:
## (Intr)
## time -0.380

Because we are confident that this model is appropriate for our data, we can
interpret the statistical output. The Satterthwaite error degrees of freedom are
200, so we can construct a 95% confidence interval by finding the appropriate
t-value.

qt(0.025, df = 200)

## [1] -1.971896

The 95% confidence interval for the effect of time is then from −3.349 −
1.97 × 0.1368 to −3.349 + 1.97 × 0.1368, so from -3.62 to -3.08. We can report:

A linear mixed model was run on the headache levels, using the
numeric predictor variable time and random effects for the variable
patient. We saw a significant linear effect of time on headache level,
t(200) = −24.42, p < 0.001. The estimated effect of time based
on this analysis is negative, −3.3, so with every hour that elapses
after aspirin intake, the predicted headache score decreases with 3.3
points (95% CI: 3.08 to 3.62 points).

11.3 Linear mixed models and interaction effects

Suppose we carry out the aspirin and headache study not only with a random
sample of NY Times readers that suffer from regular headaches, but also with
a random sample of readers of the Wall Street Journal that suffer from regular
headaches. We’d like to know whether aspirin works, but we are also interested
to know whether the effect of aspirin is similar in the two groups of readers. Our
null-hypothesis is that the effect of aspirin in affecting headache severity is the
same in NY Times and Wall Street Journal readers that suffer from headaches.

H0 : The effect of aspirin is the same for NY Times readers as for Wall Street
Journal readers.

297
Suppose we have the data set in Table 11.4 (we only show the first six patients),
and we only look at the measurements before aspirin intake and 3 hours after
aspirin intake (pre-post design).

Table 11.4: Headache measures in NY Times and Wall Street Journal readers
in wide format.
patient group pre post
1 NYTimes 55 45
2 WallStreetJ 63 50
3 NYTimes 66 56
4 WallStreetJ 50 37
5 NYTimes 63 50
6 WallStreetJ 65 53

In this part of the data set, patients 2, 4, and 6 read the Wall Street Journal,
and patients 1, 3 and 5 read the NY Times. We assume that people only read
one of these newspapers. We measure their headache before and after the intake
of aspirin (a pre-post design). The data are now in what we call wide format:
the dependent variable headache is spread over two columns, pre and post. In
order to analyse the data with linear models, we need them in long format, as
in Table 11.5.

Table 11.5: Headache measures in NY Times and Wall Street Journal readers
in long format.
patient group measure headache
1 NYTimes pre 55
1 NYTimes post 45
2 WallStreetJ pre 63
2 WallStreetJ post 50
3 NYTimes pre 66
3 NYTimes post 56

datalong <- datawide %>%

pivot_longer(cols = -c(patient, group),
names_to = "measure",
values_to = "headache")

The new variable measure now indicates whether a given measurement of

headache refers to a measurement before intake (pre) or after intake (post).
Again we could investigate whether there is an effect of aspirin with a linear
mixed model, with measure as our categorical predictor, but that is not really
what we want to test: we only want to know whether the effect of aspirin (being
small, large, negative or non-existent) is the same for both groups. Remember

298
that this hypothesis states that there is no interaction effect of aspirin (measure)
and group. The null-hypothesis is that group is not a moderator of the effect
of aspirin on headache. There may be an effect of aspirin or there may not,
and there may be an effect of newspaper (group) or there may not, but we’re
interested in the interaction of aspirin and group membership. Is the effect of
aspirin different for NY Times readers than for Wall Street Journal readers?
In our model we therefore need to specify an interaction effect. Since the
data are clustered (2 measures per patient), we use a linear mixed model. First
we show how to analyse these data using dummy variables, later we will show
the results using a different approach.

We recode the data into two dummy variables, one for the aspirin intervention
(dummy1: 1 if measure = post, 0 otherwise), and one for group membership
(dummy2: 1 if group = NYTimes, 0 otherwise):

datalong <- datalong %>%

mutate(dummy1 = ifelse(measure == "post", 1, 0),
dummy2 = ifelse(group == "NYTimes", 1, 0))
datalong %>% head(3)

## # A tibble: 3 x 6
## patient group measure headache dummy1 dummy2
## <int> <chr> <chr> <dbl> <dbl> <dbl>
## 1 1 NYTimes pre 55 0 1
## 2 1 NYTimes post 45 1 1
## 3 2 WallStreetJ pre 63 0 0

Next we need to compute the product of these two dummies to code a dummy
for the interaction effect. Since with the above dummy coding, all post measures
get a 1, and all NY Times readers get a 1, only the observations that are post
aspirin and that are from NY Times readers get a 1 for this product.

datalong <- datalong %>%

mutate(dummy_interact = dummy1*dummy2)
datalong %>% head(3)

## # A tibble: 3 x 7
## patient group measure headache dummy1 dummy2 dummy_interact
## <int> <chr> <chr> <dbl> <dbl> <dbl> <dbl>
## 1 1 NYTimes pre 55 0 1 0
## 2 1 NYTimes post 45 1 1 1
## 3 2 WallStreetJ pre 63 0 0 0

With these three new dummy variables we can specify the linear mixed
model.

299
model5 <- datalong %>%
lmer(headache ~ dummy1 + dummy2 + dummy_interact + (1|patient),
data = .,
REML = FALSE)
model5

## Linear mixed model fit by maximum likelihood [’lmerModLmerTest’]

## Formula: headache ~ dummy1 + dummy2 + dummy_interact + (1 | patient)
## Data: .
## AIC BIC logLik deviance df.resid
## 1201.6855 1221.4754 -594.8427 1189.6855 194
## Random effects:
## Groups Name Std.Dev.
## patient (Intercept) 5.177
## Residual 2.855
## Number of obs: 200, groups: patient, 100
## Fixed Effects:
## (Intercept) dummy1 dummy2 dummy_interact
## 59.52 -10.66 0.32 0.60

In the output, we recognise the three fixed effects for the three dummy
variables. Since we’re interested in the interaction effect, we look at the effect
of dummy interact. The effect is in the order of +0.6. What does this mean?
Remember that all headache measures before aspirin intake are given a 0
for the intervention dummy dummy1. A reader from the Wall Street Journal
gets a 0 for the group dummy dummy2. Since the product of 0 × 0 equals 0, all
measures before aspirin in Wall Street Journal readers get a 0 for the interaction
dummy dummy interact. Therefore, the intercept of 59.52 refers to the expected
headache severity of Wall Street Journal readers before they take their aspirin.
Furthermore, we see that the effect of dummy1 is -10.66. The variable dummy1
codes for post measurements. So, relative to Wall Street Journal readers prior
to aspirin intake, the level of post intake headache is 10.66 points lower.
If we look further in the output, we see that the effect of dummy2 equals
+0.32. This variable dummy2 codes for NY Times readers. So, relative to Wall
Street Journal readers and before aspirin intake (the reference group), NY Times
readers score on average 0.32 points higher on the headache scale.
However, we’re not interested in a general difference between those two
groups of readers, we’re interested in the effect of aspirin and whether it is
different in the two groups of readers. In the output we see the interaction
effect: being a reader of the NY Times AND at the same time being a measure
after aspirin intake, the expected level of headache is an extra +0.60. The effect
of aspirin is -10.66 in Wall Street Journal readers, as we saw above, but the
effect is −10.66 + 0.60 = −10.06 in NY Times readers. So in this sample the
effect of aspirin on headache is 0.60 smaller than in Wall Street Journal readers
(note that even while the interaction effect is positive, it is positive on a scale
where a high score means more headache).

300
Table 11.6: Expected headache levels in Wall Street Journal and NY Times
readers, before and after aspirin intake.
measure group dummy1 dummy2 dummy interact exp mean
pre WallStreetJ 0 0 0 60
post WallStreetJ 1 0 0 60 + (−11) = 49
pre NYTimes 0 1 0 60 + 0.3 = 60.3
post NYTimes 1 1 1 60 + (−11) + 0.3 + 0.6 = 49.9

Let’s look at it in a different way, using a table with the dummy codes, see
Table 11.6. For each group of data, pre or post aspirin and NY Times readers
and Wall Street Journal readers, we note the dummy codes for the new dummy
variables. In the last column we use the output estimates and multiply them
with the respective dummy codes (1 and 0) to obtain the expected headache
level (using rounded numbers):
The exact numbers are displayed in Figure 11.4. We see that the specific
effect of aspirin in NY Times readers is 0.60 smaller than the effect of aspirin
in Wall Street Journal readers. This difference in the effect of aspirin between
the groups was not significantly different from 0, as we can see when we let R
plot a summary of the results.

model5 %>% summary()

## Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite’s
## method [lmerModLmerTest]
## Formula: headache ~ dummy1 + dummy2 + dummy_interact + (1 | patient)
## Data: .
##
## AIC BIC logLik deviance df.resid
## 1201.7 1221.5 -594.8 1189.7 194
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.34085 -0.43678 -0.02942 0.50880 2.22534
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 26.80 5.177
## Residual 8.15 2.855
## Number of obs: 200, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error df t value Pr(>|t|)
## (Intercept) 59.5200 0.8361 125.9463 71.192 <0.0000000000000002 ***
## dummy1 -10.6600 0.5710 100.0000 -18.670 <0.0000000000000002 ***
## dummy2 0.3200 1.1824 125.9463 0.271 0.787

301
 62

headache 58

measure
difference = −10.66 difference = −10.66+0.60 = −10.06
pre
54
post

50

WallStreetJ NYTimes
group

Figure 11.4: Expected headache levels in NY Times readers and Wall Street
Journal readers based on a linear mixed model with an interaction effect.

## dummy_interact 0.6000 0.8075 100.0000 0.743 0.459

## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Correlation of Fixed Effects:
## (Intr) dummy1 dummy2
## dummy1 -0.341
## dummy2 -0.707 0.241
## dummy_ntrct 0.241 -0.707 -0.341

The null-hypothesis that the effect of aspirin is the same in the

two populations of readers cannot be rejected, t(100) = 0.743, p =
0.459. We therefore conclude that there is no evidence that aspirin
has a different effect for NY Times than for Wall Street Journal
readers.

Note that we could have done the analysis in another way, not recoding
the variables into numeric dummy variables ourselves, but by letting R do it
automatically. R does that automatically for factor variables like our variable
group. The code is then:

model6 <- datalong %>%

lmer(headache ~ measure + group + measure:group + (1|patient),
data = .,
REML = FALSE)
model6 %>% summary()

302
## Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite’s
## method [lmerModLmerTest]
## Formula: headache ~ measure + group + measure:group + (1 | patient)
## Data: .
##
## AIC BIC logLik deviance df.resid
## 1201.7 1221.5 -594.8 1189.7 194
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -2.34085 -0.43678 -0.02942 0.50880 2.22534
##
## Random effects:
## Groups Name Variance Std.Dev.
## patient (Intercept) 26.80 5.177
## Residual 8.15 2.855
## Number of obs: 200, groups: patient, 100
##
## Fixed effects:
## Estimate Std. Error df t value
## (Intercept) 49.7800 0.8361 125.9463 59.542
## measurepre 10.0600 0.5710 100.0000 17.619
## groupWallStreetJ -0.9200 1.1824 125.9463 -0.778
## measurepre:groupWallStreetJ 0.6000 0.8075 100.0000 0.743
## Pr(>|t|)
## (Intercept) <0.0000000000000002 ***
## measurepre <0.0000000000000002 ***
## groupWallStreetJ 0.438
## measurepre:groupWallStreetJ 0.459
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Correlation of Fixed Effects:
## (Intr) mesrpr grpWSJ
## measurepre -0.341
## grpWllStrtJ -0.707 0.241
## msrpr:grWSJ 0.241 -0.707 -0.341

R has automatically created dummy variables, one dummy measurepre that

codes 1 for all measurements before aspirin intake, one dummy groupWallStreetJ
that codes 1 for all measurements from Wall Street Journal readers, and one
dummy measurepre:groupWallStreetJ that codes 1 for all measurements from
Wall Street Journal readers before aspirin intake. Because the dummy coding
is different from the hand coding we did ourselves earlier, the intercept and
the main effects of group and measure are now different. We also see that the

303
significance level of the interaction effect is still the same. You are always free
to choose to either construct your own dummy variables and analyse them in
a quantitative way (using numeric variables), or to let R construct the dummy
variables for you (by using a factor variable): the p-value for the interaction
effect will always be the same (this is not true for the intercept and the main
effects).
Because the two analyses are equivalent (they end up with exactly the same
predictions, feel free to check!), we can safely report that we have found a non-
significant group by measure interaction effect, t(98) = 0.74, p = 0.46. We
therefore conclude that we found no evidence that in the populations of NY
Times readers and Wall Street Journal readers, the short-term effect of aspirin
on headache is any different.

11.4 Mixed designs

The design in the previous section, where we had both a grouping variable and
a pre-post or repeated measures design, is often called a mixed design. It is a
mixed design in the sense that there are two kinds of variables: one is a between-
individuals variable, and one variable is a within-individual variable. Here the
between-individuals variable is group: two different populations of readers. It
is called between because one individual can only be part of one group. When
we study the effect of the group effect we are essentially comparing the scores of
one group of individuals with the scores of another group of individuals, so the
comparison is between different individuals. The two groups of data are said to
be independent, as we knew that none of the readers in this data set reads both
journals.
The within-variable in this design is the aspirin intervention, indicated by the
variable measure. For each individual we have two observations: all individuals
are present in both the pre condition data as well as in the post condition
data. With this intervention variable, we are comparing the scores of a group
of individuals with the scores of that same group of individuals at another time
point. The comparison of scores is within a particular individual, at time point
1 and at time point 2. So the pre and post sets of data are not independent:
the headache scores in both conditions are coming from the same individuals.
Mixed designs are often seen in psychological experiments. For instance, you
want to know how large the effect of alcohol intake is on driving performance.
You want to know whether the effect of alcohol on driving performance is the
same in a Fiat 600 as in a Porsche 918. Suppose you have 100 participants for
your study. There are many choices you can make regarding the design of your
study. Here we discuss 4 alternative research designs:

1. One option is to have all participants participate in all four conditions:

they all drive a Fiat with and without alcohol, and they all drive a Porsche,
with and without alcohol. In this case, both the car and the alcohol are
within-participant variables.

304
 2. The second option is to have 50 participants drive a Porsche, with and
without alcohol, and to have the other 50 participants drive the Fiat,
with and without alcohol. In this case, the car is the between-participants
variable, and alcohol is the within-participant variable.

3. The third option is to have 50 participants without alcohol drive both

the Porsche and the Fiat, and to have the other 50 participants drive the
Porsche and the Fiat with alcohol. Now the car is the within-participant
variable, and the alcohol is the between-participants variable.

4. The fourth option is to have 25 participants drive the Porsche with alcohol,
25 other participants drive the Porsche without alcohol, 25 participants
drive the Fiat with alcohol, and the remaining 25 participants drive the
Fiat without alcohol. Now both the car variable and the alcohol variable
are between-participant variables: none of the participants is present in
more than 1 condition.

Only the second and the third design described here are mixed designs,
having at least one between-participants variable and at least one within-participant
variable.
Remember that when there is at least one within variable in your design,
you have to use a linear mixed model. If all variables are between variables, one
can use an ordinary linear model. Note that the term mixed in linear mixed
model refers to the effects in the model that can be both random and fixed. The
term mixed in mixed designs refers to the mix of two kinds of variables: within
variables and between variables.
Also note that the within and between distinction refers to the units of
analysis. If the unit of analysis is school, then the denomination of the school is
a between-school variable. An example of a within-school variable could be time:
before a major curriculum reform and after a major curriculum reform. Or it
could be teacher: classes taught by teacher A or by teacher B, both teaching at
the same school.

11.5 Mixed design with a linear effect

In an earlier section we looked at a mixed design where the between variable
was group and the within variable was measure: pre or post. It was a 2 by 2
design (2 × 2) design: 2 measures and 2 groups, where we were interested in
the interaction effect. We wanted to know whether group moderated the effect
of measure, that is, the effect of aspirin on headache. We used the categorical
within-individual variable measure in the regression by dummy-coding it.
In an earlier section in this chapter we saw that we can also model linear
effects of numeric variables in linear mixed models, where we treated the time
variable numerically: 0hrs, 3hrs after aspirin intake and 24 hrs after intake. Here
we will give an example of a 3 × 20 mixed design: we have a qualitative group
(between-individuals) variable with 3 levels and a quantitative time (within)

305
variable with 20 levels. The example is about stress in athletes that are going
to partake in the 2018 Winter Olympics. Stress can be revealed in morning
cortisol levels. In the 20 days preceding the start of the Olympics, each athlete
was measured every morning after waking and before breakfast by letting them
chew on cotton. The cortisol level in the saliva was then measured in the lab.
Our research question is by how much cortisol levels rise in athletes that prepare
for the Olympics.
Three groups were studied. One group consisted of 50 athletes who were
selected to partake in the Olympics, one group consisted of 50 athletes that
were very good but were not selected to partake (Control group I) and one
group consisted of 50 non-athlete spectators that were going to watch the games
(Control group II). The research question was about what the differences are in
average cortisol increase in these three groups: the Olympians, Control group I
and Control group II.
In Table 11.7 you see part of the fictional data, the first 6 measurements on
person 1 that belongs to the group of Olympians.

Table 11.7: Cortisol measures over time.

person group measure cortisol
1 Olympian 1 19
1 Olympian 2 20
1 Olympian 3 22
1 Olympian 4 23
1 Olympian 5 24
1 Olympian 6 22

When we plot the data, and use different colours for the three different
groups, we already notice that the Olympians show generally higher cortisol
levels, particularly at the end of the 20-day period (Figure 11.5).
We want to know to what extent the linear effect of time is moderated by
group. Since for every person we have 20 measurements, the data are clustered
so we use a linear mixed model. We’re looking for a linear effect of time, so
we use the measure variable numerically (i.e., it is numeric, and we do not
transform it into a factor). We also use the categorical variable group as a
predictor, but in a qualitative way. It is a factor variable with three levels, so
that R will automatically make two dummy variables. Because we’re interested
in the interaction effects, we include both main effects of group and measure
and their interaction in the model. Lastly, we control for individual differences
in cortisol levels by introducing random effects for person.

model7 <- datalong %>%

lmer(cortisol ~ measure + group + measure:group + (1|person),
data = .,
REML = FALSE)
model7 %>% summary()

306
 40

35
group
cortisol

Control group I
30
Control group II
Olympian
25

20

5 10 15 20
measure

Figure 11.5: Cortisol levels over time in three groups.

## Linear mixed model fit by maximum likelihood . t-tests use Satterthwaite’s

## method [lmerModLmerTest]
## Formula: cortisol ~ measure + group + measure:group + (1 | person)
## Data: .
##
## AIC BIC logLik deviance df.resid
## 8969.8 9017.8 -4476.9 8953.8 2992
##
## Scaled residuals:
## Min 1Q Median 3Q Max
## -3.5823 -0.6405 -0.0116 0.6621 3.3154
##
## Random effects:
## Groups Name Variance Std.Dev.
## person (Intercept) 0.9655 0.9826
## Residual 0.9960 0.9980
## Number of obs: 3000, groups: person, 150
##
## Fixed effects:
## Estimate Std. Error df t value
## (Intercept) 20.094018 0.153650 202.476943 130.778
## measure 0.596989 0.005473 2850.000014 109.078
## groupControl group II -0.226802 0.217293 202.476944 -1.044
## groupOlympian -0.402950 0.217293 202.476944 -1.854
## measure:groupControl group II 0.006383 0.007740 2850.000006 0.825
## measure:groupOlympian 0.411896 0.007740 2850.000006 53.216
## Pr(>|t|)

307
## (Intercept) <0.0000000000000002 ***
## measure <0.0000000000000002 ***
## groupControl group II 0.2978
## groupOlympian 0.0651 .
## measure:groupControl group II 0.4096
## measure:groupOlympian <0.0000000000000002 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
##
## Correlation of Fixed Effects:
## (Intr) measur grCgII grpOly m:CgII
## measure -0.374
## grpCntrlgII -0.707 0.264
## groupOlympn -0.707 0.264 0.500
## msr:grpCgII 0.264 -0.707 -0.374 -0.187
## msr:grpOlym 0.264 -0.707 -0.187 -0.374 0.500

In the output we see an intercept of 20.09, a slope of 0.60 for the effect
of measure, two main effects for the variable group (Control group I is the
reference group), and two interaction effects (one for Control group II and one
for the Olympian group). Let’s fill in the linear model equation based on this
output:

cortisolij = 20.09 + personi + 0.60 measure − 0.23 ContrGrII −

0.40 Olympian + 0.006 ContrGrII measure + 0.41 Olympian measure + eij
personi ∼ N (0, σp2 = 0.97)
eij ∼ N (0, σe2 = 1.00)
0.9655
We see a clear intraclass correlation of around 0.9655+0.9960 = 0.49 so it’s
a good thing we’ve included random effects for person. The expected means
at various time points and for various groups can be made with the use of the
above equation.
It’s easier to look at what linear effects we have for the three different groups.
Filling in the above equation for Control group I (the reference group), we get:

cortisolij = 20.09 + personi + 0.60 × measure + eij

For Control group II we get:

cortisolij = 20.09 + personi + 0.60 measure − 0.23 + 0.006 measure + eij

= 19.86 + personi + 0.606 × measure + eij

And for the Olympians we get:

308
 40

35
group
cortisol

Control group I
30
Control group II
Olympian
25

20

5 10 15 20
measure

Figure 11.6: Cortisol levels over time in three groups with the group-specific
regression lines.

cortisolij = 20.09 + personi + 0.60 measure − 0.40 + 0.41 measure + eij

= 19.69 + personi + 1.01 × measure + eij
In these three equations, all intercepts are close to 20. The slopes are about
0.6 in both Control groups I and II, whereas the slope is around 1 in the group
of Olympian athletes. For illustration, these three implied linear regression lines
are depicted in Figure 11.6.
So based on the linear model, we see that in this sample the rise in cortisol
levels is much steeper in Olympians than in the two control groups. But is this
true for all Olympians and the rest of the populations of high performing athletes
and spectators? Note that in the regression table we see two interaction effects:
one for measure:groupControl group II and one for measure:groupOlympian.
Here we’re interested in the rise in cortisol in the three different groups and to
what extent these sample differences extend to the three populations.
A possible answer we find in the F -statistic of an ANOVA. When we run an
ANOVA on the results,

model7 %>% anova()

## Type III Analysis of Variance Table with Satterthwaite’s method
## Sum Sq Mean Sq NumDF DenDF F value Pr(>F)
## measure 54095 54095 1 2850.00 54313.3049 <0.0000000000000002 ***
## group 3 2 2 202.48 1.7285 0.1802
## measure:group 3703 1852 2 2850.00 1859.1587 <0.0000000000000002 ***
## ---
## Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

309
 we see a significant measure by group interaction effect, F (2, 2850) = 1859.16, p <
.001. The null-hypothesis of the same cortisol change in three different populations
can be rejected, and we conclude that Olympian athletes, non-Olympian athletes
and spectators show a different change in cortisol levels in the weeks preceding
the games.

310
Chapter 12

Non-parametric alternatives
for linear mixed models

12.1 Checking assumptions

In previous chapters we discussed the assumptions of linear models and linear
mixed models: linearity (in parameters), homoscedasticity (equal variance),
normal distribution of residuals, normal distribution of random effects (relevant
for linear mixed models only), and independence (no clustering unaccounted
for).
The problem of non-linearity can be solved by introducing quadratic terms,
for instance by replacing a linear model Y = b0 + b1 X + e by another linear
model Y = b0 + b1 X + b2 X 2 + e.
If we have non-independence, then you can introduce either an extra fixed
effect or a random effect for this clustering. For example, if you see that cars
owned by low income families have much more mileage than cars owned by high
income families, you can account for this by adding a fixed effect of an income
variable as predictor. If you see that average mileage is rather similar within
municipality but that average mileage can vary quite a lot across municipalities,
you can introduce a random effect for municipality (if you have data say from
30 different municipalities).
Unequal variance of residuals and non-normal distribution of residuals are
harder to tackle. Unequal variance can be tackled sometimes by using linear
models, but with more advanced options, or by making corrections to standard
errors and p-values that make inference more robust against model violations.
Violations of normality are even a bigger problem. Non-normality can sometimes
be solved by using generalised linear models (see Chapter 13). A combination
of non-normality and unequal variance can sometimes be solved by using a
transformation of the data, for instance
√ not analysing Y = b0 + b1 X + e but
analysing log(Y ) = b0 + b1 X + e or Y = b0 + b1 X + e.
If these data transformations or advanced options don’t work (or if you’re

311
 28

24
Time in minutes

20

16

EuropeanChampionships WorldChampionships Olympics

Occasion

Figure 12.1: Boxplot of the imaginary speed skating data.

not acquainted with them), and your data show non-equal variance and/or non-
normally distributed residuals, there are non-parametric alternatives. Here we
discuss two: Friedman’s test and Wilcoxon’s signed rank test. We explain them
using an imaginary data set on speed skating.

Suppose we have data on 12 speed skaters that participate on the 10 kilometres

distance in three separate championships in 2017-2018: the European Championships,
the Winter Olympics and the World Championships. Your friend expects that
speed skaters will perform best at the Olympic games, so there she expects the
fastest times. You decide to test the null-hypothesis that average times are the
same at the three occasions. In Figure 12.1 we see a box plot of the data.
In order to test this null-hypothesis, we run a linear mixed model with
dependent variable time, and independent variable occasion. We use random
effects for the differences in speed across skaters. In Figure 12.2 we see the
residuals. From this plot we clearly see that the assumption of equal variance
(homogeneity of variance) is violated: the variance of the residuals in the World
Championships condition is clearly smaller than the variance of the European
Championships condition. From the histogram of the residuals in Figure 12.3 we
also see that the distribution of the residuals is not bell-shaped: it is positively
skewed (skewed to the right).
Since the assumptions of homogeneity of variance and of normally distributed
residuals are violated1 , the results from the linear mixed model cannot be
trusted. In order to answer our research question, we therefore have to resort
1 Remember that assumptions relate to the population not samples: often-times your data

set is too small to say anything about assumptions at the population level. Residuals for a
data set of 8 persons might show very non-normal residuals, or very different variances for two
subgroups of 4 persons each, but that might just be a coincidence, a random result because
of the small sample size. If in doubt, it is best to use non-parametric methods.

312
 5
Residual

−5
EuropeanChampionships WorldChampionships Olympics
Occasion

Figure 12.2: Residuals of the speedskating data with a linear mixed model.

4
Count

0
−5 0 5
Residual

Figure 12.3: Histogram of the residuals of the speedskating data with a linear
mixed model.

313
to another kind of test. Here we discuss Friedman’s test, a non-parametric test,
for testing the null-hypothesis that the medians of the three groups of data
are the same (see Chapter 1. This Friedman test can be used in all situations
where you have at least 2 levels of the within variable. In other words, you
can use this test when you have data from three occasions, but also when you
have data from 10 occasions or only 2. In a later section the Wilcoxon signed
ranks test is discussed. This test is often used in social and behavioural sciences.
The downside of this test is that it can only handle data sets with 2 levels of
the within variable. In other words, it can only be used when we have data
from two occasions. Friedman’s test is therefore more generally applicable than
Wilcoxon’s. We therefore advise to always go with the Friedman test, but for
the sake of completeness, we will also explain the Wilcoxon test.

12.2 Friedman’s test for k measures

Similar to many other non-parametric tests for testing the equality of medians,
Friedman’s test is based on ranks. Table 12.1 shows the speed skating data in
wide format.

Table 12.1: The speed skating data in wide format.

athlete EuropeanChampionships WorldChampionships Olympics
1 14.35 15.79 16.42
2 17.36 14.26 18.13
3 19.01 18.37 19.95
4 27.90 15.12 17.78
5 17.67 17.17 16.96
6 17.83 15.30 16.15
7 16.30 15.63 19.44
8 28.00 15.69 16.23
9 18.27 15.65 15.76
10 17.00 14.99 16.18
11 17.10 15.83 13.89
12 18.94 14.77 14.83

We rank all of these time measures by determining the fastest time, then the
next to fastest time, etcetera, until the slowest time. But because the data in
each row belong together (we compare individuals with themselves), we do the
ranking row-wise. For each athlete separately, we determine the fastest time
(1), the next fastest time (2), and the slowest time (3) and put the ranks in
a new table, see Table 12.2. There we see for example that athlete 1 had the
fastest time at the European Championships (14.35, rank 1) and the slowest at
the Olympics (16.42, rank 3).
Next, we compute the sum of the ranks column-wise: the sum of the ranks
for the European Championships data is 31, for the Olympic data it’s 15 and
for the World Championships data it is 26.

314
 Table 12.2: Row-wise ranks of the speed skating data.
athlete EuropeanChampionships WorldChampionships Olympics
1 1 2 3
2 2 1 3
3 2 1 3
4 3 1 2
5 3 2 1
6 3 1 2
7 2 1 3
8 3 1 2
9 3 1 2
10 3 1 2
11 3 2 1
12 3 1 2

From these sums we can gather that in general, these athletes showed their
best times (many rank 1s) at the World Championships, as the sum of the ranks
is lowest. We also see that in general these athletes showed their worst times
(many rank 2s and 3s) at the European Championships, as the relevant column
showed the highest sum of ranks.
In order to know whether these sums of ranks are significantly different from
each other, we may compute an Fr -value based on the following formula:
 
12
Fr = Σkj=1 Sj2 − 3n(k + 1) (12.1)
nk(k + 1)
In this formula, n stands for the number of rows (12 athletes), k stands for
the number of columns (3 occasions), and Sj2 stands for the squared sum of
column j (312 , 262 and 152 ). If we fill in these numbers, we get:

 
12
Fr = × (312 + 152 + 262 ) − 3 × 12(3 + 1)
12 × 3(3 + 1)
 
12
= × 1862 − 144 = 11.167
144

What can we tell from this Fr -statistic? In order to say something about
significance, we have to know what values are to be expected under the null-
hypothesis that there are no differences across the three groups of data. Suppose
we randomly mixed up the data by taking all the speed skating times and
randomly assigning them to the three contests and the twelve athletes, until we
have a newly filled data matrix, for example the one in Table 12.3.
If we then compute Fr for this data matrix, we get a different value. If we do
this mixing up the data and computing Fr say 1000 times, we get 1000 values
for Fr , summarised in the histogram in Figure 12.4.

315
 Table 12.3: The raw skating data in random order.
athlete EuropeanChampionships WorldChampionships Olympics
1 15.76 14.26 17.78
2 19.44 17.10 17.83
3 16.18 14.83 15.63
4 15.30 17.00 16.42
5 15.79 16.23 17.36
6 14.77 28.00 15.65
7 14.35 16.15 16.30
8 27.90 15.12 18.94
9 16.96 19.01 19.95
10 15.83 17.17 13.89
11 14.99 18.27 15.69
12 18.37 17.67 18.13

So if the data are just randomly distributed over the three columns (and
12 rows) in the data matrix, we expect no systematic differences across the
three columns and so the null-hypothesis is true. So now we know what the
distribution of Fr looks like when the null-hypothesis is true: more or less like
the one in Figure 12.4. Remember that for the true data that we actually
gathered (in the right order that is!), we found an Fr -value of 11.167 . From
the histogram, we see that only very few values of 11.167 or larger are observed
when the null-hypothesis is true. If we look more closely, we find that only 0.2%
of the values are larger than 11.167 , so we have a p-value of 0.004. The 95th
percentile of these 1000 Fr -values is 6.1666667, meaning that of the 1000 values
for Fr , 5% are larger than 6.1666667. So if we use a significance level of 5%, our
observed value of 11.167 is larger than the critical value for Fr , and we conclude
that the null-hypothesis can be rejected.
Now this p-value of 0.004 and the critical value of 6.1666667 are based on our
own computations2 . Actually there are better ways. One is to look up critical
values of Fr in tables, for instance in Kendall M.G. (1970) Rank correlation
methods. (fourth edition). The p-value corresponding to this Fr -value depends
on k, the number of groups of data (here 3 columns) and n, the number of rows
(12 individuals). If we look up that table, we find that for k = 3 and n = 12
the critical value of Fr for a type I error rate of 0.05 equals 6.17. Our observed
Fr -value of 11.167 is larger than that, therefore we can reject the null-hypothesis
that the median skating times are the same at the three different championships.
So we have to tell your friend that there are general differences in skating times
at different contests, Fr = 11.167, p < 0.05, but it is not the case that the fastest
times were observed at the Olympics.
2 What we have actually done is a very simple form of bootstrapping: jumbling up the data

set many times and in that way determining the distribution of a test-statistic under the
null-hypothesis, in this case the distribution of Fr . For more on bootstrapping, see Davison,
A.C. & Hinkley, D.V. (1997). Bootstrap Methods and their Application. Cambridge, UK:
Cambridge.

316
 200

150
count

100

50

0
0 5 10
F_r

Figure 12.4: Histogram of 1000 possible values for Fr given that the null-
hypothesis is true, for 12 speed skaters.

Another way is to make an approximation of the distribution of Fr . Note

that the distribution in the histogram is very strangely shaped. The reason is
that the data set is quite limited. Suppose we have not data on 12 speed skaters,
but on 120. If we then randomly mix up data again and compute 1000 different
values for Fr , we get the histogram in Figure 12.5.
The shape becomes more regular. It also starts to resemble another distribution,
that of the χ2 (chi-square). It can be shown that the distribution of the Fr for
a large number of rows in the data matrix, and at least 6 columns, approaches
the shape of the χ2 -distribution with k − 1 degrees of freedom. This is shown
in Figure 12.6.
The line of the χ2 -distribution with 2 degrees of freedom approaches the
histogram quite well, but not perfectly. In general, for large n and k > 5, the
approximation is good enough. In that way it gets easier to look up p-values
for certain Fr -values, because the χ2 -distribution is well-known3 , so we don’t
have to look up critical values for Fr in old tables. For a significance level of
5%, the critical value of a χ2 with 2 degrees of freedom is 5.991. This is close to
the value in the table for Fr in old books: 6.17. The part of the χ2 -distribution
with 2 degrees of freedom that is larger than the observed 11.167 is 0.004, so
our approximate p-value for our null-hypothesis is 0.004.

# critical value for chi-square distribution

# with 2 degrees of freedom and alpha = 0.05:
qchisq(df = 2, p = 0.95)

3 The χ2 -distribution is based on the normal distribution: the χ2 -distribution with k degrees

of freedom is the distribution of a sum of the squares of k independent standard normal random
variables.

317
 200

150
count

100

50

0
0 5 10 15
F_r

Figure 12.5: Histogram of 1000 possible values for Fr given that the null-
hypothesis is true, for 120 speed skaters.

0.5

0.4

0.3
density

0.2

0.1

0.0
0 5 10 15
F_r

Figure 12.6: The distrbution of Fr under the null-hypothesis, overlain with a

chi-square distribution with 2 degrees of freedom.

318
## [1] 5.991465

# p-value for statistic of 11.167 based on

# chi-square distribution with 2 degrees of freedom:
pchisq(11.167 , df = 2, lower.tail = F)

## [1] 0.003759385

12.3 How to perform Friedman’s test in R

First of all, you need your data to be in long format. If your data happens to be
in wide format, use the pivot longer() function to get the data in long format.
Suppose your data is in wide format, as in Table 12.1. Then the following code
turns the data into long format:

datawide %>%
pivot_longer(cols = -athlete, names_to = "occasion", values_to = "time")

This creates the long format data matrix in Table 12.4:

Table 12.4: The raw skating data in long data format.

athlete occasion time
1 EuropeanChampionships 14.35
1 Olympics 16.42
1 WorldChampionships 15.79
2 EuropeanChampionships 17.36
2 Olympics 18.13
2 WorldChampionships 14.26

We can then specify that we want Friedman’s test by using the friedman.test()
function and indicating which variables we want to use:

datalong %>%
friedman.test(time ~ occasion | athlete, data = .)

##
## Friedman rank sum test
##
## data: time and occasion and athlete
## Friedman chi-squared = 11.167, df = 2, p-value = 0.00376

This code says, ”use time as dependent variable, occasion as independent

variable, and the data are clustered within athletes (data with the same athlete
ID number belong together)”.

319
 In the output we see a chi-squared statistic, degrees of freedom, and an
asymptotic (approximated) p-value. Why don’t we see an Fr -statistic?
The reason is, as discussed in the previous section, that for large number of
measurements (in wide format: columns) and a large number of individuals (in
wide format: rows), the Fr statistic tends to have the same distribution as a
chi-square, χ2 , with k − 1 degrees of freedom. So what we are looking at in this
output is really an Fr -value of 11.167 (exactly the same value as we computed
by hand in the previous section). In order to approximate the p-value, this value
of 11.167 is interpreted as a chi-square (χ2 ), which with 2 degrees of freedom
has a p-value of 0.004.
This asymptotic (approximated) p-value is the correct p-value if you have a
lot of rows (large n) and at least 6 variables (k > 5). If you do not have that,
as we have here, this asymptotic p-value is only what it is: an approximation.
However, this is only a problem when the approximate p-value is close to the
pre-selected significance level α. If α equals 0.05, an approximate p-value of
0.002 is much smaller than that, and we do not hesitate to call it significant,
whatever its true value may be. If a p-value is very close to α, it might be a
good idea to look up the exact critical values for Fr in online tables4 . If your
Fr is larger than the critical value for a certain combination of n, k and α, you
may reject the null-hypothesis.
In the above case we can report:

A Friedman’s test showed that speed skaters have significantly

different median times on the 10 kilometre distance at the three
types of contests, Fr = 11.167, p < 0.05.

12.4 Wilcoxon’s signed ranks test for 2 measures

Friedman’s test can be used for 2 measures, 3 measures or even 10 measures. As
stated earlier, the well-known Wilcoxon’s test can only be used for 2 measures.
For completeness, we also discuss that test here. For that test we compare the
measures at the Olympics and at the World Championships.

For each athlete, we take the difference in skating times (Olympics - WorldChampionships)
and call it d, see Table 12.5. Next we rank these d-values, irrespective of sign,
and call these ranks rank d. From Table 12.5 we see that athlete 12 shows
the smallest difference in skating times (d = 0.06, rank = 1) and athlete 2 the
largest difference (d = 3.78, rank = 12).
Next, we indicate for each rank whether it belongs to a positive or a negative
difference d and call that variable ranksign.
Under the null-hypothesis, we expect that some of the larger d-values are
positive and some of them negative, in a fairly equal amount. If we sum the
ranks having plus-signs and sum the ranks having minus-signs, we would expect
that these two sums are about equal, but only if the null-hypothesis is true. If
4 https://www.jstor.org/stable/3315656?seq=1

320
Table 12.5: The raw skating data and the computations for Wilcoxon signed
ranks test
athlete Olympics WorldChampionships d rank d ranksign
1 16.42 15.79 0.63 5.00 5.00
2 18.13 14.26 3.87 12.00 12.00
3 19.95 18.37 1.58 8.00 8.00
4 17.78 15.12 2.66 10.00 10.00
5 16.96 17.17 -0.21 3.00 -3.00
6 16.15 15.30 0.85 6.00 6.00
7 19.44 15.63 3.81 11.00 11.00
8 16.23 15.69 0.54 4.00 4.00
9 15.76 15.65 0.11 2.00 2.00
10 16.18 14.99 1.19 7.00 7.00
11 13.89 15.83 -1.94 9.00 -9.00
12 14.83 14.77 0.06 1.00 1.00

the sums are very different, then we should reject this null-hypothesis. In order
to see if the difference in sums is too large, we compute them as follows:

T+ = 5 + 12 + 8 + 10 + 6 + 11 + 4 + 2 + 7 + 1 = 66
T− = 3 + 9 = 12
To know whether T + is significantly larger than T − , the value of T + can be
looked up in a table, for instance in Siegel & Castellan (1988). There we see
that for T + , with 12 rows, the probability of obtaining a T + of at least 66 is
0.0171. For a two-sided test (if we would have switched the columns of the two
championships, we would have gotten a T − of 66 and a T + of 12!), we have to
double this probability. So we end up with a p-value of 2 × 0.0171 = 0.0342.
In the table we find no critical values for large sample size n, but fortunately,
similar to the Friedman test, we can use an approximation using the normal
distribution. It can be shown that for large sample sizes, the statistic T + is
approximately normally distributed with mean

n(n + 1)
µ= (12.2)
4
and variance:

n(n + 1)(2n + 1)
σ2 = (12.3)
24
If we therefore standardise the Tp+ by subtracting the µ and then dividing
by the square root of the variance (σ 2 ) = σ, we get a z-value with mean 0
and standard deviation 1. To do that, we use the following formula:

T+ − µ T + − n(n + 1)/4
z= =p (12.4)
σ n(n + 1)(2n + 1)/24

321
 Here T + is 66 and n equals 12, so if we fill in the formula we get z = 2.118.
From the standard normal distribution we know that 5% of the observations lie
above 1.96 and below -1.96. So a value for z larger than 1.96 or smaller than
-1.96 is enough evidence to reject the null-hypothesis. Here our z-statistic is
larger than 1.96, therefore we reject the null-hypothesis that the median skating
times are the same at the World Championships and the Olympics. The p-value
associated with a z-score of 2.118 is 0.034.

12.5 How to perform Wilcoxon’s signed ranks

test in R
If you want to use the Wilcoxon test, your data needs to be in wide format.
If your data are in long format, you can transform them with the function
pivot wider():

datawide <- datalong %>%

pivot_wider(values_from = "time", names_from = "occasion")

Next, you use the wilcoxon.test() function. You select the two variables
(occasions) that you would like to compare, and indicate that the Olympic and
World Championship data are paired within rows (i.e., the Olympic and World
championship data in one row belong to the same individual).

wilcox.test(datawide$Olympics, datawide$WorldChampionships,
paired = TRUE)

##
## Wilcoxon signed rank exact test
##
## data: datawide$Olympics and datawide$WorldChampionships
## V = 66, p-value = 0.03418
## alternative hypothesis: true location shift is not equal to 0

The output shows a V -statistic, which corresponds to our T + above. The

standard output yields an approximate p-value. This p-value is by default two-
sided. Wilcoxon’s T + -statistic (or V ) under the null-hypothesis approaches a
normal distribution in case we have a large number of observations (many rows
in wide format). If n > 15, the approximation is good enough so that if we
standardise this statistic, it can be interpreted as a z-score (standardised score
with a normal distribution). That means that a z-score of 1.96 or larger or -1.96
or smaller can be regarded as significant at the 5% significance level. Since the
standard normal distribution is only an approximation, and we have n = 12, it
is safer here to look at the exact significance level. That can be done with the
option exact = TRUE:

322
wilcox.test(datawide$Olympics, datawide$WorldChampionships,
paired = TRUE,
exact = TRUE)

##
## Wilcoxon signed rank exact test
##
## data: datawide$Olympics and datawide$WorldChampionships
## V = 66, p-value = 0.03418
## alternative hypothesis: true location shift is not equal to 0

In this case we see that the exact p-value is equal to the approximate p-value.
Note that we use a two-sided test, to allow for the fact that random sampling
could lead to a higher median for the Olympic Games or a higher median for
the World Championships. We just want to know whether the null-hypothesis
that the two medians differ can be rejected (in whatever direction) or not.

Let’s compare the output with the Friedman test, but then only use the
relevant variables in your code:

datalong %>%
filter(occasion != "EuropeanChampionships") %>%
mutate(occasion = as.integer(occasion)) %>%
friedman.test(time ~ occasion | athlete,
data = .)

##
## Friedman rank sum test
##
## data: time and occasion and athlete
## Friedman chi-squared = 5.3333, df = 1, p-value = 0.02092

Note that the friedman.test() function does not perform well if some
variables are factors and you make a selection of the levels. Here we have the
factor occasion with originally 3 levels. If the friedman.test() function only
finds two of these in the data it is supposed to analyse, it throws an error.
Therefore we turn factor variable occasion into an integer variable first. The
friedman.test() function then turns this integer variable into a new factor
variable with only 2 levels before analysis.
In the output we see that the null-hypothesis of equal medians at the World
Championships and the Olympic Games can be rejected, with an approximate
p-value of 0.02.
Note that both the Friedman and Wilcoxon tests come up with very similar
p-values. Their rationales are also similar: Friedman’s test is based on ranks and
Wilcoxon’s test is based on positive and negative differences between measures

323
1 and 2, so in fact ranks 1 and 2 for each row in the wide data matrix. Both
can therefore be used in the case you have two measures. We recommend to
use the Friedman test, since that test can be used in all situations where you
have 2 or more measures per row. Wilcoxon’s test can only be used if you have
2 measures per row.

In sum, we can report in two ways on our hypothesis regarding similar skating
times at the World Championships and at the Olympics:

1. A Friedman test showed a significant difference between the

10km skating times at the World Championships and at the
Olympics, Fr = 5.33, p = 0.02.

2. A Wilcoxon signed ranks test showed a significant difference

between the 10km skating times at the World Championships
and at the Olympics, T + = 66, p = 0.03.

How do we know whether the fastest times were at the World Championships
or at the Olympics? If we look at the raw data in Table 12.1, it is not
immediately obvious. We have to inspect the T + = 66 and T − = 12 and
consider what they represent: there is more positivity than negativity. The
positivity is due to positive ranksigns that are computed based on d = Olympics−
W orldChampionships, see Table 12.5. A positive difference d means that the
Olympics time was larger than the WorldChampionships time. A large value
for time stands for slower speed. A positive ranksign therefore means that the
Olympics time was larger (slower!) than the WorldChampionships time. A
large rank d means that the difference between the two times was relatively
large. Therefore, you get a large value of T + if the Olympic times are on the
whole slower than the World Championships times, and/or when these positive
differences are relatively large. When we look at the values of ranksign in Table
12.5, we notice that only two values are negative: one relatively large value and
one relatively small value. The rest of the values are positive, both small and
large, and these all contribute to the T + value. We can therefore state that the
pattern in the data is that for most athletes, the Olympic times are slower than
the times at the World Championships.

12.6 Ties
Many non-parametric tests are based on ranks. For example, if we have the data
sequence 0.1, 0.4, 0.5, 0.2, we give these values the ranks 1, 3, 4, 2, respectively.
But in many data cases, data sequences cannot be ranked unequivocally. Let’s
look at the sequence 0.1, 0.4, 0.4, 0.2. Here we have 2 values that are exactly
the same. We say then that we have ties. If we have ties in our data like the
0.4 in this case, one very often used option is to arbitrarily choose one of the
0.4 values as smaller than the other, and then average the ranks. Thus, we rank
the data into 1, 3, 4, 2 and then average the tied observations: 1, 3.5, 3.5, 2. As

324
another example, suppose we have the sequence 23, 54, 54, 54, 19, we turn this
into ranks 2, 3, 4, 5, 1 and take the average of the ranks of the tied observations
of 54: 2, 4, 4, 4, 1. These ranks corrected for ties can then be used to compute
the test statistic, for instance Friedman’s Fr or Wilcoxon’s z. However, in many
cases, because of these corrections, a slightly different formula is to be used. So
the formulas become a little bit different. This is all done in R automatically.
If you want to know more, see Siegel and Castellan (1988). Non-parametric
Statistics for the Behavioral Sciences. New York: McGraw-Hill.

325
326
Chapter 13

Generalised linear models:

logistic regression

13.1 Introduction
In previous chapters we were introduced to the linear model, with its basic form

Y = b0 + b1 X1 + · · · + bn Xn + e (13.1)
e ∼ N (0, σe2 ) (13.2)

Two basic assumptions of this model are the additivity in the parameters,
and the normally distributed residual e. Additivity in the parameters means
that the effects of intercept and the independent variables X1 , X2 , . . . Xn are
additive: the assumption is that you can sum these effects to come to a predicted
value for Y . So that is also true when we include an interaction effect to account
for a moderation,

Y = b0 + b1 X1 + b2 X2 + b3 X1 X2 + e (13.3)
e ∼ N (0, σe2 ) (13.4)

or when we use a quadratic term to account for another type of non-linearity

in the data:

Y = b0 + b1 X1 + b2 X1 X1 + e (13.5)
e ∼ N (0, σe2 ) (13.6)

In all these models, the assumption is that the effects of the parameters (b0 ,
b1 , b2 ) can be summed.

327
 0.20

0.15

Inflexion point Inflexion point

density

0.10

0.05

0.00
−6 −5 −4 −3 −2 −1 0 1 2 3 4 5 6
residual

Figure 13.1: Density function of the normal distribution, with mean 0 and
variance 4 (standard deviation 2). Inflexion points are positioned at residual
values of minus 1 standard deviation and plus 1 standard deviation.

The other major assumption of linear (mixed) models is the normal distribution
of the residuals. As we have seen in for instance Chapter 7, sometimes the
residuals are not normally distributed. Remember that with a normal distribution
N (0, σ 2 ), in principle all values between −∞ and +∞ are possible, but they tend
to concentrate around the value of 0, in the shape of the bell-curve. Figure 13.1
shows the normal distribution N (0, σ 2 = 4): it is centred around 0 and has
variance 4. Note that the inflexion point, that is the point where the decrease
in density tends to decelerate, is exactly at the values -2 and +2. These are
equal to the square root of the variance, which is the standard deviation, −σ
and +σ.
A normal distribution is suitable for continuous dependent variables. For
most measured variables this is not true. Think for example of temperature
measures: if the thermometer gives degrees Celsius with a precision of only 1
decimal, we can never have values of say 10.07 or -56.789. Our actual data will
in fact be discrete, showing rounded values like 10.1, 10.2, 10.3, but never any
values in between.
Nevertheless, the normal distribution can still be used in many such cases.
Take for instance a data set where the temperature in Amsterdam in summer
was predicted on the basis of a linear model. Fig 13.2 shows the distribution
of the residuals for that model. The temperature measures were discrete with
a precision of one tenth of a degree Celsius, but the distribution seems well
approximated by a normal curve.
But let’s look at an example where the discreteness is more prominent. In
Figure 13.3 we see the residuals of an analysis of exam results. Students had
to do an assignment that had to meet 4 criteria: 1) originality, 2) language, 3)

328
 0.15

0.10
density

0.05

0.00
−10 −5 0 5 10
residual

Figure 13.2: Even if residuals are really discrete, the normal distribution can be
a good approximation of their distribution.

structure, and 4) literature review. Each criterion was scored as either fulfilled
(1) or not fulfilled (0). The score for the assignment was determined on the
basis of the number of criteria that were met, so the scores could be 0, 1, 2, 3
or 4. In an analysis, this score was predicted on the basis of the average exam
score on previous assignments, using a linear model.
Figure 13.3 shows that the residuals are very discrete, and that the continuous
normal distribution is a very bad approximation of the histogram. We often see
this phenomenon when our data consist of counts with a limited maximum
number.
An even more extreme case we observe when our dependent variable consists
of whether or not students passed the assignment: only those assignments that
fulfilled all 4 criteria are regarded as sufficient. If we score all students with a
sufficient assignment as passed (scored as a value of 1) and all students with
an insufficient assignment as failed (scored as a value of 0) and we predict this
score by the average exam score on previous assignments using a linear model,
we get the residuals displayed in Figure 13.4.
Here it is also evident that a normal approximation of the residuals will not
do. When the dependent variable has only 2 possible values, a linear model
will never work because the residuals can never have a distribution that is even
remotely looking normal.
In this chapter and the next we will discuss how generalised linear models
can be used to analyse data sets where the assumption of normally distributed
residuals is not tenable. First we discuss the case where the dependent variable
has only 2 possible values (dichotomous dependent variables like yes/no or
pass/fail, heads/tails, 1/0). In Chapter 14, we will discuss the case where the
dependent variable consists of counts (0, 1, 2, 3, 4, . . . ).

329
 3

2
density

0
−1 0 1 2 3
residual

Figure 13.3: Count data example where the normal distribution is not a good
approximation of the distribution of the residuals.

10.0

7.5
density

5.0

2.5

0.0
−0.25 0.00 0.25 0.50 0.75 1.00
residual

Figure 13.4: Dichotomous data example where the normal distribution is not a
good approximation of the distribution of the residuals.

330
 1.00

0.75
score

0.50

0.25

0.00
75 80 85
age in months

Figure 13.5: Data example: Exam outcome (score) as a function of age, where
1 means pass and 0 means fail.

13.2 Logistic regression

Imagine that we analyse results on an exam for third grade children. These
children are usually either 6 or 7 years old, depending on what month they were
born in. The exam is on February 1st. A researcher wants to know whether
the age of the child can explain why some children pass the test and others fail.
She computes the age of the child in months. Each child that passes the exam
gets a score of 1 and all the others get a score of 0. Figure 13.5 plots the data.
She wants to use the following linear model:

score = b0 + b1 age + e (13.7)

e ∼ N (0, σe2 ) (13.8)
Figure 13.6 shows the data with the estimated regression line and Figure
13.7 shows the distribution of the residuals as a function of age.
Clearly a linear model is not appropriate. Here, the assumption that the
dependent variable, score in this case, is scattered randomly around the predicted
value with a normal distribution is not reasonable. The main problem is that
the dependent variable score can only have 2 values: 0 and 1. When we have
a dependent variable that is categorical, so not continuous, we generally use
logistic regression. In this chapter we cover the case when the dependent variable
takes binary values, like 0 and 1.

13.2.1 Bernoulli distribution

Rather than using a normal distribution, we could try a Bernoulli distribution.
The Bernoulli distribution is the distribution of a coin flip. For example, if the

331
 0.8
score

0.4

0.0

75 80 85
age in months

Figure 13.6: Example exam data with a linear regression line.

0.5
residual

0.0

−0.5

−1.0

75 80 85
age in months

Figure 13.7: Residuals as a function of age, after a linear regression analysis of

the exam data.

332
probability of heads is 0.1, we can expect that if we flip the coin 100 times, on
average we expect to see 10 times heads and 90 times tails. Our best bet then
is that the outcome is tails. However, if we actually flip the coin, we might see
heads anyway. There is some randomness to be expected. Let Y be the outcome
of a coin flip: heads or tails. If we have a Bernoulli distribution for variable Y
with probability p for heads, we expect to see heads p times, but we actually
observe heads or tails (Y ).

Y ∼ Bern(p) (13.9)
The same is true for the normal distribution in the linear model case: we
expect that the observed value of Y is exactly equal to its predicted value (b0 +
b1 X), but we observe Y that it is most often different.

Y ∼ N (µ = b0 + b1 X, σe2 ) (13.10)
In our example of passing the exam by the third graders, the pass rate could
also be conceived as the outcome of a coin flip: pass instead of heads and fail
instead of tails. So would it be an idea to predict the probability of passing the
exam on the basis of age? And then for every predicted probability, we allow for
the fact that actually the observed success can differ. Our linear model could
then look like this:

pi = b0 + b1 agei (13.11)
scorei ∼ Bern(pi ) (13.12)

So for each child i, we predict the probability of success, pi , on the basis

of her/his age. Next, the randomness in the data comes from the fact that a
probability is only a probability, so that the observed success of a child scorei ,
is like a coin toss with probability of pi for success.
For example, suppose that we have a child with an age of 80 months, and
we have b0 = −3.8 and b1 = 0.05. Then the predicted probability pi is equal to
−3.8 + 0.05 × 80 = 0.20. The best bet for such a child would be that it fails the
exam. But, although 0.20 is a small probability, there is a chance that the child
passes the exam. This model also means that if we would have 100 children of
age 80 months, we would expect that 20 of these children would pass the test
and 80 would fail. But we can’t make exact predictions for one individual alone:
we don’t know exactly which child will pass and which child won’t. Note that
this is similar to the normally distributed residual in the linear model: in the
linear model we expect a child to have a certain value for Y , but we know that
there will be a deviation from this predicted value: the residual. For a whole
group of children with the same predicted value for Y , we know that the whole
group will show residuals that have a normal distribution. But we’re not sure
what the residual will be for each individual child.
Unfortunately, this model for probabilities is not very helpful. If we use
a linear model for the probability, this means that we can predict probability

333
values of less than 0 and more than 1, and this is not possible for probabilities.
If we use the above values of b0 = −3.8 and b1 = 0.05, we predict a probability
of -.3 for a child of 70 months and a probability of 1.2 for a child of 100 months.
Those values are meaningless, since probabilities are always between 0 and 1!

13.2.2 Odds and logodds

Instead of predicting probabilities, we could predict odds. The nice property of
odds is that they can have very large values, much larger than 1.
What are odds again? Odds are a different way of talking about probability.
Suppose the probability of winning the lottery is 1%. Then the probability of
loosing is 99%. This is equal to saying that the odds of winning against loosing
are 1 to 99, or 1 : 99, because the probability of winning is 99 times smaller
than the probability of loosing.
As another example, suppose the probability of being alive tomorrow is equal
to 0.9999. Then the probability of not being alive tomorrow is 1 − 0.9999 =
0.0001. Then the probability of being alive tomorrow is 0.9999/0.0001 = 9999
times larger than the the probability of not being alive. Therefore the odds of
being alive tomorrow against being dead is 9999 to 1 (9999:1).
If we have a slightly biased coin, the probability of heads might be 0.6. The
probability of tails is then 0.4. Then the probability of heads is then 1.5 times
larger than the probability of heads (0.6/0.4=1.5). So the odds of heads against
tails is then 1.5 to 1. For the sake of clarity, odds are often multiplied by a
constant to get integers, so we can also say the odds of heads against tails are 3
to 2. Similarly, if the probability of heads were 0.61, the odds of heads against
tails would be 0.61 to 0.39, which can be modified into 61 to 39.
Now that we know how to go from probability statements to statements
about odds, how do we go from odds to probability? If someone says the odds
of heads against tails is 10 to 1, this means that for every 10 heads, there will
be 1 tails. In other words, if there were 11 coin tosses, 10 would be heads and 1
would be tails. We can therefore transform odds back to probabilities by noting
that 10 out of 11 coin tosses is heads, so 10/11 = 0.91, and 1 out of 11 is tails,
so 1/11 = 0.09.
If someone says the odds of winning a gold medal at the Olympics is a
thousand to one (1000:1), this means that if there were 1000 + 1 = 1001
opportunities, there would be a gold medal in 1000 cases and failure in only
one. This corresponds to a probability of 1000/1001 for winning and 1/1001 for
failure.
As a last example, if at the horse races, the odds of Bruno winning against
Sacha are four to five (4:5), this means that for every 4 winnings by Bruno,
there would be 5 winnings by Sacha. So out of a total of 9 winnings, 4 will be
by Bruno and 5 will be by Sacha. The probability of Bruno outrunning Sacha
is then 4/9 = 0.44.

If we would summarise the odds by doing the division, we have just one number.
For example, if the odds are 4 to 5 (4:5), the odds are 4/5 = 0.8, and if the odds

334
 natural logarithm of the odds

−4

0.00 0.25 0.50 0.75 1.00

probability

Figure 13.8: The relationship between a probability and the natural logarithm
of the corresponding odds.

are a thousand to one (1000:1), then we can also say the odds are 1000. Odds,
unlike probabilities, can have values that are larger than 1.
However, note that odds can never be negative: a very small odds is one to
a thousand (1:1000). This can be summarised as an odds of 0.000999001, but
that is still larger than 0. In summary: probabilities range from 0 to 1, and
odds from 0 to infinity.
Because odds can never be negative, mathematicians have proposed to use
the natural logarithm 1 of the odds as the preferred transformation of probabilities.
For example, suppose we have a probability of heads of 0.42. This can be
transformed into an odds by noting that in 100 coin tosses, we would expect
42 times heads and 58 times tails. So the odds are 42:58, which is equal to
42
58 = 0.7241379. The natural logarithm of 0.7241379 equals -0.3227734 (use the
ln button on your calculator!). If we have a value between 0 and 1 and we take
the logarithm of that value, we always get a value smaller than 0. In short: a
probability is never negative, but the corresponding logarithm of the odds can
be negative.
Figure 13.8 shows the relationship between a probability (with values between
0 and 1) and the natural logarithm of the corresponding odds (the logodds).
The result is a mirrored S-shaped curve on its side. For large probabilities close
to one, the equivalent logodds becomes infinitely positive, and for very small
probabilities close to zero, the equivalent logodds becomes infinitely negative.
1 The natural logarithm of a number is its logarithm to the base of the constant e, where

e is approximately equal to 2.7. The natural logarithm of x is generally written as lnx or

log e x. The natural logarithm of x is the power to which e needs to be raised to equal x.
For example, ln(2) is 0.69, because e0.69 = 2, and ln(0.2) = −1.6 because e−1.6 = 0.2. The
natural logarithm of e itself, ln(e), is 1, because e1 = e, while the natural logarithm of 1,
ln(1), is 0, since e0 = 1.

335
A logodds of 0 is equal to a probability of 0.5. If a logodds is larger than 0,
it means the probability is larger than 0.5, and if a logodds is smaller than 0
(negative), the probability is smaller than 0.5.

In summary, if we use a linear model to predict probabilities, we have the

problem of predicted probabilities smaller than 0 and larger than 1 that are
meaningless. If we use a linear model to predict odds we have the problem of
predicted odds smaller than 0 that are meaningless: they are impossible! If on
the other hand we use a linear model to predict the natural logarithm of odds
(logodds), we have no problem whatsoever. We therefore propose to use a linear
model to predict logodds: the natural logarithm of the odds that correspond to
a particular probability.

Returning back to our example of the children passing the exam, suppose we
have the following linear equation for the relationship between age and the
logarithm of the odds of passing the exam

logodds = −33.15 + 0.42 × age,

This equation predicts that a child aged 70 months has a logodds of −33.15+
0.42 × 70 = −3.75. In order to transform that logodds back to a probability, we
first have to take the exponential of the logodds2 to get the odds:

odds = exp(logodds) = elogodds = e−3.75 = 0.02

An odds of 0.02 means that the odds of passing the exam is 0.02 to 1 (0.02:1).
So out of 1 + 0.02 = 1.02 times, we expect 0.02 successes and 1 failure. The
0.02
probability of success is therefore 1+0.02 = 0.02. Thus, based on this equation,
the expected probability of passing the exam for a child of 70 months equals
0.02.
If you find that easier, you can also memorise the following formula for the
relationship between a logodds of x and the corresponding probability:

exp(x)
px = (13.13)
1 + exp(x)

Thus, if you have a logodds x of −3.75, the odds equals exp(−3.75) = 0.02,
0.02
and the corresponding probability is 1+0.02 = 0.02.

2 If we know ln(x) = 60, we have to infer that x equals e60 , because ln(e60 ) = 60 by

definition of the natural logarithm, see previous footnote. Therefore, if we know that ln(x) = c,
we know that x equals ec . The exponent of c, ec , is often written as exp(c). So if we know
that the logarithm of the odds equals c, logodds = ln(oddsratio) = c, then the odds is equal
to exp(c).

336
13.2.3 Logistic link function
In previous pages we have seen that logodds have the nice property of having
meaningful values between −∞ and +∞. This makes them suitable for linear
models. In essence, our linear model for our exam data in children might then
look like this:

logoddspass = b0 + b1 age (13.14)

Y ∼ Bern(ppass ) (13.15)

Note that we can write the odds as p/(1 − p), p is a probability (or a
proportion). So the logodds that corresponds to the probability of passing the
ppass
exam, ppass , can be written as ln 1−p pass
, so that we have

ppass
ln = b0 + b1 age (13.16)
1 − ppass
Y ∼ Bern(ppass ) (13.17)

Note that we do not have a residual any more: the randomness around
the predicted values is no longer modelled using a residual e that is normally
distributed, but is now modelled by a Y -variable with a Bernoulli distribution.
Also note the strange relationship between the probability parameter ppass for
the Bernoulli distribution, and the dependent variable for the linear equation
b0 +b1 age. The linear model predicts the logodds, but for the Bernoulli distribution,
we use the probability. But it turns out that this model is very flexible and useful
in many real-life problems. This model is often called a logit model: one often
writes that the logit of the probability is predicted by a linear model.

logit(ppass ) = b0 + b1 age (13.18)

Y ∼ Bern(ppass ) (13.19)

In essence, the logit function transforms a p-value into a logodds:

p
logit(p) = ln( )
1−p
So what does it look like, a linear model for logodds (or logits of probabilities)?
In Figure 13.9 we show a hypothetical example of a linear model for the logit
of probabilities of passing an exam. These logits or logodds are predicted by
age using a straight, linear regression line.
When we take all these predicted logodds and convert them back to probabilities,
we obtain the plot in Figure 13.10. Note the change in the scale of the vertical
axis, the rest of the plot is the same as in Figure 13.9.
Here again we see the S-shape relationship between probabilities and the
logodds. We see that our model predicts probabilities close to 0 for very young

337
logit(p) = logarithm of the odds 4

intercept: −33.15, slope: 0.42

−2

70 75 80 85
age in months

Figure 13.9: Example of a linear model for the logit of probabilities of passing
an exam.

1.00
probability of passing the exam

0.75

0.50

0.25

0.00
70 75 80 85
age in months

Figure 13.10: Example with logodds transformed into probabilties (vertical

axis).

338
 1.00
probability of passing the exam

0.75

0.50

0.25

0.00
70 75 80 85
age in months

Figure 13.11: Transformed regression line and raw data points.

ages, and probabilities close to 1 for very old ages. There is a clear positive effect
of age on the probability of passing the exam. But note that the relationship is
not linear on the scale of the probabilities: it is linear on the scale of the logit of
the probabilities, see Figure 13.9, but non-linear on the scale of the probabilities
themselves, see Figure 13.10.
The curvilinear shape we see in Figure 13.10 is called a logistic curve. It is
based on the logistic function: here p is a logistic function of age (and note the
similarity with Equation 13.13):

exp(b0 + b1 age)
p = logistic(b0 + b1 age) =
1 + exp(b0 + b1 age)
In summary, if we go from logodds to probabilities, we use the logistic
exp(x)
function, logistic(x) = 1+exp(x) . If we go from probabilities to logodds, we
p
use the logit function, logit(p) = ln 1−p . The logistic regression model is a
generalised linear model with a logit link function, because the linear equation
b0 +b1 X predicts the logit of a probability. It is also often said that we’re dealing
with a logistic link function, because the linear equation gives a value that we
have to subject to the logistic function to get the probability. Both terms, logit
link function and logistic link function are used.
If we go back to our data on the third-grade children that either passed or
failed the exam, we see that this curve gives a description of our data, see Figure
13.11. The model predicts that around the age of 78 months, the probability of
passing the exam is around 0.50. We indeed see in Figure 13.11 that around this
age some children pass the exam (score = 1) and some don’t (score = 0). On
the basis of this analysis there seems to be a positive relationship between age
in third-grade children and the probability of passing the exam in this sample.
What we have done here is a logistic regression of passing the exam on age. It

339
is called logistic because the curve in Figure 13.11 has a logistic shape. Logistic
regression is one specific form of a generalised linear model. Here we have
applied a generalised linear model with a so-called logit link function: instead
of modelling dependent variable Y directly, we have modelled the logit of the
probabilities of obtaining a Y -value of 1. There are many other link functions
possible. One of them we will see in the chapter on generalised linear models
for count data (Chapt. 14). But first, let’s see how logistic regression can be
performed in R, and how we should interpret the output.

Table 13.1: Taking the train to Paris data.

train age sex male income business
1 35.12 1 7544.00 1
1 66.66 1 7096.00 0
0 42.77 1 29261.00 1
0 72.63 0 24977.00 0
1 76.25 0 876.00 1
0 19.87 1 126943.00 1

13.3 Logistic regression in R

Imagine a data set on travellers from Amsterdam to Paris. From 1000 travellers,
randomly sampled in 2017, we know whether they took the train to Paris, or
whether they used other means of transportation. Of these travellers, we know
their age, sex, yearly income, and whether they are travelling for business or
not. Part of the data are displayed in Table 13.1. A score of 1 on the variable
train means they took the train, a score of 0 means they did not.
Suppose we want to know what kind of people are more likely to take the
train to Paris. We can use a logistic regression analysis to predict whether
people take the train or not, on the basis of their age, sex, income, and main
purpose of the trip.
Let’s see whether income predicts the probability of taking the train. The
function that we use in R is the glm() function, which stands for Generalised
Linear Model. We can use the following code:
model . train < - data . train % >%
glm ( train ∼ income ,
data = . ,
family = binomial ( link = logit ))
train is our dependent variable, income is our independent variable, and
these variables are stored in the data frame called data.train. But further
we have to specify that we want to use the Bernoulli distribution and a logit
link function. So link = logit. But why a binomial distribution? Well, the
Bernoulli distribution (one coin flip) is a special case of the Binomial distribution
(the distribution of several coin flips). So here we use a binomial distribution

340
for one coin flip, which is equivalent to a Bernoulli distribution. Actually, the
code can be a little bit shorter, because the logit link function is the default
option with the binomial distribution:
model . train < - data . train % >%
glm ( train ∼ income ,
data = . ,
family = binomial )
Below, we see the parameter estimates from this generalised linear model
run on the train data.

model.train %>%
tidy()

## # A tibble: 2 x 5
## term estimate std.error statistic p.value
## <chr> <dbl> <dbl> <dbl> <dbl>
## 1 (Intercept) 90.0 32.5 2.77 0.00564
## 2 income -0.00817 0.00297 -2.75 0.00603

The parameter estimates table from a glm() analysis looks very much like
that of the ordinary linear model and the linear mixed model. An important
difference is that the statistics shown are no longer t-statistics, but Z-statistics.
This is because with logistic models, the ratio b1 /SE does not have a t-distribution.
In ordinary linear models, the ratio b1 /SE has a t-distribution because in linear
models, the variance of the residuals, σe2 , has to be estimated (as it is unknown).
If the residual variance were known, b1 /SE would have a standard normal
distribution. In logistic models, there is no σe2 that needs to be estimated (it is
by default 1), so the ratio b1 /SE has a standard normal distribution. One could
therefore calculate a Z-statistic Z = b1 /SE and see whether that value is smaller
than 1.96 or larger than 1.96, if you want to test with a Type I error rate of 0.05.

The interpretation of the slope parameters is very similar to other linear models.
Note that we have the following equation for the logistic model:

logit(ptrain ) = b0 + b1 income
train ∼ Bern(ptrain ) (13.20)

If we fill in the values from the R output, we get

logit(ptrain ) = 90.0 − 0.008 × income

train ∼ Bern(ptrain ) (13.21)

We can interpret these results by making some predictions. Imagine a

traveller with a yearly income of 11,000 Euros. Then the predicted logodds

341
equals 90.0 − 0.00817 × 11000 = 0.13. When we transform this back to a
exp(0.13)
probability, we get 1+exp(0.13) = 0.53. So this model predicts that for people
with a yearly income of 11,000, about 53% of them take the train (if they travel
at all, that is!).
Now imagine a traveller with a yearly income of 100,000. Then the predicted
logodds equals 90.0 − 0.00817 × 100000 = −727. When we transform this back
exp(−727)
to a probability, we get 1+exp(−727) = 0.00. So this model predicts that for
people with a yearly income of 100,000, close to none of them take the train.
Going from 11,000 to 100,000 is a big difference. But the change in probabilities
is also huge: the probability goes down from 0.53 to 0.00.
We found a difference in probability of taking the train for people with
different incomes in this sample of 1000 travellers, but is there also an effect
of income in the entire population of travellers between Amsterdam and Paris?
The regression table shows us that the effect of income, −0.00817, is statistically
significant at an α of 5%, Z = −2.75, p < 0.01. We can therefore reject the null-
hypothesis that income is not related to whether people take the train or not.
We conclude that in the population of travellers to Paris, a higher income is
associated with a lower probability of travelling by train.
Note that similar to other linear models, the intercept can be interpreted as
the predicted logodds for people that have values 0 for all other variables in the
model. Therefore, 90.0 means in this case that the predicted logodds for people
with zero income equals 90.0. This is equivalent to a probability of very close
to 1.

342
Chapter 14

Generalized linear models

for count data: Poisson
regression

14.1 Poisson regression

Count data are inherently discrete, and often when using linear models, we see
non-normal distributions of residuals. In Chapter 13 we discussed a data set
on the scores that a group of students got for an assignment. There were four
criteria, and the score consisted of the number of criteria that were met for each
student’s assignment. Figure 13.3 showed that after an ordinary linear model
analysis, the residuals did not look normal at all.
Table 14.1 shows part of the data that were analysed. Similar to logistic
regression, perhaps we can find a distribution other than the normal distribution
that is more suitable for this kind of data? For dichotomous data (1/0) we
found the Bernoulli distribution very useful. For count data, the traditional
distribution is the Poisson distribution.

Table 14.1: Scores on an assignment.

ID score previous
1 0 0.41
2 2 -0.47
3 4 0.07
4 0 -0.50
5 2 -0.83
6 3 0.17

The normal distribution has two parameters, the mean and the variance. The
Bernoulli distribution has only 1 parameter (the probability), and the Poisson

343
 0.20

0.15
probability

0.10

0.05

0.00

0 1 2 3 4 5 6 7 8 9 10
count

Figure 14.1: Count data example where the normal distribution is not a good
approximation of the distribution of the residuals.

distribution has also only 1 parameter, lambda or λ. λ is a parameter that

indicates tendency. Figure 14.1 shows a Poisson distribution with a tendency
of 4.
What we see is that many values centre around the tendency parameter
value of 4 (therefore we call it a tendency parameter)! We see only discrete
values, and no values below 0. We see a few values higher than 10. If we take
the mean of the distribution, we will find a value of 4. If we would compute
the variance of the distribution we would also find 4! In general, if we have a
Poisson distribution with a tendency parameter λ = 4, we know that both the
mean and the variance will be equal to λ.
A Poisson model could be suitable for our data: a linear equation could
predict the parameter λ and then the actual data show a Poisson distribution.

λ = b0 + b1 X (14.1)
y ∼ P oisson(λ) (14.2)

However, because of the additivity assumption, the equation b0 + b1 X leads

to negative values. A negative value for λ is not logical, because we then have a
tendency to observe data like -2 and -4 in our data, which is contrary to having
count data, which consists of non-negative integers. A Poisson distribution
always shows integers of at least 0, so one or way or another we have to make
sure that we always have a λ of at least 0.
Remember that we saw the reverse problem with logistic regression: there
we wanted to have negative values for our dependent variable logodds ratio, so

344
therefore we used the logarithm. Here we want to have positive values for our
dependent variable, so we can use the inverse of the logarithm function: the
exponential. Then we have the following model:

λ = exp(b0 + b1 X) = eb0 +b1 X (14.3)

y ∼ P oisson(λ) (14.4)

This is a generalized linear model, now with a Poisson distribution and an

exponential link function. The exponential function makes any value positive,
for instance exp(0) = 1 and exp(−100) = 0.
Let’s analyse the assignment data with this generalized linear model. Our
dependent variable is the number of criteria met for the assignment (a number
between 0 and 4), and the independent variable is previous, which is a standardized
mean of a number of previous assignments. We expect that the mean score on
previous assignments is associated with a higher score on the present assignment.
When we run the analysis, the result is as follows:

λ = exp(0.1576782 − 0.0548685 × previous) (14.5)

score ∼ P oisson(λ) (14.6)

What does it mean? Well, similar to logistic regression, we can understand

such equations by making some predictions for interesting values of the independent
variable. For instance, a value of 0 for previous means an average grade on
previous advanced that is around the mean value. So if we choose previous=0,
then we have the prediction for an average student. If we fill in that value, we
get the equation λ = exp(0.1576782 − 0.0548685 × 0) = exp(0.1576782) = 1.17.
Thus, for an average student, we expect to see a score of 1.17. A Poisson
distribution with λ = 1.17 is depicted in Figure 14.2.
Another interesting value of previous might be -2. That represents a
student with generally very low grades. Because the average grades were standardized,
only about 2.5% of the students has a lower average grade than -2. If we fill in
that value, we get: λ = exp(0.1576782 − 0.0548685 × −2) = 0.85. A Poisson
distribution with λ = 0.85 is depicted in Figure 14.3.
The last value of previous for which we calculate λ is +2, representing a
high-performing student. We then get λ = exp(0.1576782−0.0548685×2) = 1.6.
A Poisson distribution with λ = 1.6 is depicted in Figure 14.4.
If we superimpose these three figures, we obtain Figure 14.5, where we see
that the higher the average score on previous assignments, the higher is the
expected score on the present assignment.
We found that in this data set, previous high marks for assignments predicted
a higher mark for the present assignment. In the next section we will see how
to perform the analysis in SPSS, and how to check whether there is also a
relationship in the population of students.

345
 0.3
probability

0.2

0.1

0.0

0 1 2 3 4 5 6 7 8 9 10
count

Figure 14.2: Poisson distribution with lambda=1.17.

0.4

0.3
probability

0.2

0.1

0.0

0 1 2 3 4 5 6 7 8 9 10
count

Figure 14.3: Poisson distribution with lambda=0.85.

346
 0.3

0.2
probability

0.1

0.0

0 1 2 3 4 5 6 7 8 9 10
count

Figure 14.4: Poisson distribution with lambda=1.60.

0.4

0.3

previous
probability

low
0.2 average
high

0.1

0.0

0 1 2 3 4 5 6 7 8 9 10
count

Figure 14.5: Three different Poisson distributions with lambdas 0.85, 1.17, and
1.60, for three different kinds of students.

347
14.2 Poisson regression in SPSS
Poisson regression is a form of a generalized linear model analysis, similar to
logistic regression. However, instead of using a Bernoulli distribution we use a
Poisson distribution. For a numeric predictor like the variable previous, the
syntax is as follows.
GENLIN scores WITH previous
/MODEL previous
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION.
The output with parameter values is shown in Figure 14.6.

Parameter Estimates

95% Wald Confidence Interval Hypothesis Test

Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) .158 .0925 -.024 .339 2.908 1 .088
previous -.055 .0899 -.231 .121 .372 1 .542
(Scale) 1a
Dependent Variable: score
Model: (Intercept), previous
a. Fixed at the displayed value.

Figure 14.6: SPSS output of a generalized linear model for predicting

assignments scores from the average of previous assignments.

We see the same values for the intercept and the effect of previous as in the
previous section. We now also see 95% confidence intervals for these parameter
values. For both, the value 0 is included in the confidence intervals, therefore
we know that we cannot reject the null-hypotheses that these values are 0 in the
population of students. This is also reflected by the Wald statistics. Remember
that the Wald chi-square (X 2 ) statistic is computed by B 2 /SE 2 . For large
enough samples, these X 2 statistics follow a χ2 distribution with 1 degree of
freedom. From that distribution we know that a value of 0.372 is not significant
at the 5% level. It has an associated p-value of 0.542.
We can write:
Scores for the assignment (1-4) for 100 students were analysed
using a generalized linear model with a Poisson distribution (Poisson
regression). The scores were not significantly predicted by the average
score of previous assignments, B = −0.06, X 2 (1) = 0.37, p = 0.54.
Therefore we cannot reject the null-hypothesis that there is no relationship
between the average of previous assignments and the score on the
present assignment in the population of students.
Suppose we also have a categorical predictor, for example the degree that
the students are working for. Some do the assignment for a bachelor’s degree

348
(degree=1), some for a master’s degree (degree=2), and some for a PhD degree
(degree=3). The syntax would then look like:

GENLIN scores BY degree

/MODEL degree
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION.

Note that only the independent variable has changed and WITH is changed
into BY. The output is given in Figure 14.7.

Parameter Estimates

95% Wald Confidence Interval Hypothesis Test

Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) .354 .1459 .068 .640 5.878 1 .015
[degree=1.00] -.584 .2415 -1.057 -.111 5.852 1 .016
[degree=2.00] -.089 .2110 -.503 .325 .178 1 .673
a
[degree=3.00] 0 . . . . . .
b
(Scale) 1
Dependent Variable: score
Model: (Intercept), degree
a. Set to zero because this parameter is redundant.
b. Fixed at the displayed value.

Figure 14.7: SPSS output of a generalized linear model for predicting

assignments scores from the degree that is studied for.

We see that the parameter for the degree=3 category is fixed to 0, meaning
that it is used as the reference category. If we make a prediction for this group
of students that is studying for a PhD degree, we have λ = exp(.354 + 0) =
exp(0.354) = 1.4. For the students studying for a Master’s degree we have
λ = exp(.354 − 0.089) = 1.3 and for students studying for their Bachelor’s
degree we have λ = exp(.354 − 0.584) = 0.8. These λ-values correspond to
the expected number in a Poisson distribution, so for Bachelor students we
expect a score of 0.8, for Master students we expect a score of 1.3 and for Phd
students a score of 1.4. Are these different scores also present in the population?
We see that the effect for degree=1 is significant, X 2 (1) = 5.85, p = 0.02, so
there is a difference in score between students studying for a Bachelor’s degree
and students studying for a PhD. The effect for degree=2 is not significant,
X 2 (1) = 0.18, p = 0.67, so there is no difference in assignment scores between
Master students and PhD students.

Remember that for the linear model, when we wanted to compare more than
two groups at the same time, we used an F -test to test for an overall difference
in group means. Also for the generalized linear model, we might be interested

349
in whether there is an overall difference in scores between Bachelor, Master and
PhD students. For that we need to tweak the syntax a little bit, by stating that
we also want to see an overall test printed. The PRINT statements then also
needs the word SUMMARY. In other words, the syntax becomes

GENLIN scores BY degree

/MODEL degree
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION SUMMARY.

We then get the relevant output in Figure 14.8. There we see a Wald Chi-
Square statistic for the effect of degree. It has 2 degrees of freedom, since the
effect for the 3 categories is coded by 2 dummy variables. So this test tells us
that the null-hypothesis that the expected scores in each group of students are
the same can be rejected, X 2 (2) = 6.27, p = 0.04.

Tests of Model Effects

Type III
Wald Chi-
Source Square df Sig.
(Intercept) 1.844 1 .175
degree 6.271 2 .043
Dependent Variable: score
Model: (Intercept), degree

Figure 14.8: SPSS output of a generalized linear model for predicting

assignments scores from the degree that is studied for.

14.3 Interaction effects in Poisson models

In the previous subsection we looked at a count variable, the number of criteria
fulfilled, and we wanted to predict it from the degree that students were studying
for. Let’s look at an example where we want to predict a count variable using
two categorical predictors.
In 1912, the ship Titanic sank after the collision with an iceberg. There were
2201 people on board that ship. Some of these were male, others were female.
Some were passengers, others were crew, and some survived, and some did not.
For the passengers there were three groups: those travelling first class, second
class and third class. There were also children on board. If we focus on only
the adults, suppose we want to know whether there is a relationship between
the sex and the counts of people that survived the disaster. Table 14.2 gives the
counts of survivors for males and females separately.

350
 Table 14.2: Counts of adult survivors on the Titanic.
count
Male 338
Female 316

Let’s analyse this data set with SPSS. In SPSS we assign the value sex=1
to Females and sex=2 to Males. Our dependent variable is count, and the
independent variable is sex.

GENLIN count BY sex

/MODEL sex
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION.

Parameter Estimates
95% Wald Confidence Interval Hypothesis Test
Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) 5.823 .0544 5.716 5.930 11460.858 1 .000
[sex=1.00] -.067 .0783 -.221 .086 .740 1 .390
[sex=2.00] 0a . . . . . .
(Scale) 1b
Dependent Variable: count
Model: (Intercept), sex
a. Set to zero because this parameter is redundant.
b. Fixed at the displayed value.

Figure 14.9: SPSS output of a generalized linear model for predicting numbers
of men and women on board the Titanic.

From the output in Figure 14.9 we see that the expected count for females
is exp(5.823 − 0.067) = 318.3 and the expected count for males is exp(5.823) =
340.4. These expected counts are close to the observed counts of males and
females. The only reason that they differ from the observed is because of
rounding errors (SPSS shows only the first three decimals). From the Wald
statistic, we see that the difference in counts between males and females is not
significant, X 2 (1) = 0.74, p = 0.391 .
The difference in these counts is very small. But does this tell us that women
were as likely to survive as men? Note that we have only looked at those who
1 Note that a hypothesis test is a bit odd here: there is no clear population that we want

to generalize the results to: there was only one Titanic disaster. Also, here we have data on
the entire population of those people on board the Titanic, there is no random sample here.

351
survived. How about the people that perished: were there more men that died
than women? Table 14.3 shows the counts of male survivors, female survivors,
male non-survivors and female non-survivors. Then we see a different story:
on the whole there were many more men than women, and a relatively small
proportion of the men survived. Of the men, most of them perished: 1329
perished and only 338 survived, a survival rate of 20.3%. Of the women, most
of them survived: 109 perished and 316 survived, yielding a survival rate of
74%. Does this tell us that women are much more likely than men to survive
collisions with icebergs?

Table 14.3: Counts of adults on the Titanic.

sex survived count
Male 0 1329
Female 0 109
Male 1 338
Female 1 316

Let’s first run a multivariate Poisson regression analysis including the effects
of both sex and survival. The syntax is

GENLIN count BY sex WITH survived

/MODEL sex survived
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION.

where we treat sex as a categorical variable and survival as a numeric variable

(since survived is already coded as a dummy and sex is not).

Parameter Estimates
95% Wald Confidence Interval Hypothesis Test
Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) 7.044 .0286 6.988 7.100 60709.658 1 .000
[sex=1.00] -1.367 .0543 -1.473 -1.260 632.563 1 .000
[sex=2.00] 0a . . . . . .
survived -.788 .0472 -.880 -.695 279.073 1 .000
(Scale) 1b
Dependent Variable: count
Model: (Intercept), sex, survived
a. Set to zero because this parameter is redundant.
b. Fixed at the displayed value.

Figure 14.10: SPSS output of a generalized linear model for predicting numbers
of men and women that perished and survived on board the Titanic.

352
 observed predicted

1000

survived
count

0
1
500

Female Male Female Male

sex

Figure 14.11: Difference between expected and observerd numbers of passengers.

The output is given in Figure 14.10. From the parameter values, we can
calculate the predicted numbers of male (sex = 2) and female (sex = 1) that
survived and perished. For female survivors we have exp(7.04 − 1.37 − .79) =
131.63, for female non-survivors we have exp(7.04 − 1.37) = 290.03, for male
survivors we have exp(7.04 − .79) = 518.01 and for male non-survivors we have
exp(7.04) = 1141.39.
These predicted numbers are displayed in Figure 14.11. It also shows the
observed counts. The pattern that is observed is clearly different from the one
that is predicted from the generalized linear model. The linear model predicts
that there are fewer survivors than non-survivors, irrespective of sex, but we
observed that in females, there are more survivors than non-survivors. It seems
that sex is a moderator of the effect of survival on counts.
In order to test this moderation effect, we run a new generalized linear
model for counts including an interaction effect of sex by survived. This is done
in SPSS syntax by changing the MODEL part by including a sex by survived
interaction:
GENLIN count BY sex WITH survived
/MODEL sex survived sex*survived
DISTRIBUTION=POISSON LINK=LOG
/PRINT CPS DESCRIPTIVES SOLUTION.
The output is displayed in Figure 14.12. When we plot the predicted counts
from this new model with an interaction effect, we see that they are exactly
equal to the counts that are actually observed in the data, see Figure 14.13.
From the output we see that the interaction effect is significant, X 2 (1) =
91.82, p < 0.001. If we regard this data set as a random sample of all ships that

353
 Parameter Estimates
95% Wald Confidence Interval Hypothesis Test
Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) 7.192 .0274 7.138 7.246 68745.825 1 .000
[sex=1.00] -2.501 .0996 -2.696 -2.306 630.032 1 .000
a
[sex=2.00] 0 . . . . . .
survived -1.369 .0609 -1.489 -1.250 505.126 1 .000
[sex=1.00] *
survived 2.434 .1267 2.185 2.682 368.979 1 .000
[sex=2.00] *
survived 0a . . . . . .
b
(Scale) 1
Dependent Variable: count
Model: (Intercept), sex, survived, sex * survived
a. Set to zero because this parameter is redundant.
b. Fixed at the displayed value.

Figure 14.12: SPSS output of a generalized linear model for predicting numbers
of men and women that perished and survived on board the Titanic.

sink after collision with icebergs, we may conclude that in such situations, sex
is a significant moderator of the difference in the numbers of survivors and non-
survivors. One could also say: the proportion of people that survive a disaster
like this is different in females than it is in males.

14.4 Cross-tabulation and the Pearson chi-square

statistic
The data on male and female survivors and non-survivors are often tabulated
in a cross-table like in Table 14.4

Table 14.4: Counts of adult survivors and non-survivors on the Titanic.

No Yes
Male 1329 338
Female 109 316

In the previous section these counts were analysed using a generalized linear
model with a Poisson distribution and an exponential link function. We wanted
to know whether there was a significant difference in the proportion of survivors
for men and women. In this section we discuss an alternative method of
analysing count data. We discuss an alternative chi-square (X 2 ) statistic for
the moderation effect of one variable of the effect of another variable.

354

Page
 observed predicted

1000

survived
count

0
1
500

Female Male Female Male

sex

Figure 14.13: Difference between observed and predicted numbers of passengers.

First let’s have a look at the overall survival rate. In total there were 654
people that survived and 1438 people that did not survive. Table 14.5 shows
these column totals.

Table 14.5: Counts of adult survivors and non-survivors on the Titanic.

No Yes
Male 1329 338
Female 109 316
Total 1438 654

Looking at these total numbers of survivors and non-survivors, we can

calculate the proportion of survivors overall (the survival rate) as 654/(654 +
1438) = 0.31. Table 14.6 shows the totals for men and women, as well as the
overall total number of adults.

Table 14.6: Counts of adult survivors and non-survivors on the Titanic.

No Yes Total
Male 1329 338 1667
Female 109 316 425
Total 1438 654 2092

Suppose we only know that of the 2092 people, 1667 were men, and of all
people, 654 survived. Then suppose we pick a random person from these 2092
people. What is the probability that we get a male person that survived, given
that sex and survival have nothing to do with each other ?

355
 Well, from probability theory we know that if two events A and B are
independent, the probability of observing A and B at the same time, is equal
to the product of the probability of event A and the probability of event B.

P rob(AandB) = P rob(A) × P rob(B) (14.7)

If sex and survival are independent from each other, then the probability of
observing a male survivor is equal to the probability of seeing a male times the
probability of seeing a survivor. The probability for survival is 0.31, as we saw
earlier, and the probability of seeing a male is equal to the proportion of males
in the data, which is 1667/2092 = 0.8. Therefore, the probability of seeing a
male survivor is 0.8 × 0.31 = 0.24. The expected number of male survivors is
then that probability times the total number of people, 0.24 × 2092 = 502.08.
Similarly we can calculate the expected number of non-surviving males, the
number of surviving females, and the number of non-surviving females.
These numbers, after rounding, are displayed in Table 14.7.

Table 14.7: Expected numbers of adult survivors and non-survivors on the

Titanic.
No Yes
Male 1155 519
Female 289 130

The expected numbers in Table 14.7 are quite different from the observed
numbers in Table 14.4. Are the differences large enough to think that the two
events of being male and being a survivor are NOT independent? If the expected
numbers on the assumption of independence are different enough from the
observed numbers, then we can reject the null-hypothesis that being male and
being a survivor have nothing to do with each other. To measure the difference
between expected and observed counts, we need a test statistic. Here we use
Pearson’s chi-square statistic. It involves calculating the difference between the
numbers in the respective cells, and standardize them by the expected number.
Here’s how it goes:
For each cell, we take the predicted count and subtract it from the observed
count. For instance, for the male survivors, we expected 519 but observed 338.
The difference is therefore 338 − 519 = −181. Then we take the square of this
difference, 1812 = 32761. Then we divide this number by the expected number,
and then we get 32761/519 = 63.1233141. We do exactly the same thing for the
male non-survivors, the female survivors and the female non-survivors. Then
we add these 4 numbers, and then we have the Pearson chi-square statistic. In
formula form:

(Oi − Ei )2
X 2 = Σi (14.8)
Ei

356
 So for male survivors we get

(338 − 519)2
= 63.1233141 (14.9)
519
For male non-survivors we get

(1329 − 1155)2
= 26.212987 (14.10)
1155
For female survivors we get

(316 − 130)2
= 266.1230769 (14.11)
130
and for female non-survivors we get

(109 − 289)2
= 112.1107266 (14.12)
289
If we add these 4 numbers we have the chi-square statistic: X 2 = 467.57.
Note that we only use the rounded expected numbers. Better would be to use
the non-rounded numbers. Had we used the non-rounded expected numbers,
we would have gotten X 2 = 460.87.
The Wald chi-square statistic for the sex*survived interaction effect was
368.9788928, see Figure 14.12. It tests exactly the same null-hypothesis as the
Pearson chi-square: that of independence, or in other words, that the numbers
can be explained by only two main effects, sex and survival.
If the data set is large enough and the numbers are not too close to 0,
the same conclusions will be drawn, whether from a Wald chi-square for an
interaction effect in a generalized linear model, or from a cross-tabulation and
computing a Pearson chi-square. The advantage of the generalized linear model
approach is that you can do much more with them, for instance more than
two predictors, and that you make it more explicit that when computing the
statistic, you take into account the main effects of the variables. You do that also
for the Pearson chi-square but it is less obvious: we did that by first calculating
the probability of survival and second calculating the proportion of males.

14.5 Poisson regression or logistic regression?

In the previous section we analysed the relationship between the variable sex
of the person on board the Titanic, and the variable survived: whether or
not a person survived the shipwreck. We found a relationship between these
two variables by studying the cross-tabulation of the counts, and testing that
relationship using a Pearson chi-square statistic. In the section before that, we
saw that this relationship could also be tested by applying a Poisson regression
model and looking at the sex by survived interaction effect. These methods are
equivalent.

357
 There is yet a third way to analyse the sex and survived variables. Remember
that in the previous chapter we discussed logistic regression. In logistic regression,
a dichotomous variable (a variable with only two values, say 0 and 1) is the
dependent variable, with one or more quantitative or qualitative independent
variables. Both sex and survived are dichotomous variables: male and female,
and survived yes or survived no. In principle therefore, we could do a logistic
regression: for example predicting whether a person is a male or female, on
the basis of whether they survived or not, or the other way around, predicting
whether people survive or not, on the basis of whether a person is a woman or
a man.
What variable is used here as your dependent variable, depends on your
research question. If your question is whether females are more likely to survive
than men, perhaps because of their body fat composition, or perhaps because of
male chivalry, then the most logical choice is to take survival as the dependent
variable and sex as the independent variable.
The syntax for logistic regression then looks like

GENLIN survived (REFERENCE=FIRST) BY sex

/MODEL business
DISTRIBUTION=BINOMIAL LINK=LOGIT
/PRINT CPS DESCRIPTIVES SOLUTION.

Note however that the data is in the wrong format. For the Poisson regression,
the data were there in the form of what we see in Table 14.4. However, for a
logistic regression, we need the data in the format like in Table 14.8. For every
person on-board the ship, we have to know their sex and their survival status.

Table 14.8: Individual data of adult survivors and non-survivors on the Titanic.
ID sex survived
1004 Male 0
623 Male 0
934 Male 0
400 Male 0
1626 Male 1
1103 Male 0
270 Male 0
2052 Female 1
1685 Male 1
1236 Male 0

We use BY to treat the sex variable as categorical. We use (REFERENCE =

FIRST) because we want to predict whether people survive (survive=1). Then
our reference category is survive=0, which is the first value. In the output in
Figure 14.14 we see that sex is a significant predictor of the survival status,
B = 2.434, X 2 = 368, 98, p < 0.001. The logodds ratio for a male surviving the
shipwreck is −1.37, and the logodds ratio for a female surviving the shipwreck

358
is −1.37 + 2.43 = 1.06. These logodds ratios correspond to probabilities of 0.20
and 0.74, respectively. Thus, some are much more likely to survive than men.

Parameter Estimates
95% Wald Confidence Interval Hypothesis Test
Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) -1.369 .0609 -1.489 -1.250 505.126 1 .000
[sex=1.00] 2.434 .1267 2.185 2.682 368.979 1 .000
a
[sex=2.00] 0 . . . . . .
(Scale) 1b
Dependent Variable: survived
Model: (Intercept), sex
a. Set to zero because this parameter is redundant.
b. Fixed at the displayed value.

Figure 14.14: SPSS output of a generalized linear model for predicting numbers
of men and women that perished and survived on board the Titanic.

However, suppose you are the relative of a passenger on board a ship that
shipwrecks. After two days, there is news that a person was found. The only
thing known about the person is that he or she is alive. Your relative is your
niece, so you’d like to know on the basis that the person that was found lives,
what is the probability that that person is a woman, because then it could be
your beloved niece! You could therefore run a logistic regression on the Titanic
data to see to what extent the survival of a person predicts the sex of that
person. The syntax would then look like this:

GENLIN sex (REFERENCE=LAST) WITH survived

/MODEL survived
DISTRIBUTION=BINOMIAL LINK=LOGIT
/PRINT CPS DESCRIPTIVES SOLUTION.

Note that we use WITH in order to treat the dummy variable survived as
quantitative. We also use (REFERENCE=LAST) to indicate that we use the
last (second) category of sex (2) as the reference category, because that category
refers to men, because we want to predict whether a person is a female.
The output is give in Figure 14.15
From this output we conclude that survival is a significant predictor of sex,
B = −2.434, X 2 = 368, 98, p < 0.001. The logodds ratio for a surviving person
to be a woman is −4.93 + 2.43 = −2.50, and the logodds ratio for a non-
surviving person to be a woman is −4.93. These logodds ratios correspond to
probabilities of 0.08 and 0.01, respectively. Thus, if you know that there is a
person that survived the Titanic, it is not very likely that it was a woman, only
8% chance. If you think this is counter-intuitive, remember that even though a

359
 Parameter Estimates
95% Wald Confidence Interval Hypothesis Test
Wald Chi-
Parameter B Std. Error Lower Upper Square df Sig.
(Intercept) -4.934 .2141 -5.354 -4.515 531.265 1 .000
survived 2.434 .1267 2.185 2.682 368.979 1 .000
a
(Scale) 1
Dependent Variable: sex
Model: (Intercept), survived
a. Fixed at the displayed value.

Figure 14.15: SPSS output of a generalized linear model for predicting numbers
of men and women that perished and survived on-board the Titanic.

large proportion of the women survived the Titanic, there were many more men
on-board than women.

In summary if you have count data, and one of the variables is dichotomous, you
have the choice whether to use a Poisson regression model or a logistic regression.
The choice depends on the research question: if your question involves prediction
of a dichotomous variable, logistic regression is the logical choice. If you have
a theory that one or more independent variable explain one other variable,
logistic regression is the logical choice. If however your theory does not involve
a natural direction or prediction of one variable, and you are simply interested
in associations among variables, then Poisson regression is the obvious choice.

360

Page
Chapter 15

Introduction to big data

analytics

15.1 Introduction
Previous chapters looked into traditional or classic data analysis: inference
about a population using a limited sample of data with a limited number of
variables. For instance, we might be interested in how large the effect is of a
new kind of therapy for clinical depression in the population of all patients,
based on a sample of 150 treated patients and 150 non-treated patients on a
waiting list.
In many contexts, we are not interested in the effect of some intervention
in a population, but in the prediction of events in the future. For example, we
would like to predict which patients are likely to relapse into depression after
an initially successful therapy. For such a prediction we might have a lot of
variables. In fact, more variables than we could use in a straightforward linear
model analysis.
In this age of big data, there are more often too many data then too few
data about people. However, this wealth of data is often not nicely stored in
data matrices. Data on patients for example are stored in different types of files,
in different file formats, as text files, scans, X-rays, lab reports, perhaps even
videotaped interviews. They may be stored at different hospitals or medical
centres, so they need to be linked and combined without mixing them up. In
short: data can be really messy. Moreover, data are not variables yet. Data
science is about making data available for analysis. This field of research aims
to extract knowledge and insight from structured and unstructured data. To do
that, it draws from statistics, mathematics, computer science and information
science.
The patient data example is a typical case of a set of unstructured data.
From a large collection of pieces of texts (e.g., notes from psychiatric interviews
and counselling, notes on prescriptions and adverse effects of medication, lab

361
reports) one has to distil a set of variables that could predict whether or not an
individual patient would fall back into a second depressive period.
There are a couple of reasons why big data analytics is different from the data
analysis framework discussed in previous chapters. These relate to 1) different
types of questions, 2) the p > n problem and 3) the problem of over-fitting.
First, the type of questions are different. In classic data analysis, you have a
model with one or more model parameters, for example a regression coefficient,
and the question is what the value is of that parameter in the population.
Based on sample data, you draw inferences regarding the parameter value in
the population. In contrast, typical questions in big data situations are about
predictions for future data (e.g., how will the markets respond to the start of
the hurricane season), or how to classify certain events (e.g., is a Facebook
posting referring to a real event or is it ”fake news”). In big data situations,
such predictions or classifications are based on training data. In classic data
analysis, inference is based on sample data.
Second, the type of data in big data settings allows for a far larger number
of variables than in non-big data settings. In the patient data example, imagine
the endless ways in which we could think of predicting relapse on the basis
of the text data alone. We could take as predictor variables the number of
counselling sessions, whether or not a tricyclic antidepressant was prescribed,
whether or not a non-tricyclic antidepressant was prescribed, whether or not the
word ”mother” was mentioned in the sessions, the number of times the word
”mother” was used in the sessions, how often the word ”mother” was associated
with the word ”angry” or ”anger” in the same sentence, and so on. The types of
variables you could distil from such data is endless, so what to pick? And where
to stop? So the first way in which big data analytics differs from classic data
analysis is that a variable selection method has to be used. The analyst has to
make a choice of what features of the raw data will be used in the analysis. Or,
during the analysis itself, an algorithm can be used that picks those features that
predict the outcome variable most efficiently. Usually there is a combination
of both methods: there is an informed choice of what features in the data are
likely to be most informative (e.g., the data analyst a priori believes that the
specific words used in the interviews will be more informative about relapse
than information contained in X-rays), and an algorithm that selects the most
informative features out of this selection (e.g., the words ”mother” and ”angry”).
One reason that variable selection is necessary is because statistical methods,
like for example linear models, do not work when the number of variable is large
relative to the number of cases. This is known as the p > n problem, where p
refers to the number of variables and n to the number of cases. We will come
back to this problem below.
Third, because there is so much information available in big data situations,
there is the likely danger of over-fitting your model. Maybe you have enough
cases to include 1,000 predictor variables in your linear models, and they will run
and give meaningful output, but then the model will be too much focused on the
data that you have now, so that it will be very bad at predicting or classifying
new events correctly. Therefore, another reason for limiting the number of

362
 120

100
Y

80

60
80 90 100 110 120
X

Figure 15.1: Ilustration of overfitting: a data set showing all 50 data points
showing a linear relationship between variables X and Y .

variables in your model is to prevent over-fitting. Limiting the number of

variables in a model can be done by various variable selection methods, for
example the LASSO in linear regression (Tibshirani, 1996). Over-fitting can
also be countered by using so-called ensemble methods like boosting and random
forest.
In order to check that you are not over-fitting, one generally splits the data
into a training data set and a test data set (or validation data set). The training
data are used to select the variables and fit the model (i.e., determine model
parameters). For example, when a simple linear regression model is used to
predict height in school children, the least squares estimates are determined
based on the training data alone. Next, these estimates are used to predict
new values for the test data. That is, we take the test data, use the predictor
variables and plug them into the model equation and see what height values
we predict for the children in the test data. Next, we compare these predicted
values with the actually observed height measurements in these children. In the
case of over-fitting, the model will show a very good fit (good predictions) to
the training data but a poor fit to the test data (bad predictions). If there is
no over-fitting, the predictions for the test data will not be much worse for the
test data than for the training data.
As a small data example of the phenomenon of over-fitting, in Figure 15.1
we see a complete data set on variables X and Y for 100 observations. The
actual relationship between the two variables can be described by a simple
linear regression model Y = X + e, where e has a normal distribution with
standard deviation 10. But suppose we don’t know that, and we want to find
a model that gives a good description for these data. We split the data set in
half, and use a local polynomial regression fitting algorithm that best describes

363
 120
Y

100

80

80 100 120
X

Figure 15.2: Ilustration of overfitting: only showing half the data points
(training data) and a local polynomial regression model fitted to them.

the training data. The training data set and the model predictions, depicted as
a blue line is given in Figure 15.2. It turns out that the correlation between the
observed Y -values and the predicted values based on the model is pretty high:
0.78.
Next, we apply this model to the test data. These are depicted in Figure
15.3. The blue line are the predicted Y -values for the X-values in this data set,
based on the model. We see that the blue line is not a good description of the
pattern in the test data. This is also reflected in the much lower correlation
between the observed Y -values and the predicted values in the test data: 0.71.
Thus we see that the model is a good model for the training data, upon which
the model was based, but the model is a terrible model for new data, even
both data sets have the same origin. The training data were only randomly
selected. The model was simply too much focused on the details in the data.
Had we used a much simpler model, a linear regression model for example, the
relative performance on the test data would be much better. The least squares
equation predicts the Y -values in the training data with a correlation of 0.48.
That is much worse than the splines, but we see that the model performs much
better in the test data: there we see a correlation of 0.39. In sum: a complex
model will always give better predictions than a simple model for training data.
However, what is important is that a model will also show good predictions in
test data. Then we see that often, a relatively simple model will perform better
than a very complex model. This is due to the problem of over-fitting. The
trade-off between the model complexity and over-fitting is also known as the
bias-variance trade-off, where bias refers to the error that we make when we
select the wrong model for the data, and variance refers to error that we make
because we are limited to seeing only the training data.

364
 120
Y

100

80

80 100 120
X

Figure 15.3: Ilustration of overfitting: only showing half the data points (test
data) and a local polynomial regression model that does not describe these data
well because it was based on the training data.

15.1.1 Model selection

In the preceding chapters, we discussed only the linear model and the extensions
thereof: the linear mixed model and the generalized linear model. In big
data analytics one uses these models too, but in addition there is a wealth
of other models and methods, too. To name but the most well-known: decision
trees, support vector machines, smoothing splines, generalized additive models,
naive Bayes, and neural networks. Each model or method in itself has many
subversions. A very important part of big data analytics is therefore model
selection. Models vary in their flexibility, that is, how well they can fit the
data. Remember that when we looked at multiple regression, a model with
many predictors usually shows a larger R-squared than a model with fewer
predictors. That is, it explains or predicts the sample data better. But now
we know that this also brings the danger of over-fitting: it fits the sample or
training data better, but not necessarily the test data and the data that we
want to predict in future applications. The same is true for simple models and
more flexible models. A linear model with two additive predictors is a relatively
simple model that might not explain so much variance in the training data.
A neural network is an example of a very flexible model that might be much
more able to capture variance in training data that shows a large amount of
complexity, for example a complicated interaction pattern. We therefore have
to make a trade-off between model flexibility and over-fitting that is dependent
on each type of problem. If reality is complex we need a complex model to make
the right predictions, but if reality is simple and we apply a complex model, we
run the risk of over-fitting. If reality is complex and we use a simple model, we
run the risk of bad predictions. In order to strike the right balance and find the

365
optimal level of complexity for our training data, one often uses cross-validation.

15.1.2 Cross-validation
Cross-validation is a form of a re-sampling method. In re-sampling methods,
different subsets of the training data are used to fit the same model or different
models, or different versions of a model. There are different forms of cross-
validation, but here we discuss k-fold cross-validation. In k-fold cross-validation,
the training data are split randomly into k groups (folds) of approximately equal
size. The model is then fit k times, each time leaving out the data from one of
the k groups. Each time, predictions are made for the data in the group that
is left out of the analysis. And each time we assess how good these predictions
are, for example by determining the residuals and computing the mean squared
error (MSE). With k groups, we then have k MSEs, and we can compute the
mean MSE. If we do this cross-validation for several models, we can see which
model has the lowest mean MSE. That is the model that on average shows the
best prediction. This should not lead to over-fitting, because by the random
sampling into k sub-samples, we are no longer dependent on one particular
subset of the data. Usually, a value of 5 or 10 is used for k.

15.1.3 The p > n problem

Table 15.1: Small data set illustrating the p larger than n problem.
X Y
0.00 0.64
1.00 2.08
2.00 3.33

Suppose we have a data set consisting of 2 variables, X and Y , with 3

observations. This data set is tabulated in Table 15.1. So we have n = 3
and p = 2. If we plot the data and fit a regression line with the least squares
criterion, we encounter no problem, see Figure 15.4. This is because p is smaller
than n. Now let’s see what happens when p and n are of equal size. Suppose
we omit the first observation and plot only the second and third observations.
These are plotted in Figure 15.5. If we now fit a line using the least squares
criterion, we see that we can only fit a line without residuals: the line shows a
perfect fit. You can probably imagine that for any two data points, a linear line
will always show a perfect fit without residuals. The variance of the residuals
will be 0. The software that you use for fitting such a model will give you some
sort of warning. R will say that there are no residual degrees of freedom. SPSS
will say nothing special but will also show a 0 for the number of residual degrees
of freedom. More obviously, the output will give you an intercept and a slope,
but there will be no standard errors, and hence no t-values and no p-values.

366
 p<n

2
Y

0
0.0 0.5 1.0 1.5 2.0 2.5
X

Figure 15.4: p < n: There is a unique solution that fits the least squares
criterion. No problem whatsoever.

Therefore, you can say something about the intercept and slope in the sample,
but you cannot say anything about the intercept and slope in the population.
The situation will be even worse when you have only 1 data point. Suppose
we only have the second data point, which we plot in Figure 15.6. If we then
try to fit a regression line, we will see that the software will refuse to estimate
a slope parameter. It will be fixed to 0, so that an intercept only model will
be fitted (see Section 5.15). It simply is impossible for the software to decide
what regression line to pick that goes through this one data point: there is an
infinite number of regression lines that go through this data point! Therefore,
for a two-parameter model like a regression model (the two parameters being
the intercept and slope), you need at least two data points for the model to run,
and at least three data points to get standard errors and do inference.
The same is true for larger n and larger models. For example, a multiple
regression model with 10 predictor variables together with an intercept will have
11 parameters. Such a model will not give standard errors when you have 11
observations in your data matrix, and it will not run if you have fewer than 11
observations.
In sum, the number of data points, n, should always exceed the number of
parameters in your model. That means that if you have a lot of variables in
your data file, you cannot always use them in your analysis, because you simply
do not have enough rows in your data matrix to estimate the model parameters.

15.1.4 Steps in big data analytics

Now that we are familiar with the differences between classical data analysis
and big data analytics and the major concepts, we can look at how all the
elements of big data analytics fit together. In big data problems, we often see

367
 p=n

2
Y

0
0.0 0.5 1.0 1.5 2.0 2.5
X

Figure 15.5: p = n: There is a unique solution, but there are no degrees of

freedom left. Standard errors cannot be determined, so no inference regarding
the population parameters is possible.

p>n

2
Y

0
0.0 0.5 1.0 1.5 2.0 2.5
X

Figure 15.6: p > n: There is an infinite number of lines through the data point,
but there is no criterion that determines which is best. The problem of the least
squares regression line is not defined with only one data point.

368
the following steps:

1. Problem identification. You need to know what the problem is: what do
you want to know? What do you want to predict? How good does the
prediction have to be? How fast does it have to be: real-time?

2. Selection of data sources. Depending on what you want to know or predict,

you have to think about possible sources of information. Are there any
websites that already contain part of the information that you are looking
for? Are there any databases you can get access to?

3. Feature selection. From the data sources you have access to, what features
are of interest? For example, from spoken interviews, are you mainly
interested in the words spoken by the patient? Or perhaps interested in
the length of periods of silence, or perhaps in changes in pitch? There are
so many features you could extract from data.

4. Construction of a data matrix. Once you have decided what features you
want to extract from the raw data, you have to put this information into
a data matrix. You have to decide what to put in the rows (your units of
observation), and in the columns (the features, now variables). So what is
now your variable: this could be the length of one period of silence within
one interview for a particular patient. But it could also be the average
pitch for a 1-minute interval in one interview for one particular patient.

5. Training and test (validation) data set. In order to check that we are not
over-fitting, and to make sure that our model will work for future data,
we divide our data set (our data matrix) into two parts: training data and
test data. This is done by taking a random sample of all the data that
we have. Usually, a random sample of 70% is used for training, and the
remaining 30% is used for testing (validating) the model. We set the test
data aside and will only look at the training data.

6. Model selection. In data science it is very common to try out various

models or various sub-models on the training data to see which model fits
the data best. To make sure that we do not over-fit the data, we use some
form of cross-validation, where we split up the training data even further.

7. Build the model. Once we know what model works best for our training
data, we fit that model on the training model. This fitted model is our
final model.

8. Validate the model. The final test is whether this final model will work
on new data. We don’t have new data, but we have put away some of
our data as test data (validation data). These data can now be used as
a substitute for new data to estimate how well our model will work with
future data.

369
9. Interpret the result and evaluate. There will be always some over-fitting,
so the performance on the test data will always be worse than on the
training data. But is the performance good enough to be satisfied with
the model? Is the model useful for daily practice? If not, maybe the
data sources and feature selection steps should be reconsidered. Another
important aspect of statistical learning is interpretability. There are some
very powerful models and methods around that are capable of very precise
predictions. However, the problem with these models and methods is
that they are hard to interpret: they are black boxes in that they make
predictions that cannot be explained by even the data analysts themselves.
Any decisions are therefore hard to justify, which brings ethical issues.
For instance, what would you say if an algorithm would determine on the
basis of all your life’s data that you will not be a successful student? Of
course you would want to know on the basis of what data exactly that
decision is based. Your make-up? Your height? The colour of your skin?
Last year’s grades? Of course it would matter to you what variables are
used and how. Recent research has focused on how to make complicated
models and methods easier to interpret and help data analysts evaluate
the usefulness and applicability of their results and communicate them to
others.

370
Appendices

371
Appendix A

Cumulative probabilities for

the standard normal
distribution

373
 0.4

0.3

density
0.2

0.1
p
0.0
−2 −1 0 1 2
z

Table A.1: Cumulative proportions (p) for the standard normal

distribution.

z p z p z p z p z p z p
-4.00 0.0000 -1.43 0.0764 -0.71 0.2389 0.01 0.5040 0.73 0.7673 1.45 0.9265
-3.80 0.0001 -1.42 0.0778 -0.70 0.2420 0.02 0.5080 0.74 0.7704 1.46 0.9279
-3.60 0.0002 -1.41 0.0793 -0.69 0.2451 0.03 0.5120 0.75 0.7734 1.47 0.9292
-3.40 0.0003 -1.40 0.0808 -0.68 0.2483 0.04 0.5160 0.76 0.7764 1.48 0.9306
-3.20 0.0007 -1.39 0.0823 -0.67 0.2514 0.05 0.5199 0.77 0.7794 1.49 0.9319
-3.00 0.0013 -1.38 0.0838 -0.66 0.2546 0.06 0.5239 0.78 0.7823 1.50 0.9332
-2.90 0.0019 -1.37 0.0853 -0.65 0.2578 0.07 0.5279 0.79 0.7852 1.51 0.9345
-2.80 0.0026 -1.36 0.0869 -0.64 0.2611 0.08 0.5319 0.80 0.7881 1.52 0.9357
-2.70 0.0035 -1.35 0.0885 -0.63 0.2643 0.09 0.5359 0.81 0.7910 1.53 0.9370
-2.60 0.0047 -1.34 0.0901 -0.62 0.2676 0.10 0.5398 0.82 0.7939 1.54 0.9382
-2.50 0.0062 -1.33 0.0918 -0.61 0.2709 0.11 0.5438 0.83 0.7967 1.55 0.9394
-2.40 0.0082 -1.32 0.0934 -0.60 0.2743 0.12 0.5478 0.84 0.7995 1.56 0.9406
-2.30 0.0107 -1.31 0.0951 -0.59 0.2776 0.13 0.5517 0.85 0.8023 1.57 0.9418
-2.20 0.0139 -1.30 0.0968 -0.58 0.2810 0.14 0.5557 0.86 0.8051 1.58 0.9429
-2.10 0.0179 -1.29 0.0985 -0.57 0.2843 0.15 0.5596 0.87 0.8078 1.59 0.9441
-2.00 0.0228 -1.28 0.1003 -0.56 0.2877 0.16 0.5636 0.88 0.8106 1.60 0.9452
-1.99 0.0233 -1.27 0.1020 -0.55 0.2912 0.17 0.5675 0.89 0.8133 1.61 0.9463
-1.98 0.0239 -1.26 0.1038 -0.54 0.2946 0.18 0.5714 0.90 0.8159 1.62 0.9474
-1.97 0.0244 -1.25 0.1056 -0.53 0.2981 0.19 0.5753 0.91 0.8186 1.63 0.9484
-1.96 0.0250 -1.24 0.1075 -0.52 0.3015 0.20 0.5793 0.92 0.8212 1.64 0.9495
-1.95 0.0256 -1.23 0.1093 -0.51 0.3050 0.21 0.5832 0.93 0.8238 1.65 0.9505
-1.94 0.0262 -1.22 0.1112 -0.50 0.3085 0.22 0.5871 0.94 0.8264 1.66 0.9515
-1.93 0.0268 -1.21 0.1131 -0.49 0.3121 0.23 0.5910 0.95 0.8289 1.67 0.9525
-1.92 0.0274 -1.20 0.1151 -0.48 0.3156 0.24 0.5948 0.96 0.8315 1.68 0.9535
-1.91 0.0281 -1.19 0.1170 -0.47 0.3192 0.25 0.5987 0.97 0.8340 1.69 0.9545
-1.90 0.0287 -1.18 0.1190 -0.46 0.3228 0.26 0.6026 0.98 0.8365 1.70 0.9554
-1.89 0.0294 -1.17 0.1210 -0.45 0.3264 0.27 0.6064 0.99 0.8389 1.71 0.9564
-1.88 0.0301 -1.16 0.1230 -0.44 0.3300 0.28 0.6103 1.00 0.8413 1.72 0.9573
-1.87 0.0307 -1.15 0.1251 -0.43 0.3336 0.29 0.6141 1.01 0.8438 1.73 0.9582
-1.86 0.0314 -1.14 0.1271 -0.42 0.3372 0.30 0.6179 1.02 0.8461 1.74 0.9591
-1.85 0.0322 -1.13 0.1292 -0.41 0.3409 0.31 0.6217 1.03 0.8485 1.75 0.9599
-1.84 0.0329 -1.12 0.1314 -0.40 0.3446 0.32 0.6255 1.04 0.8508 1.76 0.9608
Continued on next page

374
 Table A.1: Cumulative proportions (p) for the standard normal
distribution.

z p z p z p z p z p z p
-1.83 0.0336 -1.11 0.1335 -0.39 0.3483 0.33 0.6293 1.05 0.8531 1.77 0.9616
-1.82 0.0344 -1.10 0.1357 -0.38 0.3520 0.34 0.6331 1.06 0.8554 1.78 0.9625
-1.81 0.0351 -1.09 0.1379 -0.37 0.3557 0.35 0.6368 1.07 0.8577 1.79 0.9633
-1.80 0.0359 -1.08 0.1401 -0.36 0.3594 0.36 0.6406 1.08 0.8599 1.80 0.9641
-1.79 0.0367 -1.07 0.1423 -0.35 0.3632 0.37 0.6443 1.09 0.8621 1.81 0.9649
-1.78 0.0375 -1.06 0.1446 -0.34 0.3669 0.38 0.6480 1.10 0.8643 1.82 0.9656
-1.77 0.0384 -1.05 0.1469 -0.33 0.3707 0.39 0.6517 1.11 0.8665 1.83 0.9664
-1.76 0.0392 -1.04 0.1492 -0.32 0.3745 0.40 0.6554 1.12 0.8686 1.84 0.9671
-1.75 0.0401 -1.03 0.1515 -0.31 0.3783 0.41 0.6591 1.13 0.8708 1.85 0.9678
-1.74 0.0409 -1.02 0.1539 -0.30 0.3821 0.42 0.6628 1.14 0.8729 1.86 0.9686
-1.73 0.0418 -1.01 0.1562 -0.29 0.3859 0.43 0.6664 1.15 0.8749 1.87 0.9693
-1.72 0.0427 -1.00 0.1587 -0.28 0.3897 0.44 0.6700 1.16 0.8770 1.88 0.9699
-1.71 0.0436 -0.99 0.1611 -0.27 0.3936 0.45 0.6736 1.17 0.8790 1.89 0.9706
-1.70 0.0446 -0.98 0.1635 -0.26 0.3974 0.46 0.6772 1.18 0.8810 1.90 0.9713
-1.69 0.0455 -0.97 0.1660 -0.25 0.4013 0.47 0.6808 1.19 0.8830 1.91 0.9719
-1.68 0.0465 -0.96 0.1685 -0.24 0.4052 0.48 0.6844 1.20 0.8849 1.92 0.9726
-1.67 0.0475 -0.95 0.1711 -0.23 0.4090 0.49 0.6879 1.21 0.8869 1.93 0.9732
-1.66 0.0485 -0.94 0.1736 -0.22 0.4129 0.50 0.6915 1.22 0.8888 1.94 0.9738
-1.65 0.0495 -0.93 0.1762 -0.21 0.4168 0.51 0.6950 1.23 0.8907 1.95 0.9744
-1.64 0.0505 -0.92 0.1788 -0.20 0.4207 0.52 0.6985 1.24 0.8925 1.96 0.9750
-1.63 0.0516 -0.91 0.1814 -0.19 0.4247 0.53 0.7019 1.25 0.8944 1.97 0.9756
-1.62 0.0526 -0.90 0.1841 -0.18 0.4286 0.54 0.7054 1.26 0.8962 1.98 0.9761
-1.61 0.0537 -0.89 0.1867 -0.17 0.4325 0.55 0.7088 1.27 0.8980 1.99 0.9767
-1.60 0.0548 -0.88 0.1894 -0.16 0.4364 0.56 0.7123 1.28 0.8997 2.00 0.9772
-1.59 0.0559 -0.87 0.1922 -0.15 0.4404 0.57 0.7157 1.29 0.9015 2.10 0.9821
-1.58 0.0571 -0.86 0.1949 -0.14 0.4443 0.58 0.7190 1.30 0.9032 2.20 0.9861
-1.57 0.0582 -0.85 0.1977 -0.13 0.4483 0.59 0.7224 1.31 0.9049 2.30 0.9893
-1.56 0.0594 -0.84 0.2005 -0.12 0.4522 0.60 0.7257 1.32 0.9066 2.40 0.9918
-1.55 0.0606 -0.83 0.2033 -0.11 0.4562 0.61 0.7291 1.33 0.9082 2.50 0.9938
-1.54 0.0618 -0.82 0.2061 -0.10 0.4602 0.62 0.7324 1.34 0.9099 2.60 0.9953
-1.53 0.0630 -0.81 0.2090 -0.09 0.4641 0.63 0.7357 1.35 0.9115 2.70 0.9965
-1.52 0.0643 -0.80 0.2119 -0.08 0.4681 0.64 0.7389 1.36 0.9131 2.80 0.9974
-1.51 0.0655 -0.79 0.2148 -0.07 0.4721 0.65 0.7422 1.37 0.9147 2.90 0.9981
-1.50 0.0668 -0.78 0.2177 -0.06 0.4761 0.66 0.7454 1.38 0.9162 3.00 0.9987
-1.49 0.0681 -0.77 0.2206 -0.05 0.4801 0.67 0.7486 1.39 0.9177 3.20 0.9993
-1.48 0.0694 -0.76 0.2236 -0.04 0.4840 0.68 0.7517 1.40 0.9192 3.40 0.9997
-1.47 0.0708 -0.75 0.2266 -0.03 0.4880 0.69 0.7549 1.41 0.9207 3.60 0.9998
-1.46 0.0721 -0.74 0.2296 -0.02 0.4920 0.70 0.7580 1.42 0.9222 3.80 0.9999
-1.45 0.0735 -0.73 0.2327 -0.01 0.4960 0.71 0.7611 1.43 0.9236 4.00 1.0000
-1.44 0.0749 -0.72 0.2358 0.00 0.5000 0.72 0.7642 1.44 0.9251 4.01 1.0000

375
376
Appendix B

Critical values for the

t-distribution
0.4

0.3
density

0.2

0.1
p
0.0
−2 −1 0 1 2
t

377
Table B.1: Values of the t-distribution, given the degrees of freedom (rows) and
tail probability p (columns). These can be used for critical values for a given
confidence level.
df 0.15 0.10 0.05 0.025 0.02 0.01 0.005 0.0025 0.001 0.0005
1 1.963 3.078 6.314 12.71 15.90 31.82 63.66 127.3 318.3 636.6
2 1.386 1.886 2.920 4.303 4.849 6.965 9.925 14.09 22.33 31.60
3 1.250 1.638 2.353 3.182 3.482 4.541 5.841 7.453 10.22 12.92
4 1.190 1.533 2.132 2.776 2.999 3.747 4.604 5.598 7.173 8.610
5 1.156 1.476 2.015 2.571 2.757 3.365 4.032 4.773 5.893 6.869
6 1.134 1.440 1.943 2.447 2.612 3.143 3.707 4.317 5.208 5.959
7 1.119 1.415 1.895 2.365 2.517 2.998 3.499 4.029 4.785 5.408
8 1.108 1.397 1.860 2.306 2.449 2.896 3.355 3.833 4.501 5.041
9 1.100 1.383 1.833 2.262 2.398 2.821 3.250 3.690 4.297 4.781
10 1.093 1.372 1.812 2.228 2.359 2.764 3.169 3.581 4.144 4.587
11 1.088 1.363 1.796 2.201 2.328 2.718 3.106 3.497 4.025 4.437
12 1.083 1.356 1.782 2.179 2.303 2.681 3.055 3.428 3.930 4.318
13 1.079 1.350 1.771 2.160 2.282 2.650 3.012 3.372 3.852 4.221
14 1.076 1.345 1.761 2.145 2.264 2.624 2.977 3.326 3.787 4.140
15 1.074 1.341 1.753 2.131 2.249 2.602 2.947 3.286 3.733 4.073
16 1.071 1.337 1.746 2.120 2.235 2.583 2.921 3.252 3.686 4.015
17 1.069 1.333 1.740 2.110 2.224 2.567 2.898 3.222 3.646 3.965
18 1.067 1.330 1.734 2.101 2.214 2.552 2.878 3.197 3.610 3.922
19 1.066 1.328 1.729 2.093 2.205 2.539 2.861 3.174 3.579 3.883
20 1.064 1.325 1.725 2.086 2.197 2.528 2.845 3.153 3.552 3.850
21 1.063 1.323 1.721 2.080 2.189 2.518 2.831 3.135 3.527 3.819
22 1.061 1.321 1.717 2.074 2.183 2.508 2.819 3.119 3.505 3.792
23 1.060 1.319 1.714 2.069 2.177 2.500 2.807 3.104 3.485 3.768
24 1.059 1.318 1.711 2.064 2.172 2.492 2.797 3.091 3.467 3.745
25 1.058 1.316 1.708 2.060 2.167 2.485 2.787 3.078 3.450 3.725
26 1.058 1.315 1.706 2.056 2.162 2.479 2.779 3.067 3.435 3.707
27 1.057 1.314 1.703 2.052 2.158 2.473 2.771 3.057 3.421 3.690
28 1.056 1.313 1.701 2.048 2.154 2.467 2.763 3.047 3.408 3.674
29 1.055 1.311 1.699 2.045 2.150 2.462 2.756 3.038 3.396 3.659
30 1.055 1.310 1.697 2.042 2.147 2.457 2.750 3.030 3.385 3.646
40 1.050 1.303 1.684 2.021 2.123 2.423 2.704 2.971 3.307 3.551
50 1.047 1.299 1.676 2.009 2.109 2.403 2.678 2.937 3.261 3.496
60 1.045 1.296 1.671 2.000 2.099 2.390 2.660 2.915 3.232 3.460
120 1.041 1.289 1.658 1.980 2.076 2.358 2.617 2.860 3.160 3.373
10000 1.036 1.282 1.645 1.960 2.054 2.327 2.576 2.808 3.091 3.291
70% 80% 90% 95% 96% 98% 99% 99.5% 99.8% 99.9%
Confidence level

378