L ABORATORY FOR A PPLIED S TATISTICS H ELLE S ØRENSEN

U NIVERSITY OF C OPENHAGEN AUGUST 2016

Introduction to R: Exercises
This document is a collection of exercises made for the one-day course Introduction to R given
by the Laboratory for Applied Statistics, Department of Mathematical Sciences, University of
Copenhagen. Please read Section 1 about the purpose of the course and the suggested use of
the exercises. Suggestions for improvements are most welcome. For example, send an email
to [email protected].

Contents
1 About R, RStudio, the course, and the exercises 3

2 Getting started 5
2.1 Working with R and RStudio . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Built-in mathematical functions . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Simple summary statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Vectors and sample statistics (more examples)? . . . . . . . . . . . . . . . . 7
2.6 Help pages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

3 Reading data from files and working with datasets 9

3.1 Reading data from csv files, inspection of data . . . . . . . . . . . . . . . . . 9
3.2 Reading data from Excel files . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Working with datasets: Inspection, using variables, attaching . . . . . . . . . 11
3.4 Transformation of variables, subsets of datasets . . . . . . . . . . . . . . . . 12
3.5 Merging datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6 Working with data: Tomato yields? . . . . . . . . . . . . . . . . . . . . . . . 14

4 Graphs 15
4.1 Basic high-level plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2 Modifications of scatterplots . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.3 Modifications of histograms, parallel boxplots . . . . . . . . . . . . . . . . . 17
4.4 Exporting graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Regression 19
5.1 Simple and multiple linear regression: Cherry data . . . . . . . . . . . . . . 19
5.2 Simple linear regression, prediction: Heart and body weights? . . . . . . . . 20
5.3 Multiple linear regression: Toxicity of dissolutions? . . . . . . . . . . . . . . 21
5.4 Nonlinear regression: Enzyme kinetics . . . . . . . . . . . . . . . . . . . . . 22
5.5 Logistic regression: Pneumoconiosis among coalminers . . . . . . . . . . . . 24

6 Analysis of variance (ANOVA) 27

6.1 Oneway analysis of variance: Psoriasis . . . . . . . . . . . . . . . . . . . . . 27
6.2 Oneway ANOVA: Pillbugs? . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
6.3 Twoway ANOVA: Growth of soybean plants . . . . . . . . . . . . . . . . . . 31
6.4 An analysis with categorical as well as quantitative variables: FEV . . . . . . 32

7 Principal component analysis 33

7.1 PCA: Physical measurements of crabs . . . . . . . . . . . . . . . . . . . . . 33
7.2 PCA: Ecological zones along the Doubs River . . . . . . . . . . . . . . . . . 34

8 Matrices 35
8.1 Matrix operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
8.2 Least squares estimates in linear regression? . . . . . . . . . . . . . . . . . . 37

9 Functions 38
9.1 Mathematical functions of a single argument . . . . . . . . . . . . . . . . . . 38
9.2 The mean as a least squares estimate? . . . . . . . . . . . . . . . . . . . . . 39
9.3 Mathematical functions of several arguments . . . . . . . . . . . . . . . . . 39
9.4 Non-linear least squares? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
9.5 Non-mathematical functions in R? . . . . . . . . . . . . . . . . . . . . . . . 41

2
1 About R, RStudio, the course, and the exercises
What is R and RStudio? R is a statistical software program. It has extremely useful tools
for data exploration, data analysis, and data visualization. It is flexible and also allows for
advanced programming. RStudio is a user interface for R, which provides a nice environment
for working with R.

About the course and the exercises The course is aimed at researchers in natural sciences,
who need tools for data exploration, data analysis, and data visualization, and who have no or
little experience with R. The purposes of the course are to

• get the participants started with R, so they can start using R in their own work

• give the participants an impression about the possibilities with R

Course material The course material can be downloaded from

http://www.statlab.math.ku.dk/english/courses/workshops/r-intro/

and consists of the following:

• This document with exercises.

• Data files. All datsets used in exercises are saved as csv as well as xlsx files, and all
the files are gathered in a zip-archive.

• R programs and R markdown files used for the presentations.

We recommend that you make a new folder for the course, put all files concerning the course
in that folder, and use the folder as ”working directory” in RStudio. The working directory
can be changed in the Session menu.

About the exercises The exercises in this document are meant for the hands-on sessions in
the course. Some important comments on the exercises:

• The collection of exercises does not constitute anything like a text-book type introduc-
tion to R, and the exercises cannot stand alone in the participants’ later use of R.

• The exercises do not explain much (if any at all) of the statistics involved, so the partic-
ipant should not work with the exercises on analysis of variance, say, if he or she is not
familiar with that type of data analysis.

• There are far too many exercises for one day’s work (probably even if you were an
experienced R user). The idea is that participants can choose exercises according to
their interests and experience with R, so do not drive yourself to despair if you only get
through a few exercises during the day.

3
 • Many questions are of the type “Try the following commands, and explain what hap-
pens”. The last part is of course important to be able to similar commands in other
settings. The participants are strongly encouraged to play with other commands along
the way.

• Prerequisites are given in the beginning of each exercise in terms of previous exercises.
You do not necessarily have to have solved all the exercises mentioned completely,
but you would find useful commands here and there. Hopefully, I have succeeded in
remembering the majority of the prerequisites, but I don’t give any promises...

• Some exercises are marked with a star (? ). Only few commands are written directly in
these exercises, so the reader should know them from other exercises (or from some-
where else).

Installation of R and RStudio. R packages R and RStudio are avaliable for all platforms
(Windows, Mac and Linux). They are open-source programs, and they can be downloaded
from http://www.r-project.org and http://www.rstudio.org, respectively.
Apart from many functions that are immediately avalaible after installation, there exists a
huge number of R packages, i.e. collections of functions written by users and made available
to other users. Many of those packages can be downloaded and installed from the CRAN
repository — most easily in RStudio via the Package menu in the lower right window. After
installation the package must be loaded before the functions in the package are available for
use, again via the Package menu in RStudio. Notice that packages should be installed one
time only, but must be loaded in each R session where you need it.

Material about R There are many monographs on the use of R, as well as enormous
amounts of online material. Among many others, the following books could be adequate
for the course participants: Ekstrøm and Sørensen (2011) and Dalgaard (2008) are textbooks
in (applied) statistics with comprehensive use of R. Both are written for introductory statis-
tics courses, and thus cover basic statistical tools. Martinussen et al. (2012) is about basic
use of R as well as statistical analysis with R. Some of the topics are relative advanced (ran-
dom effects, repeated measures, survival analysis), and the reader is supposed to be somewhat
familiar with the statistical methods in advance. Ekstrøm (2012) is a handbook answering
many basic as well as advanced questions about R. Venables and Ripley (2002) is a standard
reference among R users (although in principle written for the language S rather than R). It
contains data examples on a wealth of statistical methods.
Many of the exercises in this note are inspired from material from these books.

4
2 Getting started

2.1 Working with R and RStudio

Prerequisites: None
You would like to save relevant commands for later use. Therefore you should write your
commands in a file — a socalled R script or R program. The main part of this exercise learns
you how to work with such files in RStudio.

1. Start RStudio. Go to the console (lower left window) and write

3+2

at the promt. Then use R to calculate 3 · 4 and 8/2 − 3 · 4 (use * for multiplication).
Commands written directly at the prompt are not saved for later use.

2. If there is not already an editor open in the upper left window, then go to the file menu
and open a new script. Type one of the commands from before in the editor, hold the
Control buttom and push the Enter buttom. Then the command is transferred to the
console and the command is executed (just as before).

3. Make a few more commands in the editor, and run them. Also, mark several commands
with the arrow buttoms and use Control-Enter.

4. Save the file, close the file, and quit RStudio (all via the File menu). You are asked
if you want to save the workspace image. If you answer “Yes”, then everything is
saved — both relevant and irrelevant stuff — so unless you have made time-consuming
computations, you should say “No”.

5. Start RStudio again, and open the file you just saved (again via the File menu), and redo
some of the calculations.
6. Try the command
# 4+7

Nothing is computed! This is because the character # can be used for comments. Any-
thing after # is discarded by R. It cannot be recommended enough that you make ap-
propriate comments in your R programs. That makes it much easier to use the programs
again after a while.

2.2 Built-in mathematical functions

Prerequisites: None
All standard functions (and many non-standard, too) are programmed in R. Here come some
examples.

5
 1. Try the commands sqrt(16), 16^0.5. Compute 43 .

2. Try the commands log10(1000), log(1000), exp(log(1000)). Then try the com-
mand log2(64). Make sure you understand different logarithmic functions.

3. Try the command ?log. Then a help page opens in the lower right window of RStu-
dio. Read the first few lines; does the text match your observations from the previous
question?

4. Try the commands pi, round(pi), round(pi, digits=4), and trunc(pi).

5. The sine and cosine functions are implemented in sin and cos. Calculate sin(π),
cos(π), sin(π/2), cos(π/2).

2.3 Vectors
Prerequisites: None

1. Try the commands:

x <- c(3,6,8)
x
x/2
x^2
sqrt(x)

x[2]
x[c(1,3)]
x[-3]
y <- c(2,5,1)
y
x-y
x*y

x[y>1.5]
y[x==6]

4:10
seq(2,3,by=0.1)
rep(x,each=4)

An important point is that operations are carried out elementwise!

2. Assume that we have registered the height and weight for four people: Heights in cm
are 180, 165, 160, 193; weights in kg are 87, 58, 65, 100. Make two vectors, height
and weight, with the data. The bodymass index (BMI) is defined as
weight in kg
(height in m)2

6
 Make a vector with the BMI values for the four people, and a vector with the natural
logarithm to the BMI values. Finally make a vector with the weights for those people
who have a BMI larger than 25.

2.4 Simple summary statistics

Prerequisites: None
In an experiment the dry weight has been measured for 8 plants grown under certain conditions
(the values are given below).

1. Try the following commands in order to make a vector with the values and compute
various sample statistics:
dry <- c(77, 93, 92, 68, 88, 75, 100)
dry

sum(dry)
length(dry)
mean(dry)
sum(dry)/length(dry) ## Checking

sort(dry)
median(dry)

sd(dry)
var(dry)
sd(dry)^2
sum((dry-mean(dry))^2) / (length(dry)-1) ## Checking

min(dry)
max(dry)

summary(dry)

2.5 Vectors and sample statistics (more examples)?

Prerequisites: Exercises 2.3 (vectors) and 2.4 (summary statistics)

1. Assume that we have the following three observations of temperature: 23◦ C, 27◦ C,
19◦ C. Make a vector with these values. Recall the relation between the Celcius and
Fahrenheit temperature scale:
9
degress in Fahrenheit = degrees in Celcisus · + 32
5
Make a new vector with the temperatures in Fahrenheit.

7
 2. Assume that you are interested in cone-shaped structures, and have measured the height
and radius of 6 cones. Make vectors with these values as follows:
R <- c(2.27, 1.98, 1.69, 1.88, 1.64, 2.14)
H <- c(8.28, 8.04, 9.06, 8.70, 7.58, 8.34)

Recall that the volume of a cone with radius R and height H is given by 13 πR2 H. Make
a vector with the volumes of the 6 cones.

3. Compute the mean, median and standard deviation of the cone volumes. Compute also
the mean of volume for the cones with a height less than 8.5.

2.6 Help pages

Prerequisites: None
There are help pages for every function i R, but, admittedly, they are not always too easy to
read for an unexprerienced R user.

1. Try the command ?which.min. This opens a help page in the lower right window of
RStudio. What does the function do?
2. Try the commands
which.min(c(2,5,1,7,8))
which.max(c(2,5,1,1,8))

Is the output as expected?

3. You must know the name of the function in order to open the help page as above.
Sometimes (often, even) you do not know the name of the R functions; then google can
often help you. Try, for example, to search the text R minimum vector.

8
3 Reading data from files and working with datasets

3.1 Reading data from csv files, inspection of data

Prerequisites: None
Suppose that your data is saved as a csv file (csv stand for comma-separated values). We
need to read the data into R. Consider an experiment where girth, height and volume has
meen measured for 31 cherry trees. The data are saved in the file cherry.csv.

1. Open the file cherry.csv in Excel (or a similar application), and make sure you under-
stand the structure of the file.
2. Go to R, use the command
cherry <- read.csv(file.choose(), dec=’.’, sep=’;’)

Another window pops up, and you must click your way through to the relevant csv file.
Once the dataset has been constructed, the name cherry appears in the upper right box
in RStudio. Click the name, and you can see the content in the upper left box.
Notice how you can change the characters for decimal comma and field separator ac-
cording to what is used in the csv file, so the file is interpreted in the correct way (more
about this below).

3. Whenever you have constructed a new dataset (or made changes in an existing one) it is
extremely important that you check that it has been constructed correctly. The following
commands may be useful for that purpose:

head(cherry)
plot(cherry)
summary(cherry)

Try the commands!

4. Did you get lists of minimum, maximum, mean, etc., for all three variables?
If not, it is because R has coded some of the variables as categorical variables rather
than numerical variables (R reads a number with decimals as a text string). Luckily this
can be changed with the dec option in read.csv. Try instead
cherry1 <- read.csv(file.choose(), dec=’,’, sep=’;’)
summary(cherry1)

The question about coding is important, so let us say it a bit more specifically: A common
problem is that one or more numerical variables have been coded as categorical variables
because the decimal seperator has not been interpreted correctly. Then R reads a number with
decimals as a text string. This causes all sorts of trouble, and the sooner you detect it, the
better. The summary of numerical variable lists the minimum, maximum, mean, etc., for the

9
variable. If, instead, you get information about how many times specific values occur in the
dataset, then your variable is coded as a categorical value, and you have to change the dec
option in your read.csv command.

5. (Optional) You can write the file name instead of file.choose(), such that you do not
have to click your way through to the file. The path to the file must be given, starting
from the current working directory. If the file is in the current working directory, then
the relevant command is
cherry2 <- read.csv(’cherry.csv’, dec=’.’, sep=’;’)

You can change the working directory in the Session menu (Set working directory).

6. (Optional) Consider an experiment with tomatoes. Three different varieties and four
different seed densities have been tested, and there are three replications for each of the
12 combinations. The yield has been registered for each of the 36 field plots. The data
are saved in the file tomatoes.csv
Create a dataset called tomatoes, say, in R, and inspect the data with summary. Which
variables are quantitative and which variables are categorical?

3.2 Reading data from Excel files

Prerequisites: Exercise 3.1
You probably most often save your data in Excel (or something similar), and therefore need
to “transfer” your data from Excel to R. There are several ways of doing this; here comes two
if them:

(a) Read the Excel file directly into R with the function read.xlsx. This is the nicest
solution; however problems now and then occur due to different versions of Excel and
different computer systems. It also requires that you have the rights to install add-on
packages on your computer (previously this was sometimes a problem on SCIENCE-
PC’s).

(b) Save your Excel sheet as a comma separated file (.csv), and use the function read.csv
to read the file as described in Exercise 3.1. Make sure that the decimal seperator and
the field seperator in read.csv match those used to save the csv file.

Let us consider the same data as in Exercise 3.1: Girth, height and volume has meen measured
for 31 cherry trees. The data are saved in the file cherry.xlsx.
We first try method (a). Here is something that works on my computer:

1. First install the package xlsx via the Package menu (this requires connection to the
internet). Second load the package via the Package menu. See Section 1 for more
information about R packages.
2. Use the command

10
 cherry3 <- read.xlsx(file.choose(), sheetIndex=1)

and choose the relevant file.

3. Inspect the dataset cherry3. In particular, make sure that all variable are coded cor-
rectly as numerial variables.

If method (a) did not work, then try method (b):

4. Open the file cherry.xlsx in Excel. Go to the file menu, and save the data as a csv
file (use Save As). As a starting point, you can use the default options for delimiters.
5. Go to R and use the command
cherry4 <- read.csv(file.choose(), dec=’.’, sep=’;’)

and choose the csv file that you just created.

6. Inspect the dataset cherry4. In particular, make sure that all variable are coded cor-
rectly as numerial variables. If this is not the case, then try to modify the dec and sep
options so they match those from the csv file.

3.3 Working with datasets: Inspection, using variables, attaching

Prerequisites: Exercise 3.1 (reading the cherry data)
Consider the dataset cherry from Exercise 3.1 or 3.2.

1. (Inspection of data) Click the name cherry in the upper right box in RStudio (as you
probably already did in Exercise 3.1). Then try the following commands (one at a time)
and explain what you see:
head(cherry)
cherry
plot(cherry)
summary(cherry)

2. (The $ syntax) Try the following commands one at a time; notice that the first two give
you an error message:
Girth
hist(Girth)
cherry$Girth
mean(cherry$Girth)
hist(cherry$Girth)

3. (The with function) Try the following commands:

11
 with(cherry, hist(Height))
with(cherry, mean(Height))

4. (Attaching a dataset) Try the following commands one at a time:

attach(cherry)
Girth
mean(Girth)
hist(Girth)
plot(Height, Volume)
detach(cherry)
Girth

The point of questions 2–4 is that you need to tell R where to look for the variables. You
do so either by using the $ syntax every time you need the variable (which quickly becomes
tendersome), by using the with function ”outside the command”, or by attaching the dataset.
If you attach datasets, you should be very careful not to have several datasets with the same
variable names attached at the same time. This can be very dangerous because it is not always
obvious which dataset is then used.

5. (Structure of datasets) Make sure you understand the structure of a dataset. More specif-
ically, for example: What is the difference between a variable and a dataset? What do
you see in the different rows of a dataset? What do you see in the different columns in
a dataset?

3.4 Transformation of variables, subsets of datasets

Prerequisites: Exercise 3.1 (reading the cherry data)
Consider the dataset cherry from Exercise 3.1 or 3.2.

1. (Transformation of variables in a dataset) Construct a new dataset, cherry1, with the

command
cherry1 <- transform(cherry, logVolume=log(Volume), logGirth=log(Girth))

and notice that the name appears in the upper right box of RStudio. Inspect the new
dataset, either by clicking on the name in the upper right box, or by one of the com-
mands
cherry1
head(cherry1)
hist(cherry1$logVolume)

Notice that log is the natural logarithm. If you prefer the logarithm with base 2 or 10,
you should use the functions log2 or log10 instead og log.

12
 2. (Subset of a dataset) Try the following commands (one at a time). Make sure you
understand the output.
cherry[3,]
cherry[3:5,]
cherry[-c(2,4),]

Construct and inspect each of the following datasets (one at at time). Make sure you
understand the content of each dataset.
subset(cherry, Height>70)
subset(cherry, Height>=70)
subset(cherry, Height==80)
subset(cherry, Height==80, select=c(Girth,Volume))
subset(cherry, Height>80 & Girth>15)
subset(cherry, Height>80 | Girth>15)

3.5 Merging datasets

Prerequisites: Exercise 3.1 (reading the cherry data)
It sometimes happens that data comes from different sources (files), and should be merged be-
fore analysis. Consider the dataset cherry from Exercise 3.1. Always check that the datasets
are as you supposed them to be!

1. (Merging data, new datalines) Assume that data from two more trees are made avail-
able. Try the following commands that construct a dataset with the new data and merge
the two datasets:
newData <- data.frame(Girth=c(11.5, 17.0), Height=c(71, 75), Volume=c(22, 40))
newData
allData <- rbind(cherry, newData)
allData

2. (Merging data, new variables) Assume that the precipitation at the location of each tree
has also been registered and saved in a variable called precipitation. The variable
below contains random numbers with mean 50 and standard deviation 10, and are mean-
ingless — yet the commands illustrate the merging of a new variable and an existing
dataset.
precipitation <- rnorm(n=31, mean=50, sd=10)
precipitation
allData2 <- cbind(cherry, precipitation)
allData2

Of course the variables should be ordered the same way in the dataset and the new
variable. See (Martinussen et al., 2012, Section 4.2) for more advanced merging using
the merge function.

13
3.6 Working with data: Tomato yields?
Prerequisites: Exercises 3.1, 3.3, 3.4 (reading and working with datasets)
Consider an experiment with tomatoes. Three different varieties and four different seed den-
sities have been tested, and there are three replications for each of the 12 combinations.
The yield has been registered for each of the 36 field plots. The data are saved in the file
tomatoes.xlsx

1. Create a dataset called tomatoes, say, in R, and inspect the data with the functions plot
and summary. Which variables are quantitative and which variables are categorical?

2. Make a histogram of the yield variable. Moreover, compute the mean, median and
standard deviation of the yield variable.
3. Try the following commands, and explain the output:
table(variety)
table(density)
table(density, variety)

4. Make a new dataset with two more variables: one with the squareroot of the yield values
and one with the logarithm of the yield values.

5. Make a dataset which only contains the datalines corresponding to the variety Ife (use
variety==’Ife’ in a subset command).

6. Make a dataset containing datalines for the variety Pusa with density less than 25000.
What is the median of the corresponding yield values?

7. Make a dataset without observation numbers 5 and 24. Make a histogram of the yield
values from this dataset.

14
4 Graphs

4.1 Basic high-level plots

Prerequisites: Exercises 3.1 and 3.3 (reading and working with the cherry data)
This exercise introduces some basic high-level plots, which can be produced with very simple
commands. Here we use the default options (which are often fine), but it is an important point
that graphs can easily be modified, see Exercise 4.2.
Consider first the dataset cherry from Exercise 3.1, and recall that it has variables Girth,
Height and Volume.

1. (Scatterplots) The most important graphics function plot function. It does different
things depending on the type of argument(s) supplied to the function.
Try the commands and explain what you see:
plot(cherry)
plot(Girth, Volume)
plot(log(Girth), log(Volume))
plot(Height)

2. (Histograms and boxplots) Try the commands

hist(Volume)
boxplot(Volume)

3. (Barplots) The temperature in New York was measured daily for five month (May to
September). The average temperatures in degrees Fahrenheit were 65.5, 79.1, 83.9,
84.0, and 76.9, respectively. Try the commands
temp <- c(65.5, 79.1, 83.9, 84.0, 76.9)
barplot(temp, names=5:9)

4.2 Modifications of scatterplots

Prerequisites: Exercises 3.1 and 3.3 (reading and working with data)
The temperature in New York was measured daily for five month (May to September). The
data from July and August are saved in the files ny-temp.xlsx and ny-temp.csv. There
are three variable: Month with values 7 and 8, Day with values 1–31, and Temp with the
temperature in degrees Fahrenheit.

1. Read the data into a dataset, NYtemp. Then make two sub-dataset, july and august
with the data from July and August, respectively.
2. (Plotting symbols, lines, axes, title) Try the following commands and explain what hap-
pens:

15
 attach(july)
plot(Day, Temp)
plot(Day, Temp, pch=16)
plot(Day, Temp, pch=2)
plot(Day, Temp, type="l")
plot(Day, Temp, type="b")
plot(Day, Temp, type="l", lty=2)
plot(Day, Temp, type="l", lwd=2)
plot(Day, Temp, type="l", col="blue")
plot(Day, Temp, type="l", xlab="Day in month", ylab="Temperature (F)",
main="Temperatur in New York")
plot(Day, Temp, type="l", cex.lab=1.3)
plot(Day, Temp, type="l", cex.axis=1.3)
plot(Day, Temp, type="l", xlim=c(0,40), ylim=c(50,100))
detach(july)

3. (Adding data points) Try the following commands and explain what you see:
range(july$Temp)
range(august$Temp)
plot(july$Day, july$Temp, type="l", ylim=c(70,100))
points(august$Day, august$Temp, col="red")
plot(july$Day, july$Temp, type="l", ylim=c(70,120))
lines(august$Day, august$Temp, col="green")

4. (Legends) Add a legend to the previous plot:

legend(5,120, c("July", "August"), text.col=c("Black","Green"))

5. (Adding straight lines) Try the commands:

plot(july$Day, july$Temp, type="l", ylim=c(70,100))
abline(h=73, col="red")
abline(v=14, col="blue")
abline(84.652, -0.0468)

Notice how the final command adds a line to the plot with intercept and slope as spec-
ified in the command. The actual values are the estimates from a linear regression, see
Exercise 5.1, question 1, on how to make R compute those numbers.
6. (Points coloured according to third variable) In the R package MASS there is a dataset
called cats with variables Sex, Bwt (body weight in kg), Hwt (in g). Run the following
commands:
library(MASS)
data(cats)
plot(cats)

Then try the command and explain what you see:

16
 Cats

20
Female

18
Male ●
●

16
Heart weight (g)
●
●
●
● ●
● ●
● ● ●

14
●
●
● ●
●
● ● ●
● ● ● ●
● ● ●
● ● ● ●
●

12
● ● ●
● ● ● ●
● ● ●
● ● ● ● ●
● ●
● ● ●
●
● ●
● ●
● ●
● ● ● ●
● ●

10
● ● ● ● ● ● ●
● ● ● ●
● ● ● ●
● ● ●
● ●
● ● ● ●
● ● ● ● ● ●
● ● ●
● ● ●
● ● ● ●
● ● ●
●
●

8
● ● ● ● ●
● ● ●
● ● ● ●
● ●
●
●
●
●
6

2.0 2.5 3.0 3.5

Body weight (kg)

Figure 1: Scatterplot with regression lines for male and female cats.

plot(cats$Bwt, cats$Hwt, col=cats$Sex)

7. Modify the figure such that it looks similar to Figure 1. The lines are regression lines
for linear regression models fitted to male and female cats, respectively, namely

Male cats: HWt = −1.184 + 4.313 · Bwt

Female cats: HWt = 2.981 + 2.636 · Bwt

(See Exercise 5.1, question 1, on how to fit such linear regressions).

4.3 Modifications of histograms, parallel boxplots

Prerequisites: None

1. In the R package MASS there is a dataset called cats. Run the following commands:
library(MASS)
data(cats)
plot(cats)

The variables Bwt and Hwt contain the weight of the body (kg) and the heart (g), respec-
tively. There are both male and female cats.
2. (Modification of histograms) Try the following commands and explain what happens:
attach(cats)
hist(Hwt[Sex=="M"])
hist(Hwt[Sex=="M"], prob=T)
hist(Hwt[Sex=="M"], breaks=c(5,10,15,20,25)) ## Not nice

Use the xlab and main arguments to change the x-label and the title to something more
appropriate.

17
 3. (Parallell boxplots) Try the following commands and explain what happens:
boxplot(Hwt[Sex=="F"])
boxplot(Hwt[Sex=="M"])
boxplot(Hwt ~ Sex)

4.4 Exporting graphics

Prerequisites: Exercise 4.2 (the temperature data)
Consider the data from Exercise 4.2.
There are (at least) two different ways to save your plots as files as pdf files that can be used
for other documents. The plots can also be saved in other formats, and it is possible to set the
size (width and height) of the graphs if the default is not satisfactory.

1. (Save an existing plot) Try the following commands:

plot(july$Day, july$Temp)
dev.print(file="plotfile1.pdf", device=pdf)

The file plotfile1.pdf should now appear in the working directory. You can get
information about the current working directory with the command getwd(). Find the
file and open it.
2. (Plot directly in a pdf file) Try the following commands:
pdf("plotfile3.pdf")
plot(july$Day, july$Temp, col="blue")
dev.off()

The commands open a file in the working directory, makes the graph in the file, and
closes the file again. Find the file and open it.

18
5 Regression

5.1 Simple and multiple linear regression: Cherry data

Prerequisites: Exercises 3.1 and 3.3 (reading and working with the cherry data)
Consider the dataset cherry from Exercise 3.1. We will fit an examine a linear regression
modelling the expected value of volume as a linear function of girth.

1. (Fit of model and plots) Use the following commands:

plot(cherry$Girth, cherry$Volume)
lm(Volume ~ Girth, data=cherry)

This gives you the estimated intercept and slope. Notice how we specify the dataset;
then it is not necessary to attach the data or use the $ syntax (cf. Exercise 3.3.) It is
often convenient to give the object with the fit a name. Try
linreg1 <- lm(Volume ~ Girth, data=cherry)
linreg1
abline(linreg1)

2. (Estimates, standard errors, tests) We need more information than just the estimated
coefficients. Try
summary(linreg1)

Some explanation of the output is probably needed. The “Coefficients” part is the most
important, and it is important to understand its structure. It has two lines — one for
each parameter in the model — and four columns. The first line is about the intercept,
the second line is about the slope. The columns are

• Estimate: The estimated value of the parameter (intercept or slope)

• Std. Error: The standard error (estimated standard deviation) associated with
the estimate in the first column
• t value: The t-test statistic for the hypothesis that the corresponding parameter
is zero. Computed as the estimate divided by the standard error.
• Pr(>|t|): The p-value associated with the hypothesis just mentioned. In partic-
ular the p-value in the second line is for the hypothesis that there is no effect of
girth on volume.

Below the “Coefficients” part you find, among others, the “Residual standard error”, i.e.
the estimated standard deviation for the observations.
Identify the estimates and corresponding standard errors in the output. Is there a sig-
nificant effect of girth on volume? What is the estimated standard deviation for the
observations?

19
 3. (Confidence intervals) Try the command confint(linreg1), which gives you 95%
confidence intervals for the intercept and slope.

4. (Model validation) The easiest way to make model validation plots for the model fit is
as follows:

par(mfrow=c(2,2))
plot(linreg1)

Notice how the command par(mfrow=c(2,2)) splits the plot window into 4 subplots.
Does the model seem to be appropriate for the data?
Fitted values, raw residuals and standardized residuals from the fit are extracted with the
functions fitted, residuals and rstandard, respectively, so the classical model val-
idation plots can also be obtained as follows (the commands with abline add relevant
lines to the plots):

plot(fitted(linreg1), residuals(linreg1))
abline(h=0)
plot(fitted(linreg1), rstandard(linreg1))
abline(h=0)
qqnorm(rstandard(linreg1))
abline(0,1)

5. (Transformation) Fit a new linear regression model where you use log(Volume) as the
response variable and log(Girth) as covariate.
Is the model appropriate for the data? Is the effect of log(Girth) significant? What
is the estimated relation between log(Girth) and log(Volume)? Which relation be-
tween Girth and Volume does this correspond to?
6. (Multiple linear regression) Fit a multiple linear regresison model where log(Girth)
as well as log(Height) are used as covariate as follows:
linreg3 <- lm(log(Volume) ~ log(Girth) + log(Height), data=cherry)

Are both covariates significant (use summary)? Finally try the command
anova(linreg3, linreg2)

which carries out the F-test for comparison of the two models. Did you see the p-value
before (explain where and why)?

5.2 Simple linear regression, prediction: Heart and body weights?

Prerequisites: Exercises 3.3 (working with datasets) and 5.1 (linear regression)

1. In the R package MASS there is a dataset called cats. Run the following commands:

20
 library(MASS)
data(cats)

Have a look at the dataset. The variables Bwt and Hwt give the weight of the body (kg)
and the heart (g), respectively. There are both male and female cats. Make a dataset
with the data from males only.

2. Make a scatterplot of the data for the male cats (Bwt on x-axis, Hwt on y-axis). Does it
look reasonable to use a linear regression model for the data?

3. Fit a linear regresison model for the male cats, that allows for prediction of the heart
weight given the body weight. Add the fitted regression line to the scatterplot from the
previous question.

4. Find the coefficients of the fitted line. How large is the expected difference in heart
weight for two cats with a difference of 1 kg in bodyweight? Find a confidence interval
for this difference? How large is the expected difference in heart weight for two cats
with a difference of 100 g in bodyweight?

5. Use model validation plot to examine if the model is appropriate for the data.
6. Use the estimates to find the expected heart weight for a male cat that weighs 3 kg.
Then try the commands (where you replace the name regModel with whatever name
you gave the the model fit in question 2).
newObs <- data.frame(Bwt=3)
newObs
predict(regModel, newObs)
predict(regModel, newObs, interval="predict")

5.3 Multiple linear regression: Toxicity of dissolutions?

Prerequisites: Exercises 3.1, 3.3 (reading and working with data) and 5.1 (linear regression)
Data from 24 chemical dissolutions have been collected in order to examine the association
between the toxicity of the dissolution on the one side and three explanatory variables on
the other side. The data are saved in the files lser.xlsx and lser.csv with the following
variables:

• tox: Toxicity of the dissolution

• base: Ablility to accept hydrogen ions

• acid: Ablility to liberate hydrogen ions

• colour: Ability to change colour

1. Make an R dataset called lser with the data, and use the command plot(lser) to get
an overview of the data.

21
 2. Fit a multiple linear regression model with tox as response and base, acid and colour
as explanatory variables. Is the model appropriate for the data? What is the interpreta-
tion of the parameter estimates?
3. Are all three explanatory variables significant? Remove insignificant variables (one at
a time) until all terms are significant.
4. Calculate the expected toxicity for a solvent which has base=0.60, acid=0.95, and
colour=0.52.

5.4 Nonlinear regression: Enzyme kinetics

Prerequisites: Exercises 3.1, 3.3 (reading and working with data), and 4.1 (scatter plot)
In a chemical experiment the enzyme activity has been measured for different concentrations
of the substrate and different concentrations of an inhibitor. The enzyme activity is measured
as a reaction rate. There are three measurement series corresponding to three concentrations
of the inhibitor (no inhibitor, 50 µM, 100 µM). For each series the reaction rate has been
measured for 6 different substrate concentrations ranging from 10 µM to 600 µM. There are
two replications and therefore 36 observations in total.
Data are available in the files inhib.xlsx and inhib.csv with the following variables:

• S: Concentration of substrate
• I: Concentration of inhibitor
• R: Reaction rate

1. Make a scatterplot of the data with S on the x-axis and R on the y-axis. Why is this plot
not very illustrative? Try the commands
grp <- c(rep(1,times=12), rep(2,times=12), rep(3,times=12))
plot(S, R, col=grp)
plot(S, R, pch=grp)

When there is no inhibitor, the association between substrate concentration and reaction rate
is most often described by the so-called Michaelis-Menten relation:
Vmax · S
R≈ (1)
K +S
where Vmax and K are parameters that should be estimated from the data. This is typically done
with least squares. The function nls can do this, but needs starting values for the parameters.

2. Make a dataset, dat0, which only contains the data with no inhibitor (I=0). Then try
the commands
mm0 <- nls(R ~ Vmax * S / (K + S), start = list(Vmax=3, K=100), data=dat0)
summary(mm0)

22
 Make sure you understand the output.
3. It is not possible to extract standardized residuals from a nls fit, so we have to evaluate
the raw residuals instead. The residualplot with raw residuals against fitted values is
made in the usual way:
plot(fitted(mm0), residuals(mm0))

Does the model seem to be appropriare for the data?

4. Make a scatterplot for the data with no inhibitor. Then try the following commands:
f <- function(S) 2.9811 * S / (35.802 + S)
plot(f, from=0, to=620, add=T)

The first command defines the fitted function, and the second adds a graph to the scat-
terplot.

Now we want to make fit a model to the complete dataset. Consider the association
Vmax · S
R≈ (2)
K1 · (1 + I/K2 ) + S

between inhihitor concentration, substrate concentration and reaction rate. Here Vmax , K1 , and
K2 are parameters to be estimated from the data.

5. Fit the model with nls. You can for example use Vmax = 3, K1 = 100, and K2 = 25 as
starting values. What are the estimated coefficients?

6. Make the scatterplot from question 1 again, with different colours for different inhihitor
concentrations. Add three fitted curves to the plot, one for each inhibitor concentration.
Use the same colours for the graphs as you did for the datapoints.
7. Alternatively, we could use separate Michaelis-Menten models for each inhibitor con-
centration, i.e. use (1), but with three different values of Vmax and K. This model can be
fitted with the command
mm2 <- nls(R ~ Vmax[grp] * S / (K[grp] + S),
start = list(Vmax=c(3,3,3), K=c(100,100,100)), data=inhib)

Notice that starting values are needed for each parameter in the model. Fit the model.
8. The model given by (2) is nested in the model just fitted (explain why), and the models
can be compared with an F-test using anova. If mm1 is the model from question 5, then
try
anova(mm1, mm2)

Is the model given by (2) appropriate for the data?

23
5.5 Logistic regression: Pneumoconiosis among coalminers
Prerequisites: None (although Exercise 5.1 is perhaps useful for understanding the output)
Binary variables are variables with two possible outcomes, for example dead or alive, healthy
or sick, germinated or not. Logistic regression is relevant when the response variable is binary.
The aim is to relate the probabilities of the two outcomes to one or more explanatory variables.
In the simplest case with one continuous covariate, x, and binary response, y, the logistic
regression model assumes that log-odds is a linear function of x. Let p = P(y = 1), and recall
that the odds is defined as P(y = 1)/P(y = 0) = p/(1 − p). Then the assumption is
 
p
log = α + β x. (3)
1− p
Just as in ordinary linear regression (with continuous response) the first aim is to estimate the
parameters α and β .
This exercise is about pneumoconiosis among coalface workers. Data were collected in order
to examine the relationship between exposure time (years) and risk of disease. Severity of
disease was originally rated into three categories, but here we will use only two (normal and
diseased):

Exposure time Normal Diseased

5.8 98 0
15 51 3
21.5 34 9
27.5 35 13
33.5 32 19
39.5 23 15
46 12 16
51.5 4 7

The data are saved in the files coalworker1.xlsx and coalworker1.csv.

Notice how, for each person, it has been observed whether he was diseased or not. This
corresponds to a binary variable. In the following we will consider a model with log-exposure
as covariate, i.e. with log(exposure) as the x-variable in (3).

1. Read the data into R; call the dataset coalworker1. Then try the following commands
and make sure you understand what happens:
total <- normal + diseased
relativeFreq <- diseased/total
logOdds <- log(relativeFreq / (1-relativeFreq))
plot(log(exposure), logOdds)

What does the plot tell you about the appropriateness of the logistic regression model
with log-exposure as explanatory variable?
2. Fit the logistic regression model with the following commands:

24
 status <- matrix(c(diseased,normal), ncol=2)
status
logreg1 <- glm(status~log(exposure), family=binomial)

Notice the option family=binomial. It tells R to interpret values in the matrix status
as outcomes from a binomial distribution.
3. Try the commands
summary(logreg1)
abline(logreg1)

What are the estimated values of α and β ? Does the model seem to fit the data reason-
ably well?

Above, the data was represented by the numbers of diseased and normal for each level of
exposure with 8 observations in total. The data could also be represented with one observation
per person (371 observations in total). The files coalworker2.xlsx and coalworker2.csv
contain the data in this representation with the variable y telling whether the person was
diseased (y=1) or normal (y=0). Apart from this there is as variable exposure2 with the
exposure time.

4. Read the data into R in the new form; call the dataset coalworker2. Make sure
you understand the structure of the new dataset, and make sure you understand that
coalworker1 and coalworker2 contain the same information.
5. When the data is represented as in coalworker2, the logistic regression model is fitted
as follows:
logreg2 <- glm(y ~ log(exposure2), family=binomial, data=coalworker2)
summary(logreg2)

Fit the model and make sure that you get the same estimates as you did with logreg1
before.

The last questions are about the interpretation of the estimates and the models.

6. Consider a person with exposure time equal to 30 years. What is the estimated log-
odds of this person being diseased? What is the estimated probability that the person is
diseased?
Hints: For the estimation of log-odds, remember that the explanatory variable is the
log-transformed exposure time. For the probability you should solve expression (3) for
p.
7. How many years of exposure gives a 50% risk of having developed the disease?
Hint: We are looking for the exposure time corresponding to p = 0.5. What is the cor-
responding value of log-odds? Use the estimates to compute the (logarithmic) exposure
time.

25
8. What happens with the odds if the exposure time is doubled?
Hint: What happens to log-exposure if exposure time is doubled? What is the effect on
log-odds? Which effect on odds does that correpsond to?

26
6 Analysis of variance (ANOVA)

6.1 Oneway analysis of variance: Psoriasis

Prerequisites: Exercises 3.1 and 3.3 (reading and working with datasets)
Psoriasis is an immune-mediated disease that affects the skin. Researchers carried out an
microarray experiment with skin from 37 people in order to examine a potential association
between the disease and a certain gene (IGFL4). For each of the 37 samples the gene express-
sion was measured as an intensity.
There were three different types of skin samples: 15 skin samples were from psoriasis patients
and from a part of the body affected by the disease (psor); 15 samples were from psoriasis
patients but from a part of the body not affected by the disease (psne); and 7 skin samples
were from healthy people (control). The data is saved in the files psoriasis.xlsx and
psoriasis.csv with variables intensity and type. There is also a variable typeNum,
which is not used until the very last question.
The scientific question is whether the gene expression level differs between the three types/groups,
and the natural type of analysis is thus a oneway analysis of variance (ANOVA).

1. (Stripchart, boxplots) Try the following commands to get an overview of the data, and
explain what you learn from the plots:
stripchart(intensity ~ type)
boxplot(intensity ~ type)

2. (Group means) It is well-known that the group means are essential ingredients in the
analysis. Find the groups means with the commands
mean(intensity[type=="healthy"])
mean(intensity[type=="psne"])
mean(intensity[type=="psor"])

Notice that you would rarely do this as part of the analysis, but the values are useful to
understand what happens in the next questions.
3. (Fit of oneway ANOVA, reference group) The model for oneway analysis of variance is
fitted with
oneway1 <- lm(intensity ~ type, data=psoriasis)

Notice that this is analogous to fitting a simple linear regression: intensity is the
response variable; type is an explanatory variable.
Then make a summary of the model:
summary(oneway1)

27
 Some explanation is most likely useful at this point. The “Coefficients” part of the
output has three lines — one per group or parameter in the model.
The lines are denoted (Intercept), denoted typepsne and typepsor. No line is de-
noted typehealthy. This is because R has selected healthy as the reference group,
and compares the other groups to this reference group. More specifically the (Intercept)
line concerns the expected value of the reference group, so the estimate is simple the
mean of the 7 observations from healthy people, whereas the estimate in the typepsne
line is the difference between the mean of the psne observations and the healthy ob-
servations. Similarly for the typepsor line.
All four values in the same line contain information about the estimation of the same
parameter:

• Estimate: The estimated value of the parameter (mean for reference group or
difference between group mean and reference group mean for the other groups)
• Std. Error: The standard error associated with the estimate in the first column
• t value: The t-test statistic for the hypothesis that the corresponding parameter
is zero. Computed as the estimate divided by the standard error.
• Pr(>|t|): The p-value associated with the hypothesis just mentioned.

In the typepsne line, for example, you find the estimate of the difference between the
psor and healthy group, the correponding standard error, as well as information about
the test for the hypothesis that this difference is zero, i.e. that the expected value is the
same for psne observations and healthy observations.
At first glance, the parameterization may seem may seem annoying, but there is a point:
Differences between groups means — not the groups means themselves — are the pri-
mary matters of interest in a oneway ANOVA, and the above version of the model gives
directly the interesting quantities regarding comparison to the reference group.
Below the “Coefficients” part you find, among others, the “Residual standard error”, i.e.
the estimated standard deviation for the observations (not the parametere estimates).
4. (Parameterization in terms of group means) Try
oneway2 <- lm(intensity ~ type -1, data=psoriasis)
summary(oneway2)

and find the three group means as the estimates.

Notice that the residual standard error in oneway1 and oneway2 are the same. This is
because they fit the same model! We say that we have different versions, or parameter-
izations, of the model. The one fit is not more or less correct than the other, but they
are useful for different purposes: oneway1 gives you all the useful information about
comparison to a reference group, whereas oneway2 gives you all the useful information
about the group means themselves.

5. (Interpretation of output) Let us make sure that we understand the output correctly:

• Is there a significant difference between the healthy group and the psne group?

28
 • Is there a significant difference between the healthy group and the psor group?
• Is the p-value for the comparison between the psne and psor groups available in
the current output?
• Which hypothesis is tested in the typepsor line in oneway2? Is this a relevant
hypothesis?
• Why are the standard errors for psne (and psor) not the same in oneway1 and
oneway2?
• Why are the standard errors for healthy and psne not the same in oneway2 even
though they both concern group means?
6. (Test for homogeneity between groups) It is standard to carry out an F-test for the overall
effect of the explanatory variable. To be precise, the hypothesis is that the expected
values are the same in all groups. One way to carry out the test is to fit two models —
the model with as well as the model without the variable in question — and compare
them with the anova function:
noEffect <- lm(intensity ~ 1, data=psoriasis)
anova(noEffect, oneway1)

Notice the way the model with no type effect is fitted: with 1 on the right-hand side of
the ~, meaning that there are no explanatory variables in the model, such that all obser-
vations are assumed to have the same distribution (this is true under the hypothesis).
Is there a significant difference between the three type of skin samples?
7. (Model validation) Try the commands
par(mfrow=c(2,2))
plot(oneway1)

Does the model seem to be appropriate for the data? Check also the stripchart and the
boxplot from question 1 again, and discuss how they can be used to assess the validity
of (some of) the assumptions behind the analysis.
8. (Change of reference group) As default R sorts the groups in alphabetical order and
chooses the first one as the reference group. In our case this happened to be the healthy
group, but this was a coincidence. Luckily, it is easy to change reference group. Try the
following commands and explain what you see:
type
newType <- relevel(type, ref="psor")
newType
newType
oneway3 <- lm(intensity ~ newType, data=psoriasis)
summary(oneway3)

9. (Categorical variables coded with numeric values) Finally, we are going to consider the
variable typeNum. Try the following commands:

29
 typeNum
table(type, typeNum)
reg <- lm(intensity ~ typeNum, data=psoriasis)
summary(reg)

Notice how typeNum gives the same group structure as type, but with numbers instead
of letter names.
Which model is fitted here in reg? Why does the model not make sense in the current
set-up? Hint: Look at the name of the model.
So what should you do if a categorical variable is coded with numeric values? Try the
following commands:

newVariable <- factor(typeNum)

newVariable
newModel <- lm(intensity ~ newVariable -1, data=psoriasis)
summary(newModel)

Explain the difference between the variables typeNum and newVariable. Compare the
estimates from newModel and oneway2.

6.2 Oneway ANOVA: Pillbugs?

Prerequisites: Exercises 3.1, 3.3 (reading and working with datasets), and 6.1 (ANOVA)
An experiment on the effect of different stimuli was carried out with 60 pillbugs. The bugs
were split into three groups: 20 bugs were exposed to strong light, 20 bugs were exposed to
moisture, and 20 bugs were used as controls. For each bug it was registered how many seconds
it used to move six inches. The data are saved in the files pillbugs.xlsx and pillbugs.csv
with variables time and group.

1. Make stripcharts and/or parallell boxplots where you use time as response. Do the
same where you use log(time) as response. Explain why it is more reasonable to use
log(time) than time as the response in a oneway ANOVA.

2. Fit a oneway ANOVA model with log(time) as response, and carry out model valida-
tion. Is the model appropriate?

3. Is the expected value for log-time the same for all three treatment groups?

4. What is the estimated expected log-time for the control group? For the group with light
exposure? For the group with moisture exposure?

5. What is the estimated difference in expected log-time between the group with light
exposure and the control group? What is the interpretation of the exponential of this
value?

6. Does the light exposure have a significant effect on log-time? How about the moisture
exposure?

30
 7. Finally, fit also the oneway ANOVA with time (not log(time)) as response, and carry
out model validation. What do you see?

6.3 Twoway ANOVA: Growth of soybean plants

Prerequisites: Exercises 3.1, 3.3 (reading and working with datasets), and 6.1 (ANOVA)
An experiment with 52 soybean plants was carried out in order to examine the effect of
light and stress on plant growth. There were two different levels of light exposure (low and
moderate), and two different levels of stres (no or yes, where yes means that the plant has
been shaken daily for 20 minutes). The 52 plants were divided into four groups correspons-
ing to the combinations of the light and stress treatments. After a periode the leaf areas was
measured for each plant. The data is saved in the files soybean.xlsx and soybean.csv.

1. It is natural to start with a twoway ANOVA with interaction between stress and light.
This model can be fitted in several different ways, for example:
twowayWithInt <- lm(leafarea ~ stress * light, data=soybean)

Fit the above model, and carry out model validation.

2. Make a summary of the model, and make sure you understand the estimates. In partic-
ular: Find, for each of the four stimuli combinations, the expected value of leafarea.
Hint: Notice how reference levels are selected for each of the two factors (low and
no, respectively), such that the intercept is to be interpreted as the expected value for
this combination of stimuli. This estimate should be “corrected” for the other stimuli
groups.
3. Make an interaction plot as follows:
interaction.plot(stress, light, leafarea)

Make sure you understand what has been plotted. Does the graph indicate an interaction
effect between light and stress stimuli or not?
A test of the hypothesis that there is no interaction is carried out by fitting the model
without interaction (with main effects only), and comparing the two models with anova:

twowayWithoutInt <- lm(leafarea ~ stress + light, data=soybean)

anova(twowayWithoutInt, twowayWithInt)

What is your conclusion regarding interaction? Make a summary of the model without
interaction, and make sure you understand the estimates.
4. Is there a significant difference between the two light exposure levels? Is there a signif-
icant effect of stress?
Hint: For the light exposure, say, make a model with stress as the only explanatory
variable, and compare to the model without interaction. Alternatively, look at the sum-
mary from the model without interaction.

31
6.4 An analysis with categorical as well as quantitative variables: FEV
Prerequisites: Exercises 3.1, 3.3 (reading and working with datasets), 5.1 (linear regression),
and 6.1 (ANOVA)
In Chapter 5 we considered models with only quantitative explanatory variables, whereas in
this chapter we have so far considered categorical explanatory variables only. In this exercise
we would like to use both types.
The primary objective of the analysis is to examine if children exposed to smoking have lower
respiratory function than children who are not exposed, but we will also account for other vari-
ables that may influence respiratory function. The dataset contains information on more than
600 children. The measured outcome of interest is forced expiratory volume (FEV), which is,
essentially, the amount of air an individual can exhale in the first second of a forceful breath.
The data is saved in the files fev.xlsx and fev.csv and include the following variables:
FEV (liters), Age (years), Ht (height, measured in inches), Gender, and Smoke (exposure to
smoking, 0 = no, 1 = yes).

1. Read the data into R, and make a scatterplot matrix of the data: If you have called the
dataset fevdata, then use the command plot(fevdata).
2. Which of the variables in the dataset are quantitative and which variables are categori-
cal? Make factor-type versions of the categorical variables, i.e. define
GenFac <- factor(Gender)
SmokeFac <- factor(Smoke)

3. Fit a model as follows:

fevdata <- transform(fevdata, HtSqr=Ht^2)
fevModel0 <- lm(FEV ~ Age + Ht + HtSqr + GenFac + SmokeFac +
SmokeFac*GenFac + GenFac*Age, data=fevdata)

Make sure that you understand all the terms in the model (including the interactions and
the term HtSqr).

4. Is the model appropriate for the data, or is some transformation of the response needed?

5. Simplify the model (possibly after transformation) as much as possible, i.e., test the
significance of the terms in the model and remove non-significant terms. Remember
that you use should only remove one term (variable or interaction) from the model at a
time.

6. What is your conclusion regarding smoking: Does smoking status influence respiratory
function? If yes, how much? Discuss also he effect of the other variables.

32
7 Principal component analysis
There are several functions in R that can be used for principal component analysis. Below we
will use the princomp function as working horse. Other possibilities include the rda function
from the vegan package, which has advanced options for scaling, among others.

7.1 PCA: Physical measurements of crabs

Prerequisites: Exercise 3.3 (working with datasets)
The data for this exercise come from 200 specimens of a certain type of crabs. The crabs
come in two colours (blue and orange). In the experiment 100 of each type were collected,
50 males and 50 females, and for each of the 200 crabs, five quantities were measured: The
carapace/shell length (CL), carapace/shell width (CW), size of frontal lobe (FL), rear width (RW),
and body depth (BD). The experimenters are interested in characterization of the colour types
(and sexes) in terms of the variables.
The data are available as the dataset crab in the MASS package.

1. We are going to work on the log-measurements. Try the following commands and
explain what happens:
library(MASS) # Load package
head(crabs) # Just looking at the data
logcrabs <- log(crabs[,4:8]) # Dataset with log-values
head(logcrabs) # The log-dataset

group <- crabs$sex : crabs$sp # Group variable

group
plot(logcrabs, col=group)

In particular, does the raw data make it possible to easily distinguish between the
four groups? Notice that black/red/green/blue corresponds to Female-Blue/Female-
Orange/Male-Blue/Male-Orange.
2. A PCA can be carried out with the princomp function. If we use the option cor=T, then
the correlation matrix (rather than the covariance function) is used. This correspods to
a scaling of the variables. Try the following commands and discuss the output:
pca <- princomp(logcrabs, cor=T)
pca
summary(pca)
plot(pca)
loadings(pca)
pca$scores

3. Try the following commands and discuss the graphs. I particular, is it possible to use
the principal components (the scores) to distinguish between the four groups? Which

33
 aspects of crab characteristics relate to the three first components (recall the association
between colours in the graphs, and colour/sex of the crabs from question 1)?

scorData <- data.frame(pca$scores)

plot(scorData)
plot(scorData, col=group)
plot(scorData[,1:3], col=group)
plot(logcrabs[,1], scorData[,1], col=group)

7.2 PCA: Ecological zones along the Doubs River

Prerequisites: Exercise 7.1 (PCA)
As part of a large project on characterization of ecological zones, 11 environmental variables
were measures at 30 sites along the Doubs River. The variables were distance from the source,
i.e. from the start location (das), altitude (alt), slope (pen), mean minimum discharge (deb),
pH of water (pH), concentration of calcium, phosphate, nitrate, ammonium, respectively (dur,
pho, nit, amm), dissolved oxygene (oxy), biological oxygen demand (dbo). See Borcard et al.
(2011) for more details.
The data is saved in the files DoubsEnv.xlsx and DoubsEnv.csv.

1. Read the data into an R dataset called doubs.

2. The first variable in the dataset, das, is not an environmental variable, so we will only
use the remaining variables. Run a PCA on the those data, i.e. on doubs[,-1].

3. How many components are needed to explain 75% and 90%, respectively, of the total
variation? Make a plot of the first two principal components.
4. One of the aims of the study was to identify ecological zones, i.e. groups of sites that are
similar in certain senses. Based on the abundance of different fish species, the 30 sites
were allocated to four clusters. This allocation is given by the variable clus4 below.
clus4 <- c(rep(1,10), rep(2,9), rep(3,3), rep(4,3), rep(3,5))
clus4

For example,the first site in included in cluster 1, whereas the last site is included in
cluster 3.
Make the plot of the two first principal components again, this time with the points
coloured according to the clusters. Do the first two principal components contain infor-
mation about the clusters?

34
8 Matrices

8.1 Matrix operations

Prerequisites: None

1. (Construction of matrices) Try the following commands, and explain what you see:
A <- matrix(c(1,4,2,7,5,3), nrow=2, ncol=3)
A
dim(A)

Notice how the matrix as default is filled column-wise. Try instead

B <- matrix(c(1,4,2,7,5,3), nrow=2, ncol=3, byrow=T)
B
dim(B)

Make the matrices C and D in R, and check that you get the right thing:
   
1 −1 6 4
C= and D =
−1 3 3 2

2. (Indices) Try the following commands and explain what you see:
B[1,2]
B[2,2] <- 6
B
B[1,]
B[,3]
B[,3] <- c(8,9)
B

3. (Matrix computations) Try the following command and explain what you see:
C+D
C-D
C*D

In particular, notice how C*D is not the usual matrix product, but element-wise multi-
plication, which is only rarely relevant. Matrix multiplication goes like this (check that
the result is correct):
C %*% D

4. (Transpose) Try the command t(A). What happens? For the above matrices, calculate
the matrix products A0 A and B0 B, where a prime means transpose.

35
5. (Diagonal elements and matrices) Try the following commands and explain what you
see:
diag(C)
diag(D)
diag(c(2,6,7))
diag(4)

6. (Determinant) The determinant of a square matrix is computed with the det function.
Find the determinants of C and D. Are the matrices regular or singular? What happens
if you write det(A)?
7. (Inverse matrix) The inverse of a regular matrix is computed with solve. Try the com-
mands and explain what you see:
solve(C)
solve(C) %*% C
C %*% solve(C)
solve(D)

8. (Solve matrix equations) Say that you want to solve the equation Cx = y for a known
square matrix C and a known vector y. Then write solve(C,y).
Try the command solve(C, c(1,1)), and check that the answer is correct. Then try
the following commands and compare to the output of solve(C):
solve(C, c(1,0))
solve(C, c(0,1))

9. (Eigen decomposition) Eigen values and eigen vectors for for a square matrix are com-
puted with eigen. Try
eigen(C)
lambda <- eigen(C)$values
lambda
Q <- eigen(C)$vectors
Q

Then lambda contains the eigen values and the columns of Q are the corresponding
eigen vectors. Check that this is correct, i.e. check that CQ1 = λ1 Q1 and CQ2 = λ2 Q2
where λ j is the jth eigen value and Q j is the jth column in Q.
Also compute Q %*% Delta %*% t(Q). What do you get?
Finally, try the following commands and explain why you get complex numbers:
F <- matrix(c(-1,-1,-1,1,2,0,0,1,2), nrow=3, ncol=3)
eigen(F)

36
8.2 Least squares estimates in linear regression?
Prerequisites: Exercises 8.1 (matrices) and 5.1 (linear regression, only for question 3)
Consider data consisting of pairs (x1 , y1 ), . . . , (xn , yn ), and assume that the pairs are indepen-
dent and that the conditional variance of yi given xi is the same for all i. The linear regression
on y of x assumes that the expected value of yi is a linear function of xi :

Eyi = β0 + β1 xi

If y1 , . . . , yn are collected in a vector Y , the so-called the design matrix X is defined by

 
1 x1
1 x2 
X =  .. ..  ,
 
. . 
1 xn

and the unknown parameters, β0 and β1 , are collected in a vector β , then the assumption on
the expected values can be written on matrix form as EY = Xβ .
The least squares estimate of β is given by

β̂ = (X 0 X)−1 X 0Y, (4)

and
1  0 
σ̃ 2 = Y Y − (X β̂ )0 (X β̂ ) (5)
n−1
is an unbiased estimate of the conditional variance of yi given xi . Furthermore, the estimated
variance matrix of β̂ is
Var(β̂ ) = β̂ σ̃ 2 (X 0 X)−1 (6)

Consider the cherry data from Exercise 3.1 and 5.1. In the following you should use volume
as response variable, y, and girth as explanatory variable, x.

1. Make the designmatrix X the cherry data.

2. Use formulas (4)–(6) to calculate β̂ , σ̃ 2 and Var(β̂ ).

3. Fit the linear regression with lm, and recognize that you get the same results. More
specifically:

• Compare the estimates of the intercept/slope obtained from your matrix compu-
tatations and from lm.
• Compare the variance estimate from your matrix computations to the “Residual
standard error” from lm. What is the relation?
• The diagonal elements of Var(β̂ ) are the estimated variances of β0 and β1 , respec-
tively. How do they relate to the standard errors listed in the summary of the lm
fit?

37
9 Functions

9.1 Mathematical functions of a single argument

Prerequisites: None
There are many pre-programmed mathematical functions, for example log and sqrt, but
sometimes you need to examine functions that are not already implemented.

1. (Definition) Consider the function f (x) = 2x2 − 0.9x − 1, and calculate f (0) and f (1.5).
Then try the following commands:
f <- function(x) return(2*x^2 - 0.9*x - 1)
f(0)
f(1.5)

The first line defines the function: It has a name, f, takes an argument, x, and returns a
value that depends on the argument.
2. (Graphs) Try the following commands and see what happens:
plot(f)
plot(f, from=-1, to=2)

Next, define the function g(x) = −4x2 + 0.2x + 4 in R, and try the commands:
plot(g, from=-1, to=2, add=T, col="red")

3. (Minima and maxima) From the graphs we see that f has a mimimum around 0.25. We
can find the actual minimum point with optimize as follows.
optimize(f, interval=c(-1,2))

Notice that we should specify an interval in which R should look for a minimum.
From the graph we also see that g has a maximum around 0. Use optimize to find
this maximum. Hint: You now want to find a maximum rather than minimum. Which
function has its minimum the same place as g has its maximum?
4. (Roots) The function f is zero around −0.5 and 1. The function uniroot can give us
the exact values. Try
uniroot(f, interval=c(-1,2)) ## Gives error message
uniroot(f, interval=c(-2,0))

This gives you the root around −0.5. Find the root which is close to 1.
Then find the roots of g. Finally, find the values of x for which f (x) = g(x). Hint: Which
function is zero if and only if f (x) = g(x)?

38
9.2 The mean as a least squares estimate?
Prerequisites: Exercise 9.1 (functions)
Consider an experiment where the biomass has been measured for 8 random plants grown
under certain conditions. The sample values are denoted y1 , . . . , yn . Assume that we are
interested in an estimate of the population average, denoted µ.
It is well known that the sample mean ȳ is an estimate of the population average. The sample
mean is also the least squares estimate: Consider the sum of squared deviations from µ,
regarded as a function of µ:
f (µ) = (y1 − µ)2 + · · · + (yn − µ)2
This function has its minimum for µ = ȳ.
Consider in the following the sample consisting of
24.7 32.5 22.6 23.9 19.6 21.6 19.9 20.9

1. Make a vector with the sample values, denoted y. Then make a function that takes µ
(mu) as argument and calculates the f (µ). Call the function f.
2. Find the minimum of f with optimize, and compare with mean(y).
3. Recall that the sample standard variance is defined as
1  
s2 = (y1 − ȳ)2 + . . . + (y1 − ȳ)2
n−1
Use the output from optimize to compute this value, and compare to the output from
var(y).
4. Make a plot of f with the following commands, and check it the results from optimize
seems to be correct:
muVal<- 10:40
fVal <- sapply(muVal, f)
plot(muVal, fVal)
plot(muVal, fVal, type="l")

9.3 Mathematical functions of several arguments

Prerequisites: Exercise 9.1 (preferably, on functions)
Consider the function
f (x1 , x2 ) = −5 − 3x2 + 4x2 + x12 − x1 x2 + x22

1. (Definition) We can define the function in two ways. For the first one the arguments
x1 and x2 are considered as two numbers whereas the second one considers x = (x1 , x2 )
as a vector. Try the following commands and make sure you understand the difference
between f1 and f2.

39
 f1 <- function(x1,x2) return(-5-3*x1+4*x2+x1^2-x1*x2+x2^2)
f1(0,0)
f1(1,2)

f2 <- function(x) return(-5-3*x[1]+4*x[2]+x[1]^2-x[1]*x[2]+x[2]^2)

f2(c(0,0))
f2(c(1,2))

2. (Contour plot) For functions of two arguments, it is often convenient to plot the contours
of the functions. A contour plot is a plot with the arguments (x1 and x2 ) at the axes and
with lines/curves corresponding to different values of the function. All points of a given
contour plot give rise to the same value of the function.
Try the following commands and explain what is going on:
x1 <- seq(0,2,length=51)
x1
x2 <- seq(-3,1, length=51)
x2
fVals <- outer(x1,x2,f1) ## Computes f1 in all grid points
dim(fVals)
contour(x1,x2,fVals)

It is not obvious how to illustrate functions that take three or more arguments.
3. (Minimum) The R function optim can be used to find the minimum (or maximum) of a
function. The simplest possible optim command goes as follows:
optim(par=c(0,0), fn=f2)

Compare to the contour plot from the previous question.

Notice that optim requires that the function is defined as f2 above, i.e. that the argument
is considered as a vector rather than as several values. Moreover, initial values should
be provided where the optimization algorithm starts its search.
In general multi-dimensional optimization is a difficult numerical problems, and one
should be careful and examine the properties of the function as much as possible before
applying optim (is there a minimum at all; is it unique; what is a reasonable place to
start the algorithm?). It is often necessary to make use of the extra arguments which
can be supplied to optim, see the help page. An alternative to optim is nlminb. It is
generally slower than optim, but also more reliable for functions with many parameters
(more than 6–8, say).

9.4 Non-linear least squares?

Prerequisites: Exercise 9.3 (functions)
In a plant experiment duckweed plants were treated with different amounts of the herbicide
glyphosate, and the relative growth rate was measured. Eight different non-zero doses were

40
used with two replicates, and there were five control plants (no glyphosate treatment). In
total, this gives 21 observations. The data are saved in the files glyphosate.xlsx and
glyphosate.csv.

1. Make a R dataset with the data. Make two scatterplots: Relative growth rate (y-axis)
against dose (x-axis), and relative growth rate against logarithmic dose.

The logistic function is often used to describe dose-response relations. If y is the relative
growth rate and d is the dose, then
M
y≈
1 + (d/d50 )a
where M > 0, d50 > 0 and a > 0 are parameters that should be estimated from the data. This
is done with least squares: Write the sum of squared deviations between data and function
values as a function of the uknown parameters:
 M 2  M 2
f (M, c50 , a) = y1 − + · · · + yn −
1 + (d1 /d50 )a 1 + (dn /d50 )a
The estimates M̂, ĉ50 and â are the values that make f as small as possible.

2. Convince yourself that the parameter M can be interpreted as the largest possible re-
sponse (when the dose is extremely large), and that d50 can be interpreted as the dose
where the expected response is half of the largest possible, i.e. M/2. Use graphs to find
some loose guesses for M and d50 .
3. Make a function f that takes the parameter vector as arguments. Use optim to find the
estimates. You can use the guesses from the previous question as initial values for M
and d50 , and the value 1, say, for a.
4. Non-linear least squares has been implemented in the nls function. Try the following
command and compare with the minimum you just found:
nls(rgr ~ M/(1+(dose/d50)^a), start=list(M=1.7, d50=exp(12), a=1))

5. Add a graph of the estimated relation onto the scatterplot with log-dose on the x-axis.
Hint: Make a long vector with dose-values, a vector with the corresponding estimated
values of the rekative growth rate, and use the lines function.

9.5 Non-mathematical functions in R?

Prerequisites: Exercises 3.1 (reading data), 3.4 (subsets of datasets), 5.1 (linear regression)
In the previous exercises we have implemented and examined mathematical functions, that is,
functions that take one or more arguments and return a value. However, output of of an R
function could be a graph (just think of the plot function), a model object (think of the lm
function), or something else.
In the following we consider the cherry data from Exercise 3.1 and 5.1 again and use volume
as response variable, y, and girth as explanatory variable, x in a linear regression.

41
 1. Make an R function that takes an observation number, obsNo, as arguments and does
the following (all in the same function):
• Makes a sub-dataset that consists of all observations except the one given by the
argument value obsNo
• Fits the linear regression for the sub-dataset
• Makes s scatterplot for the sub-dataset and adds the fitted regression line
• Returns the estimates coefficients (use coef() on the lm-object)
Your code could look something like this (where the ... should be filled by you):
cherryFct1 <- function(obsNo)
{
myData <- ...
myReg <- lm(...)
... ## Make scatterplot
... ## Add the regression line
return(...)
}

2. Make a function that returns a matrix of dimension 31 times 2. Row i should contain the
estimated coefficients from the regression where observation i is omitted. The function
should take no arguments.
Your code could look something like this:

cherryFct2 <- function()

{
estMat <- matrix(...) # Initialize the matrix
for (i in 1:31) estMat[i,] <- ... # Use cherryFct1
return(estMat)
}
cherryFct2()

References
Borcard, D., Gillet, F., and Legendre, P. (2011). Numerical Ecology with R. Springer, New
York.
Dalgaard, P. (2008). Introductory Statistics with R (2nd ed.). Springer, New York.
Ekstrøm, C. (2012). The R Primer. Taylor & Francis group/CRC Press.
Ekstrøm, C. and Sørensen, H. (2011). Introduction to Statistical Data Analysis for the Life
Sciences. Taylor & Francis group/CRC Press.
Martinussen, T., Skovgaard, I. M., and Sørensen, H. (2012). A First Guide to Statistical
Computations in R. Biofolia.
Venables, W. and Ripley, B. (2002). Modern Applied Statistics with S. Springer, fourth edition.

42