You are on page 1of 248

Introductory Statistics for

Biology Students

JOIN US ON THE INTERNET VIA WWW, GOPHER, FTP OR EMAil:


WWW:
GOPHER:
FTP:
EMAIL:

http://www.thomson.com
gopher.thomson.com
ftp.thomson.com
findit@kiosk.thomson.com

Aservice of

iT\

I(!)P

Introductory Statistics
for Biology Students
Second edition

T.A. Watt

Environment Department,
Wye College, University of London, Wye, Ashforo, uK

la !11

SPRINGER-SCIENCE+BUSINESS MEDIA, B.v.

First edition 1993


Reprinted 1994
Second edition 1997

1993, 1997 T.A. Watt


Originally published by Chapman & Hall in 1997
Typeset in 101I2pt Times by AFS Image Setters Ltd, Glasgow

ISBN 978-0-412-80760-2

ISBN 978-1-4899-3166-5 (eBook)

DOI 10.1007/978-1-4899-3166-5

Apart from any fair dealing for the purposes of research or private study,
or criticism or review, as permitted under the UK Copyright Designs and
Patents Act, 1988, this publication may not be reproduced, stored, or
transmitted, in any form or by any means, without the prior permission in
writing of the publishers, or in the case of reprographic reproduction only
in accordance with the terms of the licences issued by the Copyright
Licensing Agency in the UK, or in accordance with the terms of the
licences issued by the appropriate Reproduction Rights Organization
outside the UK. Enquiries concerning reproduction outside the terms
stated here should be sent to the publishers at the London address printed
on this page.
The publisher makes no representation, express or implied, with regard
to the accuracy of the information contained in this book and cannot
accept any legal responsibility or liability for any errors or omissions that
may be made.
A catalogue record for this book is available from the British Library
Library of Congress Cataloging-in-Publication: 97-'7 3551

8 Printed on permanent acid-free text paper, manufactured in accordance


with ANSlfNISO Z39.48-1992 and ANSI/NISO Z39.48-1984 (Permanence
of Paper).

To Colyear Dawkins

Contents

Preface
Preface to first edition
Note to students
How long is a worm?
1.1 Introduction
1.2 Sampling a population
1.3 The Normal distribution
1.4 Expressing variability
1.5 Exercise

Xl

xiii
xv
1
1
2
3
8

15

Confidence intervals, computers and comparisons


2.1 The importance of confidence intervals
2.2 Introduction to MINITAB
2.3 MINITAB - high-quality graphics
2.4 Comparing two populations - hypothesis testing and
the (-test
2.5 Exercises

18
18
22

Sampling
3.1 First, catch your worm
3.2 Random sampling
3.3 Stratified random sampling
3.4 Systematic sampling
3.5 Further methods of sampling
3.6 Practical problems of sampling
3.7 Exercise

42
42
42
45
49

Planning an experiment
4.1 Replication
4.2 Randomization
4.3 Controls
4.4 Objectives

55

27
27
36

51
52
53

55
56

58
58

Vlll

CONTENTS

4.5
4.6
4.7
4.8
5

Laying out the experiment


Recording data
MINITAB
Exercise

59
61
62
64
67
67
69
71

Accounting for background variation and constructing a model


5.1 Sources of variation
5.2 The model
5.3 Blocking
5.4 Exercise

77

Analysing your results - Is there anything there?


6.1 Wholly randomized design
6.2 Analysis-of-variance table
6.3 Randomized complete block designs
6.4 Which treatment is different?
6.5 MINITAB
6.6 Exercise

79
80
84
87
89
89
94

7.1
7.2
7.3
7.4

Consider the treatments of the field


Confidence intervals
Factorial structure of treatments
MINITAB
Exercises

98
98
101
105
116

Relating one thing to another


8.1 Linear regression
8.2 The model
8.3 Assumptions
8.4 Further example of linear regression
8.5 The importance of plotting observations
8.6 Confidence intervals
8.7 Correlation
8.8 Exercise

121
122
123
129
131
134
142
143
147

What to do when data are skewed, or are ranks or scores, or


are counts in categories
9.1 Introduction
9.2 Mann-Whitney test
9.3 Kruskal-Wallis test
9.4 Friedman's test
9.5 Chi-squared contingency test
9.6 Chi-squared goodness-of-fit test
9.7 Exercises

152
152
153
155
159
161
166
167

CONTENTS

I I

10

Summarizing data from an observational study


10.1 Case study 1
10.2 Case study 2
10.3 Specialized programs for analysing vegetation data

172
173
179
185

11

Your project
11.1 Choosing a topic and a supervisor
11.2 What can happen if you do not seek advice at the start
11.3 General principles of experimental design and execution
11.4 General principles of survey design and execution
1l.5 Health and safety

186
186
187
188
195
200

12

Preparing a report - What's it all for?


12.1 Computers
12.2 Basics
12.3 Illustrating results
12.4 Language
12.5 Exercise

202
202
203
207
209
209

Appendix A

Choosing how to analyse data from a replicated,


randomized experiment

213

Appendix B References and Further reading

215

Appendix C

219

Index

Statistical tables

229

IX

Preface

The structure and format of the second edition remain very similar to
those of the first. However, in response to suggestions from both teachers
of statistics to biologists and from my own students, several important
additions have been made:

Chapter 2 has been greatly expanded to include the independent t-test.


Statistical tables are included in Appendix C.
Exercises with answers are included at the end of most chapters
The section on questionnaire design and analysis has been linked to
chi-squared contingency table analysis and the chi-squared goodness-offit test has also been included.
MINITAB analyses and graphics have been updated to releases 10
and II.
Minor alterations have been made to update and clarify the text and
several short supplements have been added.

The aim continues to be to familiarize students with the most important


aspects of the design and analysis of experiments and surveys, enabling
them to use a computer package with confidence and to treat its output
with caution.
I thank Mark Pollard, Stephanie Harding and Geoff Amor for their help
with the second edition.

Preface to first edition

Students of biological subjects in further and higher education often lack


confidence in their numerical abilities. At the beginning of their course
mature students commonly describe their experience in dealing with
numbers as 'rusty', while those students coming straight from school may
have few science A-levels and often have only a modest pass in GCSE
mathematics. Many individuals from both groups express their surprise a~
having to study statistics as part of their course in biological sciences.
This book aims to calm their fears; to show why biologists need statistics
and to provide a painless way into the subject. MINITAB Statistical
Software removes drudgery, allows easy data checking and exploratory
analysis and enables complex analyses to be undertaken with ease.
[MINIT AB is a registered trademark of Minitab Inc. whose cooperation I
gratefully acknowledge. *]
My own interest in statistics was inspired by the late Colyear Dawkins
whose enthusiasm for communicating the principles of experimental design
and analysis was infectious. I also value discussions with many colleagues
over the years, including: Alan Clewer, John Palmer, Jean Power and
Howard Wright. I am grateful to Brian Carpenter, Keith Kirby, Mark
Macnair, Dominic Recaldin and Eunice Simmons for their constructive
comments on draft material.

*Minitab Inc. is based at 3081 Enterprise Drive, State College, PA 16801-3008, USA.
Telephone: 001-814-238-3280, fax: 001-814-238-4383.

Note to students

Statistics is an essential tool for all life scientists. Unfortunately, a first-year


statistics course at college or university is often seen as an inconvenient
hurdle that must be jumped. Mathematical equations can appear daunting,
so why bother? Because statistics really will be useful and important to you,
as this book aims to show. Furthermore, I aim to cover the important ideas
of statistics without using forbidding mathematical equations.

How long is
a worm?
I would not enter on my list offriends . .. the man who needlessly sets
foot upon a worm.
William Cowper

1.1 INTRODUCTION
School experiments in physics and chemistry often have known answers.
If you don't record a value of 9.8 metres per second per second for 'the
acceleration with which an object falls to the Earth' then you know it must
be because there was something wrong with your equipment or with how
you used it. Similarly, the molar mass of calcium carbonate is 100.09
grams per mole, so any other value would be wrong. The idea that there is
a single clear-cut answer to a question isn't relevant in biology. 'How
heavy is a hedgehog?' or 'What is the length of an earthworm?' do not
have just one answer. However we need to know the answers because, for
example, the weight of a hedgehog in autumn largely determines whether
or not it will survive its over-winter hibernation. The aim of this chapter is
to show how to obtain a useful answer to such questions.
We will simplify life by concentrating on just those earthworms of one
species living in one particular field. Since earthworms are both male and
female at the same time we don't need to specify which sex we wish to
measure. Even so, individuals occur with widely differing lengths. Why is
this? Like all animals, earthworms vary in their genetic makeup - some
inherit a tendency to be short and fat and others to be long and thin.
Earthworms can live for a long time and young worms are likely to be
shorter than old ones.
Those which live in the moister part of the field at the bottom of the
slope might be more active and have a better food supply so they will grow
more quickly and may tend to be longer than those in a less favourable
part of the field. Meanwhile perhaps those living near the footpath along
one side of the field tend to be shorter because they are infested with a
parasite or because they have recently escaped from a tussle with a bird.
How then should we measure and describe the length of worms in this
field?

k-_2___1

H_O_W__L_O_N_G__
IS_A
__W_O_R
__
M_?_________________

L I_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

1.2

SAMPLING A POPULATION

We would like to obtain information about all the worms in the field
because they form the population in which we are interested, but it is
impossible to measure them all. Therefore we measure a few - our sample
- and generalize the results to the whole popUlation. This is only a valid
solution provided that the worms in our sample are representative of the
whole population.
1.2.1 Measuring worms

Let's imagine that we have collected ten worms and propose to measure
them, using a ruler, before returning them to their habitat. First, how do
we make one keep straight and still? We will need to coax it into a straight
line with one end at 0 mm and read off the measurement. We must avoid
stretching it or allowing it to contract too much. This is difficult and will
need practice. No doubt slime or soil will add to the problem of deciding
the correct reading. I suspect that, although we would like to say that 'this
particular worm is 83 mm long' and mean it, we will realize that this is
unlikely to be a very accurate measurement.
Before you can even start to worry about analysing your results,
therefore, you should think carefully about the errors and uncertainties
involved in actually making the measurements. Would you be quite so
careful in measuring worms, for example, on a dripping wet day as when
the sun was shining?
We can but do our best to standardize the process and, in view of the
various uncertainties, decide to measure earthworms only to the nearest
5 mm, i.e. 0.5 cm. Here are the results in cm:
11.5, 10.0, 9.5, 8.0, 12.5, 13.5,9.5, 10.5,9.0,6.0
If we arrange our measurements in order we see at once the enormous
variation in length:
6.0,8.0,9.0,9.5,9.5,10.0, 10.5,11.5, 12.5, 13.5
Some of the worms are twice as long as others. It is sensible to check our
set of observations at this stage. For example, if one of them was 70 cm,
do we remember measuring this giant (worms this size do exist in
Australia), or is it more likely that we have misread 10cm as 70cm?
So how long is a worm? We could work out the average or mean length
of our ten worms and say that it was the length of a 'typical' worm in
our sample. If we add up the lengths and divide the sum by 10 we get
10 cm. Is this, however, the length of a 'typical' worm in the field?
Let's imagine that our sample of ten worms had shown that each worm
measured 10 cm. This is very consistent (not to mention very suspicious -

T_H_E_N_O_R
__
M_A_L_D
__
IS_T_R_I_B_U_T_IO_N
______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _

we should check that someone is not making up the results). The mean
length is 10 cm and it would seem, in common-sense terms, very likely that
the mean length of all the worms in the field (the population) is very close
to this value. If we had sampled 20 worms and they were all correctly
measured as 10 cm, this would be amazingly strong evidence that the
population mean is very close to 10 cm. The more worms we measure
(increasing replication), the more information we have and so the greater
our confidence that the estimate of the mean that we have obtained from
the sample is close to the real population mean which we want to know
but cannot measure directly.
However, the ten worms in our sample were not all the same length.
Here they are again:
6.0,8.0,9.0,9.5,9.5, 10.0, 10.5, 11.5, 12.5, 13.5
The mean length is:
(6.0 + 8.0 + ... + 12.5 + 13.5)/10 = 100/10 = IOcm
With this amount of variation within the ten values, how confident can
we be that the mean length of all the worms in the field is 10 cm? To answer
this question we need a way of expressing the variability of the sample
and of using this as an estimate of the variability of the population, and
for that you need 'statistics'.
1.2.2

How reliable is our sample estimate?

Statistical methods allow you to obtain an estimate of some characteristic


(for example, the mean length of the worms) of a large population (all
the worms in the field) from only a small sample (ten worms). The
population's characteristic is called a parameter, while the estimate of it,
obtained from the sample, is called a statistic. So, if the mean length of all
the worms in the field is actually 9.8 cm, what we actually get from our
measurements is an estimate of this parameter. First we took a sample of
ten worms and calculated a sample mean of 10cm. This is the sample
statistic. We could of course take another sample of ten worms, in which
case we would probably obtain a different estimate, perhaps 9.7 cm, and so
we might go on.
However, we usually only take one sample and so have only one sample
statistic, so it is important to be able to express the reliability of our
measurements in estimating the parameter (the real mean length of all the
worms). The following sections outline how this is done.
1.3 THE NORMAL DISTRIBUTION
If we summarize the information about the lengths of our ten worms in a
graph, we see that there is a tendency for a large number of worms to be

__

1~_3 ~

L-_4__~1

H_O_W__L_O_N_G__IS_A
__W_O_R
__
M_?________________~

L I_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

No.

No.

worms

worms

of

of

30

20
10

a
(a)

6 7 8 910 11121314
Mid-point of class (cm)
Length

o
(b)

67891011121314
Mid-point of class (cm)
Length

Figure 1.1 The number of worms in eaeh size class: (a) few worms (ten in our
sample); (b) many worms.
of medium length and fewer to be either very big or very small. Look at
Figure I.Ia, where the frequency of occurrence (or actual number) of worms
in a range of size classes is shown. If we measure more worms, the number
in the sample increases, and so the shape of the graph becomes less ragged
(Figure 1.1 b). For the whole population of worms in the field (assuming that
the field does not contain any localized abnormality) it is likely to follow
a smooth 'bell-shaped' curve (Figure 1.2), called the Normal distribution.
(Normal is a technical word here - it doesn't just mean 'ordinary'.)
The spread of the curve depends on the natural variability in the
population, being greater if the variability is large, i.e. there are relatively
more values a long way from the centre line. This is a common shape of
distribution for measurements like length and weight. However, not all
populations follow the Normal distribution. For example, counts of the

10 000
No.
of
worms
5000

a
Length (em)

Figure 1.2 The Normal distribution.

T_H_E_N
__
O_R_M_A_L_D
__
IS_T_R_IB_U_T_I_O_N______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _

No.
of
thistles

30
20
10

4
5
2
3
No. of butterflies/thistle

Figure 1.3 A positively skewed distribution.

number of butterflies per thistle might produce mainly zeros (most thistles
lack butterflies), a few ones and twos and the occasional larger number
where the thistle is in a sunny spot and attracts many butterflies. Such a
distribution would be skewed to the right (Figure 1.3).
Although other distributions are used in statistics (Box 1.1) the Normal
distribution is the most important and so it is the one upon which we
concentrate here.
BOX 1.1

Discontinuous distributions
The Normal distribution is a continuous one in that it is used for data
which are measurements and so can take any value on a continuous
scale (e.g. mm, kg). However data may also be counts and so can
only be whole numbers, i.e. they are discontinuous. There are two
main types of distribution which may be used for counts (for more
detail see Rees, 1995):

First, the Binomial distribution. This is relevant when we have only


two mutually exclusive categories. For example, alive and dead
or male and female. We often convert these data into percentages
or proportions, e.g. percentage mortality. The probability of
obtaining say 20 live individuals can be predicted from this
distribution by knowing the total number of individuals and the
probability that anyone of them is alive .
Second, the Poisson distribution. This is relevant for example when
we count microorganisms within a square on a microscope slide.
The probability of obtaining a count of a certain size can be
predicted from this distribution by knowing simply the mean
number of individuals per square.

LI__5__~

~_6__~1 ~I

_________________H_O_W__L_O_N_G__IS_A__W_O_R__M_?________________~
These two distributions assume that individuals are only behaving
according to chance. Sometimes we may suspect that there is a
biological reason for individuals (for example microorganisms on a
slide) to be arranged in a pattern rather than just according to
chance. We then compare our observations with the predictions from
these models (see section 9.6) to discover whether there is evidence
to reject the idea that chance is the only factor.
However, if data are counts and fairly large numbers it is often
reasonable and convenient to treat them as if they followed an
approximately Normal distribution. For example, in Chapters 5 and
6 we discuss an experiment in which the numbers of spiders in field
margins managed in different ways are compared with a method
which uses the Normal distribution.

Instead of having a scale of frequency of worms up the side of our


graph, we could have one of probability (how likely it is that a worm's
length will fall into a particular length class). This probability scale would
go from 0 to 1. If all worms in the field were 10 cm long, the probability
that a worm is in the size class 9.5-10.4 would be 1. This represents 100%
certainty. However, in reality the worms vary in length so the different size
classes have different probabilities, with the highest being for the central
size class as in the Normal distribution curve in Figure 1.2. The convenient
aspect of this curve is that mathematicians have worked out how to
describe it as an equation. The equation takes into account the mean (the
central, highest point) and also its spread through an estimate of
variability of the values from one individual to another - the standard
deviation. We will see how to calculate the standard deviation from a
sample of observations a few pages further on. For the moment it is
enough to know that the larger the variability of a population the larger is
the standard deviation.
If the equation of a curve is known, a mathematical technique
(integration) can be used to work out the area underneath the curve, which
represents the whole population. In addition, integration can be used to

8%Or

~
~ 0.08

--I
6

10

Length (em)
Figure 1.4 The probability of a worm being between 6 em and 7 em in length.

__

~_____________T_H__E_N_O_R_M
__A_L_D
__
IS_T_R_I_B_U_T_IO_N
______________~I ~I 7__~

9.5

em

10.5

10

Standard deviation

= 0.5 em

Figure 1.5 A consistent population.

find out the area between any two length values. The ratio of such an area
to the total area under the curve (representing the whole population) gives
you the probability that you will encounter worms of a particular length.
For example, if we wanted to know the probability of a worm being
between 6cm and 7 cm long we would find that it is 0.08 (or only 8% of the
population, Figure 1.4).
Fortunately you do not need to know anything about either the equation
of the curve or about integration to be able to answer such questions.
The information needed is already available in tables (statistical tables)
and most statistically oriented computer programs already incorporate
them, so that we don't even need to consult the tables themselves very
often. However, it is useful to remember that the probability of a worm's
length being in the range between the two values where the curve changes
direction (points of inflection) is about 0.68 (68%) (Figure 1.5). These
values are the mean plus or minus one standard deviation. For example, if
the mean is 10 cm and the standard deviation was calculated to be 0.5 cm
we would know that 68% of worms would have lengths between 9.5 cm
and 10.5 cm. If the population was more variable and had a standard
deviation of 2 cm we would expect 68% of worms to have lengths between
8 cm and 12 cm (Figure 1.6).
How in practice therefore do we summarize the variability of a
population in a way that will tell us something about the spread of the
curve, so that we can work out the probability of finding worms of
different lengths?

10

12

Standard deviation

= 2.0 em

Figure 1.6 A variable population.

em

~_8__~1

H_O_W
__
L_O_N_G__
IS_A
__W_O_R
__
M_?________________~

L I_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Table 1.1

The sum of the differences for the worms data

a
Observation (cm)

b
Distance from the
mean value (cm)

c
Difference
(positive) (cm)

6.0
8.0
9.0
9.5
9.5
10.0
10.5
11.5
12.5
13.5

-4.0
-2.0
-1.0
-0.5
-0.5
0.0
0.5
1.5
2.5
3.5

4.0
2.0
1.0
0.5
0.5
0.0
0.5
1.5
2.5
3.5

Sum

0.0

16.0

1.4 EXPRESSING VARIABILITY


1.4.1 First step - the sum of the differences
A first attempt to describe the variability of our initial sample of ten
worms is to note how far away each value is from the mean and then to
add up these distances or differences (Table 1.1). It doesn't matter whether
the values are greater or less than the mean, only how far away they are,
so all the differences are treated as positive values in this exercise. The
further away the values are from the mean, the greater their sum.
Adding up the values in column b always gives a sum of zero, because
by definition the mean is the value at which the sum of the distances below
it equals the sum of those above it. However, when the differences are all
treated as positive (often called absolute values) their sum is, in this
instance, 16.
If five of the worms in our sample were each 9 cm long and the other five

Figure 1.7 Ten differences of I cm.

10

em

11

12

E_X_P_R_E_SS_I_N_G_V_A
__R_IA_B_I_L_I_T_Y______________~IIL _ _9__~

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

were 11 cm long, this method would give a mean of 10 cm. Intuitively we


might consider such a set of observations less variable. What would be the
sum of the differences (Figure 1.7)? So, the less the variability in the
observations, the smaller the sum of the differences. This seems a
promising way of summarizing variability but we can improve upon it.
1.4.2

Second step - the sum of the squared differences

By our 'sum of the differences' method, a sample containing just two


worms, one of 6cm and the other of 14cm in length, is no more variable
than one which has four worms of 9 cm and four worms of 11 cm. Both
have a sum of differences of 8 (Figure 1.8). But, from looking at the values,
the first sample would seem to be more variable than the second.
In the second sample we have many observations which all fall only a
small distance from the mean, which is evidence of consistency, whereas in
the first sample the results suggest inconsistency; there are only a few
values and these fall far away from the mean. A neat way of taking this
consistency into account is to square each difference (to multiply it by
itself) before adding the resulting values up (Table 1.2). This means that
extreme values will contribute more to the total sum.
The sum of the squared differences between the observations and their
mean is 8 for the consistent sample (sample 2) but 32 for the inconsistent

(a)

10
em

12

14

(b)

10
em

11

Figure 1.8 Sum of differences = 8: (a) variable; (b) consistent.

10

I I

HOW LONG IS A WORM?


Table 1.2 The sum of the squared differences for two other samples
of worms

Sample 1

Sample 2

Measured
values (cm)

Difference
(all positive)

Difference
squared

6
14

4
4

16
16

Sum

32

9
9
9
9
II
11
11
11

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

Sum

one (sample 1) so this revised method gives a much better picture of the
variability present than just the sum of the differences used earlier. The
phrase 'the sum of the squared differences between the observations and
their mean' is usually abbreviated to 'the sum-of-squares'. However, this
abbreviation sometimes leads to it being miscalculated. The 'sum-ofsquares' does not mean the sum of the squares of each observation, as the
following example should make clear.
The sum-of-squares for a sample of three worms measuring 2, 3 and
4 em respectively is not (2 x 2) + (3 x 3) + (4 x 4) = 29. Instead, the mean
of the values 2, 3 and 4 is 3, and so the sum-of-squares is:
(2 - 3)2 + (3 - 3)2 + (4 -

3i = 12 + 02 + 12 = 2

1.4.3 Third step - the variance


So far so good, but we have not allowed for the number of worms which
we have in our sample in our calculation. In general we can refer to this as
the number of observations (each worm measured counts as an
observation). The more worms we measure the greater will be the 'sum-ofsquares' simply because we are adding up more squared differences. To
take account of this fact and so to provide a standard estimate of
variability (called the variance), we divide the sum-of-squares by the
number of observations minus one. In this case, we divide the sum-ofsquares by 9 if we have ten worms or by 7 if we have eight worms. Why do
we divide by one less than the number of observations rather than by the
number itself?

E_X_P_R_E_S_SI_N_G__
V_A_R_IA_B_I_L_IT_Y______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _

If we have a sample of only one worm and it is 9 cm long, we have an


estimate of the population mean (9 cm), but we have no information about
how variable or unreliable this estimate is. As soon as we select a second
worm (of length 10cm) we have two values (9 and 10cm). We can revise
our estimate of the population mean to (9 + 10)/2 = 9.5 cm and we have
one piece of information about its variability. This is the difference
between the lengths of the two worms. With three worms we can revise the
estimate of the mean again and we have two independent pieces of
information about the variability:
worm lengths 9.0, 10.0, 11.0

new mean = 10.0

The two independent pieces of information about variability are


(10.0 - 9.0) and (11.0 - 10.0). The third difference (11.0 - 9.0) is not
independent because both values already feature as part of the other
differences.
It is common to refer generally to the number of observations in a
sample as n. So, in our main example, where n = 10, we divide the sum-ofsquares of our sample by n - 1 (which is 9) to obtain a variance. The
number n - 1 is referred to as the degrees of freedom because it refers to
the number of independent items of information about variability that we
can derive from the sample.
To reinforce the above justification for using n - 1 to calculate the
variance, statisticians have shown (using many lines of algebra) that this
standard estimate of variability, the variance, calculated from a sample by
dividing the sum-of-squares by n - 1 is unbiased. This means that if we
took many samples and calculated for each a variance value in this way,
then their average would be close to the real population variance; it would
be neither consistently too big nor consistently too small.

1.4.4 Fourth step - the standard deviation

We have now obtained variances (measures of variability within a


sample) which can be compared with one another, but they are in units
which are different from those used in the initial measurement. If you go
back to section 1.4.2 you will see that we took the differences in length
(cm) and then squared them. Thus, in the case of our example, the units
of variance are in cm2 We wouldn't naturally think of the variability of
the length of a worm in such units. So we take the square root of the
variance to return to cm. The result is a statistic called the standard
deviation. This was referred to earlier in connection with the Normal
distribution curve.
We can now work out the standard deviation from our original data
set. The sum-of-squares is 42.5 (Table 1.3), so the variance is 42.5 -:- 9 =

11

12

II

HOW LONG IS A WORM?

~----------------------------------------------------~

Table 1.3 The sum-of-squares for the worms data


Observation

Difference
(all positive)

Difference
squared

6.0
8.0
9.0
9.5
9.5
10.0
10.5
11.5
12.5
13.5

4.0
2.0
1.0
0.5
0.5
0.0
0.5
1.5
2.5
3.5

16.0
4.0
1.0
0.25
0.25
0.0
0.25
2.25
6.25
12.25

Sum

16.0

42.5

4.722 and the standard deviation, which is the square root of this, is
2.173. This value helps us to judge the extent to which worm length varies
from one individual to another, but its usefulness becomes clearer when
we put this piece of information together with our knowledge of the
Normal distribution (Figure 1.5). This told us that 68% of the worms in
the field will have lengths between:
the mean + the standard deviation
In our sample of ten worms that is between
10cm - 2.173cm and lOcm +2.173cm
which works out to be between
7.827cmand 12.173cm
This sounds amazingly precise - it implies that we can measure the
length of a worm to the nearest 0.01 of a millimetre. Since we only
measured our worms to the nearest 0.5 cm (5 mm) it is better to express
this result as:
68% of the worms in the field have lengths between 7.8 cm and
12.2cm.
If the worms we measured still had a mean length of 10 cm but had been
much less variable, for example, with lengths mainly in the range from 9
to 11 cm, the standard deviation would be much less. Try working it out
for these values:
8.5,9.0,9.5,9.5, 10.0, 10.0, 10.5, 10.5, 11.0, 11.5
(You should get 0.913.)

~______________E_X_P_R_E_S_SI_N_G__V_A_R_IA_B_I_L_IT_Y______________~I
1.4.5

Fifth step - the sampling distribution of the mean

Measuring ten worms gives us an estimate of their mean length and of


their variability but we have already implied that if we took a second
sample of wor11l6 the estimate of the mean would be slightly different.
Thus, in this section we look at how the mean varies from one sample to
another. This is referred to as a sampling distribution.
To illustrate what this means imagine that the population that we wish
to sample consists of just six worms and that we are going to estimate the
mean length of a worm by sampling just two of them at a time (the worms
being returned to the population after each sampling). We could find the
following possible outcomes:
Population (mean = 10)
Lengths 8 9 10 10 11

12

Sample

ABC

Worm 1
Worm 2

8 8 8 8 9
10 10 11 12 10

9
10

9
11

9
12

10
10

10
11

10
12

10
11

10
12

11
12

For convenience, where the two worms in a sample differ in length, the
shorter has been called worm 1. There are 15 possible samples (A to 0),
because that is the maximum number of different combinations of pairs of
worms that you can get from six worms. Because two of the worms share
the same length (10 cm) some samples will give the same mean (see samples
Band C, for example).
We can see that, by chance, our sample might have provided us with
an estimate of the mean that was rather extreme: sample A would give 8.5,
while sample 0 would give 11.5, compared with a population mean of
10.0. The means of the 15 possible samples can be summarized as in
Table 1.4.
Table 1.4

Sampling two worms out of a population of six

Sample
numbers

Number of
samples

Sample
mean

Number of
samples x sample
mean

A
B,C
D,F,G
E, H,J
I,K,M
L,N
0

I
2
3
3
3
2
1

8.5
9.0
9.5
10.0
10.5
11.0
11.5

8.5
18.0
28.5
30.0
31.5
22.0
1.5

Total
Mean

15

150
10.0=150115

13

14

H_O_W_L_O_N_G_IS_A_W_O_R_M_?_ _ _ _ _ _ _---..J

L I_ _ _ _ _ _ _ _

No.
of
samples

3
2

o
9

9.5

10

10.5

11

Sample mean (em)

Figure 1.9 Means from the 15 possible samples each of two worms.
The second column in this table shows that if we take a sample of two
worms we have a 1 in 15 chance of selecting sample A and so getting an
estimated mean of 8.5. However, we have a 3 in 15 (or 1 in 5) chance of
having a sample with a mean of 9.5 (samples D, F or G) or of 10.0 (E, H
or J) or of 10.5 (I, K or M). This is the sampling distribution of the sample
mean. It shows us that we are more likely to obtain a sample with a mean
close to the true population mean than we are to obtain one with a mean
far away from the population mean.
Also, the distribution is symmetrical (Figure 1.9) and follows the
Normal curve so that if we take a series of such samples we will obtain an
unbiased estimate of the population mean. The mean of all 15 samples is
of course the population mean (10.0, bottom of right-hand column of the
table) because all worms in the population have been sampled equally,
with each worm occurring once with each of the other five worms in the
population.

1.4.6 Sixth step - the standard error of the mean


In reality, we will only have one sample mean and we will not know the
true population mean because that is one of the things we usually want to
estimate. Therefore we need a way of expressing the variability from one
sample mean to another. For example, it is very common in experiments
(Chapter 4) to want to compare say the mean length of ten worms fed on
one type of food with the mean length of ten worms fed on a different type
of food.
In earlier steps we showed how the standard deviation could be used to
describe the variability from one worm to another. The standard deviation
is obtained by taking the square root of the variance. It describes the
dispersion of our observations. To estimate the variability of a mean a

~________________E_X_ER_C_I_SE________________~I
slightly different statistic is used - the standard error. In contrast to the
standard deviation the standard error describes our uncertainty about the
mean length of a worm.
This uncertainty is caused by the fact that if we took several samples,
as we have illustrated above, they may each provide a different estimate of
the mean. To obtain the standard error (of the mean) we divide the
variance by the sample size before taking the square root. This can be
expressed in three equivalent ways, but mathematically they are all the
same and give the same result:
variance
sample size

Jvariance

standard deviation

Jsample size

Jsample size

The standard error gets smaller as the sample size increases because
the standard deviation is divided by the square root of the sample size.
So, if our sample size was 16 the standard error of the mean would be
the standard deviation divided by 4. If we took a sample of 25 we would
divide the standard deviation by 5 and so on. This is because the lengths
of occasional tiny or huge worms have much less influence on the mean
when many worms are measured than if only a few are measured. Thus,
the mean of a larger sample is more likely to be closer to the population
mean.
Back in section 1.4.4 the standard deviation of the sample of ten worms
was found to be 2.173 cm. The standard error of the mean will be less
because we divide the standard deviation by the square root of 10 (which
is 3.162):
2.173/3.162 = 0.69

(we should perhaps round this to 0.7 em)

Because the distribution of sample means follows the Normal curve then
there is a 68% chance that the range between the estimated mean plus
one standard error and the estimated mean minus one standard error will
contain the population mean. So if the mean is 10.0 cm and the standard
error of the mean is 0.7 cm, there is a 68% chance that the range between
10 cm plus 0.7 cm and lOcm minus 0.7 cm, i.e. from 9.3 cm to 10.7 cm,
contains the population mean.

1.5 EXERCISE
Histogram, sample mean, standard deviation and standard error
Here are the circumferences of a random sample of ten apples taken from
each of two contrasting varieties:

15

16

I LI________________H_O_W__L_O_N_G__IS_A__W_O_R__M_?______________~
Apple
number

Tyler's
Kemal
(cm)

Red
Charles
Ross
(cm)

22.0
24.5
25.5
27.5
22.5
27.5
24.0
26.5
23.5
25.0

18.3
18.4
20.2
22.0
17.5
18.1
17.6
16.8
18.8
18.9

2
3

5
6
7
8

9
10
n
df
mean
SD

SE

Plot a histogram of the data from each variety (as in Figure l.1a).
Choose class ranges which suitably summarize the variability (i.e. avoid
putting all the observations into one class or having a number of classes
almost equal to the number of apples).
Copy out the table. Use your calculator to complete the entries at the
bottom of each column in your table:
Switch on your calculator and put it into statistical mode.
Clear out the statistical memory.
Enter the ten values.
(b)

(a)

No.

No.

of
apples 4

of
apples 4

.-

23 25 27
Mid-point of class (cm)
Circumference

17 19

21 23

Mid-point of class (cm)


Circumference

Figure 1.10 The number of apples in each size class: (a) Tyler's Kemal; (b) Red
Charles Ross.

E_X_E_R_C_I_SE______________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Check that n = 10. Write this number in your table and follow it with
the number of degrees of freedom (df) (n - 1).
Obtain the sample mean and standard deviation (SD) from the
appropriate keys xand O"n_1 or s and enter them into your table.
Calculate the standard error (SE) by: squaring the standard deviation,
dividing the result by n and taking the square root. Enter the value into
your table.
Answer

The histograms are shown in Figure l.1O. The values you should have
obtained in your table are:

n
df
mean
SD
SE

Tyler's
Kernal

Red
Charles
Ross

10
9
24.85
1.930
0.610

10
9
18.66
1.492
0.472

17

Confidence intervals, computers


and comparisons

2.1

THE IMPORTANCE OF CONFIDENCE INTERVALS

Any estimate that we make of the mean value of a population mean should
be accompanied by an estimate of its variability - the standard error. As
we have seen in Chapter 1 this can be used to say within what range of
values there is a 68% chance that the population mean will occur. This
range is called a 68% confidence interval. However, this is a bit weak; we
usually want to be rather more confident than only 68%. The greater the
chance of our range containing the population mean, the wider the
confidence interval must be.
Think about it: if we want to be absolutely certain that the range
contains the population mean we need a 100% confidence interval and this
must be very wide indeed; it must embrace the whole of the distribution
curve, not just the middle part. It is standard practice to calculate a 95%
confidence interval (excluding 2.5% at each end) which is usually taken as
a good compromise. The 68% confidence interval was obtained by
multiplying the standard error by 1; therefore to calculate the 95%
confidence interval we need a 'multiplier' of greater than 1. Statistical
tables, summarizing the Normal distribution curve, tell us that we need to
multiply by 1.96 for a 95% confidence interval (see Table C.l, where the
'multiplier' = z = 1.96 for p = 0.975, i.e. 97.5% of the area lies below this
point and so 2.5% lies above it). There is however just one further
modification to be made. The Normal distribution assumes that we have a
large number of observations (usually taken to be more than 30) in our
sample. If, as in our case, we have only ten, our knowledge is more limited
so we need to take this fact into account.
Having less information leads to more uncertainty and so to a wider
confidence interval. A distribution which is appropriate for small samples
is the t distribution. Again this exists in table form, so from the tables we
can find out the appropriate value of t with which to multiply the standard

IM
__P_O_R_T_A_N_C_E_O_F__C_O_N_F_ID
__
EN
__
C_E_IN
__T_E_R_V_A_L_S________~I

L -________

Table 2.1 Values of t for various numbers of observations and confidence intervals
Number of worms Degrees of freedom
10
20
30

19
29

95% confidence

99% confidence

2.26
2.09
2.04'

3.25
2.84
2.76

This is very close to the value of 1.96 which comes from the Normal distribution. For most
practical purposes, if our sample contains 30 or more individuals a 95% confidence interval
is given by the mean twice its standard error.

error m order to obtain a confidence interval. Our equation for the


confidence interval is:
confidence interval = mean (t x SE mean)
The value of t increases as the number of observations decreases - there
is more uncertainty - and, as with the Normal distribution, the value of t
increases as the amount of confidence we require increases. So, for
example, with 30 worms in the sample and only 95% confidence required
t = 2.04, but if 99% confidence is needed and the sample is only ten worms
t increases to 3.25 (Table 2.1).
2.1.1 So what's so important about a confidence interval?
The statement that 'the mean length was 94 cm' has some value, but not
a great deal. It indicates that we are probably not referring to whales, but
it doesn't give us any idea about the variability around that mean. It is this
variation which is essential information if we are to take decisions.
For example, imagine that you are the owner of a factory which
prepares and packages a special herbal tea which is in great demand. The
distinctive element in the product is an extract from a fruit which is
produced on a particular species of shrub in the local forest and which is
harvested by hand. This forest is remote and its species composition is not
well known. Market research indicates that the demand for the product
is likely to continue to grow and you would happily invest several million
pounds in expanding the factory, but you would only do this if you were
confident that there was enough of the shrub present within a certain
distance from the factory to keep a new and larger factory fully supplied
in a sustainable way. If you decide to invest the money and then run out of
fruit you will go bankrupt and the local people will lose their jobs. The
way to gain information is to take a sample (Chapter 3).
The results from the sample survey are that there is a mean of 20 bushes
per hectare (ha). If every single bush (i.e. the population) had been found
we would know that 20 bushes/ha was the true value. However, as we
have only sampled some of the popUlation we have only an estimate of the

19

20

II

CONFIDENCE INTERVALS

~----------------------------------------------------~

12

14

16

18
20
22
Sample mean

24

26

28

Figure 2.1 There is a 95% chance that this range includes the mean number of
bushes per hectare over the entire forest.

mean and we know that there is some uncertainty about how close the true
population mean is to this estimate. If we calculate a confidence interval,
we can quantify this uncertainty. The 95% confidence interval might be
from 15 to 25 bushes per hectare (Figure 2.1). So we are 95% confident
that the population mean lies in this range. If we have worked out that, to
make full use of the factory's capacity, we must have at least 15 bushes/
ha we might decide that we are happy to proceed since there is only a small
chance (2.5%, which is the area in the left-hand tail of the distribution)
that the population mean will be less than this.
However, we may decide that a 2.5% chance of bankruptcy from this
source of risk is too much and that we are only prepared to take a 0.5%
risk. Then a 99% confidence interval for the population mean should be
used instead. This gives the range which is 99% likely to include the
population mean. This might be between 12 and 28 bushes. Then, there is
a 0.5% chance of the population mean being less than 12 bushes per
hectare (the 0.5% chance of there being more than 28 bushes per hectare
does not influence the decision). This confidence interval now includes
uneconomic densities of shrubs (12 to 14.9). Thus we may decide not to
expand the factory.
In the real world there are likely to be areas of the forest where the bushes
are dense and other areas where they are sparse. The mean of 20 bushes
per hectare might be built up from a few areas at a great distance from the
factory where there are many shrubs and areas nearby where there are few.
This would be vital additional information since it would affect the cost of
collecting the fruit. In Chapter 3 we look at how sampling methods can be
made more sophisticated to take account of this type of variation.
2.1.2 Consolidation of the basic ideas
You have now encountered some very important concepts. Don't be
surprised if you don't understand or remember them after just one reading.

IMPORTANCE OF CONFIDENCE INTERVALS

L -____________________________________________________

II

Most people find that they need to come across these ideas many times
before they are comfortable with them. The best way to understand the
subject is to have some observations of your own which you wish to
summarize, then you are motivated to put these methods to good use. In
the mean time, use the flow chart below to obtain the standard error from
the data which follow it. Check that you understand what is going on
and why. If you are puzzled, refer back to the appropriate section and
re-read it. Reading about statistical methods cannot be done at the same
fast pace as reading a light novel or a science fiction book - but it will
repay your attention.
FLOW CHART

n observations
y,. y2. Yn

~
sum of squares (n - 1)

J,

variance of the
observations = Vy

('Ly)/n

------t)

=Y

sum of squares =

'L(Y-W

Vy/n

variance of the
mean=Vy

J,

J,

KY

..ffy

J,

J,

standard deviation = SOy

mean

--4)

SOy

.;n

standard error of the


mean = SEy

The symbol 'L means 'add up all the values of the item which follows'. For example. all n
values of y.
The bar above y indicates that this is the mean of all the values of y.

2.1.3 Using the statistical mode on your calculator

Many calculators have a statistical mode. This allows us to short-circuit


the above pathway by entering the data and obtaining the standard
deviation directly by pressing the appropriate key, usually either a button
marked s or one marked:
or
(depending on the make of calculator).
Tryout your calculator using this sample from a medical doctor's
records of the ages (years) at which 12 people developed the symptoms of
a particular disease:

27,32,36,38,45,48, 55,61,

68,71,7~78

21

22

I LI________C_O_N_F_ID_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ _---'
To calculate the mean and a 95% confidence interval go through the
following procedure. Put your calculator into statistical mode and then
enter the observations. Press the button marked n (the number of
observations) to check that you have, in fact, entered 12 values. Press the
button marked x to obtain the mean.
To obtain the standard error easily we then square the standard deviation
(to get the variance), divide by n (the number of observations) and take the
square root. Multiply the standard error by the value of t for 11 degrees of
freedom (n - 1) from the tables (95%) which is 2.201. Add the result to, and
subtract it from, the mean to give the required confidence interval.
Here are the results: mean = 52.8, standard deviation = 17.6986,
standard error = 5.1092, t = 2.201, confidence interval = 52.8 11.245, i.e.
from 41.5 to 64.0 years. We conclude that there is a 95% chance of the
population mean lying within this range. What would be the range for a
99% confidence interval? Remember however that the population consists
of the patients of one particular doctor. It would be misleading to
generalize from these results to those patients on the list of a doctor in a
different part of the country. No matter how good the statistical summary,
your conclusions must always be reviewed in the light of how the
observations were originally collected.
2.2 INTRODUCTION TO MINITAB
We can improve on the calculator by using a computer package. The
MINITAB package is available on PCs and is very popular. The latest
versions (releases 10 and 11) allow you to select commands from menus
and have excellent graphics. They are very easy to use and provide quick
and useful ways of looking at the sample of observations or set of data we
have collected and of summarizing it. We will see how MINITAB could
be used to deal with our original sample of the lengths of ten worms. We
type the ten values directly into column 1 of the worksheet and name the
column 'length'. Then we can ask for the data to be printed (File, Display
Data):
length

6.0

8.0

9.0

9.5

9.5

10.0

10.5

11.5

12.5

13.5

Because there are only a few values they are printed across the page
instead of in a column, to save space.
We can see the position of each observation on a scale of length (Graph,
Character Graphs, Dotplot):

..

- - + - - - - + - - - + - - - - - + - - - - + - - - + ---length
6.0
7.5
9.0
10.5
12.0
13.5

I_N_T_R_O_D_V_C_T_I_O_N_T_O__M_I_N_IT_A_B______________~I

L -_____________

This shows us the length of each worm very clearly.


We can group the observations into classes (or groups) by length
(Graph, Character Graphs, Histogram):
Histogram of length
Midpoint Count
6
1
7
0

9
10
11
12
13
14

1
3
1
1
1
1

N=10

MINITAB produces histograms on their side (you may be more used to


seeing them with the midpoints of the classes running along the bottom).
Each star or asterisk represents one worm. The class with a midpoint of 8
contains any worm with a length of from 7.5 to 8.4 cm. In this case there
is only one. There are no worms in our sample with lengths in the range
from 6.5 to 7.4 cm, but there are three in the range from 9.5 to 10.4 cm.
A more detailed version of the above histogram is provided by the
stem-and-Ieafplot (Graph, Character Graphs, Stem-and-Leaf):
Stem-and-leaf of length
Leaf unit = 0 .10
1
6
7
1
2
8
5
9
5
10
11
3
2
12
13
1

N=10
0
0
055
05
5
5
5

This has three columns of numbers. The central column is called the stem
and the numbers to its right are the leaves. There is one leaf for each
observation. Here it represents the value of the number after the decimal
point in the observation. There can be many leaves on each stem. The stem
here represents the whole number part of an observation, before the
decimal point. The top row of the middle and right-hand columns shows
that there is one observation with a stem of 6 and a leaf of 0 (this is
equivalent to a worm 6.0 cm in length). The fourth row shows that there is
one worm 9.0cm long and two worms are each 9.5 cm long.
The column of numbers on the left is the cumulative number of worms
up to the midpoint of the histogram. Thus, starting from the top these are
1 (one worm in first class), 1 again (no more worms added from second
class), 2 (one more worm added -length 8.0cm) and 5 (three more worms

23

24

I I

CONFIDENCE INTERVALS

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

added in fourth class). The same process goes on from the bottom upwards
as well, so that we can see that there are three worms whose lengths are
greater than or equal to 11.5 cm and five worms with lengths of at least
10 cm. In our sample the observations happen to fall into two groups of
five counting from both the high end and from the low end, but the
distributions in small samples are not always quite so well balanced.
The advantage of the stem-and-Ieaf plot is that, in addition to giving
us the shape of the distribution, the actual values are retained. The plot
would also suggest to us (if we did not know already) that the lengths were
recorded only to the nearest 0.5 cm which might affect how we interpret
the results or the instructions we give to anyone who wanted to repeat the
survey.
The following instruction produces many summary statistics (Stat, Basic
Statistics, Descriptive Statistics):
length

N
10

Mean
10.000

Median
9.750

TrMean
10.063

length

Min
6.00

Max
13.500

Q1
8.750

Q3
11. 750

StDev
2.173

SEMean
0.687

N is the number of observations (the same as n used in the text previously)


and is followed by the Mean. Further on there are minimum (Min) and
maximum (Max) values. Notice that, in addition, MINITAB gives a
'trimmed mean' (TrMean). This excludes the smallest 5% and largest 5%
of the values (rounded to the nearest whole number) before calculating the
mean. Here, since 5% of a sample of ten is 0.5, one value has been
excluded from each end. If there were a big difference between the mean
and the trimmed mean it would suggest that one very large (or very small)
value was greatly affecting the mean value. You might at this stage wish
to review such values, to double-check that they have not been wrongly
recorded or wrongly entered. Perhaps they may have some other
characteristics which mean that they should be excluded from the sample.
You might find on checking, for example, that the very long worm was, in
fact, a different species.
The values headed StDev and SEMean are the standard deviation and
the standard error of the mean. These are the same as we obtained by
laborious calculation earlier. That leaves us with three other statistics:
Median, QI and Q3. These belong to a slightly different way of
summarizing data which is particularly useful when observations do not
follow a Normal distribution. In such circumstances the mean is still an
estimate of the average value but it may be a misleading one to use. For
example, in most organizations there is a hierarchy of salaries from the
most junior employee to the most senior one. Usually there are many more
people earning relatively low salaries than high ones but if we calculate
the mean salary it is greatly affected by the very high salaries of just a few

I_N_T_R_O_D_U_C_T_I_O_N_T_O__
M_I_N_IT_A_B______________~I

L -_____________

individuals. The mean is then not the best measure of the salary of a
'typical' employee. Let's look at some actual figures. The salaries of 13
employees have been entered into column 1 of a MINIT AB worksheet and
printed out, as for the worm lengths, as follows (File, Display Data):
salary (pounds per year)
8500
12000

9000
12000

9500
18000

10000
17500

10000
25000

10500
54000

11500

Then a stem-and-leaf plot is produced (Graph, Character Graphs,


Stem-and-Leaf):
Stem-and-leaf of salary
Leaf unit=1000
3
(6)
4

2
2
1
1
1
1
1

o
1
1
2
2
3
3

N = 13

899
000122
78
5

4
4

MINITAB chooses an appropriate method of rounding for the data


set, so that the shape of the distribution is clear. The classes are in 5000
blocks. The stem gives tens of thousands and each leaf represents 1000.
Therefore, the smallest three values appear by a stem of zero (no tens of
thousands) as 8, 9 and 9. The next line contains the six values between
10000 and 14999.
The median value is that value of salary which is exceeded by 50% of
individuals. It is a good way of estimating the average salary for this data
because it is not unduly affected by the few very high values. We know that
the median is contained in the second class because the left-hand column
shows the number of observations in that class (6) in parentheses.
Starting from the high end of the range we see that one person has a
salary of 54000 and then there is a long gap before there is another one
who earns 25000, making two cumulatively who earn 25000 or more,
whereas most people earn far less.
The mean salary is 15952. If we use the trimmed mean this is reduced
to 13182 reflecting the influence of the one very high salary. However,
the observations are not symmetrical around the mean but have a much
wider spread (tail) of high values than of low ones. This is called positive
skewness. If there were more high values than low ones this would be
called negative skewness. In our case the median is a much better summary
statistic than is the mean (Stat, Basic Statistics, Descriptive Statistics):

25

26

I I

CONFIDENCE INTERVALS
salary

13

Mean
15962

Median
11500

Min
8500

Max
54000

Ql

Q3

salary

9750

17750

TrMean
13182

StDev
12360

SEMean
3428

This command tells us that the median is 11500. QI stands for lower
quartile. This is the value which has one-quarter (25%) of observations
below it. Obviously some rounding is necessary where the number of
observations is not exactly divisible by four. Similarly Q3 is the salary
which is exceeded by one-quarter (25%) of people. So three-quarters of
people have salaries below it. It is called the upper quartile. MINIT AB
produces a very helpful graph which contains these features, called a boxand-whisker plot (or boxplot) (Graph, Character Graphs, Boxplot):
minimum

01 median 03

\tL_~
-1+

outlier

maximum

\,

1---

-----+---+------+-----+---+-----salary
10 000
20 000
30 000
40 000
50 000

The cross shows the median. The box around it encloses the central 50%
of the observations and so it excludes the largest 25% and smallest 25% of
values. It is defined at the lower end by Ql and at the upper end by Q3.
The lines from each side of the box (called whiskers) extend as far as the
minimum and maximum values except that very small or very big values
(outliers) are marked separately beyond the ends of the lines as '*' or '0' to
draw attention to them. Thus the 54000 point is so marked in our
example. This also makes it very easy to spot mistakes made in entering
the data (for example if an extra zero had been added onto a salary of
10000).
Contrast the box plot for the salaries data with that for the lengths of
the ten worms (cm) we measured earlier:
6.0, 8.0, 9.0, 9.5, 9.5, 10.0, 10.5, 11.5, 12.5, 13.5
+

1-------

--+-----+---+---+----+--------+ ---length
6.0
7.5
9.0
10.5
12.0
13.5

Here the median is nearer the middle of the box and the whiskers are
symmetrical. It is often helpful to display the data in this way early on in
the proceedings, to get an idea as to whether they do show a roughly
Normal distribution (as for worm lengths) or not (salaries), because some
of the tests that will be discussed later are based upon the assumption that
the data are normally distributed.

C_O_M_P_A_R_I_N_G_TW
___O_P_O_P_U_L_A_T_I_O_N_S____________~I

L -____________

2.3 MINITAB - HIGH-QUALITY GRAPHICS


Releases 10 and 11 of MINITAB allow high-quality graphics. Using the
salary data try (Graph, Histogram) and (Graph, Boxplot). This will
produce the graphics in Figures 2.2 and 2.3. They are not saved with the
text in the Session Window. They may be saved using (File, Save Window
As) or printed (before or after editing) using (File, Print Window). They
may be retrieved using (File, Open Graph).
9 -

>.7 u
c 6 <Ll

:J
IT
<Ll

L-

4 -

u... 3

2 -

10000 20000

Figure 2.2

I
30000

salary

I
40000

I
50000

Histogram of salary data.

50000

>.40000
eo
rn({) 30000
L-

20000
10000

Figure 2.3 Boxplot of salary data.

2.4 COMPARING TWO POPULATIONS - HYPOTHESIS TESTING


AND THE t-TEST
Martin Hollis (1985) pointed out that: 'The "art of discovery" requires
imagination and conjecture: the "logic of validation" requires an objective

27

28

I L-I________C_O_N_FI_D_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ _--'
test, which a hypothesis is not guaranteed to pass. Scientific method works
by discarding or amending hypotheses when the predictions extracted from
them fail.'
If we have two random samples, one from each of two groups, we start
by assuming that they come from the same population, and so have a
common mean (the null hypothesis 'Ho'). If we are satisfied that the
evidence from our sample data is strongly against this idea we reject it and
instead accept that the samples come from populations with different
means (the alternative hypothesis 'HI '). We proceed in this way by rejecting
a hypothesis as false instead of accepting it as true, although it seems
cumbersome, because it is logically sound.
2.4.1

Introduction

The purpose of testing and comparison is to:


test the validity of,
or weigh the evidence for,
or detect the presence of, some specified hypothesis or judgement or
causation.
Thus:
1. The observations may conform with some hypothesis. This is evidence
for (but not proof of) the hypothesis.
2. The observations may differ sufficiently from the hypothesis (as judged
by our test). The hypothesis is probably false. The falsity may not
always be provable if the observations are only a sample (almost always
the case in biology).
3. The observations may possess the form, system, pattern, trend or
grouping that is expected if a particular cause-effect relationship holds.
This is evidence for that relationship, but, contrary to newspaper headlines, such relationships can never be proved by statistics.
The validity of our conclusions, i.e. how far we can rely on our data to
support I or 2 or 3 above, is measured by the size or frequency of
deviations from the hypothesis or from randomness. So we need a means
of quantifying deviations or differences.
Therefore, the confidence in a sample estimate and the support for a
hypothesis depend on the uniformity or consistency of the observations (or
how well they conform to expectations) which depends on measurement
of scatter or variation or dispersion.
Hence statistics works on the measurement and study of variability in
the form of: sum-or-squares and variance and standard deviation and in
particular through standard error, confidence limits and significance
testing.

C_O_M
__P_A_R_IN
__
G_T_W
__
O_P_O_P_U_L_A_T_I_O_N_S____________~I

L -____________

The simplest form of a significance test is to compare two samples, one


from each of two populations, using a t-test in which we calculate at-ratio.
2.4.2 The t-ratio

A ratio is obtained when one value is divided by another. The t-ratio


measures the distance from a sample mean to some point of interest,
usually in units of standard error. It answers the question: 'How far from
the mean is the point in question, when measured in units of standard
error?'
We can then calculate, by referring to a t-table, the probability that an
observation above the critical point might arise from the same population
from which the mean was taken.
The t-ratio takes the form:
t

sample mean - the point of interest


= ---=------.......:.------

standard error
and the probability level is discovered by comparing the calculated t-ratio
with the values in a t-table for the number of degrees of freedom
appropriate to the standard error in the denominator.
2.4.3

Independent t-test

This compares the means of two random samples. It is sometimes called


a 'two-sample t-test'. The 'null hypothesis' (Ho) used as a starting point is
that the samples come from the same population and we test this
hypothesis by looking at the means of the two samples and the variability
around each of them. The further apart the means and the less the overlap
in the individual values around each of them, the more likely it is that

9 -

units

8-

7-

6 5-

0
0

0
0

40

3-

Figure 2.4

Complete overlap.

29

30

I LI________C_O_N_FI_D_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ _

----I

9 -

B7 _

units 6 -

5 -

o
o
o

3 2

0r-'

----L-,--_ _ _ _ _ _ _ _ _ _ _ _ _

Figure 2.5 No overlap.


the samples come from populations with different means. In that case we
can reject the null hypothesis at a given level of confidence (e.g. p = 0.05)
although never with absolute certainty.
It is important to plot data first:
If the two sets of observations overlap each other a great deal it may
be so obvious that the null hypothesis cannot be rejected that it is
pointless carrying out a t-test (Figure 2.4).
If the observations in one sample are all clearly separated from those in
the other sample it may be so obvious that the null hypothesis is false that
a t-test is superfluous - this assumes that you have a reasonably large
number, say five or more, of observations in each group (Figure 2.5).
Often our data are intermediate - with some overlap between the two
samples (Figure 2.6). Here we need to carry out a statistical test. Before
embarking on a simple t-test, check whether the variability within each

9 -

8 7 -

units

6 -

3 A

Figure 2.6 Partial overlap.

COMPARING TWO POPULATIONS

L-____________________________________________________

uni t s

8 -

7-

6 0

5 -

43 -

o
o

o
o
o

Figure 2.7

II

Unequal variances.

of the two groups is assumed to be similar. If your plot shows great


differences in the variabilities you may have to do something else (see
transformation - later!) (Figure 2.7). This test is identical to 'one-way
analysis of variance' (see Chapter 6) but the latter is more flexible in that
it can compare samples from two or more groups at once.
2.4.4

Calculation of independent t-test

Suppose we have 14 sheep, seven of which were selected at random to


receive a new diet (X 2) while the other seven were given the old diet (XI).
We have recorded the change in weights (kg) of the animals after two
months on these feeding regimes and from this calculated the mean value
for each group. We wish to judge the magnitude of the difference between
the two means compared to a difference of zero (i.e. if they were the same
diet).
We can do this by at-test:
.

t-ratzo

difference between two means - zero


.
standard error of difference

The standard error of difference is often abbreviated to SED.


BOX 2.1

A note on variances
Note that even if the mean value has not changed, the variation in
weights about the mean might have done so. If the two variances are
very different this is evidence in itself that the populations behave
differently (i.e. the new diet is having a different effect and there is
little point in continuing with the comparison). We can test for this
by dividing the bigger variance by the smaller one. If this variance

31

32

IL________C_O_N_F_I_D_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ _

---l

ratio is larger than the value in F-tables for p = 0.025 for respective
degrees of freedom (6 and 6 in our example) we have evidence that
the variances are significantly different. We will return to this in
Chapter 6.
The numerator (on the top) of equation (2.3) (see Box 2.2) is straightforward to calculate:
3.671 - 4.714 = -1.043
To get the denominator (on the bottom) of equation (2.3) we must
proceed as follows. First, the pooled (or average or combined) variance is
obtained. We need to calculate the sum-of-squares of the observations in
XI and add this to the sum-of-squares of the observations in X 2 (equation
(2.1)). A quick method is to obtain the standard deviation from the
calculator (O"n_1 or XO"n_1 or s), square it to obtain the variance and multiply
by (n - 1) to find the sum-of-squares (see section 2.1.3 and the flow chart
preceding it):
(0.5648)2 x 6 = 1.9143 = first sum-of-squares

(0.8275i x 6 = 4.1086

= second sum-of-squares
1.9143 + 4.1086 = 6.0229 = total of both sums-of-squares

Divide the result by the sum of the degrees of freedom for XI and X 2 :
6+6

= 12

6.0229/12

= 0.5019 = the pooled variance

Then the standard error of the difference (equation (2.2)) is obtained by


multiplying the pooled variance by 1/7 + 1/7 (because 7 is the number of
sheep in each group) and taking the square root:
SED

= '/0.5019 x (0.1428 + 0.1428) = 0.37868

Finally the t-ratio equation (2.3) is t = (-1.043 - 0)/0.37868 = -2.75.


The negative sign indicates that diet 2 has the greater mean.
We ignore the sign and consult t-tables to see whether 2.75 is greater
than the critical value for the combined df, i.e. 12 df and for 95%, i.e.
p = 0.05 (Table C.2, row 12, column 5).
BOX 2.2

Equations
The equation for the pooled variance is:
S2 _
p -

Sxx l +Sxx2
(n1 - 1) + (n2 - 1)

(nl -l)si +(n2 -l)s~


(n1 - 1) + (n2 - 1)

(2.1)

COMPARING TWO POPULATIONS

L -____________________________________________________

II

then the standard error of the difference may be calculated:


(2.2)
and, finally:
t-ratio

(2.3)

As our t- ratio is bigger than the table value of 2.179 we can be


reasonably confident (95% confident) that the difference between the
population means is not zero. Formally, this means that we can reject the
null hypothesis (of no difference) with 95% confidence. In everyday
language we conclude that the populations probably (95% confidence or
p < 0.05) have different means. The p value is the probability, given our
data, that there is only one population. Here it is less than 0.05 or 5 in 100
or 5%.
If the t-ratio is much larger than the table value we can check it against
columns of the t-tables for 99% (p = 0.01) and 99.9% (p = 0.001) levels
for the same degrees of freedom.
Some computer packages and many papers in scientific journals
abbreviate the three conventional significance levels using asterisks:
%

Symbol

Phrase

<95
95
99
99.9

>0.05
0.05
0.01
0.001

ns

non-significant
significant
highly significant
very highly significant

*
**
***

Many computer packages (like MINIT AB) give the exact value of p
instead, e.g. p = 0.03.
This is preferable to consulting conventional levels in tables as it is more
accurate.
BOX 2.3

One-tailed and two-tailed tests


Usually we are asking the question: Is there any evidence that the
two treatments are not the same? In other words diet Xl could result
in either bigger or smaller sheep than diet X 2 This is known as a
two-tailed test since the mean difference could lie in either 'tail' of the
distribution, i.e. away from the zero in the middle.
We find the standard appropriate t value in the statistical table

33

34

I ~I_______________C_O_N_F_ID_E_N_C_E__IN_T_E_R_V_A_L_S______________~
(Table C.2) by looking down the column headed 5. This means that
there is 2.5% in each tail, giving 5% in total (i.e. p = 0.05). (Beware,
other sets of tables may not behave this way. You want the column
with 1.96 at the bottom - for infinite degrees of freedom!)
Occasionally we might be certain that one treatment could only
give sayan increase in weight, not a decrease. This would certainly
be the case in many medical trials - where we may only be testing
treatments which might increase survival. We are then asking if there
is any evidence that the mean of treatment XI is bigger than that of
treatment X 2 This is a one-tailed test. In this case we use the t value
in the column of Table C.2 headed 10. This means that there is 5%
in each tail and we are only interested in one of them.

2.4.5 MINITAB carries out a two-sample t-test


First we display the data (File, Display Data):
Row
1
2
3
4
5
6
7

old
3.4
3.9
4.2
4.5
3.6
2.9
3.2

new
4.5
4.8
5.7
5.9
4.3
3.6
4.2

Then we ask for a two-sample t-test (Stat, Basic Statistics, 2-Sample t- Test,
Samples in Different Columns):
Twosample T for old
N
Mean
old
7
3.671
new
7
4.714

vs new
StDev
0.565
0.828

SEMean
0.21
0.31

95% C.L for mu old - mu new: (-1.87, -0.22)


T-Test mu old=mu new (vs not =): T=-2.75 P=0.017 DF=12
Both use Pooled StDev= 0.708

Note that MINITAB refers to the population means as 'mu' (mu old
and mu new). This is the sound of the Greek letter J1 which is used to
represent the population mean(s) which we have estimated by the sample
means.
It concludes that we can reject the hypothesis of no difference between
the population means with 98.3% confidence (Le. p = 0.017). Another way
of thinking about this is to realize that 1. 7% of the means from samples
with eight observations in each group drawn from two populations with

C_O_M_P_A_R_I_N_G_T_W
__O_P_O_P_U_L_A_T_IO_N
__S____________~I

L -____________

the same mean will produce a difference between the sample means, which
indicates that the difference between the two population means is zero. Is
it more likely that you were lucky enough to draw one of the few sets of
samples that indicate that the mean difference is equal to zero; or that the
mean difference is really not equal to zero?
The confidence interval calculation tells us that we can be 95% confident
that the two population means differ by between 0.2 and 1.9 kg with the
new diet having the greater weight gain.
If the sheep were actually seven pairs of twins we can take these
relationships into account by asking MINITAB to calculate the differences
in weight gain for each of the seven pairs of twins and then carrying out
a 'one-sample t-test' also called a 'paired t-test' on these differences (Calc,
Mathematical Expression, New Variable=C3, Expression= Cl - C2,
File, Display Data):
Row
1
2
3
4
5
6
7

old
3.4
3.9
4.2
4.5
3.6
2.9
3.2

new
4.5
4.8
5.7
5.9
4.3
3.6
4.2

diff
1.1
0.9
1.5
1.4
0.7
0.7
1.0

Now the t-test on the difference (Stat, Basic Statistics, I-Sample t-Test):
Confidence Intervals
Variable
diff

N
7

Mean
1. 043

StDev
0.315

SEMean
0.119

95.0% C.I.
(0.751,1.335)

T-Test of the Mean


Test of mu=O.OOO vs mu not=O.OOO

Variable
diff

N
7

Mean
1. 043

StDev
0.315

SEMean
0.119

8.75

P-Value
0.0001

We are testing the null hypothesis that the mean difference is zero.
Now that this extra information is taken into account we can be even
more confident that the null hypothesis is untrue. The p value is now
0.0001. So the chance that the two groups of animals come from the same
population is only I in 10000.
The confidence interval calculation tells us that we are 95% confident
that the two population means differ by between 0.75 and 1.31 kg with
the new diet being an improvement. It is a narrower (more precise)
interval than from the two-sample t-test which ignored the pairing. This
shows the benefit of accounting for variation between the seven pairs
of twins. We will explore this concept again later in Chapter 5 when we
discuss the advantages of blocking treatments, i.e. of pairing or
grouping them.

35

36

1 ,-I________C_O_N_FI_D_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ ___
2.5 EXERCISES
Confidence intervals and boxplots

We can now use the statistics of the sample mean and its standard error,
which we calculated for each of two apple varieties in the exercise at the
end of Chapter I, to obtain confidence intervals for the population means
of the two varieties.
Answer
The t Tfllue

Look up the appropriate value of t from the statistical tables (Table C.2,
row = df, column headed 5). Double check: the correct column for a 95%
confidence interval value of t has 1.96 in the bottom row of the table.

mean
t

SE

Tyler's
Kernal

Red
Charles
Ross

24.85
2.262
0.610

18.66
2.262
0.472

Confidence interval

as
follows: multiply t by SE; subtract the result from the sample mean (to
obtain the bottom of the range); then add the same value onto the sample
mean (to obtain the top of the range).
The 95% confidence interval for population mean circumference of
Tyler's Kernal is:

N ow obtain the 95% confidence interval for the population mean

24.85 + (2.262 x 0.610)

i.e. from 23.5 to 26.2 cm

This means that we are 95% confident that the mean circumference of the
whole population of Tyler's Kernal apples is in this range. Another way of
thinking of this is to imagine taking another 99 random samples each of
ten apples so we have 100 samples in total. If we constructed such
confidence intervals round each of the 100 sample means, 95 of the
intervals would contain the population mean but five of them would not
contain it. In reality we only usually have the resources to take one sample
and so the confidence interval gives us a way of relating its summary
statistics to the mean of the whole population.
Similarly, the 95% confidence interval for population mean circumference of Red Charles Ross is:
18.66 (2.262 x 0.472)

i.e. from 17.6 to 19.7 cm

E_X_E_R_C_IS_E_S____________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

These two confidence intervals do not overlap, suggesting that these


are two distinct populations and that Red Charles Ross apples are larger.
In the next section we will see how to carry out a formal test on these data
to test whether we can reject the idea that there is only one population
here.

MINITAB exercise and answer


Enter the data into one column and name it 'em'. Enter code numbers '1'
and '2' into another column to indicate the two varieties.

Display the data


Data Display
Row

variety

1
2
3
4
5
6
7
8
9
10

1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2

11

12
13
14
15
16
17

18
19
20

cm

18.3
18.4
20.2
22.0
17.5
18.1
17.6
16.8
18.8
18.9
22.0
24.5
25.5
27.5
22.5
27.5
24.0
26.5
23.5
25.0

Obtain basic descriptive statistics for each variety


Descriptive Statistics
Variable
Tyler
Red

10
10

Variable
Tyler
Red

22.000
16.800

Min

Mean

24.850
18.660

Max
27.500
22.000

Median
24.750
18.350

Q1
23.250
17.575

TrMean

24.875
18.475

Q3

26.750
19.225

StDev

SEMean

1.930
1.492

0.610
0.472

37

38

II

CONFIDENCE INTERVALS

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Obtain stem-and-leaf plots for each variety


Character Stem-and-Leaf Display
Stem-and-leaf of em
Leaf Unit=0.10
1
3
(5)
2
2
1
1

16
17
18
19
20
21
22

22
23
24
25
26
27

N=10

variety=2

N=10

8
56
13489
2
0

Stem-and-leaf of em
Leaf Unit = 0 .10
2
3
5
5
3
2

variety = 1

05
5
05
05
5
55

Obtain dotplots for each variety


Character Dotplot
variety 1

-------+---------+---------+---------+---------+---------cm
variety 2

-------+---------+---------+---------+---------+---------em
18.0

20.0

22.0

24.0

26.0

28.0

Obtain high-quality boxplots for each variety

(Note: use variety 1 as the first 'y' and variety 2 as the second 'y', Frame,
Multiple Graphs, Overlay Graphs.) The result is shown in Figure 2.8.
Obtain a high-quality histogram for each variety

(Note: insert 'em' as 'x', select 'For each group' and insert 'variety' under
'Group variables'.) The result is shown in Figure 2.9.

L -_ _ _ _ _ _ _ _
E_XE_R_C_IS_E_S_ _ _ _ _ _ _ _---.l1

28

E 23 u

18 -

2
variety

Figure 2.8

High-quality boxplot for each variety.

1-11

>.4

u
c

~ 3

()
(l.)
Co.

LL

17

Figure 2.9

19

21

em

23

25

27

High-quality histogram for each variety.

The t-test

Here are summary statistics from the samples of two apple varieties which
we calculated in the last section:

n
df
mean
SD
Sxx

Tyler's
Kernal

Red
Charles
Ross

10
9
24.85
1.930
33.52

10
9
18.66
1.492
20.03

Carry out a t-test with Ho (null hypothesis) that 'the population means

39

40

I LI________C_O_N_F_ID_E_N_C_E_IN_T_E_R_V_A_L_S_ _ _ _ _ _ _

----l

are equal' and HI (alternative hypothesis) that 'the population means


differ', first using the calculator and then using MINITAB.

Calculator answer

Square each standard deviation (to get the variance) and mUltiply this by
the degrees of freedom to obtain the sum of squares for each sample and
check that you obtain the values of 33.52 and 20.03 in the above table.
Calculate the pooled variance by adding these two sums of squares and
dividing by the sum of the two degrees of freedom (pooled df). You should
obtain: 2.975.
We can now obtain the test statistics t by substituting for the two sample
means, the pooled variance and the two values of n in the equation:
.

t-ratlO

(x - x) - 0
I

s~(J..+J..)
n
n
l

You should obtain t = 8.02 (see below). If you obtain -8.02 this simply
indicates the direction of the difference (it depends which apple variety you
put into the equation first).
Compare this value with t-tables (Table C.2) using pooled degrees of
freedom. What is your conclusion?

The t-test using MINITAB

Carry out a two-sample t-test, using a null hypothesis of no difference


between the two variety means and an alternative hypothesis of inequality.
Assume that the variances are equal.
Twosarnple T for ern
variety
1

N
10
10

Mean
18.66
24.85

StDev
1. 49
1. 93

SEMean
0.47
0.61

95%C.I. forrnul-rnu2: (-7.81, -4.57)


T-Test rnu l=rnu 2 (vs not =): T=-8.02 P=O.OOOO DF=18
Both use Pooled StDev= 1. 73

As before, we have a very large value of t (8.02). So we reject the null


hypothesis with over 99.9% confidence and instead conclude that the two
populations have different means (p < 0.001). Note that MINITAB

~_______________E_X_E_RC_I_SE_S______________~I
presents a 95% confidence interval for the mean difference of between
-7.81 and -4.47. If this had included zero it would show that a difference
of zero was a possibility (i.e. the two varieties having the same population
mean). However, this confidence interval excludes zero, which is consistent
with the conclusion that one variety (here Tyler's Kernal) is larger than
the other.

41

Sampling

3.1 FIRST, CATCH YOUR WORM


In Chapter 1 I cheated and said, 'Let's imagine that we have collected
ten worms'. Not a word was said about how we would do this. But we
must now consider this challenge.
First, there are practical problems. If we go out into the field during
the day the worms will be buried and we will need to dig them up. How
deep should we go? Also, we run the risk of cutting some of them by
accident. It would probably be better to go out on a wet night with a torch
and collect them into a bucket while they are squirming on the surface.
Second, how do we decide which ten worms to pick up from the many
thousands available? If we collect them from near the footpath, because
this is convenient and we won't get our shoes too dirty, we might have
individuals which tend to be shorter than those in the rest of the field
because that was where they happened to be infected by parasites. We will
have failed to obtain a sample of ten worms which is representative of the
population of thousands of worms in the field. So deciding which ten
worms to pick is important and the process is called sampling. This chapter
will describe three alternative methods of sampling (random, stratified
random and systematic), together with their strengths and weaknesses.
3.2 RANDOM SAMPLING
3.2.1 Why do we need a random sample?
The purpose of random sampling is to ensure that each worm in the
population we are studying has an equal chance of being selected. If we
have been successful in taking a random sample then it will provide an
unbiased estimate of the mean. In other words if you took all possible
random samples, each of ten worms, from the field, the sample means

R_A_N_D_O
__
M_S_A_M
__
PL_I_N_G________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

would vary - some would be bigger than the unknown population mean
and some would be smaller. However, there would be as many 'high'
values as 'low' ones, and if you took the average of all of their means, that
value would be the true population mean.
Unfortunately, if we usually take only one random sample of, say, ten
observations we may obtain an unrepresentative result because, by chance,
no observations have been selected from a particularly distinctive part of
the population. Nevertheless, if we start off knowing nothing about the
worms and how they vary in length and abundance throughout the field, a
random sample is more likely to give us a representative sample than is
any other approach.
3.2.2

How to select a random sample

We have decided that our population consists of the worms on the surface
of a field on a damp night. Note that this immediately rules out extrapolating our results to worms that have stayed below ground. If we are
about to set out on a sampling scheme we should think about whether our
method automatically precludes some of the people, species, individuals
or events in which we may be interested. Telephone surveys exclude any
household that does not have a working telephone; surveys of woodland
plants will under-record plants such as celandines and anemones if they
take place in the late summer or autumn, after these species have died
back. But back to the worms!
We can make a random sample of surface-crawling worms by selecting
ten positions in the field at random, going to these and then picking up the
nearest worm. Make a rough plan of the field and mark the length and

100

Length
(m)

48 -

34

100

Width (m)

Figure 3.1 Position of a random sample.

43

44

II

SAMPLING

~----------------------------------------------------~

ROW

COLUMN

1 2

1 8 7 12 9 6 8 12
2 16 18 9 JO 7 9 14
3 22 [B] 17 14 12 13 16
4 19 16 13 10 8 9 12
5 16 12 12 9 5 7 10
6 19 16 II 7 8!Illi 14
7 22 18 16 19 12 13 17
8 25 14 16 12 13 15 16
9 20 17 14!m 17 19 00
10 I2SJ 22 19 25 29 32 36
II 29 30 27 32 37 42 45
12 31 35 38 45 47 44 49
13 28 33 30 1TIl40 39 42
14 26 29 32 35 35 31 29
15 22 28 31 29 28 27 25
16 21 22 28 26 31 27 22
J7 17 26 30 31 31 1211 25
18 17 24 31 33 27 22 19
19 19 19 21 17 15 16 16
20 23 1TIl17 14 13 19 23

14
10
9
9
12
10
12
14
24
30
45
51

Figure 3.2

9 10 11 12 13 14 15 16 17 18 19 20
7
6
4
8
6
16
18
23
25
35
43

4 5 4
3 4 8
5 6 5
7 9 10
8 7 6
15 17 16
ITIl 20 19
25 28 33
27 32 40
40 111147
51 53 52
[3DJ 58 56 50
46 51 48 44 40
32 40 37 35 31
27 16 22 19 16
24 00 14 10 8
33 12 8 7 6
17 14 10 9 7
19 16 13 10 12
17 12 II 6 9

3
5
7
9
6
18
25
30
42
45

7
6
7
8
5
17
22
27
35
40

48 39

41 43
37 30
16 19
8 4
6 6
6 4
3 0
8 3
7 3

4
5
6
3
6
19
24
31
35
32
42
30
35
20
2
4
2
3

2 4 6
3 3 4
5 4 5
3 7 JO
4 6 7
22 24 20
27 [2ffi' 17
25 22 20
37 29 28
33 27 27
!TIl 33 29
27 29 23
22 21 17
II 12 10
3 6 5
7 4 3
0 0
I 0 4
0 3 2
0

8
6
7
13
9
19
22
24
27
31
27
21
14
13
7
I

0
2
5
2

7
6
9
10

10
15
14
29

[33]
30
20

[j]]
16
12
10
5
0
0
3
4

Random sample of 16 patches.

width of it in metres (Figure 3.1). Then press the random number button
on the calculator (RAN #) to give the number of metres across the field
and press the random number button again to give the number of metres
down the field that we should go to find a sample point. For example, if
the numbers generated are 0.034 and 0.548 these can be interpreted as
34m across and 48m up (using the last two digits). Alternatively, we can
use the first two values (like 10 and 27) from a table of random numbers,
ignoring any which are too large (see Table C.7).
Make sure that the range of coordinates which are available will allow
any point in the field to be selected. If the field was more than 100 m in one
direction for example, at least three digits would be needed. Having found
one point we repeat this process until we have found ten points which lie
within the field. Let's look at a practical example which for simplicity will
be a small square field, 400 m2 in size. Figure 3.2 shows the yield of fruit
(g) from a particular shrub on each 1 m2 patch, conveniently arranged in
20 rows and 20 columns.
To select a position at random we could use the last two digits in the
numbers generated by the random number button on our calculator. If this
gives 0.302 and then 0.420 this would identify the position in row 2 and
column 20. If we select a random sample of 16 positions at which to

S_T_R_A_T_I_F_IE_D__R_A_N_D_O_M
__S_A_M_P_L_IN
__
G____________~I

L -___________

harvest the crop and measure its yield we could get the selection shown
in Figure 3.2. One of the most common problems in sampling is deciding
on how many observations should be made. A general rule is that the more
variable is the population, the more observations must be taken from it.
This is necessary if we are to obtain a 95% confidence interval which is not
so large as to be useless. Here we have taken 16 observations. This could
be regarded as a pilot study. If the confidence interval turns out to be very
wide we will take more observations. At least we will be unlikely to have
taken an 'excessive number of observations.
The standard error of the mean is then obtained by putting all the
observations into the calculator on statistical mode, pressing the standard
deviation button and dividing the answer by the square root of 16:
mean = 18.87 g, SE mean = 2.979 g
We notice that, by chance, in our sample of 16 points no positions were
selected from the top right or bottom right of the field, where crop yields
happen to be low. So this sample is unrepresentative although there is no
way that, without some other information, this could have been predicted
in advance. However, if we started off knowing which areas were
particularly low-yielding and which high-yielding then we could improve
our sampling technique. We do this in the next section.
3.3

STRATIFIED RANDOM SAMPLING

Where there is existing information or good reason to believe that the


variation in the feature we want to sample is not evenly distributed, then it
is sensible to divide the population into sections called strata (singular
stratum). For example, if we could see that the field in the example above
has a slope, with a stream at the bottom, we could make one stratum at
the top, two in the middle and one at the bottom of the slope. This would
be likely to give four strata which have soils with different moisture
contents, but with a reasonably consistent amount of soil moisture within
a particular stratum (conditions are homogeneous within it) and we might
expect that soil moisture is likely to influence yield. Also, although we
wouldn't be able to see all the crop yields, in real life we might well notice
that the shrubs are biggest in the central band of the slope and especially
small in the top right and top left of the field. We therefore could split each
of the four horizontal strata into two down the middle to form eight strata
- with the aim of having roughly even shrub size within anyone of them.
This is called stratification (Figure 3.3).
We have successfully identified a pattern in the population and formed
different SUb-popUlations which will enable us to improve the effectiveness
of our sampling of the yield of the shrubs in the field.

45

46

II~________________S_A_M_P_LI_N_G______________~
ROW

1 234
1 8 7 12 9
2 16 18 9 10
3 22 15 17 14
4 19 16 13 10
5 16 12 12 9
6 19 16 11 7
7 22 18 16 19
8 25 14 16 12
9 20 17 14 16
10 25 22 19 25
11 29 30 27 32
12 31 35 38 45
13 28 33 30 37
14 26 29 32 35
15 22 28 31 29
16 21 22 28 26
17 17 26 30 31
18 17 24 31 33
19 19 19 21 17
20 23 21 17 14

5
6
7
12
8
5
8
12
13
17
29
37
47
40
35
28
31
31
27
15
13

678
8 12 14
9 14 10
13 16 9
9 12 9
7 10 12
10 14 10
13 17 12
15 16 14
19 20 24
32 36 30
42 45 45
44 49 51
39 42 46
31 29 32
27 25 27
27 22 24
27 2533
22 19 17
16 16 19
19 23 17

COLUMN
9 10 11 12 13
7 4 543
6 3 485
4 5 657
8 7 9 10 9
6 8 766
16 15 17 16 18
18 21 20 19 25
23 25 28 33 30
25 27 32 40 42
35 40 41 47 45
43 51 53 52 48
50 58 56 50 41
51 48 44 40 37
40 37 35 31 16
16 22 19 16 8
20 14 10 8 6
12 8 7 6 6
14 10 9 7 3
16 13 10 12 8
12 11 6 9 7

6
8
9
13
9
7
10
13
15
19
32
42
44
39
31
27
27
27
22
16
19

COLUMN
9 10 11 12
7 4 5 4
6 3 4 8
4 5 6 5
8 7 9 10
6 8 7 6
16 15 17 16
18 21 20 19
23 25 28 33
25 27 32 40
35 40 41 47
43 51 53 52
50 58 56 50
51 48 44 40
40 37 35 31
16 22 19 16
20 14 10 8
12 8 7 6
14 10 9 7
16 13 10 12
12 11 6 9

14
7
6
7
8
5
17
22
27
35
40
39
43
30

19
4
6
4
0
3
3

15
4
5
6
3
6
19
24
31
35
32
42
30
35
20
2
4
2
3
1
1

16 17 18 19 20
2 468 7
3 346 6
5 4 5 7 9
3 7101310
4 6 7 9 10
22 24 20 19 15
27 20 17 22 14
25 22 20 24 29
37 29 28 27 35"
33 27 27 31 30
37 33 29 27 20
27 29 23 21 19
22 21 17 14 16
11 12 10 13 12
3 6 5 7 10
7 4 3 1 5
1 0 0 0 0
1 0 4 2 0
0 3 2 5 3
0 1 1 2 4

15
4
5
6

16
2
3
5
3
4
22
27
25
37
33
37
27
22
11
3
7
1
1
0
0

Figure 3.3 Eight strata.

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

1
8
16
22
19
16
19
22
25
20
25

29
31
28
26
22

21
17
17
19
23

2
7
18
15
16
12
16
18
14
17
22
30
35
33
29
28
22
26
24
19
21

3
12
9
17
13
12
11
16
16
14
19
27
38
30

32
31
28
30
31
21
17

4
9
10
14
10
9
7
19
12
16
25
32
45
37
35
29
26
31
33
17
14

5
6
7
12
8
5
8
12
13
17
29
37
47
40
35
28
31
31
27
15
13

7
12
14
16
12
10
14
17
16
20
36
45
49
42
29
25
22
25
19
16
23

Figure 3.4 Poor stratification.

8
14
10
9
9
12
10
12
14
24

30

45
51
46
32
27
24
33
17
19
17

13 14
3 7
5 6
7 7
9 8
6 5
18 17
25 22
30 27
42 35
45 40
48 39
41 43
37 30
16 19
8 4
6"6
6 4
3 0
8 3
7 3

6
19
24
31
35
32
42
30
35
20
2
4
2
3
1
1

17
4
3
4
7
6
24

20

22

29
27
33
29
21
12
6
4
0
0
3
1

18
6
4
5
10
7
20
17
20
28
27
29
23
17
10
5
3
0
4
2
1

19
8
6
7
13
9
19
22
24
27
31
27
21
14
13
7
1
0
2
5
2

20
7
6
9
10
10
15
14
29
35
30
20
19
16
12
10
5
0
0
3
4

S_T_R_A_T_I_F_IE_D__
R_A_N_D_O_M
__S_A_M_P_L_IN
__
G____________~I

L -___________

Stratifying in the opposite direction in this case would not be sensible.


The stratification system we use should be based on features that are
important for what we want to observe or measure, not on arbitrary lines
or in ways that do not reflect the underlying variation. For example we
could have divided up our field as in Figure 3.4. This would lead to some
strata having a mixture of big and small shrubs and is unlikely to improve
the effectiveness of our sampling if (as is likely) big and small shrubs differ
considerably in their yields.
When the boundaries of the eight strata have been fixed, then two (or
more) sample positions are selected at random from within each of them.
This is called stratified random sampling (Figure 3.5a). To select the points
for each stratum you can use the same procedure as before except that
once you have two points in a stratum you pass over any further ones that
fall in it, and keep going until you have two points in each. Alternatively
you may find it quicker to work out sets of coordinates for each stratum
separately. For irregularly shaped strata this can be done by overlaying
them with a grid of points (Figure 3.5b). The grid must of course be larger
than the whole area and the grid points outside the stratum are not used.
Here we are dealing with strata that are approximately equal in size.
The method can be generalized to unequal-sized strata.
When the area has been stratified into several roughly equally sized
strata we use a method which differs slightly from that adopted for simple
random sampling to obtain the standard error of the mean. The data for
a given stratum are entered into the calculator which, as before, is on
statistical mode. We then work out the variance of the mean for that
stratum by calculating the standard deviation from our calculator button,
squaring it (a quick way to return to the variance) and dividing by the
number of observations (n = 2) in the stratum. (See the flow chart in
section 2.1.2.) In Table 3.1 we calculate this for Figure 3.5.
The sum of the eight stratum means is 151 which when divided by 8
gives the sample mean 18.87, the same as for a random sample with the
same observations. What, however, is its standard error?
The sum of the variances of the stratum means is 52. There are eight
strata but to obtain the variance of the overall mean we divide not by 8
but by 82 which is 64. This gives 0.8125. It may seem odd to divide the sum
of the eight variances of the stratum means by 82 instead of just by 8 as
we did for the mean. This is because the variance is in squared units,
whereas the mean is in ordinary units. The standard error of the mean is
then found by taking the square root of the variance of the mean to give
0.9014. This is much smaller than the standard error of the mean from our
random sample (3.768).
So we have improved our estimate of the mean since the 95% confidence
interval derived by multiplying the standard error by an appropriate value of
t will be smaller. For example we might now be able to state with a particular

47

48

I ~I________________SA_M_P_L_IN_G________________J
(a)

ROW

COLUMN
I

I
2
3
4
5

8
16
22
19
16

7
18
15
16
12

12 9 6 8 12 14
9 10 7 [2] 14 10
17 14 12 13 16 9
[TI] 10 8 9 12 9
12 9 5 7 10 12

6
7
8
9
10

19
22
25
20
25

16
18
14
17
22

II 7
16 19
16 [I2)
(M] 16
19 25

11
12
13
14
15

29
31
28
26
22

30
35
33
29
28

16
17
18
19
20

21
17
17
19
23

8
12
13
17
29

10
13
15
19
32

14
17
16
20
36

10
12
14
24
30

9 10 I I 12 13 14 15 16 17 18 19 20
7
6
4
8
6

4
3
5
7
8

5 4
4 8
6 5
9 10
7 6

16
18
23
25
35

15
21
25
27
40

17 16
20 19
28 33
3240
41 47

3 7
5 []]
7 7
9 8
6 5
18
25
30
42
45

17
22
27
35
40

4
5
6
3
6
19
24
31
35
32

2 4 6 [}D 7
3 3 4 6 6
5 4 5 7 9
3 7 10 13 10
4 6 7 9 10
22
27
25
37
33

27 32 37 42 45 45 @] 51 53 52 48 39 42 37
[E) 27
30 37 40 39 42 46 51 48 44 40 37 30 35 22
32 35 35 31 29 32 40 37 35 31 16 19 20 II
31 29 28 27 25 27 16 22 19 16 8 4 2 3

~ 45 47 44 49 51 50 58 56 SO 41 43

22 28
30
24 31
19 21
21 17

26
31
33
17
14

31
31
27
15
13

27
27
22
16
19

22
25
19
16
23

24
33
17
[l2)
17

20
12
14
16
12

14 10 8 6 6
8 7 6 []:I 4
10 9 7 3 0
13 10 12 8 3
II 6 9 7 3

4 7
2 1
3 I
I 0
1 0

24 20
20 17
22 20
122128
27 27
33
29
21
12
6

19 15
22 14
I24l 29
27 35
31 30

29 27
23 I2Il
17 14
\0 13
5 7

4 3
0 0
0 131
3 2
I

1
0
2
5
2

20
19
16
12
10
5
0
0
3
4

(b)

Figure 3.5 (a) Stratified random sampling of two patches in each of eight strata.
(b) Select random points from those within the area marked '.'.

degree of confidence that the population mean lies within the range from
17.9 to 19.9 instead of from 16.9 to 20.9 derived from the random sample.
The value of t for a stratified random sample with two samples per
stratum is found for degrees offreedom equal to the number of stratum. Here
there are eight degrees of freedom. This is because we add together the

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _S_Y_S_T_E_M_A_T_IC
__
SA_M
__
P_L_IN_G______________

~I

Table 3.1 Working out the sum of the variances of the


stratum means
Observations
(g/m2)

Stratum
mean

Variance of
stratum mean"

9 13
12 14
38 43
19 26
6
8
24 29
21 30
4
6

11.0
13.0
40.5
22.5
7.0
26.5
25.5
5.0

4.0
1.0
6.25
12.25
1.0
6.25
20.25
1.0

Total = 302

151.0

52.0

For example, standard deviation = 2.8284 (from calculator), variance =


(2.8284)2 = 8.0, variance of the mean for the stratum = 8.0/2 =4.0.

degrees of freedom for each stratum (1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 = 8)


just as we added together the estimates of variability from within each
stratum (above).
A stratified random sample which has been sensibly allocated to a
population with some pattern in it will always give rise to a smaller
standard error than that from a random sample. This is defined as
improving the precision of our estimate. Why does this happen?
Stratified random sampling splits the total variation in the sample into
two parts: that caused by variation within each stratum and that caused by
variation between the strata. The latter is then excluded and the standard
error is calculated from within-stratum variation only. This should be
relatively small, since strata have been chosen to be internally homogeneous. In other words the differences between the observations in each
stratum are relatively small compared with the differences between all the
observations in a fully random sample. That is why the stratification in
Figure 3.4 is considered poor - the mean value may still be the same or
similar but the variance and hence the standard error and confidence
interval are larger. If strata are naturally of different sizes or of different
variability we should make more observations in the bigger and/ or more
variable strata.
In summary we can think of random sampling as providing an insurance
policy against bias and of stratification as a way of improving precision.

3.4 SYSTEMATIC SAMPLING


Another form of sampling that you may come across is systematic
sampling. In systematic sampling the sample units are chosen to achieve

49

50

I ~I_________________SA_M_P_L_IN_G________________~
maximum dispersion over the population. They are not chosen at random
but regularly spaced in the form of a grid (Figure 3.6a).
Systematic sampling is very efficient for detecting events because we
are less likely to miss one (perhaps a fallen tree within a wood or a molehill
in grassland) than if the sampling has a random element. Also, because
you are more likely to include both very small and very large individuals,
the mean of a homogeneous population is often close to the true mean, but
has a large standard error.
(a) ROW

I
2
3
4
5
6
7
8
9
10

II
12
13
14
15
16
17
18
19
20

COLUMN

8
16
22
19
16
19
22
25
20
25
29
31
28
26
22
21
17
17
19
23

7
18
15
16
12
16
18
14
17
22
30
35
33
29
28
22
26
24
19
21

12
9
17
13
12
II
16
16
14
19
27
38
30
32
31
28
30
31
21
17

9 6 8
\0 [2] 9
14 12 13
\0 8 9
9 5 7
7 8 10
19 [UJ 13
12 13 15
16 17 19
25 29 32
32 37 42
45 IDJ44
37 40 39
35 35 31
29 28 27
26 31 27
31[TI] 27
33 27 22
17 15 16
14 13 19

12
14
16
12
\0
14
17
16
20
36
45
49
42
29
25
22
25
19
16
23

14
10
9
9
12

9 10 II 12 13 14 15 16 17 18 19 20

10

12
14
24
30
45
51
46

32
27
24
33
17
19
17

7
6
4
8
6
16
18
23
25
35
43
50
51

5
4
5 6
7 9
8 7
15 17
12Il20
25 28
27 32
40 41
51 53
l3ID 56
48 44
40 37 35
16 22 19
20 14 10
12 [8J 7
14 10 9
16 13 10
12 II 6

OJ

4 3 7 4 2
8 5 6 OJ 3
5 7 7 6 5
10 9 8 3 3
6 6 5 6 4
16 18 17 19 22
19 25 22 ~ 27
33 30 27 31 25
40 42 35 35 37
47 45 40 32 33
52 48 39 42 37
50 41 43 00 27
40 37 30 35 22
31 16 19 20 II
16 8 4 2 3
8 6 6 4 7
6 6 4 Dl I
7 3 0 3 I
12 8 3
0
9 7 3
0

4
3
4
7
6
24
20
22
29
27
33
29
21
12
6
4
0
0
3
I

6
4
5
\0
7
20
17
20
28
27
29
23
17
10

5
3
0
4
2

8
6
7
13
9
19
22
24
27
31
27
21
14
13
7
I
0
2
5
2

(b) (ii)

(b) (i)

10

11

12

13

15

18

19

21

----..

LOW

~~--------~
MEDIUM
HIGH

Figure 3.6 (a) Systematic sample of 16 patches. (b)(i) Observations from a


systematic sample; (ii) stratified random sampling boundaries.

[Q]
9
\0
\0
15

Illl
29
35
30
20

II2l
16
12
10
5

[QJ
0
3
4

F_U_R_T_H_E_R_M_E_T_H_O_D_S_O_F_SA_M_P_L_IN_G
_ _ _ _ _ _---'1

L -_ _ _ _ _ _

If we have decided to use a systematic sampling grid then the first point
should be chosen at random. Once this is chosen (for example, column
15, row 2 in Figure 3.6a) the other sample points are chosen in a fixed
pattern from this point according to the scale of the grid, and so they are
not independent from each other. It is important to locate the first point at
random because otherwise there is a risk that the grid will be positioned
such that all the points miss (say) the edges of the sample area.
Much use is made of systematic sampling and the data are often treated
as if they were from a random sample. For example in work in forestry
plantations every 10th tree in every 10th row may be measured. As long as
the number of sample units is high there is little risk of coinciding with
any environmental pattern which might affect tree growth.
Similarly in social surveys, every 50th name on the electoral roll might
be selected as a person to be interviewed. This is very convenient.
However, it is important to be aware of the possibility of bias in systematic
surveys. In the social survey for example, some flats might be in blocks
of 25 and all occupied by couples, so we could end up only interviewing
people who lived on the ground floor. In the forestry survey every 10th tree
might coincide with the spacing of the forest drains so that all the sampled
trees were growing a little bit better than their neighbours on the wet
site.
Systematic sampling is excellent for mapping an unknown area however
and for looking for patterns that you may wish to investigate in later
samples. The yields of fruit per shrub taken from 16 plants distributed
evenly as points in a grid can be used to divide the area into parts of low,
medium and high productivity (Figure 3.6b). Such parts could be used as
strata in subsequent stratified random sampling, to obtain an unbiased,
precise confidence interval for the population mean yield.

3.5 FURTHER METHODS OF SAMPLING


The basic methods of sampling may be made more sophisticated. For
example, strata may naturally be of different sizes or we may choose to
sample a very variable stratum more intensively. This is called irregular
stratified random sampling. We may select individuals in groups or clusters
for convenience by choosing several National Trust properties at random
and then recording the height of all specimens of Wellingtonia trees at
each.
Sampling may be two-stage. Here we take a random sample of, for
example, trees in a wood and then count the insects we find on a random
sample of leaves on each of these trees. In addition, we might decide to
select the trees with 'probability (of selection being) proportional to size'.

1 51

52

I I~________________S_A_M_PL_I_N_G______________~
In other words we select so that a bigger tree stands a greater chance of
being selected than a smaller one. This can be an efficient method of
sampling because the population mean depends more on the means of the
large trees than on those of the smaller ones. Sampling is a huge subject
area in itself. As a beginner, you won't need to use any of these more
sophisticated methods, but it is always a good idea to discuss your
proposed method with someone who has experience of sampling. If
nothing else they may be able to provide you with a few practical tips.
(For example, when doing fieldwork never be separated from your
waterproof clothing and your sandwiches!)
3.6 PRACTICAL PROBLEMS OF SAMPLING
In real life, areas of land tend to be strange shapes rather than square.
Then we must mark the boundaries of homogeneous areas as strata on a
map. Draw a baseline along the bottom of the map (the x axis) and a line
up the side at right-angles to it (the y axis). Mark distances in metres along
the x and y axes of the map. Then select the positions of the required
number of small sampling areas (these are usually called quadrats) in each
stratum using random numbers to identify the x and y grid coordinates.
In the field the quadrats can be sited in turn by pacing 1 m strides up
and across, placing, say, the bottom left corner of the quadrat where the
toe of the last foot falls.
We can make sample plots circular, square (rectangular) or in the shape
of a long, thin strip (called a transect) (Figure 3.7a-c). The advantage of

(a)

(b)

(c)

(d)

Figure 3.7 Types of sampling plot: ( a) circular; (b) square; (c) transect; (d) use
of transects from paths.

EXERCISE

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

II

a circular plot is that we need to mark its position only at one central point
where we can then fix a tape measure rather than having to mark all the
four corners of a square. We can use a transect when it is difficult to move
through the vegetation. Strips cut into very dense vegetation at rightangles from tracks and positioned at random starting points along the
paths are more efficient than random quadrats covering the same area,
because we do not waste time in getting to them (Figure 3.7d).
We must mark quadrats securely if we want to re-record them later. A
couple of wooden pegs is not likely to be sufficient to survive several years
of field wear and tear or vandalism. A buried metal marker (15 cm of thick
metal piping) provides added security and we can re-Iocate it using a metal
detector. If we intend to make observations at the same place on several
dates a square quadrat is advantageous. It is easier to decide whether, say,
new seedlings are inside or outside it and, if there are large numbers of
them, we can decide to sample only part of the quadrat (subsampling). For
example, although we might record cover of established plant species over
the whole quadrat, we might count seedlings present in the bottom right
quarter only.
If we divide aIm square quadrat into, say, 25 squares each of
20 x 20 cm we can then record the presence or absence of particular species
within each of these small squares. Then each species will have a frequency
out of 25 for each quadrat. This is an objective method which encourages
careful scrutiny of the whole quadrat and allows us to re-record subquadrats over time. It gives figures which are likely to be more reliable
than subjective estimates of percentage ground cover for each species.
Such subjective estimates of cover vary greatly from one observer to
another. Also, the appearance of the same species under different
managements or at different seasons may lead us to under- or overestimate its importance.

3.7 EXERCISE
Choice of sampling method

1. Why do we sample populations?


2. Why is it important to use random numbers to select a sample?
3. In what circumstances is it desirable to stratify a population before
selecting random samples from each stratum?
4. When would a systematic sample be useful?

Answers
1. A population often consists of too many individuals for us to measure
all of them. Therefore we wish to select a few individuals and measure

53

54

I ~I_________________SA_M_P_L_IN_G________________~
them. We want this sample to be representative of the whole
population.
2. If we use random numbers to select the individuals to be measured we
ensure that each one has an equal chance of being selected. This means
that the sample is not biased and should therefore be representative of
the whole.
3. If the population is heterogeneous it may be that a single random
sample will be unrepresentative because say, by chance, no individuals
have been selected from a distinctive part of the population. So it will
be advantageous to divide the population up into internally homogeneous strata or sub-populations. This also reduces the variability
within each stratum and hence the standard error of the mean and the
width of the confidence interval for the population mean.
4. A systematic sample runs the risk of coinciding with a pattern in the
population and so obtaining a biased sample. However, if this is
thought unlikely to be a problem we may sample, say, every 10th
person on a list. The problem is that we can then never be sure that this
is not a biased sample. However, if we wish to map an area about which
we know nothing a grid sample is sensible.

Planning an
experiment

So far we have learned how to estimate the mean of a large population


by observing a small sample of individuals taken from it. Very often,
however, we wish to compare the mean performance of several different
categories. For example:
Which of three fertilizers gives the best crop yield?
Which of four species of worm moves the fastest?
Which drug cures the disease most rapidly?
In this chapter we will see how to design experiments which will answer
such questions.
Let's start with a comparison of the yields of barley obtained from
applying three increasing amounts of fertilizer (A, Band C where A is the
lowest amount and C the highest). The term treatment is used to describe
that which we are varying in our experiment. In the experiment that we
are going to design we have three treatments.
4.1

REPLICATION

Our first thought might be just to split a field in three and apply a different
amount of fertilizer to each third (Figure 4.1a). This could give misleading
results however. Suppose that the natural fertility of the soil is higher at
the bottom of the slope, then whichever fertilizer is allocated to that
position will appear to be better than it really is in comparison with the
others. However, we could divide the field into, say, 12 parts or
experimental units (usually called plots) and allocate each fertilizer
treatment at random (see section 4.2) to four of them.
This will improve matters if there is variation in fertility in the field to
start with because it is unlikely that all of one treatment will end up in a
very high- or very low-fertility patch. Rather, the underlying variation is

56

II

PLANNING AN EXPERIMENT

~----------------------------------------------------~

(a)

Top

(b)

c
Botlom

Figure 4.1

(a) No replication. (b) Four replicates of each treatment.

likely to be spread between the treatments (Figure 4.lb). We now have


four replicates of each treatment.
We now compare the yields from just the plots that receive fertilizer
treatment A. Any variation in the yields between the four plots is caused
by random variation in conditions across the site. The same is true for
variation in the four yields from treatment B and, separately, from
treatment C. So replication has given us three separate estimates of the
background or random variation in the yields of plots receiving the same
treatment. This provides a basis for comparison with the differences
between the treatments that might be due to the fertilizer treatments. It is
essential to have at least two replicates of each treatment but four
replicates is commonly considered to be a minimum in field experiments.
The greater the replication we have (perhaps we could have six plots of
each treatment instead of four), the better is our estimate of the random
variation effects within the plots in that treatment. As in the previous
chapter our estimate of the mean yield of treatment A, for example, is
improved and the confidence intervals attached to the mean are reduced.
Thus we should be able to detect smaller real differences in yield between
the treatments from which we have taken our samples. We can achieve
increased precision in our estimates of the population mean yields and of
differences between them by having more replication.

4.2 RANDOMIZATION
The four replicate plots of each treatment must be allocated to positions
in the field at random. This is achieved by numbering the 12 plots from I

R_A_N_D_O
__
M_IZ_A_T_I_O_N__________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

to 12 (Figure 4.2a). Then we use the random number button on the


calculator (or use random number tables) to select four of these plots for
treatment A. For example, using the last two digits from the numbers
0.809, 0.312, 0.707, 0.836 and O.lOl allocate this treatment to plot
numbers 9, 12, 7 and 1 respectively; we ignore the value 36 from 0.836
because it is greater than 12) (Figure 4.2b). Then we select four more
numbers for treatment B and treatment C must go on the four plots which
remain.
As with selecting a sample from one population (Chapter 3) we can
think of randomization in allocating treatments to experimental units as
an insurance policy. It protects us from obtaining estimates of treatment
means which are biased (consistently higher or lower than the population
mean - say, because all plots from one treatment were in the corner of the
field where the crop was damaged by frost). Randomization is also
necessary because it helps to ensure that we can carry out a valid statistical
analysis of the data (Chapter 6).
The replicates of each treatment must be interspersed (mixed up) over
the site. Randomization is a good way of achieving this. Very rarely,
randomization may produce an arrangement of plots in which say the four
replicate plots of one treatment are grouped together at the bottom of
the slope. Although this would be acceptable if this were one of a series of
similar experiments, in real life this is unfortunately seldom the case. To
avoid the possibility of our one-and-only experiment giving unrepresentative results, we could re-randomize the positions of the treatments, so that
they are satisfactorily interspersed. However do not go to the other
extreme of imposing a systematic layout of the plots such that treatment B
is always between A and C. This may also create problems (as well as
(b)

(a)

10

11

12

B
B

Figure 4.2 (a) Allocation of plot numbers. (b) Allocation of treatments to plots
using random numbers.

57

58

I ~I______________P_L_A_N_N__IN_G__A_N_E_X_P_E_R_I_M_E_N_T______________~
being statistically invalid) if, for example, it means that B is always being
shaded by the taller growing C.

4.3 CONTROLS
Suppose we carry out our fertilizer experiment and there are differences
between the treatments, with the yields from A being the least. How will
we know that A is in fact having any effect on growth at all? What we need
is a control. A control is the name given to a treatment in which (usually)
nothing is applied to the plot. We can then see what changes take place
naturally during the experiment. A slight variation on this idea is a
procedural control. For example, suppose the experiment is taking place
not in a field but in a glasshouse and the treatments are different chemicals
applied in water. If we have a treatment where nothing is applied the
results we get may be simply the effect of the water applied with chemicals.
Therefore we might apply water only as a control. This allows us to assess
any effect of the chemicals separately from that of the water.
So far we have been considering a simple experiment in which fertilizers
are applied to barley. Imagine a more complicated one in which we are
interested in the effect of sheep grazing at different times of year on the
number of wild flower species in grassland at different sites across the
country. We can fence off each plot to make sure that the sheep are in the
right place at the right time each year for, say, five years. But the grassland
present at the beginning will contain a range of species characteristic of
its location, soil type and previous management. These may differ from
one place to another. For example, one site had always been grazed with
sheep whereas another was growing maize two years ago.
If we want to carry out the same experiment on several sites, we must
have control plots on each site. They tell us what happens on each site in
the absence of any grazing treatments (in this instance) and provide a
standard comparison between the two sites. In the same way, if we carry
out the same experiment in different years the results from the control
plots in each year provide a standard basis for comparison of the effect of
different treatments.

4.4 OBJECTIVES
The above gives you an idea of the key factors in experimental design replication, randomization and controls. But there is an important stage
we have skipped - precisely what are you trying to test in your experiment.
It is always good practice to write down the background to your
experiment. This consists of why you are interested in the problem and

LAYING OUT THE EXPERIMENT

L-____________________________________________________

II

your general objective. For example, farmers may be paid to sow grass in
strips of land round the edges of their corn fields to encourage wildlife.
The seed used to establish these strips could also contain wild flower seeds
(but at much greater cost). The wild flowers will attract butterflies and bees
or spiders and may be generally thought to be better for nature
conservation than strips without them. However the strips may also
harbour weeds that could spread into the crop and reduce crop yield or
quality. How often and at what times of year should the grass strips be cut
if we want to encourage the growth of the desirable wild flower species
without increasing the competitive weed species. Is it possible to
recommend the best solution? Let us take this question and see where it
leads us in terms of trying to design and layout an experiment. We must
not forget the practical issues of how and what we record that will need to
be faced before we ever get results to analyse.
4.5 LAYING OUT THE EXPERIMENT

Talking to farmers we have discovered that they think two cuts a year
are needed to stop thistles flowering and setting seeds which then become
weeds in the crop. However, wildlife advisers believe that two cuts may be
too much for the sown wild flowers. Our first questions then become:
1. What is the effect of sowing or not sowing wild flowers?
2. What is the difference between one cut and two cuts?
We have decided to compare four treatments:

With flower seeds


No flower seeds

Cutting once a year

Cutting twice a year

Fl
NFl

F2
NF2

At this stage we might well also sensibly decide to include a control


treatment. But, for simplicity of explanation here we will deal only with
the above four treatments. We could decide to have four replicate lengths
of strip (plots) for each of our treatments. We must decide the size of each
strip. Its depth from the edge of the field to the crop will be as used by
farmers (say, 10 m), but how long should a plot be? It is important to have
enough room to turn mowing machines round at the ends of each plot.
Also it's possible that wild flower seeds may blow a metre or so into a
neighbouring plot at sowing.
When a particular plot has received its one cut of the year it will grow
quite tall in the rest of the year and may shade the edge of the neighbouring plot which receives a further cut. All these facts point to the idea
of having a buffer zone round the edge of each plot (Figure 4.3a). This is

59

60

II

PLANNING AN EXPERIMENT
(a)

(b)

Buffer zone

Fl
N

F
2
F
2

F
1

(e)
1m

(
1m

(d)

.-

F
2

11
20em
D20em

Figure 4.3 (a) Plot with buffer zone. (b) Field layout of plots - wholly
randomized design. (c) An individual quadrat. (d) Presence or absence recorded
in each small square.

treated as far as possible like the rest of the plot but we will not make
any recordings from it. It is unlikely to be a good representation of the
treatment because it may be affected by the neighbouring treatment. Its
function is to protect the inside of the plot.
It is common to use aIm x 1 m quadrat for recording vegetation. There
will probably be variation within each plot - perhaps caused by soil
composition, shading by occasional large shrubs in the nearby hedge or
rabbit grazing. Just as it was desirable to have replicates of each treatment,
so it is desirable to take more than one sample from within each strip.
We now decide how we are going to sample each particular treatment
strip. We will take the mean value from several quadrats as a fairer representation of the plot as a whole than if we had recorded only one quadrat
which happened by chance to have been on the regular pathway of a
badger.
Choosing the number of replicates and samples within each replicate is
a problem. On the one hand the more we have, the more precision we will
have in our results. On the other hand we will have more work to do. We
need to consider the cost and availability of resources: people, land,
equipment and the time and effort involved. Tired people will start to
make mistakes at the end of a hard day's work, no matter how dedicated
they are. Somewhere there is a compromise to be reached. Let's assume
that we decide to have three quadrats on each plot. This will allow us to
decide on the length of the plot, after allowing room between each quadrat

RECORDING DATA

L -____________________________________________________

II

to move around and place equipment. We now have 16 x 3 =48 quadrats


to record in total.
Our four replicate plots of each treatment have been allocated at
random to 16 positions round the field (Figure 4.3b). They have all been
ploughed and made level, with eight being sown to grass and eight to grass
and wild flowers, as appropriate, and we know that we will be recording
three 1 m xl m quadrats in each plot. What, however, will we record in
each?
4.6 RECORDING DATA

To monitor the effect of the treatments on species composition we could


make a list of the species occurring in each quadrat. It is a good idea to do
this as soon as the plants have established themselves and are recognizable,
before any cutting is done. Any differences in establishment may help to
interpret the results that we see later on once the differences in cutting
regimen start to take effect. As well as a simple list of plants per quadrat
however it would be more informative to have an estimate of what
percentage of the quadrat area was covered by each species.
This is rather subjective and different people would give very different
estimates. A good way to be more objective is to have a grid of 25 small
squares on a I m xl m frame placed on the quadrat (Figure 4.3c). Then we
record the presence or absence of each species in each small square (Figure
4.3d). A species present in every square receives a value of 25125 = 1,
whereas one present in five squares receives 5/25 = 0.2.
Taking care in the conduct of any experiment is essential. You want to
be able to tell at the end whether or not a treatment has an effect, but there
are all sorts of other errors and sources of variation that could creep in.
For example, the treatments must be applied uniformly which is not
always as easy as it sounds. The results must be recorded accurately. So if
several people are involved they may need a training session and should
then each record or treat one complete set of all treatments (otherwise the
differences between treatments may be mixed up with the differences
between recorders).
If you can you should try to prevent people knowing which treatment
goes with which plot, for example by using plot numbers as codes. This
should prevent any unconscious bias towards the treatment that someone
thinks ought to come out best.
Take account of the human factor. Recording data in the field can be
difficult and also boring if there is a large amount to be done. We must be
prepared to carryon in strong sun or wind and even in drizzle - so it's wise
to have suitable clothing and refreshments. The more comfortable you
are the more likely you are to make good records. Therefore, it is useful to

61

62

II

P_L_A_N_N_I_N_G_A_N_E_X_P_E_R_I_M_E_N_T_ _ _ _ _ _ _-'

L _______

Date _ _ _ _ __

Recorder _ _ _ _ __

Quadrat No. _ _

Square

1
2
1 234 567 890 1 234 5 6 789 0 1 2 345

/./,/./1

./.f "

Lolium perenne

Figure 4.4 Top of record sheet for field edge experiment.

have something soft to kneel on when peering at the plants. If you are
studying the behaviour of ducks on a pond take a folding stool to sit on. If
you can persuade someone to write down the observations as you call
them out, this can speed the process up a great deal. Spend time designing
and testing a recording sheet before photocopying or printing it for use.
In our case the sheet needs space for the date, recorder's name and plot
code, plus a column of species names down the left-hand side (Figure
4.4). The next 25 columns are used to tick the presence of a species in one
of the 25 small squares in the quadrat. The column at the far right is used
to enter the total for each species at the end of the day's work. We need
a minimum of 48 sheets for each time we do the recording and, in practice,
a few spares.
Because of all the work that has gone into designing and laying out the
experiment and recording the quadrats the data are precious. Ideally they
should be transferred to a computer as soon as possible, whilst always
keeping the original sheets. If the data cannot be input immediately the
record sheets should be photocopied and the copies kept in a separate
building. Always remember, what can go wrong will go wrong!

4.7 MINITAB
The observations made in the fertilizer experiment described at the
beginning of this chapter can be easily compared using MINITAB. We
code the three treatments (A = 1, B = 2 and C = 3) and put these codes
directly into one column of the worksheet (headed 'fert' here) and the
yields from each plot in the next column (headed 'yield') (File, Display
Data):
ROW
1
2
3
4
5

fert
1
1
1
1
2

yield
7.8
8.3
8.7
7.5

8.0

I I

MINITAB
2
2
2
3
3
3
3

7
8
9
10
11
12

8.9
9.3
9.7
10.3
11.8
11.0
10.9

Note that MINITAB puts in the row numbers to tell us how many
observations we have. We can then ask for a boxplot to be drawn
separately for each of the three treatments (Graph, Character Graphs,
Boxplot):
fert
--I

1-

-I

1-

-I

1-

--+---+---+---+---+----yield
8.00
8.80
9.60
10.40
11.20

Or we could ask for the yield from each plot to be put on a graph
against its treatment code (Graph, Character Graphs, Scatter Plot):
12.0+
yield

10.5+

9.0+

7.5+
--+---+---+----+-__+---yield
1.20
1.60
2.00
2.40
2.80

63

64

I LI______________P_L_A_N_N_I_N_G__A_N_E_X_P_E_R_I_M_E_N_T______________~
The codes for fertilizer (1, 2, 3) are along the bottom axis. MINITAB
expresses the values on this axis as though it were possible to have values
which were not whole numbers. This is not very elegant.
To make this clearer we can ask for the points to be labelled A, Band
C (Graph, Character Graphs, Scatter Plot, Use Labels):
12.0+
C

yield

C
C

10.5+

9.0+

A
A
A
A

7.5+

B
B
B
B

---+-------+---+-----+----+1.20
1.60
2.00
2.40
2.80

fert

These displays show that there is some overlap between the yields from
treatments A and B. but that the yields from treatment C seem generally
higher than those from the other two. We need some way of quantifying
this description and assessing the evidence objectively. Some of the
variability between yields is caused by the different amounts of fertilizer
and some by random variation. How confident can we be that if we
repeated this same experiment in exactly the same conditions we would
obtain similar results? In Chapter 5 we will see how to describe these ideas
in a model. This then will enable us (Chapter 6) to test whether or not
these fertilizers affect yield and to say how certain we are about our
conclusion.
General guidelines on the choice of method of data analysis are given
in Appendix A.

4.8 EXERCISE
Planning an experiment on diet and growth in mice

You are asked to carry out an experiment to compare the effects of six
different diets (A to F) on the growth of mice. You have 18 mice available
- each in a separate cage. Allocate three mice per treatment to their three

EXERCISE

L -____________________________________________________

II

cages - using either a table of random numbers (Table C.7) or the random
number button on your calculator. One way of doing this is to represent
each cage by its row and column coordinates. So that the first mouse (diet
A) would be allocated as shown if the first two random numbers selected
were 2 (for row) and 4 (for column):

At the end of the experiment the following weight gains (g) were
recorded:
Diet A

Diet B

DietC

Diet D

Diet E

Diet F

24.7

20.4

16.6

16.0

19.0

20.5

23.3

20.9

17.5

17.2

18.3

19.7

23.7

21.4

17.0

17.0

18.7

19.6

Enter the data into a MINIT AB worksheet. In column 1 you will need
codes to identify the diets (l to 6) (try Calc, Set Patterned Data), while the
18 weight gains go into column 2. Save the worksheet. You will need it
again in Chapter 5. Try summarizing the data using high-quality graphics:
Data Display
Graph Plot (weight on y axis, diet on x axis)
Graph Boxplot (y

= weight, x = diet)

Answer

A possible answer is shown below - although yours may look very


different:
C

Data display
Row
1

2
3
4
5

Diet
1
1
1

2
2

Weight
24.7
23.3
23.7
20.4
20.9

65

66

I I

PLANNING AN EXPERIMENT
2
3
3
3
4
4
4
5
5
5
6
6
6

6
7
8
9
10
11

12
13
14
15
16
17
18

21.4
16.6
17.5
17.0
18.0
17.2
17.0
19.0
18.3
18.7
20.5
19.7
19.6

The results are shown in Figures 4.5 and 4.6. Note that Figures 4.5 and
4.6 are shown exactly as they first appear on the screen. They may be
edited, for example to add a title, to add units to the vertical axis and to
make the label 'Weight' horizontal. A 'double click' on the graph produces
the editing toolbars.
25
24 23 -

0
0
0

22 0>21 -

-c

~ 20
19
18
17

0
0
0

0
@

16
2

80

0
0
0

Diet

Figure 4.5

High-quality plot of weights of mice on six diets.


25
24 23 ~

-c

22 -

0>21 -

~ 20 -

Q
E3

19 18 -

B D

17 16

Figure 4.6

Die t

El

High-quality boxpJot of weights of mice on six diets

Accounting for background


variation and constructing
a model

Experimental evidence about possible differences between treatments can


only be evaluated if we first define the model we are constructing to
account for the variation in our observations. In this chapter we will see
how variation caused by treatments and the environment can be separated
from random variation in a model. Then we will start to explain how to
take account of environmental variation in an experiment. In Chapter 6
we will use the model to analyse our results.
5.1

SOURCES OF VARIATION

We design and carry out an experiment because we want to discover the


important causes of variability within the system we are studying.
For example, in Chapter 4 we described an experiment to compare the
effect of four treatments on the species richness of field margins:

With flower seeds


No flower seeds

Cutting once a year

Cutting twice a year

FI
NFl

F2
NF2

We wanted to find out whether cutting the vegetation once a year instead
of twice affects the numbers of spiders which live on the site. Do they
prefer vegetation which has been seeded with wild flowers to the natural
grasses?
It is important to realize that even if we did not impose any treatments
in the experiments we would find that the number of spiders of a particular
species would not be the same in each quadrat. The natural variation
across the site will mean that some plots have damper or more fertile soil
or are more protected from wind and so they may have more vigorous
grass growth. One quadrat may, by chance, contain many poppies because

68

1 1.. 1_ _ _ _
A_C_C_O_U_N_T_I_N_G_F_O_R_B_A_C_K_G_R_O_U_N_D_V_A_R_I_A_T_IO_N
_ _ _ _---I

the grass had been dug up by rabbits the previous year so the poppy
seeds which had been dormant in the soil for years were able to
germinate at last. The number of spiders in a quadrat will be affected
by these differences in the vegetation in different quadrats. Such natural
variability from one quadrat or sampling unit to another is called
random variation.
In a laboratory experiment such variation is often quite small. For
example, the effect of different concentrations of a chemical on the growth
rate of a bacterium may be studied by adding the chemical solutions to
inoculated agar in Petri dishes which are then kept at a constant
temperature. The conditions within each Petri dish will be extremely
similar, apart from the deliberately imposed treatments (the chemical
concentrations) being different. In contrast, in a field experiment this
random variation is often very large so the quadrats and hence whatever is
measured within them will differ in many ways from each other, apart
from the deliberate differences introduced by the cutting and sowing
treatments.
That is why if we wish to determine the factors which influence say the
number of spiders per plot we need to set up a model of our system. This
must separate background or random variation from that caused by the
treatments. Let us look at the results we might collect from our field
experiment. First we count the number of spiders in each 1 m x 1 m
quadrat in our plot. We then add together the values from each of the
three quadrats per plot and divide by 3 to give the number of spiders per
square metre. This is because it is the plots which were randomized and so
they are the experimental units. Recording three quadrats within each plot
is like subsampling the plots. It gives us more information and allows us
to obtain a better estimate of the number of spiders per square metre of
plot than if we had only recorded one quadrat. However, because each set
of three quadrats are within one plot they are not independent
observations and therefore should not be used separately in the analysis of
variance that is described below. Thus we can summarize the variation in
spider numbers across the 16 plots as shown in Table 5.1.
Table 5.1

Number of spiders per plot

Replicate

1
2
3

4
Mean

Treatment
FI

F2

NFl

NF2

21
20
19
18
19.5

16
16
14
14
15.0

18

14

15
16
16.5

12
13.0

17

13
13

T_H_E_M_O_DE_L______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

The mean number of spiders per plot in Table 5.1 is not the same in all
plots. How can we make sense of the variation in terms of our model? The
variation might be caused by the effects of the four treatments or it might
represent random variation or, more likely, it is composed of both
elements.
If we want to predict how many spiders we would find on a plot which
has received a certain treatment, this experiment provide us with an
estimate. It is the mean of the values from the four plots which received
that treatment, the bottom row in Table 5.1, and is called the treatment
mean.
We find that the mean number of spiders on plots which received wild
flower seed and were cut once (F1) was 19.5, whereas on those which were
not sown and were cut twice a year (NF2) the mean was only 13.0. Not
every plot receiving treatment F1 however will contain 19.5 spiders (this is
simply an average - we cannot have half a spider in reality). There is a
considerable amount of random variation around the treatment mean.
Some plots have more and some less: we cannot say why they differ, except
that the differences are caused by chance.

5.2 THE MODEL

We can set up a model which describes these ideas. If we want to predict


the number of spiders on a plot the simplest estimate would be the mean
of all 16 values. This is called the grand mean. Here it is 16.0.
We can do better than this. We observe that if a plot is in treatment
F1 its mean number of spiders will be greater than the grand mean. If we
calculate the difference between the treatment mean and the grand mean it
is 19.5 - 16.0 = 3.5. In contrast, for treatment NF2 this difference is
13.0 - 16.0 = -3.0. The expected number of spiders in a plot of a
particular treatment is called the expected value. We can represent this
more generally as:
expected
value

grand

= mean

difference between treatment

+ mean and grand mean

This simple model predicts that each of the four plots in treatment F2 is
expected to have 16 + (15 - 16) = 16 + (-1) = 15 spiders per square
metre. However they do not, so the model needs further work. We can find
out by how much our model fails to fit our observations on each plot in
treatment F2 (Table 5.2).
Two of the replicate plots of this treatment (Table 5.2) have observed
numbers greater than the mean for the treatment and two have values
which are less. The differences between observed and expected values are
called residuals. They represent random variation. Residuals can be

69

70

ACCOUNTING FOR BACKGROUND VARIATION


Table 5.2 Observed values, expected values and residuals
Replicate

Observed
number

Expected
number

16
16
14
14
15

15
15
15
15
15

2
3

Mean

Difference

= Residual
I
I
-I
-I

positive or negative. They always add up to zero for each treatment and
must also have a mean of zero.
A simple model to explain what is going on in our experiment is:
observed number of spiders per plot (observed value) = expected number of
spiders per plot + residual. In more general terms:
observed value = expected value + residual
Since we already know how to calculate an expected value (see above) we
can include this information as well to give the full equation:
observed _ grand difference between treatment
.d I
value
- mean + mean and grand mean
+ res} ua
We can make this clearer by using the term treatment effect to represent
the difference between the treatment mean and the grand mean:
observed value = grand mean + treatment effect + residual
Note that both the treatment effect and residuals may be positive or
negative.
Box 5.1
The model using symbolic equations
Some textbooks use equations to represent the same idea:

Y;j

= ji + tj +

eij

The small letters i and j are called subscripts. The letter i stands for
the treatment number. In our experiment we could replace it by 1,2,
3 or 4. The letter j stands for the replicate number. We could use 1,
2, 3 or 4 here. If we wished to consider the number of spiders in
treatment 2, replicate 3 we would replace these SUbscripts of y (the
observed value) by 2 and 3 respectively.' Thegraad mean is
represented by y with a bar above it. The treatment effect is given by
t with a subscript i which represents the appropriate treatment. In
our case we would use 2 for treatment 2. Finally, the residual is

B_LO
__
C_K_IN_G
____________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

represented by the letter e, which again has subscripts for the


appropriate treatment and replicate because the random variation
differs from plot to plot.
Note that the letter e is used to represent the residual because an
alternative name for the residual is the error. This does not mean
that we have made a mistake. It comes from the Latin word errare
which means 'to wander'. It conveys the idea that random variation
represents 'chance wandering above and below the treatment mean'.
You may also come across the term fitted value or fit instead of
expected value, but they are the same thing.
The simple model we have constructed splits up the variability in our
data into two parts: that which can be accounted for (expected value) and
that which cannot be accounted for (residual). Residuals can be thought
of as what is left (the residue) after we have done our best to account for
the variation in our data. Another way of thinking of this is to consider
how we can control some sources of variation (the residuals) which are
due to randomness.
5.3

BLOCKING

There are four replicates of each treatment in our experiment and the 16
plots were arranged at random with four of them on each side of the field
(see Figure 4.3b). However, the four sides of the field may well provide
slightly different environments. We will now see how to take this
information into account in a revised layout known as a randomized
complete block design and so reduce random variation and improve our
ability to detect differences between treatments.
For example, the field may be on a slope; the side at the top may have
drier or sandier soil than that at the bottom; perhaps a tall hedge runs
along one side of the field, whereas there may be only a low fence along
the other side. With the completely randomized distribution of treatments
in Chapter 4 it is very likely that we will have allocated two replicates of
a particular treatment to the field margin at the top of the slope and none
to the margin at the bottom (Figure 4.3b). (Can you calculate the
probability of each treatment being present by chance only once on all
four sides?) Such differences can be important because, for example, in a
dry season, grass growth will be less vigorous at the top of the slope and
this may mean that low-growing wild flowers have a better chance of
becoming established. Thus, if our results from the treatment in which we
sow flower seeds and cut once (FI) show it to have a large number of
herbs, this may partly be because of its over-representation on the

71

72

IL____A_C_C_O_U_N_T_I_N_G_FO_R_B_A_C_K_G_R_O_U_N_D_V_A_R_IA_T_I_O_N_ _ _---'
favourable ground at the top of the slope. If we were to carry out such
an experiment many times, such effects would even out since perhaps next
time the same treatment might be over-represented at the bottom of the
slope. However, if we have only the resources for one or two experiments
we need to find a way of overcoming variations like this.
Ideally we want one replicate of each treatment to be on each of the four
sides of the field. This type of design is called a randomized complete block.
Each side of the field is a 'block' and each block contains a complete set
of all the treatments within it. Each block is selected so that the conditions
are even or homogeneous within it but that the conditions differ between
one block and another. So in our field the top block is on slightly sandier
soil and the block on one side is more shaded because of the hedge.
However, these differences should not be too great. For example, one
block can have a higher soil moisture content than another, but if one
block is a bog then the other must not be a desert. The whole experiment
should be carned out on a reasonably uniform site.
How should we allocate the four treatments to the four plots within each
block? This must be done using random numbers (Figure S.Ia). This
ensures that each treatment has an equal probability of occurring on each
plot. We number the four plots in a block from 1 to 4. Then we use the
random number button on our calculator. If the last digit is 3 we allocate
treatment FI to plot 3; if number 2 appears next we allocate treatment F2
to plot 2; number 1 would mean that treatment NFl goes on plot 1,
leaving plot 4 for treatment NF2 (Figure S.Ib).
Such blocking can also be helpful in enabling us to apply treatments
(b)

(a)

Randomized complete
block design
Block 1
F2 I NF21 Fl I NFl
N
F

~F

N
F

2
l- Block 2

+.

Block
N
4
F

F
1

1-

F
2

F2 INFll Fl

Block 3

I NF2

F
2

NF1

3rd
2

2nd

1st
4th

Plot
numbers

NF2
Treatments

Order of
obtaining the numbers
1-4 at random

Figure 5.1 (a) Field layout of plots. (b) Allocation offour treatments to block 2.

I I

BLOCKING
(a)

(b)

Block

Block
2

III]
Figure 5.2

mSown

Recorded

(a) Sowing - day 1. (b) Recording - person 1.

and to record results sensibly and fairly. Just as we have seen how to
associate environmental variation with blocks, so we can do the same with
our own behaviour. Perhaps it is possible to sow only four plots with
flower seeds in a working day, so it will take two days to sow the eight
plots required. We should sow two plots in each of two of the blocks on
the first day (Figure 5.2a) rather than sowing only one plot in all four
blocks. Then, if it rains overnight and the soil is too wet to allow the
remainder to be sown until a few days later, any differences in the
establishment of plants from the two different times of sowing will be
clearly associated with blocks. The same applies to recording the species
present. There are 48 quadrats to be recorded. If two people are sharing
the work they will have different abilities to see and to identify species
correctly. One person should record results from blocks 1 and 2 and the
other from blocks 3 and 4 (Figure 5.2b). This means that each treatment
will have half of its data recorded by each person so that any differences
in the recorders' abilities affect all treatments similarly on average. In
addition, we can account for differences between recorders; this becomes
part of the differences between blocks.
Blocking should be used wherever there may be a trend in the
environment which could affect the feature in which you are interested.
For example, in a glasshouse heating pipes may be at the rear of the bench,
so one block should be at the rear and another at the front of the bench.
Even in laboratories and growth cabinets there can be important gradients
in environmental variables which make it worth while arranging your plots
(pots, trays, Petri dishes, etc.) into blocks. It is common to block feeding
experiments with animals by putting the heaviest animals in one block and
the lightest ones in another block. This helps to take account of differences
in weight at the start of the experiment.
The concept of blocking in an experiment is similar to that of stratifying

73

74

I L-I____A_C_C_O_U_NT_I_N_G_F_O_R_B_A_C_K_G_R_O_U_N_D_V_A_R_I_A_T_IO_N_ _ _

----.J

the samples taken in a sampling exercise. However, in an experiment we


impose different treatments on randomly allocated plots within each block
and record their effects. In a sampling exercise we simply record what is
present in randomly allocated quadrats within each stratum, without
imposing any treatments.
In a randomized block design we can reduce the random variation in
our data by accounting for variation between blocks as well as variation
between treatments. Just as plots in one treatment may tend to have more
or fewer spiders than the grand mean (positive or negative treatment effect)
so plots in one block may tend to have more or fewer spiders than the
grand mean (positive or negative block effect). Perhaps a block at the
bottom of the hill may be damper and therefore, whichever treatment a
plot in that block receives, the vegetation grows more vigorously and tends
to be more attractive to the spiders.
We now have an extra classification in our model:
.
observatIOn

grand

treatment

= mean + effect
Y;j

block

+ effect + resIdual

= y + ti + bj + eij

Remember that the treatment effect is the difference between treatment


mean and grand mean. The block effect is defined similarly as the
difference between the block mean and the grand mean.
The observations of the number of spiders per plot are shown again in
Table 5.3. This time we have calculated both treatment and block means.
The grand mean is shown at the bottom right of the table. It is 16.0. This
has not changed.
We can now use our model to calculate a table, of the same layout, but
showing expected values:
expected value = grand mean + treatment ,effect + block effect
The block and treatment means are as we calculated them from our data.
We are now predicting the expected value for each plot from our model.
Table 5.3

Block means and treatment means


Mean

Treatment

Block
Fl

F2

NFl

NF2

3
4

21
20
19
18

16
16
14
14

18
17
15
16

14
13
13
12

17.25
16.5
15.25
15.0

Mean

19.5

15.0

16.5

13.0

16.0

1
2

I I

BLOCKING
Table 5.4 Calculation of expected value for
each plot
Block

Treatment
Fl

F2

NFl

Mean
NF2

20.75 16.25
20.0

2
3
4

Mean

19.5

15.0

16.5

13.0

17.25
16.5
15.25
15.0
16.0

Let's concentrate on treatment Fl in block 1 (top left in Table 5.4). To


calculate the expected value we must know the grand mean (16) and both
the treatment and block effects. As before:
treatment effect = treatment mean - grand mean
= 19.5 - 16 = 3.5
Block effect is calculated in the same way as for the treatment effect:
block effect

= block mean - grand mean


= 17.25 - 16 = 1.25

So these give us:


expected value

= grand mean + treatment effect + block effect


= 16 + 3.5 + 1.25 = 20.75

We can now start to construct the table of expected values. Use the model
to calculate the expected values for treatment F2 in block 1 and for
treatment Fl in block 2. You should obtain 16.25 and 20.0 respectively
(Table 5.4).
There is a quicker way to calculate expected values once the first one
has been calculated. It also sheds light on what the model is doing. On
average all plots in treatment F2 have 4.5 fewer spiders than those in
treatment Fl (19.5 -15.0). To obtain the expected value for treatment F2
in block 1 we subtract 4.5 from the expected value for treatment Fl in
block 1. This gives 16.25. Similarly, to obtain the expected value for
treatment NFl in block 1 we add 1.5 (the difference between the treatment
means for F2 and NFl) to the expected value for F2 in block 1 and obtain
17.75.
The same idea works for calculating expected values in the same
column. On average all plots in block 1 have 0.75 more spiders than those

75

76

II

ACCOUNTING FOR BACKGROUND VARIATION


Table 5.5

Expected values per plot

Block

Treatment

Mean

Fl

F2

NFl

NF2

1
2
3
4

20.75
20.0
18.75
18.5

16.25
15.5
14.25
14.0

17.75
17.0
15.75
15.5

14.25
13.5
12.25
12.0

17.25
16.5
15.25
15.0

Mean

19.5

15.0

16.5

13.0

16.0

in block 2. Therefore the expected value for treatment Fl in block 2 is


20.75 - 0.75 = 20.0, etc. (Table 5.5).
Now we know what our model predicts. How good is it at explaining
variation in spider numbers? If it was a perfect fit we would find that the
tables of observed and expected values (Tables 5.3 and 5.5) were the same
as each other. This is very unlikely. Usually there are differences which
are the residuals. We remember that: residual = observed value - expected
value (Table 5.6).
The residuals in Table 5.6 seem quite small. This suggests that our
model is quite good at explaining variation in spider numbers between
plots; there is not much that cannot be explained either in terms of a
treatment or a block effect. But how big do residuals have to be before you
start to be concerned that the model is not a good fit? How do you decide
whether differences between the treatments are big enough to mean
anything? After all, we carried out this experiment to answer the question
'Is there any difference between the effects of the four treatments on spider
numbers' (because spiders are part of the wildlife that might benefit from
these treatments)? In this chapter we have constructed a model to describe
our experiment which seems to be relevant. This is the first step towards
answering our question, but in the next chapter we need to go a step
further and decide whether this model does fit the data satisfactorily and
how to reach a preliminary conclusion on the significance of the apparent
treatment differences.
Table 5.6

Residuals per plot

Fl
1
2
3
4
Mean

Mean

Treatment

Block

F2

NFl

NF2

0.25 -0.25
0.25 -0.25
-0.5
0.5
0
0
0.75
0.25 -0.25 -0.75
-0.5
0.5
0
0
0

0
0
0
0
0

EXERCISE

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

II

5.4 EXERCISE
Blocks and treatments

We need to consolidate the concept that each observation in an experiment


with blocks and treatments is predicted by which block and which
treatment the observation is taken from and that the difference between
this observation and the value predicted or fitted from the model is called
the residual and represents chance variation. We will do this by using a
simplified data set where the effect of cutting on the number of daisy plants
per plot has been recorded. The observations are as follows:

Block 1
Block 2
Block 3
Block 4
Block 5

None

Low

High

12
9
3
3
8

31
14
7
11
12

18
8
0
5
8

Mean

Change

Mean
Change

Answer

Work out the mean of each row and each column and check your working
by calculating the grand mean which should be both the mean of the
column means and the mean of the row means. Then work out the change
expected when moving from one block to another and from one row to
another. You should obtain:
None

Low

High

Mean

Block 1
Block 2
Block 3
Block 4
Block 5

12
9
3
3
8

31
14
7
11
12

18
8
0
5
8

20.3
10.3
3.3
6.3
9.3

Mean
Change

7.0
+8 --+

15.0
-7.2 --+

7.8

9.9

Change

.J,
.J,
.J,
.J,

-10
-7
+3
+3

Next, calculate the expected value for the top left cell in the table - using
the general principle:
fitted value = grand mean + block effect + treatment effect
where an effect is the difference between the relevant block (or treatment)
mean and the grand mean. You should obtain:
block 1 'none' fitted value = 9.9 + (20.3 - 9.9) + (7.0 - 9.9) = 17.4

77

78

II~_______A__CC_O_U__N_T_IN_G__F_O_R_B_A_C_K_G_R_O_U_N_D__V_A_R_I_A_T_IO_N______~
Then complete the table by using the 'changes' shown above, for example
block 1 'low' fitted value = 17.4 + 8 = 25.4
block 2 'none' fitted value = 17.4 - 10 = 7.4
You should obtain the following expected, predicted or fitted values based
on our model:
None

Low

High

Mean

Block I
Block 2
Block 3
Block 4
Block 5

17.4
7.4
0.4
3.4
6.4

25.4
15.4
8.4
1l.4
14.4

18.2
8.2
l.2
4.2
7.2

20.3
10.3
3.3
6.3
9.3

Mean

7.0

15.0

7.8

9.9

Finally calculate the residuals using:


residuals = observations - expected values
For example
top left cell, residual = 12 - 17.4 = -5.4
You should obtain the following residuals:

Block 1
Block 2
Block 3
Block 4
Block 5
Mean

None

Low

High

Mean

-5.4
l.6
2.6
-0.4
l.6

5.6
-1.4
-1.4
-0.4
-2.4

-0.2
-0.2
-l.2
0.8
0.8

0
0
0
0
0

Rows and columns sum to zero.


In the next chapter we will see that MINIT AB will calculate expected
and residual values for us and display them graphically to allow us to
interpret the effects of treatments and the validity of our model.

Analysing your results Is there anything there?

In Chapter 5 we saw how to construct a model to fit our experimental


design. Here we will use it to analyse the data from the experiment.
We want to know whether it is reasonable to conclude that the
treatments do affect spider numbers and the technique we will use is the
analysis of variance (often abbreviated to Anova). It was published by
R.A. Fisher in 1923 and it works by splitting up the variation in all the
data into pieces (components) which are attributable to (can be explained
by) particular sources (for example treatments or blocks).
Analysis of variance quantifies this variation by calculating sums-ofsquares. First we calculate the sum-of-squares for all the data taken
together. This is called the total sum-of-squares: the sum of the squares of
the differences between each observation and the grand mean. Do you
remember these from Chapter I? If we want to avoid the long calculation
involved we can enter all 16 observations (Table 6.1) into our calculator
on statistical mode; press the standard deviation button (O"n_I); square it to
find the variance; and multiply this by n - 1 (15) to find the total sum-ofsquares. This reverses the procedure needed to calculate the standard
deviation in section 1.4.4. We obtain the result 106.0.

Table 6.1 Observed values


Replicate

1
2
3

4
Mean

Treatment

Fl

F2

NFl

NF2

21
20
19
18
19.5

16
16
14
14
15.0

18
17
15
16
16.5

14
13
13
12
13.0

80

II

ANALYSING YOUR RESULTS

~----------------------------------------------------~

6.1

WHOLLY RANDOMIZED DESIGN

We will start by assuming that our data came from a wholly randomized
experiment with four replicates per treatment. We will incorporate the idea
of block variation later. Textbooks often refer to the analysis of a wholly
randomized design as one-way analysis of variance. This refers to the fact
that there is only one source of variation involved (treatments).
The analysis of variance splits up the variation:
total
treatment
residual
sum-of-squares = sum-of-squares + sum-of-squares
We have just calculated the total sum-of-squares; therefore if we can
calculate the residual sum-of-squares then we can find out the treatment
sum-of-squares which is what we require because it tells us how much of
the total variation is due to treatments. Therefore we need to know what
the 'residuals' are. As we saw in Chapter 5 they are the differences between
the observed and expected values and in section 5.3 we saw how to
calculate expected values (Table 5.5). Thus the values in Table 6.2 and 6.3
can be built up.

Table 6.2

Expected values

Replicate

Treatment
Fl

F2

NFl

NF2

19.5
19.5
19.5
19.5

15.0
15.0
15.0
15.0

16.5
16.5
16.5
16.5

13.0
13.0
13.0
13.0

Mean

19.5

15.0

16.5

13.0

1
2
3

Table 6.3

Residuals

Replicate

Treatment
Fl

F2

NFl

NF2

1
2
3
4

1.5
0.5
-0.5
-1.5

1.0
1.0
-1.0
-1.0

1.5
0.5
-1.5
-0.5

1.0
0
0
-1.0

Mean

W_H_O_L_L_y__R_A_N_D_O_M
__IZ_E_D__
D_E_SI_G_N____________~I

L -_ _ _ _ _ _ _ _ _ _ _ _

We can now calculate the sum-of-squares of the residuals by the same


method as for the total sum-of-squares. That is, we calculate the sum of
the squares of the differences between the residual values in each cell of
Table 6.3 and their grand mean (which is always zero). We obtain the
result 16.0.
We can now find the treatment sum-of-squares because we know the
other two quantities in the equation:
total
sum-of-squares

treatment

residual

= sum-of-squares + sum-of-squares

106.0 = treatment
160
sum-of-squares + .
(a)

No.
spiders
per
plot

20

10

o~~~~~~~~~~~~~~~

2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
F1
F2
NF1
NF2
Treatment

Figure 6.1

Replicate

(a) Treatments explain nothing. (b) Treatments explain everything.

81

82

I ~I_______________A_N_A_L_y_S_I_N_G_y__O_U_R_R__E_SU__LT_S______________~
Therefore,
treatment
sum-of-squares

total

= sum-of-squares = 106.0 -

residual
sum-of-squares

16.0 = 90.0

We see that in this case the treatments account for most of the total
variation. The experiment has been reasonably successful in that other
sources of variation have been kept to a minimum. This won't always be
the case however, so let's consider the two extreme situations we could
encounter: where treatments explain nothing or everything (Figure 6.1) .
Treatments explain nothing
If all the treatment means are the same, knowing the treatment mean
does not help you to predict the plot mean. Each treatment may contain
a great deal of random variation (Table 6.4) and the total variation
would be entirely explained by this variation within treatments.
In Table 6.4 the total sum-of-squares equals the sum-of-squares of
the residuals. The treatment sum-of-squares is zero. It is just as if we
had selected 16 values from one population at random and allocated
them as the treatment results. Note that there are many experiments
where a zero treatment effect is a very desirable result; suppose for
example you are testing for possible side-effects of herbicides on fish in
nearby streams.
Table 6.4

Treatments explain nothing

Replicate

Treatment
Fl

F2

NFl

NF2

1
2
3
4

16.0
15.0
17.0
16.0

15.0
17.0
16.0
16.0

16.0
16.0
17.0
15.0

17.0
16.0
15.0
16.0

Mean

16.0

16.0

16.0

16.0

Treatments explain everything


At the other extreme, if treatment effects explained everything then
the treatment variation would account for all the total variation and
the random variation would be zero.
In this situation, if you know the treatment mean you can predict
the plot mean perfectly (Table 6.5). There is no random variation within
any treatment. All residuals are zero because the observed and expected
values are equal in each plot. The total sum-of-squares equals the

WHOLLY RANDOMIZED DESIGN


Table 6.5

I I

Treatments explain everything


Treatment

Replicate
FI

F2

NFl

NF2

I
2
3
4

19.5
19.5
19.5
19.5

15.0
15.0
15.0
15.0

16.5
16.5
16.5
16.5

13.0
13.0
13.0
13.0

Mean

19.5

15.0

16.5

13.0

Total

78

60

66

52

treatment sum-of-squares. The sum-of-squares of the residuals (also


called the residual sum-of-squares) is zero.
So far we have calculated the treatment sum-of-squares by subtracting
the residual sum-of-squares from the total sum-of-squares. However, we
could have calculated the treatment sum-of-squares directly. We will do
this once, to show that the method works, since it can then be extended to
more complicated designs (Chapter 7).
The first stage is to look at how the treatment means vary. Our four
treatment means are: 19.5, 15.0, 16.5 and 13.0. To find the treatment
sum-of-squares we enter the treatment totals into the calculator on
statistical mode. These are 78, 60, 66 and 52. We then calculate their
variance (using the short-cut noted earlier) by pressing the standard
deviation button (<Tn-I) and squaring it. We then multiply by n - 1 (3).
This gives us the sum-of-squares of these four numbers (the treatment
totals), but it ignores the fact that each of these numbers is itself the
total of four observations (the four replicates). To obtain the treatment
sum-of-squares on a 'per-plot' scale (like the total and residual sums-ofsquares), we must divide our result by 4 (the number of observations
in each treatment total):
treatment totals 78, 60, 66, 52

standard deviation = 10.954


variance = 120
variance x 3 = 360
treatment sum-of-squares = 360/4 = 90
This is in agreement with our previous calculation for 'treatment sum-ofsquares', where we obtained it by subtracting the residual sum-of-squares
from the total sum-of-squares.

83

84

I 1L-______________A_N_A_L_y_S_I_N_G_y__O_U_R_R__E_SU__LT_S______________~
6.2

ANALYSIS-OF-VARIANCE TABLE

Whichever way we have calculated the treatment sum-of-squares we can


now put it into an analysis-of-variance table. The standard method of
presentation is as follows. We start by writing down the analysis-ofvariance plan. This gives the structure of the experiment in terms of the
sources of variation, together with their degrees of freedom:
Source of variation

df

Treatments
Residual

3
12

Total

15

(Some textbooks and computer packages, including MINITAB, refer to


Residual as 'Error'.)
The convention is that the total variation is always on the bottom line
with the residual line immediately above it. Sources of controlled variation
appear above them. There are 15 degrees of freedom (df) for the totalline
because there are 16 plots and df = n - 1. Similarly there are 3 df for
treatments because there are four treatments, so n - 1 = 3. Residual df are
obtained by subtracting 3 df from 15 df. They are the residue. Another
way of thinking about residual df is to spot that each of the four
treatments has four replicates. So within each treatment there are
n - 1 = 3 df which represent random variation. So the residual variation
has 3 df for each of the four treatments, giving 12 df.
It is very important to prepare an analysis-of-variance plan before
carrying out any experiment. Otherwise, we may find out, too late, that
the experiment either cannot be analysed, has too little replication to be
able to detect differences between population means, or has excessive
replication and so is inefficient.
We can now fill in the next part of the analysis-of-variance table (Table
6.6). First we put in the sums-of-squares calculated above. Then we have
a column which is called mean square. This is the conventional name for
Table 6.6

The analysis-of-variance table

Source

df

Sum-of-squares

Mean square

Variance ratio'

Treatments
Residual

3
12

90.0
16.0

30.0
1.3333

22.5

Total

15

106.0

Variance ratio = treatments mean square/residual mean square. Our value here is 30/
1.3333 =22.5.

A_N_A_L_y_S_I_S-_O_F_-V_A_R_I_A_N_C_E_T_A
__
BL_E____________~I

L -____________

variance when we are carrying out analysis of variance. We obtain it, as

usual, by dividing the sum-of-squares by degrees of freedom. We only need


to do this for the treatment and residual lines. So, the treatment mean
square = 90.0/3 = 30.0 and the residual mean square = 16.0112 = 1.25. The
treatment and residual mean squares (or variances) can now be compared
directly because both are expressed in terms of the number of independent
pieces of information which they contain (n - 1 = df).
If the treatment mean square was zero we would be in the first situation
in section 6.1 above - 'treatments explain nothing'. If the residual mean
square was zero, treatments would explain everything (the second situation
above). We can quantify where we are between these two extremes by
dividing the treatment mean square by the residual mean square. This
gives a ratio of two mean squares (or variances), a variance ratio. If
treatments explained nothing the variance ratio would be o. If our
observations had been selected from one population and allocated to
treatments at random we would expect a variance ratio of 1 since
treatment and residual mean squares would be equal. If treatments
explained everything the variance ratio would be infinity. The larger our
variance ratio, the more evidence we have that the treatments differ from
each other more than just by chance.
We now need some way of quantifying the strength of evidence which
this provides, which we do by a hypothesis test. It is part of the logic of
statistical method that we can only assess the extent to which evidence
rejects a hypothesis rather than asking to what extent our data agree with
a hypothesis. We therefore specify a hypothesis such as the following, that
the four treatments come from one population. In other words our
hypothesis is that the four treatments should have the same mean and if
our samples have different means this is only due to sampling error - they
differ only by chance. This is called a null hypothesis because it presumes
no effect. It is sometimes abbreviated to Ho. The alternative hypothesis
(HI) is that the four treatments do not all come from one population. If
this were true at least one of the treatments would have a different
population mean from at least one of the others.
Our problem is that, in carrying out an experiment, we are taking small
random samples from the populations of several treatments and
comparing the results. There is natural variation in all populations. Even
if all the populations were identical (the null hypothesis is correct) there is
a small chance that the biggest four values from a population might go
into one treatment and the smallest four into another in our samples. Thus
there would appear to be strong evidence from our experiment that the
population means differed, when in fact they are the same. It is far more
likely however that four large values from one treatment and four small
values from another is evidence of a real difference between the means of
the populations which we are sampling. But how much more likely?

85

86

I I

ANALYSING YOUR RESULTS

~----------------------------------------------------~

Degrees of freedom for treatments

2
df
for
residual

1
2

12

Figure 6.2

3.49

Part of the table of the F-distribution, 5% (Table C.3).

With a variance ratio of 22.5 in the Table 6.6, with what confidence
might we reject the null hypothesis of no difference between the four
treatments? If we are not using a computer program we will need to find
the appropriate table in the back of this textbook. The table we want is
called an F-table. By comparing our calculated variance ratio (22.5) with
the appropriate values of F in the table we can estimate the probability p
of obtaining our particular results if the null hypothesis of no difference
between the treatments is correct. If this probability is very small we
decide to reject the null hypothesis and instead conclude that the
treatments do not all come from one population.
However, what is the appropriate value (usually called the critical value
of F) with which we should compare our variance ratio? It is an accepted
standard to use the table of F labelled '5%' or 'p = 0.05' (p stands for
probability; it is a proportion between 0 and I). The complete F-table is
given as Table C.3. The numerous possible values within it depend on what
combination of degrees of freedom have been used in calculating the
variances of treatment and residual effects. We choose the column in the
F-table which has degrees of freedom for the item in which we are
interested - in this case the treatment effect. Here, this is 3 df. We choose
the row according to the degrees of freedom which go with the residual, in
this case 12 df. This value of F - 'for 3 on 12 df' - is 3.49 (Figure 6.2).
Our variance ratio of 22.5 is much greater than this. Therefore we may
conclude that we have evidence to reject the null hypothesis.
Because the table is labelled '5%' or 'p = 0.05' we can say that if the null
hypothesis were true (and the populations of the four treatments were the
same) there is a less than 5% chance of our obtaining our set of results by
a random selection from the popUlations. We are usually content to
conclude that this is so unlikely that we will accept the alternative
hypothesis, HI> and state that there is likely to be a difference between the
treatment population means.
Our variance ratio is so much bigger than 3.49 that it is worth while

RANDOMIZED COMPLETE BLOCK DESIGNS

L -____________________________________________________

II

checking the F-table which is labelled '1%' or 'p = 0.01' (Table C.4) or
even the one labelled '0.1%' or 'p = 0.001' (Table C.5). These two tables
give critical values of F of 5.95 and 10.8 respectively (for degrees of
freedom 3 on 12). Our variance ratio is even bigger than 10.8. Thus we can
conclude that there is very strong evidence to reject the null hypothesis.
This is because 'p = 0.001' tells us that there is only a one-in-a-thousand
chance of obtaining our particular set of experimental results if the
treatments really all have the same effect (and so the observations come
from only one population). This is such a small chance that we should
prefer to conclude that the treatments do not all have the same effect on
spiders.
Now we can state our degree of confidence in our conclusion in a
formal way. Here we are 99.9% confident that we should reject
hypothesis Ho and accept HI. Computer packages like MINITAB give
the value of p for each grouping in an analysis of variance from a builtin database, so we do not need to consult F-tables. This means that
we obtain exact probabilities. For example, 'p = 0.045' means we are
less confident in rejecting the null hypothesis than if we saw 'p = 0.012'.
Often however p values will be represented (particularly in published
papers) by asterisks or stars: * represents p < 0.05, ** represents
p < 0.01 and *** represents p < 0.001. However, unless you are
specifically asked to present probabilities in this way you should give
the exact probabilities. Note also that when results are described as
'statistically significant' this is simply a convention for describing the
situation when the p value is less than 0.05.

6.3

RANDOMIZED COMPLETE BLOCK DESIGNS

We can now introduce an extra line to the analysis-of-variance table. It


will represent the amount of variation accounted for by the blocks.
Textbooks often call this a two-way analysis of variance because there are
two classifications - treatments and blocks. By accounting for variation
caused by differences within the site through blocking we will reduce the
random or residual variation. Therefore the residual mean square will
decrease and the variance ratio for treatment effects will increase.
First, the analysis-of-variance plan:
Source of variation
Blocks
Treatments
Residual
Total

df

3
3
9
15

87

88

II

ANALYSING YOUR RESULTS

~----------------------------------------------------~

Because there are four blocks there are 3 df for blocks. We add together
3 df for blocks and 3 df for treatments and subtract the answer (6) from
the total df (15) to obtain the df for the residual (9). This is three fewer
than it was before because 3 df are now accounted for by blocks. Just as
the old df for the residual has now been split into df for blocks and new df
for residual so the sum-of-squares for the residual will be split: part will
now belong to blocks and the remainder will be the new, smaller, residual
sum-of-squares.
We must now calculate a sum-of-squares for blocks. The method is just
the same as for treatments. Our four block means are 17.25, 16.5, 15.25
and 15.0 and the block totals are 69, 66, 61 and 60. To find the block sumof-squares we enter the block totals into the calculator on statistical mode.
We then press the standard deviation button (Un-I) and square it to find
the variance. We then multiply by n - I (3). This gives us the sum-ofsquares of these four totals, but, as with the treatment sum-of-squares, it
ignores the fact that each of them is itself derived from four observations
(one plot from each treatment). To obtain the block sum-of-squares on a
'per plot' scale (like the total, treatment and residual sums-of-squares), we
must divide our result by 4 (the number of observations in each block
total):
block totals 69, 66, 61, 60
standard deviation (of block totals) = 4.243
variance = 18
variance x 3 = 54
block sum-of-squares = 54/4 = 13.5
We now include the block sum-of-squares in the analysis-of-variance table
(Table 6.7). The revised residual sum-of-squares is obtained by subtracting
13.5 and 90.0 (treatment sum-of-squares) from 106.0 to give 2.5.
The blocks mean square is obtained by dividing its sum-of-squares by
degrees of freedom as before. The revised residual mean square is then
obtained by dividing 2.5 by 9. The revised variance ratio for treatments is
obtained by dividing the treatments mean square by the revised residual
mean square to give 108.0. This value is then compared with the critical
Table 6.7 The analysis-of-variance table
Source
Blocks
Treatments
Residual
Total

df

Sum-of-squares

Mean square

Variance ratio

3
3
9

13.5
90.0
2.5

4.5
30.0
0.278

16.2
108.0

15

106.0

~________________M_IN_I_TA_B________________~I
value in F -tables for '3 on 9 df'. Our variance ratio is very much bigger
than the critical F value for p = 0.001. Therefore we have strong evidence
for rejecting the null hypothesis. A common way of expressing this in
scientific papers is to say that treatment means differ (p < 0.001).
The < sign means 'less than'.
The blocks mean square (4.5), though smaller than that for treatments,
is also very much bigger than the residual mean square. Therefore we have
strong evidence that the blocks are not merely random groups. They have
accounted for site variation well, in that plots in some blocks tend to
contain more spiders than plots in other blocks. This source of variation
has been identified and separated from the residual. Thus, the amount of
random variation which remains unaccounted for is quite small.
6.4 WHICH TREATMENT IS DIFFERENT?
So far we have concentrated on hypothesis testing. (Are the treatments all
the same?) This is commonly only one aspect of the problems in which we
are interested and not the most important. When we carry out experiments
we usually choose treatments which we are already reasonably sure will
have effects which differ from each other. Formally rejecting the null
hypothesis that there is no difference between the treatments is thus often
a routine matter. What is frequently of greater interest is the comparison
of the results for different treatments in the experiment. The treatment
means are only estimates based on the samples in our experiment, so we
need to calculate confidence intervals if we wish to know the range of
values within which we are 95% confident that:
1. the treatment population mean lies, or
2. a difference between two population means lies.

We will explore this in Chapter 7, together with an examination of


how we might ask more specific questions about treatment effects.
6.5 MINITAB
The effect of four different combinations of sowing and cutting on the
number of spiders will be investigated using MINIT AB.
6.5.1 The data
We type our data and code numbers for blocks and treatments (see below)
into columns in the MINITAB worksheet, check and edit it and then save
it onto a disk. We can ask MINITAB to print out a copy of our

89

90

II

ANALYSING YOUR RESULTS

~----------------------------------------------------~

information (File, Display Data). MINITAB provides the row number


on the left-hand side (column 0). Each treatment is given a code number 1
to 4 (column 1) and so is each block (column 2). This means that the
computer program knows the correct block and treatment for each
observation - in this case the number of spiders (column 3).

6.5.2

Row

treat

block

spiders

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

1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4

1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4

21
20
19
18
16
16
14
14
18
17
15
16
14
13
13
12

One-way analysis of variance

This assumes that the plots were not blocked and so ignores the block
codes in column 2 (Stat, Anova, Balanced Anova, Response = C3,
Model=CI, Options, Display Means for CI). This asks for the data in
column 3 to be explained in terms of the codes in column I and the means
for each treatment to be printed. MINIT AB responds by first summarizing
the model: the factor 'treat' is 'fixed' (controlled by us) and has four levels
which are coded 1, 2, 3 and 4.
Factor
treat

Type
fixed

Levels

Values

Analysis of Variance for spiders


Source
treat
Error
Total

OF

SS

MS

3
12
15

90.000
16.000
106.000

30.000
1.333

22.50

0.000

MINIT AB gives the heading 'F' to the variance ratio column (often
labelled VR in textbooks) because the variance ratio is compared with the
appropriate value in an F-table. The p value (0.000) is extremely small
0.001). The chance of obtaining our sample of results if all four
treatments are samples from the same population is extremely unlikely

~________________M_I_N_IT_A_B________________~I
(less than one chance in 1000). we conclude that the treatments do not all
have the same effect on spider numbers.
Means
treat
1
2
3
4

6.5.3

4
4
4
4

spiders
19.500
15.000
16.500
13.000

Two-way analysis of variance

Now we introduce blocking. Thus there are now two sources of variation
which we can identify, treatments and blocks (Stat, Anova, Balanced
Anova, Response C3, Model Cl, C2, Options, Display Means for
Cl, C2):
Factor
treat
block

Type
fixed
random

Levels
4
4

Values
1 2 3
1 2 3

4
4

Analysis of Variance for spiders


Source
treat
block
Error
Total

DF
3
3
9
15

SS
90.000
13.500
2.500
106.000

MS
30.000
4.500
0.278

F
108.00
16.20

0.000
0.001

Means
treat

4
4
4
4

2
3
4
block
1
2
3
4

4
4
4
4

spiders
19.500
15.000
16.500
13.000
spiders
17.250
16.500
15.250
15.000

The variance ratio for treatments* (F = 108) is now much higher than it
was in the one-way analysis. We have removed variation between blocks
from the random variation (error). The treatment mean square is therefore
now compared with a much smaller mean square for randomness (0.278).

* Note:

MINITAB labels the variance ratio column 'F', which represents the theoretical values
found in the F-table, against which the calculated variance ratio is compared.

91

92

I I~______________A_N_A_L_Y_S_IN_G__Y_O_U_R_R_E_S_U_L_T_S____________~
We now have even stronger evidence for rejecting the null hypothesis of
no difference between the effects of the four treatments.

6.5.4 Boxplots for treatments


As with the sampling data we can ask MINITAB to display the results
for each treatment as in Figure 6.3 (Graph, Boxplot, y = spiders,
x = treat). We can see how plots receiving treatment 1 have a high number
of spiders while those receiving treatment 4 have a low number. We could
also ask for a boxplot of the block means if these were of special interest.
21 20 -

I9 ~ 18 -

~ 17 -

8
B

~ 16 -

15 14 13 12 -

Figure 6.3

8
treat

High-quality boxpJot of spider numbers by treatment.

Plotting mean values against treatments

Each plot's number of spiders can be shown against its treatment on a


chart. The four treatments are coded 1,2,3 and 4 as in Figure 6.4 (Graph,
20
Ul

'-

r---

Q)

-0
0.

r-r---

Ul

o 10
c

ro

Q)

:L

Figure 6.4

treat

High-quality chart of mean number of spiders by treatment.

M_IN_I_TA_B________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Chart, Function, Mean, y = spiders, x = treat, Frame, Min and Max,


Yminimum=O).

Two-way analysis of variance, asking for residuals and expected values


When we carry out an analysis of variance we have to make certain
assumptions about the observations and about the residuals. If these
assumptions are untrue then the conclusions we reach are not valid. It is
vital that the replicates of each treatment were independent of each other
and that the experiment was randomized. We also assume that the
variability within each treatment is similar. This is because we use its
average (the residual mean square) to represent random variation. If this
is much too small for some treatments and much too large for others we
are breaking the rules. We can examine this assumption by plotting the
residual for each plot against its fitted or expected value. We should see no
pattern, simply a random arrangement of dots.
We now ask for MINITAB to calculate residuals and expected values
and store them (Stat, Anova, Balanced Anova, Response = spiders,
Model = block treat, Storage Residuals Fits):
Factor
treat
block

Type
fixed
random

Levels
4
4

Values
1 2 3
1 2 3

4
4

Analysis of Variance for spiders


Source
treat
block
Error
Total

OF
3
3
9
15

SS
90.000
13.500
2.500
106.000

MS
30.000
4.500
0.278

F
108.00
16.20

0.000
0.001

MINITAB labels the residuals RESII and fits FITS 1:


ROW
1
2
3
4
5
6
7
8
9
10
11
12
13

treat
1
1
1
1
2
2
2
2
3
3
3
3
4

block
1
2
3
4
1
2
3
4
1
2
3
4
1

spiders
21
20
19
18
16
16
14
14
18
17
15
16
14

RESIl
0.25
0.00
0.25
-0.50
-0.25
0.50
-0.25
0.00
0.25
0.00
-0.75
0.50
-0.25

FITS1
20.75
20.00
18.75
18.50
16.25
15.50
14.25
14.00
17.75
17.00
15.75
15.50
14.25

93

94

ANALYSING YOUR RESULTS

I I
14
15
16

4
4
4

13
13
12

2
3
4

-0.50
0.75
0.00

13.50
12.25
12.00

Plotting a histogram of the residuals and a graph of residuals against fitted


values

We now ask MINITAB to display these as in Figure 6.5. (Stat, Anova,


Residual Plots, Residuals = RESIl, Fits = FITSl):
We can ignore the top right plot in Figure 6.5 The histogram at the
bottom left is slightly skewed but not too worrying, considering that we
only have 16 values. If this followed a perfect Normal distribution the line
of points at the top left would be a perfect 45 line. The bottom right plot
shows that the variability in the residuals is similar irrespective of the size
of the fitted values. Therefore we know that our data 'obey this rule of
analysis of variance' and the conclusions we have reached from the
analysis are reliable.

6.6 EXERCISE
You have carried out an experiment to compare the yields (tlha) of four
varieties of a crop (VI to V4). The layout plan (as viewed from the air)
also shows the yield from each plot:

Residuals versus fits: spiders


I Chart of Residuals

Normal Plot of Residuals

2~-------~

0.7

---------IUCl1.551

ro

:::>
~

.,

~1

1i.,

0.2

If>

0:::

0::-0.3

V 1\v .c">,~
'-I

">.,......

"'.000

-I
- - - - - - - - - LCt"1.551

- 0.6 ' ' ' r _ - - r - - - r - - - - . - - - . '


2
1
0
1
2
Normal Score

-2~--~--r---~

>-7

0.7

0-

u6

ro

r-

&4

:::>
~
f-----,

~3

LL.2
1

-0.6

-0.3
0.2
Residual

15

Residuals vs. Fits

Histogram of Residuals
c:
illS

10

Observation Number

0.7

0.2

If>

III

~---------~

0:::-0.3

12 13 14 15 16 17 18 19 20 21
Fit

Figure 6.5 High-quality residual plots for data on spiders.

II

EXERCISE
Block I
Block 2
Block 3
Block 4
Block 5

VI 5.5
V3 6.3
V35.6
V45.3
V43.9

V36.6
V46.8
V2 5.1
VI 3.2
VI 2.7

V4 7.1
V26.0
VI 4.6
V34.7
V23.4

V26.4
VI 5.7
V46.7
V23.6
V34.0

Enter the data into a MINITAB worksheet. Column 1 should contain


block codes (numbers 1 to 5). Column 2 should contain variety codes
(numbers I to 4). Finally put the yields into column 3 and save the
worksheet.
Now carry out an analysis of variance using (Stat, Balanced Anova).
The Response is 'yield' and the Model is 'block' and 'variety'. Save
residuals and fitted values. Use the options screen to ask for the means of
'block' and 'variety' to be printed.
Finally - some graphics. To show the means: (Graph, Chart, Function,
Mean, Y = Yield, X = Variety, Frame, Min and Max, Min Y =0). To plot
the residuals: (Stat, Anova, Residual Plots, Residuals = ... ). Select the
appropriate residuals and fitted values and enter a title.

Answer
Row

block

variety

yield

1
1
1
1
2

1
2
3
4
1
2

5.5
6.4
6.6
7.1
5.7
6.0
6.3
6.8
4.6
5.1
5.6
6.7
3.2
3.6
4.7
5.3
2.7
3.4
4.0
3.9

3
4
5
6
7

8
9
10
11
12
13
14
15
16
17
18
19
20

2
2

2
3
3
3
3
4
4
4
4
5
5
5
5

4
1
2
3
4
1
2
3
4
1
2
3
4

Analysis of Variance (Balanced Designs)


Factor
block
treat

Type
fixed
fixed

Levels
5
4

Values
1 2 3
1

4
4

95

96

YOUR
RESULTS
I ~I_ _ _ _ _ _ _ _ _ANALYSING
______
___
_ _ _ _ _ _ _ _ _ _~

Analysis of Variance for yield


Source
block
treat
Error
Total

DF
4
3
12
19

SS

MS

25.6480
7.2920
1.0080
33.9480

6.4120
2.4307
0.0840

F
76.33
28.94

0.000
0.000

Means
block

yield

1
2
3
4
5

4
4
4
4
4

6.4000
6.2000
5.5000
4.2000
3.5000

treat

yield

1
2
3
4

5
5
5
5

4.3400
4.9000
5.4400
5.9600

The results are shown in Figures 6.6 and 6.7. We conclude that there is
very strong evidence that the variety population means are not all the same
(p < 0.001). The residual plots reassure us that the model is valid as the
residuals are approximately Normally distributed and show similar
variability over the range of fitted values.

6
~5
Cl)

>-4

'"-

c 3 -

co

.;!.2
2

3
variety

Figure 6.6

High-quality chart of mean yield of each variety.

E_X_E_R_C_I_SE____________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Residual plots for varieties

Normal Plot of
Residuals
0.4
_ 0.3
~ 0.2
-0 0.1
'iii 0.0
a.>-O.I
0::: -0.2
- 0.3
- 0.4

I Chart of Residuals

..

0.7 r=========~

.
..
.
. ...

-6

0.2
"X-O.OOO

Ul

a.>

0:::_

-1
0
1
Normal Score

-2

0.3

- 0.8

- - - - - - - - - LCL-'O.6BB7

10
Observation

..

r----

r---

a.> 4
::>

0-3

a.>

~2

r---

r---

-0.4 -0.2 0.0 0.2


Residual

Figure 6.7

20

Residuals vs.
Fits

Histogram of Residuals
>u 5
c

0.4

UCL-O.6BB7

co

0.4
_ 0.3
~ 0.2
-0 0.1
'iii 0.0 -+-'-----------1
a.>_ 0.1
0::: -0.2
'0.3

..

- 0 . 4 '----r--..,.:--,-----,r-----,---'
7
3
5
6
Fit

High-quality residual plots of variety data.

97

Consider the treatments


of the field

In Chapter 6 we concentrated on testing the null hypothesis that 'the


treatments come from the same population'. However if the results suggest
that they do not, how do you decide which are the most important
differences between them? In this chapter we will tackle this question by
finding out how to estimate confidence intervals for population treatment
means and for differences between them. Then we will see how a skilful
choice of treatments allows us to obtain more specific and unambiguous
information about the nature of the treatment effects.
7.1
7.1.1

CONFIDENCE INTERVALS
Treatment mean

As we saw in Chapter 2, a confidence interval is the range within which


we are, for example, 95% confident that the true or population mean lies.
We can calculate it by working out the estimated treatment mean from
our sample and then adding and subtracting from it an amount 't times
the standard error of the mean':
confidence interval = mean (t x SE mean)
The size of the confidence interval reflects the uncertainty that exists in
our knowledge of the population mean because we are extrapolating from
information from a few plots. If we had carried out the experiment in a
slightly different part of the field, or using a different group of individual
animals from the same population, we would have obtained a different
value for the sample mean.
Let's calculate the 95% confidence interval for a treatment where a field
margin received a seed mixture and was cut once a year (treatment Fl,
section 4.5). Our sample mean, from the four replicates of this treatment
in the experiment, is 19.5 spiders per plot.

C_O_N
__
F_ID_E_N_C_E__
IN_T_E_R_V_A_L_S______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _

The value of t is looked up in t-tables, for residual degrees of


freedom = 12 (from the analysis-of-variance plan) and for the required
confidence level (here the standard 95% or p = 0.05). It is 2.179.
Now we need the standard error of the mean. Unlike with a simple
sample we do not find the variance by using only the four observations for
a particular treatment. Rather, we use the residual mean square (variance)
from the whole experiment. This represents the pooled or average random
variation within all treatments, after accounting for differences both
between treatments and between blocks. It is derived from all 16
observations and therefore contains more information than if we used only
the four observations for each treatment separately. So it is a better
estimate of the overall variability. If we divide it by the number of
replicates of a particular treatment we find the variance of the mean of
that treatment. Then we take the square root to give the standard error of
the mean:
treatment mean = 19.5
t=2.179
residual mean square = variance = 0.278
number of replicates of the treatment = n = 4
variance of the treatment mean =varianceln = 0.278/4 = 0.0695
standard error of the mean = JO.0695 = 0.264
Therefore the confidence interval is:
19.5 (2.179 x 0.264)
or 19.5 0.575
or from 18.9 to 20.1
Thus, we are 95% confident that the population mean number of spiders
per plot is in the range 18.9 to 20.1 (Figure 7.1). There is only a 5% chance
that the population mean is larger than 20.1 or smaller than 18.9.
7.1.2

Difference between two treatment means

Treatment NF2 (no seed mixture, cut twice per year) has an estimated
mean of 13 spiders per plot. Our best estimate of the difference between
the two treatments is 19.5 - 13, i.e. 6.5 spiders per plot. We now wish to

<
18

19

>
20

21

Figure 7.1 The 95% confidence interval for the mean number of spiders per plot
on treatment Fl.

99

100

II

CONSIDER THE TREATMENTS OF THE FIELD

~----------------------------------------------------~

calculate a 95% confidence interval for the difference between the two
population means:
difference (t x standard error of the difference)
The difference is 6.5. As usual, t is for residual degrees of freedom (12),
so it is 2.179 as before.
The standard error of the difference has to take into account the
variance of the first mean and the variance of the second mean. If we are
estimating how far apart two population means are and we have uncertain
estimates of each of them, both amounts of uncertainty will have an effect
on the range of size of the difference.
Both treatments have four replicates. Therefore, in this case, the two
treatment mean variances are the same as each other. We add the two
variances of the means together:
0.0695 +0.0695 =0.139
We then take the square root, to give 0.373 as the standard error of the
difference between the two means. Putting the values into the formula, the
confidence interval is:
6.5 (2.179 x 0.373)
or 6.5 0.813
or from 5.7 to 7.3
We are therefore 95% confident that the mean difference in spider numbers
between these two populations is in this range. This can be put another
way: There is no more than a 5% chance that the mean difference lies
outside this range and, since the probabilities are equal that it lies above as
below, no more than a 2.5% chance that the difference is less than 5.7.
With these results, it is very unlikely that the mean difference is really zero.
Thus we can be very confident that the difference between the two
treatments is real and not caused by chance.
A common way of comparing treatment means after analysis of
variance is to ask whether the difference between them is greater than a
certain amount. If it is, then we can say that the two treatments are
significantly different - in other words that we are (usually 95%) confident
that they come from populations which have different means. The 'certain
amount' is called the least significant difference (LSD):
LSD=(tx standard error of the difference) = tx SED
So, in our example above
LSD =

x SED = 2.179 x 0.373 = 0.813

The difference between our two treatment means is 6.5. As 6.5 is greater
than 0.813 we can be 95% sure that the samples come from populations
with different means.

'-----_ _ _ _F_A_C_T_O_R_IA_L_ST_R_V_C_T_V_R_E_O_F_T_R_E_A_T_M_E_N_T_S_ _ _ _----'I

As 6.5 is very much greater than 0.813 we can use a value of t for a
higher level of confidence: t for 99.9% confidence and for 12df is 4.318
(Table C.2). So the new LSD = 4.318 x 0.373 = 1.611. As 6.S is greater
than 1.611 we can be 99.9% confident that the samples come from
populations with different means.
Because we have 3 df we are allowed to make three particular
comparisons between treatments to determine what is having the most
effect on spider numbers. In this well-designed experiment there is one set
of three comparisons which ,is most informative. We will now outline the
reasons for this.

7.2 FACTORIAL STRUCTURE OF TREATMENTS


We can think of a treatment as the experience which an animal or plant
(the experimental unit) receives from us. Some treatments may be the result
of several 'things' being applied. Thus we now need to introduce a new
technical term - factor. For example, in our experiment we have four
treatments. However these are in fact the four combinations resulting from
the application of two factors (sowing and cutting) which are each present
at different levels. Levels refer to how a factor is applied. In this case
sowing is applied at two levels (sown or unsown) and cutting is similarly
applied at two levels (cut once or cut twice) (Table 7.1). In other
experiments more levels may be involved. For example a drug may be
given at four levels (0, 2, 4 or 8 mg per patient) or animals might receive
one of three levels of feed supplement (0, 6 or 12 g per day). The zero level
is as important as any other level.
Returning to our vegetation management experiment, this has a 2 x 2
factorial structure (because each of the factors is present at two levels).
Since we have four replicates of each treatment the experiment has 16
observations (2 x 2 x 4).
To summarize:
factor = something which may have an effect (for example: herbicide,
moisture)
Table 7.1

Factorial structure of treatments

Factor 2 Cutting

Level I Cut once


Level 2 Cut twice

Factor I Sowing
Levell Sown

Level 2 Vnsown

Treatment I FI
Treatment 3 F2

Treatment 2 NFl
Treatment 4 NF2

101

102

II

CONSIDER THE TREATMENTS OF THE FIELD

~----------------------------------------------------~

Table 7.2 A 3 x 4 factorial, with 12 treatment


combinations
Fibre in diet

Protein in diet

low
medium
high
very high

low

medium

high

5 reps
5 reps
5 reps
5 reps

5 reps
5 reps
5 reps
5 reps

5 reps
5 reps
5 reps
5 reps

level = state of a factor (for example: present or absent; low, medium


or high)
treatment = a particular combination of one level of one factor plus
one level of another factor
It will help to remember that, when there is a factorial structure, the word

'treatment' can be replaced by the phrase 'treatment combination'.


Another example of a factorial structure is shown in Table 7.2. In this
case the number of levels is different for the two factors, being three for
fibre and four for protein. Each treatment combination then has five
replicates so the structure for this experiment can be summarized as a
'3 x 4 factorial' with five replicates.
Our vegetation management experiment has 3 df for treatments (one
fewer than the number of treatment combinations). If the variance ratio
for treatments is significant we have evidence to show that at least one of
the four treatments has a population mean which is different from the
population mean of at least one other treatment. This is ambiguous.
Immediately we want to ask more specific questions. We can ask three
independent questions because we have 3 df. Thus if we take the treatment
line from the analysis-of-variance table at the start of section 6.2 it can
be broken down as follows:
Source of variation

df

Sown versus un sown


Cut once versus cut twice
Interaction sow x cut
Four treatments

Each line represents a separate question which we can ask. First, in


general, is there evidence that the sown plots have a different number of
spiders per plot than the un sown ones? This component of the treatment
sum-of-squares represents the main effect of sowing averaged over the two
types of cutting. It ignores the fact that four of the eight sown plots were

F_A_C_T_O_R_I_A_L_S_T_R_V_C_T_V_R_E_O_F_T_R_E_A_T_M_E_N_T_s_ _ _ _----11

L..-_ _ _ _

cut once while the other four were cut twice, and similarly for the unsown
plots. It simply asks, in general, did sowing a wild flower mixture have
any effect?
The second line represents the main effect of cutting. In other words,
we compare the spider numbers on the eight plots which were cut once
with those on the eight plots which were cut twice. Does cutting once, in
general, have a different effect from cutting twice?
The third line represents a very important idea. It enables us to see
whether cutting once instead of twice changes spider numbers similarly
on all plots irrespective of whether they were sown with wild flower
seed or not. If this test is not significant we can assume that the two
factors are independent. If this test is significant we say that the two
factors interact; in other words the effect of cutting is not consistent,
rather it depends on whether seed has been sown or not. For example,
suppose that spiders liked to hunt on the wild flowers (because flies
are attracted to the flowers) but the second cut in the plots that receive
it removes the flower heads - hence no flies, hence no great abundance
of spiders. In this case an interaction would be expected - sowing wild
flowers would give more spiders than on unsown plots when plots were
also cut only once but this would not be so in the plots which were
cut twice.

7.2.1 How to split up the treatment sum-of-squares to assess main effects


and interaction
We could calculate the sums-of-squares for the sowing main effect using
the same method as for treatments. In this case we use two totals: that for
the eight plots cut once (36) and that for the eight plots cut twice (28).
We enter these into the calculator; square their standard deviation;
multiply the result by 1 (degrees of freedom for sowing) and divide it by 8
(as there are eight observations in each total). This gives a value of 64
for the cutting sum-of-squares. Following the same process for the totals
for the eight unsown and eight sown plots gives a sowing sum-of-squares
of25.
Inserting these into the analysis-of-variance table (Table 7.3) shows us
that the interaction sum-of-squares can then be calculated by subtracting
both the sowing and cutting sums-of-squares from the treatment sum-ofsquares:
treatment sum-of-squares = sowing sum-of-squares
+ cutting sum-of-squares
+ sowing
x cutting interaction sum-of-squares

103

104

II

CONSIDER THE TREATMENTS OF THE FIELD

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Table 7.3

The analysis-of-variance table

Source

df

Sum-of-squares

Sown versus unsown


Cut once versus cut twice
Interaction sow x cut

I
I
I

25
64

Four treatments

90
16
106

Residual
Total

12
15

Therefore
interaction sum-of-squares = treatment sum-of-squares
- sowing sum-of-squares
- cutting sum-of-squares.
interaction sum-of-squares = 90 - 25 - 64 = 1
We find the mean square for each of the three component lines of the
treatment line by dividing their sums-of-squares by their degrees of
freedom. As in this case df are always 1 the mean squares equal the sumsof-squares. If, as with the structure in Table 7.2, factors have more than
two levels then the df will not be 1 and hence mean squares will not equal
sums-of-squares. For example, in Table 7.2 the main effect for fibre in
the diet has 2 df as there are three levels.
Then each mean square is divided by the residual mean square from
the original analysis (Table 6.6) to produce three variance ratios which can
each be compared with the critical F value from tables for 1 on 9 df
(5.12). The annotated MINITAB printout in section 7.3 shows how the
computer package carries out all these calculations for us.

7.2.2 Why bother with a factorial structure for your experiment?


A factorial structure is a very efficient way of investigating biological
(and other) phenomena. The cunning trick is that there is 'hidden
replication'. For example, although there are only four replicates of each
of the four combinations of sowing and cutting, there are eight replicates
for 'with seed' and eight replicates for 'without seed' in the above
experiment. Also, we save time and money by carrying out one experiment
in which two factors are explored, instead of examining each separately.
Above all we have the opportunity to find out whether these factors act
independently of each other or whether they interact. If the former is true,
life is simple and we can make general recommendations like 'we should
sow flower seed to increase spider numbers'. If the latter is true it is
important to discuss main effects only in the light of their interactions - we

'--_ _ _ _ _ _ _ _
M_IN_I_TA_B_ _ _ _ _ _ _ _-----',

cannot generalize about sowing seed; its effect on spider numbers depends
on how frequently we cut the vegetation. A complete factorial structure
can have many factors, each at many levels but they must be balanced: all
possible combinations of factors at all levels must be present. If any are
missing then the analysis described above will not work. However,
MINITAB can be used to analyse such unbalanced data whether the lack
of balance is intentional or accidental (as when one or more plots are lost
because of an accident). An example of such an analysis is given in section
7.3.3.
A useful check on the degrees of freedom in the Anova plan is that the
number of df for an interaction can be found by multiplying together the
number of df for each of the component factors (hence 1 x 1 = 1 above or
2 x 3 = 6 in Table 7.2). A more obvious way of obtaining the number of
interaction df is to take away the df for all main effects (and any other
interactions) from the treatment df. Therefore in Table 7.2 this gives
11-2 - 3 = 6.

7.3
7.3.1

MINITAB
The effect of cutting and sowing on spider numbers

Factorial analysis of variance, using codes for sowing and cutting levels
We display the data:
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

treat

block

spiders

seeds

1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4

1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4

21
20
19
18
16
16
14
14
18
17
15
16
14

1
1
1
1
1
1
1
1
0
0
0
0
0

13
13
12

0
0
0

cut
1
1
1
1
2
2
2
2
1
1
1
1
2
2
2
2

(Stat, Anova, Balanced Anova, Response = spiders, Model = block


seeds I cut, Storage Residuals Fits, Options, Display Means of block
seeds I cut).
The symbol between seeds and cut is called the pipe symbol. Its meaning

105

106

I I~_______C_O__N_SI_D_E_R__T_H_E_T_R_E_A_T_M_E_N_T_S_O__F_T_H_E_F_I_E_L_D________~
here is: please examine the main effect of seeds, the main effect of cut
and the interaction between them.
Analysis of Variance (Balanced Designs)
Factor
block
seeds
cut

Type
fixed
fixed
fixed

Levels

Values

4
2
2

0
1

2
1
2

Analysis of Variance for spiders


Source
block
seeds
cut
seeds*cut
Error
Total

OF

SS

3
1
1
1
9
15

13.500
25.000
64.000
1.000
2.500
106.000

MS

4.500
25.000
64.000
1.000
0.278

16.20
90.00
230.40
3.60

P
0.001
0.000
0.000
0.090

The p values for seeds and for cut are both very small, showing that
there is very strong evidence that both factors affect spider numbers.
However, the p value for the interaction is greater thaiJ.0.05, showing that
there is no evidence for an interaction between the factors. The main
effects may be interpreted independently of one another: sowing a seed
mixture increases spider numbers as does only cutting once.
Means
block

1
2
3
4

4
4
4
4

spiders

seeds

0
1

8
8

14.750
17.250

cut

spiders

1
2

8
8

18.000
14.000

17.250
16.500
15.250
15.000

spiders

seeds

cut

0
0
1
1

1
2
1
2

spiders

4
4
4
4

16.500
13.000
19.500
15.000

We can plot the number of spiders in each plot against the two types
of sowing and cutting (Stat, Anova, Main Effects Plot, Factors Seeds Cut,
Response Data in Spiders). We can do the same for cutting. The unedited
results are shown in Figure 7.2.

---'I I

M_IN_I_TA_B_ _ _ _ _ _ _ _

L---_ _ _ _ _ _ _ _

Main Effects Plot - Spiders

18
17
Q)

~ 16
Ul

15

14
seeds

Figure 7.2

cut

Main effects plot for spiders (unedited).

To edit this graph, maximize its size and double-click to produce the
toolbars. Use these to amend the size, appearance, position and nature of
text and to add lines or shading. Also use them to delete unwanted items.
An example of edited graphs is shown in Figure 7.3.
Main effects of seeding and cutting on mean number of
spiders per plot

18
17
mean
number
of
16
spiders
per
plot 15

14
no

Figure 7.3

yes
seed mixture

no

yes
cut ting

Main effects plot for spiders (edited).

7.3.2 The effect of amount of phosphate and contact with plant tissue on
bacterial growth
We have designed an experiment to investigate the importance of increasing amounts of phosphate (1, 2, 3 and 4mg per culture) and of
contact with living plant tissue (present or absent) on the extension of a
bacterial colony (mm per day).
The experiment was carried out by inoculating a carbohydrate medium

107

108

II

CONSIDER THE TREATMENTS OF THE FIELD

~----------------------------------------------------~

on Petri dishes with the bacterium. One set of each of the eight treatments
was established on each of three days, making 24 dishes in all. The
amounts of growth are analysed in MINITAB:
Row

expt

1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3

1
2
3
4
5
6
7
8
9
10
11
12
13
14

15
16
17
18
19
20
21
22
23
24

contact

phos

1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1

exten

1
1
2
2
3
3
4
4
1
1
2
2
3
3
4
4
1
1
2
2
3
3
4
4

10
6
13
11
14
20
16
22
12
10

treat
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7

13

13
14
14
14
18
14
10
12
10
10
14
16
18

(Stat, Anova, Balanced Anova, Response = exten, Model = expt contact I


phos, Storage Residuals Fits, Options, Display Means for expt
contact I phos).
Factor
expt
contact
phos

Type
random
fixed
fixed

Values

Levels

1
0
1

3
2
4

2
1
2

3
3

Analysis of Variance for exten


Source
expt
contact
phos
contact* phos
Error
Total

OF

SS

MS

2
1
3
3
14
23

4.000
2.667
166.000
57.333
68.000
298.000

2.000
2.667
55.333
19.111
4.857

0.41
0.55
11. 39
3.93

0.670
0.471
0.000
0.031

There is very strong evidence for phosphate having an effect


(p < 0.001) in general but this is modified by the presence or absence of
plant material (interaction p = 0.031).

I I

MINITAB
MEANS
expt

ext en

1
2
3

8
8
8

14.000
l3.500
13.000

contact

ext en

0
1

12
12

13.833
l3.167

phos

exten

6
6
6
6

10.333
12.000
14.333
17.333

1
2
3
4

contact

phos

exten

0
0
0
0
1
1
1
1

1
2
3
4
1
2
3
4

3
3
3
3
3
3
3
3

8.667
11.333
16.000
19.333
12.000
12.667
12.667
15.333

We need to plot the eight treatment means to see the interaction effect
- code numbers 1 to 8 in column 5 (Graph, Character Graphs, Boxplot).
The result is shown in Figure 7.4.
The bacteria not in contact with plant tissue (even-numbered treatments: 2, 4, 6 and 8) show a great response to increasing amounts of
phosphate. In contrast, the bacteria which are in contact with plant tissue
(odd-numbered ones: 1, 3, 5 and 7) show much less response. Perhaps this
is better displayed in an interaction diagram (Figure 7.5). We might
speculate on the biological reasons for such a difference.
We can plot a histogram of the residual~ and a graph of residuals
against fitted values (Figure 7.6). If, as here, the histogram follows an
approximately Normal distribution and there is no pattern in the other
plot this reassures us that the assumptions are reasonable.
Other particular comparisons between treatment means or groups of
treatment means can be made but this should be done with care. Ask a
statistician for advice before carrying out your experiment. He or she will
help you answer your questions efficiently and validly.
7.3.3 What to do if you have missing observations

Observations of zero
It is important to realize the difference between missing values caused by
accidents and values of zero in a data set. The latter may represent death

109

110

I LI_________C_O_N_S_ID_E_R__T_H_E_T_R_E_A_T_M_E_N_T_S__O_F_T_H_E_F_I_E_L_D________~
treat

-I

--I

1-

-I +

-I

1-

--I

1--

-I

7
8

+
+

1--

--+---+---+---+----il---+-exten
6.0
9.0
12.0
15.0
1B.0
21.0

Figure 7.4 Boxplot of eight treatments.

of an individual and this may be related to the treatment. If so, the 0 g is


as valid an observation as 5.3 g or 10 g and should not be omitted.
Problems can arise when there are quite a few zero values in a data set. If
these are clustered in one or two treatments it is best to exclude those
treatments from the analysis - since they are obviously different from the
rest.

Mistakes and disasters

These will happen. Despite taking care with your experiment, something
may well go wrong. You accidentally drop a test tube on the floor, the
tractor driver accidentally ploughs the wrong plot, or your helper throws

M_IN_I_TA_B_ _ _ _ _ _ _ _

L -_ _ _ _ _ _ _ _

-----'I I

Interaction Plot - Means of


Extension Growth

contact

----

19

0
I
0
I

c
rn

v 14

-----~--

-------~

phos

Figure 7.5 The effect of increasing amounts of phosphate and contact with plant
tissue on the extension rate of a bacterial colony.

away a paper bag containing the shoots from one plant before you have
weighed them. You now have the problem of one (or more) missing
observations. It is sensible to represent these by a '*' on your record sheet
(with a footnote describing the problem) so that you do not confuse them
with a real value of zero.
Residual plots from bacterial growth experiment
Normal Plot of Residuals
I Chart of Residuals
4 ~------------------,

....

(ij2

-6

Ui 0

.. ...

~-I

-2
-3
-2

..

'

-I
Normal Score

10

15

20

25

Observation Number

Histogram of Residuals

Residuals vs. Fits


4

9
>-8

0 7
c6
v5

5-4

v3 Ir----

f----,

.... 2

u.. 1

3
(ij2
-6 1
iii 0
~-I

'.

-2
-3

-3

-I
Residual

Figure 7.6

10

15

20

Fit

High-quality residual plots from bacterial growth experiment.

111

112

I LI_________C_O_N_S_ID_E_R__T_H_E_T_R_E_A_T_M_E_N_T_S_O__F_T_H_E_F_I_E_L_D________~
Since you have fewer observations you have less information and so
the experiment is less likely to be able to detect differences between the
populations which you are comparing. Another problem is that you now
have more information about some treatments than about others. A oneway analysis of variance can be carried out in MINITAB as usual. It takes
into account the differing replication between the treatments.
In a wholly randomized experiment we compared the effects of six
different plant hormones on floral development in peas. A standard
amount was applied to the first true leaf of each of four plants for each
hormone, making 24 plants in the experiment. Unfortunately, two plants
were knocked off the greenhouse bench and severely damaged: one from
hormone 2 and one from hormone 6. So we have data on the number of
the node (the name given to places on the stem where leaves or flowers
may appear) at which flowering first occurred for only 22 plants. This is in
C2 ofa MINITAB worksheet, with treatment codes in Cl.
ROW

hormone

node

1
2
3
4
5
6
7
8
9
10

1
2
3
4
5
6
1
3
4
5
1
2
3
4
5
6
1
2
3
4
5
6

49
53
49
53
51
58
52
50
57
56
51
54
52
54
51
54
48
57
49
51
54
55

11

12
13
14
15
16
17
18
19
20
21
22

(Stat, Anova, Balanced Anova, Response = node, Model = hormone,


Options, Display Means for hormone).
Factor
hormone

Type
fixed

Levels

Values

Analysis of Variance for node


Source
hormone
Error
Total

DF

SS

5
16
21

101. 008
70.083
171. 091

MS
20.202
4.380

4.61

0.009

I I

MINITAB

Means
hormone
1

node
50.000
54.667
50.000
53.750
53.000
55.667

4
3
4
4
4
3

3
4
5
6

The analysis of variance provides strong evidence to reject the null


hypothesis that all six hormones affect floral initiation similarly. If we wish
to calculate standard errors for treatment means we must remember that
they are based on different numbers of replicates. Remember the formula
IS:

SE mean =

residual mean square


n

For hormones 2 and 6, where we have only three replicates, the SE for
each mean is .J4.38/3 = 1.46; whereas, for the remaining hormones it is
.J4.38/4 = 1.095, which is considerably smaller.
Perhaps we realized that our glasshouse was not a homogeneous
environment and so laid out our experiment in four randomized complete
blocks. We now have a two-way analysis of variance. However, this is not
balanced:
Hormone

Block

1
2
3
4

49
52
51
48

53

49
50
52
49

53
57
54
51

51
56
51
54

58

*
54
57

54
55

If we want to compare the mean of hormone 2 with that of hormone 4


we have a problem. It may differ because the hormone has a different effect
but it may also differ because hormone 2 was not represented in block 2
which could be a location in the glasshouse which is especially good or
especially bad for plant growth. Here are the data with block and
treatment codes in MINITAB:
Row
1
2

3
4
5
6

block
1
1
1
1
1
1

hormone
1
2

3
4
5
6

node
49
53
49
53
51
58

113

114

I I

CONSIDER THE TREATMENTS OF THE FIELD


7
8
9
10

2
2
2
2
3
3
3
3
3
3
4
4
4
4
4
4

11

12
13
14
15
16
17
18
19
20
21
22

1
3
4
5
1
2
3
4
5
6
1
2
3
4
5
6

52
50
57
56
51
54
52
54
51
54
48
57
49
51
54

55

We can examine the data (Graph, Boxplot, y=hormone, x = node).


The result is shown in Figure 7.7.
If we try to analyse the data using the usual Balanced Anova command
in MINITAB we receive a surprise:

* ERROR * Unequal cell counts


This is because there isn't one representative of each treatment in each
block. To analyse these data we need to use a more flexible program.
MINITAB has a command 'glm' which stands for general linear model. If
this is used in place of Anova, all will be well (Stat, Anova, General Linear
Model, Response = node, Model = block hormone, Storage Residuals Fits,
Options, Display Means for block hormone).
58 -

<1)

-0

053 c

48 -

~ ~~~

B0
2

3
hormone

Figure 7.7 High-quality boxplot of number of nodes for each of six hormones.
Factor
hormone
block

Levels
6
4

Values
1 2 3
1 2 3

4
4

M_I_N_IT_A_B__________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Analysis of Variance for node


Source
hormone
block
Error
Total

DF
5
3
13
21

Seq SS
101.008
23.465
46.618
171.091

Adj SS

Adj MS

117.632
23.465
46.618

23.526
7.822
3.586

6.56
2.18

0.003
0.139

The puzzling thing about this response is that there are two columns of
sums-of-squares. These are 'sequential' (Seq) and 'adjusted' (Adj). When
an experiment is balanced the order in which we ask the terms in the model
(here hormones and blocks) to be fitted doesn't affect the outcome. There
is only one value for the sum-of-squares due to hormones, whether we ask
it to be fitted either before or after blocks. However, now we have an
unbalanced experiment the order in which the terms are fitted is very
important.
In the sequential sum-of-squares column, hormone has been fitted first
and then blocks. This was the order in which we asked for the two terms in
the glm line. However, in the adjusted sum-of-squares column each term
has the sum-of-squares appropriate to it if it were fitted last in the model.
Here, if hormones is fitted before blocks its SS is 101.0, but if it is fitted
after blocks its SS is 117.6. It is sensible to use the adjusted sum-of-squares
here. This is appropriate since it represents evidence for differences
between the hormone effects after taking into account the fact that some
hormones were not represented in all of the blocks.
Let's now look at the hormone means. Although the data are identical
with the figures used in the one-way analysis, the means are not the
same:
Means for node
block
1
2
3
4

Mean

52.17
55.14
52.67
52.33

StDev

0.7731
0.9981
0.7731
0.7731

hormone

Mean

StDev

1
2
3
4
5
6

50.00
55.35
50.00
53.75
53.00
56.35

0.9468
1.1270
0.9468
0.9468
0.9468
1.1270

For example, the mean for hormone 2 has increased from 54.667 to
55.35. This adjustment takes account of the fact that the representative of
hormone 2 in block 2 was accidentally lost. Block 2 was, in general, one
in which, for those treatments which were present, flowering was initiated

115

116

II~_________C_O_N_S_ID_E_R__T_H_E_T_R_E_A_T_M_E_N_T_S__O_F_T_H_E_F_I_E_L_D________~
Residual plots - hormones
Normal Plot of Residuals

I Chart of Residuals

3;---------------~

..

ro
::>1

"0

iii 0

r:========1

UClS.IJ8

;'1.001

If)

III

... -

0:::_1

-2

5
ro
::>
~O

-2

0::

-5

-I
o
Normal Score

- - - - - - - - - lClS.ZJS

'2

10
20
Obser va t ion Numbe r

Residuals vs. Fits

Histogram of Residuals
5

3~--------~

.--

'2

r--

ro
::>1

f---

"0

Vi 0
v

f---,

0::_

-l-.---~~-'-----j

-2

- '2

-I

2
Residual

Figure 7.8

49 50 51 52 53 54 55 56
Fit

High-quality residual plots of hormone experiment.

at a higher node number. It is only fair that we should adjust the value
for hormone 2 upwards. This method provides us with an estimate of the
results we might have expected to obtain if all hormones were present in
all blocks.
We plot a histogram of the residuals and a graph of residuals against
fitted values. These are shown in Figure 7.8. The residuals are slightly
skewed to the right but the variance is similar across the range of fitted
values.

7.4 EXERCISES
Confidence interval and least significant difference

(i) Calculate a 95% confidence interval for the population mean yield of
variety 4 in section 6.6.
(ii) Calculate a 95% least significant difference for the difference between
any two variety means in section 6.6. Is there evidence for a significant
difference between varieties 1 and 2?

EXERCISES

L-____________________________________________________

II

Answer

(i) Confidence interval


A confidence interval is the sample mean (tx SE mean). In this case the
sample mean is 4.2. The value of t for 12 df (error df) at 95% is 2.179. The
standard error of the mean is

SE mean = Jerror mean square/number of replicates

= J(0.084/4) = 0.l449
Therefore the confidence interval is
4.2 (2.179 x 0.1449) = 4.2 0.32

We are 95% confident that the population mean yield of variety 4 is


between 3.9 and 4.5 tlha.
(ii) Least significant difforence

The LSD is t x SE difference. Here t for 12 df at 95% is 2.l79. The


standard error of the difference is
SE difference = ./2 x error mean square/number of replicates

= J(2 x 0.084/4) = 0.205


Therefore the LSD is
t x SED

= 2.179 x 0.205 = 0.446

The difference between the sample means of varieties 1 and 2 is


4.90 - 4.34 = 0.56 tlha. This is greater than 0.446 tlha. So we conclude
that the two varieties come from popUlations with different means with
variety 2 having the higher yield.
Factorial analysis

An experiment was carried out to estimate the effect of fertilizer and


irrigation on the height of maize plants (cm). It was a completely
randomized design with four replicates of each of the four combinations
of: with and without irrigation, and with and without fertilizer. Enter the
data into a MINITAB worksheet, together with the codes for the two
factors (try Calc, set patterned data to enter the codes quickly) as shown:
Row
1
2
3
4
5
6

irrig

fert

0
0
0
0

0
0
0
0
0
0

1
1

height
17
16
15
18
14
16

117

118

CONSIDER THE TREATMENTS OF THE FIELD

7
8
9
10

1
1
0
0
0
0
1
1
1
1

11

12
13
14
15
16

0
0
1
1
1
1
1
1
1
1

16
14
13
13
14
12
21
20
19
18

Carry out an analysis of variance with a model accounting for the effects
of the two factors and their interaction and display the relevant means.
Produce an interaction plot (Stat, Anova, Balanced Anova, Interactions
Plot, Factors irrig fert, Response data height) and a residual plot and
interpret them.
Answer
Analysis of Variance (Balanced Designs)
Factor
irrig
fert

Type
fixed
fixed

Levels
2
2

Values

o
o

1
1

Analysis of Variance for height


DF
1
1
1
12
15

Source
irrig
fert
irrig* fert
Error
Total

SS
25.000
1. 000
64.000
16.000
106.000

MS
25.000
1. 000
64.000
1.333

18.75
0.75
48.00

0.001
0.403
0.000

Means
irrig
0
1

fert
0
1

irrig
0
0
1
1

height
14.750
17.250

8
8

height
15.750
16.250

8
8

fert
0
1
0
1

4
4
4
4

height
16.500
13.000
15.000
19.500

While in general irrigation increases height from 14.7 to 17.2 cm


(p = 0.001) this is misleading because these figures are averaged over with
and without fertilizer and there is strong evidence that these factors

-----'I I

E_X_E_R_C_IS_E_S_ _ _ _ _ _ _ _ _

L -_ _ _ _ _ _ _ _ _ _

Irrigation and fertiliser


irrigation
19.B
18.B

0
I
0

--- I

17.8
~ 16.8
([)

2::: 15.B

14.8
13.8
12.8

fertilizer

Figure 7.9

Interaction plot for irrigation and fertilizer experiment.

interact. In the absence of irrigation fertilizer decreases height (from 16.5


to 13.0 cm), whereas in the presence of irrigation height is increased (from
15.0 to 19.5 cm) (Figure 7.9), p < 0.001. The residual plots (Figure 7.10)
show that the residuals are only slightly skewed and have similar variance
over the range of fitted values. Thus the model is valid.

Irrigation and fertiliser


Normal Plot of Residuals

I Chart of Residuals

1.5
ro

-6

g;

0.5

- - - - - - - - - 1 UCL3.723

"DI

(f)

(f)

III

~-0.5

0
_I

a::: _2

-3
-4 l-===:::;===;::::==~ LCL-3.723

-1.5 .'-r---..,.---.---~
-1.5
'0.5
0.5
1.5
Normal Score

5
10
15
Observation Number

Histogram of Residuals

r----

Residuals vs. Fits


1.5

r--

;---

ro

-6

0.5

(f)

~-0.5
-1.5

-1.5

Figure 7.10

-0.5
0.5
Residual

1.5

13 14 15 16 17 18 19 20
Fit

High-quality residual plots for irrigation and fertilizer experiment.

119

120

I LI_________C_O_N_S_ID_E_R__T_H_E_T_R_E_A_T_M_E_N_T_S_O__F_T_H_E_F_I_E_L_D________~
General linear model analysis for missing data

You have discovered that unfortunately the plants in the last plot in your
experiment have been destroyed. Erase row 16 from your worksheet and
reanalyse the data using the general linear model option instead of
balanced Anova.

Answer
Analysis of Variance for height
Source
irrig
fert
irrig* fert
Error
Total

Seq SS
21. 376
0.665
66.692
13.000
101. 733

OF
1
1
1
11
14

Adj SS
27 . 923
2.077
66.692
13.000

Adj MS
27.923
2.077
66.692
1.182

F
23.63
1. 76
56.43

0.000
0.212
0.000

Means for height


irrig
0
1

Mean
14.75
17.50

StOev
0.3844
0.4151

fert
0
1

Mean
15.75
16.50

StOev
0.3844
0.4151

irrig* fert
0
0
1
0
1
0
1
1

Mean
16.50
13.00
15.00
20.00

StOev
0.5436
0.5436
0.5436
0.6276

Note that there are now only 15 observations and so the total degrees of
freedom have been reduced to 14. MINITAB calculates standard errors
(SE) for all the means but labels them StDev (standard deviation). The SE
for means with 8 replicates (fert 0 and irrig 0) is 0.3844. This is obtained
by taking the adjusted MS of 1.182, dividing by 8 and taking the square
root. The SE values of 0.4151 are obtained by dividing by 7 (for seven
replicates) and those of 0.5436 and 0.6276 by dividing by 4 and 3
respectively (according to the replication).

Relating one thing to


another

The previous chapters have been concerned with estimating the effects of
different treatments (for example frequency of cutting vegetation) and
testing the null hypothesis that they all have the same effect (for example
on the number of spiders). Here we will see what to do when our
treatments consist of increasing amounts or 'levels' of a factor (for
example water on plants, doses of a drug given to humans).
In this case we want to describe the nature of the relationship (if any)
between the amount of water and plant growth or between the dose of the
drug and the time taken for symptoms of disease to disappear.
The first thing to do is to plot our observations. The usual practice is
to put the response variable (plant weight or time for symptoms to
disappear) on the left-hand or vertical or y axis (also called the ordinate)
and the amount (of water or drug) on the horizontal or x axis (also called
the abscissa). The response variable is also known as the dependent variable

because it may depend on (be affected by) the amount of the substance
applied (the independent variable).
If there is no relationship between the two variables the plot of response
against amount will show a random scatter of points (Figure S.la). The
points do not form any pattern. If we find this, we do not need to carry
out any further analysis, as knowing the amount of substance applied does
not help us to predict the size of the response - the 'best' line through these
points is a horizontal one (Figure S.1b). However, we may see a pattern.
The simplest pattern is where the points seem to lie about a straight line
with a slope. The line may slope up (a positive relationship in which more
of x tends to produce more of y) or down (a negative relationship in which
more of x leads to less of y) (Figures S.lc and d). We then need to find a
way of describing the relationship so that we can make use of it in future
to predict just how much more (or less) y we will get for a given amount of
x.
Another possibility is that the points seem to lie on a curve. Perhaps

122

I I

RELATING ONE THING TO ANOTHER

(bl

(a)

Response

Response

Amount

Amount

(e)

Response

/-

(d)

Response

~

Amount

Amount

(el

(I)
Response

Response

Amount

Amount

(a) No relationship. (b) No relationship, showing line. (c) Positive


relationship, with line. (d) Negative relationship, with line. (e) Response
increases, then levels off. (f) Response decreases, then levels off.
Figure 8.1

plant growth is increased by giving them more water, but only up to a


certain amount (Figure 8.1e) of water, above which growth cannot
increase further. Similarly, the percentage body fat in a person may
become less with increasing exercise but there is a minimum percentage fat
which cannot be reduced (Figure 8.lf).
First we will consider how to test whether or not our results provide
evidence for a significant straight-line relationship (one which will help us
to explain or predict the response by knowing the amount of the factor
applied). Remember we have only obtained observations from a few
individuals but we wish to be able to extrapolate to the whole population.
The name given to this process of providing the best-fit straight line which
fits the points is linear regression.

8.1 LINEAR REGRESSION


Sir Francis Galton (1822-1911) thought that the mean height within a
family of children should tend to equal the mean height of their parents.
In other words that tall parents would produce tall children.
What he found was that the mean height of the children tended to

T_H_E__M_O_D_E_L__________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

'regress' (go back) towards the mean population height. The name
regression has since stuck to describe the relationship that exists between
two or more variables (in Galton's case, parents' height and height of
child; in the example above, plant growth and amount of water).
To develop the idea further let us take the example of increasing activity
affecting a person's percentage of body fat. First we have to have obtained
appropriate permission for the clinical trial to be carried out. We have to
assume that the people chosen are representative of the population we wish
to study (say, women in a county who are both aged between 30 and 40
and are overweight). We will ask them to take different amounts of exercise
and we will take blood samples monthly to record the concentration of a
chemical which is known to reflect percentage body fat very well. We will
select individual women at random, discuss the project with them and ask
them if they are willing to take part. If so, we allocate ten at random to
the lowest amount of exercise and ten each to take each of three higher
amounts, and ask them to record their actual amount of exercise.
Ideally, we do not inform the women whether they are taking a
relatively high or low amount of exercise (this is known as a blind trial).
This ensures that there is no psychological bias. Otherwise the knowledge
that they were, for example, taking the greatest amount of exercise might
subconsciously affect the manner in which they carried out the exercise
and, hence, the results. In the same way, if you are recording the results of
an experiment try to avoid thinking about what results you expect (want)
from each treatment as you are recording it.
We can now work out how linear regression operates by developing a
model as we did for analysis of variance (Chapter 6). First, imagine that
we knew the concentration of the chemical in the blood at the end of the
trial for each woman, but we did not know the amount of exercise which
had been taken. The sum-of-squares (of the differences between the
observations and their mean) can be calculated (see Figure 8.2, which
shows only a few observations, for clarity). It represents the total amount
of variability in our data.

8.2 THE MODEL


We now want to divide that variability into two parts: the amount that
we can account for by knowing which amount of exercise the patient
received and the rest (random variation). This is similar to dividing up the
variation in analysis of variance into treatment and residual effects.
However, we ask the question in a slightly different way. Specifically, if the
plot of the data shows an approximately 'straight-line' relationship with
a slope, how much of the total variability in 'concentration of chemical in

123

124

II

R_E_L_A_T_IN_G
__O_N_E_T_H_I_N_G_T_O__
A_N_O_T_H_E_R__________~

L ___________

chemical

~--~L---~---~.---.---,rl

mean

exercise

Figure 8.2 Calculation of sum-of-squares of y (total sum-of-squares).


the blood' (y) can be explained by a linear relationship with 'increasing
amount of exercise' (x)?
If the real relationship is a close one (the amount of exercise does have
a major effect on the concentration of the chemical in the blood) we will
find that almost all of the variation in y can be accounted for by knowing
x. In this case, the 'best-fit line' (the line that best describes the
relationship) will be close to all the points on the graph (Figure 8.3). If
there are many other factors which may affect the concentration of the
chemical in the blood (for example, age and body weight) then the scatter
of points about the line will be much wider.
Each woman has taken a particular amount (x) of exercise and we have
then noted the observed value of the chemical 'y'. The 'line of best fit'
calculated to describe this relationship will enable us in future to read off
an 'expected' or 'fitted' or 'predicted' value of y for any particular value of
x. This is unlikely to be identical to the observed value of y for that value
of x. The difference between the two is called a residual (Figure 8.4), just
as with analysis of variance. It tells us the discrepancy between our model
(the straight line) and the data:
residual = (observed y-fitted y)

Chemical
(y)

Figure 8.3 Best-fit line = regression line.

T_H_E__
M_O_D_E_L____________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Chemical
concentration
(g/lnre)

regression line
;
. - ~ observed value of chemical for
10 minutes exercise

-I T1-

+--

residual. observed minus expected

expected value of chemical


for 10 minutes exercise

10
Exercise (min)

Figure 8.4 Observed value, fitted value and residual.

A residual may be a positive or negative number. The method used to


obtain the line of best fit is to minimize the sum-of-squares of the distances
of the observations from the line, in a vertical direction. Another way of
saying this is: to minimize the sum-of-squares of the residuals. This is
known as a least-squares fit. The vertical direction is used because we are
trying to account for variation in y which is on the vertical axis.
Box S.l
Describing a straight line by an equation
First we put a y axis and an x axis on our graph. To specify a
particular straight line we need to know two things: the value of y
when x is zero (where the line cuts the y axis), and by how much the
line goes up (or down) for each increase of one unit in x (the slope
or gradient of the line).
You may already have met the equation of a straight line in
mathematics as y = mx + c. In this form the letter c represents the
value of y when x is zero and the letter m represents the slope. In
statistics it is standard practice to use the letter a to represent the
value of y when x is zero and the letter b to represent the slope. The
equation is thus written:
y

= a +bx

with a = estimate of the regression constant or intercept and b =


estimate of the regression coefficient or slope (Figure 8.5).
Notice that the terms of the equation are conventionally presented
with the intercept preceding the term containing the slope. This latter
symbolism is used because it is easily extended to more complicated
models like multiple regression, in which, for example, we might wish
to try and explain changes in percentage body fat by knowing both
amount of exercise and type of diet.

125

126

I LI___________R_E_L_A_T_IN_G__O_N_E_T_H_I_N_G_T_O__A_N_O_T_H_E_R___________
y

" ' - - b,. slope


a

=regression coefficient

=intercept =regression constant


x

Figure 8.5 The slope and intercept of a regression line.

Let us use MINITAB to examine the data from our clinical trial. We
have entered the x values (amount of brisk walking in minutes per day)
and the y values (concentration of a chemical in the blood in grams per
litre) into columns 1 and 2:
Row
1
2
3
4
5
6
7

8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

exercise
5
5
6
6
7
7
7
9
10
10
10
10
12
12
15
15
17
18
18
19
20
20
20
22
23
25
25
25
28
30

chemical
0.90
0.85
0.79
0.85
0.87
0.82
0.86
0.80
0.81
0.75
0.70
0.74
0.83
0.73
0.71
0.75
0.69
0.65
0.70
0.68
0.61
0.65
0.70
0.68
0.67
0.65
0.58
0.68
0.59
0.56

I I

THE MODEL
31
32
33
34
35
36
37
38
39
40

30
30
32
34
35
35
35
38
39
40

0.54
0.55
0.60
0.57
0.50
0.48
0.41
0.50
0.47
0.45

First we plot the observations to see whether they look as though a


straight line with a slope is a sensible model (Graph, Plot, y = chemical,
x = exercise). This is shown in Figure 8.6. Since this looks reasonable (it
looks as though the relationship might be summarized by a straight line
with a slope rather than being either a random scatter of points or better
summarized by a curve), we ask for the regression line to be fitted and
tested, with residuals and fitted values to be stored, so that we can ask for
them to be plotted later (Stat, Regression, Regression).
First, this produces the equation of the line in the form y = a + bx:
The regression equation is
chemical = 0.899 - 0.0113 exercise

Notice that the coefficient (b) has a negative value here (-0.0113). This
tells us that the line has a negative slope; it goes from top left to bottom
right.
Second, MINITAB tests whether the intercept (a) and the slope (b) are
significantly different from zero:
Predictor
Constant
exercise

Coef
0.89927
-0.0112574
0.9 0.8 -

CD

.~
E 0. 7 -

StOev
0.01344
0.0005899

t-ratio
66.89
-19.08

0.000
0.000

0
008
0
0
00
0

8
0

<1)

..c:

o 0.6 -

0
0

0 0 0 0 00 0
0
00
0
0

0.5 -

0
0
0

0.4 -

0
0

10

20

30

40

exercise

Figure 8.6 High-quality plot of chemical concentration against exercise


duration.

127

128

I LI___________R_E_L_A_T_I_N_G_O_N__E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
This printout shows that the intercept is significantly different from zero
(Constant, Coef, p < 0.000, which is much less than 0.05). So, when a
woman takes no exercise this regression line predicts that the concentration of the chemical in her blood will be significantly greater than zero the estimated value is 0.899 grams per litre. Also the slope of the line
(exercise, Coef) is significantly different from zero (p < 0.000 again). So
we can be very confident that there is a strong linear relationship between
chemical concentration and exercise.
Third, MINITAB gives the values of sand r2:
s=0.04009

R-sq=90.6%

R-sq(adj) =90.3%

The first (s) is the standard deviation (the square root of the residual or
error variance) and the second (R-sq) - pronounced R squared - is called
the coefficient of determination. It is the proportion of the variation in y
accounted for by variation in x (from 0 to 1, which is the same as from 0
to 100%). If r2 is 100% it indicates that the regression line goes through all
the points on the graph. Our model then explains all the variability in the
response variable: there is no random variation, all the residuals are zero.
In contrast, if r is 0% it is consistent with a random arrangement of points
on the graph. The larger the value of r2, the more useful the independent
variable is likely to be as a predictor of the response variable. (Do not
worry about the adjusted r2 (R-sq(adj value on the printout. It is only
helpful in more complicated models.)
Fourth, the analysis-of-variance table appears. This is of the same
general form as for when we are comparing the effects of several
treatments in an experiment. However, here the treatments line is replaced
by one called 'regression'. This represents the evidence for a linear relation
between the chemical and exercise.
Analysis of Variance
SOURCE
Regression
Error
Total

OF
1
38
39

SS
0.58518
0.06106
0.64624

MS
0.58518
0.00161

364.18

0.000

The evidence is very strong (p < 0.001 which is much less than 0.05). We
can reject the null hypothesis of no linear relationship between chemical
concentration and amount of exercise with great confidence. MINIT AB is
carrying out the same test here as the test shown above for whether the
slope of the line is significantly different from zero (the slope of a
horizontal line). We can see this by noting that the value of the variance
ratio (or F, 364.18) is the square of the t-ratio given above for the slope
(-19.08). This will always be the case. It is probably easier to concentrate
on the analysis of variance when interpreting the analysis.
Notice that the value of (90.6% or 0.906) can be obtained by dividing
the regression sum-of-squares by the total sum-of-squares.

A_S_S_U_M_P_T_I_O_N_S__________________~I

L -__________________

Regression of chemical on exercise


0.95
0.85

..~

...

m 0.75
u

.c:
U

. .",

0.65

.... ~
.....- ...
... . ....

0.55

.....

0.45

"'

....

"

0.899273 - 1.13E-02
R-Squaced 0.906

....

Regression

95t CI
0.35
10

20

exercise

Figure 8.7

30

40

High-quality plot showing fitted line and 95% confidence interval.

Next MINITAB notes any observations which are 'unusual':


Unusual Observations
Fit StDev.Fit Residual St.Resid
Obs. exercise chemical
10.0 0.70000 0.78670
0.00870 -0.08670
11
-2.22R
0.41000 0.50527
37
35.0
0.01084 -0.09527
-2.47R
R denotes an obs. with a large st. resid.

This tells us that observations number 11 and 37 have rather large


residuals. This can be useful in helping us to spot copying errors in our
data. We should always check such values. Here we will assume that we
find that they are correct.
We can ask for a plot of the data with the regression line superimposed
upon it, together with the 95% confidence interval for the line (Stat,
Regression, Fitted line). This is shown in Figure 8.7.
In interpreting the output it is important to realize that, just because
you have shown a significant relationship, this does not automatically
mean that x is causing the variation in y. Both could be responding
independently to some third factor (z) (see later). Also the model makes
assumptions about the nature of the data. If these assumptions are not
correct then the interpretation may again be invalid.

8.3 ASSUMPTIONS
1. Normal distribution
A Normal distribution of residuals is assumed, since they represent
random variation. There should be many residuals with very small
absolute values (near zero) and only a few with very large ones (far
from zero). We can test this by asking for residual plots in MINITAB

129

130

I I~__________R_E_L_A_T_I_N_G_O_N__E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
Residual plots: chemical on exercise
I Chart of Residuals

Normal Plot of Residuals

.. -..

_0.05
co
::>

~O.OO

...

(I)

o
~0.05

- 0 .1

..,...

, ./

_ 0.1
co

'iii 0.0
o

n::

- 0.1

~.,.---.,.---.,.---.,.---,...-J

-2

-1

Normal Score

10

>-10

30

40

Residuals vs. Fits

_ 0.05

C,)

....

co

c
o

g-

20

Observation Number

Histogram of Residuals

::>

UCLO.1145

-0

~.

r=========l

::>

::>
~

(f)

0. 0 0

. .. ..

-+-----:,....--..-=.--~'---........_=_1

o
~0.05

'-

u..

- 0. 10 ~---'----,--,-----.----r'
0.45

Figure 8.8

0.55

0.65

Fit

0.75

0.B5

High-quality residual plots for regression of chemical on exercise.

(Stat, Regression, Residual Plots). This is shown in Figure 8.8. We see


that the residuals (bottom left plot) have an approximately Normal
distribution (albeit a bit skewed to the left).
2. Homogeneity a/variance
Homogeneity of variance of residuals is assumed. The residuals should
not show a tendency to increase (or decrease) as x increases. The

x
Figure 8.9 Variability in y increases with x. This is the 'shotgun effect'.

FURTHER EXAMPLE OF LINEAR REGRESSION

L -____________________________________________________

II

variability in the residuals is similar for all four amounts of exercise


(bottom right plot). This is fine. There is no obvious pattern. A
common problem is shown in Figure 8.9 where there is much more
variability at high levels of x than at low levels. If this occurs consult a
statistician.
3. Independence
Every observation of y (chemical concentration in the MINITAB
example) must be independent of every other y. If, for example, we
included several women from the same household in our clinical trial
their observations will not be independent of each other because the
women may well have the same diet and be closely genetically related
to each other.
We need to ensure independence by carrying out our sampling or
experimentation in an appropriate way. For example, in an experiment
to study the effect of water on plant growth each observation should
come from a plant growing in a separate pot.
4. Continuity
Ideally both x and y variables should be measured on a continuous
scale (like kilograms, minutes or centimetres) rather than being counts,
proportions or ranks. Counts and proportions can also often be
analysed using regression. It is important to examine the graph of the
residuals and if in any doubt you should ask advice.
Other special techniques for ranked observations are introduced in
Chapter 9.
5. Absence of 'errors' in values
Ideally the x observations should be 'without error'. Obviously we want
our measurement of the x values to be as accurate as possible although
we realize that they will not be perfectly correct. If there is no bias,
slight mistakes either way in measuring x values will merely increase
the residual mean square and make the regression less precise.
However, if there is reason to suspect bias in the observation of x such
that it is generally recorded as too high (or generally recorded as too
low) this will affect the estimation of the slope.
8.4 FURTHER EXAMPLE OF LINEAR REGRESSION

In an experiment such as the clinical trial described above we allocated


women at random to take different amounts of exercise. However, the
principles of regression analysis extend to situations where we simply
observe pairs of values for each individual as they occur.
Let us take an example of data from a survey in which the amount of
lead in parts per million (y) on the vegetation alongside 19 roads depends
on the amount of traffic in hundreds of vehicles passing per day (x).

131

132

I I

RELATING ONE THING TO ANOTHER


80
0

70 -

00

60 -

0 0
0

~ 50 Q)

00
0

40 -

0
0

30 -

20 I0

0
0

0
0

40

140

90

traffic

Figure 8.10

High-quality plot of lead against traffic.

We can enter the data into columns 1 and 2 of a MINITAB worksheet:


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

lead
44
58
43
60
23
53
48
74
14
38
50
55
14
67
66
18
32
20
30

traffic
79
109
90
104
57
111
124
127
54
102
121
118
35
131
135
70
90
50
70

Now examine the printout in Figure 8.10 and interpret it.


The regression equation is
lead= -10.7 + 0.569 traffic

What is the intercept? What is the slope? [-10.7, 0.569] (Figure 8.11)
Predictor
Constant
traffic

Coef
-10.736
0.56893

StOev
5.813
0.05923

t-ratio
-1.85
9.61

P
0.082
0.000

Is the intercept significantly different from zero? [no, p > 0.05] Note that
the estimate of -10.736 must be meaningless since we cannot have a

F_U_R_T_H_E_R__E_X_A_M_P_L_E_O_F__L_IN_E_A_R__R_E_G_R_E_S_SI_O_N________~I

L -_ _ _ _ _ _ _

Lead

.,

Figure 8.11

40

Traffic

Relationship between lead and traffic.

negative value for lead. It may well be that if our sample included lower
levels of traffic we would find that the relationship was curved (Figure
8.11) .
Is the slope significantly different from zero? [yes, p < 0.05)

Residual plots for lead on traffic


[ Chart of Residuals

Normal Plot of Residuals


20

10

to
::J
"tJ

to
::J
"tJ

.... ...

(f)

<1l

..

c::

-10
-2

(f)

<1l

c::

000

l=====:;::====J
10

~ 5

&4

.--

f--

-12 -7

Figure 8.12

-2

Residual

lCl19.29

20

Residuals vs. Fits

i;'6

-10

Y\.yv
A b .-. x.a
V .
Observation Number

Histogram of Residuals
7

3
~2
1

0
-20

-1

Normal Score

<1l

10

UCl19.29

13

High-quality residual plots for lead and traffic data.

133

134

II

R_E_L_A_T_I_N_G_O
__
N_E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~

L ___________

Fitted line plot of lead against traffic

..

80
60
LJ

~ 40

y -\

0.7362 ' 0.5689

R-Squared 0.844

20
_

Regression l

95% CI

140

90

40

traffic

Figure 8.13 Relationship between lead and traffic .

Is this equation of a straight line a useful model to describe the


relationship between roadside lead and traffic numbers?
s=7.680

R-sq= 84.4%

R-sq (adj) = 83.5%

What proportion of the variation in lead concentration is accounted


for by amount of traffic? [84%]
Analysis of Variance
SOURCE

Regression
Error
Total

DF

SS

5442.0
1002.8
6444.7

17
18

MS
5442.0
59.0

92.26

p
0.000

With what confidence can you reject the null hypothesis of no linear
relationship between lead concentration and amount of traffic? [99.9%,
p < 0.001]

Do the residuals show an approximately Normal distribution? [yes]


(Figure 8.12)
Do the residuals show similar variability over the range of x values?
[yes]
A plot of the observations with the regression line superimposed
shown in Figure 8.l3.
8.S

IS

THE IMPORTANCE OF PLOTTING OBSERVATIONS

The lead deposition-traffic data set which we have just analysed was a
typical example of a relationship which is well modelled by linear
regression. However, it is possible for the computer to give you the same

'--_ _ _T_H_E_I_M_P_O_R_T_A_N_C_E_O_F_P_L_O_TT_IN_G_O_B_S_E_R_V_A_T_IO_N_S_ _ _---'I

equation (and the same values for p and r2) from very different data sets
for which the model would be inappropriate. This is an alarming thought.
It is important to plot the observations to identify such problems.
We will use four data sets devised by Anscombe (1973) to illustrate this
problem. The x values for data sets 1, 2 and 3 are the same and are in
column 1 while those for data set 4 are in column 5. The corresponding
responses are in columns named Yb Y2, Y3 and Y4:
Row
1
2
3
4
5
6
7
8
9
10
11

x123
10
8
13
9
11
14
6
4
12
7
5

y1
8.04
6.95
7.58
8.81
8.33
9.96
7.24
4.26
10.84
4.82
5.68

y2
9.14
8.14
8.74
8.77
9.26
8.10
6.13
3.10
9.13
7.26
4.74

y3
7.46
6.77
12.70
7.11
7.81
8.84
6.08
5.39
8.15
6.42
5.73

x4
8
8
8
8
8
8
8
19
8
8
8

y4
6.58
5.76
7.71
8.84
8.47
7.04
5.25
12.50
5.56
7.91
6.89

All plots have been produced using (Graph, Character Graphs, Scatter
Plot).
8.5.1 A straight-line relationship with some scatter (data set 1)
In carrying out the regression we can ask for MINITAB to calculate fitted
values and automatically put them into a column called 'FITS'. It will call

y1

x
x

10.0+
x
x
x

7.5+

x
x

5.0+

-+
4.0

+
6.0

Figure 8.14 Data set 1: YI against x 123

+
8.0

+
10.0

12.0

+-x123
14.0

135

136

I ~I____________R_E_L_A_T_IN__G_O_N_E__T_H_I_N_G_T_O__A_N_O_T_H_E_R____________~
the fitted values from the first analysis 'FITS1' and those from the next
analysis 'FITS2'. Instead of asking for ordinary residuals we have asked
for what are called 'standardized residuals' to be calculated and put
automatically into a column named 'SRES1'. The values of ordinary
residuals will depend on the units in which y was measured. If we ask for
them to be standardized by dividing them by an estimate of their
variability we can judge their size more easily. If we find that a
standardized residual is greater than 2 or less than -2 we should be alert
to possible errors in the data or to the possibility that we are not fitting a
sensible model.
Data set 1 is plotted in Figure 8.14.
The regression equation is
yl=3.00+0.500 x123

Note that the values of the intercept (3.0) and the slope (0.5) are the same
in this and the next three analyses. The regression line fitted is exactly the
same for allfour data sets.
Predictor
Constant
x123
s=1.237

Coef
3.000
0.5001
R-sq=66.7%

StDev
1.125
0.1179
R-sq (adj)=62.9%

t-ratio
2.67
4.24

P
0.026
0.002

Note that r2 (66.7%) is the same in this and the next three analyses.
Analysis of Variance
DF
1
9
10

SOURCE

Regression

Error
Total

F
17.99

MS
27.510
1. 529

SS
27.510
13.763
41.273

P
0.002

SRES1

1.2+
-

0.0+

x
-1.2+

-+----+---+---+---+---+-FITS1
5.0
6.0
7.0
8.0
9.0
10.0

Figure 8.15 Data set I: standardized residuals against fitted values.

'--_ _ _
T_H_E_I_M_P_O_R_T_A_N_C_E_O_F_P_L_O_TT_IN_G_O_B_S_ER_V_A_T_I_O_N_S_ _ _---',

Note that the probability (p = 0.002) is the same in this and the next three
analyses. We now ask for standardized residuals to be plotted against
fitted values (Figure 8.15). There is no pattern in the residuals. A straightline model seems satisfactory.
8.5.2 A curved relationship (data set 2)
Although the response increases from left to right, the rate is not steady,
the response levels off. Here, the fit of the regression line is poor. Predicted
values are too high at extreme values of x and too low in the middle.
Data set 2 is plotted in Figure 8.16.
The regression equation is
y2=3.00+0.500 x123
Predictor
Constant
x123
s=1.237

Coef
3.001
0.5000
R-sq=66.6%

StOev
1.125
0.1180
R-sq (adj) = 62.9%

t-ratio
2.67
4.24

P
0.026
0.002

Analysis of Variance
SOURCE
Regression
Error
Total

OF

SS
27.500
13.776
41.276

10

MS

27.500
1. 531

17.97

0.002

The poor fit of the regression line is confirmed by the pattern of the
standardized residuals when plotted against the fitted values. They are not
10.0+

y2

x
x

x
7.5+

x
x

5.0+
x

x
2.5+
-+
4.0

Figure 8.16 Data set 2:

+
6.0
Y2

+
8.0

against XI23'

10.0

+
12.0

+-x123
14.0

137

138

I LI___________R_E_L_A_T_I_N_G_O__N_E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
x

1.0+

SRES2

x
0.0+

-2.0+

_ + _ _ _+I _ _ _ + _ _ _++ _ _ _ +_ _ _ +_FITS2


5.0

Figure 8.17

-1.0+

6.0

7.0

8.0

9.0

10.0

Data set 2: standardized residuals against fitted values.

randomly distributed. They are positive in the middle and negative at


either end of the range of fitted values (Figure 8.17).

8.5.3 All points except one follow one line (data set 3)
One point is clearly very different from the others. MINITAB identifies it
as having a particularly large standardized residual (3.0, it is an outlier).

12.5+

y3

10.0+
x
7.5+

5.0+
-+
4.0

Figure 8.18

+
6.0

+
8.0

Data set 3: Y3 against x 123

+
10.0

12.0

+-x123
14.0

T_H_E_I_M_P_O_R_T_A_N_C_E_O_F_P_L_O_TT_IN_G_O_B_S_E_R_V_A_T_IO_N_S_ _ _---'I

L -_ _ _

There is no pattern in the residuals. We should check to see whether there


is an error in the data. If the data point is correct subsequent analysis
and interpretation both with and without the point seems sensible.
Data set 3 is plotted in Figure 8.18.
The regression equation is
y3=3.01+0.498 x123
Coef
3.012
0.4983
R-sq=66.9%

Predictor
Constant
x123
s=1.225

StOev
1.114
0.1168
R-sq (adj) = 63.2%

t-ratio
2.70
4.27

P
0.024
0.002

Analysis of Variance
SOURCE
Regression
Error
Total

OF
1
9
10

MS
27.310
1. 500

SS
27.310
13.498
40.808

F
18.21

P
0.002

Unusual Observations
Obs.
3

x123
13.0

y3
12.700

Fit

9.489

StOev.Fit
0.595

Residual
3.211

St.Resid
3.00R

R denotes an obs. with a large st. resid.

The plot of standardized residuals against fitted values clearly identifies


the same problem (Figure 8.19).

x
SRES3

1.5+

x
0.0+

x
X

-1.5+
-+
5.0

+
6.0

+
7.0

+
8.0

9.0

Figure 8.19 Data set 3: standardized residuals against fitted values.

+-FITS3
10.0

139

140

I LI___________R_E_L_A_T_I_N_G_O__N_E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
x

12.5+
y4

10.0+

x
x
x

7.5+

x
2

x
x

5.0+

--+----+----+----+---+-----+-x4
7.5
10.0
12.5
15.0
17.5
20.0

Figure 8.20 Data set 4:

Y4

against

X4-

8.5.4 All points except one are clustered together on the x axis (data set 4)
Here, one observation is determining the nature of the regression line. It
is important to check that the observation is correct. If so, it again seems
sensible to analyse the data with and without the value, to quantify the
difference it makes. In the absence of the one isolated value there is no
significant linear relationship between y and x. MINITAB identifies such
points on the printout as having much 'influence'. Note that '2' indicates
two points are very close together.
Data set 4 is plotted in Figure 8.20.
The regression equation is
y4 = 3.00 + 0.500 x4
Predictor
Constant
x4
s=1.236

Coef
3.002
0.4999
R-sq=66.7%

StOev
1.124
0.1178
R-sq (adj) = 63.0%

t-ratio
2.67
4.24

P
0.026
0.002

Analysis of Variance
SOURCE
Regression
Error
Total

OF
1

9
10

SS
27.490
13.742
41.232

MS
27.490
1. 527

18.00

0.002

Unusual Observations
Obs.
8

x4
19.0

y4
12.500

Fit StOev. Fit


12.500

Residual
1.236

St.Resid
0.000

X denotes an obs. whose X value gives it large influence.

*X

---'I I

T_H_E_I_M_P_O_R_T_A_N_C_E_O_F_P_L_O_TT_IN_G_O_B_S_E_R_V_A_T_IO_N_S
___

L -_ _ _

SRES4

1.0+

x
x

0.0+

x
-1.0+

X
X
X

+-----+----+------+------+---+--FITS4
o
10
20
30
40
50

Figure 8.21 Data set 4: standardized residuals against fitted values.


The plot of standardized residuals against fitted values (Figure 8.21) omits
the value derived from the eighth because the standardized residual is
calculated by dividing the residual by its standard deviation which happens
to be zero, thus resulting in an impossible calculation (division by zero).
8.5.5 The difference between using linear regression in an experiment and
in a survey
The values of x are imposed in an experiment: for example we could
control the amount of exercise which 40 women take by allocating times at
random. In contrast, in a survey we are not in control of the x values. They
are whatever they happen to be. In both cases our conclusions must be
restricted to the range of x values we have included in our study. But,
whereas in the experiment we have evidence to show that a change in the
amount of x causes a change in the amount of y, in a survey it may simply
be that x and yare linearly associated because they are both related to a
third variable.
8.5.6 Extrapolation is dangerous
If we predict a y value for an x value outside our range of data
(extrapolation) we cannot be sure that the relationship will still be of the
same form. For example, a plant which is growing steadily faster over the
range of fertilizer from 0 to 300 kg/ha is unlikely to continue growing at
the same rate if fertilizer were applied at the rate of 1000 kg/ha - it will
probably show signs of toxicity!

141

142

I ~I___________R_E_L_A_T_IN_G__O_N_E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
Therefore we should not extrapolate from our data; rather we should
extend the range of treatments in a subsequent experiment.

8.6

CONFIDENCE INTERVALS

A regression line is an estimate of the relationship between y and x. If


earlier we had taken a different random sample of roads or plants our
estimated line would have been somewhat different. For example we might
have obtained
lead= -9.90 + 0.623 traffic

8.6.1

For the slope of the line

As with the estimate of a mean (Chapter 2) we should consider calculating


the 95% confidence interval for the slope of the line. This gives us the range
within which we are 95% confident that the slope of the line appropriate
to the whole population will lie. These can be represented by:

bt x JRMS
SXX

with t taken from tables for p = 0.05 and for n - 2 df (where n is the
number of points on the graph). Here RMS = residual (or error) mean
square and Sxx = the sum-of-squares of the x values.
This equation is of the same general form as that for a population mean
(Chapter 2) in that we have an estimated value plus or minus t times the
standard error of that value.
8.6.2 For the regression line
We often subsequently wish to predict the value of y (y') for a particular
individual with a value of x (x'). This can be achieved as follows:

ytx

xi)

I (x' RMSx ( I+~+ Sxx

with t taken from the tables for n - 2 df and an appropriate probability


level (where n = number of points on the graph). Here RMS=residual (or
error) mean square, Sxx is the sum-of-squares of the x values and x is the
mean of the x values.
Again this follows the format: estimate plus or minus t times the
standard error of that estimate. It is just that here the standard error is
more complex.

C_O_R
__
R_EL_A_T_I_O_N__________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

If we work out confidence intervals for three values of x (Figure 8.22a)


we can join them together to provide a confidence interval for the line
(Figure 8.22b). We are 95% confident that the line which represents the
relationship for the entire population is within this range. Notice how the
confidence interval is narrowest in the centre of the graph and becomes
wider at the extremes. This is because outside the region of our
observations we know nothing about the relationship and so as we
approach this area we become less and less certain about the location of
the line.

8.7

CORRELATION

8.7.1 Various correlation coefficients


So far in this chapter we have worked with data where it is clear that one
variable may depend on the other. Sometimes however we may simply
wish to know to what extent two variables of equal status increase or
decrease with each other in a straight-line relationship. We are not
concerned with predicting one from the other but we want to know if they
are varying in a related way, i.e. are correlated.
For example, the concentration (ppm) of two chemicals in the blood
might be measured from a random sample of 12 patients suffering, to
various extents, from a particular disease. If a consequence of the disease is
that both chemicals are affected we should expect patients with high values of
one to have high values of the other and vice versa. The extent of the linear
correlation between the chemicals can be estimated by calculating a
correlation coefficient (r). This can be positive or negative, with r ranging
(b)

(a)

Response

X1

X2

X3

Amount

Amount

Note: Confidence interval smallest in centre

Figure 8.22 (a) Confidence interval for a particular value of y. (b) The
regression line.

143

144

II

RELATING ONE THING TO ANOTHER

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

between -1 and +1. A value near + 1 indicates a strong positive linear


correlation while a value near -1 shows a strong negative correlation. A
value of 0 may either indicate randomness or that a more complicated
relationship is present (a curve, for example); a plot helps here. The
correlation coefficient r is the square root of r2, the coefficient of
determination which we saw in the printout from regression analysis. Let
us examine a set of data.
The concentrations of chemical A in the blood are in the first column
of a MINITAB worksheet and the concentrations of chemical B are in the
second column. Both are in micrograms per litre. There is one row for each
of 15 patients.
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

46.1
23.6
23.7
7.0
12.3
14.2
7.4
3.0
7.2
10.6
3.7
3.4
4.3
3.6
5.4

8
24.7
15.2
12.3
10.9
10.8
9.9
8.3
7.2
6.6
5.8
5.7
5.6
4.2
3.9
3.1

First, we ask for a scatter plot (Graph, Character Graphs, Scatter Plot).
This is shown in Figure 8.23. We can ask for the correlation coefficient, r
(Stat, Basic Statistics, Correlation).
Correlation of A and 8=0.939

This is very high, suggesting a strong, positive linear correlation. We can


find out whether this value of r indicates a significant correlation by
comparing it with a critical value found in a statistical table of the linear
correlation coefficient, r (see Table C.8). The null hypothesis is that the
correlation coefficient is zero. If our estimate of it (which we make positive
if it happens to be negative, as it is only the size of the coefficient rather
than its sign which is important) is greater than the value found in the
table for the appropriate n (number of points on the graph) and p = 0.05
we conclude that there is a significant correlation. In this case the critical
value for p = 0.05 is 0.514, while for p = 0.01 it is 0.641, so we have strong
evidence for a significant correlation.
However, the point at the top right of the plot may well have a
considerable influence on this result. We could simply run the analysis with
and without the point and compare the results, but by omitting the result

II

CORRELA nON

~----------------------------------------------------~

30+

15+

------+----------+-------+-----+----+------------+---8
4.0
B.O
12.0
16.0
20.0
24.0

Figure 8.23 Scatter plot for correlation of chemicals A and B.


we are not making as much use of the data as we might. After all it is
(we believe) a real result. An alternative is therefore to opt to use a
different correlation coefficient: Spearman's r. (The 'classical' one which
we have been using up to now can similarly be called Pearson's r, or,
sometimes, the product moment correlation coefficient. This is the square
root of the r2 value we met earlier in linear regression.)
Spearman's r is calculated using the ranks of the observations as
follows. The lowest value for A is 3.0 so this has a rank of 1, while the
highest value is 46.1, so this has a value of 15. The amounts of B are also
ranked, with 3.1 receiving a rank of 1 and 24.7 a rank of 15. MINITAB
can do this for us and puts the results in columns 3 and 4 (Manip, Rank
data in Cl, Store ranks in C3, etc.):
ROW

1
2
3
4
5
6
7
8
9
10
11

12
13

14
15

A
46.1
23.6
23.7
7.0
12.3
14.2
7.4
3.0
7.2
10.6
3.7
3.4
4.3
3.6
5.4

B
24.7
15.2
12.3
10.9
10.8
9.9
8.3
7.2
6.6
5.8
5.7
5.6
4.2
3.9
3.1

rA
15
13
14
7
11

12
9
1
8
10
4
2
5
3
6

rB
15
14
13
12
11

10
9
8
7
6
5
4
3
2
1

145

146

I ~I___________R_E_L_A_T_I_N_G_O_N__E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
15.0+
rA

10.0+

5.0+

------+-----------+---------+-------+-----+-----+ ---rB
2.5
5.0
7.5
10.0
12.5
15.0

Figure 8.24 Scatter plot for correlation of rA and rB.

We can now ask for a correlation coefficient to be calculated for the


ranked data (Spearman's r) (Stat, Basic Statistics, Correlation):
Spearman's Rank Correlation coefficient
Correlation of rA and rB=O. 764

The outlier clearly had a considerable effect on the correlation coefficient


because the r value has come down to 0.764. We can see the difference by
plotting the ranked data (Graph, Character Graphs, Scatter Plot). This is
shown in Figure 8.24.
This examines whether there is a tendency for the two components of
the matched pairs to increase and decrease together (or for one to increase

,
,,

,,

,
,,

,
,,

,
I

.. , ,

I
I

x
Pearson's r
Spearman's r

,,

0.93
1

Figure 8.25 The effect of curvature on two types of correlation coefficient.

~___________________E_X_E_R_C_I_SE____________________~I
as the other decreases). Ranking the data maintains the order of the
observations on each axis. However, the size of the difference between one
observation and the next biggest one is standardized. Now the patient
who was an outlier is still the point at the top right of the graph but,
because the data are now ranked, he or she is not so far removed from the
other data points as before. Spearman's r is described as being more robust
because it is less sensitive than Pearson's r to occasional outliers or to
bends in the relationship.
Figure 8.25 illustrates this for two data sets showing how the value of
Spearman's r remains at 1 for a positively (or negatively) curved
relationship whereas Pearson's r gives a lower value, because of the lack of
linearity.
Again we can test the significance of Spearman's r. A table of critical
values gives one of 0.521 for p = 0.05 and 0.654 for p = 0.01. So we have
strong evidence of correlation.

8.7.2

Correlation is not causation

It is commonly found that there is a strong positive linear relation between

the number of doctors in a city and the number of deaths per year in that
city! At first sight we may be tempted to conclude that having more
doctors leads to more deaths. Therefore if we cut the number of doctors
we might expect fewer deaths. However, we have overlooked the fact that
both the number of doctors and the number of deaths in a city depend
upon the population of a city. We can calculate correlations between any
pair of variables but we must always be wary of assuming that one causes
variation in the other.

8.8

EXERCISE

Feeding lambs

We wish to investigate how the amount of a certain feed given to young


lambs may be related to their weight. Different amounts of feed (kg) were
eaten by each animal during the experiment and their weights (kg) at the
end of the experiment were recorded:
Data Display
Row
1
2
3
4

weight
7.4
8.5
7.1
8.2

feed
53
58
49
56

147

148

II

RELATING ONE THING TO ANOTHER

~----------------------------------------------------~

5
6
7
8
9
10
11
12

8.4
8.0
8.5
8.0
8.6
7.6
7.6
7.3

60
52
62
56
54
49
50
46

Enter the data into a MINITAB worksheet and save it.


What is the correlation coefficient (Pearson's correlation coefficient)
between these data sets (Stat, Basic Statistics, Correlation)?
What is the Spearman's rank correlation coefficient between these data
sets? (Use Manip, Rank to put ranks for weights into C3 and ranks of
feed into C4, then obtain the correlation coefficient as before.)
Plot the data to see whether there is likely to be a linear relationship
(Graph, Plot).
Carry out a linear regression of weight on feed (Stat, Regression,
Regression) and produce a graph with a fitted line and 95% confidence
interval.
Produce and examine a residual plot.
Interpret your results.

Answer
Correlations (Pearson)
Correlation of weight and feed = 0.825
Row
1
2
3
4
5
6
7
8
9
10
11
12

weight
7.4
8.5
7.1
8.2
8.4
8.0
8.5
8.0
8.6
7.6
7.6
7.3

feed
53
58
49
56
60
52
62
56
54
49
50
46

rankwt
3.0
10.5
1.0
8.0
9.0
6.5
10.5
6.5
12.0
4.5
4.5
2.0

rankfeed
6.0
10.0
2.5
8.5
11. 0
5.0
12.0
8.5
7.0
2.5
4.0
1.0

Correlations - although this will say Pearson it is now a Spearman's


coefficient.
Correlation of rankwt and rankfeed = 0.808

It is slightly smaller than the Pearson's coefficient but still very large.

~___________________E_X_E_R_C_I_SE____________________~I
Regression Analysis
The regression equation is
weight=3.17+0.0887 feed
Predictor
Constant
feed
s=0.3092

StOev
1. 038
0.01924
R-sq (adj) = 64 .8%

Coef
3.167
0.08867
R-sq= 68.0%

t-ratio
3.05
4.61

P
0.012
0.000

Analysis of Variance
SOURCE
Regression
Error
Total

OF
1
10
11

MS
2.0306
0.0956

SS
2.0306
0.9560
2.9867

Unusual Observations
Obs.
feed
weight
9
54.0
8.6000

Fi t
7.9555

21.24

0.000

StOev. Fi t
0.0894

Residual
0.6445

St.Resid
2.18R

R denotes an obs. with a large st. resid.

Interpretation
The initial plot (Figure 8.26) and high correlation coefficient (r = 0.825)
lead us to believe that a straight-line relationship between weight and feed
is appropriate. The linear regression analysis shows that there is very
strong evidence for a linear relationship with more feed leading to heavier
lambs. The p value p < 0.001 shows that we can reject the null hypothesis
of no linear relationship with 99.9% confidence and the graph shows that
the 95% confidence interval for the relationship covers a narrow band

8.5-

0
0

.::; 8.0 -

OJ

Q)

7.5 -

0
0

7.0
45

50

55

feed

Figure 8.26 High-quality plot of weight against feed.

60

149

150

II~___________R_E_L_A_T_I_N_G_O_N__E_T_H_I_N_G_T_O__A_N_O_T_H_E_R__________~
Regression of weight on feed intake
9.5

......, 8.5

..c

Ol
(j)

3.16712 8.87[-02
R-Squared ' 0.68

7.5

Regression

95t Cl

6.5

45

50

55

feed

60

Figure 8.27 High-quality plot showing fitted straight line and 95% confidence
interval for weight and feed data.

(Figure 8.27). The r2 value (coefficient of determination) indicates that


about two-thirds of the variation in lamb weights may be explained by
variation in the amount of feed eaten. The residual plots (Figure 8.28) are
reassuring in that the residuals follow an approximately Normal
distribution and their variability is not related to the size of the fitted

values.

E_X_E_R_C_I_SE____________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Residual plot - lamb growth


I Chart of Residuals

Normal Plot of Residuals


_ 0,5
ro
::>

::>

...

'D

~ 0:0

'D
(J)
Q)

<X

<X

- 0,5

r==========l
---t"4--+~
d-A~-+/\_~"""I
rv~

ro

-I

--L.:...----,--~--.,-l

5
10
Observation Number

Residuals

Histogram of Residuals
9
>-8

::> 5

'D

'- 3

Ct:

2'4

~ 0, a ...j...!.------=--------1

u...2
I

Figure 8.28

Fits

m
::>

VS,

- a,s

g7
Q)

x,o.ooo

- - - - - - - - - lCl'-O,990B

-0,5
0,5
1.5
Normal Score

-1.5

UCl'O,990B

-0,5

0,0
0,5
Residual

-a,5 --!,----,---=----,.-----,!
7,2

7.7

8,2

Fit

High-quality residual plots for weight and feed data,

8.7

151

What to do when data are


skewed, or are ranks or
scores, or are counts in
categories

9.1 INTRODUCTION

Most of the statistical techniques we have considered so far are parametric


tests (t-tests, analysis of variance and linear regression). This means that
they make certain assumptions about parameters of the distribution of the
population(s), e.g. that the data come from populations which are
Normally distributed and have similar variability. If data do not conform
to these assumptions we need to use non-parametric or distribution-free
tests instead. These do not make such strict assumptions about the shape
of the population distributions.
The method of box-and-whisker plots (Chapter 2) which calculates a
median and summarizes data in a way which highlights any departures
from the assumptions for parametric tests provides a useful starting point
for considering which techniques may be helpful.
Non-parametric tests can be useful when:
1. Underlying distributions are not Normal (for example they may be
skewed - showing 'straggle' to the right or to the left) (Figure 9.1a). It
may be possible to modify the data so that parametric tests can be used.
For example, if leaf areas have been measured we could analyse the
square root of each value, since this provides an alternative measure of
leaf size. This is called transformation. However, if we wish to analyse
the raw data we need to consider non-parametric tests.
2. There are extreme observations (outliers) (Figure 9.1b).
3. Observations are merely ranked or arbitrarily scored. For example,
first, second, third or a score of 0 to 10 for depth of colour.

M
__
A_N_N_-_W_H_I_T_N_E_Y_T_E_S_T________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

(a)

(b)
No.

01

No.

individuals

01

10

indiv- 10
iduals

No. 01 fleas I individual

10

20

30

40

50

60

Mid points 01 weight classes (kg)

Figure 9_1 (a) Positive skewness; (b) outliers.


In the last chapter we described Spearman's rank correlation coefficient
rs, which is a non-parametric method. We can use this to estimate the
correlation between two variables when our data fail to meet the
assumptions for using the equivalent parametric method (Pearson's
product moment correlation coefficient r).
In this chapter we will consider techniques that are alternatives to the
parametric tests which we have already used. These are:
1. Mann-Whitney. This is sometimes also known as the Wilcoxon rank

sum test. This is the non-parametric equivalent of analysis of variance


when there are two treatments.
2. Kruskal-Wallis. This is the equivalent of one-way analysis of variance
when there are two or more treatments. This can also cope with data of
type I.
3. Friedman. This is the equivalent of two-way analysis of variance when
there are blocks as well as two or more treatments.
In addition, we will discuss the chi-squared contingency test. This is used
for categorical data. Categorical refers to the fact that the data are the
number of individuals in each treatment which have been classified into
each of two or more categories (like dead or alive, rich or poor) rather
than being measures of some feature. For example, we have transplanted
200 plants of one species into a field: 100 with red flowers and 100 with
white flowers. One year later we can categorize both the red- and whiteflowered plants as either living or dead. We might wish to know whether
the red-flowered plants were more or less likely to survive than the whiteflowered ones.
9.2

MANN-WHITNEY TEST

If we have an experiment with two treatments in which our data are either
not measured on an absolute scale (for example they are scores or ranks

153

154

I ~I_______D_A_T_A_A_R__E_S_K_E_W_E_D_,_R_A_N_K_S_,_S_C_O_R_E_S_O_R_C_O__U_N_T_S______~
rather than centimetres or kilograms) or they are skewed rather than
Normally distributed we should use a Mann-Whitney test. This tests
whether there is a difference between the population medians. The
assumptions we must be able to make are: first, the observations must be
random and independent observations from the populations of interest;
and second, the samples we compare are assumed to come from
populations which have a similar shaped distribution (for example, both
are positively skewed).
The growth form of young trees grown in pots can be scored on a scale
from I (poor) to 10 (perfect). This score is a complex matter as it
summarizes the height, girth and straightness of the tree. We have 16 trees.
Eight of the trees were chosen at random to be grown in a new compost
while the remaining eight were grown in a traditional compost.
Unfortunately, one of the latter was accidentally damaged, leaving only
seven trees. The null hypothesis is that the two composts produce trees
with the same median growth form. The alternative hypothesis is that the
composts differ in their effect.
After six months growth the scores were recorded and entered into
columns 1 and 2 of a MINITAB worksheet:
Row
1
2
3
4
5
6
7

new
9
8
6
8

old
5
6
4
6

6
8
7

7
6

(Stat, Non-Parametrics, Mann-Whitney, First Sample new, Second


Sample old). First the number of observations (N) and the median for each
treatment is given:
Mann-Whitney Confidence Interval and Test
new
old

N=8
N=7

Median = 7.500
Median = 6.000

Now MINITAB calculates an estimate of the difference between the two


medians and provides a confidence interval for this difference:
Point estimate for ETAl - ETA2 is 2.000
95.7 pct c.L for ETAl - ETA2 is (l.000, 3.000)

This seems rather odd at first sight. ETAI - ETA2 represents the
difference between the 'centre' of the new compost scores and the 'centre'
of the old compost scores. ETA stands for a Greek letter of that name. If
we take away one median from another we have 7.5 - 6 = 1.5. However,

K_R_U_S_K_A_L_-_W_A_L_L_I_S_T_E_ST________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

the computer uses a more sophisticated calculation to estimate this


difference. It calculates the median of all the pairwise differences between
the observations in the two treatments. This gives a value of 2.0.
The confidence interval (1.0-3.0) is at 95.7%. This is the range of values
for which the null hypothesis is not rejected and is calculated for a
confidence level as close as possible to 95% (again the method involves
repetitive calculations which can only sensibly be carried out by a
computer program).
Finally a test statistic, W, is calculated. All the scores are ranked
together, with the smallest observation given rank 1 and the next largest
rank 2, etc. If two or more scores are tied the average rank is given to
each. Then the sum of the ranks of the new compost is calculated:
score
compost
rank

4
0
1

5
0
2.5

5
0
2.5

6
0
6

6
0
6

6
0
6

6
N
6

6
7
7
NON
6 10 10

7
N
10

8
N
13

8
N
13

8
N
13

9
N
15

(For example, the two trees with a score of 5 each receive a rank of 2.5
because this is the average of the next two available ranks, 2 and 3). The
sum of the ranks for the new compost trees (the smaller sample size) is
6 + 6 + 10 + 10 + 13 + 13 + 13 + 15 = 86

This is W
w= 86.0

Then the p value for the test is given (p

= 0.0128):

Test of ETAl=ETA2 vs. ETAI n.e. ETA2 is significant at 0.0128

This calculation assumes that there are no tied values in the data. Since
there are some ties, in this case an adjustment has to be made to the
calculation of the probability level to give p = 0.0106:
The test is significant at 0.0106 (adjusted for ties)

We can conclude that the two composts differ in their effect on tree growth
form. The new compost is preferable as it produces trees with a higher
median score. We now need to consider the relative costs and benefits of
using the two composts.

9.3 KRUSKAL-WALLIS TEST


Suppose that we have data from an experiment with more than two
treatments. We want to ask questions of it as we would in analysis of
variance but we should not because the observations suggest that the
populations are not Normally distributed. We can then consider using a
Kruskal-Wallis test because this test makes less stringent assumptions

155

156

II

DATA ARE SKEWED, RANKS, SCORES OR COUNTS

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

about the nature of the data than does analysis of variance. However, it still
makes the same assumptions as the Mann-Whitney test does: first, the
observations must be random and independent observations from the
populations of interest; and second, the samples we compare are assumed to
come from populations which have a similar-shaped distribution. This does
not have to be 'Normal'. It could be that both/all tend to have a few large
values and so have 'positive skewness' (Figure 9.1a) or they could both/all
be negatively skewed; but whatever the shape they must both/ all share it.
We will illustrate the use of this test on two data sets. First we will reexamine the data we analysed in Chapter 6 using one-way analysis of
variance.
We should note that data which are suitable for analysis of variance
can always be analysed by Kruskal-Wallis but the reverse is not always
so. Where the option of using analysis of variance exists however it will
normally be preferred since it is a more sensitive and elegant technique
(for example, the analysis of a factorial structure of treatments is easily
achieved).
To remind you, the data are from four treatments (types of vegetation
management by sowing and cutting) each replicated four times on plots
whose positions were randomly selected around a field margin. The data
are the number of spiders per plot.
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

treat
1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4

spider
21
20
19
18
16
16
14
14
18
17
15
16
14
13

13
12

We ask MINITAB to carry out the Kruskal-Wallis test on the data (Stat,
Non-Parametrics, Kruskal-Wallis, Response spider, Factor treat):
LEVEL

1
2
3
4

OVERALL

NOBS

4
4
4
4
16

MEDIAN
19.50
15.00
16.50
13.00

AVE. RANK
14.4
7.0
9.9
2.7
8.5

Z VALUE
2.85
-0.73
0.67
-2.79

-----'I I

K_R_V_SK_A_L_-_W_A_L_L_I_S_T_ES_T_ _ _ _ _ _ _

L -_ _ _ _ _ _ _

First, the number of observations in each treatment is presented


(NOBS), together with their median values. All 16 observations are ranked
and the mean rank of each treatment is given. (If there are ties the mean
rank is given to each individual.) Finally, a 'z value' is calculated for each
treatment. This shows how the mean rank for each treatment differs from
the mean rank for all 16 observations. Details of its calculation are given
in the box.
Box 9.1
Calculation of the z value

For convenience, the mean rank for all observations has been
converted to zero. This overall mean has been subtracted from each
treatment's mean rank and the result has been divided by the
standard deviation of that treatment's ranks to give a z value for
each treatment. These show the location of each treatment's mean
rank around zero, the overall mean of this 'standard' Normal
distribution (Figure 9.2). We remember from Chapter 2 that 95% of
the values would be expected to lie between + 1.96 and -1.96 units
from the mean. Here we see that two treatments have z values
outside this range suggesting that not all treatments are likely to
come from one population.
~

-4

Treatments

-2
;!

Figure 9.2

mean of ranks of all observations

value

Position of four treatments (mean ranks).

The program then prints a test statistic (H) (just as we have met t and
F before):
H = 12 . 66

df = 3

= 0 .006

The degrees of freedom are as usual one less than the number of
treatments. The p value is given (0.006), so we don't need to consult

157

158

II

DATA ARE SKEWED, RANKS, SCORES OR COUNTS

~----------------------------------------------------~

statistical tables. We can reject the null hypothesis that the observations
all come from the same population, because the value of p is very much
less than 0.05. We have strong evidence that at least one of the four
treatments differs from at least one of the other four in terms of its spider
population.
A second version of the test statistic, H, is also given:
H=12.85

df=3

p=O.005 (adj. for ties)

This differs slightly from the first in that an adjustment has been made
to account for the presence of tied observations (for example there are
several plots with 16 spiders). This has to be done because the test assumes
that the observations come from a continuous distribution (which could
contain values like 2.13459) whereas we have used it to analyse counts
where ties are more likely to occur. We should use this second, corrected
version of the statistic.
Let us now analyse the results of a different experiment. A student has
scored the amount of bacterial growth on each of 20 Petri dishes. A score
of 0 indicates no growth while one of 5 shows that the bacterium covers
the entire dish. Five days previously, ten randomly selected dishes had
received a standard amount of an established antibiotic used to inhibit
bacterial growth (treatment 1) while the remaining ten dishes had received
the same amount of a newly discovered inhibitory compound (treatment
2).
Row
1
2
3
4
5
6

7
8
9
10
11
12
13
14
15
16
17
18
19
20

treat
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2

score
1
2
3
2
3
4
2
1
2
3
3
3
5
4
2
5
4
3
4
3

A preliminary plot of the data suggests that the new compound does
not seem to be an improvement (Graph, Character Graphs, Boxplot):

F_R_I_ED
__
M_A_N_'_S_T_E_ST__________________~I

L -_________________

treat

------+

1----

------1

1-------

+ -------+-------+------+------+-----+ ----score
0.80
1.60
2.40
3.20
4.00
4.80

The Kruskal-Wallis test tests the null hypothesis of no difference between


the two populations:
LEVEL

NOBS

MEDIAN

10
10
20

2.000
3.500

1
2

OVERALL
H=6.0
H=6.46

d. f. = 1
d. f. = 1

AVE. RANK
7.2
13.8

Z VALUE
-2.46
2.46

p=0.014
p=O.Ol1 (adj. for ties)

We have evidence to reject the null hypothesis (p < 0.05). We conclude


that the new chemical is a worse inhibitor of bacterial growth.
9.4 FRIEDMAN'S TEST
In analysis of variance we saw that blocking was a useful way of taking
account of some of the random variation and so making it easier to detect
the effect of treatments (if there are any such effects). Friedman's test
allows us to do the same thing. However it is a non-parametric test and, as
with the Kruskal-Wallis test, it allows us to analyse data from populations
which are not Normally distributed but which have a similar shape.
In the field experiment with sowing and cutting management treatments
we allocated one of each treatment to each of four blocks, with a block
being one side of the field.
Row
1
2
3
4
5
6
7
8

treat
1
1
1
1
2
2
2
2

spider
21
20
19
18
16
16
14
14

block
1
2
3
4
1
2
3
4

159

160

I I

DATA ARE SKEWED, RANKS, SCORES OR COUNTS


9
10
11
12
13
14
15
16

3
3
3
3
4
4
4
4

18
17

15
16
14
13
13
12

1
2
3
4
1
2
3
4

(Stat, Non Parametrics, Friedman, Response spider, Treatment treat,


Blocks block):
Friedman test of spider by treat blocked by block
S=12.00
treat
1

2
3
4

d.L =3
N
4
4

4
4

p=0.008

Est. Median
19.344
14.844
16.469
12.719

Sum of RANKS
16.0
8.0
12.0
4.0

Grand median = 15.844

As with the Kruskal-Wallis test, the number of observations in each


treatment is given together with their median values. Here the sum of the
ranks of each observation within each block is then calculated. The sum of
16 for treatment 1 shows that it received the maximum rank of 4 in each
of the four blocks. Above this summary lies the test statistic, S, together
with the degrees of freedom (number of treatments minus one =4 - 1) and
the p value (0.008). This is slightly smaller than the p value (0.014)
obtained by ignoring blocking (Kruskal-Wallis). In other words, as we
might expect in a well-designed experiment blocking has caused an
increase (albeit small in this case) in the efficiency of detecting differences
between treatments.
Returning to our second example of the effect of two chemicals on
bacterial growth we can see that a Friedman's test might also be
appropriate here. Suppose that it takes about 30 seconds to score a Petri
dish for bacterial growth. Those that are last to be scored will have had
longer to grow than those scored first. Such a difference could be
important if, for example, all the treatment 1 dishes were scored first, then
all the treatment 2 ones. It would be much better to arrange our dishes in
ten pairs, each containing one of each treatment. We then score each pair,
randomly choosing whether to score treatment 1 or treatment 2 first within
each pair. We can then take time of assessment into account in the analysis
as a blocking effect. The two dishes which were scored first are given a
block value of 1 while the last pair are each given a value of 10. This will
remove any differences in growth due to time of assessment from the
analysis.

CHI-SQUARED CONTINGENCY TEST

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

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

treat
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2

score
1
2
3
2
3
4
2
1
2
3
3
3
5
4
2
5
4
3
4
3

II

time
1
2
3
4

5
6
7
8
9
10
1
2
3
4
5
6
7

8
9
10

Friedman test of score by treat blocked by time


S = 4.90
S = 5.44

d.f. =1
d.f. =1

treat
1
2

N
10
10

p=0.027
p=0.020 (adjusted for ties)

Est. Median
2.000
4.000

Sum of RANKS
11.5
18.5

Grand median = 3.000

Here the test statistic, S (as for the H statistic in the Kruskal-Wallis
test), is still highly significant (p < 0.05), but the p value has not been
reduced compared with that from the Kruskal-Wallis test. It appears that
time of assessment is not very important here. After all, with only 20 plates
in total the assessment should be complete in 10 minutes. However, if the
bacteria grew very fast and if there were 100 plates to assess, taking at
least 100 minutes, then blocking for time could be very important.

9.5

CHI-SQUARED CONTINGENCY TEST

Up to now the results of our experiments have provided us with an


observation for each individual (plot or Petri dish) which is quantitative.
They may be continuous data (weights or lengths or speeds) or counts
(how many spiders) or scores (the amount of bacterial growth on a 0-5
scale).
In contrast we may have an experiment in which each observation for
each treatment is categorical. This means that it can be placed in one of

161

162

I ~I_______D_A_T_A_A_R__E_S_K_E_W_E_D_,_R_A_N_K__S,_S_C_O_R_E_S_O_R__C_O_U_N_T_S______~
two or more categories, for example: green or brown; live or dead; small,
medium or large. If we prepare a table with two or more rows and two or
more columns to summarize the data, each item in the table will represent
the number of observations in that category. We then want to ask
questions like: Is the proportion of green individuals the same in each
treatment? Let us clarify this with an example.
A new drug has been developed which may be more or less effective at
clearing all parasites from the blood of humans within 36 hours. In an
experiment 287 individuals took part to compare its effectiveness with that
of chloroquinine (the standard). Of the 184 individuals receiving
chloroquinine, 129 were cleared of parasites within 36 hours while 55 were
not. We can summarize these observations and those for the new drug in
a table of observed values (0):
Cleared in 36 h
Chloroquinine
New drug
Total

Not cleared in 36 h Total

129

55
23

209

78

80

184
103
287

Note that the numbers of individuals taking the new drug and those
taking the old one do not have to be equal although the analysis will be
more robust if they are of similar magnitude. Our null hypothesis is that
the two variables are statistically independent: the proportion of
individuals from whose blood the parasite has been cleared is the same for
both drugs. Because we have only a sample from each of the two
populations we need statistically to assess these data. The first step is to
calculate the number of individuals we would expect to be in each category
(or 'class' or 'cell') if the null hypothesis is true.
In general, the expected number in each category or 'cell' is:
expected vaIue

row total x column total


d
I
gran tota

The maximum number of people who could appear in the top left 'cell'
is 184 because this is the number who were treated with chloroquinine.
The proportion 209/287 is the overall proportion of individuals from
whose blood the parasite had cleared in 36 hours, irrespective of which
drug they had received. This proportion is then multiplied by the number
of individuals who received chi oro quinine (184) to give the number of
individuals we would expect to find who had both received chloroquinine
and whose blood had cleared in 36 hours assuming both drugs have the
same effect. This value represents R o, the null hypothesis.
For the top left class of the above data set (Chloroquinine, Cleared)
we therefore obtain:

c_H_I-_s~Q~U_A_R_E_D_C_O__N_T_IN_G__EN__C_Y_T_E_S_T__________~I

L -___________

expected value

=E=

209 x 184
287
= l33.99

This test is called a chi-squared contingency test because the expected


values are contingent upon (dependent upon) the row and column totals
and the grand total. The table of such expected values (E) is then:
Cleared in 36 h

Not cleared in 36 h

Total

50.01

184

133.99

Chloroquinine
New drug
Total

27.99
78

75.01

209

103

287

The general idea is that if the null hypothesis is true, the observed and
expected counts will be very similar whereas if it is false the counts will be
very different. To enable us to calculate the strength of evidence against
the null hypothesis we calculate a test statistic - chi-squared (chi is a Greek
letter pronounced 'ky' to rhyme with sky):

l = L(O~Ei
This equation states that l is calculated as the sum of (the squares of
the differences between each observed count and its expected value,
divided by its expected value). Here we have:
(129 - l33.99)2 /l33.99 + (55 - 50.01)2/50.01
+(80 - 75.01)2/75.01 + (23 - 27.99i /27.99

= 1.91

The bigger the value of l the greater the chance that we will reject the
null hypothesis. Here, l = 1.91. We compare this with the critical value
from a statistical table (Table C.6) of l taken from the column headed
95% (p = 0.05). If 1.91 is bigger than this critical value then the null
hypothesis is rejected with 95% confidence. As always, we need to know
the degrees of freedom to find the critical value in a statistical table. In a
contingency table there are (r - 1) x (c - 1) degrees of freedom where
r = number of rows and c = number of columns. Thus for a 5 x 4 table the
df = 12. Here in our 2 x 2 table the df = 1. So we need the critical value
from the column headed 95% for 1 df in Table C.6. The calculated value of
l = 1.91 is less than the table value of 3.841 so we have no reason to reject
the null hypothesis of no difference in the efficacy of the two drugs.
MINIT AB can carry out a l analysis for us. The contents of the table
of observed counts is placed into the worksheet:
Row
1

clear
129
80

not
55

23

(Stat, Tables, Chisquare Test, Columns clear not)

163

164

I LI______D__A_TA__A_R_E_S_K_E_W_E_D_,_R_A_N_K_S_,_S_C_O_R_E_S_O_R_C_O_U_N_T_S______~
Expected counts are printed below observed counts
1
2
Total

clear
129.00
133.99

not
55.00
50.01

Total
184

80.00
75.01

23.00
27.99

103

209

78

287

ChiSq = 0 .186 + 0.499 + 0.332 + 0.891 = 1. 908


df=l, p=0.168

MINITAB prints the expected values in each category below the


observed values and presents the row and column totals and grand total.
Then follow the components of l (that is the value of 'observedexpected', squared and divided by expected) which are added up to give
the l statistic, together with its degrees of freedom.
Here the l statistic is very small, indicating that the observed values
are very close to those we expected to find if the null hypothesis were true.
Formally, following our MINITAB analysis we should compare our value
of 1.908 with the value of l found in a statistical table for 1 df and
p = 0.05 (see Table C.6, row 1, column 5). This is 3.84. Since 1.908 is less
than 3.84 we conclude that a similar proportion of people is cured by the
two drugs. MINITAB gives us the exact value of p = 0.168.
9.5.1 Some important requirements for carrying out a valid
test

contingency

1. The observations must be counts not percentages or proportions or


measures. It is important that the grand total is the number of
independent individuals in the study as this gives information about the
reliability of the estimates.
2. In general the expected count or frequency in each class should exceed
2 and also 80% of the classes should have expected frequencies greater
than 5. If this is not so then either we should collect more data or we
can combine neighbouring classes (if this is sensible).
For example, if we applied fertilizer to 30 of 60 plants at random
and classified the plants after one month's growth we might find:
Small
No fertilizer

With fertilizer

20

10

Medium

15

Large

The expected value for the top right class (large, no fertilizer) would
be (30 x 8)/60 = 4. For the bottom right class (large, with fertilizer) it

----'I I

C_H_I-_S_Q_V_A_R_E_D_C_O_N_T_IN_G_E_N_C_y_T_E_S_T_ _ _ _ _

L -_ _ _ _ _ _

would be the same. To fulfil the requirements for a X2 contingency test


we can combine the medium and large categories:
Small

20

No fertilizer
With fertilizer

Large
10

20

10

Now there will be no problem with expected values being too small.
They all happen to be (30 x 30)/60 = 15.
The reason for ensuring that expected values are greater than 5 is
that the test is over-sensitive to small differences when the expected
value is small. This is because dividing by a very small expected value
(imagine dividing by 0.1) will give rise to a ridiculously high component
of/.
3. When the contingency table is larger than the 2 x 2 case look at the
component of / which comes from each class. Those classes with large
values are mainly responsible for the rejection of the null hypothesis
and are therefore the ones to concentrate on when it comes to
interpreting your results, as we can see in the following example.
A further example of a

9.5.2

contingency test

We have carried out an experiment to compare the effectiveness of three


termite repellents in preserving fencing stakes. Each chemical was applied
to a sample of 300 fencing stakes, giving 900 stakes in total. The number
of stakes which were attacked was recorded after a year in the ground.
Row
1

2
3

attack
112
82

123

avoid
188
218

177

oil
creosote
copper arsenate

(Stat, Tables, Chisquare Test, Columns attack avoid)


Expected counts are printed below observed counts
attack

112.00
105.67

188.00
194.33

avoid

Total

82.00
105.67

218.00
194.33

300

123.00
105.67

177.00
194.33

300

317

583

900

Total

300

ChiSq= 0.380 + 0 .206 +5.301 +2.882 +2.843+ 1. 546= 13 .158

df = 2, P = 0 . 001

165

166

I LI_______D_A_T_A_A_R_E__SK__E_W_E_D_,_R_A_N_K_S_,_S_C_O_R_E_S_O_R_C_O__U_N_T_S______~
The table value of l for 2 df at p = 0.05 is 5.99 and for p = 0.01 it is
9.21. We therefore have strong evidence to reject the null hypothesis that
all three chemicals are equally effective. MINITAB provides the value
p = 0.001.
If we examine the components of l we see that the two coming from
creosote are especially high (5.301 and 2.882). This tells us that it is
creosote which has a different effect from the others. Comparing the
observed and expected values we see that creosote was more effective than
the other two treatments with only 82/300 = 27% of stakes being attacked
compared with 37% and 41% for the others.
An important assumption we have made is that the allocation of
treatments to the 900 stakes was carried out at random. If the 300 stakes
for each treatment had been grouped, any differences between treatment
might have been due to environmental differences such as one treatment
being on a sandier soil which is better shaded, thus possibly affecting the
density of termites in the area. As with many other forms of sampling,
therefore, randomization is the key.

9.6

CHI-SQUARED GOODNESS-OF-FIT TEST

Tests of goodness of fit are often required in genetic studies. In a genetic


crossing experiment we may want to know whether the observed ratios in
phenotypes of the offspring deviate significantly from the expected ratios.
For example, from genetic theory we might expect a ratio of 1: 2: 1 of the
phenotypes A, Band C and actually observe frequencies of 70, 134 and 75
in a sample of size 279. We calculate that theory predicts expected
frequencies of 1/4, 112 and 1/4 of 279 respectively: 69.75, 139.5 and
69.75.
We can now calculate l, using the same formula as for the contingency
test. The degrees of freedom are found by starting from the number of cells
(3) and deducting the number of constraints on the expected values (here
there is only one, namely that the expected values must sum to the same
total as the observed values: 279). So there are 2 df.
We can use M 1NITAB to calculate this test. This is done in the 'oldfashioned' way of typing in instructions because at the moment MINITAB
does not have a menu-based way of asking for this test.
Enter the expected values for each phenotype in column 1 and the
expected values in column 2:
Row
1
2
3

observed
70

134
75

expected
69.75
139.50
69.75

EX
__
ER
__
C_IS_E_S____________________~I

L -____________________

Now click to the right of the MINITAB prompt in the Session window
'MTB>' and type the commands in each line exactly as shown, followed
by the 'ENTER' key.
The first line is an instruction to calculate the l value, using Calc,
mathematical expressions):
MTB

>

let

kl

sum((cl

c2)**2/c2)

The next three lines ask for the p value to be calculated and stored:
MTB > cdf k1 k2;
SUBC > chisquare 2.
MTB > let k3 = 1 -

k2

Now we can ask for the values of X2 and its p value to be displayed along
with the observed and expected value by returning to the usual commands
and asking for c1, c2 , kl and k3 to be displayed:
Data Display
Row

observed
70
134
75

1
2

K1
K3

expected
69.75
139.50
69.75

0.612903
0.736054

The kl value (Kl) is l and the k3 value (K3) is the p value. The p value
is very large, indicating that there is no reason to doubt the hypothesis of
a 1 :2: 1 ratio of phenotypes A, Band C.
9.7 EXERCISES
Non-parametric tests

(aJ Comparing two treatments

Enter the observations for five replicates of each of two treatments A and
B into two columns of a MINIT AB worksheet:
Row
1
2
3
4
5

12
10
9
8
9

9
8
5
5
4

Carry out a Mann-Whitney test to compare the two sets of data (Stat,
non-parametric tests). What is your conclusion?

167

168

II~______D__A_T_A_A_R_E_S_K_E_W_E_D_,_R_A_N_K_S_,_S_CO__R_E_S_O_R_C_O_U_N_T_S______~
(b) Including a third treatment

The appropriate test is a Kruskal-Wallis test. For this the data need to
be in one column and treatment codes in a second column:
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

score
12
10
9
8
9
9
8
5
5
4
7

treat
1
1
1

1
1
2
2
2
2
2
3
3
3
3
3

5
4
3

Carry out this test and interpret the results.


(c) Adding in the effect ofa blocking system

This requires a third column of block codes. Then the Friedman test can
be carried out. What is your conclusion?
Row

score

12

2
3
4
5
6
7
8
9

10

treat

block

1
1
1
1
1

5
2
3
4
5
1
2

10

9
8
9
9
8
5
5
4

11

12
13
14
15

2
2
2
2
2
3
3

5
4
3

3
3
3

Answers
(a) Mann- Whitney confidence interval and test
A
N= 5
Median =
B
N= 5
Median =
Point estimate for ETAI-ETA2 is

9.000
5.000
4.000

2
3
4

3
4

E_X_E_R_C_IS_E_S____________________~I

L -____________________

96.3 Percent C.l. for ETA1-ETA2 is (0.001,7.000)


W=37.5
Test of ETA1 = ETA2 vs. ETA1 ~= ETA2 is significant at 0.0472
The test is significant at 0.0432 (adjusted for ties)

We have evidence that the two populations differ in their median values
(p = 0.047). Treatment A has a significantly higher median value (9
compared to 4).
(b) Kruskal- Wallis test
LEVEL
1
2
3
OVERALL
H=8.15
H=8.29

NOBS
5
5
5
15
d.f.=2
d.f.=2

MEDIAN
9.000
5.000
5.000

AVE. RANK
12.5
6.8
4.7
8.0

Z VALUE
2.76
-0.73
-2.02

p=0.017
p=0.016 (adjusted for ties)

We have strong evidence that the three populations do not all have the
same median value (p = 0.02). Treatment A has a higher median than the
other two (9 compared to 5 and 5).
(c) Friedman test
Friedman test of score by treat blocked by block
S=9.10
S=9.58

d.f. =2
d.L =2

treat
1
2
3

N
5
5
5

p=O.Oll
p=0.009 (adjusted for ties)

Est.
Median
9.000
6.000
5.000

Sum of
RANKS
15.0
9.5
5.5

Grand median = 6.667

The inclusion of blocking in this experiment has allowed us to say that


we have even stronger evidence that the three treatments do not all have
the same median (p = 0.01).

Chi-squared test
The medical officer of a large factory administered five different influenza
vaccines (see rows 1 to 5) to randomly chosen employees in December.
Next March she recorded the incidence of flu and carried out a l test on
the data as shown in the MINITAB worksheet:

169

170

I IL-_____D__A_TA__A_R_E_S_K_E_W_E_D_,_R_A_N_K_S_,_S_C_O_R_E_S_O_R_C_O_U_N_T_S______~
MTB>print cl c2
Row
1
2
3
4
5

flu
43
52
25
48
57

noflu
237
198
245
212
233

(a) How many people were expected to contract flu when given vaccine
3?
(b) What is the component of X2 from the cell for 'no flu, vaccine 5'.
(c) What are the degrees of freedom?
(d) Enter the data into a MINITAB worksheet and carry out the l
analysis.
(e) Use (Calc, mathematical expressions, New variable=c3, Expression =
cll(c1 +c2 to work out the proportion of people with flu for each
vaccine.
(f) Interpret the output.
(g) Remove vaccine 3 from the worksheet, reanalyse the data and
interpret the results.
Answers

(a) Expected value = (270 x 225)/1350 = 45.0


Ei /E = 0.311
(c) Degrees of freedom = (rows - 1) x (columns - 1) = 4 x 1 = 4
(d) Chi-squared test for all five vaccines

(b) (0 -

Expected counts are printed below observed counts


flu
43
46.67

noflu
237
233.33

Total
280

52
41. 67

198
208.33

250

25
45.00

245
225.00

270

48
43.33

212
216.67

260

57
48.33

233
241.67

290

Total

225

1125

1350

ChiSq= 0 .288 + 0.058 +


2.563+0.513+
8.889+1. 778+

~____________________E_X_E_R_C_IS_E_S____________________~I
0.503+0.101+
1. 554 + O. 311 = 16. 555
df=4, p=0.002

(e) Proportion with flu for each vaccine


Row
1
2
3
4
5

flu
43
52
25
48
57

noflu
237
198
245
212
233

prop

o.15357l

0.208000
0.092593
0.184615
0.196552

Vaccine 3 has a much lower proportion of people with flu (0.09) than
do the others (between 0.15 and 0.20). This is shown by the large
component of l (8.889) coming from this cell.
(f) The calculated X2 value of 16.55 is greater than the value in the l table
(Table C.6) for row 4 (df) and column 1 of 13.3. So we can reject the
null hypothesis of each vaccine having the same proportion of people
contracting flu with 99% confidence. MINITAB gives a more accurate
value of p = 0.002 - showing that we are 99.8% confident in this
conclusion.
(g) Chi-squared test for four vaccines
Expected counts are printed below observed counts
flu
43
51.85

nof1u
237
228.15

Total
280

52
46.30

198
203.70

250

48
48.15

212
211.85

260

57
53.70

233
236.30

290

Total

200

880

1080

ChiSq= 1. 511 + 0.343 +


0.703+0.160+
0.000 + 0.000 +
0.202+0.046=2.966
df=3, p=0.397

We conclude that there is no reason to reject the null hypothesis that


vaccines 1, 2, 4 and 5 have the same effectiveness. This is because p is
greater than 0.05. Thus it was vaccine number 3 which, in differing from
the other four, caused the first l test to be significant.

171

10

Summarizing data from an


observational study

In an experiment we must have an observation from each of several


experimental units from each treatment (replication). These replicates must
not have any effect on each other (independence) and they should be

arranged during the experiment so that each one has an equal chance of
being in any particular position (randomization). However, the name
'experiment' is often mistakenly applied to investigations which are really
observational studies. For example, the investigator may make
observations without applying any treatments or in which the so-called
replicate plots of each treatment are grouped together, i.e. they are neither
randomized nor interspersed (Figure 10.1).
In this chapter we will outline an exploratory technique which can
simplify the results from such observational studies - principal components
analysis. Its great strength is that it can summarize information about
many different characteristics recorded from each individual in the study.

Site 1

Figure 10.1

Site 2

Random sampling within each of the three sites.

Site 3

C_A_SE_S_T_U_DY
__
1______________~1

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

We will illustrate its use by reference to two data sets: first, a study of
the economic and social characteristics of a group of countries; and
second, a study of the vegetation on a hillside.

10.1 CASE STUDY 1


Assessing the economic and social characteristics of 14 countries
We have obtained information (from the OEeD Observer, vol. 109, March
1981) about seven characteristics (population density; proportion in
agricultural employment; national income per head; capital investment in
machinery, etc.; infant mortality rate; energy consumption per head; and
number of TV sets per 100 of the population) for 14 countries (Australia,
France, West Germany, Greece, Iceland, Italy, Japan, New Zealand,
Portugal, Spain, Sweden, Turkey, the UK and the USA). This has been
entered on MINITAB in columns I to 7 in the same order as listed above.
Each column has been given a suitable short name (from 1 to 8
characters). We use (File, Display data):
Row
1
2
3
4
5
6
7
8
9
10
11
12
13
14

popden

2
97
247
72
2
189
311
12
107
74
18
56
229
24

agemp

nat inc

6
9
6
31
13
15
11
10
31
19
6
61
3
4

8.4
10.7
12.4
4.1
11.0
5.7
8.7
6.8
2.1
5.3
12.8
1.6
7.2
10.6

capinv
10.1
9.2
9.1
8.1
6.6
7.9
10.9
8.0
5.5
6.9
7.2
8.8
9.3
7.3

infmort

12
10
15
19
11
15
8
14
39
15
7
153
13
13

energy

tvsets

5.2
3.7
4.6
1.7
5.8
2.5
3.3
3.4
1.1
2.0
6.3
0.7
3.9
8.7

36
28
33
12
25
22
24
26
9
21
37
5
39
62

We then obtain summary statistics for each of the variables (see section
2.2 for an explanation of the columns) (Stat, Basic Statistics, Descriptive
Statistics):
pop den
agemp
nat inc
capinv
infmort
energy
tvsets

Mean

Median

TrMean

14
14
14
14
14
14
14

102.9
16.07
7.671
8.207
24.6
3.779
27.07

73.0
10.50
7.800
8.050
13.5
3.550
25.50

StDev

SEMean

93.9
13.42
7.750
8.208
15.3
3.625
26.00

101.3
15.73
3.619
1.462
37.7
2.211
14.40

27.1
4.20
0.967
0.391
10.1
0.591
3.85

173

174

I I

SUMMARIZING DATA FROM A STUDY


pop den
agemp
natinc
capinv
infmort
energy
tvsets

Min
2.0
3.00
1. 600
5.500
7.0
0.700
5.00

Max
311. 0
61. 00
12.800
10.900
153.0
8.700
62.00

Q1
16.5
6.00
5.000
7.125
10.8
1. 925
18.75

Q3
199.0
22.00
10.775
9.225
16.0
5.350
36.25

The MINIT AB printout shows that there is a great deal of variability


between countries, but the data are not easy to digest. If we had only
obtained information about one variable, such as population density, we
might draw a boxplot. Even with seven variables it is sensible to start by
looking at each in this way, to spot outliers (Graph, Character Graphs,
Boxplot). The resulting boxplots are shown in Figure 10.2.
We see that Turkey had an unusually high proportion of agricultural
employment and that both it and, to a much lesser degree, Portugal had a
high infant mortality rate, while the USA had a high number of TV sets
per 100 people.
Some of these variables are probably correlated. We can check for this
by asking for correlation coefficients (Chapter 8) (Stat, Basic Statistics,
Correlation):
agemp
nat inc
capinv
infmort
energy
tvsets

popden
-0.150
0.019
0.490
-0.131
-0.255
-0.069

agemp

natinc

capinv

infmort

energy

-0.786
-0.183
0.890
-0.715
-0.783

0.196
-0.602
0.830
0.722

0.002
0.009
0.134

-0.494
-0.526

0.915

Here we see, for example, a strong positive correlation between energy


and TV sets, while there is a strong negative correlation between the
proportion of the population in agricultural employment and TV sets.
It would be useful to be able to give each country a score which took
into account all seven of its characteristics. Then these scores would
summarize all the relative similarities and differences between them. We
might consider simply adding up the seven values for each country.
However, this would give undue weight to variables with large values (like
population density). So we could start by standardizing the data. If we
subtract the mean population density from each individual value and then
divide by the standard deviation of population density, the new values
have a mean of 0 and a variance of 1. They have been standardized. The
same method is applied to the remaining six variables so that they all now
have means of 0 and variances of 1. In this way they all have equal
status.
We can, however, still improve upon this by adding together these new

C_A_S_E__
ST_U_D
__
Y_l__________________~1

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

(a)

-I

+----+---+---+---+----+--popden

60

120

180

240

300

(b)

--I

1------

- + - - - - - - + - - - - + - - - - - + - - - - + - - - + ---agemp

12

24

36

48

60

(c)

------1

1-----

-+---+----+-----+---+-----+-natinc

2.0

4.0

6.0

8.0

10.0

12.0

(d)

--------1

- - - + - - - - - + - - - + - - - - + - - - + - - - + -capinv

6.0

7.0

-I +-

8.0

9.0

10.0

11.0

(e)

+ - - - + - - - - + - - - - - + - - - + - - - - - + ---infmort

30

60

90

120

150

(I)

---I

---+----+---+---+----+----energy
1.5
3.0
4.5
6.0
7.5

(g)

---------1

1-

+---+-----+----+---+------+--tvsets

12

24

36

48

60

Figure 10.2 Boxplots for Case study I: (a) population density; (b) agricultural
employment; (c) national income per capita; (d) capital investment; (e) infant
mortality rate; (f) energy consumption per capita; (g) number of TV sets per 100
people.

175

176

II

SUMMARIZING DATA FROM A STUDY

~----------------------------------------------------~

values for each variable for each country to obtain a score. We can ask
for our scores to be combined in such a way as to explain as much of the
variation present in our data set as possible. This is the basis for principal
components analysis (PCA). It provides us with the 'best' scores.
We ask for a PCA to be carried out on the seven variables in columns
1 to 7, with the 'best' scores for the first principal component to be put in
column 15 (with a further set of scores which accounts for some of the
remaining variation being put into column 16 - the second principal
component and so on). We use (Stat, Multivariate, Principal Components,
Variables Cl-C7, Number of Components 7, Type of Matrix Correlation,
Storage Coefficients C8-CI4, Scores C15-C2l). We name columns 8 to
14 PCAl-PCA7 and columns IS to 21 scorel-score7.
The resulting printout can seem rather overwhelming, but we often need
only concentrate on small parts of it.
Eigenanalysis of the Correlation Matrix
Eigenvalue 3.9367 1.5639 0.8100
Proportion 0.562 0.223
0.116
Cumulative 0.562 0.786 0.902

0.3572 0.2703 0.0451 0.0169


0.051 0.039 0.006 0.002
0.953 0.991
0.998 1.000

The first part of the output (above) has seven columns. Each represents
one of seven principal components which summarize the variability in the
data and each contains an eigenvalue. We can think of the eigenvalue as
the amount of variation in the data set which is explained by that
particular principal component. For example, the first principal component
has an eigenvalue of 3.9367. This represents 56.2% (shown underneath the
eigenvalue in row 2, but as a proportion 0.562) of the sum of all seven
eigenvalues. Principal component number two accounts for a further
22.3%, so components I and 2 account for 78.6% (given in row three as
0.786) of all the variation between the two of them.
This is good news; we have accounted for over three-quarters of the
variation in the data by two sets of scores whereas the original data
contained seven variables. If we include the third principal component we
can account for 90% of the variation. However we shall probably obtain
enough insight into the structure of the data by examining only the first
two principal components.
The printout then shows the coefficient or weight allocated to each
variable in each of the seven principal components to account for the
maximum amount of variation and to produce the 'best scores':
Variable
popden
agemp
natinc
capinv

PCl
-0.009
0.476
-0.452
-0.083

PC2
0.717
-0.108
0.015
0.639

PC3
0.252
-0.257
-0.146
-0.538

PC4
PC5
0.634
0.060
0.158
0.156
0.022
0.805
-0.519 -0.116

PC6
-0.116
-0.598
0.189
-0.114

PC7
-0.059
0.537
0.301
0.003

CASE STUDY I
infmort
energy
tvsets

0.394
-0.450
-0.452

-0.067
-0.234
-0.085

-0.626
-0.311
-0.267

I I

0.379 0.097
0.227
0.016
0.329 -0.549

0.462
-0.577
0.182

-0.286
-0.511
0.524

What are these coefficients? How are they used? Let us take the
column headed PCl. We will use these coefficients to obtain the score
for the first country (Australia) on principal component 1. To do this
we will copy down the coefficients and then put the actual data for
Australia's population density, etc., beside them (check that you can see
that the data come from row I in the printout at the beginning of this
section):
Variable

PCI

Data

popden
agemp
natinc
capinv
infmort
energy
tvsets

-0.009
0.476
-0.452
-0.083
0.394
-0.450
-0.452

2
6
8.4
10.1
12
5.2
36

We now standardize each data value. This is achieved by subtracting


the mean of that variable and then dividing by its standard deviation. For
example, for population density we subtract 102.9 and then divide by
101.3. For Australia this gives -0.996. The standardized values for the
other variables are given below. Check that you can obtain the
standardized value for agricultural employment in the same way by
referring to the data summary at the start of this section to find the mean
and standard deviation:
Variable

PCI

Data

Standardized data

popden
agemp
natinc
capinv
infmort
energy
tvsets

-0.009
0.476
-0.452
-0.083
0.394
-0.450
-0.452

2
6
8.4
10.1
12
5.2
36

-0.996
-0.640
0.201
1.295
-0.334
0.643
0.620

Now we multiply the coefficient and the standardized value for each
variable to give a final column of numbers which, when added together,
give the score for Australia:

177

178

II~___________SU_M__M_A_R_I_Z_IN_G__D_A_T_A_F_R_O_M__A_S_T_U_D_Y__________~
Variable

PCI

Standard data

PCI x Standard data

popden
agemp
natinc
capinv
infmort
energy
tvsets

-0.009
0.476
-0.452
-0.083
0.394
-0.450
-0.452

-0.996
-0.640
0.201
1.295
-0.334
0.643
0.620

0.008964
-0.304640
-0.090852
-0.107485
-0.131596
-0.289350
-0.280240

SUM = SCORE on PCI axis

-1.195199

We ask MINITAB to display this score and those for the other 13
countries for the first two principal components whose columns we have
named scorel and score2. Notice that our value for the score of Australia
on PC1 (-1.195199) agrees with the MINITAB value (-1.19523) to three
figures after the decimal point.
Row
1
2
3
4
5
6
7
8
9
10
11

12

13

14

scorel
-1.19523
-0.81330
-1.41233
1. 74510
-0.89650
0.54333
-0.43256
-0.05447
2.56455
0.91468
-1. 88907
4.74817
-0.93008
-2.89230

score2
0.00515
0.48198
1.39261
-0.06283
-1. 55911
0.65681
2.78599
-0.62952
-0.91376
-0.56365
-1.24519
-0.17634
1. 39480
-1.56695

We now put the two- or three-letter codes for each country into column
22. We ask for a labelled plot of score2 against scorel with each country
being labelled according to its code in column 22 (Graph, Plot, y = score2,
x=scorel, Annotation Data Labels, Show Data Labels, Use labels C22).
Figure 10.3 summarizes our data. We see that the countries are well
spread out. Let us take the first principal component (x axis). What aspects
of the data does it represent? The weights or coefficients for agricultural
employment and infant mortality are large and positive. This means that
countries with a high proportion of people in agricultural employment and
with high infant mortality (Turkey) are at the right-hand side of the graph.
The weights for national income, energy and TV sets are large and
negative. This means that countries with high national income, high energy
consumption and a large number of TV sets per 100 people are at the
left-hand side of the graph (Sweden and the USA). With this information
we can reasonably summarize the meaning of scorel as a summary

C_A
__
SE__
ST_U_D
__
Y_2__________________~1

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Principal components
analysis of socio-economic
data from 14 countries
o

Jap

WG UK
o

Ib8

FrB
0
Aus
0

NZ

Swe
0

Sga

Gre
0

TI5'
Por
0

Ice
0

-2

-1

score 1

Figure 10.3 Plot of score2 against scorel for data for 14 countries in Case
study I.
statistic of economic development on which a low score represents a high
level of development.
Looking at score2 we see that countries at the top of the graph are
characterized by a high population density and a high level of capital
investment (Japan) whereas those at the bottom of the graph have a low
population density and a lower amount of capital investment (Iceland,
USA). This graph enables us to see the similarities and differences between
the 14 countries at a glance. In this sense it is a great improvement on
the original table of data. We could if we wanted continue with the other
components. For example, we would next look at component 3 (score3).
However it is not very worth while because it only accounts for a further
11 % of the variation and the remaining components only account for
about the same amount of variation again between them.
10.2

CASE STUDY 2

Species composition of grassland on three parts of a site


A common exercise on many ecology courses is to put ten 0.5 m x 0.5 m
quadrats down at random on the top of a hill, another ten on one face and
another ten at the bottom of the hill. We then estimate the percentage
cover of each species in each quadrat. The aim is to find out whether and,
if so, how the vegetation differs between the areas.
It is tempting to see this as an experiment with three treatments (sites)
and ten replicates but really it is separate random sampling within each of
three separate areas (Figure 10.1). The results can be summarized

179

180

I I

SUMMARIZING DATA FROM A STUDY


20
No. of
species
per
quadrat

x
x

:1
x

:1

10

= Standard
deviation

:r

x
x

x
x
0

Site

Figure 10.4 The mean number of species per quadrat on each of three sites in

Case study 2.

separately for each area; for example by giving estimates of the mean
number of species per quadrat with a standard deviation to show the
variability within the site (Figure 10.4). However, what we should not do is
use the method of analysis of variance to answer the question: 'Is there
evidence of a difference in percentage cover of a particular species between
the three areas?' This is because we have not started from a common
population and then applied the treatments (top, middle, bottom) at
random.
Instead we should use an exploratory technique, such as principal
components analysis, to summarize the variability. This has the added
advantage of using the information about all the species together. It is a
type of multivariate analysis.
We enter the data into MINITAB, with 37 columns (= species) and 30
rows (=quadrats).
Here are the data (collected by Wye College students) for the first six
species (= columns) which represent Agrostis capillaris, Brachypodium
pinnatum, Carex caryophyllea, Carex jlacca, Dactylis glomerata and
Festuca ovina:

a
a
a
a

17
13
100
50

20

a
a

40
66
70
53
83
87
100

a
a

50
90
40

a
a
a
a
a
a
a
a
a
a

40
40

a
a
a
a
a
a
a
a
a
a

100
60

53
100
70
47
22
87
80
40

60

a
a

100
100
90
87
89
93
90
80
100
90
100
60

CASE STUDY 2
0
0
0
0
0
0
0
0
0
0
40
60
0
20
40
0
40
40

50
88
50
100
10
60
80
60
0
0
0
0
8
20
0
0
0
50

25
38
10
20
80
40
0
0
0
0
0
0
0
0
0
0
0
0

I I

88
63
30
50
100
80
20
0
0
0
0
0
0
20
0
0
0
0

0
0
5
50
0
0
0
40
60
80
40
0
0
40
20
0
40
50

100
88
65
100
90
70
0
90
40
80
40
0
100
0
100
0
100
90

We ask for a principal components analysis (Stat, Multivariate,


Principal Components, Variables CI-C37, Number of Components 6,
Type of Matrix Correlation, Storage Coefficients C38-C43, Scores C44C49); and name column 44 'scorel' and column 45 'score2'.
The output below shows the proportion of variation accounted for by
each of the first six principal components (eigenvalues). Then the weights
or coefficients for each species in each of the first six principal components
have been displayed.
Eigenanalysis of the Correlation Matrix
Eigenvalue
Proportion
Cumulative

11. 544
0.312
0.312

5.144
0.139
0.451

2.639
0.071
0.522

2.257
0.061
0.583

2.016
0.054
0.638

1.773
0.048
0.686

Here the first two axes only account for 45% of the total variation but,
as we shall see, it will still provide a useful summary.
Variable
ag.cap
brac.pin
car.car
car.fla
dac.glo
fest.ov
hol.lan
lol.per
ach.mill
agr.eup
cam. rot
cir.ac
cir.ar
cir.pal
crep.spp

PC1
-0.121
0.120
0.239
0.268
-0.173
0.040
-0.096
-0.177
-0.080
-0.085
0.163
0.183
-0.088
-0.062
-0.072

PC2
-0.032
0.283
-0.005
-0.022
0.190
0.196
0.207
-0.185
0.175
0.173
-0.023
0.057
0.219
0.203
-0.231

PC3
-0.248
-0.114
-0.060
-0.049
-0.171
-0.089
-0.299
-0.062
0.293
-0.349
0.124
0.021
0.189
-0.339
-0.090

PC4
-0.152
-0.086
0.063
0.077
-0.095
-0.079
-0.173
0.085
0.219
-0.037
-0.198
0.171
0.277
0.022
0.127

PC5
0.277
-0.034
0.011
-0.051
-0.156
-0.259
0.108
0.019
0.147
-0.219
-0.282
0.228
-0.047
-0.130
-0.023

PC6
-0.257
-0.024
0.046
-0.015
0.172
-0.307
0.054
0.193
-0.050
0.121
-0.110
-0.005
-0.102
-0.055
0.093

181

182

I
cru.lae
cyn.cri
gal.ver
gen.am
gle.hed
hel.num
hie .pi!
hyp.per
leon.sp
lot. cor
ori.vul
pim. sax
plan.lan
pol. vul
pot. rep
pru.vul
ran. spp
san.min
thy. spp
tri.rep
ver.cha
vio.hir

SUMMARIZING DATA FROM A STUDY


-0.037
-0.080
0.012
0.176
-0.116
0.273
0.242
0.172
0.216
-0.069
0.147
0.250
-0.168
0.135
-0.146
-0.025
-0.069
0.281
0.254
-0.224
-0.087
0.232

0.243
-0.197
-0.143
-0.095
0.247
0.017
-0.003
-0.029
-0.074
-0.324
0.072
-0.006
-0.313
-0.144
0.164
-0.182
0.002
0.021
0.021
-0.193
0.192
-0.027

0.137
0.116
-0.078
-0.194
0.031
0.069
-0.044
-0.197
-0.107
-0.017
0.113
-0.025
0.044
-0.233
-0.189
-0.117
-0.310
0.046
-0.043
-0.101
0.167
-0.124

0.130
0.010
0.282
0.263
0.173
-0.092
0.145
0.286
0.005
-0.033
-0.388
0.039
0.073
0.258
0.204
-0.090
-0.014
-0.064
-0.020
0.020
0.327
0.054

0.088
-0.229
-0.247
0.260
0.155
-0.003
0.059
-0.130
0.003
-0.093
0.061
-0.137
-0.084
0.114
-0.364
-0.147
0.368
-0.056
-0.071
0.042
-0.100
-0.046

-0.008
0.023
0.375
0.025
-0.176
0.134
0.095
-0.112
-0.112
-0.315
0.209
-0.078
0.070
-0.093
0.136
-0.501
0.061
0.063
0.053
0.024
-0.203
-0.082

Each column (species) was given an abbreviated Latin species name


(maximum of eight letters allowed), according to Clapham et al. (1981).
We ask for a printout of the scores of each quadrat on the first two
principal components:
Row
1

2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

score1
-2.65742
-3.11976
-2.50492
-2.57511
-1.33581
-0.97591
-2.70955
-2.10436
-1.74847
-1. 88399
6.87997
3.72506
6.68716
5.48874
2.73888
4.11199
6.12395
3.77876
3.09141
2.23702
-2.96016
-3.19158
-2.29824
-2.34998

score2
3.23235
3.64558
2.44219
3.52153
2.05404
2.16692
3.06575
1. 72044
0.68887
1.47517
0.59776
1.42079
-1.77234
-0.24516
-0.02376
1.25905
-0.83651
-0.01016
-0.28989
0.76591
-1.76356
-3.29659
-1.86741
-4.23661

C_A_S_E_ST_U_D
__
Y_2__________________~1

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

-1.54096
-2.96905
-2.46572
-1. 66073
-1. 67677
-2.13446

25
26
27
28
29
30

-3.42761
-2.81131
-3.48355
-2.09768
-1.73013
-0.16408

We insert ten 'l's, followed by ten '2's and ten '3's in column 39 so that
we can then obtain a labelled plot showing the three areas (Graph, Plot,
y=score2, x=scorel, For each Group, Group Variables C39). This is
shown in Figure 10.5. We can edit the graph to add a title and name the
three areas.
We can now interpret the meaning of the two axes. For each of the first
two components I have selected species with relatively large (positive or
negative) coefficients and with whose ecology I am familiar.
Principal component I
Large weights or coefficients
(positive or negative)

Short name

Full names

-0.224

tri.rep

-0.177

lol.per

-0.168

plan.lan

0.281

san.min

0.273

hel.num

Trifolium repens
(white clover)
Lolium perenne
(perennial ryegrass)
Plantago lanceolata
(ribwort plantain)
Sanguisorba minor
(salad burnet)
Helianthemum nummularia
(rock rose)

Principal components analysis of species


composition of 3 areas

N
(j)
L

'-'
(f)

0g

3 2 1 -

o-

0 0

00

+
++ +

-1 x

-2 -3 -

x
x

-4 -

x
x

-5

-4

Top
Middle
Bot tom

score 1

Figure 10.5 Plot of score2 against score! for plant species data in Case study 2.

183

184

I LI___________S_V_M_M_A__R_IZ_I_N_G_D__A_T_A_F_R_O_M__A_S_T_V_D_y__________~
Principal component 2
Large weights or coefficients
(positive or negative)

Short name

Full names

-0.324

lot.cor

Lotus corniculatus

-0.313

plan.lan

Plantago lanceolata

0.283

brac.pin

Brachypodium pinnatum

0.247

gle.hed

Glechoma hederacea

(birds foot trefoil)

(ribwort plantain)
(tor grass)

(ground ivy)

By referring to textbooks which tell us about the conditions in which


these species thrive we can build up a picture of the two axes. The
horizontal axis (scorel) goes from a fertile and moist environment on the
left (characterized by ryegrass, clover and plantain) to an infertile and dry
environment on the right (characterized by salad burnet and rock rose).
The vertical axis (score2) goes from a heavily grazed environment
(negative scores) (characterized by birds foot trefoil and plantain) to a less
heavily grazed environment (positive scores) (characterized by tor grass
and ground ivy). We can now characterize the areas from which we
sampled the vegetation (Figure 10.6).
We can conclude that at the top of the hill there is a relatively moist
and fertile area but that it is not heavily grazed. In the middle of the slope
there is an infertile, dry area which is moderately grazed. At the bottom
of the slope is an area which is fertile and moist like the top of the hill but,
in contrast, it is heavily grazed.
Principal components analysis of species
composition of 3 areas

o Top
+

3
2

o
Axis 2 0

-1

-5

++ +
~

-2
-3
-4

light g?azing,O
00

heavy g~aZing
fertile moist ....
oo([--------'>~ infertile dry

"""'-r-------,-------.-----'

-4

Axis 1

Figure 10.6 An edited version of the previous figure, showing the


characterization of the areas.

Middle
Bot tom

--'I I

S_PE_C_I_A_L_IZ_E_D_P_R_O_G_R_A_M_S_ _ _ _ _ _ _

L - -_ _ _ _ _ _ _

This type of interpretation has proved to be a very helpful way of


summarizing the complex data (a table of 30 x 37 = 1110 percentage cover
values) which we have obtained.

10.3 SPECIALIZED PROGRAMS FOR ANALYSING VEGETATION


DATA

In Case study 2 we found PCA to be a useful way of exploring vegetation


data. However, such data are better analysed by specialist programs which
fit more appropriate models. For example, de-trended correspondence
analysis (DCA) is included in the VESPAN III package (Dr A.J.C.
Malloch, University of Lancaster).

185

Your project

11

Most undergraduate courses include a project. This chapter will give


you some useful guidelines to help you to complete your project
successfully.
11.1

CHOOSING A TOPIC AND A SUPERVISOR

If you are short of ideas for a project ask a lecturer for help. Perhaps the
lecturer has an idea for a project which may appeal to you. Find out if last
year's project reports are available for you to consult, to gain an idea of
the sort of topics which might be suitable. There are also books which
contain project suggestions (Appendix B).
Start thinking about your project early. You need to find an appropriate
supervisor who is willing and able to advise you. There may be only one
person who has the experience to supervise a project in the particular
subject area which interests you. However, if there is a choice, ideally you
should find someone whom you believe you will get on with and who will
also be available to give you feedback at critical times.
It is important to realize that your potential supervisor will probably
be a very busy person and you should always arrange to see him or her by
appointment and have a well-prepared list of points ready to discuss. A
supervisor will have his or her own other deadlines for teaching, research,
administration or pastoral care. He or she will be pleased to help you but
you should each be aware of the other's preoccupations because this will
help you to plan your joint use of time efficiently.
With this background in mind it is useful to find out whether your
potential supervisor will be available to give advice at intervals throughout
your project. Will you be able to meet for regular discussion? If not, is
there someone else who can help you if you need advice when your
supervisor is unavoidably absent?

S_E_E_K_A_D
__
V_IC_E_A_T
__T_H_E_S_T_A_R_T______________~I

L -_____________

Pay attention to any deadline for agreeing a project title and outline
with your supervisor.

11.2 WHAT CAN HAPPEN IF YOU DO NOT SEEK ADVICE AT THE


START
Your supervisor will usually have sufficient knowledge of the nature of
your study to be able to recommend a satisfactory method of obtaining
and analysing data. However, if there are any doubts he or she may
suggest that you should clarify your proposed methodology with a
statistician.
It is important to feel confident about your approach to the study.
You may be put off seeking advice at the beginning of your project either
because you have only just completed a first-year statistics course or, more
likely, because you have forgotten the statistical principles you covered in
your first-year course. The subject is still 'unreal' because you do not have
your own data in front of you. You may be afraid to reveal your own lack
of knowledge of statistics and experimental design or you may be
overwhelmed by the sight of mathematical symbols in textbooks. You may
even feel that deciding on your objectives and treatments is quite enough
of a problem and it is the last straw to discuss the" design.
The result of this approach is that you will collect a mass of data and
then have to face the awkward question: 'What can I do with these
results?'
There are four main ways in which the conversation with your project
supervisor will then develop:
At best. A satisfactory analysis is possible which yields reasonable
answers to the questions asked. (Unfortunately this is rarely the case.)
Second best. All is well but a more efficient design could have been used
and saved a great deal of work.
Third best. Analysis is possible but the experiment was not sensitive
enough to answer the question with sufficient precision. Perhaps more
replicates were required or extra measurements should have been
taken.
Worst. The project is a disaster because the experimental design was
badly flawed. Either no inferences can be made, or the results are not
relevant to the required questions. In other words, your work was a
waste of time and your project is a failure.
It is part of human nature to worry about what you don't know and to
fear that other people will ridicule your lack of knowledge. Think carefully
about what is concerning you, write down some key points, and make an

187

188

II~___________________y_O_U_R__PR__O_JE_C_T__________________~
appointment to ask advice as soon as possible. It is always a great relief
to share a concern and to be able to plan ahead with confidence.
Remember, biological material is variable. We need to draw conclusions
from experiments or surveys by induction from a sample to the whole
population. Statistical theory allows definite statements to be made with a
known probability of being correct.
We should note that rigorous statistical inferences can only be drawn
from properly designed experiments. In these, the experimenter has the
power to allocate different treatments to particular individuals. Observational studies in which comparisons are made between individuals which
happen to have naturally encountered different conditions are always open
to qualified interpretation.
11.3 GENERAL PRINCIPLES OF EXPERIMENTAL DESIGN AND
EXECUTION
In Chapters 3 and 4 we covered the importance of randomization,
replication and blocking in designing an experiment in some detail. We
will now see how these topics fit into the broader considerations involved
in designing your project.
11.3.1 Why are you carrying out this experiment?
OK - it is a project which fulfils part of the requirement for your degree.

However, are you just generally curious about whether changing the
nutrients which a plant receives will change its growth rate; do you want
to compare the effects of two or more ages of insect host on the rate of
parasite development; are you working in the controlled conditions of a
glasshouse in the hope that your results may help in the selection of
treatments for a subsequent field experiment?
11.3.2 What population are you studying?
You should ensure that the experimental material is representative of the
population about which you wish to make inferences. The plant ecologist
John Harper has pointed out the 'trilemma' experienced by an experimental scientist. He or she may seek:

Precision
Precision can be achieved by using unique genotypes and carrying out the
work in controlled environments. This should provide repeatable estimates
with narrow confidence intervals. But how relevant are the results to the
'real world'?

----li

E_X_P_E_R_IM_EN_T_A_L_D_E_SI_G_N_A_N_D_E_X_E_C_V_T_IO_N
____

L--_ _ _ _

Realism
This may be achieved by studying individual plants in the field. This leads
to low precision. Estimates will have very large confidence intervals
because of the large amount of random variation. Only large differences
between populations will be detected as statistically significant.
Generality
It may be desirable to have information about treatment effects on a wide
range of, for example, different soil types. With limited resources there is
a danger of sacrificing both precision and realism and of ending up with a
shallow study.
11.3.3 What are the experimental units and how are they grouped?
Experimental units (test tubes, plots, pots, plants or animals) should not
be able to influence each other and should be of a practical size and shape
for the study. For example, if you are interested in the weight gain made
by insects feeding on different species of plants, the experimental unit is
the plant. Sensibly, you may choose to have ten insects feeding on each
plant. Their average weight gain will provide a good estimate of the plant's
value as a food source but you should not be tempted to consider the
weight of each insect as an independent value because they have been
competing with each other for food, so you must use the mean weight of
the ten insects as the unit of assessment.
The plant pots should be large enough to allow for plant growth
throughout the time of the experiment and spaced widely enough apart so
that the plants do not shade each other.
It may be possible to group experimental units so that members of the
same group will experience similar background conditions. For example,
put the biggest insects or the tallest plants into group one, medium-sized
individuals into a second group and the smallest ones into a third.
Consider applying each treatment (and a control) to one individual
selected at random from within each of these groups. This will improve
precision (in a similar way to stratified random sampling). It will result in
narrower confidence intervals and in a greater chance of detecting
differences between treatment populations (if they exist).
Treatments should be applied and records should be made group by
group rather than treatment by treatment. In the analysis you would then
account for variation between these groups or blocks.
11.3.4 What are the experimental treatments?
Is there a naturally defined control? If so, it should be included. For
example, in estimating the effects of different levels of a chemical on an

189

190

II

YOUR PROJECT

~----------------------------------------------------~

organism it is wise to include the standard recommended amount(s) as well


as applying none (the control). In addition, you may well apply a wide
range of increasing amounts deliberately chosen to include a 'ridiculously'
high amount.
Consider carefully how treatments are related to one another.
Treatments may appear to be entirely unrelated to each other. For
example, you may apply six totally different chemicals. However, if you
consider their nature you will often find that there are natural groups
within the treatments. Perhaps among six chemical treatments there is one
'traditional' chemical whose effect should be compared with the average
effect of the other five 'new' chemicals.
In contrast, some treatment groupings may be very clear. They may
represent increasing amounts or levels of a factor (shade, water or
nutrients). It is important to consider the number of levels of the factor,
whether it is biologically important for them to be equally spaced or not,
and how many replicates there should be of each level. In this case the
analysis should concentrate on testing whether there is evidence for a
steadily increasing (or decreasing) response to increasing amounts of the
factor or, perhaps, for a curved response.
Finally your treatments may involve different combinations of factors.
As we have seen, the efficient way of selecting treatments is to include all
possible combinations of all levels of each factor (a factorial experiment,
Chapter 7).
11.3.5 Will randomization be across all the experimental units or
constrained within groups?
You need to allocate treatments to material at random to avoid bias and
to be able to generalize the results. Remember that 'random' is not the
same as 'haphazard'. To allocate treatments to experimental units at
random requires the use of random numbers. The experimental units are
numbered sequentially and the treatments are allocated to them in the
order in which the plot number occurs in a sequence of random numbers.
A plan of the treatment allocations should be made. This is helpful in
making records and may also be used to explain any unexpected
environmental variation which affects the results.
If you decided to group the experimental units into blocks (section
11.3.3) then you should allocate the treatments at random to the experimental units within each block.
11.3.6 What are the aims and objectives of the experiment?
The general aims (purpose) should be clear and the objectives should be
lucid and specific, and expressed as:

~________E_X_P_E_R_I_M_E_N_T_A_L_D__ES_I_G_N_A__N_D_E_X_E_C_U_T_I_O_N________~I
Questions to be answered. (How does pH affect reaction time?)
Hypotheses to be tested. (Null hypothesis is that there is no linear
response.)
Effects to be estimated. (The mean increase in temperature is 5C with
a 95% confidence interval of between 4 and 6C.)
If you have both a primary and a secondary objective you should make
sure that the design of the experiment is effective and efficient for the
primary objective and, ideally, also for the secondary objective.
Think carefully about what you will be measuring or counting. Some
variates will be relatively easy to observe: the number of live seedlings in a
pot, for example. Others may be more difficult: the leaf area of a plant,
say. Decide which variates are of most interest to you. If they prove to be
very time-consuming to record you may choose not to record any others.
Consider whether you will analyse each variate separately or whether you
will combine any of them before analysis. For example, you might decide
to multiply leaf width by leaf length to obtain an approximate estimate of
leaf area.
If you make the same measurement on each plant on several occasions
(for example measuring height), these will not be independent
observations; they are repeated measures. A simple approach is to
subtract, say, the first height from the last height and to analyse the
increase in height over time.
It may well be sensible to obtain some estimate of the size of individuals
before they have been exposed to any treatments. For example, the weight
of each animal or the height of each plant might be recorded and called a
covariate. These values can then be used to account for some of what
would otherwise be classified as simply random variation present in the
experimental results. This may well increase the precision of your

experiment. You should ask advice about how to make use of such
information.
11.3.7 How many replicates should you have?
This is a common question. To answer it you need to know or be able to
make a sensible guess at two things:
I. The minimum size of the difference between any two treatment means
which you would regard as being of practical importance (for example,
5 kg per plot difference in mean yields or 2 mm per day difference in rate
of growth).
2. The likely variability of the experimental material. A pilot study or
the results of other people's work on similar systems is useful here.
They will probably have presented an estimate of the variability in their
experiment in the form of a standard error of the mean or of a

191

192

I 1L-__________________y_O_U_R__P_R_O_JE_C_T__________________~
difference between two means, a confidence interval, or a least
significant difference. All of these contain the standard error of a mean
which can be used to calculate the standard deviation. To do this we
multiply the standard error by the square root of the number of
replicates which they used in their experiment. For example, if the
standard error of their mean was 0.5 kg and there were four replicates
then the standard deviation will be: 0.5 x -/4 = 1.0kg.
The standard deviation can be used to calculate the number of replicates
you need to stand a reasonably good chance of detecting the required
difference (if it exists) between two treatment populations. You may need
to ask a statistician for advice about this because the required calculation
depends on the design of the proposed experiment. If you have too few
replicates there will be no hope of detecting your required difference. If
you have too many replicates you may be wasting valuable time and
money. Your project will in all probability have a small budget and you
must remain within it.
Once the number of replicates has been fixed it is important to
remember to assess and record them separately. Sometimes people
inadvertently combine the information about one treatment from all
replicates. This makes the results impossible to analyse.
11.3.8 What resources (and constraints) do you have?
Will your project need to take place at a certain time of year? It is no good

proposing to study the physiology of hibernating hedgehogs if the


fieldwork must take place in summer! Similarly, many vegetation studies
are best carried out in spring and early summer when plants are coming
into flower and are relatively easy to identify. If your work is to be carried
out in a laboratory it may well be possible to be there only when it is not
in use for teaching classes.
Check that you can fit in your project satisfactorily with your other
responsibilities, allowing for your taught course timetable and deadlines
for continuous assessment work. Do you need transport to get to your
experimental site? If so, this needs to be organized and reliable. Does your
study need any equipment and, if so, is there any money to pay for it? If
equipment needs to be made, is this going to be possible in the time
available?
11.3.9 How should you record your results?
Records of results from any scientific study should be accurate and
unbiased. Design a record sheet which is clear and easy to use, both when
recording and when putting data into the computer (Chapter 4). Each

EXPERIMENTAL DESIGN AND EXECUTION

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

II

experimental unit should have a unique number. Ideally, this should be


used when recording so that you are unaware of the treatment at that
stage.
Make sure you know exactly what you are going to record before you
start. Make a note of anything odd. Check for gross errors (for example
an ant which weighs 200 g) and keep the number of significant figures to a
minimum.
If you are working out of doors, be prepared for the wind and rain. A
clipboard on which the record form is secured by bulldog clips and which
is placed in a polythene bag to protect it from drizzle is helpful. A pencil
(and a pencil sharpener) is best for writing on damp paper. If you make a
mistake, cross out the incorrect value and write the replacement one
clearly labelled to one side of it.
You may have access to a tape recorder. This seems ali attractive option
but they are not infallible and may break down. Also, it may be more
difficult to keep track of where you are in your sequence of recording.
More recently, hand-held data-loggers (small computers) have become
popular. Data collected on them can be transferred straight into a
computer package like MINITAB via a cable connection between the two
machines. They are very valuable if your work involves regular recording
of the same experiment. Otherwise the effort required (usually from an
expert) to write the program to prompt you for the appropriate data in
sequence may be out of proportion to the possible gain in time.
Most importantly, keep a copy of results in a separate place. Remember,
what can go wrong will go wrong. It is a good idea to use a notebook
whose pairs of numbered sheets can be separated by a piece of carbon
paper. The duplicate can be torn out and immediately stored in a separate
file. Alternatively, use a hardback notebook and make regular photocopies
of each page of notes.
Ideally, transfer your results to a computer file immediately after you
have recorded them. You should have a minimum of two computer disks
with your data on them. One is a backup copy. Disks occasionally fail or
can be damaged. Professionals have copies of data on three disks. This is
to guard against the following possible (if very rare) sequence of events.
You find that the computer cannot read a file from your disk. You
assume that your disk has failed and put in your second disk.
Unfortunately, the computer cannot read the file from this disk either. It is
only now that the possibility dawns on you that perhaps the computer's
disk drive is damaging the disks. If you have a third copy on another disk
you know that your data are still safe and you will not put this disk into
the same machine.
Another reason for having a third disk is that it is good practice to keep
one copy in a separate place, in case you lose the two other copies while
they are kept together, through fire or theft. If there is a third disk with a

193

194

II

YOUR PROJECT

~----------------------------------------------------~

friend you can update it, say, once a week as an insurance policy. Also, if
you have a brainstorm and remove copies of a file from both of your dayto-day disks in error (easy to do when you are tired), your third disk is
not there so, thankfully, you cannot make the same mistake with that one
immediately. If ever you delete a file and regret it because it was your only
copy, ask for help from a computer expert immediately. As long as you
have not tried to save anything else onto the disk it will usually be possible
to retrieve the file.
It is vital to make comprehensive notes of everything that happens as
you go along. It is infuriating trying to puzzle out exactly what you did
and when some months after the event. Take a few photographs of key
features to illustrate the report. For example, striking differences in
appearance between the treatments or unexpected events like a flood on
part of the site may help to brighten your presentation. Ask a friend to
take a photograph of you applying treatments and recording results. When
you read a useful paper, make a note not only of its contents but also of
its full bibliographical reference (Chapter 12). You may wish to store the
references in a file on your word processing package on the computer so
that you can edit a copy of this file at the end of your project to provide
the references section of your report.
11.3.10

Are your data correct?

It is tempting to assume that you haven't made any mistakes either in

recording the data or in transferring them to the computer. This is highly


unlikely. It is essential to carry out a systematic check.
First, are all the data present, correctly labelled and in sequence? Are
any codes you have used in the correct range? For example, if you have
fixed treatments coded 1 to 5 there should not be a 6, and if there are six
replicates of each treatment there should not be five codes of '1' and seven
codes of '2'. What code have you used for a missing value? (MINITAB
uses a *.) Are the missing values correct or should some of them be zeros?
Are any of your observations too big or too small? Obvious errors can be
spotted by asking for boxplots of each variate, when they will appear as
outliers. Plots of repeated observations over time can also reveal potential
errors. For example, it is highly unlikely that a plant will really have
shrunk in height over time.
If you can persuade a friend to help you it is good practice for one of
you to call out data from your original notebook and for the other to
check it against a printout from the computer file. This will identify any
transcription errors. Whenever you have edited your data consider
whether you wish to overwrite the previous, incorrect file of data or to save
it as a separate file. The number of files you have can grow surprisingly
quickly. When you start all seems clear; two months later it is difficult to

S_V_R_V_E_y_D
__
E_SI_G_N__
A_N_D_E_X_E_C_V_T_I_O_N____________~I

L -_ _ _ _ _ _ _ _ _ _ _

remember what is in each file. You should plan a consistent system of


naming your files. For example, 'elarea05' (there is commonly a limit of
eight letters or numbers for filenames) might stand for experiment 1, leaf
area in May. It is important to keep a note of what each file contains and
to what extent the data have been checked.
11.3.11 How will you analyse your data?
You should define the specific questions which interest you as particular
comparisons before carrying out the experiment. For example does
treatment 3 have the same effect as treatment 4? Check that you
understand the proposed method of analysis and its constraints and that
you will be able to carry it out using whatever you have available, whether
this is a calculator, MINITAB on a personal computer, or even perhaps
a more advanced package like Genstat, SAS or SPSS. Becoming familiar
with your computer package will take longer than you think but will prove
very satisfying in the end. If it will take some time before your experiment
will produce results, practise the method of analysis using invented data
(call your datafile 'rubbish' so that you don't later incorporate the results
because you think that they are genuine). This will enable you to gain
confidence.
11.4 GENERAL PRINCIPLES OF SURVEY DESIGN AND
EXECUTION
Most of the principles we have looked at with respect to experiments also
apply to surveys. However there are some points which should be
especially stressed as well as some important extra points to consider.
We will illustrate these points by reference to a proposed project
involving a survey of the popularity of a collection scheme for different
types of waste (glass, paper, batteries, drinks cans and kitchen waste) from
households in a small village. The aim is to estimate the level of
cooperation in the scheme; to collect views on how it might be improved;
and to see whether participation in the scheme is related to any
characteristics of the households (for example, age of residents and
approximate income band).
11.4.1 Is there any bias in sampling?
You may decide to use the voting register to provide a list of all the
households in the village. This might be out of date and so exclude the
latest housing development from your population. If you decide to sample
every tenth household it may be that you accidentally select only ground-

195

196

I LI__________________y_O_U_R__P_R_O_JE_C_T__________________~
floor flats because flats are grouped in two-storey blocks with five flats
per storey. A random sample is preferable.
If you visit each selected household you may find that in many cases
the inhabitants are out when you call. It is important to call back many
times (perhaps at a different time of day) to ensure that you catch them in.
For example, if you call only between 9 am and 5 pm on a weekday your
sample will under-represent households where all adults are working
outside the home. Some people may not wish to take part in the survey. It
may be easier, quicker (and safer?) simply to deliver a questionnaire to
each selected household. In this case there should also be a covering letter
which clearly states where you come from and how you may be contacted
as well as the purpose and importance of the survey, together with a
stamped addressed envelope for the return of the completed questionnaire.
It will undoubtedly be necessary and worth while to send duplicate
questionnaires and return envelopes to those who have not responded by
the date requested. This will minimize non-response errors; perhaps those
who tend not to respond quickly are also those who tend not to participate
in the waste collection scheme?

11.4.2 Is your questionnaire well designed?


The design of the questionnaire must be carefully considered. People will
be put off if it is very long or if it contains sensitive or ambiguous
questions. It is essential to tryout your questionnaire on some of your
friends first and then on a few villagers to identify any ambiguities from
the respondents' point of view.
Ideally, most questions should be answered by selecting an answer from
a I to 5 scale. For example, from 1 =1 never do this, to 5 =1 always do
this; or 1 = yes and 2 = no. Boxes beside each question on the right-hand
side of the page are used for coding the answers (for example, 1 to 5 for an
answer and 9 for a missing value or don't know) ready for entering the
data into the computer (Figure 11.1). In addition to such 'closed'
questions, you may wish to include a few 'open' questions in which the
respondent can say what he or she thinks about the topic. In this case we
might ask: 'In what ways do you think that the waste collection service
might be improved?'
It is important to realize that even if you obtained completed
questionnaires from all the households in the village (a complete
enumeration) there would still be errors in your data. These would not be
sampling errors because we have data from the entire population, but they
would be response errors.
In a postal questionnaire response errors may be caused by the
following:

0
0

0 8

0
0
4

5.

4.

0
0
0
0
0
0
0
0
0

0
0
0
0

Please return your questionnaire in the


enclosed stamped addressed envelope.

Thank you very much lor completing this


questionnaire. We will display a copy of our
report in the village shop in May.

....

Terrace house
Semi-detached house
Detached house
Flat
Other (please specify)

What type 01 house do you live in?


Please tick one box

Newspapers
Magazineslwaste paper
Cardboard
Glass
Batteries
Aluminium cans
Tin cans
Kitchen waste
Garden waste
Other (please specify)

Please place a tick in the box next to each


type of waste material which you have
placed out for collection in the last month.

Questionnaire about recycling household waste in a small village.

Please tick one box

Do not have any suitable material


Too much trouble to sort material and have
it ready lor collection
Please tick one box
Any other reason?
Please specify here:

material for collection by the local


recycling group:

Figure 11.1

L_

Yes
No
If yes go to question 4
If no go to question 3

Do you put out any waste material for


collection by the local recycling group
(not the dustmen)?

How many people live in your house?


Children (under 18 years old)
Adults
Please put a number in each box

3. Please say why you do not put out any waste

2.

1.

000

Office
use only
Ref 1-3

020
021
022
023

o 19

0 9
010
o 11
o 12
013
014
015
016
017
018

c:::

CZl

z<3

>-l

c:::

\.)

tTl

:><

tTl

:>

CZl

atTl

-<
~

:;d

198

I LI___________________y_O_U_R__PR__O_JE_C_T__________________~
1. Misreading handwriting (age 23 instead of 28).
2. Lapses of memory. (When did you start participating in the scheme?)
It is better to give a range of time periods: less than six months ago;
between six months and one year ago; and so on; and ask people to
select one.
3. The tendency of people to want to appear 'socially acceptable' and so
to overestimate, say, the proportion of household waste they put out
for recycling.
A pilot study will help to minimize response errors but it is also
important to try to validate the answers independently. For example, you
could monitor the actual amount of each type of waste collected from the
village each week for a few weeks.
11.4.3 Analysis of questionnaire data
In many cases the results of the questionnaire may simply be presented
as the percentage of respondents who fall in a certain category (e.g. '55%
put out waste for recycling'), but ensure that the total number of
respondents is also given (an estimate based upon 200 respondents will be
more precise than one based upon ten). However, if it was the intention
to compare the characteristics of different types of respondent, this can be
done using a X2 contingency test (section 9.5). (Remember to use the actual
numbers not percentages.)
Suppose we obtained the following results:
Type of house

Recycles waste

Does not recycle waste

1. Terrace

30
24
14

20
15
3
30

2. Semi-detached
3. Detached
4. Flat

10

Analysing the whole table of results in MINITAB:


Expected counts are printed below observed counts
Yes
30
26.71

No
20
23.29

Total
50

24
20.84

15
18.16

39

14
9.08

3
7.92

17

10
21.37

30
18.63

40

S_V_R_V_E_y_D
__
ES_I_G_N_A
__
N_D_E_X_E_C_V_T_I_O_N____________~I

L -___________

Total

78

68

146

ChiSq=O. 405+0. 464 +


0.481+0.551+
2.663+3.054+
6.049+6.939=20.606
df = 3, p = 0.000

The p value is less than 0.001, indicating very strong evidence that there
is a difference in the behaviour of people living in different types of home.
The components of l which come from the fiats (6.0 and 6.9) are by far
the largest, suggesting that this group differ from the other three. We can
break down the table in two ways to examine this further. First we can
group together the first three types of home (group 1) and compare the
results from their respondents to those from the fiats (group 4):
Expected counts are printed below observed counts
Yes
68
56.63

No
38
49.37

Total
106

10
21. 37

30
18.63

40

Total

78

68

146

ChiSq=2.283+2.618+
6.049+6.939=17.890
df=l, p=O.OOO

Again, there is very strong evidence for a difference in behaviour,


namely that a smaller proportion of those living in fiats recycles waste
(10/40 =25% compared with 68/106 =64%). Perhaps those in fiats are not
prepared to take the rubbish down many flights of stairs? We can refer

back to their answers to question 3 to see whether this is an important


factor in their decision.
Finally we can exclude those living in fiats (group 4) from our analysis:
Expected counts are printed below observed counts
Yes
30
32.08

No
20
17.92

Total
50

24
25.02

15
13.98

39

14
10.91

3
6.09

17

Total

68

38

106

ChiSq:::;0.134 +0 .240 +
0.041+0.074+
0.878+1.571=2.939
df=2, p=0.230

199

200

I ~I___________________Y_O_U_R__P_R_O_JE_C_T__________________~
Now, the p value is 0.23 so there is no reason to doubt our null
hypothesis of no difference in behaviour between those living in the
remaining three types of home.
11.4.4 Ethical considerations
Proposed experiments in medical research (clinical trials) must be
approved by an independent committee of responsible people. For
example, the allocation of treatments at random to people who are
suffering from a disease may not be considered to be ethical if there is
already evidence that one of the treatments is likely to be more efficacious
than another. There are also strict controls on research work involving
animals. Such matters are the responsibility of your supervisor.
However, there are also more general ethical considerations. It is
unethical to waste resources, to carry out a badly designed experiment, or
to surveyor to analyse it incorrectly and so mislead those reading a report
of your results.
11.5 HEALTH AND SAFETY
Your institution will have guidelines which you must read and follow.
There will be booklets on this subject available from your library. Half-anhour spent considering possible problems in advance is time well spent.
Accidents do happen but sensible precautions will minimize the risk of
their occurrence.
11.5.1 In the field
Fieldwork can be dangerous and you should discuss with your supervisor
whether you need to be accompanied or not. You should always leave a
note of your route with a responsible person, together with a time by
which you should have contacted them to confirm that you have returned
safely. You should also always obtain permission to enter and work on
private land (this includes nature reserves). Plants must not be dug up and
removed from a site.
Wear sensible protective clothing and take a map, compass (which you
know how to use!), whistle, first-aid kit, and emergency food supplies with
you. If your last tetanus inoculation was ten or more years ago it is
sensible to have a booster in case you cut yourself. If you carry out fieldwork near water which is frequented by rats you should be aware of the
possibility of contracting leptospirosis. If you work in areas where there
are sheep you may risk contracting Lyme disease which is spread by means
of ticks. Such diseases may first express themselves by 'flu-like' symptoms,

----li I

H_E_A_L_T_H_A_N_D_S_A_F_ETY
________

L - -_ _ _ _ _ _ _ _

so be aware of these possibilities and contact a medical doctor if you have


any suspicions.
11.5.2 In the laboratory

If work in a laboratory is a necessary part of your project you should


already have a reasonable amount of experience in this area. You should
meet the laboratory supervisor, with your supervisor, and discuss the
proposed work. If you plan to use any chemicals which might pose a
hazard to health you should complete, in conjunction with your supervisor, a COSHH form to outline the procedures which will be followed.
COSHH stands for the Control of Substances Hazardous to Health. The
idea is that any possible hazards in your proposed work should be
identified. The risk involved is then assessed by considering both the
perceived hazard and the probability of its occurrence. For your own
protection you should:

Make full use of protective clothing and equipment such as overalls,


laboratory coats, safety shoes, dust masks, machine guards.
.Make full use of relevant 'control measures' - dust extractors, fume
cupboards.
Report to your supervisor any shortcomings in control measures or
protective clothing.
Ensure that you understand the job you are about to do and the risks
involved.
If you are in any doubt seek advice from your institution's Safety
Officer.

201

Preparing a report - What's


it all for?

12

This chapter provides some guidelines on how to produce a project report.


Do check with your supervisor for any particular requirements about
presentation and deadlines for your project report. These will usually be
listed in a reference document.
There are numerous books and articles which provide advice on how
to produce a written report on an investigation (Appendix B). It is
valuable to consult several but, since procrastination is one of the most
likely problems to beset a writer, you should not get so carried away that
the report doesn't appear. The key thing is to put something down on
paper. It then provides raw material for revision. A large notebook in
which you jot down dates, thoughts, references, methods and results as the
study proceeds is invaluable. But remember, the aim is to produce a report
by the deadline.
12.1

COMPUTERS

The utility of using a word processing package cannot be over-emphasized.


The technique is not just an up-market typewriter on which mistakes can
be corrected easily; it allows you to be more creative. It is easier to
compose at a keyboard which allows passages to be exchanged, inserted
and deleted at will than it is to do so with a pen. If a computer is a
relatively new experience for you, do take advantage of the opportunity to
use one for your project. When you receive comments on a draft from
another student (or in some institutions, from your supervisor) you can
edit your report very easily.
Computer packages such as MINITAB enable you to store, edit and
analyse your data. You can import summary output from MINITAB into
a word processing package like 'Word' or 'WordPerfect'. These can also
check your spelling, count your words and suggest alternatives for over-

BASICS

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

II

used ones. These or other computer packages will check your grammar
and suggest how to improve your style.
In recent years computers have revolutionized access to the literature.
It is now possible to search databases for published papers which include a
keyword or set of keywords. For example, you might ask for references
to papers which include the keyword 'blood' or only for those which
contain the phrase 'blood pressure'. You can obtain a copy of the output
on disk, so that you can read the titles and abstracts at leisure and edit out
the unwanted references before printing out details of the interesting ones
you wish to follow up.

12.2 BASICS
Writers' block is very common. The only way around it is to work back
from the final deadline and set down a schedule of work to achieve it. This
should include some spare time to allow for emergencies.

12.2.1 Structure
A report should have a clear structure. If you are not sure what that
structure should be, start with the basics. Useful section headings are as
follows:
Introduction

This presents the background to the study, and reviews the relevant
published literature before outlining the general aims and specific
objectives of the project.
Materials and methods

This contains a description of:

any pilot study,


the experimental or survey design,
the study populations,
the treatments (including any control),
the reason for choosing the number of replicates,
the method of randomization,
the methods of accounting for other sources of variation,
the criteria for assessing the outcome of the study,
any possible sources of bias,
any errors in execution (for instance missing observations),

203

204

II

PREPARING A REPORT

~----------------------------------------------------~

where and when the study took place,


a description of the apparatus and of any analytical methods used (or
a reference to where these are fully described),
statistical analyses (or a reference to where these are fully described).
You should also state the statistical package which you used (if any).
The idea is to enable someone else to be able to repeat your work, if they
so desire.

Results
This presents the data obtained with a brief explanation of the major
trends revealed. You should illustrate the results section with tables and
graphs (see section 12.3) and with reference to statistical analyses as
appropriate. Ideally you should provide exact 'p values' rather than simply
'p < 0.05'. It is important to concern yourself with your original
objectives. It is bad practice to 'dredge the data', in other words to make
every possible comparison between treatments in the hope of coming up
with a 'significant' difference!

Discussion
How do the results affect our existing knowledge of the subject? The
discussion provides the setting for interpretation of your results. You
should try to avoid repeating the results in this section. Here you may be
more imaginative and speculative in your writing, perhaps outlining
possible reasons for the findings and giving suggestions for further work.
It is important to check that your conclusions are justified by the analysis
of the data and that you are not trying to make the results fit some
preconceived idea of what you think ought to have happened.
This section is the place also to compare and contrast your results with
those of previous studies, perhaps suggesting reasons for any discrepancies. It is important not to attempt to extrapolate the results of a
sub-population which you have studied to the whole population. For
example, what you have discovered about the behaviour of worms on the
local farm may well not be true for worms of a different species, or on a
different soil type, or 200 kilometres further north.

References
This is a list of the material cited in the report. You should quote sources
of reference from the published literature or personal communication in
the text where appropriate and list them fully in the references section.
You may be instructed to follow a particular 'house style'. This may seem

---'I I

B_A_S_IC_S_ _ _ _ _ _ _ _

L - -_ _ _ _ _ _ _ _

fiddly to you but it is essential to give accurate references and to doublecheck that the list contains all the required references and no others.
Consider how annoyed you would be if you wanted to follow up an
incorrect reference.
It is important to give credit to other people for their work. You can
refer to their publications in your text by giving the surname(s) of the
author(s) and the year of publication:
Our findings agree with those of Watt (1992) and differ markedly
from the results of work in the USA (Smith, 1990; Jones and Smith,
1994; Smith et al., 1995).
Notice how grouped references in the text are given in date order rather
than alphabetical order of first author's name. Also, it is common for
publications with three or more authors to be referred to by the first
author's name followed by 'et al.' at the second or subsequent time of
mention. This is an abbreviation for 'et alia' which is Latin for 'and others'
and is written in italics to indicate that it is in a foreign language.
If you attempt to represent other people's work as your own, either by
copying parts of their text (with or without editing it), or by rephrasing
their ideas without acknowledging the source, you are guilty of plagiarism.
At best this is bad manners and at worst it is criminal. If you are
submitting your report as part of the requirements of a degree you will find
that your institution has clear rules forbidding plagiarism and stating the
disciplinary action which may be taken if your report contains such
material. Failure to acknowledge sources is also unprofessional in that you
should enable your readers to read the original material if they wish.
In the list of references at the end of the report it is important to provide
enough detail to enable readers to locate the works. There are different
ways of doing this and journals will specify their particular 'house style'. A
commonly used system has the following formats for three different types
of publication:
Paper in ajournal
Bannister, N.R., and Watt, T.A. (1995) Effects of cutting on the growth
of Crataegus monogyna (hawthorn) in hedges. Journal of Environmental
Management, 45, 395-410.
Chapter in a book
Bannister, N.R., and Watt, T.A. (1994) Hedgerow management: history
and effects. In Hedgerow Management and Nature Conservation, eds
T.A. Watt and G.P. Buckley, Wye College Press, Wye, Kent, pp. 7-15.
Book
Watt, T.A. (1997) Introductory Statistics for Biology Students, 2nd edn,
Chapman & Hall, London.
Notice how the surname(s) and initials of the author(s) are always

205

206

I LI________________P_R_E_P_A_R_I_N_G_A__R_E_P_O_R_T_________________
followed by the year of publication. The title of the document follows with
the journal or book title then being given. Italics are used for journal or
book titles. Editors' surnames and initials are given for books with
chapters written by different authors and the publisher's name and place
of publication are included for the books. Finally, page numbers are
required where only part of a book or journal is relevant.
The best way to become familiar with these principles is to read journal
articles and to see how to summarize information clearly and to present
cogent arguments justified by relevant references. Then have a go for
yourself!
Acknowledgements

You should thank all those people or organizations who have helped
you: your supervisor, members of the statistical and computing staff, the
laboratory supervisor and your friends. Don't forget the land-owners if
you have been working in the field and do send them a copy of your
finished report.
Appendixes

These are optional and contain items which are not required to follow
the flow of the argument. Examples include: raw data, calibration curves,
species lists, mathematical equations and statistical tables.
Abstract

A brief statement of aims, methods and results is helpful in all but the
briefest of reports. It should focus on the 'take-home message'. This may
well be the last part which you write. In many journals this is placed at the
beginning of the paper so that it can be readily scanned by those who
may not have time to read any more.

12.2.2 Drafts
When the above outline is in place, a first draft can appear. At this point
it becomes obvious that there are gaps which you need to fill or that points
should be made in a different order to emphasize connections. It is
important to immerse yourself in this task so that the momentum is
maintained. Write 'REPORT WRITING' in your diary and put a notice
saying 'REPORT WRITING - KEEP OUT' on your door! A first draft
will say things like: **must find comparative figures** or **check
analysis** or **find reference** or **yuck! - rewrite this!**. These points
are attended to in preparing a second draft which should then be given to

IL_L_U_S_T_R_A_T_IN
__
G_R_E_S_U_L_T_S______________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _

a friend whose opinion you value and (if it is the custom) to your
supervisor for comments.
This is a vital stage and such comments wi111ead to great improvements
in the clarity of the third draft. Don't be surprised if the reviewer covers
it with comments in red ink. We all have to endure this because it is easier
for others to spot our errors and inconsistencies than it is for us to see
them ourselves. You can usefully have your revenge by constructively
criticizing another student's draft report.
As you are writing you should consider who will read your report. What
will they know already? What do you want them to understand and to
remember? Browse through the literature in your field. When you find a paper
which is clear, brief and stimulating, use it as a model for your approach.
When you find yourself absorbed in typing your report onto a computer
you should be sure to take a short break after an hour's work, and a long
break after a further hour, otherwise you will find that your body will
protest, your mind become tired, and you will make mistakes. If you
find that your neck or your arms are beginning to ache ask advice about
your body position relative to the computer. Perhaps the chair needs
adjustment.
12.3 ILLUSTRATING RESULTS

The question of whether to use graphs or tables can be a difficult one to


resolve.
12.3.1

Graphs

Graphs are excellent for communicating qualitative aspects of data like


shapes or orders of magnitude. They are invaluable in exploring the data.
For example, they can help you to decide whether a straight-line model is
sensible or whether there are outliers. Graphs are also helpful in showing a
broad picture, or in reminding a reader of principles.
It is important to label the axes of the graph, giving the units of
measurement, and to present a suitable legend (caption). When you
present a graph you should also include a measure of the variability of the
data. This is commonly achieved by presenting a vertical bar labelled 'SE
mean' (standard error of a treatment mean after analysis of variance,
Chapter 7). Remember that when each treatment has an equal number of
replicates the same SE mean applies to each treatment and so only one SE
bar is needed on the graph (Figure 12.1). Alternatively, you may put a
least significant difference (LSD) on your graph or you may put 95%
confidence intervals around each mean. It is important to make clear
which of these three possible measures of variability you have used.

I 207

208

I LI________________P_R_E_P_A_R_I_N_G__A_R_E_P_O_R_T________________~
10
Yield
t/ ha

SE meanI

OL-__-L-L____
control

~~

____

organic

~~

___ _

conventional

fertilizer

Figure 12.1

The mean yield of barley (tha-') from each fertilizer treatment.

12.3.2 Tables
Although graphical plots can show general trends, tables are required to
convey quantitative features. Ideally, a good table will display patterns
and exceptions at a glance but it will usually be necessary to comment in
the text on the important points it makes. Some useful guidelines for
producing tables have been suggested by Ehrenberg (1977):
1. Round data to two significant figures. This refers to digits which are
'effective', in other words, which vary from one observation to
another.
2. Provide row and column averages or totals on the right-hand side of
the rows and at the bottom of the columns.
3. If you want to compare a series of numbers it is easier to do so if they
are in a column rather than in a row. This is because the eye finds it
easier to make comparisons vertically than horizontally.
4. Ordering columns and rows by the size of the values in each cell helps
to show any patterns.
5. Single spacing between cells guides the eye down a column and gaps
between rows guide the eye across the table.
If you have carried out a factorial analysis of variance there is a neat
way of presenting the treatment means in a table, with their standard
errors. You may remember that in Chapter 7 we discussed a wholly
randomized design experiment with two levels of sowing and two levels of
cutting the field margins. The factorial analysis of variance of the mean
number of spiders per quadrat in each plot produced an error or residual
mean square of 1.333 (Table 6.6). Each of the four treatments had four
replicates, so each main effect has eight replicates. Use this information
and the formula for calculating a standard error (above) to check
Table 12.1.

E_X_E_RC_I_SE________________~I

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Table 12.1

plot

The effect of sowing and cutting on the mean number of spiders per

Cut once
Cut twice
Mean

Sown

Unsown

Mean

19.5
15.0
17.25

16.5
13.0
14.75

18.0
14.0
16.0

SE sowing mean
SE treatment mean
12.4

SE cutting
mean
0.408

0.408
0.577

LANGUAGE

The report must be legible and well written. It is normal in scientific


reports to write in the past tense rather than in the present tense and to use
'we' rather than 'I'. It is traditional to use the passive rather than the
active voice although many journals are now starting to encourage the
latter. So you should write 'The vegetation was sampled (or we sampled
the vegetation) using five randomly positioned quadrats in each plot',
rather than 'I will choose five quadrats selected at random from my field
plots'.
Generic names of organisms take an upper-case initial letter and
subspecific names commonly have a lower-case initial letter and both are
always in italics (or underlined): Beta maritima ssp. maritima. To be
complete you should follow the binomial name by its authority (an
abbreviation for the name of the person credited with describing the
species first): Lolium perenne L. (L. stands for Linnaeus). You should
check spelling with a standard list and state which list you have used in the
'Materials and methods' section.
12.5

EXERCISE

EXAMPLE OF A SHORT REPORT

Read the following example of a short report and comment on its


strengths and weaknesses.

209

210

II

P_R_E_PA_R
__
IN_G
__
A_R_E_P_O_R_T______________~

L ________________

The effect of fertilizer on forage rape: a pilot study


A.N. Other
Small College, Largetown, Midshire
Abstract

An experiment was carried out to investigate the response of forage


rape seedlings (Brassica napus L. ssp. oleifera) to the major nutrients
on a soil which was thought to be nutrient-poor. It was concluded that,
for this soil, the addition of 67.5 g m -2 of aN: P: K 7: 7 : 7 compound
fertilizer would be adequate and that the effect of lower levels should
be investigated in a subsequent pot experiment before carrying out a
field experiment.

Introduction
Forage rape (Brassica napus ssp. oleifera) is a crop grown for feeding
to livestock in autumn and winter. An average fertilizer requirement
for the crop is 125 kgha- I N, 125 kgha- I P20 and 125 kgha- I K 20
(Lockhart and Wiseman, 1978). We wished to investigate the optimal
amount of a compound fertilizer containing these nutrients to apply
to seedlings growing on what was reputed to be a very nutrient-poor
soil. A pilot study was carried out to find out the range of amounts
of fertilizer which should be used in a more detailed pot experiment
before carrying out a field experiment.
Materials and methods
On 10 September 1996 a 1 cm layer of gravel was put into each of
80 pots (lOcm x 10cm in cross-section and 12cm deep). The pots
were then filled with sieved soil from a nutrient-poor site near
Largetown, Midshire. There were 20 replicate pots for each of four
fertilizer treatments: the equivalent of 67.5, 101, 135 and 270gm-2
(1 gm- 2 is equivalent to 10kgha- l ) of Growmore granular fertilizer
(N : P: K 7: 7 : 7) which was ground to a powder and mixed throughout the soil. The recommended rate for an 'ordinary garden soil' is
135 gm- 2 The treatments were identified by colour codes (unknown
to the assessors to prevent bias) painted on the pots. The pots were
arranged in a wholly randomized design in the glasshouse.
Ten seeds of Brassica napus were sown in each pot and covered
with a thin layer of soil. The pots were examined daily and watered
using rain water as required. The plants were thinned to six per pot
on 16 September and to four per pot on 20 September.
At harvest on 15 October the number of plants per pot was
counted and their health scored (0 = perfect to 5 = dead). Plant
height and the number of true leaves per plant were noted and leaf

EXERCISE

L -____________________________________________________

width and length (cm) were put into a formula to obtain an estimate
ofleaf area (cm 2 ) (H. Moorby, pers. comm.):
LA = -0.35 + 0.888(LW x LL)
where LA=leaf area, LW=maximum leaf width and LL=midrib length.
Shoot material from each pot was put into a labelled paper bag
and oven-dried at 80 a C for 48 hours before being weighed.
This report concentrates on the effect of fertilizer on shoot drymatter yield. The data were analysed using analysis of variance in
MINITAB Release 11 (Minitab Inc., 1996).
Results
By 16 September germination had occurred in all pots but with
increasing fertilizer there were fewer plants and these were of smaller
size and were yellower. There was very strong evidence that fertilizer
affected shoot dry-matter yield (p < 0.0001, Table 1).
Table 1 Analysis of variance for total shoot dry weight per pot
Source
Fertilizer
Error
Total

DF
3
76
79

SS
l.l0257
0.94326
2.04584

MS
0.36752
0.01241

F
29.61

p
0.000

Shoot dry weight decreased with increasing fertilizer (Table 2,


Figures 1 and 2).
Table 2 Mean shoot dry weight yields per pot
Fertilizer (gm- 2)
Dry weight (g)

67.5

0.4770

101

0.3655

135

0.3235

270

0.1505

SEmean
0.02491

Boxplots of shoot dry weight


with increasing fertilizer

0.8
0.7
shoot
d,y
weight
g/pot

0.6
0.5
0.4
0.3
0.2
0.1

B$ $

0.0
67.5

10 1

135

270

Fer ti!izer g/sq m

Figure 1 Boxplots of the total shoot dry weight per pot for the four
fertilizer levels. There was a negative linear relationship between leaf area
and shoot dry weight.

I I

211

212

I LI________________P_R_E_PA_R_I_N_G__A_R_E_P_O_R_T________________
Shoot dry weight against leaf
area, labelled by fertilizer level
0.8
0.7 o

0.6 shoot

0.5 -

weight

0.4 -

dry

g/pol

o
o
x

00

t++ +

0.3 -

fertilizer
g/sq m

o 67.5
+ 101
x
135
* 270

0.2 0.1 0.0 0

50

100150200250300350
leaf area per pot

sq em

Figure 2 Graph of shoot dry weight per pot against leaf area per pot,
labelled by fertilizer level: A=67.5, B= 101, C= l35 and D=270g m- 2

Discussion

The soil tested was of a surprisingly high nutrient status as indicated


by the performance of the plants in the pots receiving the lowest
amount of fertilizer, which gave the highest yield. Greater amounts
had a toxic effect. Perhaps even less than 67.5 g m-2 would give an
even higher yield. For the next pot experiment, levels of fertilizer
should cover the range 0 to 120 gm- 2 It is possible that plants grown
for longer than one month might have a greater fertilizer requirement.
Acknowledgements

I thank the glasshouse staff for maintenance of the plants and the
first-year students on Course 199 for harvesting them and recording
the data, with help from P.Q. Smith. I thank E.O. Jones for
computing advice.
References

Lockhart, J.A.R. and Wiseman, A.J.L. (1978) Introduction to Crop


Husbandry, 4th edn, Pergamon Press, Oxford, p. 302.
MINITAB Inc. (1996) MINITAB Reference Manual, Release 11 for
Windows, MINITAB Statistical Software, Pennsylvania, USA.

Appendix A
Choosing how to analyse data
from a replicated, randomized
experiment

First plot the data - histograms, stem-and-Ieaf diagrams, boxplots, plots


of one variable against another. Then choose from the options below.
1. Data are continuous (kg, mm, etc.; or, if they are counts, these are large
values per experimental unit). Use analysis of variance (with blocks, if
appropriate) to compare treatments:

Consider whether there is a factorial structure - if so, account for


main effects and any interactions in your model.
Consider whether a factor is present at increasing levels - if so, use
linear regression to see whether there is a linear response or not.
Test the assumptions of analysis of variance/regression by plotting
histograms of residuals (to check Normality) and residuals against
treatments (to check for similar variability).
If the data do not meet the assumptions, either ask advice about
transforming your data or consider a non-parametric test (see next).
2. Data are scores or ranks or are continuous but from skewed populations.
Use non-parametric equivalents of analysis of variance (MannWhitney for two treatments, Kruskal-Wallis or Friedman for more
than two treatments, the latter with blocks).
3. Data are categorical (each observation can be placed in a category).
Use a chi-squared contingency test to compare the proportion of
individuals in a category for the different treatments.
It may help to imagine that someone has presented you with some data
and asked you: 'How should I analyse these data?'

214

I I

APPENDIX A

12
14
20

15
18
24

21
25
32

32
29
31

You need to ask some questions.


First what do the numbers represent? Are they from 12 independent
experimental units (animals, plants, Petri dishes)? Are they continuous
measurements (e.g. g or cm) or counts (e.g. number of aphids per plant).
Or, if they are counts, does the grand total of all the numbers represent the
number of independent observations (e.g. 273 people were asked questions
in a survey). You should also consider the possibility that the data are in
the form of percentages.
If there are only 12 independent values then what do A-D and E-G
represent? If A-D are four treatments, are E-G three replicates from a
completely randomized experiment or do they represent blocks? In either
case an appropriate analysis of variance (for example considering whether
the treatments have a factorial 2 x 2 structure) or non-parametric
equivalent (Kruskal-Wallis or Friedman) will be required. The comparisons of interest are those between the column means (treatment
means).
Again, if the data are from a survey of 273 people we need to know
about the classifications. Perhaps E-G represent the responses 'Like,
Indifferent, Dislike' to a particular question. The columns A-D might
represent four groups of people (e.g. children, young adults, middle-aged
and old people). In this case we are interested in comparing the proportions of people in each age group falling into the three response classes.
So we are comparing the 'shape' of response within each column between
the four columns - we are not interested in comparing the column means.
This suggests that a chi-squared contingency table analysis would be
appropriate.

Appendix B
References and Further reading

REFERENCES

Anscombe, F.J. (1973) Graphs in statistical analysis. American Statistician,


27,17-2l.
Clapham, A.R., Tutin, T.G., and Warburg, E.F. (1981) Excursion Flora
of the British Isles, Cambridge University Press, Cambridge.
Ehrenberg, A.S.C. (1977) Rudiments of numeracy. Journal of the Royal
Statistical Society, A, 140,277-297.
Hollis, M. (1985) Invitation to Philosophy, Blackwell, Oxford.
Minitab Inc. (1994) MINITAB Reference Manual, Release 10 for Windows,
Minitab Statistical Software, Pennsylvania, USA.
Minitab Inc. (1996) MINITAB Reference Manual, Release 11 for Windows,
Minitab Statistical Software, Pennsylvania, USA.
Rees, D.G. (1995) Essential Statistics, 3rd edn, Chapman & Hall, London.

FURTHER READING

There are vast numbers of articles and books written about statistics and
carrying out research at a wide range of levels. I have selected a few which
are modern and relatively easy to understand.
Elementary statistics

Campbell, R.C. (1989) Statistics for Biologists, 3rd edn, Cambridge


University Press, Cambridge. [A revision of a classic textbook which
covers all the topics in this book except PCA but at a more technical level,
including some MINITAB instructions (and ones for two other statistical
packages, Genstat and SPSS).]

216

I LI____________________A_P_P_E_N_D_I_X_B____________________~
Clegg, F. (1983) Simple Statistics: A Course Book for the Social Sciences,
Cambridge University Press, Cambridge. [An excellent, simple and
humorous book.]
Heath, D. (1995) An Introduction to Experimental Design and Statistics
for Biology, UCL Press, London. [Strong on experimental design; includes
Poisson and binomial distributions but not two-way Anova.]
Mead, R., Curnow, R.N., and Hasted, A.M. (1993) Statistical Methods
in Agriculture and Experimental Biology, 2nd edn, Chapman & Hall,
London. [A revision of a very popular textbook. It is more sophisticated
and covers more complex methods than the present book but is highly
recommended as a next step if you have found this book easy to follow.]
Neave, H.R., and Worthington, P.L. (1989) Distribution-Free Tests,
Unwin Hyman, London. [This covers non-parametric statistical methods
which are useful when your data do not meet the assumptions (like
Normality of residuals) required for parametric tests.]
Porkess, R. (1991) Dictionary of Statistics, 2nd edn, Collins, London.
[Provides definitions of technical terms.]
Rees, D.G. (1989) Essential Statistics, 3rd edn, Chapman & Hall, London.
[Very clear. Covers many of the same topics as this book but with a
slightly more mathematical approach; omits analysis of variance and PCA
but includes a final chapter on MINIT AB examples.]
Samuels, M.L. (1989) Statistics for the Life Sciences, Maxwell Macmillan,
San Francisco. [Very thorough. Includes basic analysis of variance and
linear regression but not PCA.]
Sokal, R.R., and Rohlf, F.J. (1981) Biometry: The Principles and Practice
of Statistics in Biological Research, W.H. Freeman, New York. [More
advanced; many research scientists would regard this as their standard
reference book and it is fine if you are really keen.]
Medical statistics

Anyone of these three texts would provide you with a straightforward


explanation of most of the subjects covered in this book (and a few
others), but in the context of medical research.
Campbell, M.J., and Machin, D. (1990) Medical Statistics Commonsense Approach, John Wiley, Chichester.

Mould, R.F. (1989) Introductory Medical Statistics, 2nd edn, Adam


.'
Hilger, Bristol.
Petrie, A. (1987) Lecture Notes on Medical Statistics, 2nd edn, Blackwell,
Oxford.

APPENDIXB

L -____________________________________________________

II

Exploring and interpreting data


Afifi, A.A., and Clark, V. (1996) Computer-Aided Multivariate Analysis,
3rd edn, Chapman & Hall, London. [More advanced; includes principal
components analysis.]
Anderson, A.J.B. (1989) Interpreting Data, Chapman & Hall, London. [A
useful next step; more advanced.]
Chatfield, C. (1995) Problem Solving - A Statistician's Guide, 2nd edn,
Chapman & Hall, London. [Very useful for developing a 'feel' for the
problems involved in analysing real-life data.]
Marsh, C. (1988) Exploring Data: An Introduction to Data Analysis for
Social Scientists, Blackwell, Oxford. [Excellent for complete beginners
with an arts or social science background, but concentrates on descriptive
statistics rather than testing hypotheses.]
Minitab
Minitab Inc. (1994, 1996) MINITAB Reference Manual, Releases 10 and
11 for Windows, Minitab Statistical Software, Pennsylvania, USA. [Useful
to browse through and see what MINITAB can do. Remember that you
can obtain some of this information by using the HELP command on the
computer.]
Rees, D.G. (1995) Essential Statistics, 3rd edn, Chapman & Hall, London.
[Final chapter contains numerous examples ofMINITAB analyses.]
Projects
Barnard, C., Gilbert, F., and McGregor, P. (1993) Asking Questions in
Biology: Design Analysis and Presentation in Practical Work, Longman,
Harlow. [Ideal for introductory practical courses in biology and as a guide
for assessed projects.]
Bell, J. (1987) Doing Your Research Project: A Guide for First-Time
Researchers in Education and Social Science, Open Univerity Press, Milton
Keynes. [Excellent practical advice - albeit with a social science bias.
Includes principles of questionnaire design.]
Chalmers, N., and Parker, P. (1986) The OU Project Guide, Field Studies
Council, Taunton. [Well worth consulting if you are carrying out an
ecological project in the field.]

Fowler, J., and Cohen, L. (1990) Practical Statistics for Field Biology,
Open University Press, Milton Keynes. [Very popular, especially with
ecologists; a slightly more mathematical approach than this book.]

217

218

I IL____________________A_P_P_E_N_D_I_X_B____________________~
Fry, J.C. (ed.) (1993) Biological Data Analysis: A Practical Approach,
IRL Press, Oxford. [An advanced text which is designed for biologists ideal for postgraduate research students.]
Pentz, M., Shott, M., and Aprahamian, F. (1988) Handling Experimental
Data, Open University Press, Milton Keynes. [A beginners' guide: very
useful.]
Report writing

Cooper, B.M. (1964) Writing Technical Reports, Penguin, Harmondsworth. [Particularly helpful on correctness and style.]
O'Connor, M. (1991) Writing Successfully in Science, Harper Collins,
London. [An excellent book which covers all aspects of communicating in
science, including presenting posters and talks as well as writing a paper.]
Wheatley, D. (1988) Report Writing, Penguin, London. [A good starting
point.]
Statistical tables

The following two books contain useful tables, including those reproduced
in this book.
CauIcutt, R. (1991) Statistics in Research and Development, 2nd edn,
Chapman & Hall, London.
Mead, R., and Curnow, R.N. (1983) Statistical Methods on Experimental
Biology, Chapman & Hall, London.
Other books

Do look at any reference books which your lecturers recommend. They


will, no doubt, have good reason for making their suggestions. It is very
helpful to read about the same method of approaching a problem in
several books as different authors have different approaches and it is
important to find one which makes the topic clear for you.

Appendix C

Statistical tables

220

I I

APPENDIXC

TABLE C.l

THE STANDARDIZED NORMAL DISTRIBUTION

The distribution tabulated is that of the Normal distribution with mean 0


and standard deviation 1. For each value of the standardized Normal
deviate Z, the proportion P of the distribution less than Z is given. For a
Normal distribution with mean J1. and variance (i, the proportion of the
distribution less than some particular value x is obtained by calculating
Z = (x - J1.)/u and reading the proportion corresponding to this value of z.
z

-4.00
-3.50
-3.0
-2.95
-2.90

0.00003
0.00023
0.0014
0.0016
0.0019

-1.50
-1.45
-1.40
-1.35
-1.30

0.0668
0.0735
0.0808
0.0885
0.0968

0.00
0.05
0.10
0.15
0.20

0.5000
0.5199
0.5398
0.5596
0.5793

1.55
1.60
1.65
1.70
1.75

0.9394
0.9452
0.9505
0.9554
0.9599

-2.85
-2.80
-2.75
-2.70
-2.65

0.0022
0.0026
0.0030
0.0035
0.0040

-1.25
-1.20
-1.15
-1.10
-1.05

0.1056
0.1151
0.1251
0.1357
0.1469

0.25
0.30
0.35
0.40
0.45

0.5987
0.6179
0.6368
0.6554
0.6736

1.80
1.85
1.90
1.95
2.00

0.9641
0.9678
0.9713
0.9744
0.9772

-2.60
-2.55
-2.50
-2.45
-2.40

0.0047
0.0054
0.0062
0.0071
0.0082

-1.00
-0.95
-0.90
-0.85
-0.80

0.1587
0.1711
0.1841
0.1977
0.2119

0.50
0.55
0.60
0.65
0.70

0.6915
0.7088
0.7257
0.7422
0.7580

2.05
2.10
2.15
2.20
2.25

0.9798
0.9821
0.9842
0.9861
0.9878

-2.35
-2.30
-2.25
-2.20
-2.15

0.0094
0.0107
0.0122
0.0139
0.0158

-0.75
-0.70
-0.65
-0.60
-0.55

0.2266
0.2420
0.2578
0.2743
0.2912

0.75
0.80
0.85
0.90
0.95

0.7734
0.7881
0.8023
0.8159
0.8289

2.30
2.35
2.40
2.45
2.50

0.9893
0.9906
0.9918
0.9929
0.9938

-2.10
-2.05
-2.00
-1.95
-1.90

0.0179
0.0202
0.0228
0.0256
0.0287

-0.50
-0.45
-0.40
-0.35
-0.30

0.3085
0.3264
0.3446
0.3632
0.3821

1.00
1.05
1.10
1.15
1.20

0.8413
0.8531
0.8643
0.8749
0.8849

2.55
2.60
2.65
2.70
2.75

0.9946
0.9953
0.9960
0.9965
0.9970

-1.85
-1.80
-1.75
-1.70
-1.65

0.0322
0.0359
0.0401
0.0446
0.0495

-0.25
-0.20
-0.15
-0.10
-0.05

0.4013
0.4207
0.4404
0.4602
0.4801

1.25
1.30
1.35
1.40
1.45

0.8944
0.9032
0.9115
0.9192
0.9265

2.80
2.85
2.90
2.95
3.00

0.9974
0.9978
0.9981
0.9984
0.9986

-1.60
-1.55

0.0548
0.0606

0.00

0.5000

1.50

0.9332

3.50
4.00

0.99977
0.99997

I I

APPENDIXC

THE STUDENT'S t-DISTRIBUTION

TABLE C.2

This table gives the value of t for which a particular percentage P of the
Student's t-distribution lies outside the range -t to +t. These values of t
are tabulated for various degrees of freedom.
Degrees of
freedom

50

20

lO

1
2
3
4
5

1.00
0.82
0.76
0.74
0.73

3.08
1.89
1.64
1.53
1.48

6.31
2.92
2.35
2.13
2.02

12.7
4.30
3.18
2.78
2.57

31.8
6.96
4.54
3.75
3.36

6
7
8
9
10

0.72
0.71
0.71
0.70
0.70

1.44
1.42
1.40
1.38
1.37

1.94
1.89
1.86
1.83
1.81

2.45
2.36
2.31
2.26
2.23

12
15
20
24
30

0.70
0.69
0.69
0.68
0.68

1.36
1.34
1.32
1.32
1.31

1.78
1.75
1.72
1. 71
1.70

40
60

0.68
0.68
0.67

1.30
1.30
1.28

1.68
1.67
1.64

00

0.2

0.1

63.7
9.92
5.84
4.60
4.03

318
22.3
10.2
7.17
5.89

637
31.6
12.9
8.61
6.87

3.14
3.00
2.90
2.82
2.76

3.71
3.50
3.36
3.25
3.17

5.21
4.79
4.50
4.30
4.14

5.96
5.41
5.04
4.78
4.59

2.18
2.13
2.09
2.06
2.04

2.68
2.60
2.53
2.49
2.46

3.05
2.95
2.85
2.80
2.75

3.93
3.73
3.55
3.47
3.39

4.32
4.07
3.85
3.75
3.65

2.02
2.00
1.96

2.42
2.39
2.33

2.70
2.66
2.58

3.31
3.32
3.09

3.55
3.46
3.29

221

THE F-DISTRIBUTION: 5% POINTS

4.76
4.35
4.07
3.86
3.71

5.14
4.74
4.46
4.26
4.10

3.89
3.68
3.49
3.40
3.32

3.23
3.15

5.99
5.59
5.32
5.12
4.96

4.75
4.54
4.35
4.26
4.17

4.08
4.00

6
7
8
9
10

12
15
20
24
30

40
60

2.84
2.76

3.49
3.29
3.10
3.01
2.92

19.2
9.28
6.59
5.41

19.0
9.55
6.94
5.79

18.5
10.1
7.71
6.61

2
3
4
5

n2

2.45
2.37

2.34
2.25

3.00
2.79
2.60
2.51
2.42

3.11
2.90
2.71
2.62
2.53

3.26
3.06
2.87
2.78
2.69
2.61
2.53

4.28
3.87
3.58
3.37
3.22

19.3
8.94
6.16
4.95

4.39
3.97
3.69
3.48
3.33

19.3
9.01
6.26
5.05

4.53
4.12
3.84
3.63
3.48

19.2
9.12
6.39
5.19

nl

2.25
2.17

2.91
2.71
2.51
2.42
2.33

4.21
3.79
3.50
3.29
3.14

19.4
8.89
6.09
4.88

2.18
2.10

2.85
2.64
2.45
2.36
2.27

4.15
3.73
3.44
3.23
3.07

19.4
8.85
6.04
4.82

2.08
1.99

2.75
2.54
2.35
2.25
2.16

4.06
3.64
3.35
3.14
2.98

19.4
8.79
5.96
4.74

10

These tables give the values of F for which a given percentage of the F-distribution is greater than F.

TABLE C.3

2.00
1.92

2.69
2.48
2.28
2.18
2.09

4.00
3.57
3.28
3.07
2.91

19.4
8.74
5.91
4.68

12

1.79
1.70

2.51
2.29
2.08
1.98
1.89

3.84
3.41
3.12
2.90
2.74

19.5
8.64
5.77
4.53

24

(")

t:I

tTl

"tI

>
"tI

98.5
34.1
21.2
16.3

13.7
12.3
11.3
10.6
10.0

9.33
8.68
8.10
7.82
7.56

7.31
7.08

6
7
8
9
10

12
15
20
24
30

40
60

5.18
4.98

6.93
6.36
5.85
5.61
5.39

10.98
9.55
8.65
8.02
7.56

99.0
30.8
18.0
13.3

4.31
4.13

5.95
5.42
4.94
4.72
4.51

9.78
8.45
7.59
6.99
6.55

99.2
29.5
16.7
12.1

3.83
3.65

5.41
4.89
4.43
4.22
4.02

9.15
7.85
7.01
6.42
5.99

99.2
28.7
16.0
11.4

3.51
3.34

5.06
4.56
4.10
3.90
3.70

8.75
7.46
6.63
6.06
5.64

99.3
28.2
15.5
11.0

THE F-DISTRIBUTION: 1% POINTS

2
3
4
5

n2

TABLE C.4

3.29
3.12

4.82
4.32
3.87
3.67
3.47

8.47
7.19
6.37
5.80
5.39

99.3
27.9
15.2
10.7

nl

3.12
2.95

4.64
4.14
3.70
3.50
3.30

8.26
6.99
6.18
5.61
5.20

99.4
27.7
15.0
10.5

2.80
2.63

4.30
3.80
3.37
3.17
2.98

4.50
4.00
3.56
3.36
3.17
2.99
2.82

7.87
6.62
5.81
5.26
4.85

99.4
27.2
14.5
10.1

10

8.10
6.84
6.03
5.47
5.06

99.4
27.5
14.8
10.3

2.66
2.50

2.29
2.12

3.78
3.29
2.86
2.66
2.47

7.31
6.07
5.28
4.73
4.33

7.72
6.47
5.67
5.11
4.71
4.16
3.67
3.23
3.03
2.84

99.5
26.6
13.9
9.47

24

99.4
27.1
14.4
9.89

12

tTl

t:l
><!

)'"t:J
'"t:J

27.0
21.7
18.5
16.4
14.9

13.0
11.3
9.95
9.34
8.77

8.25
7.77

35.5
29.3
25.4
22.9
21.0

18.6
16.6
14.8
14.0
13.3

12.6
12.0

6
7
8
9
10

12
15
20
24
30

40
60

6.59
6.17

10.8
9.34
8.10
7.55
7.05

23.7
18.8
15.8
13.9
12.6

999
141
56.2
33.2

999
149
61.3
37.1

999
167
74.1
47.2

5.70
5.31

9.63
8.25
7.10
6.59
6.12

21.9
17.2
14.4
12.6
11.3

999
137
53.4
31.1

8.38
7.09
6.02
5.55
5.12
4.73
4.37

5.13
4.76

20.0
15.5
12.9
11.1
9.93

20.8
16.2
13.5
11.7
10.5
8.89
7.57
6.46
5.98
5.53

999
133
50.5
28.8

nl

999
135
51.7
29.8

THE F-DISTRIBUTION: 0.1% POINTS

2
3
4
5

n2

TABLE C.5

4.44
4.09

8.00
6.74
5.09
5.23
4.82

19.5
15.0
12.4
10.7
9.52

999
132
49.7
28.2

4.21
3.86

7.71
6.47
5.44
4.99
4.58

19.0
14.6
12.1
10.4
9.20

999
131
49.0
27.7

3.87
3.54

7.29
6.08
5.08
4.64
4.24

18.4
14.1
11.5
9.87
8.74

999
129
48.1
26.9

10

3.64
3.32

7.00
5.81
4.82
4.39
4.00

18.0
13.7
11.2
9.57
8.44

999
128
47.4
26.4

12

3.01
2.69

6.25
5.10
4.15
3.74
3.36

16.9
12.7
10.3
8.72
7.64

1000
126
45.8
25.1

24

\.)

><:

t:I

tTl

"0
"0

;J>

I I

APPENDIXC

TABLE C.6

THE CHI-SQUARED DISTRIBUTION

This table gives the values of X2 for which a particular percentage P of


the chi-squared distribution is greater than l. These values of l are
tabulated for various degrees of freedom.
Degrees of
freedom

50
1
2
3
4
5

0.45
1.39
2.37
3.36
4.35

6
7
8
9
10

5.35
6.35
7.34
8.34
9.34

10

2.5

3.84
5.99
7.82
9.49
11.1

5.02
7.38
9.35
11.1
12.8

6.64
9.21
11.3
13.3
15.1

10.8
13.8
16.3
18.5
20.5

10.6
12.0
13.4
14.7
16.0

12.6
14.1
15.5
16.9
18.3

14.5
16.0
17.5
19.0
20.5

16.8
18.5
20.1
21.7
23.2

22.5
24.3
26.1
27.9
29.6

2.71
4.61
6.25
7.78
9.24

0.1

12
15
20
24
30

11.3
14.3
19.3
23.3
29.3

18.5
22.3
28.4
33.2
40.3

21.0
25.0
31.4
36.4
43.8

23.3
27.5
34.2
39.4
47.0

26.2
30.6
37.6
43.0
50.9

32.9
37.7
45.3
51.2
59.7

40
60

39.3
59.3

51.8
74.4

55.8
79.1

59.3
83.3

63.7
88.4

73.4
99.6

225

6077
1796
25 16
8451
3428

1982
77 51
0096
8884
5972
5750
1680
5432
2897
3231

6444
8838
5880
82 15
71 00

71 85
9666
1837
6434
0647

8088
3355
6600
7925
63 18

7370
33 18
13 33
6754
5385

9383
4202
4242
6671
7305

5076
3963
1948
3920
5866

8634
4487
4898
2139
14 19

6917
9721
2355
5809
77 21

11 32
9973
7488
81 23
5022

1971
1954
11 75
0083
7822

3006
91 76
5684
2996
5891

9929
2561
21 88
1748
6334

8332
2717
4089
51 66
9522

2329
8761
41 10
2764
4389

4343
3501
9982
2873
3061

7450
5466
6528
72 96
1922

8608
44 46
3297
3826
4020

6655
1569
64 78
1793
6156

9737
0340
5289
72 76
6050

1843
61 71
2993
1770
61 13

64 17
91 14
2771
1296
5430

2802
0613
9753
31 89
11 23

5014
7469
5951
1244
2994

5451
0284
2027
21 16
1975

3623
9570
6950
7611
9560

5054
2276
0755
53 11
7759

2371
5290
9049
7574
7056

5396
2258
2988
2252
5420

Random numbers

10 27
8590
4433
4757
0320

Table C.7

3895
4976
1537
6290
7083

9333
9928
3412
1771
1290

7654
7430
9262
3075
0351

1562
3680
41 91
72 81
8239

1624
2376
9752
78 16
6009
3741
2422
7981
8346
4925

8887
0650
2226
4775
1447

1229
9886
5704
1691
5939

(")

><:

ti
.....

;I>
'"C
'"C

1224
8132
6484
0507
2194

8736
2677
0912
2007
5308

7454
2689
4327
2408
9737

2851
5308
9945
7496
1876

0280
2289
9445
10 30
7339

4791
3924
7729
0478
8381

9767
5280
8069
0048
1491

5062
1759
7317
3795
7695

9410
5806
8528
2544
7628

9614
13 38
3356
7392
5340

4569
6237
4177
17 15
0923

3478
3272
1728
2134
0184

15 18
8000
6317
9566
1860

6328
7008
3988
3719
4626

0301
8246
2957
6812
2261

2250
4892
0657
5918
2820

0602
7571
9931
4202
44 92

9811
7322
7331
6987
2925

2425
9396
3489
3826
9941

5702
9575
7196
8595
5035

3994
9513
2462
3148
7609

1833
6470
2444
9179
9642

1364
8275
9495
0190
2790

0701
8856
81 36
21 87
5782

13 91
7691
7582
8221
4696

8260
8344
8733
8627
5722

4274
7516
4570
6830
3543

13 00
7553
3765
7316
8813

5450
2455
7889
7687
3937

9041
4924
0821
4791
9434

3667
9505
5985
8380
0709

7880
7917
4262
7837
5253

6027
3409
2759
8675
2712

33 11
2093
4006
3170
5971

77 67
3068
3804
1989
6226

9493
5381
4384
1598
7673

2668
9712
1862
0795
3044

77 59
1259
77 91
5352
2359

0092
8302
0480
9865
4583

1453
5417
4523
1666
6822

tT1

(j

><

Z
0
.....

)'"C
'"C

228

I I

APPENDIXC

TABLE C.S CRITICAL VALUES OF THE PRODUCT MOMENT


CORRELATION
Degrees of
freedom

Two-sided test

One-sided test

5%
(0.05)

1%
(0.01)

5%
(0.05)

1%
(0.01)

2
3
4
5

0.950
0.878
0.811
0.754

0.990
0.959
0.917
0.875

0.900
0.805
0.729
0.669

0.980
0.934
0.882
0.833

6
7
8
9
10

0.707
0.666
0.632
0.602
0.576

0.834
0.798
0.765
0.735
0.708

0.621
0.582
0.549
0.521
0.497

0.789
0.750
0.715
0.685
0.658

11
12
13
14
15

0.553
0.532
0.514
0.497
0.482

0.684
0.661
0.641
0.623
0.606

0.476
0.457
0.441
0.426
0.412

0.634
0.612
0.592
0.574
0.558

20
30
40
60

0.423
0.349
0.304
0.250

0.537
0.449
0.393
0.325

0.360
0.296
0.257
0.211

0.492
0.409
0.358
0.295

Index

Abscissa 121
Absolute values 8
Abstracts, reports 206
Accidents 110-16, 200
Acknowledgements section, reports
206
Active v. passive voice 209
Adjusted sums-of-squares 115
Aims and experimental design 58-9,
188, 190-1
Alternative hypothesis 28
analysis of variance 85, 86
Analysis of results 79
analysis-of-variance table 84--7
choice of method 213-14
different treatments 89
exercise 94--7
experimental design 195
MINITAB 89-94
questionnaires 198-200
randomized complete block designs
87-9
wholly randomized design 80-3
Analysis of variance (Anova) 78, 213,
214
exercise 94--7
factorial structure of treatments 103,
104, 105-7, 108, 118
linear regression 128
MINITAB 89-94, 105-7, 108, 118
non-parametric equivalents 153
one-way 80-3
independent I-test 31
MINITAB 90-1,112-13
non-parametric equivalent 153,
156
plan 84, 87
table 84--7, 88

linear regression 128


two-way 87-9
MINITAB 91-2, 93-4,113-16
non-parametric equivalent 153
Animal research 200
Appendixes, reports 206
Backup disks 193
Best-fit line 122, 124
Bias
linear regression 131
and randomization of experiments
57
recording data 61
sampling
survey design 195
systematic 51
Bibliographies 194
Binomial distribution 5, 6
Blind trials 123
Block effect 74, 75
Blocking 71-8
analysis of variance 87-9
MINITAB 91-2
experimental design 190
Friedman test 159-61,168,169
Block mean 74, 75
Block sum-of-squares 88
Boxplots (box-and-whisker plots) 26
experimental design 194
MINITAB 26, 27,38,39
analysis of variance 92-4
factorial analysis 110
missing observations 114
observational studies 174, 175
planning an experiment 63, 66
non-parametric tests 152
Buffer zones 59-60

230

II

INDEX

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Calculators 21-2
block sum-of-squares 88
random numbers 44
random sampling 45
stratified 47
total sum-of-squares 79
treatment sum-of-squares 83
t-test exercise 40
Carbon copies of results 193
Categorical data 153, 161-6,213
Chi-squared 163, 164, 166, 167, 171
Chi-squared contingency test 153,
161-6,213,214
exercise 169-71
questionnaire data 198-200
Chi-squared distribution 225
Chi-squared goodness-of-fit test 166-7
Circular sampling plots 52-3
Clinical trials
ethical issues 200
linear regression 123
Closed questions 196
Clusters, irregular stratified random
sampling 51
Coefficient of determination
correlation coefficients 144
linear regression 128, 150
Computers and computer packages
analysis of variance 87
data entry errors 194
familiarity with 195
filing system 194-5
independent t-test 33
recording data 62, 193-4
reports 202-3, 207
vegetation data analysis 185
see also MINITAB
Confidence intervals
exercises 36-7, 116, 117
graphs 207
importance 18-22
linear regression 142-3
MINITAB 35
treatment means 98-9
difference between two 99-101
t-test 40-1
Constraints, and experimental design
192
Continuous distributions 5, 213
linear regression 131
Control of Substances Hazardous to
Health 201

Controls 58
experimental design 189-90
Correlation 143-7
exercise 148-9
observational studies 174
Correlation coefficients 143-7
observational studies 174
COSHH201
Covariates 191
Critical values, analysis of variance 86,
88-9
Data-loggers 193
Degrees of freedom 11
analysis of variance 84, 88
factorial structure of treatments
104,105
F-table 86
chi-squared contingency test 163,
164,170
Kruskal-Wallis test 157
stratified random sampling 48-9
Dependent variables 121
Design, experimental 55, 188-95
controls 58
exercise 64-6
laying out the experiment 59-61
MINITAB 62-4
objectives 58-9
randomization 56-8
recording data 61-2
replication 55-6
De-trended correspondence analysis
(DCA) 185
Disasters 110-16
Discontinuous distributions 5-6
Discussion section, reports 204
Disks 193-4
Distribution-free (non-parametric) tests
152-3,213
Dotplots 22, 38
Draft reports 206-7
Dust extractors 201
Ehrenberg, A.S.C. 208
Eigenvalues 176, 181
Errors
experimental design 193, 194-5
linear regression 129, 131, 139
non-response 196
response 196-8
see also Residuals

---'I I

'----_ _ _ _ _ _ _ _I_N_D_EX
________
Ethical issues 200
Expected values 69, 70, 71, 75-6
analysis of variance 93-4
chi-squared contingency test 162,
163,164-5,170
chi-squared goodness-of-fit test 166
exercise 77-8
Experimental units 55, 189
Experiments
design and execution 188-95
linear regression 141
planning 55
controls 58
exercise 64-6
laying out the experiment 59-61
MINITAB 62-4
objectives 58-9
randomization 56--8
recording data 61-2
replication 55--6
Extrapolation 141-2
Factorial analysis 101-5,213
exercises 117-19
experimental design 190
MINIT AB 105-9
tables 208
Factors 101
F-distribution 222-4
analysis of variance 86, 87
Fieldwork safety 200-1
Fisher, R.A. 79
Fit see Expected values
Fitted values see Expected values
Friedman test 153, 159-60,213,214
exercise 168, 169
Fume cupboards 201
Galton, Sir Francis 122-3
Generality of experiments 189
General linear model (glm) 114-15,
120
Generic names 209
Goodness of fit, chi-squared test 166--7
Grand mean 69, 70, 74, 75
exercise 77
Graphs 207-8
H 157, 158
Harper, John 188
Hazardous substances 201
Health and safety 200--1

High significance 33
Histograms
exercise 16
MINITAB 23, 27, 38, 39
factorial analysis 109, 111
missing observations 116
Homogeneity of variance 130-1
Hypothesis testing 27-9
analysis of variance 85
see also t-test
Illustrating results 207-9
Independent factors 103, 104
Independent t-test 29-31
calculation 31-4
MINITAB 34-5
Independent variables 121
linear regression 131
Induction 188
Inference 188
Integration 6-7
Interaction diagrams 109, Ill, 118,
119
Interaction of factors 103-4
MINITAB 106, 109
Introduction to report 203
Irregular stratified random sampling 51
Kruskal-Wallis test 153, 155-9,213,
214
exercise 168, 169
Laboratory health and safety 201
Language of reports 209
Laying out the experiment 59-61
Least significant difference (LSD) 100,
101
exercises 116, 117
graphs 207
Least-squares fit 125
Leptospirosis 200
Levels 10 1, 102
Linear regression 122-3, 213
assumptions 129-31
confidence intervals 142-3
correlation 143-7
example 131-4
exercise 147-51
model 123-9
plotting observations, importance
134-42
Literature searches 203

231

232

I I~________________IN_D_E_X________________~
Lower quartile 26
Lyme disease 200
Main effects 102-4, 106-7
Mann-Whitney test 153-5,213
exercise 167, 168-9
Materials and methods, report section
203-4
Maximum values 24
Mean 2-3
block 74,75
confidence intervals 36
exercise 17
grand 69, 70, 74, 75, 77
limitations 24--5
MINITAB 24, 34--5
Normal distribution 6, 7
sampling distribution 13-14
standard error see Standard error
statistical mode, calculators 22
treatment 69
analysis of variance 83, 85
confidence intervals 98-101
trimmed 24
Mean square 84--5
factorial structure of treatments
104
see also Variance
Measurements 2-3
Median 25
Mann-Whitney test 154--5
MINITAB 25-6
Minimum values 24
MINITAB 22-6,37-9
analysis of variance 89-94, 95-7,
118
missing observations 112-16
chi-squared contingency test 163-4,
165-6, 170-1
questionnaire analysis 198-9
chi-squared goodness-of-fit test
166-7
correlation 144, 145-6, 148
data-loggers 193
factorial analysis 105-9, 117-19
Friedman test 159-60,161, 168,
169
Kruskal-Wallis test 156, 157, 158-9,
168, 169
linear regression 126-30, 132-4
confidence intervals 143
exercise 148-51

importance of plotting
observations 135-41
Mann-Whitney test 154--5,167,
168-9
planning an experiment 62-4, 65
principal components analysis
case study 1: 173-4, 175, 176-7,
178,179
case study 2: 180-3, 184
t-test 33, 34--5, 40-1
word processing packages 202
Missing observations 109-16
exercise 120
experimental design 194
Mistakes 11 0-16
Models 68-71
blocking 71-6
exercise 77-8
Multiple regression 125
Multivariate analysis 180
Negatively skewed distributions 25
Negative relationship 121, 122
Non-parametric tests 152-3,213
Non-response errors 196
Non-significance 33
Normal distribution 3-7, 220
checking for 26
confidence intervals 18
factorial analysis 109
linear regression 129-30
Notebooks 202
Notes 194
Null hypothesis 28,89
analysis of variance 85, 86, 87
t-test 29-30, 33, 35
Objectives, and experimental design
58-9, 188, 190-1
Observational studies
interpretation 188
summarizing data from 172-3
case study 1: 173-9
case study 2: 179-85
specialized programs for analysing
vegetation data 185
Observations 10
number of 45, 60
Observed values 70, 76
One-sample t-tests 35
One-tailed t-tests 34
One-way analysis of variance 80-3

-----'I I

IN_D_EX
________

L - -_ _ _ _ _ _ _ _ _

independent t-test 31
MINITAB 90-1
missing observations 112-13
non-parametric equivalent 153, 156
Open questions 196
Ordinate axis 121
Outliers
boxplots 26, 194
correlation 146, 147
linear regression 138
non-parametric tests 152, 153
Paired t-tests 35
Parameters 3
Parametric tests 152
Passive v. active voice 209
Past v. present tense 209
Pearson's r (Product moment
correlation coefficient) 145, 146, 147
critical values 228
exercise 148
Photocopies of results 193
Photographs 194
Pilot studies
experimental design 191
random sampling 45
survey design 198
Pipe symbol 105-6
Plagiarism 205
Planning an experiment 55
controls 58
exercise 64-6
laying out the experiment 59-61
MINITAB 62-4
objectives 58-9
randomization 56-8
recording data 61-2
replication 55-6
Plots 55
Plural v. singular 209
Poisson distribution 5-6
Pooled variance 32, 40
Population
experimental design 188-9
mean 3, 34-5
confidence intervals 36
sampling 2
Positively skewed distributions 5, 25
Positive relationship 121, 122
Precision of estimates
experimental design 188
and replication 56

stratified random sampling 49


Present v. past tense 209
Principal components analysis (peA)
172-3
case study 1: 176-9
case study 2: 180-5
specialized programs for analysing
vegetation data 185
Probability 6, 7
Procedural controls 58
Product moment correlation coefficient
(Pearson's r) 145, 146, 147
critical values 228
exercise 148
Projects
advice-seeking at start, importance
187-8
choice of topic and supervisor 186-7
experimental design and execution
188-95
health and safety 200-1
survey design and execution 195-200
Protective clothing and equipment 201
QI26
Q326
Quadrats 52-3
Questionnaires 196-200
Randomized complete block designs
71,72
analysis of variance 87-9
MINITAB 91-2
Randomized experiments 56-8
analysis of variance 80-3
design 190
Random numbers 226-7
blocking 72
random sampling 44, 54
Random sampling 42-5
chi-squared contingency test 166
stratified 45-9, 54
blocking compared to 73-4
irregular 51
and systematic sampling 51
Random variation 68
analysis of variance 85
blocking 74
linear regression 123, 128
residuals 69, 71
Rat-borne diseases 200
Realism, experimental design 189

233

234

II

INDEX

L -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

Recording data 61-2


experimental design 192-4
Record sheets 62, 192
References section 194, 204-6
Reliability of sample estimate 3
Repeated measures 191, 194
Replication 3, 55-6
analysis of variance 84
experimental design 191-2
factorial structure of treatments
104
Reports 202
basics 203-7
computers 202-3
exercise 209-12
illustrating results 207-9
language 209
Representative samples 42
random sampling 43, 45
Residuals 69-71,76
analysis of variance 96-7
one-way 80-2, 83
MINITAB 93-4
table 84, 85
two-way 88, 93-4
exercise 77, 78
factorial analysis 104
MINITAB 109, Ill, 119
linear regression 124-5, 129-31,
133
exercise 150-1
standardized 136-8,139,141
Resources, and experimental design
192
Response errors 196-8
Response variables 121
Results
analysis of 79
analysis-of-variance table 84-7
choice of method 213-14
different treatments 89
exercise 94-7
experimental design 195
MINITAB 89-94
questionnaires 198-200
randomized complete block
designs 87-9
wholly randomized design 80-3
reports 204
illustrations 207-9
Reviews, report 207
Rounding of data 208

S 161
Safety 200-1
Sample mean 34-5
confidence intervals 36
Sampling 2-3, 42
bias 195-6
confidence intervals 19-20
distribution 13-14
exercises 53-4
practical problems 52-3
random 42-5
chi-squared contingency test 166
irregular stratified 51
stratified 45-9
systematic 49-51
two-stage 51-2
Scatter plots 63
correlation 144, 145, 146
linear regression 135, 136, 137-8,
139-40, 141
Sequential sums-of-squares 115
Sheep-borne diseases 200
'Shotgun effect' 130
Significance 33, 87
Significant difference 100
Significant figures 208
Singular v. plural 209
Skewed distributions 5, 25, 152, 153
Mann-Whitney test 154
Spearman's r 145-7,148,153
Square sampling plots 52
Standard deviation 11-12
exercise 17
experimental design 192
linear regression 128
MINITAB24
Normal distribution 6, 7
statistical mode, calculators 21, 22
Standard error 14-15
confidence intervals 18,99
exercise 17
flow chart 20
graphs 207
MINITAB24
missing observations 113
random sampling 45
stratified 47,49
statistical mode, calculators 22
Standard error of difference (SED)
confidence intervals 100
independent t-test 31, 32, 33
Standardized data 174, 177

INDEX

L -____________________________________________________

Standardized residuals 136-8, 139, 141


Statistics 3
Stem-and-leaf plot 23--4, 25, 38
Straight line, equation 125-6
Strata 45
Stratified random sampling 45-9,54
blocking compared to 73--4
irregular 51
and systematic sampling 51
Structure of reports 203-6
Student's't-distribution 18-19,221
Subsampling 53
Summarizing data from observational
studies 172-3
case study 1: 173-9
case study 2: 179-85
specialized programs for analysing
vegetation data 185
Sum of differences 8-9
Sum of squared differences (sum-ofsquares) 9-10
analysis of variance 79,80-3,88
factorial structure of treatments
103--4
linear regression 123, 124, 125
missing observations 115
t-test 32, 40
Supervisors 186-8
Surveys
design and execution 195-200
linear regression 141
Systematic sampling 49-51,54
Tables 208-9
Tape recorders 193
(distribution 18-19,221
Tense 209
Tetanus inoculations 200
Topic, choice of 186-7
Total sum-of-squares 79, 82-3
Transcription errors 194
Transects 52, 53
Transformation 152, 213
t-ratio 29
calculation 31, 33
exercise 40
Treatment effect 70
Treatment mean 69
analysis of variance 83, 85
confidence intervals 98-9
difference between two treatment
means 99-10 1

II

Treatments 55
experimental design 189-90
factorial structure 101-5
Treatment sum-of-squares
analysis of variance
one-way 80, 81-2, 83
two-way 88
factorial structure of treatments
103-4
Trimmed mean 24
t-test 29
exercises 36, 39--41
independent 29-34
MINITAB 34-5
stratified random sampling 48
Two-sample (independent) (-test 29-31
calculation 31--4
MINITAB 34--5
Two-stage sampling 51-2
Two-tailed t-tests 33--4
Two-way analysis of variance 87-9
MINITAB 91-2, 93--4
missing observations 113-16
non-parametric equivalent 153
Unbalanced analysis of variance 113,
114,115
Unrepresentative samples 42
random sampling 43, 45
Upper quartile 26
Validation of data 198
Variance 10-11
analysis of see Analysis of variance
homogeneity 130-1
independent t-test 31-2
pooled 32, 40
ratio 85, 86-7, 88-9
factorial structure of treatments
102
MINITAB 90,91
stratified random sampling 47,49
Variates 191
Variation 3
expressing 8-15
Normal distribution 6, 7
sources 67-9
blocking 71-6
exercise 77-8
model 69-71
Very high significance 33
VESPAN III 185

235

236

II

INDEX

~----------------------------------------------------~

W155
Wilcoxon rank sum test see MannWhitney test
Word 202
WordPerfect 202

Word processing packages 202-3


Writers' block 203
Zero, observations of 109-10
z value 157