Biostatistics

Brian Williams
London School of Hygiene and Tropical Medicine, UK

CHAPMAN & HALL

London· Glasgow· New York· Tokyo . Melbourne . Madras
Published by Chapman & Hall, 2-6 Boundary Row, London SEt 8HN, UK

Chapman & Hall, 2-6 Boundary Row, London SE1 8HN, UK

Blackie Academic & Professional, Wester Cleddens Road, Bishopbriggs,
Glasgow G64 2NZ, UK
Chapman & Hall GmbH, Pappe1allee 3, 69469 Weinheim, Germany
Chapman & Hall USA, One Penn Plaza, 41st Floor, New York, NYI0119,
USA
Chapman & Hall Japan, ITP - Japan, Kyowa Building, 3F, 2-2-1 Hirakawacho,
Chiyoda-ku, Tokyo 102, Japan
Chapman & Hall Australia, Thomas Nelson Australia, 102 Dodds Street, South
Melbourne, Victoria 3205, Australia
Chapman & Hall India, R. Seshadri, 32 Second Main Road, crr East, Madras
600035, India

First edition 1993

Reprinted 1994

© 1993 Brian Williams

Typeset in 10/12pt Palatino by Thomson Press (India) Ltd, New Delhi

Printed in Great Britain by St Edmundsbury Press, Bury St Edmunds, Suffolk

ISBN 0 412 46220 6

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 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 Data

Williams, Brian, 1945-
Biostatistics I Brian Williams.-lst English language ed.
p. em.
Includes bibliographical references (p. ) and index.
ISBN 0-412-46220-6 (acid-free paper)
1. Biometry. I. Title.
QH323.5.W55 1993 93-3336
574'.01'5195-dc20 CIP
t§ Printed on acid-free text paper, manufactured in accordance with
ANSIINISO Z39.48-1992 and ANSIINISO Z39.48-1984 (permanent Paper).
Contents
~
~

Preface 1~ vii
ACknowledgements xi

1 Why athematiCS 1
1.1 So e problems 5
1.2 A rief outline 8
1.3 Ex
M
~
rcises 10

2 Proba ility 12
2.1 Th concept of probability 12
2.2 Th laws of probability 14
2.3 Su mary 23
2.4 Ex rcises 23

3 Repres ntations 25
3.1 Di crete random variables 26
3.2 Co tinuous random variables 29
3.3 Su mary 36
3.4 Ex rcises 37
~
4 Meas es 39
4.1 M asures of location 39
4.2 M asures of scale 42
4.3 M asures of shape 45
4.4 Su mary 49
4.5 Ap endix 49
4.6 Ex rcises 51

5 Distri ution 52
5.1 Th binomial distribution 53
5.2 Th Poisson distribution 60
5.3 Th normal distribution 63
5.4 Th X2 t and F distributions 67
5.5 Cu ulative distribution functions 71
5.6 Ta les of distribution functions 73
5.7 Su mary 74
5.8 Ex rcises 75
,
~
6 Testin hypotheses 77

g
6.1 Si nificance levels and the power of a test
~ Thf~
78
~
vi Contents
63 Contingency tables 87
64 Summary 90
65 Appendix 91
66 Exercises 92

7 omparisons 94
7 lOne-and two-tailed tests 94
7 2 Comparison of means 96
73 Non-parametric tests 101
7 4 Multiple range tests 103
75 Variance ratios 107
7 6 Confidence limits 107
77 Several factors 109
7 8 Experimental design 113
7 9 Summary 117
710 Appendix 119
7 II Exercises 12 I
~
8 AnalySis of variance 123
811 One-way ANOVA 124
8 2 Unequal sample sizes 13 I
83 Model I and Model II ANOV A 133
84 Two-way ANOV A 135
8 5 Summary 150
8.6 Exercises 15 I
~
~
9Regression 154
91.1 Defining the fit 157
9.2 ANOV A 159
9.3 Error estimates and confidence limits 160
9.4
Residuals 163
9.5 Correlation 166
9.6 Polynomial regression 169
9.7 Multiple regression 171
9.8 Summary 174
9.9 Appendix 176
9.10 Exercises 177

10 Tables 183

References 187
Index 193
~
 ~

Pref Ice

This book iS an introduction to statistics for students of the life sciences, particularly
for field a d laboratory biologists, but also for those studying agriculture and medicine.
Students ommonly face the statistical analysis of their data with anxiety, and yet
statistics c n be elegant and exciting. Statistical analysis can reveal subtle patterns in
otherwise uninformative data, reduce large data sets to a few key parameters, give
researcher confidence in their conclusions and ensure that they do not deceive
themselve when interpreting their data.
This bo k stresses the concepts and ideas that underlie the most important statistical
methods u ed in biology. I have placed these concepts and ideas in a narrative structure
and consi er the same problems from different angles and in progressively more detail.
Once you understand the general principles and are able to apply them to specific
problems, ou should be able to analyse new problems and to discover new applica-
tions for s atistics in your own work. To ensure that the text remains firmly rooted
in real pro lems, I have taken examples and illustrations from the biological literature
wherever os sible.
When using statistics to help answer a biological question, it is important to under-
stand the question clearly. Once you know what it is that you wish to gain from
your data, statistical techniques become powerful tools for revealing information that
is otherwi e hidden. Always remember that the statistics is simply a way to help you
think clea ly about the biology.

FURTHE READING

Many bo ks on biostatistics are available; some can be used to complement this

book. I ha e drawn on the following sources.

Bailey N .. (1974) Statistical Methods in Biology, English Universities Press, London.

Glantz, S. A. (1987) Primer of Biostatistics, McGraw-Hill, New York.
Parker R. . (1986) Introductory Statistics for Biologists, Edward Arnold, London.
Phillips, J. L. Jr. (1982) Statistical Thinking, Freeman, New York.
Porkess, R' (1988) Dictionary of Statistics, Collins, London.
viii Preface

Blis, C. I. (1967) Statistics in Biology-Volume I, McGraw-Hill, New York.

Bul er, M. G. (1979) Principles of Statistics, Dover, New York.
Ca pbell, R. C. (1987) Statistics for Biologists, Cambridge University Press, Cambridge.
Flei s, J. L. (1986) The Design and Analysis of Clinical Experiments, John Wiley, New York.
Fre nd, J. E. (1972) Mathematical Statistics, Prentice-Hall, London.
Hays' W. L. (1988) Statistics, Holt, Rinehart and Winston, Orlando.
JeffS' J. N. R. (1978) An Introduction to Systems Analysis: With Ecological Applications,
E ward Arnold, London.
Ma Iy, B. (1986) Multivariate Statistical Methods: A Primer, Chapman & Hall, London.
Mead, R. (1988) The Design of Experiments: Statistical Principlesfor Practical Applications,
C mbridge University Press, Cambridge.
Me d, R. and Curnow, R. N. (1983) Statistical Methods in Agriculture and Experimental
B ology, Chapman & Hall, London.
Sieg I, S. and Castellan Jr. N. J. (1988) Nonparametric Statistics for the Behavioral Sciences,
cGraw-Hill, New York.
Sok I, R. R. and Rohlf F. J. (1981) Biometry, Freeman, New York.
Sok I, R. R. and Rohlf F.J. (1987) Introduction to Biostatistics, Freeman, New York.
Spre t P. (1990) Applied Nonparametric Statistical Methods, Chapman & Hall, London.
Zar . H. (1984) Biostatistical Analysis, Prentice-Hall, Englewood Cliffs.

Gre n P. E. (1978) Analyzing Multivariate Data, Dryden Press, Hinsdale, Illinois.

Mas n, R. L. Gunst, R. F. and Hess, J. L. (1989) Statistical Design and Analysis of
E periments: With Applications to Engineering and Science, John Wiley, New York.
Seb r, G. A. F. (1989) Linear Regression Analysis, John Wiley, New York.
Sne ecor, G. W. and Cochran, W. G. (1980) Statistical Methods, Iowa State University
P ess, Ames, Iowa.
Win r B. J. (1971) Statistical Principles of Experimental Design, McGraw-Hill, New York.

ST TISTICAL SOFTWARE

M y statistical packages are available for personal computers and spreadsheet pack-
age offer basic statistical routines. Using one of these you will avoid having to carry

A
out detailed and tedious calculations. It is essential to have access to at least one
suc package to use in conjunction with this book. This is particularly important for
the ast two chapters on the analysis of variance and regression.
T ere are two very large packages called SAS and SPSS. These will do everything
you ever thought of and a lot more besides. There are also a range of medium-sized
pac ages that should be adequate for your needs and several small packages that are
che p but limited. Some vendors offer cut-down versions at cheap rates for students.
Pric s for the big packages range up to a thousand pounds and you may have to
 Preface ix

pay an an uallicence fee; for smaller packages you should expect to pay up to several
hundred p unds.
It is dif 'cult to make firm recommendations because statistical software packages
are devel ing rapidly. In this age of microcomputers, I believe that menu-driven
software mbined with publication quality, interactive graphics and a reasonable
range of s andard analytical procedures is essential. If you use a PostScript printer,
you shoul check the printer drivers available in each package - even now some
software d es not support PostScript. Based on my own experience I would advise
you to co sider, at the very least, MiniTab, StatGraphics, Stat View and SyStat. You
should re .ew all the packages carefully and make a decision based on your needs.

BASS fro Bass Institute Inc., P.O. Box Chapel Hill, NC 27514, USA. Telephone
(919) 4 9 0729.
C-Stat fro Cherwell Scientific Publishing, 27 Park End Street, Oxford OX1 1HU,
UK. Tel phone (0865) 794884/794664.
Crunch St tistical Package from Crunch Software, 5335 College Avenue, Suite 27,
Oaklan , CA 94618, USA. Telephone (415) 4208660.
CSS from tatSoft, 2325 East 13th Street, Tulsa, OK 74104, USA. Telephone (918)
583 41 9.
Data Desk rom Odesta Corp., 4084 Commercial Avenue, Northbrook, IL 60062, USA.
Teleph e (800) 3235423.
Exstatix fr m Strategic Mapping, 4030 Moorpark Avenue, Suite 250, San Jose, CA
95117, SA. Telephone (408) 985 7400.
GUM fro NAG Ltd., Wilkinson House, Jordan Hill Road, Oxford, UK. Telephone
(0865) 11245. Fax (0865) 310139.
Mac 55/5 atistica from StatSoft, 2325 E. 13th Street, Tulsa, OK 74104, USA.
Telepho e (918) 583 4149.
MiniTab 5 atistical Software from Minitab Inc., 3081 Enterprise Drive, State College,
PA 168 1, USA. Telephone (800) 448 3555 and CLE Com Ltd., The Research
Park, Vi cent Drive, Edgbaston, Birmingham B15 2SQ, UK. Telephone (021) 471
4199.
NWA Stat ak from Northwest Analytical Inc., 520 N. W. Davis Street, Portland, OR
97209, SA. Telephone (503) 224 7727.
S-Plus fro Statistical Sciences Inc., 1700 Westlake Avenue N., Suite 500, Seattle,
WA 98 09, USA. Telephone (206) 283 8802. Fax (206) 283 8691 and Statistical
Sciences UK Ltd., 52 Sandfield Road, Oxford OX3 7RJ. UK. Telephone (0865)

I
61000. ax (0865) 61000.
SAS and J P from SAS Institute, SAS Circle, Box 8000, Cary, NC 27512, USA.
Telepho e (919) 467-8000 and SAS Software Ltd., Wittington House, Henley Road,
Marlow SL7 2EB, UK. Telephone (0628) 486933.
Solo 101 a d BP-90 from BMDP Statistical Software Inc., 1424 Sepulveda Boulevard,
Suite 3 6, Los Angeles, CA 90025, USA. Telephone (213) 4797799.
SPSS/PC from SPSS Inc., 444 N. Michigan Avenue, Chicago, IL 60611, USA.
Telepho e (312) 329 3300, and SPSS International, P.O. Box 115, 4200 AC,
x Preface
orinchem, The Netherlands. Telephone (31) 183036711. Fax (31) 183035839.
Sta Graphics from STGC Inc., 2115 E. Jefferson Street, Rockville, MD 20852, USA.
elephone (301) 984 5000, (301) 592 0050.
Sta istix II from NH Analytical Software, 1958 Eldridge Avenue, Roseville, MN 55113,
SA. Telephone (612) 631 2852.
Sta Pac Gold from StatPac Inc., 3814 Lyndale Avenue S., Minneapolis, MN 55409,
SA. Telephone (612) 822 8252 and Perifemalia, Snoekstraat 69, AIken, B-3570,
elgium. Telephone (32) 11 313754.
Sta view II/Super ANOVA and Statview SE and Graphics from Abacus Concepts, 1984
onita Avenue, Berkeley, CA 94704, USA. Telephone (800) 666 7828; (415) 540
949.
SyStat/SyGraph from SyStat Inc., 1800 Sherman Avenue, Evanston, IL 60201, USA.
lTelephone (708) 864 5670.
1

Wh mathematics?

It is inte esting to contemplate a tangled bank, clothed with many plants of many kinds,
with bi ds singing on the bushes, with various insects flitting about, and with worms
crawlin through the damp earth, and to reflect that these elaborately constructed forms,
so diffe ent from each other, and dependent on each other in so complex a manner,
have all een produced by laws acting around us. These laws, taken in the largest sense,
being G owth and Reproduction; Inheritance which is almost implied by reproduction;
Variabil ty from the indirect and direct action of the conditions of life, and from use
and dis se; a Ratio of Increase so high as to lead to a Struggle for Life, and as a
conseq ence to Natural Selection, entailing Divergence of Character and the extinction
of the I ss-improved forms. C. Darwin (1906, p. 669).

Mathemat cs and biology are two of the oldest sciences, both dating back to the
beginning of recorded history, but it is only in this century that they have come
together i the creation of modem mathematical biology. Well before the development
of our mo em view of natural history, much effort had been devoted to the systematic
cIassificati n of plants and animals; by the seventeenth century many species had
been desc ibed and identified. In 1686 John Ray defined species as groups of similar
individual characterized by 'distinguishing features that perpetuate themselves in
propagati n from seed' and noted that 'One species never springs from the seed of
another' ( urkhardt, 1981). In 1735 Linnaeus developed the two-name system of
cIassificati n for plants and animals that is used to this day. Believing in the constancy
of species, he said, 'There are as many species as the Creator produced different forms
in the beg nning' (Burkhardt, 1981). By the nineteenth century this view of species as
immutable entities was beginning to weaken; although naturalists accepted that species
were 'gen ine entities in nature, constant in their essential characters', they now added
'subject to non-essential, accidental variation' (Burkhardt, 1981). Although the classi-
fication of species cannot be called mathematical, we cannot do anything until we have
defined th entities we wish to study.
Having defined and classified species, it was natural to ponder the reasons why
these parti ular species were found in the world. The idea of change within the other-
wise fixed species was unavoidable in a world in which selective breeding of plants
and anima s was widely practised, and this led to intense debates concerning the origin
of the nat ral world: was the world created as we see it now or was it created in a
simpler fo from which our present world has evolved? Nearly all religions postulate
,,~ '''' Why mathematics?
a pr cess of creation and the idea of a progressive, albeit rather rapid, creation of
the orld is found in all cultures. But if the world has evolved from a 'simpler' state,
do e need to invoke a 'higher power' to guide and direct the process of evolution
or c n we find a theory that allows us to explain evolution without recourse to
exte al powers? If we can find laws of evolution, what kind of laws should they
be? hould the laws of evolution be causal and mechanistic as are the laws of cla5'sical
phy ics? The only physical law that depends explicitly on the direction of time is'
~tr'e,,,econd law of thermodynamics and this law predicts increasing disorder as time
.
passsHw can we reconcile this with the apparent increase in order as time passes
that we seem to see in biological evolution?
T 0 thousand years ago Aristotle reflected on the same problem (Ross, 1952): 'Why
sho ld not nature work, not for the sake of something, nor because it is better so,
but j st as the sky rains, not in order to make com grow, but of necessity? ... Why ...
sho d it not be ... that our teeth come up of necessity-the front teeth sharp, fitted
for taring, the molars broad and useful for grinding down the food ... [and] survived,
bein organized spontaneously in a fitting way; whereas those which grew otherwise
peris ed and continue to perish?' Since Aristotle believed that chance would destroy
rath r than preserve the adaptations we find in members of a given population, he
rejec ed his own proposal and continued: 'Yet it is impossible that this should be the
true iew. For teeth and all other natural things either invariably or normally come
abo t in a given way; but of not one of the results of chance or spontaneity is this
true' In this discourse Aristotle is concerned with the problem of selection rather than
evol tion, but he touched on the idea that the best adapted organisms should survive
whil those less well adapted should perish. Two thousand years later we find Darwin
won ering how undirected chance events could have produced the apparently directed
chari e found in the fossil record.
D",rwin' sinterest lay in the ways in which changes arose in natural populations and '"
he ewthatsubstantial changes can be brought about in animal and plant populations
by oss-breeding and selection. But while the work of breeders, who systematically
selec desirable characteristics, hinted at a mechanism for the emergence of novelty,
Da in could not see how such apparently directed change could be brought about
in n ture, if we assume that she is neutral and does not regard any particular change
as m re desirable than any other.
T e answer came to Darwin from a quite unexpected source, as he recorded in his
auto iography: 'In October 1838, ... fifteen months after I had begun my systematic
inqu'ry, I happened to read for amusement Malthus on Population, and being well
prep red to appreciate the struggle for existence which everywhere goes on, from
long continued observation of the habits of animals and plants, it at once struck me
that under these circumstances favourable variations would tend to be preserved,
and nfavourable ones to be destroyed. The result of this would be the formation
of a ew species. Here, then, I had at last got a theory by which to work .. .' (Darwin,
195 ). Wallace (1905) took his inspiration from the same source: 'One day something
brou ht to my recollection Malthus' Principle of Population ... I thought of his clear
exp sition of 'the positive checks' to increase ... which kept down the population ....
 Why mathematics? 3
It then oc urred to me that these causes or their equivalents are continually acting in
the case 0 animals also; and, as animals usually breed much more rapidly than does
mankind, t e destructions every year from these causes must be enormous in order to
keep dow the numbers of each species, since they evidently do not increase regularly
from year to year, as otherwise the world would long ago have become densely
crowded ith those that breed most quickly .... Why do some die and some live?
And the swer was clearly, that on the whole the best fitted live. From the effects
of disease the most healthy escaped; from enemies the strongest, the swiftest, or the
most cun ing; from famine, the best hunters or those with the best digestion; and
so on ..... The more I thought over it the more I became convinced that I had at length
found the I ng-sought-for law of nature that solved the problem of the origin of species.
(Cronin, 1 91, provides a fascinating account of the development of Darwinian theory
from Da in and Wallace to the present day.)
Malthu (1970) realized that if the growth of a natural population were unchecked
it would i crease exponentially, i.e. the numbers would double in a fixed time and
then doub e again and again in each equivalent time period. He also argued that our
ability to ncrease food production would increase only arithmetically, i.e. we can at
best incre se food production only by the same absolute amount in any fixed time.
Therefore, population growth would always tend to outstrip increases in food produc-
tion. Sinc human populations do not always increase exponentially, Malthus sought
to identify the factors that act to limit population growth. After examining the evidence
from a nu ber of countries, Malthus concluded that 'vice' (including abortion and
infanticide and 'misery' (including hunger and disease) acted as checks, although in
later wor s he added 'moral restraint' (abstinence), which he hoped might one day
replace 'vi e' and 'misery' as a check to population growth. In modem terms, Malthus
understoo that the growth of all populations must eventually be limited by an
increase i mortality or a decrease in fecundity with increasing density. Darwin and
Wallace n w took Malthus's essentially mathematical arguments and realized that for
all biologi al organisms an exponential increase in the number of individuals would
lead to a s ruggle for life in which many would die and few would survive. Given the
natural va iation between individuals in a given population, those that are in some
sense bett r adapted to their environment than the others will be more likely to survive.
We ca illustrate the consequences of exponential growth quite simply. We know,
for examp e, that one tsetse fly weighs about 30 mg. We also know that the maximum
rate of gr wth of a tsetse fly population amounts to a doubling in the numbers each
month (m king the tsetse fly a very slow grower by insect standards). So if the world
contained only one pregnant tsetse fly, and if this initial population grew at the maxi-
mum pos ible rate, at the end of one year the numbers would have doubled 12 times,
making a otal of 212, or about 4000, tsetse flies. At the end of 10 years the population
would ha e increased by this same factor of 4000 another ten times, making a total
of 400010 or 1036, tsetse flies, weighing about 3 x 1034g, which is five million times
the mass f the earth. Since tsetse flies have been around for rather more than 10 years,
the numb r of flies in existence today is only a tiny fraction of the number that could
have bee produced. Clearly there must be a constant thinning of the population, and
4 Why mathematics?
it s ms reasonable to assume that if the individuals of a population vary, those indivi-
dua s that have slightly advantageous characteristics will tend to survive and pass on
thei genes while those that possess characteristics that are slightly disadvantageous
will tend to be eliminated along with their genes.
hile evolution has resulted in the production of the most varied and wonderful
org nisms, just as remarkable is the extraordinary stability that we see in the wings
of t e dragon fly, for example, which have remained unchanged for millions of years.
To ccount for this we argue that if a particular characteristic is optimal in the sense
that any small change from that state is disadvantageous, there will be a tendency for
nat ral selection to maintain that state. Darwin's theory of natural selection therefore
atte pts to explain both speciation and stasis.
arwin's theory provided a convincing and detailed description of the processes by
whi h biological evolution occurs and it removed the need for a teleological theory of
evo uti on. Nevertheless, Darwin was aware that he did not have a mechanism that
wo ld explain the process of evolution at a deeper level. The laws governing inheri-
tanc are for the most part unknown', he wrote. 'No one can say why the same
pe liarity in different individuals of the same species, or in different species, is some-
tim s inherited and sometimes not' (Darwin, 1906, p. 15). The key to the mechanism
that Darwin lacked was discovered by Gregor Mendel, who in 1866 published the
resu ts of his work on inheritance in garden peas, Pisum sativum (Mendel, 1866). The
mo em theory that derives from Mendel's work is essentially 'atomistic'. Characteris-
tics f what we now call the phenotype are conferred on organisms by a small number
of i herited 'particles' that we now call genes. But although Mendel's papers were
sent to the Royal and Linnaean Societies in England (Fisher, 1936), his work was almost
enti ely neglected, perhaps because his observations contrasted so strongly with the
'con inuum' theory of pangenesis, proposed by Aristotle and accepted by Darwin
(Ay la and Kiger, 1980) in which inheritance involves a blending of fluids created in the
bod'es of the parents. To put Darwin and Mendel's work into prespective, it is worth
rem mbering that Schwann established the theory that living organisms were composed
of s parate cells only in 1839 and that cell division reached wide acceptance only after
the ork of Virchow in 1858 (Ronan, 1983). It was not until the middle of the nineteenth
cent ry that atomism achieved universal acceptance in physics and chemistry.
t the turn of the century Mendel's work was rediscovered, perhaps because the
intel ectual climate had become more favourable to his ideas. Biologists soon realized
that the new science of genetics provided a quantitative basis for Darwin's essentially
qual tative theory, leading to what we now call the neo-Darwinian synthesis. Genetics
itsel has been placed on a sound chemical basis.during this century with the elucidation
of t e structure of DNA, our increasing understanding of the biochemical basis of
gen tic replication and the establishment of molecular biology and genetic engineering.
s biological thought has evolved over the past three centuries, biology has become
mor quantitative and precise, Although Darwin was able to develop his theory without
mat ematics, the essentially mathematical idea of exponential growth was crucial to
the evelopment of his ideas. Mendel was obliged to use statistics to analyse his experi-
men s on peas, Watson and Crick needed advanced mathematics to determine the
 Some problems 5

structure of DNA from Franklin's X-ray photographs and geneticists now rely on
sophistic ted mathematics in the analysis and interpretation of their experiments. Now
we have orne full circle: we are applying mathematics to organisms and their inter-
actions, t e basis of much of Darwin's work, in the modelling of ecosystems. Having
started fr m Darwin's essentially qualitative theory of evolution, we have worked our
way do to an understanding of the process of evolution at a molecular level. Now
we are w rking our way back up again, synthesizing ideas and information from many
different isciplines, at many different levels, and creating in the last few decades a
quantitati e theoretical ecology.
Over 00 years ago, Galileo Galilei revolutionized natural philosophy by systemati-
cally app ing mathematics to the analysis of experimental data. 'Philosophy', he said,
'is writte in that great book which lies before our gaze-I mean the universe-but
we cann t understand it if we do not first learn the language and grasp the symbols
in which it is written. The book is written in the mathematical language, ... without
the help f which ... one wanders in vain through a dark labyrinth' (Needham, 1972,
p. 32). Bi logy is now following the same path and little can be done without mathe-
matics. espite the difficulties many of us have in learning mathematics, the power
it gives s more than repays the effort it takes to learn it.

E PROBLEMS

Before la nching fully into biostatistics, I shall introduce you to some of the problems
discussed in this book. Here I hope to persuade you that the answers are worth knowing;
in the cha ters that follow we will address these problems and many others. By the time
that you ave finished reading the book, I hope that you will agree that a knowledge
of statisti s will help you to analyse otherwise intractable problems and to draw reliable
and cons stent conclusions from your experimental data.

Suppose hat a friend of yours offers to playa game of chance with you. She will throw
a pair of ice and for £1 you are to guess the sum of the two numbers that turn up. If
your gue s is wrong, she keeps the pound; if it is right, she pays you as many pounds as
the num er you guessed. What number should you guess? Should you play with her at
all? If yo make your best guess, will you win or lose, and if so, how much?

1.1.2
In his w rk on inheritance in peas, Mendel (1866) crossed a number of pure-bred tall
plants ( ich have two 'tall' genes, T T) with pure-bred short plants (which have two
'short' g nes, tt). The plants of the first filial generation, Fl' inherited a 'tall' gene from
one pare t and a 'short' gene from the other. Since the 'tall' gene is dominant all of
the Fl p nts were tall. Mendel then crossed 1064 Fl plants and in the second filial
generation' F2' he obtained 787 tall plants and 277 short plants, as shown in Table 1.1.
6 Why mathematics?
Table 1.1 The results of an experiment conducted by Mendel in which he
compared the number of tall and short pea plants in the F2 generation to the
numbers he expected on the basis of his theory of inheritance. The expected
numbers are 1064 x 314 = 798 tall plants and 1064 x 114 = 266 short plants

Tall Short Ratio

plants plants Total talll short
Observed 787 277 1064 2.84
Expected 798 266 1064 3.00

N w if the probability that either of the two Fl parents will contribute a 'tall' gene to
the F generation is the same as the probability that either will contribute a 'short'
gene, and given that the 'tall' gene is dominant, we can show that there should be three
tall pI ts for every short plant, or 798 tall plants and 266 short plants in the F2 genera-
tion. n section 2.2.1 we shall see why Mendel decided that there should be three tall
plant for every short plant, but for the moment we note that the ratio of the number
of tal to short plants that he observed was 2.84 rather than 3. Does this refute his
theo that the ratio should be 3 or was it just chance that the observed ratio was
not q ite equal to the expected ratio? Suppose, for example, he had obtained 760 tall
and 04 short plants, giving a ratio of 2.5. Would these data support or refute his
theo ? Indeed, by how much could the ratio deviate from 3 before we would have
to sa that the data do not support the theory?

1.1.3 The polio vaccine

In th early 1950s poliomyelitis was recognized as a major disease causing crippling

paral sis and death, especially in children. In 1954 Jonas Salk developed a vaccine
again t polio (Snedecor and Cochran, 1989, p. 13). These days there is much talk about
the p ssibility of developing an effedive vaccine against malaria (Targett, 1991). If
you ere to carry out a trial of such a vaccine, how would you proceed?
In he study carried out to test the Salk vaccine, the results of which are given in
Table 1.2, 200745 children were vaccinated, of whom 33 developed polio. Those
cond ding the study needed to determine how many of these children would have
devel ped the disease if they had not been given the vaccine. The vaccinated children

Table 1.2 The number of paralytic cases of polio that developed among two
groups of children, one of which was given the Salk polio vaccine and the
other of which was given a saline solution

Number Paralytic Cases per

Group treated cases 100 000

Vaccinated 200 745 33 16

Placebo 201229 115 57
 Some problems 7

obviousl could not be used to discover what would have happened if they had not
been vac inated, so a separate control group of children was injected with a saline
solution lacebo rather than the vaccine. Of 201229 children given the placebo, 115
develope polio. The proportion of children who developed polio was greater among
those giv n the placebo (57 per 100000) than among those given the vaccine (16
per 100 0 0), so the vaccine appears to be effective in reducing the incidence of polio,
although in both cases only a small number of children developed the disease. Any
intervention, including vaccination, carries some risk. On the basis of these results,
would y u have your child vaccinated against polio? A trial involving nearly half a
million c ildren is both costly and time-consuming; could the same result have been
obtained ore quickly and at less cost by treating only 20 000 children in each category?
Anum er of other questions arise. How could the scientists carrying out the trial
be sure t at, as far as the vaccine and the disease were concerned, the control group
of childr did not differ significantly from the treated group of children? There is
a moral d lemma: since the trial would not have been carried out unless the scientists
already b lieved that the vaccine was effective in preventing polio, how should they
decide w ich children to vaccinate and which to treat as the control? The design of
experime ts is crucial and forms an entire area of biomathematics in itself, although
it is too ften overlooked. Months of hard work may be devoted to doing an experi-
ment in b ology only to discover that because of a flaw in the design of the experiment
or the pr sence of a factor that should have been included but wasn't, the data do not
provide t e answer to the question being asked. Indeed, the hallmark of a good scientist
is the abi ity to design experiments that will elicit the subtle changes and effects one
is seekin in research.

1.1.4 C ntrolling armyworm

The conv ntional way of controlling the African armyworm, Spodoptera exempta, is to
spray cro s with pyrethrum, but this pollutes the environment and kills many of the
predators that feed on armyworms. A more attractive control strategy involves the use
of biolog cal control agents, such as the bacterium Bacillus thuringiensis var. aizawai
(B.t.), wh'ch can be stored as a powder and applied in a water suspension.
In June 1988 an outbreak of the African armyworm was detected in a 40-acre wheat
field at gungugu Farm, near Nakuru, in Kenya (Brownbridge, 1988). A suspension
of B.t. in ater was applied at concentrations of 0.5, 1.0 and 2.0% weight by volume,
to three ifferent parts of a field infested with armyworm, with the results shown in
Table 1.3 B.t. was not applied in the control area. In the control area there is no obvious
reductio in the number of armyworm larvae over 4 days. Treatments A and B both
appear t reduce larval numbers by a factor of three after 4 days, while treatment C
reduces I rval numbers by a factor of about 75 after 4 days and appears to be very
effective. But are the slight reductions under treatments A and B significant or are
they sim Iy due to chance? Do treatments A and B differ from each other? For how
long wo Id we have to apply each of the treatments to achieve a 99% reduction in
the num er of larvae?
8 Why mathematics?
Table 1.3 Mean counts of armyworm larvae on each day after applying
three different treatments: Control-no treatment, A-0.5%, B-l.O% and
C- 2.0% R.t. suspended in water

Treatment
Time! days Control A B C
0 497 320 294 295
1 463 203 213 93
2 506 155 118 33
3 487 125 124 12
4 480 101 111 4

1.1. Summary

Whe you have finished reading this book, I hope you will be able to answer all of
thes questions without difficulty. These problems were chosen because they are easy
to st te and to appreciate but we will encounter more complicated problems. For
exa pIe, we have only considered experiments in which the outcome is determined by
a sin Ie factor: whether or not the children were vaccinated, whether the peas had tall
or sh rt genes. But more often than not we will be analysing data in which there are
man factors that may determine the outcome of an experiment, and we will want to
sepa te out the contributions of the various factors as best we can. Whenever possible,
we ill design our experiments in such a way that the contributions of the various
facto s that determine the outcome are independent of one another so that we can
asses their effects independently.

1.2 BRIEF OUTLINE

In this book you will encounter new words-Poisson distributions, cumulative distri-
butio functions, analysis of variance, regression, and many others. Just as we use Latin
nam s to indicate particular species and thus avoid lengthy taxonomic descriptions,
we n ed to do the same in biostatistics so that we can express sophisticated ideas
clear and concisely. This means thay you will have to learn a new language of words
and ymbols. We will, for example, define probability rather carefully, so that when
we s y, 'If I spin a coin, the probability that it falls heads is one half', we know precisely
what this means. We will then abbreviate the statement to: P(H) = 0.5.
o ten the variables we are studying are beyond our control. For example, the number
of ts tse flies caught in traps is high during the rains (Dransfield et aI., 1990). But the
humi ity is also high during the rains, so is it the rain or the humidity causing the
incre se in the catch? In fact, when it rains the temperature is generally low and perhaps
the igh catch has nothing to do with rain but rather is related to low temperatures.
Mu of biostatistical theory is devoted to developing techniques for separating out
the ffects of different factors, in this case rainfall, temperature and humidity, on the
 A brief outline 9
variable f interest, in this case the number of tsetse flies caught in a trap, the variation
of which we hope to understand.
The b ok is divided into the following main sections.

• Proba ility-the fundamental concepts and ideas underlying the application of

statisti s to biology.
• Repre entations-ways of picturing distributions of random numbers.
• Meas res-ways to measure and describe distributions of random numbers.
• Basic istributions-the most important distributions of random numbers.
• Testin hypotheses-how to frame questions and test predictions.
• Comp risons-how to make reliable comparisons among the results of experi-
ments.
• Analy is of variance-how to analyse variations in data in terms of contributions
from t e separate factors that may affect and influence the data.
• Regre sion-how changes in one variable are related to changes in another.

It is u fortunate, but true, that many biologists find statistics difficult and there is
no doubt hat the statistical analysis of biological experiments can become very compli-
cated. H wever, the number of concepts underpinning biostatistics are few.

• There are two fundamental laws of probability and two laws for combining
proba iIities.
• We ne d to understand the concept of probability distributions and their measures.
The t 0 most important measures are the mean and the variance. The mean is the
avera e value; the variance is a measure of the 'spread' of the observed values about
the man.
• When everal factors may affect the result of an experiment, we want to know how
much f the variation in the result can be ascribed to each factor. For example, if
we fee male and female rats on fresh and rancid lard, we might want to know how
the se of the rats as well as the freshness of the lard affects the weight gained by
the ra s.
• When we have several things going on at once, for example, crop growth, fruit
produ tion, rainfall and insect infestation, each of which may be important, we want
to kno how they are related, and which ones influence the others. We might say,
for ex mple, that rainfall and crop growth tend to increase or decrease together,
while i sect infestation and fruit production are inversely related, so that as the one
increa es the other decreases.
• We w nt to be able to measure variables, such as rainfall and temperature, and use
these 0 predict other variables, such as the growth of a crop. But we want to be
able t pick out those factors that affect the outcome, in this case the growth of
a crop, most strongly. If we then decide that the key factors are rainfall and tempera-
ture, e want to find the combination of the two that gives us the 'best' prediction
of cro growth, and we need to think about what we mean by the 'best' prediction.

These few ideas really contain everything that is covered in this book. If you can
get a fi grasp of the key underlying ideas, then the subtle and complex applications
10 Why mathematics?
sho ld not be difficult. If you lose sight of the underlying generalities, however, even
relat vely simple applications will seem confusing and difficult. On the other hand,
it is also important to develop technical skills because you do not always want to
start from the laws of probability. The English philosopher and mathematician Alfred
Whi ehead (1928) put it like this: Without a doubt, technical facility is a first requisite
for aluable mental activity: we shall fail to appreciate the rhythm of Milton, or the
pass on of Shelley, so long as we find it necessary to spell the words and are not
quit certain of the forms of the individual letters. In this sense there is no royal
road to learning. But it is equally an error to confine attention to technical processes,
excl ding consideration of general ideas. Here lies the road to pedantry.'
T ylor and Wheeler (1963) gave the following advice to young physicists: 'Never
mak a calculation until you know the answer. Make an estimate before every
calc lation, try a simple physical argument ... before using every derivation, guess
the nswer to every puzzle. Courage: no one else needs to know what the guess is.
The fore, make it quickly, by instinct. A right guess reinforces this instinct. A wrong
gues brings the refreshment of surprise.' I hope that you too will have the courage
and he confidence to guess.

1.3

1. I an experiment growing duckweed under controlled conditions in the laboratory,

it w s found that the number of fronds doubles every 2 days. Starting with one frond,
how many would there be after I, 2 and 3 months if the growth continued unchecked?

2. T e formation of locust swarms tends to occur after drought-breaking rains. The

proc ss of swarm development involves increased rates of multiplication, concentra-

Table 1.4 The results of a series of hypothetical experiments carried out to

test Mendel's hypothesis that the ratio of tall to short plants in the F2 genera-
tion is 3

Number of Number of
tall plants short plants Ratio

30 10 3.0
28 12 2.3
26 14 1.9
24 16 1.5
22 18 1.2
20 20 1.0
300 100 3.0
280 120 2.3
260 140 1.9
240 160 1.5
220 180 1.2
200 200 1.0
 Exercises 11
tion of t e locusts into small areas and behavioural and morphological changes
from the 'solitarious' to the 'gregarious' phase. In a study of the origin of plagues
of the So th American locust Schistocerca cancel/ata in Argentina, Hunter and Cosenzo
(1990) s owed that plagues began with a Season when three generations per year
were pos ible and plagues were sustained when only two generations were possible.
Plagues ypically took two years to increase to maximum size. The number of
generatio s per year is therefore an important factor in detennining the outbreak
and dedi e of plagues. Assuming that each female locust lays about 80 eggs in a pod
of which bout 30 survive to adulthood (Farrow, 1979), calculate the factor by which
the numb r of locusts could increase in two breeding seasons with one, two and three
generatio s per breeding season.

3. Imagi e that you repeated Mendel's experiment (section 1.1.2). Guess which of
the (hyp thetical) results given in Table 1.4 support his hypothesis that the expected
ratio of t 11to short plants in the F2 generation is 3 to 17 (You will find the answers
in Chapt r 6, Exercise 1.)

4. Plot g aphs of the number of annywonns and of the logarithm of the number of
annywo s against time from the data in Table 1.3 for the control and the three
treatmen s. What do you notice about the two sets of plots?
2

Probability

... young students, familiarwith animals,... are often disappointed that their painstaking
analysis,above allthe dull statisticalevaluation,finallyshows nothing more than what
a sensibleperson with eyes in his head and a good knowledge of animalsknows already.
There is, however, a difference between seeing and proving, and it is this difference
which divides art from science. K. Lorenz (1967)

A curious feature of science is that the foundations, which most people assume are
well established, are the least secure parts of the edifice. When we begin a new study,
we generally make a few vague statements to get us going and then race happily
onwards. The theory of probability and statistics is no exception and the philosophical
foundation of the theory is the subject of intense debate. A reasonably thorough
discussion of the foundations of statistics would require an entire course in itself, so
I shall simply give you my own biased view of how best to get going. (Hacking,
1979, discusses the foundations of statistics in his book Logic of Statistical Inference.)

2.1 THE CONCEPT OF PROBABILITY

Suppose that I spin a coin and ask, 'What is the probability of getting heads rather
than tails'? Most people would say 50% or one-in-two. But suppose we try to be more
precise and ask ourselves what we mean by this. There are several ways in which
we could expand the answer further.
One way would be to say that we are just as likely to get heads as to get tails.
But this only says the same thing in different words and what we probably mean is
that we simply have to take the definition as given, that we cannot explain it further.
Another explanation might be to say that if we throw a coin very many times,
we will get heads on about half the throws. To test this, I threw a penny one million
times (I confess with the aid of a computer pretending to be a penny) and had
500737 heads and 499263 tails. The proportion of heads is 50.07%, which is almost,
but not exactly, equal to the expected 50%. So how close to 50% do we have to be
before we can say that heads and tails are indeed equally likely?
Still another explanation might be that if someone throws a coin, it is worth betting
that it will land heads only if the odds you are offered are better than 1 to 1.
These answers represent three ways of looking at probability. The first, a priori,
approach says that the concept cannot be reduced further and so we simply have to
 The concept of probability 13

accept that a thing called probability exists and then proceed to define laws for com-
bining probabilities, derive theorems about probabilities (much as you may have done
with Euclidean geometry at school) and then try to match our theoretical predictions
with the outcome of whatever experiments we do.
The second, empirical or experimental, approach says that we should define proba-
bility in terms of long-run frequencies. This is certainly a useful way to think about
probability and is the way most people do think about it. In the first edition of his
book Logic, the English philosopher John Mill wrote (Bulmer, 1979, p. 5), 'Why in
tossing up a halfpenny do we reckon it equally probable that we shall throw heads
or tails? Because we know that in any great number of throws, heads and tails are
thrown equally often; and that the more throws we make, the more nearly the equality
is perfect.' But this long-run approach is not without its problems. We might agree
that there is a definite probability that I shall be run over by a bus tomorrow: but
if it does happen, I shall hardly be in a position to repeat the experiment many times.
The third way of looking at probability, which we might call the casino definition,
is particularly suited to those who gamble. While I would not like to suggest that
Mill led a dissipated life, he explicitly rejected the frequency definition in the later
editions of his book and gave the following definition instead (Bulmer, 1979, p.6):
'We must remember that the probability of an event is not a quality of the event
itself, but a mere name for the degree of ground which we ... have for expecting
it ... Every event is in itself certain, not probable: if we knew all, we should either
know positively that it will happen, or positively that it will not. But its probability
means to us the degree of expectation of its occurrence, which we are warranted
in entertaining by our present evidence.' This would seem to be a useful definition
for actuaries and insurance brokers.
In science you should always follow the approach that best suits you. In the first,
a priori, view of probability, we simply define an unbiased coin as one for which the
probability of heads is 0.5 and use our laws of probability to work out whether
500737 heads in one million throws is a reasonable outcome given our hypothesis
that the coin is unbiased. In other words, we make an hypothesis, do an experiment and
then see whether the resulting data are in reasonable agreement with our expectations:
if they are not we must reconsider the hypothesis and possibly our data as well. As
we shall see in this book, we do not sit around looking blankly at our data and then
suddenly jump up, saying, 'Eureka! tsetse flies like blue traps.' Usually we have a'
pretty good idea of what we think might happen and we are more likely to say, 'I
wonder if the colour of the traps affects the number of tsetse flies that we catch?
Let us catch as many flies as we can with blue traps and yellow traps, and then see
whether the data we collect supports our hypothesis that the catch is affected by
the colour of the trap.' In this book we will talk about testing hypotheses. This not
only makes for more rigorous statistics but helps us to remember that the reason we
apply mathematics to biology is not to generate large amounts of obscure numbers,
but to help to formulate clear and precise hypotheses to test.
Sometimes the second, the long-run frequency idea, is the most useful: if you are
struggling to decide on how to assign frequencies or probabilities, or indeed to decide
14 Probability
on what might be the important factors in an experiment, you can always ask: What
would I expect to happen if I repeated this experiment many times? We usually hope
that the results of our experiments on new insect trap designs, improved insecticides,
or high yielding strains of wheat, for example, will be used by other people in other
places and this might help you to think about all of the possible things that might
happen when the experiment is repeated many times.
The third, casino, approach can also be useful because it reminds us that the random
nature of our observations arises, at least in part, from our lack of knowledge of the
precise factors affecting our experiments. Given some information, we might then make
certain predictions concerning the outcome of an experiment. But more information
might lead us to change our prediction and this reflects well the provisional nature
of scientific knowledge.

2.2 THE LAWS OF PROBABILITY

Once we have agreed on some concept of probability, we need to develop our nota-
tion. We could make statements, such as, 'If I toss an unbiased coin, the probability
that it will land heads is 50%.' But this is long, and we will want to discuss considerably
more complicated situations. So let us use the letter P for probability and define our
scale of probability in such a way that anything that is certain to happen has a
probability of 1 and anything that is certain not to happen has a probability
of O. We will then say, for example, that if we toss an unbiased coin,

P(H) = 0.5, 2.1

which is read, The probability of getting heads is 0.5'. Sometimes we will want to
include explicitly the conditions under which our statement holds, and then we use
a '1' (bar) to indicate a conditional probability and write instead

P(Hlunbiased coin) = 0.5, 2.2

which reads The probability of getting heads, given that the coin is unbiased,
is 0.5'. In another situation we might then write

P(HI double-headed) = 1, 2.3

which reads, The probability of getting heads, given that coin is double-headed, is 1'.
We can combine numbers in two important ways: we can add them and we can
multiply them. The next thing we must do is to determine the equivalent laws for
combining probabilities.

2.2.1 The law of addition

I think you will agree that if we spin an unbiased coin

P(H) = P(T) = 1/2. 2.4

And I think you will also agree that with an unbiased die the probability of throwing
 The laws of probability 15
any number between 1 and 6 is 1/6, that is

P(l) = P(2) ... = P(6) = 1/6. 2.5

So if I were to throw a die and ask you to guess the number that turns up, you
would consider that the probability of guessing correctly is 1 in 6, of guessing
incorrectly is 5 in 6. You should demand odds of at least 5 to 1 to make it worth
your while playing the game.
So far so good. But suppose I throw a die and let you guess any two numbers
and we agree that if either of them comes up you win. What odds would then make
it worth your while to play? For example, you might now bet that the result will
be either a 2 or a 5, and so we want to know the probability of getting a 2 or a 5,
or in our new notation, we want to know the value of P(2 or 5).
Let us examine the problem more carefully. When we throw an unbiased die, the
reason for concluding that P(4) =
1/6 is as follows: There are six possible outcomes,
1, 2, ... 6. Since we assume that each outcome is equally likely, we argue that the
probability of getting 4, say, is simply equal to the number of ways we can get 4
(1) divided by the total number of ways the die can fall (6).
Suppose then that we throw a die and are allowed two bets. If you bet on 2 or
5, say, there are two ways in which you can win-namely, if the die falls 2 or
5-and four ways in which you can lose-namely, if the die falls 1,3,4 or 6. The
total number of outcomes is still 6, so that the chance of winning is 2 out of 6 (and
of losing is 4 out of 6) and we can write

P(2 or 5) = 2/6 = 1/3. 2.6

Clearly, if you are going to play this game you should demand odds of at least 2
to 1. Now if you think about this argument, it should be clear that we have simply
added up the number of (equally likely) ways in which you can win and divided that
number by the total number of (equally likely) ways the game can come out, so that
we can generalize Equation 2.6 and write

P(2 or 5) = (1 + 1)/6 = 1/6 + 1/6 = P(2) + P(5) = 1/3, 2.7

since adding the number of ways we can win and then dividing that number by the
total number of ways the dice can fall is just the same as adding the probabilities
of each outcome. But we must be careful: if you bet two numbers but both bets are
for 2, that is, if you make the silly bet that the outcome will be 2 or 2, then the
probability of a win is P(2 or 2) = 1/6, not 1/3, so that the rule only applies if the
alternatives are mutually exclusive: that is to say, either of the two outcomes might
happen but certainly not both.
Let us then summarize our law of addition. For two mutually exclusive events,
A and B, the probability that either A or B will occur is equal to the sum of their
separate probabilities: that is,
P(A or B) = P(A) + P(B). 2.8

Indeed, we have come full circle, for we effectively used our law of addition to derive
16 Probability
the results given in Equations 2.4 and 2.5. If the probability of getting heads is equal
to the probability of getting tails, and if our law of addition holds so that they add
up to 1, then each of them must be equal to 0.5.
We can now use this law to take another look at Mendel's experiments on peas
(Mendel, 1866). Mendel crossed homozygous TT peas with homozygous It peas,
where T indicates the gene for tallness and I the gene for shortness. Each plant in the
first filial generation. Fl' receives one gene from each of its parents and so they must
all have one T and one t gene. Thus the FI plants are all heterozygous, with genetic
constitution Tt. Since all of the plants in the first filial generation were tall, the gene
for tall plants is dominant over the gene for short plants. Mendel then cross-fertilized
these heterozygous plants and found that about 3/4 of them were tall and 1/4 were
short. In the second filial generation, Fl' there are four ways in which the Fl plants
can inherit genes from the male and female FI parents, as indicated in Table 2.1.
Since each of the four outcomes in Table 2.1 is equally likely and the tall gene is
dominant, the ratio of tall to short plants should be 3 to 1. We have effectively used
the law of addition to determine the predicted ratio of tall to short plants since

P(Fl is tall) = P(TT) + P(Tt) + P(IT) = 1/4 + 1/4 + 1/4 = 3/4, 2.9

while

P(Fl is short} = P(lI) = 1/4. 2.10

There are three important points to note. First of all, the four outcomes in Table
2.1 are equally likely only if the Fl plants are just as likely to inherit T genes as
they are to inherit t genes. When we use statistics, we must be aware of the assumptions
that underlie our analyses, especially when the assumptions are nol explicitly stated,
so that if the data do not agree with our expectation we have some idea of where
to look for an explanation. The second point is that Tt and IT are distinguishable
genetic constitutions: in the first case T came from the male and t from the female,
while in the second case the reverse was true. We must be careful to distinguish
events which, though seemingly the same, are in fact different. Lastly, we can apply
the law of addition because the outcomes are mutually exclusive. For example. we
could have TT or T/, but we could not possibly have TT and TI in the same plant.
In the previous chapter we saw that Mendel's tall and short plants did occur in
the ratio of about 3 to 1 in the Fl generation. However. being a good scientist. he

Table 2.1 The four possible (equally likely. mutually exclusive) genetic
constitutions of F.2 peas produced from heterozYKouS F1 males and females

F1 male F, female Fl

T T TT
T I TI
I T IT
I I II
 The laws of probability 17
Table 2.2 Mendel's data for pairs of charaders as they occurred in the F2 generation
of peas. The table gives the number of plants showing the dominant charader, the
number of plants showing the recessive character and the ratio of the two. The last
row gives the total number of plants, showing the dominant or recessive character
and the ratio of these two numbers

Character Dominants Recessives Ratio

Round vs wrinkled seeds 5474 1850 2.96
Yellow vs green seeds 6022 2001 3.01
Purple vs white flowers 705 224 3.15
Smooth vs constricted pods 882 299 2.95
Axial vs terminal flowers 651 207 3.14
Green vs yellow unripe pods 428 152 2.82
Tall vs dwarf stems 787 277 2.84
Total 14949 5010 2.98

did not stop there but went on to study several other pairs of characteristics, some
of which are given in Table 2.2 (Mendel, 1866). The ratios of the numbers of peas
having the dominant character to the numbers having the corresponding recessive
character cluster around a value of 3, in support of Mendel's theory, although none
of the numbers is precisely 3. The data for yellow and green seeds give a ratio very
close to 3 while the data for green and yellow unripe pods give a ratio less close
to 3. In later chapters we shall examine in detail the question 'How close is close7',
and we shall develop statistical tests that will enable us to make precise statements
about the meaning of 'closeness'. The theory that allows us to say just how close
is close was not widely understood in Mendel's day, but he stated this aspect of the
problem clearly: The true ratios of the numbers can only be ascertained by an average
deduced from the sum of as many single values as possible; the greah~ r the number,
the more are merely chance effects eliminated' (Mendel, 1866). This is in essence a
statement of our long-run frequency definition of probability. We see from Mendel's
data that the ratio of the total number of dominants to the total number of recessives
is very dose to 3:1.

2.2.2 Conditional probability

Before we can establish the second important law of probability, we need to introduce
the concept of conditional probabilities. We have agreed that if I throw one die,
the probability that it will come up showing 2, say, is 1/6, which we can write as

P(2) = 1/6. 2.11

Suppose I now throw a die and look at it but hide it from you. I tell you that it has
come up with an even number. What would you say is the probability that it is
showing 27 Since there are now only three possible outcomes-2, 4 and 6-the
18 Probability
probability that it is showing 2 is 1/3, which we can write as
P(2Ieven) = 1/3, 2.12
and Equation 2.12 reads 'The probability of getting a 2, given that the number is
even, is 1/3'.
We can now use the notion of conditional probability to define another important
concept, that of statistical independence. Suppose we throw two dice, which we
call a and b. Then the probability that a shows 2 has nothing at all to do with the
number that b shows and it must be the case that
P(a shows 21b shows even) = P(a shows 2) = 1/6, 2.13
while we know that the probability that a shows 2 does depend on whether or not
a shows even, since
P(a shows 21a shows even) = 1/3 ¥: P(a shows 2) = 1/6. 2.14
These observations provide us with an important test of statistical independence,
because if we find that
P(A I B) = P(A), 2.15
we can conclude that A and B are statistically independent. We should also note that
in general
P(AIB) ¥: P(BIA). 2.16
This is easily seen with our die since P(2Ieven) = 1/3, while P(evenI2) = 1.

2.2.3 The law of multiplication

Our law of addition gave us a rule for calculating P(A or B). We now want a rule
for calculating P(A and B). For example, we might want to throw two dice and work
out the probability that they will both show 2. We have already seen (Equations2.9
and 2.10) that Mendel's theory predicts that for peas in the second filial generation
P(plants are tall) = 0.75, P(plants are short) = 0.25, P(seeds are round) = 0.75, and
P(seeds are wrinkled) = 0.25. How then will we calculate P(plants are tall and seeds
are round) or P(plants are short and seeds are wrinkled)?
For any two outcomes, A and B, we want an expression for P(A and B). We can
consider this problem in two stages. First, what is the probability that we will get
B? Second, given that we do get 8, what is the probability that we will then get A
also? The answer to the first question is simply P(B) and the answer to the second
question is the conditional probability P(A I B). I now want to convince you that to
determine P(A and B) we must multiply these two probabilities together, so the law
of multiplication states: the probability of getting A and B is the probability of
getting A, given B, times the probability of getting B.
P(A and B) = P(A I B) x P(B). 2.17
 The laws of probability 19
In words, the probability of getting A and B is the probability of getting B times
the probability of getting A given that we already have B. If A and B are statistically
independent so that P(A IB) = P(A), as in Equation 2.15, then Equation 2.17 becomes

P(A and B) = P(A) x P(B). 2.18

Let us examine this simpler version first

Suppose we throw two dice, one coloured red and one coloured blue so that we
can tell them apart I think we agree that unless we glue them together, what happens
to the red one will not affect what happens to the blue one, so that they are statistically
independent This time we will play a game in which you have to guess a number
for each die. Suppose you guess 2 for red and 5 for blue; what then is the probability
that the red will show 2 and the blue will show 5 and you will win? That is, we
want to evaluate P(2 on red and 5 on blue).
Since the number of possible outcomes is small, we can write them all down as
=
in Table 2.3 and we see that there are 6 x 6 36 ways in which the dice can fall.
Out of the 36 possible outcomes, the red shows 2 on six occasions, so that the
probability that the red shows 2 is 6/36 = 1/6, just as we expect. We now look at
these six outcomes, the second row of Table 2.3, and on only one of them does the
blue die show 5. So the probability that the blue die shows 5, given that the red
die shows 2, is 1/6. Again, this is just as we expect, since the number that shows
on the blue die does not depend on the number that shows on the red die. We then
see that on 1/6 of 1/6 of the throws, the red shows 2 and the blue shows 5, so that
the probability that the red shows 2 and the blue shows 5 is (1/6) x (116) = 1/36,
in agreement with Equation 2.18, which is the law of multiplication for statistically
independent events.
To illustrate the law of multiplication for events that are not statistically
independent, imagine that we throw a die and ask: what is P(even and ~ 5)7 There
are two outcomes, namely 2 and 4, that are both even and less than or equal to 5 so
the answer must be 1/3. To confirm that the law of multiplication gives the same
answer we note that

P(even and~ 5) = P(evenl 5

~) x P( ~ 5) = (2/5)(5/6) = 1/3 2.19

Table 2.3 The 36 ways in which two dice can fall. The numbers in the
table give the sum of the numbers on the dice

Blue
1 2 3 4 5 6
1 2 3 4 5 6 7
2 3 4 5 6 7 8
Red 3 4 5 6 7 8 9
4 5 6 7 8 9 10
5 6 7 8 9 10 11
6 7 8 9 10 11 12
20 Probability
or, if you prefer,

P(even and::::;;5) = P(::::;; 51even) X P(even) = (2/3)(1/2) = 1/3. 2.20

In a case as simple as this we would not bother to use the law of multiplication and
would simply write down the answer directly, but in more complicated situations
we will have to use the law of multiplication. When faced with a statistical result
whose meaning may not be obvious, it is always good practice to think of a simple
example, such as the one above, and put in some numbers to help you to understand
the result.
We can now look again at the casino game we suggested in section 1.1.1, in which
we throw two dice and have to guess what the sum of the two numbers will be.
From Table 2.3 it is clear that we must guess a number between 2 and 12 and that
the dice can fall in 36 different ways. We can also see by counting the number of
ways various sums can turn up that P(2) = 1/36, P(3) = 2/36, ... , P(7) = 6/36, P(8)
= 5/36, and so on. To see how we can apply our laws to reach the same conclusion, we
note that we will get 5, for example, if we throw 1 and 4, 2 and 3,3 and 2, or 4 and 1.
We then have

P(5) = P[(I and 4) or (2 and 3) or (3 and 2) or (4 and 1)]. 2.21

Using the law of addition we can write out the 'or's so that

P(5) = P(I and 4) + P(2 and 3) + P(3 and 2) + P(4 and 1). 2.22

We can then use the law of multiplication (for independent events) to write out the
'and's, so that

P(5) = P(I) x P(4) + P(2) X P(3) + P(3) X P(2) + P(4) X P(I). 2.23

Since each of the individual outcomes has a probability of 1/6,

P(5) =(1/6) x (1/6) + (116) x (116) + (1/6) x (1/6) + (1/6) x (1/6) =4/36. 2.24

In a case as simple as this, it is easy to count the number of successes and the total
number of possible outcomes and divide the one by the other, and that is all that
our laws are doing for us. But if we have hundreds of possible outcomes, it quickly
becomes tedious if not impossible to enumerate all the possibilities; we then use our
laws to help us manipulate probabilities.
Now we can work out the best strategy for our casino game. Remember that you
pay £1 every time you guess wrong and you receive as many pounds as the number
you guess if you guess right. If you bet 5, say, your probability of wining is 4/36
and since you are then paid £5 your average winnings in 36 games are £20. Since
you pay £1 each time that you make a wrong guess and your probability of losing
is 32/36, your average losses in 36 games are £32. Table 2.4 summarizes the results
for some of the bets that you might make. Clearly, your best bet is 7 although 8 is
almost as good and you should also do well if you bet 9.
Let us return again to Mendel and his peas. We have seen that Mendel was able
to identify a number of dominant-recessive pairs of characters. This led him to
 The laws of probability 21
Table 2.4 The gains and losses in pounds for various bets in the
casino game described in section 1.2.1

Bet P(winning) Gain/36 games Loss/36 games Net gain

5 4/36 20 32 -12
6 5/36 30 31 -1
7 6/36 42 30 +12
8 5/36 40 31 +9
9 4/36 36 32 +4
10 3/36 30 33 -3

wonder if the pairs of characters were linked in any way. For example, genes for
round seeds are dominant over genes for wrinkled seeds, while genes for yellow
seeds are dominant over genes for green seeds; perhaps the round seeds will tend
also to be yellow while the wrinkled seeds will tend to be green. We can rephrase this
using our newly developed terminology and ask: is P(yellow Iround) = P(yellow)? is
P(wrinkled Iyellow) = P(wrinkled)? and so on. In other words, we can test the pairs
of characteristics to see if they are statistically independent. To do this, Mendel
proceeded as in his earlier experiments, but this time he crossed a pure line having
round yellow seeds with a pure line having wrinkled green seeds. All the resulting
hybrids had round yellow seeds, confirming the dominance of the round and yellow
characters. When these hybrids were in turn crossed, he obtained the results shown
in Table 2.5 from which we can calculate the proportions in which the various
characteristics occur. For example, we have

P(round) = 423/556 = 0.761 P(wrinkled) = 133/556 = 0.239

P(round Iyellow) = 315/416 = 0.757

P(wrinkledlyellow) = 101/416 = 0.243. 2.25

(I have used a circumflex 'A' to indicate a probability estimated as the ratio of two
frequencies to distinguish it from the 'true' underlying probability.) First of all we
see that about 3/4 of all the seeds are round. More importantly, if we consider only
the yellow seeds, about 3/4 of them are round and if we consider only the green
seeds, about 3/4 of them are round. Writing this out formally we have

P(round) = 0.761 ~ p(roundlyellow) = 0.757 ~ P(roundlgreen) = 0.771. 2.26

Table 2.5 Mendel's data on the distribution of seed colour and shape
Yellow Green Total

Round 315 108 423

Wrinkled 101 32 133
Total 416 140 556
22 Probability
Table 2.6 Bateson's data on the distribution of pollen shape and the colour
of the flowers in the F2 generation

Purple flowers Red flowers Total

Long pollen 1528 117 1645
Round pollen 106 381 487
Total 1634 498 2132

So the estimated probability that a seed will be round is about the same whether
the seed is green or yellow and the traits for colour and shape are statistically
independent (Equation 2.15).
With the rediscovery of Mendel's work, William Bateson repeated many of
Mendel's experiments (Bateson, 1909). In one experiment Bateson started from two
pure strains of peas, the first having long pollen and purple flowers, the second
having round pollen and red flowers. In the Fl generation all the plants had long
pollen and purple flowers, these being the dominant characters. In the F2 generation
he obtained the results shown in Table 2.6.
If we go through the same exercise as we did for Mendel's data, we find that

P(long pollen) = 1645/2132 = 0.772

P(round pollen) = 487/2132 = 0.229 2.27

in about the proportions expected on the basis of Mendel's theory. However,

P(long pollen Ipurple flowers) = 1528/1634 = 0.935

P(long pollen Ired flowers) = 117/498 = 0.215. 2.28

In this case, the two pairs of characters are not statistically independent, since

P(long pollen) = 0.772

=F P(long pollen Ipurple flowers) = 0.935 2.29

=F P(long pollen Ired flowers) = 0.215,

so that the probability that a pea has long pollen does depend on the colour of its
flowers.
Two points are worth noting. First of all, we have said that 0.935 and 0.215 in
Bateson's experiments differ from 0.772 but that 0.757 and 0.771 in Mendel's
experiments are not different from 0.761. Clearly Mendel's numbers are in closer
agreement with each other than are Bateson's but we are going to have to decide
when differences are and are not significant. The second point is that the statistics
merely suggest that something interesting might be going on since flower colour
and pollen shape seem to be related while seed colour and seed shape do not. But
the statistics tell us nothing at all about the biology. We now know that the reason
for these results is that genes are carried on chromosomes and if two pairs of genes
 Exercises 23
are found on the same chromosome they are likely to be inherited together, and
therefore linked, while if they are found on different chromosomes they are likely to
be inherited independently of one another. In other words, the statistics are important
and suggestive and give us hints as to where to look, but it is up to us to interpret
our data and draw out the biological implications.

2.3 SUMMARY

Probability is a subtle concept and it can be thought about in several different ways:
as something that is given, as the outcome of many repeats of the same experiment
or as a measure of our confidence or degree of belief in a certain event. How you
think about it will depend on you and on the problem at hand.
In order to progress we need to be able to calculate the probability that several
different things will happen in an experiment and the laws for combining probabilities
are central to everything that follows. I have not tried to present them in a rigorous
way; that has been done by others (Freund, 1972). What matters is that you know
how and when to apply the two laws for combining probabilities and the conditions
under which they hold and that you understand clearly when events are mutually
exclusive and the meaning of statistical independence.
We have already seen that even in simple experiments, such as the throwing of
two dice, the probabilities of each of the possible (combined) outcomes occurring
may differ. In the next chapter we will consider the probability distributions of the
outcomes of various experiments and think about ways to represent these distributions.

2.4 EXERCISES

1. Suppose that you throw one die. Write down: P(l), P(l or 2), P(l or 2 or 3), P(l
or 2 or 3 or 4), P(l or 2 or 3 or 4 or 5), P(l or 2 or 3 or 4 or 5 or 6). Write down
P(not 5), Le. the probability that it does not show 5.

2. Suppose that you throw two dice, a and b. Write down: P(a shows 1and b shows
2), P(a shows 1 and b shows 2 or 3).

3. Use Mendel's data in Table 2.5 to estimate P(round Igreen), P(wrinkledlgreen),

P(yellow), P(yellow Iround), P(yellow Iwrinkled), P(yellow and round).

4. Use Bateson's data in Table 2.6 to estimate P(round pollen), P(round pollen I purple
flowers), P(round pollen I red flowers), P(purple flowers), P(purple flowers I round
pollen), P(purple flowers Ilong pollen).

5. Use Bateson's data in Table 2.6 to estimate P(long pollen and purple flowers),
P(round pollen and red flowers).

6. Suppose that you throw three coins. How many ways can they fall so as to give
(a) three heads, (b) two heads, one tails, (c) one heads, two tails and (d) three tails?
If the coin is unbiased, what is the probability of each outcome (a) to (d)?
24 Probability
7. Suppose that we have four plants in the Fl generation as bred in Mendel's experi-
ment so that one has genetic constitution TT, two Tt, and one ft. Each of these
four plants is then crossed with each of the others. Suppose that each cross produces
four new plants. Calculate the expected number of plants with each genetic constitution
in the next generation. This illustrates the Hardy-Weinberg law (Roughgarden, 1979),
that in the absence of natural selection and with random mating the genotypic
proportions remain fixed. [Hint: You can solve this problem either by enumerating
all possible crosses or by using the laws of probability.]
3

Representations

If we fully apprehend the pattern of things of the world will it not be found that
everything must have a reason why it is as it is? And also a rule [of co-existence with
all other things] to which it cannot but conform? Is not this just what is meant by Pattern?
Hsii Heng (1209-1281) (Needham, 1972, p. 38)

When we throw a die we cannot predict with certainty which number will turn up.
In the nineteenth century scientists would have argued that our inability to predict
the outcome reflected our lack of knowledge, so that if we knew precisely how we
had thrown the die, the precise details of its shape and weight, how each tiny current
of air was moving, and so on, then we could, in principle, tell with certainty how it
would land. This was the essence of Mill's second statement of probability (section
2.1). Nowadays, uncertainty is regarded as being more fundamental than was pre-
viously thought, and quantum mechanics is based on laws that are themselves
probabilistic. Einstein never accepted quantum mechanics as a complete theory of
matter because of the probabilistic nature of its laws Gammer, 1974). 'Quantum
mechanics is very imposing,' he wrote, 'but an inner voice tells me that is still not
the true Jacob. The theory yields much, but it hardly brings us nearer to the secret
of the Old One. In any case I am convinced that He does not throw dice: Realizing
that combining quantum mechanics and general relativity in the study of black holes
leads to even greater uncertainty, Stephen Hawking comments (Boslough, 1984):
'God not only plays dice, but he sometimes throws them where they cannot be seen!'
Stewart (1990) in his book on Chaos now adds: 'For we are beginning to discover
that systems obeying immutable and precise laws do not always act in predictable
and regular ways. Simple laws may not produce simple behaviour. Deterministic laws
can produce behaviour that appears random. Order can breed its own kind of chaos.
The question is not so much whether God plays dice, but how God plays dice:
The discovery of chaotic behaviour, and its application to biology by Robert May
(1976) and others, shows that even entirely deterministic classical systems can be
inherently unpredictable and John Conway's delightful game 'Life' illustrates deter-
ministic but unpredictable dynamics (Gardiner, 1970, 1971, 1983). It is believed that
weather is chaotic so that no matter how powerful our computers become, we will
never be able to make precise weather forecasts more than a week or so into the
future. In later chapters we will have to consider the degree to which the 'randomness'
26 Representations
in our data reflects essentially unknowable features of our experiments and the extent
to which it reflects aspects of the experiment we have not understood or taken into
account.
If we throw a die many times and get 6, 3, 5, 3, 2, 5, 1, ... , say, these give us a
sequence of random numbers. Now, although we cannot say with certainty which
number will turn up next, we can make precise statements about the probability that
each number will turn up next. In other words, the probability distribution of the
outcome of even a single throw of a die can be known precisely. In our study of
statistics we are therefore concerned with determining the probability distributions
from which the numbers that we measure in such experiments come.

3.1 DISCRETE RANDOM VARIABLES

In statistics we are concerned not with the particular results of individual measurements
but with the distribution of the measured values. A great deal of work in statistics
is spent in identifying and describing the distribution associated with a particular set
of measurements or observations and the first thing we must do is to consider ways
of representing distributions of random numbers.

3.1.1 Frequency distributions

In the previous chapter we saw that when we throw a die, the probability of getting
any number between 1 and 6 is 1/6. To illustrate this, I used a computer to simulate an
experiment in which I threw a die 600 times, getting 95 ones, 101 twos, 107 threes,
112 fours, 91 fives and 94 sixes. As expected, each number came up about 100 times
or about one time in six. However, it is easier to see what is going on if we plot the
frequency distribution of outcomes in the form of a bar chart in which we draw a
series of boxes in such a way that the height of each box gives the number of times

Figure 3.1 The result of a computer simulation of an experiment in which a die is thrown
600 times. (a) The frequency distribution (b) the relative frequency distribution of the possible
outcomes. The two plots differ only in the choice of scale on the vertical axis.
 Discrete random variables 27

0.2
(a) (>' (b)
100 c
~
c
~
~
~ ~
:J J:: 0.1

~
i w ;
I
11!
Q)
a:
o M
2 4 6 8 10 12 2 4 6 8 10 12
Number Number

Figure 3.2 The result of a computer simulationof an experiment in which two dice are thrown
600 times and the pairs of numbers are added together. (a) The distribution of outcomes and
(b) the relative frequency distribution. The two plots differ only in the choice of scale on the
vertical axis.

the result for that box turns up. This is shown in Fig. 3.1a from which we see that
each number occurs with about the same frequency.
An alternative, but equivalent, way to plot the data involves dividing the number
of times each outcome occurs by the total number of events. This is called the relative
frequency distribution and is shown in Fig. 3.1b. The frequency distribution shows
us that we get a 1 in 95 out of 600 throws; the relative frequency distribution shows
us that we get a 1 with probability 0.16.
To illustrate this further I threw two (simulated) dice 600 times, added together
each. pair of numbers that turned up and then plotted the frequency distribution of
the sums. In Fig. 3.2 we can see that the two numbers add up to 7, for example, in
95 out of 600 or 16% of the throws.

3.1.2 Probability distributions

Now we know how to calculate the probabilities of each outcome in each of the
two experiments shown in Figs 3.1 and 3.2. We believe that the more times we throw
the dice the closer the observed relative frequencies will be to the theoretical
probabilities, which are plotted in Fig. 3.3. We can regard each of the relative frequency
distributions plotted in Figs 3.1 and 3.2 as estimates of the underlying 'true' probability
distributions shown in Fig. 3.3.
In passing we should note that these are examples of mathematical models. In
Fig. 3.3 we have analytical models in which we calculate the probabilities of various
outcomes from first principles; because the problems are so simple, we can do this
exactly. In biology we are seldom able to carry out exact calculations and so we often
make simulation models, which mimic our experiments, or at least what we believe
to be the most important aspects of our experiments. This is what we did to obtain
the results shown in Figs 3.1 and 3.2. We could then do an actual experiment by
throwing two real dice 600 times and comparing the outcome with the results of
28 Representations
~
2 02
- .~ ~

~ 0.1 ~
~ ~ 0.1
~ ~
~ ~

0.0 0.0
1 2 3 4 5 6 2 4 6 8 10 12
Number Number

Figure 3.3 The theoretical probability distribution functions corresponding to the two
experiments shown in Figures 3.1 and 3.2. (a) The probability distribution for the number that
shows when one die is thrown and (b) the probability distribution for the sum of the numbers
that show when two dice are thrown.

our analytical (Fig. 3.3) or simulation (Fig. 3.2) models in order to decide if there is
evidence for bias in our dice.

3.1.3 Cumulative distributions

Bar charts give us a graphical representation of the frequency or the probability with
which a particular outcome turns up in an experiment. However, instead of asking
What is the probability of getting a 37', we might ask, 'What is the probability of
getting a number less than or equal to 37' The latter question is important because
when we test an hypothesis against our data, we are usually looking for extreme
cases or unusual results and we want to know the probability of getting a result as
big as or bigger than (or as small as or smaller than) the one we actually observe.
For example, when we vaccinated the children against polio (section 1.1.3)we wanted
to know if the number who then developed polio was unusually small. If we are
testing an insect trap, we want to know if it catches an exceptionally large number
of insects.
If we throw one die, then using the law of addition we have

P(l) = 1/6 3.1

P(l or 2) = P(l) + P(2) = 1/3 3.2

P(l or 2 or 3) = P(l) + P(2) + P(3) = 1/2 3.3

and so on. In other words, to obtain the probability that the result is less than or
equal to 3, say, we add the probabilities of each outcome from i to 3. Now we
could always write these sums out in full. But I am sure you will agree that if we
had 100 possible outcomes, writing out the sums would rapidly become tedious. So
we introduce a new notation in which we use the Greek capital letter ,~ 'sigma', the
equivalent of our 5, to represent summing over various outcomes. We can then
 Continuous random variables 29
1.0 1.0
W ~

~o 0.5 :a
o
.. 05

QO QO
1 2 3 4 5 6 2 4 6 8 10 12
Number Number

Figure 3.4 The cumulative distribution functions, c.d.f.s, corresponding to the relative
frequencies given in Figures 3.1b and 3.2b.

rewrite Equation 3.3 more compactly as

3

P(I or 2 or 3) = L P(i).
i=l
3.4

The right-hand side of Equation 3.4 reads: 'the sum of P(i) from i = 1 to 3'. Since
Equation 3.4 gives us the accumulated probability for all outcomes less than 3, we
call it the cumulative distribution function or c.d.£. Writing it with a C, we have
3

C(3) = L P(i)
i=l
= 1/2 3.5

for the cumulative probability of throwing one die and getting a number less than
or equal to 3. Figure 3.4 gives the cumulative distributions for the data given in Figs 3.1
and 3.2. From the cumulative distributions we can read off directly the probability
that a number is less than or equal to any chosen value. The c.d.£. must always lie
between 0 and 1.

3.2 CONTINUOUS RANDOM VARIABLES

So far we have talked about experiments in which we measure discrete variables,

which can take on discrete values only, such as heads or tails, purple or white, and
1or 2 or 3. We will discuss other discrete variables such as the number of armyworms
in a field of maize or the number of ticks feeding on different rabbits. But we will
also measure continuous variables, which can take any value on a continuum of
numbers, such as the wing length of a fly or the weight of an engorged tick and
we will need to see how these fit into our scheme of things.
Fortunately, the ideas we have developed to describe discrete distributions all have
equivalents for continuous distributions and if we are clear about the one we should
be clear about the other. There is, however, an entire branch of mathematics that
deals with changes in continuous functions. This was developed in the eighteenth
30 Representations
century by Isaac Newton in England and Gottfried Leibniz in France. This study is
called calculus: differential calculus enables one to find the slope of a continuous
curve at any point on it and integral calculus enables one to find the area under a
continuous curve between any two points. We will not be able to develop and use
calculus in this book; we will be able to do all that we need without it. However,
for those of you who are familiar with calculus, remember that many of the results
we will obtain could be more easily obtained using the calculus.

3.2.1 Frequency distributions

As an example of a continuous distribution, consider the data shown in Table 3.1,

which give measurements in millimetres of the lengths of the wings of 100 houseflies
(Sokal and Hunter, 1955). In Table 3.1 we see that there are lots of 4.5s and only
one 3.6, but a graphical representation should make it easier to see how the lengths
of the wings are distributed. If the wing lengths occurred in discrete categories, as
in our earlier examples, we could plot them as a bar chart, which would indicate
how the lengths of the wings are distributed. We can still do this, even though the
wing lengths are continuously distributed, but we will have to create our own discrete
categories. For example, we can make a series of categories by taking 0.1 mm ranges.
Since the shortest wing is 3.63 mm long, let us take wing lengths from 3.6 to just
less than 3.7 mm for our first category, which in this example contains one fly, from
3.7 to just less than 3.8 mm for the second category, containing one fly, from 3.8
to just less than 3.9mm for the third category, containing two flies, .. ., from 4.5 to
just less than 4.6 mm for the 26th category, containing 12 flies, and so on. We can
then plot a bar chart, just as we did before and the result is shown in Fig. 3.5, which
we call a histogram. Since each interval is 0.1 mm wide, I have divided each frequency
by 0.1 so that the units are frequency per 1.0 mm rather than frequency per 0.1 mm.
The difference between a bar chart and a histogram is that in the histogram we have
chosen our own categories. (Many authors use 'bar chart' and 'histogram' interchange-
ably.) With a histogram we can make the categories as broad or as narrow as we
 Continuous random variables 31

E
E 100
(ij
a.
(I)
Q)

:: 50
o
•...
Q)
.D
E
~ 0
4.0 4.5 5.0 5.5
Wing-Iength/mm

Figure 3.5 The wing lengths shown in Table 3.1 plotted as a histogram. The vertical axis
gives the number of flies per millimetre whose wings fall into each 0.1 mm range.

E (a) E (b)
~ 100 ~ 100
Q) Q)
a. a.
~ ~
~ 50 ~ 50

I
~ 0
I
~ 0
4.0 4.5 5.0 5.5 4.0 4.5 5.0 5.5
Wing-Iength/mm Wing-Iength/mm

Figure 3.6 The histogram shown in Figure 3.5 replotted after (a) doubling and (b) halving
the widths of the categories to which the wing lengths are allocated. By plotting the data as
the frequency per millimetre, we keep the vertical axes the same in (a) and (b).

like, but unless we are very careful, we must keep the widths the same within any
one histogram. One unfortunately common way of misrepresenting data is to vary
the width of the categories to €hange the appearance of the histogram. An example
of such misrepresentation is given in Exercise 3.4.3.
I In Fig. 3.6 I have first doubled and then halved the widths of the categories in
Fig. 3.5. Doubling the width of each category to 0.2 mm produces a smoother
histogram but we are in danger of losing some of the details of the shape. Halving
the width of each category to 0.05 mm reveals more structure but the statistical
variation now makes it more difficult to pick out the underlying shape of the
distribution.
By plotting the number of flies per unit of measurement on the vertical axis we
are able to use the same vertical scale in all three cases in Figs. 3.5 and 3.6 and an impor-
tant result follows: the area under any part of the histogram is equal to the number
of flies in the corresponding range. For example, the height of the bar in Fig. 3.5
covering the range 5.0 to 5.1 mm is 70/mm while the width of the bar is 0.1 mm.
The area of the rectangle, 70 x 0.1 = 7, gives the number of flies in the range covered
32 Representations
by that bar. Finally, we can divide the area under any part of the histogram by the
total number of flies to get the probability that the length of a fly's wing falls into
a given range.
For discrete variables, we regarded each experimental frequency distribution as an
approximation to an underlying 'true' probability distribution. What then is the
equivalent of the 'true' theoretical probability distribution for continuous variables?

3.2.2 Probability density functions

The probability density function, or p.d.f., is the continuous equivalent of the

probability distribution for discrete variables. To illustrate the p.d.£., the data in
Fig. 3.5 have been replotted in Fig. 3.7, after dividing by the total number of flies,
so that the frequencies become probabilities. It then turns out that the normal distribu-
tion shown by the smooth curve in Fig. 3.7 is a good approximation to the p.d.£.
It is called the normal distribution because it is the most common distribution; things
are normally distributed in this way. Although the mathematical definition of the
normal distribution function might seem obscure, it is the most important distribution
in statistics, so we will digress for a while and consider it carefully.
We want a functional form for the bell-shaped curve in Fig. 3.7. Here we will not
derive it from first principles, but if we plot a graph of P{x) against x with

P{x) = e-x', 3.6

where e = 2.71828 ... is the number that provides the base for natural logarithms,
we will obtain a curve of the desired shape. When x = 0, P{x) = 1 and as x deviates
from 0 in either direction, P{x) becomes smaller. For very large values of x (in either
direction) P{x) becomes very small so that the curve flattens out along the x axis.
Now we want to be able to scale the height of our curve to match our data so we

Figure 3.7 The data of Figure 3.5, after dividing by the total number of flies. The line is the
probability density function for the normal distribution that best fits the data.
 Continuous random variables 33

multiply the right-hand side of Equation 3.6 by a number N to get

P(x} = Ne-:r2. 3.7

We also want to be able to shift our curve to match our data and we achieve this
by subtracting a number m from x so that Equation 3.7 becomes
P(x} = Ne-(r-m)2. 3.8

In section 5.3 we will see that m is the mean of the distribution. We also want to
be able to vary the width of our curve to match our data and we achieve this by
dividing x-m by a number 5 so that Equation 3.7 becomes
P(x} = Ne-(r-m)2/2s2 3.9

and if we include the extra factor of 2 then, as we will see in section 5.3, 5 is the
standard deviation of the distribution. Equation 3.9 is very important and you should
plot graphs of P(x} against x for various values of N, m and 5 to make sure that
you understand the effect that each has on P(x}. The values that give the best fit of
Equation 3.9 to the housefly data turn out to be 4.604 mm for m, 0.393 mm for 5
and 1.015/mm for N and these values were used to calculate the smooth curve in
Fig. 3.7.
We have to think carefully about the meaning of a p.dJ, such as the normal
distribution function. The continuous curve in Fig. 3.7 must be equivalent to the
rectangles that we use for our histograms. Since the curve is continuous, we can
imagine that it is made up of a very large number of very thin rectangles as in
Fig. 3.8 where each rectangle is 0.1 mm wide. The area under the curve between
any two values of x will then be equal to the area under the corresponding rectangles
and therefore equal to the probability that the length of a fly's wing falls into a
given range. Since the area under the entire curve, from - 00 to + 00 gives the
probability that the length of a wing takes on any value, it must be 1 and to ensure
that this is so the N in Equation 3.9 must be equal to

N = l/s.jiir. 3.10

Figure 3.8 The normal distribution curve of Figure 3.7. which we can approximate by a series
of narrow rectangles. In this plot the rectangles are 0.1 mm wide.
34 Representations
The concept of a probability density is central to much of what follows. First of
all we note that whereas a probability is always dimensionless, the units of P(x), in
this example, are probability per mm, as expected for a 'density'. Since the height
of each rectangle is equal to the probability per mm that a fly has a certain wing
length, P(x) gives the probability per unit length that a wing is x mm long. Suppose,
for example, that we choose a wing length of 5 mm. P(5) is equal to 0.67/mm for
m = 4.604 mm and 5 = 0.393 mm (note the units). If we multiply P(5) by 0.1 mm, the
units cancel and the probability that the length of a wing lies within 0.1 mm of 5 mm,
that is between 4.95 and 5.05 mm, is 0.067, or

P(4.95 mm:::;;wing length:::;; 5.05 mm) = 0.067. 3.11

If we count the number of flies in Table 3.1 whose wings are between 4.95 and
5.05 mm in length, we have seven flies, which means that the probability, estimated
directly from the data, that the length of a wing is in this range is 0.07, in good
agreement with the value obtained from the p.d.f. of 0.067. If we multiply P(5) by
0.2 mm we have

P(4.9 mm :::;;wing length:::;; 5.1 mm) = 0.12. 3.12

We can again count the number of flies in Table 3.1 whose wings are more than 4.9
and less than 5.1 mm long, and we get 14 flies, so that the probability, estimated
directly from the data, that the length of a wing lies in this range is 0.14, again in
good agreement with the value obtained from the p.dJ. of 0.12. This is valid if the
range we consider is sufficiently small, because we can then treat P(x) as constant
within this range and we are effectively working out the area under the curve
bdween 4.95 and 5.05 mm in the first case and between 4.9 and 5.1 mm in the
second1.
The two important things to remember are:

• The probability density function or p.d.f. is the probability per unit of measure-
ment that the measured variable has a particular value .
• The probability that a particular observation lies in a given range is equal to the
area under the curve of the p.d.£. over this range and if the range is small we can
calculate the area approximately by taking the value of the p.dJ. at the midpoint
of the range and then multiplying by the range.

3.2.3 Cumulative distribution functions

Just as for discrete distributions, we want to be able to evaluate the probability that
the length of the wing of a fly is less than 3.00 mm or greater than 4.00 mm, and
so on. If we approximate the continuous probability density function by a histogram,
we can proceed as we did in the case of discrete distributions and simply add up
the probabilities per unit of length for all of the bars in the histogram below the
critical level and then multiply this number by the width of the bars. If we do this
for the data on houseflies, we obtain the cumulative distribution function, or c.d.f.
 Continuous random variables 35

1.0

; 0.5
o

0.0
4.0 4.5 5.0 5.5
Wing-Iength/mm

Figure 3.9 The cumulative distribution function (c.d.f.) for the housefly data shown in
Figure 3.5.

shown in Fig. 3.9, corresponding to the histogram in Fig. 3.5. Formally we write
n

C(xn) = L P(x )t5x

i= 1
n 3.13

where Xn labels the bars and t5x is the width of the bars.
For continuous distributions the c.dJ. has several advantages over the p.dJ. First
of all, whereas the p.d.£. is a density and gives the probability per unit of measurement
that a certain result is observed, the c.d.£. is a probability and gives the probability
that, in our example, the length of a housefly's wing is less than or equal to a set
value. Secondly, although we can calculate the c.d.£. by first dividing the range of
the observed variables into discrete categories, constructing a histogram and adding
up the areas of successive histograms, there is a more direct way of obtaining the
c.d.£. We simply take the length of the shortest wing, 3.63 mm, and we have one fly
whose wing length is less than or equal to this. We then take the next smallest wing
length, 3.71 mm, and we have two flies whose wings are less than or equal to this
length. We proceed in this fashion up to the longest wing and then divide the number
of each fly by the total number of flies to get the c.d.£. You see that we do
not need to worry about setting up categories. This is illustrated in Table 3.2 and
Fig. 3.10.

Table 3.2 The flies from which the data in Table 3.1 were obtained were ordered
according to the lengths of their wings so that fly 1 has the shortest wing, fly 2
has the next shortest and so on. The c.M. is obtained by plotting the number of
each fly in the ordered sequence divided by the total number of flies against the
length of the wing for each fly as shown in Figure 3.10

Fly number 1 2 3 ... 99 100

c.d.f. 0.01 0.02 0.03 ... 0.99 1.00
Wing length 3.63 3.71 3.83 ... 5.40 5.57
 36 Representations

1.0

....;
~ 0.5
u

0.0
3.5 4.0 4.5 5.0 5.5
Wing-Iength/mm
Figure 3.10 The c.d.£. obtained by plotting the cumulative probability in Table 3.2 against
the wing length. The wiggly line gives the experimental values and the smooth line the
theoretical c.d.£. for a normal distribution, with m = 4.06 mm and 5 = 0.39 mm.

We would also like to be able to calculate the theoretical c.d.£. corresponding to

the theoretical p.d.£. given by Equation 3.9. Formally, the theoretical c.d.£. is the area
under the curve below each value of x.2 As it happens, the c.d.£. for the normal
distribution cannot be obtained analytically, even using calculus, but we will see in
Chapter 5 that the c.dJ. for the normal distribution is tabulated in books on statistics
and is easily calculated with the aid of a computer. In Fig. 3.10 I have used a computer
to calculate the theoretical c.dJ. for the housefly data and this is indicated by the
smooth line.
Finally, we note that since
n

C(xn) = L P(xj)c5x
i= 1
3.14

and
n-1

C(xn_ 1) = L P(Xj)c5x, 3.15

i= 1

C (Xn) - C(Xn-1) = P(xn)c5x, 3.16

so that we can calculate the p.d.f. simply by subtracting pairs of values from the c.d.e

3.3 SUMMARY

For both discrete and continuous random variables we can represent their distributions
in two ways. The probability density function gives us the probability that a particular
variable will fall into a given, small range. The cumulative distribution function gives
 Summary 37
the probability that a particular variable will be less than or equal to a particular
value. At first sight the p.dJ. seems to be more fundamental and easier to grasp.
However, as you will discover, we will be concerned almost exclusively with c.dJ.s
and hardly at all with p.d.f.s. The reasons for this are threefold.
First of all, the p.dJ. is a density and to specify a particular value we need also
to specify the units in which it is measured, whereas the c.d.£. is a pure probability
and does not depend on the units in which it is measured. Secondly, if we know the
c.dJ. we can immediately calculate the p.d.f. for any interval simply by subtracting
the values of the c.d.f. at either end of the interval. Thirdly, as we will discover,
when we examine our data we will be concerned to identify extreme events, ones
that are unlikely to occur on a given hypothesis. The c.d.f. gives us directly the
probability of particular extreme events occuring.
We do not always want to specify every distribution in all its detail. We therefore
need to find ways to summarize the most important properties of our data in terms of
a few key parameters. This is the subject of the next chapter.

NOTES

r
1 For those who know calculus, the exact expression is

P(a ~ x ~ b) = P(x)dx.

r
2 Using calculus

C(a) = 00 P(x)dx.

r
3 For a continuous distribution

C(b) - C(a) = P(a ~ x ~ b) = P(x)dx

3.4 EXERCISES

1. (a) Using a computer, generate 600 evenly distributed discrete random numbers
between 1 and 6 and plot the data as a bar chart. (b) Generate a second set of random
numbers, add them to the first, and then plot the sum of the numbers as a histogram.

2. Estimate from Figs 3.3b and 3.4b the probability that the sum of the two numbers
on a pair of dice shows 5 or less.

3. The data given in Table 3.3 and plotted in Fig 3.11 were published by the National
Science Foundation in the United States of America. They show a decline during the
1960s and 1970s in the number of Nobel Prizes won by Americans after the numbers
had peaked in the 1950s (Tufte, 1986). This graph was published to show that unless
funding was increased, the United States of America would rapidly fall behind its
competitors in scientific research. Why is this graph misleading? (Answer at the end
of the examples.)
 38 Representations
Table 3.3 The number of Nobel Prizes won by scientists in five countries
from 1901 to 1974

Dates
Dates USA Germany UK USSR France
France
1901-1910 1 12 5
5 2
2 6
6
1911-1920 2 7 3
3 0
0 5
5
1921-1930 4 8 7
7 0
0 4
4
1931-1940 9 8 7
7 0
0 3
3
1941-1950 14 5 7
7 0 0
0 0
1951-1960 29 3 9
9 5 0
5 0
1961-1970 26 6 12
12 4 6
4 6
1971-1974 13 1 7
7 0
0 0
0

30
USA
Germany
! ~ ----UK
''::
a.
'0 __
•• USSR
•..
mm ••

Q)

~ 10 •.•.•.•.•.•.•. France
::J
z

o o 0 0 0 0 0 0 v
.•.• C\I C') V II) <.0 ,... ,...
...-
o
I
...-
I

.•..
I
.,....
C\I
.,....
C')
I I
.,....
V
...-
••

II)
...-
<.0
...-
I

,...
Dates

Figure 3.11 Nobel Prizes in science for seleded countries, 1901-74.

4. Use Equation 3.9 and the values of m, sand N given in the text to calculate the
number of flies whose wing lengths lie between 4.5 and 5.0 mm. Compare your result
with the number given in Table 3.1. Use Fig. 3.10 to determine the number of flies
whose wing lengths lie between 4.5 and 5.0 mm.

5. The following are the wing lengths of ten of the 100 flies given in Table 3.1,
chosen at random: 4.40, 4.35, 4.25, 4.68, 5.40, 4.43, 3.63, 4.46, 4.43, 4.63. Plot the
p.d.f. and the c.d.£. for these data and compare the plots with Figs 3.5 and 3.10.

[Answer to Question 3: Each point on the graph gives the number of prizes won in
a 10 year period, except the last set of points which only cover 4 years. Not
surprisingly, the number of prizes fell in all countries at the end of the plot. Plot the
number of prizes won by scientists from each country per decade by multiplying
the- numbers in the last row by 10/4.]
4

Measures

How can it be that mathematics, being after all a product of human thought which
is independent of experience, is so admirably appropriate to the objects of reality?
Is human reason, then, without experience, merely by taking thought, able to fathom
the properties of real things? In my opinion the answer to this question is briefly
this: as far as the propositions of mathematics refer to reality, they are not certain;
and as far as they are certain, they do not refer to reality.
A. Einstein (1954)

If you were asked to describe a giraffe, you might say that it is a very tall, four-legged
mammal with a long neck and a patchy brown coat. This description is brief; you
could elaborate on it at considerable length. But it is often useful to be able to
indicate one or two key features of an object that are sufficient to describe its form
or function. Of course, the particular features we name depend on our reasons for
wanting to classify the object. You would not give a child Linnaeus's description of
an elephant and you would not tell a biologist that an elephant is a large grey beast
with tusks. If you wished to distinguish a horse from a zebra you might say that
the zebra has stripes that the horse does not have, while to distinguish a horse from
a donkey, you might say that the donkey has long ears that the horse does not
have. In biology we constantly classify the objects in the world around us and we
will need'to do the same in statistics. The 'objects' we are concerned with in statistics
are probability distributions. In the last chapter we discussed the uniform distribution
(Fig. 3.1), which we obtained when we threw one die, the triangular distribution,
which we found when we threw two dice and added the numbers together (Fig. 3.2),
and the normal distribution, obtained when we considered the length of the wings
of 100 houseflies (Fig. 3.7). In this chapter we will think about ways of classifying
probability distributions so that we do not need to specify the distributions in every
detail but rather can pick out the key properties as we need them.

4.1 MEASURES OF LOCATION

We will start by considering ways to determine the location of a distribution.
Figure 4.1 shows the even and triangular distributions from Figs 3.1 and 3.2 with
the two distributions plotted on the same scales.
40 Measures

Figure 4.1 (a) The even distribution and (b) the triangular distribution shown in Figures 3.1
and 3.2. The arrows indicate the mean values.

4.1.1 Means
The first thing we might say about the two distributions is that the triangular distribu-
tion is shifted to larger values on the horizontal axis than is the even distribution.
The most commonly used measure of location is the mean or average value. Using
a horizontal bar to indicate means we have
n

x= L x/n,
i= 1
4.1

where n is the number of measurements that we make. For the even distribution Xj
is the number that turns up on the ith throw; for the triangular distribution, xj is the
sum of the two numbers that turn up on the ith throw. In both cases n is the number
of throws. For our housefly data, Xj is the length of the ith wing. We will also use
<-. -> to indicate average values or means since (x) is sometimes easier to write than
a long bar over the top of an expression.
In Fig. 4.1 the mean of the numbers chosen from the even distribution is 3.48 and
this is less than the mean of the numbers chosen from the triangular distribution which
is 6.89. From Fig. 3.3 it is easy to see that the means of the theoretical even and
triangular distributions are 3.5 and 7, respectively, so that our estimates from the
simulation are quite close to the values for the underlying theoretical distributions.
For the housefly data the mean wing length is 4.604 mm.
We use the Greek letter Jl ('mu') for the 'true' mean of the underlying theoretical
distribution to distinguish it from m, the mean estimated from the data in a particular
experiment. For the even distribution Jl = 3.5, for the triangular distribution Jl = 7.
(We will use Roman letters for parameters estimated from data and Greek letters
for the equivalent parameters of the underlying distribution that we are trying to
estimate.)
The mean can also be expressed in terms of the probability of each possible
outcome. In the simulation in which we threw one die 600 times, we had 94 ones,
 Measures of location 4I

84 twos, 113 threes, 106 fours, 99 fives and 104 sixes, so that the mean m is

m = (1 x 94 + 2 x 84 + 3 X 113 + 4 X 106 + 5 X 99 +6 X 104)/600

=1 X (94/600) +2 X (84/600) +3 X (113/600)

+4 X (106/600) + 5 X (99/600) + 6 X (104/600)

=1 X P(I) +2 X P(2) + 3 X P(3) + 4 X P(4) + 5 X P(5) + 6 X P(6)
6
= L
i= 1
iP(i) = 3.5 4.2

and in general the mean value of any variable Xi is the sum over Xi times the
probability of getting Xi I,

Xi = LXi
i
P(x;). 4.3

Just as we learned to combine probabilities, we need also to be able to combine

means. Suppose, for example, we have calculated the mean length <I) of the left
wings of 100 houseflies and the mean length <r) of the right wings of the same
flies. If we now want <I
+ r), the mean length of all the wings, left and right, we
can either add up the lengths of all the wings, left and right, and divide by 200, or
simply note that provided we have the same number of wings in both cases, the
mean of the sum is the sum of the means (as shown in the Appendix, section 4.5.1),
so that
<I + r) = <I) + <r). 4.4
The mean has several especially desirable properties, but it is not the only measure
of the location of a dishibution and sometimes other measures are more appropriate.
An alternative to the arithmetic mean given by Equation 4.1 is the geometric mean,
G, which is the nth root of the product of the individual measurements, so that

G = (Xl X X2 ••• X xn)I/n. 4.5

Taking logarithms of both sides of Equation 4.5 gives

In G = (In Xl + Inx2• •• + Inxn)/n, 4.6

so that the logarithm of G is the arithmetic mean of the logarithms of the individual
measurements.

4.1.2 Medians

Another measure of location is the median, which is the measurement that falls in
the middle of the distribution so that there are as many items below it as above it.
(For a continuous distribution, the median is the point that divides the area under
the p.dJ. into two equal parts, which is also the point on the c.d.f. corresponding to
a probability of 0.5.) For example, the mean of 1, 3, 4, 8 and 9 is 5, while the median
42 Measures
is 4. For the housefly data given in Table 3.1, reading number 50 is 4.59 mm and
reading number 51 is 4.60mm, so we take 4.595 mm as the median wing length,
which is close to the mean wing length of 4.604 mm.
As an example of a situation in which you might wish to use the median rather
than the mean, suppose you carried out an experiment to measure the time it takes
insects to die after they have been sprayed with an insecticide. If you start with 100
insects and then measure the time it takes each one to die, you can calculate the
mean lifetime of the insects exposed to the insecticide. But if the last insect takes a
long time to die, you might spend most of your time waiting for that one to die.
To obtain the median lifetime, on the other hand, you need only to know the time
at which the 50th insect dies, after which you can stop the experiment.
A further advantage of the median is that if we have outliers in our data, that is
to say, a few points that are very much larger or very much smaller than the others,
these points will influence the mean greatly. For example, the mean of 2, 4 and 6 is
4 and the mean of 2, 4 and 100 is 53, while in both cases the median is 4. There
are many techniques that have been developed for handling outliers in data and these
are discussed in detail by Winer (1971, p. 51).
If the distribution is symmetrical, the mean and the median of the underlying
distribution are the same although the values calculated from the data may differ. One
way to test if a set of numbers comes from a symmetrical distribution is to compare
the mean and the median; if they differ significantly we conclude that the underlying
distribution is not symmetrical.

4.1.3 Modes

Another measure of the location of a distribution is the mode, which is the location
of the most probable outcome. For example, in our example of throwing two dice
(Fig. 3.2), the mode estimated from the data is 7 and of course the mode of the
theoretical probability distribution is 7. In the case of our housefly data as plotted
in Fig. 3.6, the mode is 4.6.

4.2 MEASURES OF SCALE

Looking at the frequency histograms in Fig. 4.1, we see not only that the triangular
distribution is shifted to the right of the even distribution, but the shapes of the two
distributions differ; in particular, the triangular distribution is wider than the even distri-
bution. So the next thing we need is a measure of the spread or width of our
distributions.

4.2.1 Variance
To obtain a measure of the width of the distributions, we subtract the mean, since
we have already taken that into account. For each measured point, we are then left
with the deviation from the mean. One possible measure of the spread would be the
average value of the deviations from the mean. However, the mean of 1, 2 and 3,
 Measures of scale 43

for example, is 2 so that the deviations from the mean are - 1, 0 and 1 and the mean
deviation is Of One way around this is to square the deviations before taking the
average. For the three numbers given above the squared deviations are 1, 0 and 1
and the mean squared deviation, which we call the variance, is 2/3. The variance
may be written
n
V = L (x - i)2/n. 4.7
i=l

The variance is important in biological statistics where we are usually concerned

to discover the causes of the variation in our data. In section 8.4.5 we will discuss
an experiment to compare three different designs of tsetse fly traps. The number of
flies that we catch depends on the design of the trap, the day of the experiment and
the site in which each trap is placed. By designing the experiment carefully, it turns
out that we can calculate a variance corresponding to each factor in our experiment.
This technique, which is the subject of Chapter 8, is called the analysis of variance, or
'ANOY A' for short, and it tells us which factors (trap, day or site) contribute
significantly to the variations that we observe in the data (number of flies caught).
Just as we learned to combine probabilities and nteans, we need to know the rules
for combining variances. We can show (Appendix, section 4.5.2) that if we have two
independent random variables, the variance of their sum is the sum of their variances,
so that if, for example, V(l) is the variance of the lengths of the left wings of 100
houseflies and V(r) is the variance of the lengths of the right wings, the variance of
the lengths of their right wings and their left wings, added together, is

V(l + r) = V(l) + V(r) 4.8

so that for independent random variables the variance, like the mean, is additive. If
the left and right wing lengths are not independent we have an extra term in
Equation 4.8 (see Equation 9.28).
We can also write an expression for the variance, as we did for the mean, in terms
of probabilities:
V(x) = «(x - i)2) = I: (x - i)2 P(x). 4.9

In fact, for any function of x we can always write the mean value of f(x) as the sum
over f(x) times the probability of getting x, so that

(f(x) = I:f(x)P(x). 4.10

Finally, we should note a computational formula that we will use in later chapters:

V(x) = I:(x - i)2/n = I:~

/n - 2iI:x/n + XZ
= I:/~n - (I:x/n)2 = (~) - (X)2. 4.11
We will use the right-hand side of Equation 4.11 to calculate variances in Chapter 8.

4.2.2 Standard deviation

Although ANOY A will be central to much of what we do, the units of the variance
are not the same as those of our original data: when we measure the length of the
44 Measures
wings of houseflies in mm, the variance has units of mm2, and we would like to
measure the width of the distribution in the same units as the original variable. For
this reason we define the standard deviation as the square root of the variance (for
our houseflies the units are then millimeters again) and
n 1/2

5 =[ il
~
(Xi - J1i/n
J . 4.12

The standard deviation is the root-mean-square value of the deviations about the mean.
As it stands, Equation 4.12 is correct. However, we do not usually know the true
mean; rather, we have to use the value of the mean estimated from the data. Since
the estimated mean minimizes the mean-square deviation about the mean, using m
instead of J.l in Equation 4.12 will systematically underestimate, or bias, the standard
deviation. Fortunately, it is possible to show that if we divide by n - 1 instead of
n, this will remove the bias and our equation for estimating the standard deviation
becomes
n 1/2

5 =
[
il
~
(Xi - m)2/(n - 1)
J 4.13

This argument can be made more rigorous (Bulmer, 1979, p.130). What we are
doing is dividing by the number of independent parameters we have in our estimate
of s which we call the number of degrees of freedom. In our calculation of the
standard deviation, we start with n data points, each of which is independent of the
others, so that we have n degrees of freedom. When we calculate the mean we use
up one of these degrees of freedom, leaving n - 1, and this is the number we use
in Equation 4.13. '
For the even distribution, Fig.4.1a, the standard deviation is 1.67 and for the
triangular distribution, Fig.4.1b, it is somewhat greater at 2.37. For the housefly
data, the calculated standard deviation is 0.39 mm.
We use the Greek leUer (1 to indicate the 'true' value of the standard deviation
so that the experimental standard deviation s is an estimate of (1.

4.2.3 Standard deviation of the mean

The mean value of a set of numbers is an estimate of the mean of the underlying
distribution: so how accurate is this estimate of the underlying or 'true' mean?
Suppose we have a set of measurements Xi so that their mean is

m = x~/n. 4.14

Since the variance of a sum is the sum of the variances (Equation 4.8 and Appendix,
section 4.5.2),

V(~
Xi) = ~ V(xi), 4.15

and if each Xj is chosen independently from the same distribution, they all
 Measures of shape 45
have the same variance, 52, so that

V(!;x) = ns2. 4.16

The standard deviation is the square root of the variance so that

s(!;x
j) = j;s, 4.17

and the standard deviation of the mean, sm' is

sm = s(!;x/n) = j;sln = s/.~ 4.18

(Many authors call 5 the standard deviation and Sm the standard error. To avoid
confusion, I call 5 the population standard deviation and Sm' the standard deviation
of the mean.) The importance of Equation 4.18 is that it shows that as we make
more measurements, the accuracy of each measurement, which is given by 5, will
remain constant but the accuracy of the mean, which is given by Sm' will improve
as the number of measurements increases. Since the estimate of the mean improves
as the square root of the number of measurements, we need to make four times as
many measurements if we want to double the accuracy of our estimate.

4.2.4 Ranges

Just as the mean is not the only measure of the location of a distribution, the standard
deviation is not the only measure of the dispersion. An alternative to the standard
deviation as a measure of dispersion is the range, which is the difference between
the largest and the smallest values in the data. For the even distribution (Fig. 4.1a),
the range is 5, for the triangular distribution (Fig. 4.1b), it is 10 and for the housefly
data (Table 3.1) it is 1.94 mm.
Both the range and the standard deviation, like the mean, are sensitive to the
presence of outliers and so we sometimes use the interquartile range, which has
properties rather like the median. The median is the point that occurs half-way up
the distribution, and we define the lower quartile as the point that occurs one-quarter
of the way up the distribution and the upper quartile as the point that occurs
three-quarters of the way up the distribution. For example, for the numbers 1, 3, 4,
7 and 10, the lower quartile is 3, the upper quartile is 7 and the interquartile range
is 7 - 3 = 4. If the number 10 in the series is replaced by 100, the range and the
standard deviation change considerably while the interquartile range remains the same.

4.3 MEASURES OF SHAPE

In nearly all of our work we will be concerned with means and variances or standard
deviations, but we will sometimes want to know more about the shape of a distribution.
We may want to know if a distribution is symmetrical or not. We may want to
know if it has long tails extending out on either side or if it terminates abruptly.
The two measures of the shape of a distribution that we will consider are the skewness
and the kurtosis.
46 Measures
4.3.1 Skewness

The distributions we have considered so far are all symmetrical about the mean. As
an example of a skew distribution that is not symmetrical about the mean, consider
Table 4.1, which gives the number of Prussian army officers kicked to death each
year by their horses in each of ten army corps in the 20 years from 1875 to 1894
(von Bortkewitsch, 1898).
From Table 4.1 it is clear that in most of the army corps in most years no officers
were kicked to death. However, in 1882 four officers in corps 9 were kicked to death,
and in three cells in Table 4.1 we see that three officers were kicked to death in the
same corps in the same year. The most convenient way to represent the distribution
of deaths is to calculate the frequency distribution so that we record the number of
cells in Table 4.1 in which no officers were kicked to death, the number of cells in
which one was kicked to death, and so on. The result is shown in Table 4.2.
The data in Table 4.2 are plotted as a bar chart in Fig. 4.2. The distribution is
clearly skew. Since the tail is to the right, we say that the distribution is skew to the
right; if the tail were on the other side, we would say that the distribution is skew
to the left. We now need a measure of the skewness of the distribution shown in
Fig. 4.2. We would like it to be zero for a symmetrical distribution, positive if the

Table 4.1 The number of Prussian army officers kicked to death by their
horses in each of ten army corps in each year from 1875 to 1894

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

75 1 1 1
76 1 1 1
77 1 1 2
78 2 1 1 1 1
79 1 1 2 1 1
80 2 1 1 1 2 1 3
81 . 2 1 1 1
82 1 1 2 4 1
83 1 2 1 1 1
84 1 1 2 1 1
85 2 1
86 1 1 1 3
87 2 1 2 1 1 2
88 1 1 1
89 1 1 1 1 2 2
90 2 1 2 2 1 2 2
91 1 1 1 1 1 3 1
92 2 1 1 1 1 1
93 1 2 1
94 1 1
 Measures of shape 47

Table 4.2 The frequency with which a given number of officers were
kicked to death by their horses in ten Prussian army corps in each of
the 20 years from 1875 to 1894. kj deaths occurred in each of nj of
the cells in Table 4.1

Number of Observed
deaths, kj frequency, nj

0 109
1 65
2 22
3 3
4 1
5+ 0

Figure 4.2 Bar chart of the data shown in Table 4.2.

distribution is skew to the right and negative if the distribution is skew to the left.
<
We have already defined the mean as <x) and the variance as (x - i)2), so it makes
<
sense to define the skewness as (x - i)3). But just as the variance of our housefly
wing lengths has dimensions of mm2, this definition of skewness would have units
of mm3• We could take the cube root, which for our houseflies would have units of
millimeters, but since the skewness is to be a measure of shape, we prefer it to be
dimensionless, and so instead we divide by the cube of the standard deviation to get

5= < (x - i)3)/53. 4.19

With this definition, a symmetrical distribution has skewness equal to zero. For
example, the numbers - 1, 0 and 1, for which the standard deviation is 1, have
skewness = {[(- 1)3 + (0)3 + (1)3]13 }/13 = O. Furthermore, the numbers -1, 0 and
4 have skewness 0.32 while the numbers - 4, 0 and 1 have skewness - 0.32. The
numbers - 2, 0 and 8 also have skewness 0.32, so that the skewness depends only
on. the shape of the distribution and not on the width, and it is positive when the
tail is on the right, negative when the tail is on the left.
48 Measures
Returning to our Prussian army officers, the mean of the frequency distribution is

m = r.kjP(k;) = r.kjnJn = 0.610/corps/year 4.20

where nj is the number of cells in Table 4.1 in which kj officers in anyone corps
were kicked to death in anyone year, and n is the total number of cells in Table
4.1 which is 10 x 20 = 200. In 10 years, for example, we would expect an average
of 0.61 x 10 = 6 army officers to be kicked to death in each corps. The variance is

V = r.(kj - m)2 P(kj) = r. (k j - m)2nJ(n - 1) = 0.611/corps2/year2, 4.21

so that the standard deviation is 0.782/corps/year. The skewness is

5 = r.(kj - m)3 p(k;)/S3 = {r.(kj - m)3nJn}/s3 = 1.24, 4.22

and as you can see from Figure 4.2, the distribution is skew to the right.

4.3.2 Kurtosis

The last measure we will consider is the kurtosis, K, which tells us if the distribution
has long tails or short tails. If the distribution has long low tails and a narrow peak
in the middle, the kurtosis is high and the distribution is called leptokurtic. If the
distribution has a broad hump in the middle and short tails, it is called platykurtic.
We define the kurtosis as the fourth power of the deviation about the mean divided
by the fourth power of the standard deviation,

<
K = (x - i)4 )/54, 4.23

so that it too is dimensionless.

Consider, for example, the two distributions in Table 4.3 illustrated in Fig. 4.3.
Both distributions have a mean of 0, a standard deviation of 0.834 and, since they
are symmetrical, a skewness of o. However, they are clearly different and the difference
is in the kurtosis, which is 4.2 for the first and 1.4 for the second. Note that the
distribution with the longer tails has the greater kurtosis. We will see in the next
chapter that the normal distribution has a kurtosis of 3 and this is taken as the
standard value against which we compare the kurtosis of other distributions. For the

Table 4.3 Two frequency distributions. In the first the outcome i

occurs mj times. In the second the outcome i occurs times. The "j
mean, standard deviation and skewness are the same for both but m
has a kurtosis 4.2 while for" it is 1.4

i mj "j
-2 1 0
-1 4 8
0 14 8
1 4 8
2 1 0
 Appendix 49

Figure 4.3 Bar charts of the distributions given in Table 4.3.

frequency distribution of ill-fated Prussian army officers, the kurtosis is

K = {I:(kj - m)4n/n }/S4 = 4.37, 4.24

rather more than the standard value of 3. (Many statistics packages calculate Fisher's
" statistics rather than the skewness and the kurtosis and are discussed in the Appendix,
sedion 4.5.3.)

4.4 SUMMARY
In this chapter we have developed ways of measuring the position, the width and
the shape of statistical distributions. In most of what we do we will use the mean,
the standard deviation or the variance, the skewness and the kurtosis. But we have
also seen that there are other measures, such as the median and the interquartile
range, that we could use instead, and if our data are not 'well behaved', for example
if we have outliers that we are not sure of, these other measures may be more appro-
priate. As we will see in the next few chapters, the mean and the standard deviation
enable us to make precise statements about our data but only if we know the 'true'
underlying distribution. But suppose that we do not know what the underlying
distribution is. Even worse, suppose that our data are not stridly quantifiable as
would be the case if we only had categories such as very small, small, medium, large
and very large. Provided we can order the categories, we can still determine measures
such as the median while parameters such as the mean may be of little or no use.
In these situations we will use what are called non-parametric statistics, of which the
median is one, and we will discuss these further in Chapter 8.

4.5 APPENDIX

4.5.1 The mean of a sum

The mean of a sum is the sum of the means.

Ifwemeasurexj and yJor i = 1,2, ... n and let Zj = Xj + Yj' then <z) = <x) + <y).
 50 Measures
We have

(z> = I:z;ln = I:(x + Yj)/n = I:x;ln + I:y;ln = (x> + (y>.

j 4.25

4.5.2 The variance of a sum or difference

For independent random variables, the variance of a sum or a difference is the

sum of the variances.
Suppose that we measure two random variables, x and Y, whose true means are
J1. and v. Then

V(x + y) = L)Xj
~ + Yj - J1.- V)2p(Xj and Yj) 4.26
j j

and if x and yare statistically independent,

P(Xj and Yj) = P(x;)P(Yj) 4.27

so that

V(x + y) =L (Xj - J1.)2P(Xj) L P(Yj) +L (Yj - V)2P(Yj) L P(Xj)

j j j i

- 2L(xj - J1.)P(xj)L(Yj - v)P(Yj)' 4.28

i j

The last term is the mean value of x - J1. times the mean value of Y - v, both of which
are zero. Furthermore,

P(Xj)
~ = P(Yj)
~ = 1, 4.29
I )

so that

V (x + Y) = V (x) + V (y). 4.30

If x and yare not independent, we have an extra term in the expression for the
variance of the sum. We discuss this in section 9.5.

4.5.3 Fisher's" statistics

Fisher's " statistics are similar to the skewness and kurtosis described here but are
adjusted for biases in small samples, the equivalent of dividing by n - 1 when we
calculate the variance or the standard deviation (Bliss, 1967, p. 144).
Fisher's" statistics S· and K" are

S· = SN2/[(N - I)(N - 2)], 4.31

which, if the original data are normally distributed, is itself approximately normally
distributed with mean zero and variance 6N(N -1)/[(N - 2)(N + I)(N + 3)] and

K" = KN2(N + 1)/[(N -1)(N - 2)(N - 3)] - 3(N -1)2/[(N - 2)(N - 3)], 4.32
 &e~
~ 51
which, if the original data are normally distributed, is itself approximately normally
distributed with mean zero and variance 4(N2 - 1) V (S·)/[(N - 3)(N - 5)]. Fisher's
K statistics are used to test if the skewness and kurtosis of a given distribution differ
significantly from the values obtained for a normal distribution.

NOTE

1 For continuous functions

(x) = f:oo xP(x)dx

4.6 EXERCISES

1. Calculate the means, standard deviations and skewness of (a) - 1,0,1; (b) - 1,0,4;
(c) - 4, 0, 1; (d) - 2, 0, 8.

2. Use Equations 4.3 and 4.10 to calculate the mean, standard deviation, skewness
and kurtosis of (a) the even distribution generated by throwing one die and (b) the
triangular distribution generated by throwing two dice and adding the numbers
together.

3. Mark the position of the median of the data (4.595 mm) on the p.d.f. given in
Fig. 3.5 and decide if it divides the area under the curve into two equal parts. Do
the same for the c.dJ. given in Fig. 3.10 and decide if it corresponds to a probability
of 0.5.
4. Calculate the standard deviation of the following numbers: 4.32, 2.79, 5.14, 3.87
and 2.31 using both V(x) = 1:(x - i)2 In and V(x) = <~ > - < >2.
X (These formulae
use n, the number of measurements, rather than n - 1, the number of degrees of
freedom. In both cases we can multiply the final answer by nl(n -1) to allow for this.

5. Determine the range, lower quartile, upper quartile and interquartile range for the
housefly data given in Table 3.1

6. Use a computer to generate three sets of normally distributed random numbers

with mean 0 and standard deviation 1, the first set containing ten numbers, the
second 40 and the third 160. Calculate the mean, the population standard deviation
and the standard deviation of the mean of each set of numbers. Note the relationships
between these estimates for the three sets of numbers.

7. You measure the lengths of the wings of 16 flies and obtain a mean of 4.60 mm
and a population standard deviation of 0.40 rom. What is the standard deviation of
the mean? How many wings would you need to measure to obtain an estimate of
the mean that was accurate to within 0.025 rom?

8. Work out the mean, standard deviation, skewness and kurtosis of the Prussian
army officer data in Table 4.2.
5

Distributions

I know of scarcely anything so apt to impress the imagination as the wonderful form
of cosmic order expressed by the 'Law of Frequency of Error', The Law would have
been personified by the Greeks and deified if they had known of it. It reigns with
serenity and in complete self-effacement amidst the wildest confusion. The huger the
mob, and the greater the apparent anarchy the more perfed is its sway. It is the perfed
law of Unreason. F. Galton (1889)

In the previous chapters we considered examples of random distributions: &om the

throw of a die we found the uniform distribution which is flat; from Mendel's experi-
ments on peas we obtained a distribution that can take on only one of two values;
&om the lengths of the wings of houseflies we had the 'bell-shaped' normal distribution;
and &om the fate of Prussian army officers we found a skew distribution. In this
chapter we will consider the most common distributions and their properties and we
will discover that a small number of distributions will suffice to describe most biological
situations.
We have already seen that if we throw two dice, the numbers that turn up on
each one are uniformly distributed between 1 and 6, while the sum of the pairs of
numbers on the two dice has a triangular distribution ranging from 2 to 12. We can
therefore generate a triangular distribution by adding together pairs of numbers each
taken from a uniform distribution. In this chapter we will examine several basic distri-
butions as well as the relationships among them.
To convince you thaI: we need to understand the properties of distributions,
consider the data on the polio vaccine (Table 1.2). The vaccinated children appear
to have gained some protection from polio, but few children succumbed to polio in
either group and we need to know if the effect of the vaccine is statistically significant.
In other words, does the difference represent a real effect of the vaccine or is it just
chance that one number turned out to be less than the other? If we can establish the
intrinsic variability of the data and if we can then show that the difference between
the proportion of vaccinated and the proportion of unvaccinated children who develop
polio is greater than this variability, we can reasonably argue that the vaccine did
have a significant effect. But to determine the intrinsic variability of the data, we
need to determine the underlying distribution, and this takes us into the subject of
this chapter.
 The binomial distribution 53

In this chapter we will discuss the binomial, Poisson, normal, X'1. ('chi-square'), t
and F distributions. We will start with the binomial distribution, which we get when
there are pnly two possible outcomes, such as heads or tails, true or false, success
or failure. When one of the two outcomes occurs only rarely, we will see that the
binomial distribution may be approximated by the Poisson distribution. When the
number of successes is sufficiently large, both the binomial and the Poisson distribution
may be approximated by the normal distribution. Sometimes we will want to combine
variables, each of which follows a normal distribution, and we will see that the sum
of the squares of normally distributed variables follows a X2 distribution, while the
ratio of a normally distributed variable to the square root of a variable that follows
the X2 distribution follows a t distribution. Finally, we will see that the ratio of two
X2 variates follows an F distribution.

5.1 THE BINOMIAL DISTRIBUTION

We get a binomial distribution when there are only two possible outcomes in an
experiment, such as heads or tails, tall or short, round or wrinkled. Suppose, for
example, that we want to know if a coin is biased. It is no good throwing it once
since it will certainly land either heads or tails, but if we throw it 100 times and get
45 heads and 55 tails, we can then ask if this is or is not indicative of a bias in the
coin. ihe binomial distribution will give us the probability of getting 45 heads and
55 tails, or indeed any other combination of heads and tails, and this will enable us
to decide if the observed outcome is so unlikely that we should regard the coin as
biased.
When there are only two possible outcomes, we use the words success and failure
to describe the alternatives, and we write p for the probability of success and q for
the probability of failure, bearing in mind that success and failure are arbitrary terms
and may correspond to heads and tails, tails and heads, tall and short or any other
pair of charaderistics. Then since the probabilities of all possible outcomes must add
up to 1 (section 2.2),
p + q = 1. 5.1

Now consider Table 5.1, which shows the various ways in which two or three
coins can fall. If we throw two coins and apply our Law of Multiplication, the
probability of getting one heads and one tails is pq and this can arise in two ways
(HT or TH), so that using the Law of Addition, the probability of getting one heads
and one tails in any order is 2pq. In general the probability of getting k heads in n
throws can be obtained by expanding (p + qt and then choosing the kth term in
the series. Alternatively, the probability of getting k successes in n trials may be
calculated using (Bulmer, 1979, p. 84)

n!
P(k successes In trials) = rfqn-k 5.2
k!(n - k)!

where p is the probability of success and q is the probability of failure in anyone

54 Distributions
Table 5.1 The possible outcomes when throwing either two or three coins. The table gives
the probability of each outcome, the number of ways the coins can land showing 0, I, 2 or
3 heads, the relative frequencies of each combination of heads and the probability of getting
each number of heads

2 coins

Outcome HH HT TH IT
Probability pp pq qp qq

Heads 2 1 0
Frequency 1 2 1
Probability p2 2pq q2

3 coins

Outcome HHH HHT HTH THH HIT THT TTH TIT

Probability ppp ppq pqp qpp pqq qpq qqp qqq

Heads 3 2 1 0
Frequency 1 3 3 1
Probability p3 3p2q 3pq2 q3

= =
trial. (Remember that n! n X (n - 1) X (n - 2) ... I, O! 1 and that XO = 1.) For
example, the probability of getting two heads in three throws of a coin is

P(213) = -'3' p2ql = 3p2q, 5.3

2!1!

in agreement with the value given in Table 5.1. For an unbiased coin for which
= =
p q 0.5, we expect, on average, to get two heads in three throws with probability
P(213) = 3/8.
To summarize: we get a binomial distribution whenever there are only two possible
outcomes, the probability of success or failure is constant from one trial to
the next and each trial is independent of all the others. Then if we know the
probability of success in one trial, Equation 5.2 gives us the probability of getting
k successes in n trials.
The binomial distribution is illustrated in Fig. 5.1. Keeping n, the number of trials,
constant at 10 (Figs 5.la, b and c), the distribution shifts to the right as p, the prob-
ability of success, increases from 0.1 to 0.9. When the probability of success is zero,
the number of successes is always zero and when the probability of success is one,
the number of successes is always equal to n. All three distributions peak at pn
(1, 5 and 9 in Figs 5.1a, band c, respectively) and we will see below that the mean
of the binomial distribution is in fact pn. To see the effect of changing the number
of trials, Fig. 5.ld shows the distribution for 100 trials with p set to 0.5. For
=
p 0.5, the width of the distribution for ten trials (Fig. 5.1b) is about 4, while the
width of the distribution for 100 trials (Fig. S.ld) is about 12, so that increasing the
 The binomial distribution 55

Figure 5.1 The probability of getting a given number of successesfor the binomialdistribution
where the number of trials, n, and the probability of success, p, are (a) n = 10, P = 0.1;
(b) n = 10, P = 0.5; (c) n = 10, P = 0.9; (d) n = 100, P = 0.5.

number of trials by ten times increases the width of the distribution by about three
times. We will see below that the width is in fact proportional to the square root of
the number of trials. The reason the distribution for 100 trials (Fig. 5.1d) looks narrower
than the distribution for ten trials (Fig. 5.1b) is because of the change in the horizontal
scale.

5.1.1 Families of eight

To illustrate the binomial distribution, consider the following problem. One occa-
sionally finds large families in which all of the children are boys or all of the children
are girls. Can these families of all boys or all girls be ascribed to chance alone or is
there a genetic propensity for children in large families to be of the same sex? Table
5.2, extracted from birth registers in Saxony between 1876 and 1895 (Giessler, 1889),
shows the number of boys and girls in families with eight children from a study of
53680 families and we see, for example, that 215 families had 0 boys and 8 girls
while 342 had 8 boys and 0 girls. We will treat each family as one experiment in
which we have eight trials. For each trial there are two possible outcomes: a boy or
a girl. We then repeat the experiment with each family in our sample.
56 Distributions
Table 5.2 The number of families, fi' having eight children of whom bi are boys. Pi is the binomial
probability of having bi boys. F, the total number of families, is equal to 53680 and FPi is the
expeded number of families of eight children with bi boys

Number of Number of Binomial Expected number

boys, bi families, fi probability, Pi of families, FPi

0 215 q8 165
1 1485 8pq7 1401
2 5331 28p2 q6 5202
3 10649 56p3qS 11 034
4 14959 70p4q4 14627
5 11 929 56pSq3 12411
6 6678 28p6 q2 6581
7 2092 8p7q 1994
8 342 p8 264

In our binomial formula (Equation 5.2), n is equal to 8. For each sample we then
have nine possible outcomes: 0, 1, ... ,8 boys and 8, 7, ... ,0 girls. To determine the
expected numbers of families with 0, 1, 2, ... , boys we first need to know p, the
probability of having a boy. The total number of boys is

B= 0 x 215 + 1 x 1485... 8 x 342 = 221023, 5.4

and in the same way we find that the total number df girls is
G = 208417. 5.5

The proportion of boys, which gives an estimate of p, the probability of having

a boy, is then
p = 221023/(221023 + 208417) = 0.5147, 5.6

so that q, the probability of having a girl, is

q = 1 - 0.5147 = 0.4853. 5.7

The first thing we notice is that the probability of having a boy is slightly
greater than the probability of having a girl. But is this just due to chance or is there
a genetic propensity to produce slightly more boys than girls? We can work out the
various parameters of the distribution. Let us write bi for the number of boys in a
family (column 1 in Table 5.2)and Ii for the number of families (column 2 in Table 5.2)
with bi boys. Then, the total number of families is

F= Lli = 53680. 5.8

i

(As a check, eight times the number of families must give the total number of children:
8 x 53680 = 429440 = 221023 + 208417.) The mean number of boys is then (using
Equation 4.2)
m= L biPi = L bJJF
i i
= 4.118. 5.9
 The binomial distribution 57
The variance of the number of boys is (using Equation 4.9)

V = L (hi -
i
m)2 f/(F - 1) = 2.067, 5.10

so that the population standard deviation of the number of boys is J

2.067 = 1.438.
The standard deviation of the mean number of boys is (Equation 4.18)

sm = sl JF = 0.00621. 5.11

Let us now see if the slight preponderance of boys is. significant or if it is simply
a result of the inherent variation in the data. If the probability of having a boy was
equal to the probability of having a girl. the mean number of boys per family should
be 4 whereas the observed value is 4.118, so that the observed and expected values
differ by 0.118. Since the standard deviation of the mean is 0.0062, the observed
and expected values differ by 19 times the standard deviation, the amount of spread
we expect, and the difference is almost certainly real and not a matter of chance.
We can now consider the question that we began with: does the fact that 557
families have all boys or all girls indicate that some people have a propensity to
produce children of the same sex or can it be put down to the fact that with so
many families in our sample we would expect about this many to have all boys or
all girls 7
To do this we use the binomial probabilities and the values we have calculated
for p and q to work out how many families we expect to have 0, 1,2... 8 boys
(assuming that the assumptions of the binomial distribution hold), as shown in column
4 of Table 5.2. The agreement is reasonably good although the observed numbers
of families are about 30% higher than the expected numbers at the ends of the
distribution. If there is a genetic propensity to produce all boys or all girls, it must
be rather slight and we still need to decide how big the difference should be before
we consider that it is significant.
For the binomial distribution, it is possible to derive expressions for each of the
parameters in terms of p, q and n, the number of trials. The mean, jL, is

jL = pn. 5.12

Since we don't know the 'true' probability of having a boy, we have to use our
estimated probability of having a boy (Equation 5.6) to estimate the mean and we
have 0.5147 x 8 = 4.118, the same value as before (Equation 5.9).
The variance is more interesting; we can show that for a binomial distribution the
variance may be written in terms of p, q, and n as (Bulmer, 1979, p. 90)

V = pqn, 5.13

which for our families of eight children gives 4.117 x 0.4853 = 1.998, very close to
the value of 2.067 estimated directly from the data using the squared deviations from
the mean.
58 Distributions
We can do the same for the skewness and the kurtosis. We can calculate the
skewness directly from the data as

5={ t (hi - m)3 f/F }/V312 = - 0.016, 5.14

and the kurtosis as

K={ t(b i - m)4 f/F }/V2 = 2.787. 5.15

But we can also show that for a binomial distribution the skewness is (Bulmer, 1979,
p.90)

5 = (q - p)/.;;;pq, 5.16

which for our families of eight children gives - 0.021, and the kurtosis is (Bulmer,
1979, p. 90)

K =3 + (1- 6pq)/npq 5.17

which for our families of eight children gives 2.75.We see that the variance, skewness
and kurtosis estimated from p, q, and n agree well with the values estimated directly
from the data, indicating that the data are close to a binomial distribution. The
expressions for the mean, variance, skewness and kurtosis of the binomial distribution
are summarized in Table 5.3.
We see that we are able to estimate the parameters of the distribution in two
ways: either directly from the data using the probability that a family has 0, 1, 2... 8
boys, or by using the data to estimate p, the probability of having a boy, from which
q follows immediately, and n, the number of children in each family, in this case 8.
For a given number of trials, the binomial distribution has only one free parameter,
p, since q is always 1- p, and we use this in calculating the mean. However, the
expressions in Table 5.3 enable us to calculate the other parameters once we know p
and n and if the underlying distribution is indeed binomial we should get very similar
values whichever way we calculate these parameters.
We can also use the expressions given in Table 5.3 to estimate the standard
deviations of the numbers of children in each category separately to help us to
decide if the differences between the observed and expected numbers are significant
or not. To do this, let us look at the problem in a different way and consider each

Table 5.3 The mean, variance, skewness and kurtosis for the binomial
distribution in terms of the number of trials, n, the probability of
success, p, and the probability of failure, q

m=pn
V = pqn
5 = (q - p)/ .J-npq
K = 3 + (1 - 6pq)/npq
 The binomial distribution 59
family as a trial (rather than each birth in each family) and consider only two outcomes:
o boys and 8 girls will be a success, all other combinations will be a failure. The
number of trials is now equal to the number of families, which is 53 680, and the
number of successes is the number of families having 0 boys and 8 girls. Then

p = 215/53680 = 0.0040, 5.18

q = 1 - 0.00401 = 0.9960, 5.19

n = F= 53680. 5.20

Now we have only a single repeat of this 'experiment' so we cannot deduce

the variance, skewness or kurtosis directly from the data. However, the expressions
given in Table 5.3 still apply, so that the variance is

V = pqn = 214, 5.21

and the standard deviation is

5 = fi = 14.6. 5.22

Since the observed number of families is 215 and the expected number of families
is 165 (Table 5.2), the observed and expected numbers differ by 50, which is
50/14.6 = 3.4 standard deviations, so that it appears as though the difference might
be significant. We could of course repeat this calculation for each combination of
boys and girls and hope to have a more sensitive test. We will do this in section 6.2.
Note also that the variance of the number of families having eight boys, 214, is
very dose to the number of families having eight boys, 215, which is our estimate
of the mean number of families having eight boys. Looking at Table 5.3 we see
immediately that since, in this case, p is very small, q is very dose to 1 and the mean
and the variance are both equal to pn.

5.1.2 Mendel's peas

As another application of the binomial distribution, consider again Mendel's study
on peas. In one experiment (Table 2.2),he examined 1064 plants in the F2 generation,
of which 787 were tall and 277 were short. According to Mendel's theory, the
probability of obtaining a tall plant is 0.75, and there are only two possible outcomes,
tall and short. If we assume that the events are independent, so that what happens
to one plant does not influence what happens to another, the distribution of plants
should follow a binomial distribution. We could, of course, work out the precise
probability of obtaining 787 tall plants from a total of 1064 plants with a probability
of 0.75 that each plant is tall, but most of the information we need is contained in
the mean and the variance. Since the mean of a binomial distribution is pn, the
expected value of the mean is 1064 x 0.75 = 798, so that the difference between
the observed and expected number of tall plants is 798 - 787 = 11. Furthermore,
the variance is pqn = 0.75 x 0.25 X 1064 = 199.5, so that the standard deviation is
J199.5 = 14. We see that the difference between the observed and expected number
60 Distributions
of tall plants is less than the expected spread in the number of tall plants, and we
can conclude that Mendel's data agree with the predictions of the theory within the
intrinsic variation of the experiment.

5.2 THE POISSON DISTRIBUTION

When the number of trials is relatively small the binomial distribution is easily
calculated, but for the study of families with eight children, calculating the probabilities
for each po~ sible outcome was quite tedious. In an experiment in which the number
of trials runs to hundreds or even thousands, it is impractical even to think about
calculating the binomial frequencies exactly. However, it turns out that if the
probability of success in each trial is very small but the number of trials is very large,
the binomial distribution tends to a simpler form discovered in 1837 by the French
mathematician Simeon-Denis Poisson.
For the binomial distribution, the expected number of successes in n trials is pn
where p is the probability of success. Now suppose that we have a very large number
of trials so that n ~ 00 . (We read '~' as 'tends to'. Since n cannot actually be infinite
the idea is that n ~ 00 means that n is very much bigger than anything else in the
problem. Similarly, p ~ 0 means that p is very much smaller than anything else in
the problem.) To ensure that the number of successes remains finite, we must have
a very small probability of success so that we let p ~ 0 in such a way that pn remains
finite. These conditions define the Poisson distribution, which we get whenever the
number of trials is very large but the probability of success on each trial is
very small.
Whereas the binomial distribution is specified by two parameters, the number of
trials, n, and the probability of success on each trial, p, the Poisson distribution is
specified by a single parameter, J1. = pn, the mean number of successes. The probability
of k successes is (Bulmer, 1979, p. 90)

P(k)=e-/lJ1.klk!k=0,1,2"·, 5.23

so that provided we know the mean number of successes, we can calculate the
probability of getting any given number of successes.
Figure 5.2 shows the Poisson distribution for values of the mean, J1., equal to 1, 5, 9
and 50. The distribution in Fig. 5.2a is similar to the binomial distribution shown in
Fig. 5.1a because q is small (0.1) and in both cases the mean is 1. As the mean increases,
the distribution shifts to higher values and the width of the distribution increases as
the square root of the mean while the ratio of the width to the mean decreases as
the square root of the mean.

5.2.1 Prussian army officers

The horses in the Prussian army (von Bortkewihch, 1898), which we discussed in
section 4.3.1, must have had very many occasions on which they might have kicked
their officers to death so that the number of trials in that experiment was indeed
very large and we can let n ~ 00. The probability that anyone officer was kicked
 The Poisson distribution 61

Figure 5.2 The Poisson distribution for (a) m = 1; (b) m = 5; (c) m = 9; (d) m = 50.

to death each time he happened to walk behind his horse was, fortunately for the
Prussian army, very small and we can let p - O.
We don't know the actual number of trials, which would be the number of times
an officer in each corps walked behind his horse in anyone year, but for the Poisson
distribution we need to know only the mean number of successes. In section 4.3.1
we showed that, m, the mean number of deaths per army corps per year, was 0.61.
Assuming that the data follow a Poisson distribution, the probability that no deaths
occur in anyone corps in anyone year is (Equation 5.23)

P(O)= e-O.610.61o/0! = e-O.61 = 0.5435. 5.24

Similarly, the probability that, say, three deaths occur in anyone corps in anyone
year is

P(3) = e-O.610.613/3! = 0.0205. 5.25

To get the frequency with which we expect no deaths, one death, and so on, we
need to multiply each probability by 200, the total number of events (20 years x 10
corps). Table 5.4 gives the observed and expected frequencies of deaths and the two
distributions agree very well indeed.
Since the Poisson distribution is the limiting form of the binomial distribution as
p _ 0 and q -1, we can use the binomial results (Table 5.3) to show that for the
Poisson distribution the variance is equal to the mean, J.l, the skewness is equal to
1/jP, and the kurtosis is equal to 3 + 1/J.l, as summarized in Table 5.5
62 Distributions
Table 5.4 The frequency with which between 0 and 5 officers were
kicked to death by their horses in each year from 1875 to 1894 in each
of ten Prussian army corps

Number Observed Expected

of deaths frequency frequency

0 109 108.7
1 65 66.3
2 22 20.2
3 3 4.1
4 1 0.6
5 0 0.1

Table 5.5 The variance, skewness and kurtosis for the Poisson
distribution

V=J1.
5= lIh
K = 3 + 1/J1.

We have already calculated the parameters of the distribution for the Prussian army
data. The mean was found to be 0.61 deaths/corps/year, so that on average rather
less than one officer was kicked to death by his horse in each corps in each year.
The variance we estimated from the data as 0.611 death2/corps2/year2, so that the
variance is approximately equal to the mean as expected for a Poisson distribution.
The skewness, estimated directly from the data, was 1.26, a little less than 1.28, the
value calculated from the mean, and the kurtosis estimated directly from the data
was 4.38, a little less than 4.64, the value calculated from the mean. The agreement
between the values of the variance, skewness and kurtosis calculated directly from
the data and from the expressions in Table 5.5 give us reason to believe that the
data do follow a Poisson distribution.

5.2..2. The polio vaccine

Let us consider the data on the polio vaccine given in section 1.1.3. Of the 200745
children vaccinated, 33 developed the disease. Since very many children were
inoculated, n is very large, but since few developed the disease, p is very small. We
therefore assume that the data follow a Poisson distribution and that the variance is
equal to the mean number of children who develop the disease. Since 33 vaccinated
children developed polio, the standard deviation is .j33 = 5.7. Similarly, 115
unvaccinated children developed the disease so that the standard deviation in this
number is .j1i5 = 10.7.
Now the number of children in the two categories are not the same but if we calcu-
late the number per 100000 who contracted the disease we get 16.4 ± 2.86 for those
who were vaccirtated and 57.1 ± 5.33 for those who were not vaccinated. The diffe-
 The normal distribution 63
rence between the two means, 40.7, is much greater than the expected spread in either
measurement, 2.86 and 5.33, and we conclude that the vaccine does indeed bring
about a significant degree of protection. If we want to be more precise we can apply
our formula for the variance of a difference (Appendix, section 4.5.2), which gives
J
2.862 + 5.332 = 36.59, so that the standard deviation of the difference is 36.59 =
6.04 and the observed difference is about seven times the standard deviation.
We can now see what would have happened if we had used only one tenth as
many children in the trial. We would then have vaccinated 20000 children and we
would expect to get about 33/10 or 3.3 cases of polio. Placebo would have been
given to 20000 children and we would expect to get 115/10 or 11.5 cases of polio.
The expected value of the difference would then be 8.2, but the variance of the
difference would be 3.3 + 11.5 = 14.8, giving a standard deviation of J14.8 = 3.9,
so that the observed difference is only twice the standard deviation. The intrinsic
variation in the data may well have concealed the difference between the vaccinated
and the unvaccinated children.
The fact that the variance of a Poisson distribution is equal to the mean is important
in the theory of sampling because it implies that for random sampling from large
populations, for which the probability that any particular individual is trapped or
caught is small, the standard deviation of the number sampled is given by the square
root of the number sampled.

5.3 THE NORMAL DISTRIBUTION

The most important distribution in biology, if not in all of statistics, is the normal
distribution. We have already touched on it in section 3.2.2, where we discussed the
length of the wings of houseflies. The normal distribution was first investigated by
Karl Gauss in 1809 in his study of the theory of errors in astronomical investigations
(Bulmer, 1979, p. 108). The theory of errors is dominated by the normal distribution
because the sum of any large number of random variables is normally distributed, a
result known as the Central Limit Theorem (Bulmer, 1979, p. 115). There are, as
always, some provisos: the contributing variables should be independent of each
other and their variances should be well-behaved. Since it is often the case in nature
that variables, such as the height of a person, are the result of many random effects,
each of which contributes a little to the total, we often find that biological observables
are normally distributed.
We have already seen (Equations 3.9 and 3.10) that the probability density function,
or p.d.£., for a normal distribution with mean J.l and standard deviation (J is (Bulmer,
1979, p 108)

P(x) = _l_e - {r- 1')2/2(12. 5.26

(1j"£c

The normal distribution is plotted in Fig. 5.3 for various values of the mean and
standard deviation. As the mean J.l, increases, the distribution shifts to the right and
64 Distributions

Figure 5.3 The normal distribution for (a) p, = 0, (J = 1; (b) p, = 1. (J = 1; (c) J1. = 0, (J = 2;
(d) J1. = 2, (J = 2. The distribution is symmetrical about the mean, p" and the width is proportional
to the standard deviation, (J.

as the standard deviation, U, increases, the distribution becomes wider. The normaliza-
tion constant, lIuj2n, is chosen to ensure that the area under the p.d.£. from - 00
to 00 is 1.
The normal distribution is symmetrical about the mean since we obtain the same
answer if we set x equal to Il + b or Il- b in Equation 5.26; if we set Il = 0, it will
be centred at the origin. The standard deviation determines the width of the distribu-
tion and when u is 1 and the mean is zero we refer to the standard normal distri-
bution, indicated by the letter Z, so that

Z(x) = _l_e-r'-12 5.27

,fin
The standard normal distribution is tabulated in many books on statistics. We can
use it to calculate P(x) for any normal distribution with mean Il and variance u2 by
noting that P(x) = (l/u)Z[(x -Il)/U). Since the normal distribution is symmetrical, its
skewness is zero. The kurtosis of the normal distribution is 3 and we take this for
the standard value of kurtosis.

5.3.1 Normal approximations

The Central Limit Theorem tells us that the sum of a large number of independently
distributed random numbers is nearly always normally distributed provided the
 The normal distribution 65

Figure 5.4 The distribution of numbers obtained when a coin was thrown 10000 times and
the numbers were added together in groups of ten counting 1 for heads and 0 for tails. The
curve is the normal distribution that best fits the data.

individual distributions satisfy some reasonable conditions (Bulmer, 1979, p. 115). To

illustrate this, I spun a coin (simulated on a computer) 10000 times, counting 1 for
heads and 0 for tails, and then added the numbers together in groups of ten. Figure 5.4
shows the distribution of the resulting numbers, with the normal distribution that best
fits the data, and the agreement is good.
The mean and the standard deviation, calculated from the simulation, were 4.95
and 1.57. Using Equation 4.3, the mean for one throw of the coin is

J.l = 1x P(H) +0 x P(T) = 0.5, 5.28

and since the mean of a sum is the sum of the means, the expected mean for ten
throws is 5, so that the value from the simulation (4.95) is close to the expected
value. Using Equation 4.9, the variance for one throw of the coin is

V = (1 - a.sf x P(H) + (0 - 0.5)2 X P(T) = 0.25, 5.29

and since the throws are independent, the variance of ten throws is 2.5 and the
standard deviation is .)2.5 = 1.58, so that the value from the simulation (1.57) is
again close to the expected value. The skewness and the kurtosis, calculated directly
from the data, are - 0.0062 and 2.909, close to the expected values of 0 and 3.

5.3.2 Binomial and Poisson distributions

Whenever there are two possible outcomes, success and failure, and the probability
of success or failure is constant and independent from one trial to the next, the
binomial distribution gives us the distribution of successes for any given number of
66 Distributions
trials. Furthermore, if the number of trials is very large and the probability of success
on each is very small, the binomial distribution may be approximated by the Poisson
distribution.
Both the binomial and the Poisson distributions are discrete so that they give us
the probability of observing 0, 1, 2, ... events. The normal distribution, on the other
hand, is continuous and the observable may take on a range of continuous values,
as in the case of the wing lengths of our houseflies. However, it is possible to show
that provided the variance of the distribution is not too small, the binomial and
Poisson distributions can both be approximated by a normal distribution. Referring
back to Figs 5.1d and 5.2d, both the binomial and the Poisson distribution look like
the normal distribution when the expected number of successes and the variance are
sufficiently large. In general, the approximation is quite good provided the variance,
which is pqn for the binomial and is equal to the mean for the Poisson distribution,
is greater than about 9.
For the Prussian army data of Table 5.4, pqn = 0.61, and the distribution is quite
skew, so we would not expect the distribution to approximate closely to a normal
distribution. For our data on the numbers of boys and girls in families of eight in
Saxony, we have pqn = 2, but the probability of success and failure are both close
to 0.5 so that the distribution is almost symmetrical and the skewness is small ( - 0.02).
Furthermore, the kurtosis (2.8) is close to 3, so let us see how well we can approximate
the data in Table 5.2 using a normal distribution.
Equation 5.26 gives the p.dJ for the normal distribution. Since the normalization
constant is chosen so that the total area under the curve is equal to 1, we must
multiply by 53680, the total number of families, to obtain the frequency distribution
N(x). We therefore evaluate

N(x)=_N_e-(X-m)2/2s2 5.30
sj21r,

Table 5.6 The observed and expected numbers of families of eight children with
between 0 and 8 boys. The third column gives the expected numbers after fitting the
data to a binomial distribution and the fourth column the expected numbers after fitting
the data to a normal distribution

No. of Observed no. Expected Expected

boys of families Binomial Normal

0 215 165 247

1 1485 1401 1420
2 5331 5202 5038
3 10649 11 034 11 015
4 14959 14627 14846
5 11 929 12411 12335
6 6678 6581 6318
7 2092 1994 1995
8 342 264 388
 The X2, t and F distributions 67
with a mean, m, of 4.118 and a standard deviation, s, of 1.438. The results are given
in Table 5.6, from which we see that the normal approximation is too high in the
tails (the binomial is too low), but the fit over the peaks is somewhat better and
the overall fit is about as good as that obtained using the binomial distribution. The
difference between the observed numbers and the numbers predicted using the binomial
distribution may be due in part to random effects in the way the data were collected
and recorded, which would tend to make the distribution look more like a normal
distribution.

5.4 THE X2, t AND F DISTRIBUTIONS

There are three more distributions that are widely used in biostatistics and all may
be derived from the normal distribution. In this section I will introduce them briefly
and we will consider their properties in more detail as we use them.

5.4.1 The X2 distribution

Much of our work will be concerned with analysing the variances of the numbers
we measure. When we test a new trap to see how well it catches tsetse flies, we
always expect some variation in the number of flies caught. We will then want to
know if the variation is simply due to the stochastic nature of the sampling process
or if it is due to a variable of interest such as the colour of the trap, the site in which
it is placed or even the day on which it is used. We have already seen that when
we calculate the variance, we add up the squares of the deviations of each measurement
about the mean. We generally assume that each of our measurements is normally
distributed about the mean, and we would like to know how the sum of the squares
of such normally distributed numbers will itself vary so that we can determine the
distribution of our variances.
Suppose we let x be the sum of the squares of a number of independent, standard,
normal variates, Zi' so that

x=zi+z;+ ... +z; 5.31

Then x is said to belong to a chi-square distribution with f degrees of freedom, which

is written X;. Figure 5.5 shows the X2 distribution for 1, 4, 16 and 64 degrees ot
freedom. You see that the distribution starts off being very skew but looks more
like a normal distribution as the number of degrees of freedom increases. The mean
of the X2 distribution is equal to the number of degrees of freedom, so that the
distribution shifts to higher values as the number of degrees of freedom increases.
The parameters of the X2 distribution are given in Table 5.7, from which it is clear
that as the number of degrees of freedom increases, the skewness tends to 0 and the
kurtosis tends to the normal value of 3, so that the X2 distribution tends to the
normal distribution. Of course, this is simply another example of the Central Limit
Theorem, since we have defined the X2 distribution as the sum of a number of
 68 Distributions

Figure 5.5 The X2 distribution for (a) 1, (b) 4, (c) 16 and (d) 64 degrees of freedom.

Table 5.7 Parameters of the X2 distribution

mean f
variance 21
skewness fill
kurtosis 3 + 1211

independent random variables and if we have enough of them their sum must follow
a normal distribution.

5.4.2 The t distribution

We often want to compare the mean value of a measurement with its standard devia-
tion: if the mean is less .than the standard deviation, it is clear that it does not differ
significantly from zero, while if it is much greater than the standard deviation, it is
clear that it does differ significantly from zero. For example, in section 5.1.2, in our
examination of Mendel's data on peas, we compared the difference between the
observed and expected number of tall plants (11) with the standard deviation in the
observed number (14). Because the difference was less than the standard deviation,
we concluded that the difference was not significant and that the data supported
Mendel's hypothesis. In section 5.2.2 we found that the probability of a child
developing polio was reduced if the child had received the Salk vaccine. To convince
 The X2, t and F distributions 69
ourselves that the effect was significant, we calculated the difference between the
proportion of vaccinated and the proportion of unvaccinated children who developed
polio (40.7 per 100000) and then calculated the standard deviation of the difference
(6.04 per 100000). Because the difference was seven times the standard deviation,
we concluded that the difference was real and the vaccine efficacious.
Let us restate this: if the ratio of the absolute value of the mean to the standard
deviation is much less than I, we know that the mean does not differ significantly
from zero while, if the ratio is much greater than I, we know that the mean does
differ significantly from zero. If we can determine the distribution followed by the
ratio of the mean to the standard deviation, then, as we will see in Chapter 7, we
will be able to put precise limits on such statements and determine the probability
that our observed value of the mean is simply due to chance as opposed to a real effect.
The distribution of the ratio of the mean to its standard deviation was first
investigated by William Gossett (Student, 1908) who wrote under the pen name
'Student'. Since the mean is normally distributed and the variance follows a X2 distribu-
tion, we are interested in the ratio of a normally distributed variable to the square
root of a variable following a X2 distribution; this is called Student's t distribution.
Formally, if

x = z/JYij 5.32

and z follows a normal distribution while y follows a X2 distribution with f degrees

of freedom, we say that x follows a t distribution with f degrees of freedom. At
this point we will extend our notation further and use the symbol to mean 'follows'.
IOV

Then we can summarize this result as follows:

If z '" N(O, 1) and y'" xf,

then x = z/ J y/ f '" t
f 5.33

Table 5.8 gives the parameters of Student's t distribution, from which we see that
the distribution is symmetrical about the origin and that the kurtosis tends to the
normal value of 3 as the number of degrees of freedom increases. Plotting the t distri-
bution for various numbers of degrees of freedom (Fig. 5.6), we see that the t distribu-
tion looks very much like the normal distribution except that it has long tails, which
become less pronounced as the number of degrees of freedom increases.

Table 5.8 Parameters of the t distribution. f is the number of degrees

of freedom. For f ~ 2 the variance is infinite and for f ~ 4 the kurtosis
of infinite

mean 0
variance fl(f - 2) f>2
skewness 0
kurtosis 3 +6/(f - 4) f>4
70 Distributions

Figure 5.6 The t distribution for 1, 2, 4 and 100 degrees of freedom. As the number of
degrees of freedom increases, the peak value increases and the tails become shorter.

Figure 5.7 The F distribution with (a) 4 and 4; (b) 4 and 16; (c) 16 and 4 and (d) 16 and 16
degrees of freedom.
 Cumulative distribution juncfiorzs 71

Table 5.9 The mean and the variance of the F distribution for large
numbers of degrees of freedom. The expressions for the skewness and
kurtosis are more complicated

5.4.3 The F distribution

We will use the t distribution to test a mean or to compare two means. Sometimes,
however, we will want to compare two variances. To see if two variances differ
significantly, we can calculate their ratio and then see if this ratio is greater or less
than 1. Since we know that a variance follows a X2 distribution, we will want to
know the distribution of the ratio of two numbers, each of which follows a X2 distri-
bution; this is called the F distribution, after the English statistician Ronald Fisher.
Then if

x = (y] II] )/(Y2112) 5.34

and YI and Y2 independently follow X2 distribution with II and 12 degrees of freedom,

respectively, x follows an F distribution with II and 12 degrees of freedom. We
can summarize this as follows.

If YI '" X~ and Y2 '" xt, then x = (yIIII)/(Y2112) '" F/,.!z 5.35

The F distribution is illustrated in Fig. 5.7 for various combinations of the two
degrees of freedom that define it. The mean ofax2 distribution is close to the number
of degrees of freedom so that if we divide by the number of degrees of freedom
the mean will be close to 1. The expressions for the parameters of the F distribution
are rather complicated: Table 5.9 gives the parameters for large values of II and 12
only.

5.5 CUMULATIVE DISTRIBUTION FUNCTIONS

For each of the distributions discussed in this chapter we can also determine the
corresponding c.d.£. For discrete distributions we simply sum all the probabilities up
to each observed value. For the families of eight children, given in Table 5.1, the
estimated c.d.£. is
j

Cj = I !;IF, 5.36
i= 0

where F is the total number of families and Ii is the number of families having i
boys. This is plotted in Fig. 5.8a. We can calculate the c.d.£.for the Poisson distribution
of Prussian army officers (Table 5.4) in the same way and this is plotted in Fig. 5.8b.
The importance of the c.d.£. is that we can use it to determine the probability that
an observable lies in a given range directly from graphs or tables. For example, from
72 Distributions

Figure 5.8 (a) The cumulative distribution function of the number of boys in families of eight
children. (b) The cumulative distribution function of the number of Prussian army officers
kicked to death by their horses.

Fig. 5.8 we see immediately that the probability that a family of eight children has
four or fewer boys is 0.61. Similarly, the probability that none or one officer was
kicked to death by their horses in anyone year is 0.87.
For continuous distributions, the c.d.f. is the area under the p.d.f. up to each point
(as we saw in section 3.2.3), so that for the normal distribution the value of the c.dJ.
at x is the area under the p.d.f. from - 00 to x. This is plotted for the standard normal
distribution in Fig. 5.9. If we want to know the probability that a number falls in a

Figure 5.9 The cumulative distribution function for the standard normal distribution.
 Tables of distribution functions 73
Table 5.10 Selected points from the c.dJ. of the standard normal distribution
shown in Figure 5.9

x C(x) x C(x)
-3.29 0.0005 0.00 0.50
-3.09 0.001 1.28 0.90
-2.58 0.005 1.64 0.95
-2.33 0.010 1.96 0.975
- 1.96 0.025 2.33 0.990
- 1.64 0.05 2.58 0.995
- 1.28 0.10 3.09 0.999
0.00 0.50 3.29 0.9995

certain range, say Xl to X2' then for both discrete and continuous distributions

P(xI < x < x2) = Area (- 00 to x2) - Area ( - 00 to Xl) = C(x2)- C(xl), 5.37

so that we can easily use the c.d.f. to find the probability that an observable falls
into any given range.
Suppose we select a number from a normal distribution. The c.d.f. can be used to
find the probability that the number we pick is less than 0, greater than 2, between
I and 3 or falls into any range we choose. For example, Table 5.10 gives points
from the c.d.f. for a standard normal distribution. If a number is drawn from a standard
normal distribution, the probability that it is less than 0 is 0.5. If a number is drawn
from any normal distribution the probability that it is less than the mean is 0.5.
Similarly, the probability that a number drawn from a standard normal distribution
is less than 1.96 is 0.975 (Table 5.10), so that the probability that a number chosen
from any normal distribution is less than J.l + 1.960' is 0.975, or 97.5%, and so on.
It also follows that the probability that a number drawn from a normal distribution
is greater than J.l + 1.960' is 0.025, or 2.5%. Finally, we see that the probability that
a number drawn from a normal distribution is greater than J.l + 1.960' or less than
J.l- 1.960' is 0.05, or 5%.
Tables of c.d.f.s allow us to determine the probability that a number chosen from
the appropriate distribution lies in any given range and we will use c.dJ.s extensively
in this way. And, unlike the p.dJ., the c.d.f. is always dimensionless.

5.6 TABLES OF DISTRIBUTION FUNCTIONS

At the end of this book you will find tables of the c.d.f.s of the t, X2 and F distribu-
tions and you need to be able to use them. The row labelled P in the tables
gives the value of the c.d.f., the first column gives the number of degrees of freedom
and the values in the tables, which we call critical values, are the values of the
statistics that we calculate. For example, the number in Table 10.1 corresponding to
P = 0.99 with 6 degrees of freedom is 3.14. This tells us that if we measure a variable
that follows the t distribution with 6 degrees of freedom, there is a 99% probability
 74 Distributions
that it will be less than 3.14. Similarly, Table 10.2 tells us that if we measure a
variable that follows the X2 distribution with 11 degrees of freedom, there is a
95% probability that it will be less than 19.68.
Tables of the F distribution are more extensive than the others because the F
distribution has 2 separate degrees of freedom. Table 10.3, for example, gives the
critical values for P = 0.95. For example, if we measure a variable that follows an F
distribution with 6 and 12 degrees of freedom, there is a 95% probability that it will
be less than 3.00. In the same way, Table 10.4 shows that if we measure a variable
that follows an F distribution with 6 and 12 degrees of freedom there is a 99%
probability that it will be less than 4.82.
You will notice that we do not give tables for the normal distribution. This is
because the t distribution with an infinite number of degrees of freedom is identical
to the normal distribution. Referring to Table 10.1, we see that if we choose a number
from a normal distribution, there is a 99% probability that it will be less than 2.33.

5.7 SUMMARY

The binomial, Poisson, normal, X2, t, and F distributions provide the basis for most
of what you need to know and we can summarize the conditions under which they
hold as follows.

• Binomial: two outcomes, success or failure, probability of success is constant from

one trial to the next and successive trials are independent.
• Poisson: limiting case of the binomial distribution when the probability of success
in anyone trial is very small but the number of trials is very large so that the mean
number of successes remains finite.
• Normal: sum of a large number of independent variates.
• X2: sum of squares of normal variates.
• t: ratio of a normal variate to the square root ofax2 variate.
• F: ratio of two X2 variates.

There are other distributions of importance in biology. In particular, you may

encounter the negative binomial distribution, which is often used to describe clumped
distributions Geffers, 1978). For example, if a few cows in a herd have many ticks
while most cows have few ticks, the distribution of the ticks on the cows can probably
be described by a negative binomial distribution and one of its parameters can be
used as a measure of the degree of clumping. The Poisson distribution turns out to
be a particular case of the negative binomial distribution.
For the distributions we have discussed in detail and others that you will meet,
it is important to be clear about the conditions under which they hold so that you
can relate them directly to the biology of the problem that you are considering. In
the next chapter we will use these distributions to see how we can make more rigorous
and precise statements about differences between the results of our measurements and
the predictions of our theories.
 &e~ 75
5.8 EXERCISES

1. Expand (p + q)2 and (p + q)J and confirm that the successive terms give the
probabilities listed in Table 5.1. Use Equation 5.2 to obtain the same result.

2. Use Equation 5.2 to verify the values for the expected numbers of families having
four boys in Table 5.2.

3. Use Equation 5.30 to verify the values for the expected numbers of families
having four boys in Table 5.6, assuming that the data are normally distributed with
a mean of 4.118 and a standard deviation of 1.438.

4. Table 5.11, taken from Greenwood and Yule (1920), is an accident record for 647
women working in a munitions plant and gives the number of women having no
accidents, one accident and so on. Calculate the mean number of accidents per woman.
If the probability that any woman has an accident is small and is the same for all
women, the numbers should follow a Poisson distribution. Calculate the expected
frequencies and compare them with the observed frequencies.
Greenwood and Yule conclude that the number of women having three or more
accidents is greater than expected and that there must have been two subgroups,
one more accident prone than the other. Do you agree with their conclusions?

5. Larvae of the Azuki bean weevil (Callosobruchus chinensis) enter into beans (Phaseo/us
radiatus), feed and pupate inside them and then emerge through a hole (Utida, 1943).
The number of holes per bean is therefore a good measure of the number of adults
that have emerged. If the probability that anyone bean is parasitized is small and
does not depend on whether a bean has already been parasitized, the frequency
distribution of the number of beans having 0, I, 2, ...holes should follow a Poisson
distribution. Use the data in Table 5.12 to calculate the expected numbers of beans
having a given number of holes and compare these with the observed frequencies.
The authors conclude that there are fewer beans with two or three holes than one
would expect because weevils can identify and avoid beans that are already parasitized.
Do you agree with this conclusion?

Table 5.11 The number of women

having 1, 2, 3, 4 and 5 accidents in a
five-week period

Number Number
of accidents of women

0 447
1 132
2 42
3 21
4 3
5 2
 76 Distributions
Table 5.12 The number of beans
having 1, 2 and 3 holes

6. Student (1907) counted the number of yeast cells in each of 400 squares of a
haemocytometer with the results given in Table 5.13. Assuming that the number of
cells follows a Poisson distribution, calculate the expected number of squares having
any given number of cells in them and compare the observed and expected values.
(This problem is analysed by Sokal and Rohlf, 1981, p. 84.)

7. Use the c.d.£. for the standard normal distribution (Fig. 5.9) to find the probability
that a number chosen from a normal distribution with mean 1 and standard deviation
2 is (a) greater than 3; (b) less than 2; (c) lies between 2 and 3.

8. Use the tables at the end of this book to determine the probability that

(i) a number chosen from a t distribution with 1 degree of freedom is (a) greater
~
than 6.31; (b) less than 6.31; (c) lies between ~ 6.31 and + 6.31;
2
(ii) a number chosen from a X distribution with 6 degrees of freedom is (a) less
than 1.64; (b) greater than 12.59;
(iii) a number chosen from an F distribution (a) with 3 and 4 degrees of freedom
is greater than 6.59; (b) with 7 and 10 degrees of freedom is less than 6.06.

Table 5.13 The number of squares

having 0, 1, ... cells in them
Number of Number of
cells/square squares
0 75
1 103
2 121
3 54
4 30
5 13
6 2
7 1
8 0
9 1
6

Testing hypotheses

Science is the attempt to make the chaotic diversity of our sense experience correspond
to a logically uniform system of thought. In this system single experiences must be
correlated with the theoretic structure in such a way that the resulting coordination is
unique and convincing.

The sense experiences are the given subject-matter. But the theory that shall interpret
them is man-made. It is the result of an extremely laborious process of adaptation:
hypothetical, never completely final, always subject to question and doubt.

The scientific way of forming concepts differs from that which we use in our daily life,
not basically, but merely in the more precise definition of concepts and conclusions;
more painstaking and systematic choice of experimental material; and greater logical
economy. By this last we mean the effort to reduce all concepts and correlations to as
few as possible logically independent basic concepts and axioms.
A. Einstein (1950)

The relationship between theory and experiment lies at the heart of modem science
and has been the subject of intense debates in the history and philosophy of science.
The Greeks and the Hindus established many of the most important branches of
mathematics but did not develop a quantitative experimental science. The Romans
were excellent engineers and the Chinese made the most extraordinary discoveries
concerning the natural world but neither of them developed mathematical laws to
describe natural phenomena. Kepler made a vital contribution to the development
of modem science by showing that the orbits of the planets obeyed precise algebraic
laws and Galileo's great contribution to science lay in the application of mathematics
to the analysis of carefully conducted experiments. Indeed, Needham (1988) has said
that 'Modem [as opposed to ancient or mediaeval] science is the mathematization of
hypotheses about nature'" combined with rigorous experimentation:
Newton proposed the first scientific theory based solely on the formulation of a
mathematical law, and although his law made it possible to predict the motion of
the planets with great precision, many of his contemporaries rejected Newton's theory
because it did not provide a mechanistic model of planetary motion. Darwin acknow-
ledged that his theory of natural selection did not provide a mechanistic explanation
of evolution, but he appealed to Newton's theory of gravitation in his defence, saying
(Darwin, 1906, p. 657), 'It is no valid objection that science as yet throws no light
on the far higher problem of the essence or origin of life. Who can explain what is
78 Testing hypotheses
the essence of the attraction of gravity? No one now objects to following out the
results consequent on this unknown element of attraction, notwithstanding the fact
that Leibniz formerly accused Newton of introducing occult qualities and miracles into
philosophy'. So the debate continues. The 'idealists' claim primacy for theory, the
'materialists' claim primacy for experiment. Against the idealist position, we can
argue that if you walk into a wall its reality is self-evident; against the materialist
position, we can argue that the reason why the wall, viewed in a mirror, seems to
be in front of us and not behind us is that our minds have already adopted the theo-
retical position that light travels in straight lines, and this is why we are deceived
when we look in a mirror. Einstein (1978) said that 'The object of all science, whether
natural science or psychology, is to coordinate our experiences into a logical system.'
It is important to remember that on the one hand we have our experiments and
observations, on the other hand we have our theories and our deductions. The
relationship between them is of a dialectical nature so that each informs the other.
When we perform an experiment we do it to increase our understanding; when we
build a theory we hope that it will reflect and clarify the results of our experiments.
We do not perform our experiments in a 'theoretical vacuum'; there is always a theory
or hypothesis we are trying to test. When we say that we should approach science
with an open mind, we do not mean that we should have a blank mind, only that if
our assumptions and hypotheses do not appear to be supported by the evidence, we
should be willing and able to change our assumptions, invent new ones and maintain
a flexible attitude. Indeed, whenever we look in a mirror, we should be reminded
that even when we think that we are making no hypotheses, our subconscious is
doing it for us. In this chapter we will consider the testing of hypotheses and
statistical inference and see how these relate to statistics.

6.1 SIGNIFICANCE LEVELS AND THE POWER OF A TEST

Suppose we throw a coin 100 times and it lands heads on 55 occasions and tails on
45 occasions. How shall we decide if the coin is biased?
Let us begin by stating the problem carefully. We have an idea, I, that we want
to test

I: the coin is biased. 6.1

We carry out an experiment that we hope will provide evidence for or against I and
this gives us data, D
D: 55 heads and 45 tails. 6.2

In science, it is all too easy to read into data what we want to see. We therefore
proceed cautiously and make the null hypothesis

H: the coin is not biased. 6.3

We now use our knowledge of statistics to decide if D seem reasonable if H is true.

If D seem reasonable given H, we accept H and reject I (perhaps reluctantly). If D
 Significance levels and the power of a test 79
seem unreasonable given H, we reject H, provisionally accept I and look for an explana-
tion. To explain I we might, for example, examine the coin to see if it is bent. What
we need is a way to link the hypothesis, H, to the data, D. Making this link is what
biostatistics is all about. (We are discussing a simple problem in a formal way: when
we consider more complicated problems, proceeding formally should help.)
We have already worked out how to calculate the probability of getting 55 heads
and 45 tails in 100 throws of a coin using Equation 5.2 and the answer is
nl
P(551100) = . pkqn- k
k!(n - k)!

= (100!/45!551)0.555 0.545

= 6.145 X 1028 X 7.889 X 10-31 = 0.048. 6.4

so that this particular outcome is expected to occur only five times in every
100 experiments and all we can say is that it is rather improbable. However, even
the most probable outcome, which is 50 heads and 50 tails, only occurs eight times
in every 100 sets of 100 throws. Calculating the probability of getting the observed
result doesn't really help. If we think about this, it is obvious that if there are many
possible outcomes, even the most probable outcome will occur only very rarely;
looking at the probability that particular, individual events will occur is thus of little
value.
Let us try another approach. Even if our hypothesis is true and the coin is unbiased,
we would not expect to get precisely 50 heads and 50 tails, but rather to have results
which are close to 50 heads and 50 tails. We would surely agree that 50 heads and
50 tails is evidence in favour of the coin being unbiased while 100 heads and 0 tails
is evidence in favour of the coin being biased. We might then say that if we get
anywhere between 40 and 60 heads, we will regard the coin as unbiased, while if
we get less than 40 or more than 60 heads, we will regard it as biased. We could
calculate the probability of getting between 40 and 60 heads by adding the proba-
bilities of each outcome in this range as we did in Equation 6.4. That rapidly becomes
tedious. However, we know that the expected number of heads is pn = 0.5 x 100 = 50
and that the variance is pqn = 25 so that the standard deviation is 5. Since pqn is
sufficiently large, we can approximate the binomial distribution by a normal
distribution. With an expected mean of 50 and a standard deviation of 5, the probability
of getting between 40 and 60 heads is equal to the probability that a number chosen
from a normal distribution is within ± 2 standard deviations of the mean. Table 10.1
shows us that this probability is close to 95%. Since the probability of getting any
result is 1, the probability of getting less than 40 or more than 60 heads is 1 - 0.95 =
0.05. In other words, if we repeated this experiment many times, we would expect
to get between 40 and 60 heads in 95% of the trials and less than 40 or more than
60 heads in 5% of the trials.
This test looks reasonable and this is how we will proceed, so let us tie it down
carefully. What we have done is to define an acceptance range (40 to 60 heads) and
a rejection range (0 to 39 and 61 to 100 heads). If the data, D, fall within the
80 Testing hypotheses
acceptance range, we accept Hypothesis 6.3; if they fall in the rejection range, we
reject it. Nevertheless, it is as well to bear in mind Gould's (1986) dictum that 'In
science, "fact" can only mean "confirmed to such a degree that it would be perverse
to withhold provisional assent" " for even if the null hypothesis is true and the coin
is unbiased, we know that the outcome may still be in the rejection range, and we
would then reject the hypothesis even though it is true. We call this a Type I error.
With the acceptance range of 40 to 60 heads, the probability of a Type I error is
approximately 0.05 or 1 in 20. Since errors are undesirable, why don't we simply
increase the acceptance range and in this way reduce the probability of making a
Type I error? The trouble is that if we do this, the test becomes less sensitive.
Suppose that we go to the extreme and include all results between 0 and 100 in our
acceptance range. We will then never make a Type I error but we also will never
reject Hypothesis 6.3 and we will conclude that even double-headed coins are unbiased!
The danger now is that we accept our null hypothesis even when it is false and this
is called a Type II error. We therefore have the following two kinds of error.
Type I error: reject when true.
Type II error: accept when false. 6.5

What we need is a balance: we want to keep the rejection range small to minimize
Type I errors and keep the acceptance range small to minimize Type II errors. Because
we cannot do both at the same time, our choice of range depends on our own
judgement and the reasons for performing the experiment. For example, in English
Law our hypothesis is that the defendant is innocent. The prosecution attempts to
prove that the hypothesis of innocence is false. If an innocent person is punished,
we are making a Type I error; if a guilty person is allowed to go free, we are making
a Type II error. Since we prefer to make a Type II error and let a guilty person go
free than to make a Type I error and make an innocent person suffer, the court insists
that the case against the defendant be proved 'beyond reasonable doubt', which
means that the rejection range is made very small and the probability of making a
Type I error is minimized.
We now need more terminology and we call the probability of making a Type I
error the significance level of the test. When we say that 'our data lead us to reject
the hypothesis H at the 5% significance level', we are saying that although we believe
that the hypothesis is indeed false, we acknowledge that we will be making a Type
I error five times in every 100 experiments and that on these occasions the hypothesis
really is true.
One minus the probability of making a Type II error we call the power of the
test. Unfortunately, there is a major difficulty in specifying the power of a test because
if we want to know the probability of accepting the null hypothesis when it is false
we need to know what the alternative hypothesis is. The power of a test of one
hypothesis can be defined only in relation to a second hypothesis. Let us call our
hypothesis that the coin is unbiased the null hypothesis, Ho and let us consider a
second hypothesis, HI' that the coin is biased in such a way that the probability
that it falls heads is 0.6. We then spin our coin 100 times and record the number of
 Significance levels and the power of a test 8I

Figure 6.1 The two curves give the p,d.f.sfor two hypotheses, (a)Ho: P(H) = 0,5 and (b)Hi:
P(H) = 0,65, The acceptance range at a 5% significance level is from 40 to 60 heads. The
horizontally shaded area is 0,05 and gives the probability of making a Type I error. The
vertically shaded area is 0,16 and gives the probability of making a Type II error. The power
of the test is 1 - 0.16 = 0,84,

times it shows heads. Figure 6.1 shows the probabilities of the various possible
outcomes under each of the two hypotheses. Suppose we now set the acceptance
range to include any outcome in which the number of heads is between 40 and 60,
and set the rejection range to include any outcome in which the number of heads
is less than 40 or greater than 60. The horizontally shaded area in Fig. 6.1 is the
probability of making a Type I error, in this case 0.05, and determines the significance
level. The vertically shaded area in Fig. 6.1 is the probability of making a Type II
error, in this case 0.16, and accepting Ho when HI is true and Ho is false. Since we
want the power of the test to be high when the probability of making Type II error
is small, we define the power of the test as 1- 0.16 = 0.84. If we reduce the
significance level to increase protection against Type I errors by widening the accep-
tance range, we also reduce the power of the test, giving less protection against
Type II errors.
Now if we are in that rather rare situation in which we have only two possible
hypotheses, then we can talk about both the significance level and the power of the
test and choose our acceptance and rejection ranges accordingly. More often we will
have only our null hypothesis to test. If we decide to reject it, we may think of an
alternative explanation, formulate a new hypothesis and then test the new one, but
we will generally be concerned with one hypothesis at a time. For this reason, we
will nearly always specify the significance level of our tests rather than their power.
However, when we plan experiments, we can use the power of a test to decide, in
advance, how big our samples should be. For example, we have seen in the trial of
82 Testing hypotheses
the polio vaccine that with 200000 children in each group there was little doubt
about the effectiveness of the vaccine, but we have also seen that with only 20 000
children in each group, the effect of the vaccine may have been lost in the statistical
noise. In section 7.8.2 we will show how we can use the power of a test to help us to
decide how many children we need to include in such a study.

6.1.1 Summary

Before we perform an experiment we must have an idea that we are trying to test.
The danger is that we deceive ourselves and read into the data what we want to
see there rather than what is there. We therefore err on the side of caution by
making the null hypothesis that the effect is not there and then we try to disprove
the null hypothesis. If we succeed we look more favourably on the idea we are
testing, if we fail we look less favourably on the idea we are testing.
Formally, we proceed as follows:

• We have an idea we want to examine (the coin is biased).

• We make a series of measurements (throw the coin 100 times) that give us data,
D, (H, H, H, T, H, T, T, H, T. ..).
• We make an hypothesis, H (the coin is not biased).
• We calculate a test statistic, x (65 heads, say), based on D.
• Assuming that H is true, we use our knowledge of statistics to decide on the
distribution that x should follow (binomial with p = 0.5, n = 100).
• We decide on a significance level P(5%, say) and determine critical values (40 and
60 heads) that define the acceptance range (40-60 heads), and the rejection range
(0-39 and 61-100 heads) at this level of significance.
• If x (65) falls in the acceptance range (40-60) then we provisionally accept H, and
conclude that the coin is not biased, remembering that more extensive data (more
throws of the coin) may lead us to change our minds in the future.
• If x (65) falls in the rejection range (0-39 and 61-100 heads) we reject Hat the
P% significance level and we conclude that the coin is probably biased, remembering
that although we are confident about rejecting H there is still a probability P (5%)
that we are making a Type I error and the coin is not in fact biased (since even
an unbiased coin could produce 65 heads and 35 tails).

This may seem pedantic but science proceeds by the accumulation of very many
small steps and it is vital that we do not lose our way or allow ourselves to be
seduced into seeing what we want to see.

6.2 THE X2 TEST

In Chapter 5 we examined the distribution of the number of boys in families of eight

children and compared the observed distribution with the distribution we expected
if the assumptions of the binomial distribution were satisfied. In particular we wanted
to know if there was any evidence that large families tend to produce more boys
 The X2 test 83
or girls than expected, for if this is so then some parents may have a tendency to
produce children of one sex. We also looked at the number of plants in Mendel's
experiments· having various pairs of characteristics and we wanted to know if the
observed numbers supported Mendel's theory.
When we compare observed and expected numbers, a convenient and powerful
test is the X2 test based on the X2 distribution. Before we develop the test, remember
that we will start by assuming that 'nothing is going on'; our coins are unbiased,
there is no tendency to produce boys rather than girls, there is no tendency to
produce children of the same sex, the polio vaccine does not prevent polio, and so
on. Our initial hypothesis H is therefore referred to as the null hypothesis.
For our study of the number of boys in families of eight children, our null hypothesis
is that there is no propensity to produce children of the same sex so that the observed
numbers of boys and girls should agree with the predictions of the binomial distribu-
tion calculated in Table 5.1 and repeated in Table 6.1. We can then restate Has follows:
H: the observed frequencies do not differ significantly
from those expected for a binomial distribution. 6.6
Let us first look at 0 - E, the difference between the observed frequencies 0, and
the expected frequencies E. If Ho is true, the expected value of 0 - E is O. Now, if
the probability of any given outcome, two boys and six girls, say, is small, 0, the
observed number of times each outcome occurs will follow a Poisson distribution so
that the variance of 0 is equal to its expected value E. (0 - E)/ fi
will therefore
have an expected value of 0 and a variance of 1. Provided E is not too small,
(0 - E)/Ji will be approximately normally distributed. Therefore, if we square
(0 - E)/JE for each pair of observed and expected values and add them all together,
this variable will come from a X2 distribution, provided our null hypothesis is true, so
that
n

"~ (0.I - E)2/E

I
'" l
1m' X
67
i= 1

Table 6.1 Calculation of X2 from the observed and expected numbers of

boys in families with eight children. The sum of the X2 values is 92

Number Observed Expected

of boys frequency frequency Difference X2
0 215 165 +50 15.2
1 1485 1402 + 83 4.9
2 5331 5203 + 128 3.2
3 10 649 11 035 -386 13.5
4 14959 14628 + 331 7.5
5 11929 12410 -481 18.6
6 6678 6580 + 98 1.5
7 2092 1994 + 98 4.8
8 342 264 + 78 23.0
84 Testing hypotheses
where the sum is over all our observations and m is the number of degrees of freedom.
(This result is more subtle than it appears and is discussed further in the Appendix,
section 6.5.1. In fact we only need to assume that 0 follows a binomial distribution
rather than the more restrictive assumption that 0 follows a Poisson distribution.)
We can then test Hypothesis 6.6 by comparing the left-hand side of Equation 6.7
with a suitable acceptance range chosen from the X2 distribution with m degrees of
freedom.
We have to be careful with the number of degrees of freedom. For our families
with eight children, we start off with nine possible outcomes: 0, 1, 2 ... 8. Once eight
of the probabilities have been fixed, the last is also fixed since they must add up to 1,
so this removes 1 degree of freedom. However, we have also used the data to
estimate the probability that anyone child is a boy, so we lose another degree of
freedom. The total number of degrees of freedom is therefore 7 and we use the X2
distribution with 7 degrees of freedom to determine the critical value that defines
the acceptance range.
Table 6.1 shows the data of Table 5.2 with the contributions to the X2 distribution
calculated for each observation. Adding all the numbers in the last column of Table
6.1 gives a value of 92. The next thing we need to do is to decide on our acceptance
range. Suppose we are willing to risk making a Type I error by rejecting the hypothesis
even though it is true, about 1 time in 20. We then look up the 5% significance level
for a X2 distribution with 7 degrees of freedom and Table 10.2 tells us that the critical
value is 14.1, so our acceptance range is from 0 to 14.1 and our rejection range is
anything greater than 14.1. The X2 statistic is 92, well outside the acceptance range,
and we can decisively reject our hypothesis that the observed frequencies in Table
6.1 follow a binomial distribution.
To summarize, we want to know if there is a propensity for some families to
produce children of the same sex so we start with the null hypothesis that there
is no propensity for some families to produce runs of boys and others to produce
runs of girls. We then model the results we expect to find if our null hypothesis
is consistent with the data. For the example of boys and girls in Saxony, we find
reasonably good agreement between the observed and expected numbers, but because
the sample size is very large, the data are very precise and our statistical test is able
to pick up very small differences. Although the observed distribution is close to the
distribution predicted using our null hypothesis, we find that there are small but
significant differences. The data appear to suggest that there is a slight genetic
propensity to produce children of the same sex. To pursue the matter further we
would examine the deviations from the predictions, formulate new hypotheses, make
new models and hope eventually to find a model that not only fits the data but also
makes biological sense.

6.2.1 Mendel's peas

Mendel found that the observed ratios of dominant to recessive characters in the F2
generations of his peas were always close to 3 to 1 (Table 2.2). We can now use
 The X2 test 85
the X2 test to see whether the deviations from the expected ratio of 3 to 1 are small
enough to be attributed to chance alone. Our hypothesis in each case is then

H: The ratio of the number of dominants to recessives

in the F2 generation is 3:1. 6.8

For each experiment in Table 2.2 we take the total number of peas (7324 for
the comparison of round and wrinkled seeds) and calculate the expected number of
dominants (7324 x 0.75 = 5493 for round seeds) and recessives (7324 x 0.25 = 1831
for wrinkled seeds) on the hypothesis that the expected ratio of dominants to recessives
is 3: 1. We then use Equation 6.7 to calculate the values of X2, so that for the experi-
ment involving round and wrinkled seeds we have

X2 = r. (0 - £)2/£

= (5474 - 5493)2/5493

+ (1850 - 1831)2/1831

= 0.263. 6.9

The X2 statistics for each of Mendel's experiments are given in Table 6.2. In
each experiment we have two measurements, the number of plants showing the
dominant character and the number showing the recessive character, but we have
had to use the total number of plants to calculate the expected frequencies, so we
lose 1degree of freedom, and we need to compare each statistic with a X2 distribution
with 1 degree of freedom.
To see whether each of these numbers could reasonably come from a X2 distribution
with 1 degree of freedom, we let our acceptance range correspond to the lower 95%
of the distribution, so that our significance level is 5%. From Table 10.2, the critical
value of X; at the 5% significance level is 3.84, so that the values of f in Table 6.2
are all well within the acceptance range. In fact, they are all so well within the
acceptance range that they look suspiciously good and we may begin to wonder if
Mendel did not in fact 'massage' his data to improve the agreement between the

Table 6.2 Values of X2 for Mendel's experiments on peas calculated

from the data given in Table 2.2

Characters X2

Round vs wrinkled seeds 0.263

Yellow vs green seeds 0.D15
Purple vs white flowers 0.391
Smooth vs constricted pods 0.064
Axial vs terminal flowers 0.350
Green vs yellow unripe pods 0.451
Tall vs dwarf stems 0.607
86 Testing hypotheses
experimental values and his expectations. Let us, therefore, make a new hypothesis,
H: The ratio of dominants to recessive is 3: 1 and
Mendel was an honest scientist, 6.10
and see if his data gives us cause to reject the hypothesis. Our suspicion has
been aroused because of the consistently small values of X2 obtained in Table 6.2
and we want to know if seven such small values could reasonably have arisen by
chance.
Usually our acceptance range would include all values less than the critical value,
because significant differences between observed and expected values will give large
values of X2 that we treat with suspicion. However, if Mendel tried to improve the
agreement between his observed and expected values, it is small values of X2 that
we treat with suspicion and we now choose our acceptance range so that the observed
numbers should be greater than the critical value with 95% probability and less than
the critical value with 5% probability. In other words, we take the bottom 5% rather
than the top 5% of the distribution as our rejection range. For a X2 distribution with
1 degree of freedom, the bottom 5% of the distribution corresponds to values between
o and 0.004 (Table 10.2) and all of the values in Table 6.2 lie within the acceptance
range so that we cannot reject the hypothesis that Mendel was an honest scientist
at the 5% significance level.
The fact that we have seven values of X2, which taken together seem to be
consistently small, leads us to wonder if we can combine them and obtain a more
sensitive test of our hypothesis concerning Mendel's honesty. Adding the X2 values
(each of which has 1 degree of freedom) gives 2.14 and if Hypothesis 6.10 holds,
this will be a random sample from a X2 distribution with 7 degrees of freedom. For
a i distribution with 7 degrees of freedom, the critical value for the bottom 5% of
the distribution is 2.17 (Table 10.2) and since this is slightly greater than 2.14, we
can reject the hypothesis that Mendel was an honest scientist at the 5% significance
level. The probability that we are making a Type I error (rejecting the hypothesis
when true) and accusing Mendel falsely is only 1 in 20.
The debate as to what Mendel actually did is complex but the evidence against
Mendel can be made stronger than it has been made here. Fisher carried out a
detailed analysis of all of Mendel's published data and concluded that the data are
too good to be true (Fisher, 1936). However, saying that Mendel 'cheated' does not
get us very far because his theory was correct: he had. in fact, made one of the most
important scientific discoveries of all time. If his work had been understood and ap-
preciated, the study of evolution and inheritance might have been advanced by
30 years. What we really want to know is how and why did he manipulate his data?
On the evidence available to him Fisher concluded that 'after examining various
possibilites, I have no doubt that Mendel was deceived by a gardening assistant who
knew too well what his principle expected from each trial' (Fisher, 1965). Blaming
an unnamed assistant would seem to be wishful thinking on Fisher's part and an
attempt to protect Mendel's honour. It seems to me equally likely that Mendel
deceived himself: every scientist tends to repeat experiments that give the 'wrong'
answer, stopping as soon as the 'correct' answer turns up.
 Contingency tables 87

Even if Mendel's data are falsely presented, various interpretations are possible.
It appears that Mendel did not realize that he had discovered the genetic basis
of inheritance and that his work on peas could be generalized to include the entire
living world. He seems rather to have believed that the numerical proportions he
discovered for the inheritance of characteristics in peas would have to be determined
separately for each species of plant or animal. Mendel appears not to have speculated
on the mechanism that might have given rise to his observations: if he had, he would
surely have been led to postulate pairs of units (genes), one inherited horn the male,
the other inherited from the female. In other words, Mendel may have been concerned
in his paper primarily to demonstrate his methodology and to educate others, and
may have thought that the precise numbers were unimportant. If so, it may have
seemed entirely reasonable to select those numbers that best demonstrated the result
he was trying to establish. (Broad and Wade (1985) have written about fraud in
science and Gould (1981) has addressed such questions in relation to intelligence
testing.) I have done something similar in this book by selecting examples that match
my expectations while avoiding those that are overly complicated. As Gould (1980)
says, 'Science is not an objective, truth directed machine, but a quintessentially human
activity, affected by passions, hopes and cultural biases:

6.3 CONTINGENCY TABLES

Once we agree that Mendel's theory of inheritance is essentially correct and that
the various characteristics are determined by what we now call genes, we are led
to wonder whether the genes for different characters are related: perhaps the genes
that confer 'roundness' and 'yellowness' on the seeds are linked. The question as to
whether two properties are related arises frequently in biology and is amenable to
analysis using a X2 test.
Mendel reported the results of crosses involving more than one pair of characters,
but since he had not developed a mechanistic explanation of inheritance, he did not
imagine that different pairs of characters might be linked-he assumed that the
various characters were independent and then proceeded to show that the various
combinations occurred in the appropriate proportions. It is now known that of the
genes studied by Mendel only two are linked: those that determine the colour of
the unripe pods and those that determine the length of the stem (Lambrecht, 1961).
With our present-day knowledge of genes and chromosomes, let us test Mendel's
data for linkage. To do this we use contingency tables, such as Table 2.5 repeated
here as Table 6.3. It is called a 'contingency table' because we can use it to calculate
the probability that a seed is round contingent on it also being green, and vice versa:
in other words, we can use it to test conditional probabilities.
In section 2.2.3 we estimated, for example, the probability that a seed will be
round given that it is yellow, P(round Iyellow), and showed that this is almost the same
as P(roundlgreen). We then argued that since the probability of it being round is the
same whether it is yellow or green, the gene for shape and the gene for colour must
be independent. But we need to quantify this observation and decide if the small
differences in the two probabilities are significant or not.
88 Testing hypotheses
Table 6.3 Mendel's data on the distribution of seed colour and shape

Yellow Green Total

Round 315 108 423
Wrinkled 101 32 133

Total 416 140 556

To develop our X2 test for contingency tables, we again start from the hypothesis
we want to test, in this case

H: The traits for seed colour and shape are independent. 6.11

Given this assumption, we can use our statistical theory to calculate the expected
frequencies for seeds of each colour and shape and then use a X2 test to compare
the observed and expected frequencies.
For Mendel's data, we want to calculate the expected frequency in each cell of
Table 6.3 under Hypothesis 6.11, taking the row and column totals as given. As we
saw in section 2.2.3,

P(round) = 423/556 p(yellow) = 416/556. 6.12

and we know from Equation 2.17 that if 'roundness' and 'yellowness' are statisti-
cally independent, then

P(round and yellow) = P(round) x p(yellow) = 423 x 416/5562 = 0.569. 6.13

Since we have a total of 556 plants, the expected number of plants whose seeds are
both round and yellow is 0.569 x 556 = 316.5, while the observed number is 315.
The contribution to the X2 sum is (315 - 316.5)2/316.5 = 0.0070. Repeating this for
the other three combinations of colour and shape gives the expected frequencies and
contributions to the X2 sum shown in Table 6.4. Adding together the four contributions
to the X2 sum gives 0.116.
Once again we need to decide how many degrees of freedom to use in our X2
test. We started with four measurements but we used the total number of seeds
(556), the marginal number of round seeds (423), and the marginal number of yellow
seeds (416), so that we have lost 3 degrees of freedom. We do not count the marginal
number of wrinkled seeds or the marginal number of green seeds, since once we
have fixed the three numbers mentioned above, these two are also fixed and we lose
nothing further by using them. Weare therefore left with 4 - 3 = 1 degree of
freedom. (If you calculate 0 - E for each cell in Tables 6.3 and 6.4, you will get
± 1.49 for each of them, showing that although we have four differences, there is
really only one independent difference.)
If we choose to test the experiment at the 5% significance level, the acceptance
range for xi is from 0 to 3.84 (Table 10.2), and we see that Mendel's result, 0.116,
falls well within this range (perhaps too well within it), and we accept the hypothesis
that the genes for shape and colour are independent.
 Contingency tables 89
Table 6.4 The expected frequencies and the contributions to the 1.2 sum for
the different combinations of seed colour and shape corresponding to the
observed values in Table 6.3. The 1.2 values add up to 0.116

Expected 2
1.
Yellow Green Yellow Green
Round 316.49 106.51 0.007 0.021
Wrinkled 99.51 33.49 0.022 0.066

Now we have ignored some of the information inherent in Mendel's data because
his theory predicts that the marginal totals should be in the ratio of 3 to 1, and
instead of using the observed marginal totals to calculate the expected cell frequencies,
we could have used his theory. If we make use of his theory, our hypothesis is
slightly different and we have

H: The ratio of dominants to recessives is 3 to 1 and the

genes for roundness and colour are independent. 6.14

The probability that a seed is both round and yellow under Hypothesis 6.14 is then
0.75 X 0.75 = 0.5625, and the expected cell frequency for plants whose seeds are
both round and yellow is 0.5625 x 556 = 312.8, which we see is slightly different
from the previous expected value of 316.5. We again calculate the expected frequencies
for each cell in our table and use these to calculate the value of X2, which turns out
to be 0.47, a larger number than we had before. This time, however, we have lost
only 1 degree of freedom in using the total number of plants to determine the
expected frequencies, so we are left with 4 -1 =
3 degrees of freedom. We must
now compare our result with a X2 distribution with 3 degrees of freedom, for which
the acceptance range at the 5% significance level is from 0 to 7.81 (Table 10.2). We
see that we are again well within the acceptance range and we should accept the
Hypothesis 6.14 that the ratio of dominants to recessives is 3 to 1 and that the
genes for shape and colour are independent.

6.3.1 Multiway contingency tables

There is no reason to limit ourselves to only two categories for each variable in our
contingency table. The extension to more than two is straightforward and the rules
for a contingency table with R rows and C columns are as follows.
If we need to use the marginal totals to calculate the expected cell frequencies (as
in our analysis based on Hypothesis 6.11), we proceed as follows.

1. Calculate the expected cell frequencies, by multiplying the appropriate row total
by the appropriate column total and dividing by the grand total.
2. Carry out a X2 test with (R - 1) X (C - 1) degrees of freedom.

If we have a theory that predicts the marginal totals (as in our second analysis
90 Testing hypotheses
based on Hypothesis 6.14), we proceed as follows.

1. Calculate the expected marginal totals from the grand total using the theory, and
then calculate the expected cell frequencies as in the previous case.
2. Carry out a X2 test with RC - 1 degrees of freedom.

Convince yourself of the validity of Rule 1. To understand Rule 2 in the first case,
note that we start with RC frequencies. We lose 1 degree of freedom when we use
the grand total. With R - 1 row totals and the grand total we can calculate the
remaining row total. We therefore lose another R - 1 degrees of freedom when we
use the row totals to calculate cell frequencies. Similarly, we lose C - 1 degrees of
freedom when we use the column totals. This leaves RC - (R - 1) - (C - 1) - 1 =
(R - 1) X (C - 1) degrees of freedom. For Rule 2 in the second case we only use
the grand total and hence lose only 1 degree of freedom, leaving RC - 1 degrees
of freedom.

6.3.2 Corrections for small numbers

A critical assumption in our development of the X2 test is that we can approximate

a binomial distribution by a normal distribution. If the numbers are small, this can
lead to significant bias in the results. Snedecor and Cochran (1989, p. 77) give the
following rules for applying the X2 test to small samples,

8 No class expectation should be less than 1.

8 Two expected values may be close to 1 if most of the others exceed 5.
8 If necessary, classes should be combined to meet these rules.

For a 2 X 2 contingency table, a conservative recommendation (Sokal and Rohlf,

1981, p. 711) is that no expected number should be less than 5 and a correction
factor, known as Yate's correction (Bulmer, 1979, p. 163; Hays, 1988, p. 774( can
be included in 2 X 2 tables by replacing Equation 6.7 with

X2 = (~
iO - EI- 0.5)2/E 6,15

where 1 a I indicates the absolute value of a. (The absolute value is the number with
a positive sign so that I - 31 = 1+ 31 = + 3.) Alternatively, provided that the
numbers in each cell are not too large we can calculate the probabilities of all possible
outcomes exactly which is the basis of Fisher's exact test (Siegel and Castellan, 1988,
p. 103; Hays, 1988, p. 781).

6.4 SUMMARY

When you apply statistics to the analysis of your data, there will always be some
idea or theory that you wish to test and of course it is precisely in the formulation
of such ideas or theories that the creativity of science lies. However, we choose to
proceed in a rather indirect way. If we suspect that a coin is biased we assume initially
that it is not and'then try to disprove our null hypothesis. Given our null hypothesis,
 Appendix 91
we use our knowledge of biology and statistics to decide if the observed data seem
reasonable if the null hypothesis is indeed true. If the observed data seem reasonable,
we accept our null hypothesis and conclude, provisionally, that our idea cannot be
supported by the data. It is still open to us to design a more subtle experiment, carry
out more trials or in some way try to tease out the information that we believe
is there. If the observed data do not seem reasonable given the null! hypothesis, we
conclude, provisionally, that our idea is supported by the data. It may still be the
case that the observed effects could be explained on the basis of some other idea,
one that is quite unrelated to the one we are proposing. So we try to test our ideas
in as many different ways as possible and if we still come up with the same theory
we begin to believe it. At the end of the day, all that matters is that we can provide
a sound and convincing biological explanation for the effects that we observe. The
mathematics is the handmaiden to the biology that we hope to understand.
The X2 statistic, which we have examined at some length, is only one of many
statistics that we can derive from our experimental data. In the next three chapters
we will examine other situations and other statistics that we can use to test hypotheses
about our data.

6.5 APPENDIX

6.5.1 The X2 test

Consider an experiment in which there are only two possible outcomes such as heads
or tails, tall or short plants, boys or girls, and so on. Let N be the total number of
events (throws of the coin, plants sampled, children) of which a fall into the first
category (heads, tall plants, boys), while b fall into the second category (tails, short
plants, girls). Let ri be the probability that a occurs and let f3 be the probability
that b occurs. Then

a +b =N ri + f3 = 1 6.16

E(a) = riN V (a) = rif3N 6.17

E(b) = f3N V (b) = f3riN. 6.18
Now
X2 = I:(O - £)2/E
= (a - riN)2/riN + (b - f3N)2/f3N
= {f3(a - riN)2 + ri(b - f3N)2}/rif3N. 6.19

But
b - f3N = (N - a) - (1 - ri)N = riN - a, 6.20

so that substituting for b - f3N in equation 6.19,

X2 = (a - riN)2(f3 + ri)/rif3N = (a - riN)2/rif3N. 6.21

92 Testing hypotheses
Now E(a - rxN) = 0 and V(a - rxN) = rxf3N so that (a - rxN)2Irxf3N has mean zero
and variance 1. Provided rxN is sufficiently large we can assume that its distribution
is approximately normal, so that X2 given by Equation 6.21 belongs to a X2 distribution
with 1 degree of freedom.
The key point to note is that although the variance pqN of a binomially distributed
variable is not equal to the expected value pN, we nevertheless divide by the expected
value in Equation 6.19. The sum then reduces to the correct expression for a X2
distribution with the number of degrees of freedom equal to the number of degrees
of freedom in the data. In other words, Equation 6.7 holds for binomially distributed
observations and p need not be small as would be the case if it only held for observa-
tions following a Poisson distribution as implied in the text.
It is still necessary however for each expected value to be sufficiently large for
the assumption of normality to hold and experience shows that this does hold provided
the conditions noted in section 6.3.2 are satisfied.

6.6 EXERCISES

1. Carry out X2 tests on the hypothetical data in Table 1.4 and decide which numbers
confirm Mendel's theory and at what significance level. How good were your guesses?

2. The data of Table 1.2 for the results of the trial of the polio vaccine are repeated
as a contingency table in Table 6.5. Carry out a X2 test to decide if the incidence of
polio is related to use of the vaccine.

3. Carry out a X2 test on the data of Table 5.4 and decide whether to accept or reject
the hypothesis that the number of deaths follows a Poisson distribution. (Note the
recommendations in section 6.3.2.)

4. Carry out a X2 test of Bateson's data in Table 2.6 using both the marginal totals
and Mendel's theory to predict the cell frequencies. Are we justified in concluding
that in Bateson's experiment the shape of the pollen is related to the colour of the
flowers?

5. In an experiment to investigate the possible protective effects of vitamin C against

rabies, 98 guinea-pigs were given a small dose of fixed rabies virus. Of these 48 were
also given vitamin C, the rest were not. Table 6.6 gives the number that died or
survived under each treatment (Banic, 1975). Deaths occurred in 35% of the treated and

Table 6.S The number of children that developed polio and that
remained healthy after receiving either the vaccine or a placebo

Paralytic Healthy
cases cases

Vaccine 33 200712
Placebo 115 201114
 Eu~
~ 93
Table 6.6 The number of treated and untreated guinea-pigs that died
or survived in an experiment on the efficacy of vitamin C in treating
rabies

Treatment No. of deaths No. of survivors

Vitamin C 17 31
Control 35 15

70% of the control animals. Carry out a X2 test with and without Yate's correction to
decide if vitamin C is effective against rabies.

6. In August 1989 Newsweek published a report on a US government-funded study

that showed that AZT (azidothymidine) helps to slow down the development of the
diseases AIDS (acquired immune deficiency syndrome) (Newsweek, 1989). Either AZT
or a placebo was given to 713 people. At the time of writing, 14 of those taking
AZT had developed AIDS, while 36 of those given the placebo had developed AIDS.
Do these data encourage us to believe that AZT slows the development of AIDS?
(Hint: Without knowing how many received the drug and how many received the
placebo, we can draw no conclusions. Let us therefore assume that about half were
given the drug and half the placebo.)

7. The shell of the snail Cepaea nemoralis can be yellow, brown' or pink and it can
be banded or uniform. Cain and Sheppard (1952, 1954) showed that shell colour is
adaptive in that it provides camouflage against attack by thrushes. In abitats
~ provid-
ing a uniform background the shells tend to be uniform in colour; in habitats providing
a varied background the shells tend to be banded. To confirm that this was the result
of selective predation by thrushes, Sheppard (1951) counted the numbers of banded
and unbanded snails that were living and the numbers that were killed by thrushes
in a certain area. Since a thrush breaks the snail shell by smashing it on a rock, il
was possible to sample the number killed by thrushes in each class. The results are
given in Table 6.7. Use a X2 test to decide if there is a significant association between
'bandedness' and being killed by a thrush. Do these data support Cain and Sheppard's
predation hypothesis?

Table 6.7 The numbers of banded and unbanded snails that were
living and that were killed by thrushes

Banded Unhanded

Living 264 296

Killed 486 377
7

Comparisons

It's proper to begin with the regular facts, but after a rule has been established beyond
all doubt, the facts in conformity with it become dull because they no longer teach us
anything new. Then it's the exception that becomes important. We seek not resemblances
but differences, choose the most accentuated differences because they're the most striking
and also the most instructive. R.M. Pirsig (1980)

When we analyse data we have an hypothesis that we wish to test. We then use
our knowledge of statistics to give us a measure of the probability that the observed
data will arise in an experiment if the hypothesis is true. So far we have considered
situations in which we want to know only if the measured data could reasonably
have arisen from a particular hypothetical distribution. In practice, life is usually more
complicated than this. We may have a number of different variables, perl)aps rainfall
and temperature, each of which might affect the outcome of our experiment. which
might be the growth of a crop, and we want to be able to test each of them. We
may even find that the effect of one variable depends on the particular values of
another so that high temperatures might increase crop growth when it is wet but
lead to decrease in crop growth when it is dry.
In this chapter we will consider problems that arise when we compare measurements
made under several different conditions. We will consider some aspects of the design
of experiments and this will lead us into the chapters on the analysis of variance and
regression where we formalize the way in which we design experiments and analyse
data.

7.1 ONE- AND TWO-TAILED TESTS

In Chapter 6 we compared the observed and expected number of boys in families

with eight children. To do this we used a X2 test and since a failure of the hypothesis
would lead to large values of X2, we took as our acceptance range (at the 5% signi-
ficance level) all values less than 11.1 and our rejection range corresponded to the
top 5% of the distribution. To test Mendel's data on peas, we again argued that
failure of the hypothesis would lead to large values of X2 and our rejection range
was again the top 5% of the distribution. However, when we set out to test Mendel's
honesty, we argued that if he had cheated, the observed results would have been
 One- and two-tailed tests 95

Figure 7.1 The p.d.f. for a standard normal distribution. Each of the separately shaded areas
contains 2.5% of the total area.

smaller than expected, and so we took as our acceptance range (at the 5% significance
level) all values greater than 2.26 and our rejection range then corresponded to the
bottom 5% of the distribution. On still other occasions we will want to know if two
observations differ in either direction and on these occasions we will want our rejection
range to include both tails of the distribution.
To illustrate this, Fig. 7.1 shows the p.d.f. for a standard nonnal distribution. Each
of the separately shaded areas contains 2.5% of the total area. Suppose we now
measure two numbers, let us call them a and b, and take as our null hypothesis that
their difference belongs to a standard nonnal distribution. Then we can set up various
alternative hypotheses as indicated in Table 7.1. If the alternative hypothesis is that
a > b, we take as the rejection range the two shaded areas on the right-hand side of
Fig. 7.1; if it is that a < b, we take the two shaded areas on the left-hand side
of Fig. 7.1; if it is that a =F b, we take as the rejection range the two vertically
shaded areas at the two ends of Fig. 7.1. The point to bear in mind is that in all
three cases given in Table 7.1, the significance level is the same but the acceptance

Table 7.1 Tests of the hypothesis a = b at the 5% significance level, assuming that a - b
follows a standard normal distribution, for various alternative hypotheses

Null Alternative Rejection

Hypothesis hypothesis range for a - b
a=b a>b 1.65 to CfJ
a=b a<b - CfJ to 1.65

a=b a#:b - CfJ to - 1.96 &: 1.96 to CfJ

96 Comparisons
range depends on the alternative hypothesis so that we are effectively choosing the
acceptance range in such a way as to maximize the power of the test for a given
alternative hypothesis.

7.2 COMPARISON OF MEANS

Suppose that we want to compare the results of two measurements. The first thing
we will probably want to know is whether the means are significantly different. This
is essentially what we did in section 5.2.2 in our study of the effects of the polio
vaccine where we noted that the probability of a vaccinated child developing polio
was 16.4 ± 2.9 per 100000, while the probability of an unvaccinated child developing
the disease was 57.1 ± 5.3 per 100000. Using the result (Appendix, section 4.5.2)
that for two independent random variables, the variance of their sum or difference is
the sum of their variance, we were able to calculate the standard deviation of the
difference as
5 = (2.92 + 5.32)1/2 = 6.0. 7.1
and we noted that this was much less than 40.7, the reduction in the number of
children per 100000 contracting polio when they were given the vaccination.
Now we can be more precise in our analysis. The hypothesis we want to test is
that the vaccine brings about a significant reduction in the number of children who
contract polio and so we use a one-tailed test. Since there are a large number of
trials and the numbers of children contracting polio are not too small, we can approxi-
mate the distribution of the difference in the numbers contracting polio under the
two treatments by a normal distribution. For a standard normal distribution, 0.1%
of the area under the curve is above 3.09 (Table 10.1), so that for a normal distribution
with mean zero, the hypothesis we are testing, and standard deviation 5,0.1% of the
area is above 3.095 = 3.09 x 6.0 = 18.5. The difference between the numbers of
children contracting polio in the two groups, 40.7, is well outside the acceptance
range at the 0.1% significance level so that we are able to conclude that the vaccine
does indeed bring about a significant reduction in the incidence of polio.
In this example we have used a one-tailed test. Suppose, however, we had two
different vaccines and we wanted to know if there was any significant difference
between them. We would not know in advance which was the better of the two
and we would then use a two-tailed test. Our acceptance range at a 0.1% significance
level would then be - 3.295 to + 3.295 (see Table 10.1), which in this case is - 19.4
to + 19.4, rather than anything less than 18.5, the range for the one-tailed test.
Comparing the mean, m, with the acceptance level of 3.095 is clearly equivalent
to comparing m/5 with the acceptance level 3.09 for a standard normal deviate. For
our polio data we can then compare 40.7/6.0 =6.8 with 3.09 for the 0.1% significance
level and we again find that the mean differs significantly from zero.

7.2.1 ttests

Whenever we want to compare two means we proceed as in the example above, so

let us write it out clearly, assuming that we are using a two-tailed test. To test if
 Comparison of means 97
the mean of a set of numbers differs significantly from zero, divide the mean by the
standard deviation: if the result is greater than 1.96, the mean differs significantly
from zero at the 5% level; if it is greater than 2.58, the mean differs significantly from
zero at the 1% level; if it is greater than 3.29, the mean differs significantly from zero
at the 0.1% level; and so on. To add to our notation: the three most commonly used
significance levels, 5%, 1% and 0.1%, are often indicated by", ,.,.and """, respectively,
while if a result is not significant at the 5% level we indicate this with the letters ns.
There is, however, an important correction we need to apply when we have only
a few observations. Consider the following two numbers, chosen at random from a
standard normal distribution: 0.218 and 0.332. We know that their true mean is 0
and their true variance is 1, since that is how they were chosen, and indeed both
0.218 and 0.332 lie well within the acceptance range of ± 1.96 for a 5% significance
level. Suppose we didn't know in advance how they were chosen and wanted to
test the hypothesis that their mean value differs from zero. The calculated mean, m,
of 0.218 and 0.332 is 0.275 and the calculated population standard deviation is 0.081,
so that the standard deviation of the mean, sm' is 0.057 and m/sm = 4.8. Since this
is greater than 3.29, we would say that the mean differs significantly from zero at
the 0.1% significance level and give it 3 stars! Now we know that this is not the
case, since we deliberately chose the numbers from a normal distribution with a mean
of zero. So where have we gone wrong? Our estimate of 0.275 for the mean is
reasonable since we expect it to lie in the range ± 1. However,' our estimated popula-
tion standard deviation 0.081 is 12 times smaller than the true value of 1 and this
is where our problem lies. To illustrate this further, I generated two more normally
distributed numbers and obtained - 0.969 and 0.582. For these two numbers the
mean is - 0.194 "and the standard deviation is 0.775, so that m/sm = 0.25, this time
well within the acceptance range at the 5% significance level. In this second case the
mean differs from zero by about the same amount as before but the estimate of the
standard deviation is now much closer to the true value of 1. So the problem arises
because we have had to use an estimate of the standard deviation instead of the true
value, thereby introducing a further uncertainty. If we had used the true value of 1
for the standard deviation, we would have had 0.275 and 0.194 for the two estimates
of m/sm and in both cases would have accepted the hypothesis that the numbers
came from a normal distribution with mean zero. Since we do not know the true
value of the standard deviation, we need to allow for the inaccuracy inherent in our
estimate of the standard deviation in setting the acceptance range for the ratio m/sm'
The solution to this problem was found by Student (1908), who realized that m/sm
follows a normal distribution only if Sm is known exactly. He therefore set out to
determine the distribution that is followed by m/sm' Student was able to show that
if m is the mean of n readings taken from a normal distribution whose true mean is
J1., and if sm is the estimated standard deviation of the mean, then

f=(m-J1.)/sm"'fn_I, 7.2

In words, the f statistic in Equation 7.2 follows Student's f distribution with n - 1

degrees of freedom, the number used to estimate Sm' (Equation 7.2 is derived in the
Appendix, section 7.10.1.) Of course, we do not in general know the true mean, J1.,
98 Comparisons
for if we did we would not be concerned with the estimate m. What we do is to
make an hypothesis about Jl and then test the difference (m - Jl)/5m. In our example
above, the hypothesis is that Jl = 0 and so we test ml5m against tn-I'
Let us look at our two pairs of numbers again. Since we are comparing two numbers,
we have 2 degrees of freedom, but we have used up one of them when we used
the mean to estimate the standard deviation. We therefore look up the t distribution
with 2 - 1 = 1 degree of freedom in Table 10.1. The acceptance range at a 5%
significance level is - 12.71 to + 12.71. Since our two ratios were 4.8 and 0.25,
both of them now lie well within the accepta.nce range for the 5% significance level
and we accept our null hypothesis that the two pairs of numbers come from a normal
distribution with mean zero.
Since this is a crucial result let us restate it: when we divide the sum of the squares
of the deviations by n - 1when we calculate the standard deviation, subtracting the
1allows for the fad that we have used the estimated mean, and not the true mean,
and this gives us our best unbiased estimate of the standard deviation. The use of
the t distribution rather than the normal distribution to test the ratio of the estimated
mean to the estimated standard deviation allows for the fad that we have used the
estimated rather than the true standard deviation, so that even our best estimate of
Sm may not be very good. For very large samples the estimate of the standard devi-

ation will be close to the true value and the t distribution tends to the normal distri-
bution as the number of degrees of freedom increases as illustrated in Figure 5.6.
We can therefore look on the t distribution as a corredion to the normal distribution
when the sample size is small.
To illustrate the use of the t distribution, Student (1908) analysed data on the
effed on sleep of two drugs, hyoscyamine and hyoscine. The drugs were administered
to ten patients in the Michigan Asylum for the insane at Kalamazoo and the results
were repoted by Cushny and Peebles (1905). The amount of time each patient slept
was measured for between 3 and 9 nights in each case with no drug and after the
administration of 0.6 mg of hyoscyamine or hyoscine. The results are given in Table
7.2.
To test if hyoscyamine is soporific, we calculate the mean amount of sleep gained,
as compared with the control (0.75 hours) and the standard deviation of the mean
(0.57 hours). The ratio gives us a t value of 1.32. We have ten points and lose 1
degree of freedom in calculating the mean, so we are left with 9 degrees of freedom.
Since we are interested in the drug only if it increases the amount of sleep, we use
a one-tailed test. At the 5% significance level. t9 is 1.83 (Table 10.1) so the effed of
hyoscyamine is not significant. For hyoscine, the value of t is 3.68, which exceeds
the value of t9 at the 1% significance level. 2.82, and we conclude that hyoscine does
increase the amount of sleep significantly and give it two stars.
So far we have tested each mean on its own to see if it differs from zero. We
might also like to know if the two drugs differ significantly in thier effeds. Now we
want to see if the difference between two means differs from zero. The t test for the
difference of two means is a straightforward extension of the above results. If we
have n readings 'in each set of data so that the variance of the first mean is s~In, while
that of the second is 5;ln, then the variance of the difference is 52 = s/n ~ + 5;ln.
 Comparison of means 99

Table 7.2 The number of hours for which patients slept with no drug, with hyoscyamine and
with hyoscine. The increase in the amount of sleep gained with each drug as well as the difference
in the amount of sleep gained with each drug are also given

Patient Control Hyoscyamine Increase Hyoscine Increase Difference

1 0.6 1.3 +0.7 2.5 + 1.9 1.2

2 3.0 1.4 -1.6 3.8 +0.8 2.4
3 4.7 4.5 -0.2 5.8 +1.1 1.3
4 5.5 4.3 -1.2 5.6 +0.1 1.3
5 6.2 6.1 -0.1 6.1 -0.1 0.0
6 3.2 6.6 +3.4 7.6 +4.4 1.0
7 2.5 6.2 +3.7 8.0 + 5.5 1.8
8 2.8 3.6 +0.8 4.4 + 1.6 0.8
9 1.1 1.1 +0.0 5.7 +4.6 4.6
10 2.9 4.9 +2.0 6.3 +3.4 1.4

m 0.75 2.33 1.58

s 1.79 2.00 1.23
sm 0.57 0.63 0.39
t 1.32 3.68 4.06
ns •• ••

We call 52 the 'pooled' standard deviation since it combines the estimates from
both sets of data. Dividing the difference in the two means by the standard deviation
of the difference gives our t parameter with 2n - 2 degrees of freedom. (This is
strictly valid only if the 'true' variances, C1~ and C1,~ are the same. If this is not the
case, or if the number of measurements contributing to each of the means differs, an
approximate test can be carried out in which we use the same t statistic but a rather
more complicated expression for the number of degrees of freedom. This is discussed
in the Appendix, section 7.10.2).
For Student's data on soporific drugs (Table 7.2), the two means are 0.75 ±0.57,
and 2.33 ± 0.63 hours so that

m2 - m1 = 1.58 hours

5
m,- m2 = (0.57
2
+ 0.632)1/2 = 0.85 hours 7.3

t = 1.86.

Now we need to compare this with the acceptance range for a t distribution with
2 x 10 - 2 = 18 degrees of freedom. Since either drug could be better, we need to
use a two-tailed test and since t1B at the 5% significance level is 2.10, the two drugs
do not differ significantly in their effect.

7.2.2 Paired t tests

When analysing experimental data it is important to make full use of all of the
available data. In comparing the two drugs whose effects are given in Table 7.2, we
100 Comparisons
compared the mean effects of the two drugs, and if there had been 20 different
patients this would have been all we could have done. However, each patient received
both drugs and so we could instead calculate the difference between the effects of
the two drugs on each patient separately and then see if the mean of the differences
(rather than the difference of the means) is significant.
To illustrate the problem, the data of Table 7.2 are plotted in Fig. 7.2, from which
it is clear that the individual patients vary greatly in their response to the drugs and
that this tends to conceal any difference between the effects of the two drugs. How-
ever, in all cases the second drug induces at least as much sleep as the first and by
looking at the difference between the effects of the two drugs on each patient, the
last column in Table 7.2, we might hope to have a more sensitive test of the difference
between the effects of the two drugs. The mean difference is 1.58 hours with a
standard deviation of 0.39 hours, giving a value of t equal to 4.06. We have ten
differences and have used up 1 degree of freedom in calculating the mean, leaving
us with 9 degrees of freedom. Since either drug might be more effective, we need
to use a two-tailed t test. At the 1% significance level, t9 is equal to 3.25 (Table 10.1),
so that the drugs do differ significantly at the 1% level.
The point to note is that when we compared means without taking into considera-
tion the fact that both drugs were given to each patient, their effects were not
significantly different, whereas when we included the additional information that each
patient was given both drugs, their effects were significantly different. The second
kind of test, in which we take the treatments in pairs, is called a paired t test. Obviously,
if we had used 20 different patients and given each of them one of the drugs only,

Figure 7.2 The number of hours of sleep gained by each patient after the use of hyoscyamine
and hyoscine. The points with error bars give the mean effects of the two drugs ± 2.26sm.
The diagonal lines link the response of each patient to the two drugs.
 Non-parametric tests 101
instead of using ten patients each of whom received both drugs, we would not have
been able to use a paired t test Note in particular that in Table 7.2 the mean of the
differences is precisely the same as the difference between the means. The increase
in the sensitivity of the paired t test as compared with the unpaired t test is due to
the reduction in the standard deviation from 0.85 hours, when we use the difference
between the means, to 0.39 hours when we use the mean of the differences. By
taking the differences between the responses to each drug separately for each patient,
the paired t test calculates the variation due to the effect of the drugs after allowing
for the average response of each patient

7.3 NON-PARAMETRIC TESTS

When we compare two means using a t test, we assume that the individual measure-
ments are normally distributed about their mean values. Although this assumption
may be correct in the case of Student's data on the two drugs, we may have other
sets of data for which the assumption of normality is not justified. We might even
measure the response qualitatively so that we decide in each case which drug appears
to be the better of the two without quantifying the effect in terms of hours of sleep.
In such cases, we do not make precise assumptions about the underlying distribution
or its parameters, and such tests are therefore called non-parametric.
Having carried out a parametric test on Student's data, let us see what happens
if we carry out a non-parametric test. We will start with hyoscyamine but will now
only use the fact that the drug increased the amount of sleep for five patients,
produced no effect in one patient and decreased the amount of sleep in four patients.
These observations are summarized in Table 7.3.
If hyoscyamine has no effect, we would expect an increase in the amount of sleep
to be as likely as a decrease. (Weare assuming that the distribution of the amount
of sleep is symmetrical about the mean, but we are not assuming that the distribution
is normal.) In other words

P(increase) = P(decrease) = 0.5. 7.4

Under our null hypothesis that the drug has no effect, the observations should come

Table 7.3 The number of patients for which hyoscyamine and

hyoscine increased their amount of sleep, had no effect or decreased
their amount of sleep. The last column gives positive, zero or negative
differences between the amount of sleep gained when using each of
the two drugs

Hyoscyamine Hyoscine Difference

Increase 5 9 9
No effect 1 0 1
Decrease 4 1 0
102 Comparisons
Table 7.4 Binomial probabilities P(k, n) and cumulative probabilities C(k, n) for k successes
in n eQual to 9 and 10 trials

Number of successes, k
0 1 2 3 4 5 6 7 8 9 10

P(k, 9) 0.002 0.018 0.070 0.164 0.246 0.246 0.164 0.070 0.D18 0.002
•
C(k,9) 0.002 0.020 0.090 0.254 0.500 0.746 0.910 0.980 0.998 1.000
•
P(k,10) 0.001 0.010 0.044 0.117 0.205 0.246 0.205 0.117 0.044 0.010 0.001
C(k,10) 0.001 0.011 0.055 0.172 0.377 0.623 0.828 0.954 0.989 0.999 1.000

from a binomial distribution with p = 0.5. Before we proceed, we need to decide

what to do about the one patient who neither gained nor lost sleep when using the
drug. Since the tie provides no evidence one way or the other, the usual procedure
is simply to drop it from the analysis.
We can calculate the binomial probabilities for all possible outcomes of our
experiment using Equation 5.2 and these are given in Table 7.4. We are interested
in the drug only if it increases the amount of sleep. We see from Table 7.4 that if
the drug has no effect, the probability of four or fewer successes in nine trials is 0.5,
so that the probability of five or more is also equal to 0.5 and, on the basis of the
available evidence, we conclude that hyoscyamine has no effect, which agrees with
our previous conclusion.
With five successses and four failures out of nine trials, it should have been obvious
that we would not have a significant effect. Let us therefore try the same test on
hyoscine for which nine out of ten of the patients showed an increase in sleep. We
see from Table 7.4 that the probability of eight or fewer successes in ten trials is
0.989, so that the probability of nine or more successes is only 0.011, and this
outcome is very unlikely. We can therefore reject the null hypothesis at the 5% but
not quite at the 1% level, which again agrees with our earlier conclusion.
We can now proceed to compare the effects of the two drugs. We now have nine
increases and no decreases with one tie, which we ignore (Table 7.3). This time we
need to use a two-tailed test since we do not know in advance which of the two
drugs is the more efficacious. The probability of nine successes in nine trials is equal
to the probability of zero successes in nine trials, which is 0.002. The probability of
a result as extreme as this, in either direction, is therefore 0.004 and we can reject
the null hypothesis and conclude that the two drugs differ significantly at the 1%,
but not at the 0.1% level, once again in agreement with our previous conclusions.
The test described above is called the sign test because it is based on the signs
of the differences between the effect of each drug and no drug and between the two
drugs. Although our conclusions concerning the effects of the drugs were the same
using the parametric t test and the non-parametric sign test, non-parametric tests are
generally less powerful and this is the price that we pay for making fewer assumptions
about the underlying distribution of the data. To confirm this we note that when
 Multiple range tests 103

we compared the two drugs using the two-tailed t test the significance level corres-
ponding to the critical value of 4.08 with 18 degrees of freedom was 0.07%, whereas
the significance level corresponding to zero or nine successes using the sign test was
0.4%. The t test was therefore significant at the 0.1% level while the sign test was
significant only at the 1% level. Nevertheless, it is often the case that a suitable non-
parametric test is almost as powerful as the corresponding parametric test and since
non-parametric tests make fewer assumptions about the data. they also give more
robust conclusions.
Although we will not discuss non-parametric tests further in this book, there are
non-parametric tests corresponding to all of the standard parametric tests. Non-
parametric tests are especially important in the social sciences where the data collected
often take the form of answers to questionnaires and may be nominal categories that
cannot be ranked, such as single, married, widowed or divorced, or ordinal categories
that can be ranked but not on a quantitative scale, such as very bad, bad, good or
excellent. You are also likely to encounter non-parametric tests in biology where an
animal's behaviour may again be categorized but not on a quantitative scale. In these
cases, tests may be carried out using only signs of differences, as in the examples given
above, they may take advantage of ranking in the data, they may test runs of positive
and negative values to ensure that series of data are random and not clumped. The
important thing to bear in mind is that if the data are not measured on a quantitative
scale or if you susped that the distribution of the data does not follow a known,
usually normal, distribution, you should use a non-parametric test. Good references
for further reading are the books by Siegel and Castellan (1988) and Sprent (1990).

7.4 MULTIPLE RANGE TESTS

If we are comparing only two numbers, we proceed just as in the previous section.
However, if we are comparing many numbers, we encounter a problem. Consider,
for example, the housefly data of Table 3.1 in which the longest wing was 5.57mm
and the shortest 3.63 mm. The difference in the lengths of these two wings is 1.94 mm,
while the sample standard devition~ is 0.39 mm. The standard deviation of the
difference is therefore 0.39 x )'2, so that the t statistic is 3.51 with 99 degrees of
freedom. If our hypothesis is that both measurements come from the same normal
distribution with a population standard deviation of 0.39, we should reject the
hypothesis at the 0.1% significance level!
It should be immediately clear that I have cheated by choosing the shortest and
the longest wings: if I had chosen two wings at random, the comparison would have
been valid. Of course, I might have picked out the longest and shortest wings
even if the choice had been random, but the probability of that happening is only
(2/100) x (1199) = 2 x 10- 4, and we know that we cannot exclude the possibility
of a Type I error completely. Nevertheless, if we do make several measurements
there is no reason not to compare all of them; we simply have to be more cautious
in the analysis.
There are a number of ways in which one can deal with the problem of multiple
104 Comparisons
comparisons. Sokal and Rohlf (1981, p.261) list ten different tests, along with the
conditions under which you should use them. Here I consider the problem briefly.
If, before looking at the numbers, we decide to compare only two of them, we
simply use the t test as before. However, we see that if we have 100 numbers and
make all possible comparisons, we are bound to find spuriously significant differences
when we compare the largest with the smallest numbers. To avoid this problem we
can show (Appendix, section 7.10.3) that if we measure n random variables, and if
the probability that anyone of them is less than some value, say x, is ex, then the
probability that all of them are less than x is approximately ex/no Comparing numbers
in pairs is a slightly different problem from that of testing a series of numbers
separately, but this result suggests that if we make k comparisons and want an overall
significance level of rx, say, we should use the acceptance range corresponding to a
significance level of ex/k. For ten comparisons at the 1% significance level, we would
use the acceptance range for the 0.1% significance level so that for the normal
distribution our 1% acceptance range would be ± 3.29 instead of ± 2.58 (Table 10.1).
Now if we make n measurements, we can make at most k = n(n - 1)/2 comparisons.
For this reason Fisher (Kendall et al., 1983) suggested that if we want to make all
possible comparisons between n numbers and ensure that the significance level for
each comparison is at least ex, we should use a significance level of ex/[n(n - 1)/2].
For our houseflies we can make up to 100 x 99/2 = 4950 comparisons. Testing
the difference between the biggest and the smallest numbers at the 5% significance
level, we than use the acceptance range for the 5/4950 = 0.001% significance level
with 99 degrees of freedom (the number of degrees of freedom in the calculation of
the standard deviation). At this significance level t99 = 4.5 and the calculated t statistic
for the difference between the lengths of the longest and shortest wings, 3.51, is
well within the acceptance range. We accept the hypothesis that the difference is
not significant at the 5% significance level.
We now have two extreme cases: the standard t test strictly applies only to a
comparison between two numbers or two numbers chosen at random from a larger
set of numbers and is sometimes called the least significant difference or LSD test.
If we wish to compare more than two numbers, the acceptance range will be too
narrow and it will produce too many Type I errors indicating significant differences
where none exist. On the other hand, the modified t test applies only to a comparison
between the largest and smallest of a set of numbers. If we wish to compare more
than two means, the acceptance range will be too wide, it will produce too many
Type II errors and fail to pick up significant differences.
To illustrate this, Table 7.5 shows the critical values that define the acceptance
ranges at the 5% significance level for the LSD and modified t test (the first and third
number in each triplet) for various numbers of measurements and degrees of freedom.
If we have only two measurements to compare, the critical values are the same for
both tests, as we would expect. But as we increase the number of measurements that
we wish to compare, the modified t test becomes progressively more conservative
by comparison with the LSD test since a bigger value of t is required before we can
reject the null hypothesis.
 Multiple range tests 105

Table 7.5 Critical values at the 5% significance level for the difference between the biggest
and the smallest of a set of normally distributed numbers. The number of comparisons that
are made is n and the number of degrees of freedom used to calculate the standard deviation
is f. The first number in each triplet is based on the LSD or standard t test and isj2t/O. 97 5). The
second number in each triplet is based on the Studentized range statistic q(n, f). The third
number in each triplet is based on the modified t test and is j2t f(1 - 0.025/k) where
k= n(n -1)/2

Degrees of Number of measurements, n

freedom, f 2 5 10 20

18.0 18.0 18.0 18.0

1 18.0 37.2 49.1 59.6
18.0 180 810 3421

3.64 3.64 3.64 3.64

5 3.64 5.67 6.99 8.21
3.64 6.74 9.49 12.9

3.15 3.15 3.15 3.15

10 3.15 3.58 4.52 5.50
3.15 4.66 5.60 6.47

2.77 2.77 2.77 2.77

20 2.77 3.86 4.39 5.01
2.77 3.97 4.61 5.16

The modified t test is only approximate, even when testing the difference between
the largest and the smallest numbers from a set of numbers. However, we can calculate
the so-called Studentized range statistic, q(n, f). which is the actual distribution of
the difference between the largest and the smallest of two numbers taken from a set
of n normally distributed numbers, where f is the number of degrees of freedom
used to calculate the standard deviation. The test based on the Studentized range
statistic is called Tukey's honestly significant difference, or HSD, test (Hays, 1988,
p. 420) and critical values are given by Hays (1988, p. 949). In Table 7.5 the middle
number in each triplet gives the critical value for the HSD test and this lies between
the critical values for the LSD test and the modified t test. Provided the number of
degrees of freedom is greater than the number of points that we wish to compare,
the three tests do not differ greatly. When the number of degrees of freedom is small
and the number of points to be compared is large, they do differ greatly but it would
be a bad experiment that compared, say, 20 measurements on the basis of a standard
error calculated with only 1 degree of freedom.
There are two important points to be made here. If two numbers do not differ
significantly using the LSD test they will not differ significantly on any test. Equally,
if two numbers do differ significantly using the modified t test they will differ
significantly on any other test. It is only when we find marginally significant differences
106 Comparisons
using the LSD test or marginally non-significant differences using the modified t test
that we need to exercise caution.
The second point is that even the HSD test based on the Studentized range statistic
is exact only when comparing the largest and the smallest of two numbers from a
given set of numbers. We would expect numbers that lie close together in the ordered
distribution to differ by less. To allow for this, we can use a variation on the HSD
test called the Student-Newman-Kuels test (Hays, 1988, p.420), in which we first
compare the largest and the smallest of n numbers using the Studentized range
statistic as described above. If, and only if, they differ significantly, we then compare
the two numbers that are one step closer together in the ordered sequence using the
Studentized range statistic based on k - 1 comparisons. We then proceed inwards
in this way, stopping as soon as we find the first non-significant difference.

7.4.1 Prussian army officers

Let us look again at the deaths of Prussian army officers in Table 4.1. We still feel
uneasy about army corps number 9 in which 24 officers were kicked to death by their
horses. Table 7.6 gives the number of officers kicked to death in each corps, summed
over all 20 years. To see if army corps number 9 did have significantly more deaths
than the others, our null hypothesis is

H: The number of deaths in each army corps

follows the same Poisson distribution. 7.5

The mean number of army officers kicked to death in all of the corps is 12. The
probability that a number, chosen from a Poisson distribution with a mean of 12, is
greater than or equal to 24 is 0.16%. The number of deaths in army corps 9 therefore
appears to be significant at the 1% significance level. However, we have deliberately
chosen the greatest value and since we have ten numbers we are effectively making
ten comparisons. To test the largest number at a significance level equal to P, we
need to use the critical value for a significance level of PlIO. The probability that the
largest of ten numbers chosen from our Poisson distribution is greater than 24 is
therefore equal to 0.16 x 10 = 1.6%, so that we can reject Hypothesis 7.5 at the 5%
level but not at the 1% level.
Once again, we are left to decide what it means. There may well have been a
problem with safety in army corps number 9. However, army corps number 9 might
have been bigger than the others, had more cavalry than the others or simply been
more active than the others. The statistics raise a question: we have to provide the
answer.

Table 7.6 The number of army officers kicked to death in each of ten army corps between
1875 and 1894

Corps I 2 3 4 5 6 7 8 9 10
Deaths 12 12 8 II 12 7 13 14 24 8
 Confidence limits 107

7.5 VARIANCE RATIOS

So far we have concentrated on comparisons between means and we have assumed

that the variance of the sets of numbers used to calculate the means are the same.
But suppose we want to compare two variances. We saw in section 5.4.3 that since
each variance follows a X2 distribution, the ratio of two variances will follow an F
distribution so that

V r! V2 "" Fn,-I,nz-I 7.6

where nI - 1and n2 - 1are the number of degrees of freedom associated with VI and
V2, respectively (Appendix section 7.10.4).
We can illustrate this with data collected to investigate the basis of speciation. Mayr
(1942) suggested that a new population that starts from a small group of founders
will be genetically atypical of the parent population and thus more likely to give
rise to a new species than a new population starting from a large group of founders.
To test this hypothesis, Dobzhansky and Pavlovsky (1957) took 20 populations of
fruit flies: ten having 5000 individuals each and ten having 20 individuals each. At the
start of the experiment, 50% of the individuals in each population were homozygous
for a gene P, 50% were heterozygous. At the end of the experiment, after 18 genera-
tions, the proportions of flies homozygous for the gene P were as given in Table 7.7.
The mean proportion of homozygous flies in the large populations is 0.26 with vari-
ance 0.0032, while for the small populations the mean is 0.32 with variance 0.0127. The
ratio of the two variances is 3.96 and we need to compare this with an F distribution
with 9 and 9 degrees of freedom. From Tables 10.3 and 10.4 we see that the variances
differ significantly at the 5% but not at the 1% level, giving support to Mayr's theory.

7.6 CONFIDENCE LIMITS

In the analyses developed so far we have (i) made an hypothesis, H, about the data
(e.g. two means are the same), (ii) made a series of measurements from which to
calculate a test parameter, T, (e.g., the difference between the two means divided by
the standard deviation), (iii) determined a range, R, within which T should lie if H is
true (e.g. 95% acceptance range for a normally distributed variable), (iv) accepted H
if T is contained in R; rejected H if T is not contained in R.
It is important to be clear as to what we mean when we accept or reject H. When
we accept an hypothesis, we are not saying that it is necessarily true but only that
we have failed to show that it is false; more accurate data may well lead us to reject
the hypothesis. Although we accepted the null hypothesis that hyoscyamine does

Table 7.7 The proportions of flies bred from large and small initial populations that were
homozygous for the gene P after 18 generations

Large 0.187 0,202 0,208 0.216 0.244 0.276 0.306 0.312 0.324 0.338
Small 0.153 0.172 0.208 0.306 0.321 0.333 0.440 0.306 0.450 0.462
108 Comparisons
not increase the amount of sleep significantly (section 7.2.1), more extensive data
might reveal a significant difference and lead us to reject the null hypothesis.
We are on firmer ground when we reject, rather than accept, an hypothesis, and
we can be confident that the hypothesis does not hold. In the trial of the polio
vaccine, we were able to show that the vaccine produces a statistically significant
reduction in the number of children contracting polio. Nevertheless, the data in
themselves do not tell us why the null hypothesis is false or what would be a
reasonable alternative hypothesis. In the study of children in Saxony, we decisively
rejected the hypothesis that the data follow a binomial distribution in which the
probability of having a boy or a girl is independent of the sex of its older siblings.
However, the observed numbers deviate only slightly from the binomial prediction
and any further hypothesis should take the binomial model as a starting point.
Considering hypotheses in this manner is both rigorous and sound, but rigour
alone leads to pedantry, not progress. For example, in our study of the effect of
hyoscine on sleep we found that the average sleep gained was 2.33 ± 0.63 hours so
that m/ S was about 4 and we concluded that the number of hours of sleep gained
differed significantly from zero and that the hyoscine does induce sleep. But we
should be able to say more than this. For example, I think we agree that it is unlikely
that hyoscine will increase the amount of sleep by, say, 10 hours or that it decreases
the amount of sleep. What we want is an idea of the limits within which we can be
confident that the 'true' value lies. Let us see if we can define confidence limits for
our estimates of various parameters so that we can regard the sample mean, for
example, as our 'best' estimate of the underlying population mean and use the standard
deviation as an indication of the range within which the true mean should lie.
Setting this out formally, suppose that we make a series of n measurements on a
variable that is normally distributed with true mean J-l, estimated mean m and estimated
standard deviation of the mean equal to Sm' Now we have seen (Equation 7.2) that

t = (m - J-l)/sm 7.7

follows a t distribution with n - 1 degrees of freedom, so that if t" is the upper

97.5% point of the distribution,

P( - t" < t < n = 0.95, 7.8

and there is a 95% probability that t lies in the range ±

t". We use the critical
value for a two-tailed test because we do not know if m is less than or greater than
J-l. Now if t < t", then (m - J-l)/sm < t" and J-l > m - t"sm' Similarly, if - t" < t, then
J-l < m + t"sm' so that
P(m - t"sm < J-l < m + t"sm) = 0.95, 7.9

which tells us that there is a 95% probability that the range m ± t"sm includes
the true mean. We can assert that the amount of sleep gained after taking hyoscine
is 2.33 ± 1.43 hours, that is between 0.90 and 3.76 hours, with 95% confidence.
In our study of the polio vaccine (section 7.2), we found that the difference between
the number of vaccinated and unvaccinated children who contracted polio was
40.7 ± 6.0, and we are able to reject the hypothesis that the vaccine has no effect
 Several factors 109

at the 0.1% significance level. We can now use our concept of confidence limits to
say instead that the reduction in the number of children per 100 000 who will develop
polio when given the vaccine is 40.7 ± 1.96 x 6.0, that is, in the range 29 to 52 per
100 000, with 95% confidence. We can also say that this number lies between 21
and 60 per 100 000, with 99.9% confidence.
Expressing our results in terms of confidence limits appears to be much stronger
than merely accepting or rejecting an hypothesis. However, we have to be care-
ful about the interpretation of confidence limits. We cannot strictly conclude from
Equation 7.9 that the true mean lies within the defined confidence limits with the
calculated probability because we know that it either lies within the limits with
certainty or outside the limits with certainty. What we can do is simply assert that
it lies within the given interval and we will then be right in 95% of cases and wrong
in 5% of cases, for 95% confidence limits. Alternatively, we can note that we are
effectively using our confidence limits to test all possible hypotheses about the value
of the mean and the confidence limits then correspond to those values we would
accept while excluding all those values we would reject. Regarding the polio study,
where we found a difference in the number of children, per 100 000, contracting
polio of between 29 and 52, with 95% confidence, we are saying that we would
reject the hypothesis that the true difference is 0, 1, 2, ... , 27, 28 or 53, 54, ... at
the 5% significance level but we would accept the hypothesis that the true difference
is 29, 30, ... , 51, 52 at the 5% significance level. (Bulmer, 1979, provides an excellent
discussion of this topic but see also Kendall and Stuart, 1983.)

7.6.1 The importance of 'three'

The discussion of confidence limits leads to an important observation. Whenever

you carry out an experiment or make an observation, you should always take at
least three readings. One reading gives us an estimate of the mean but no indication
of the dispersion. Two readings enable us to calculate the standard deviation, and a
95% confidence interval for the mean is then

m ± tt(0.975)slji =m ± 12.75. 7.10

Two repeats is thus the absolute minimum number of readings you should take.
However, with three repeats the confidence interval for the mean is

m ± t2(0.975)slj2 = m ± 4.3slj2 = m ± 3.05. 7.11

With 50% more effort, we have increased the precision of our estimate by a factor
of 4! If we take still more readings, we improve the accuracy further so that for four
readings we have m ± 1.65 and for five readings we have m ± 1.15. None of these
gains, however, is as striking as that made in going from two to three readings.

7.7 SEVERAL FACTORS

In many experiments the response variable that we measure depends on several

different factors and we want to examine the effects of each factor separately as well
110 Comparisons
Table 7.8 The weight (in grams) gained by male and female rats fed
on diets containing fresh and rancid lard

Fresh Rancid
171 108
Male 172 89
172 69
153 85
Female 109 64
160 82

as the interactions between them. We can illustrate this using data from an experiment
to discover whether rancid lard has a lower food value than fresh lard. Because rancid
lard contains strongly oxidizing peroxides and vitamin A is readily oxidized, Powick
(1925) felt that the consumption of rancid lard might cause rats to suffer from a
deficiency of vitamin A. Three young male and three young female rats were fed
on a diet that included fresh lard and the same number of rats were fed on an
equivalent diet but using rancid instead of fresh lard. To minimize the effects of
variables other than those arising from the difference in the diets, the rats were
selected for uniformity as to age, weight and general health. The rats each weighed
between 39 and 52 g at the start of the experiment. Table 7.8 gives the weight gained
by each rat after 73 days.
In this experiment we will treat each group of three rats as repeats of the measure-
ment of the effect of sex and lard freshness on the weight gained by the rats.
Everything we need to know about the data in Table 7.8 is contained in the mean
and the standard deviation of each set of three numbers and these are given in
Table 7.9. We want to use these data to answer the following questions:

• Does the freshness of the lard affect the weight gain of the rats?
• Does the sex of the rats affect the weight gain of the rats?
• Does the effect of lard freshness on the weight gains depend on the sex of the
rats? Or, putting it the other way round, does the effect of sex on the weight
gains depend on the freshness of the lard?

The first question is then: does the freshness of the lard affect the weight gained
by the rats? We could answer this simply by comparing the average weight gained

Table 7.9 The mean and standard deviation of the weight (in grams)
gained by male and female rats fed on a diet containing fresh and
rancid lard

Fresh Rancid

Male 171.7 ± 0.33 88.7 ± 11.25

Female 140.7 ± 15.96 77.0 ± 6.56
 Several factors 111

by all of the rats that were fed on fresh lard with the average weight gained by all
of the rats that were fed on rancid lard. This is precisely what we would do if we did not
know the sex of the rats and we would use an unpaired t test to do it. However,
we have seen that, whenever possible, we should do a paired t test as this is always
more powerful. Unfortunately, in the example of our rats we are unable to pair them
directly since each rat was fed on rancid lard or fresh lard but not both. However,
we can pair the rats by sex so that we consider the effect of lard freshness on weight
gain for male rats and the effect of lard freshness on weight gain for female rats
separately, and then calculate the effect of freshness averaged over the effects on
males and females.
From Table 7.10 we see that the male rats gain 83.0 g more weight when the lard
is fresh rather than rancid and that the standard deviation of this difference is ± 11.3 g
(remembering that the variance of a difference is the sum of the variances). Similarly,
the female rats gain 63.7 ± 17.3 g more when the lard is fresh rather than rancid. In
both cases the rats gain more weight when fed on fresh rather than rancid lard and
the average increase is 73.0 ± 10.3 g.
Before we perform t tests to determine the effect of freshness and sex on the
weight gain of the rats we need to decide how many degrees of freedom to use.
Since each mean in Table 7.9 is calculated using 3 points, each standard deviation
in that table has 2 degrees of freedom. When we compare the weight gained by
male rats eating fresh lard and rancid lard we therefore have 2 + 2 = 4 degrees of
freedom, and of course the same applies to the comparison for female rats as well
as to comparisons between the sexes fed on fresh lard and on rancid lard.
For the male rats then our t statistic for the effect of freshness on weight gain is
83.0/11.3 = 7.4 and using a two-tailed t test this is significant at the 1% level but
not at the 0.1% level (Table 10.1) and has two stars. For the female rats we see from

Table 7.10 The effect of sex and freshness on the weight gain of rats. F - fresh lard, R - rancid
lard, m - male rats, f - female rats. The first line, for example, gives the average difference
between the weight gained when eating fresh as opposed to rancid lard for male rats

Lard freshness
(F - Rim> = 83.0 ± 11.3 t4 = 7.4"
(F - RIO = 63.7 ± 17.3 t4 = 3.7'
«F - Rim> + (F - RIO> = 73.3 ± 10.3 tB = 7.1'"
Sex
(m - flF> = 31.0 ± 16.0 t4 = 1.9ns
(m - fiR> = 11.7 ± 13.0 t4 = 0.9ns
«m - flF> + (m - fiR> > = 21.3 ± 10.3 ts = 2.1 ns
Effects of sex on effect of freshness
(F - Rim> - (F - RIO = 19.3 ± 20.6 ts = 0.94 ns
Effect of freshness on effect of sex
(m - flF> - (m - fiR> = 19.3 ± 20.6 ts = 0.94 ns
112 Comparisons
the data summarized in Table 7.10 that the effect of freshness on weight gain is
significant only at the 5% level and has one star.
To determine the overall effect of lard freshness, we average the effect for males
and females and this gives us an average weight gain when the lard is fresh of
73.3 ± 10.3 g. We are now calculating the average of two numbers each of which
has 4 degrees of freedom so that the number of degrees of freedom for this comparison
is 8. Our t statistic is 73.3/10.3 = 7.12 and this is significant at the 0.1% level and
has three stars. (If we had not distinguished between the male and female rats, the
mean increase in consumption would have been the same, i.e. 73.3 g, but the standard
deviation would have been 26.4 g instead of 10.3 g. Using a two-tailed t test with
10 degrees of freedom, the value of t = 2.78 would have been significant at the 5%
but not at the 0.1% level. Although the effect of the sex of the rats is not in itself
significant, by allowing for the small effect of sex, we obtain a more sensitive test
of the effect of lard freshness.)
We can now perform the same set of calculations for the effect of sex on the
consumption of lard. The results of these calculations are also summarized in Table
7.10, from which we see that in fact none of them is significant at the 5% level and
all receive an ns.
The results of these calculations show that the weight gained by the rats is
influenced by the freshness of the lard but not by the sex of the rats. But it is also
possible that the sex of rats interacts with the effect of the freshness of the lard. The
concept of interactions in important but rather subtle. To illustrate what is meant by
an interaction, suppose that male and female rats both gained 100 g more when the
lard was fresh rather than rancid. The difference in weight gain due to lard freshness
would then be independent of the sex of the rats. But if the male rats gained 100 g
more when the lard was fresh while the females gained 100 g less when the lard was
fresh, we would conclude that the effect of freshness does depend on the sex of the
rats and we call this effect an interaction between freshness of the lard and sex of
the rats.
To calculate the interaction between the freshness of the lard and the sex of the
rats, which we write lard x sex, we take the effect of lard freshness on the weight
gain of male rats, which is 83.0 ± 11.3 g, and compare this with the effect of lard
freshness on the weight gain of female rats which is 63.7 ± 17.3 g. The difference
between the effect of freshness for male rats and the effect of freshness for female
rats is then 19.3 ± 20.6 g. Now we are comparing two numbers each with 4 degrees
of freedom so that the number of degrees of freedom in the interaction term is 8
and our t statistic is 19.3/20.6 = 0.94, which is not significant.
The sex of the rats does not affect the change in consumption due to the freshness
of the lard: is it possible that the freshness of the lard affects the change in consumption
due to the sex of the rats? The two numbers we now wish to compare are, from
Table 7.8, 31.0 ± 16.0 g and 11.77 ± 13.1g, from which we find a difference of
19.3 ± 20.6 g, exactly the same as before, so that the interadion term is the same
whichever way we calculate it.
 Experimental design 113
7.7.1 Contrasts

In the previous section we compared the effects of fresh and rancid lard, male and
female rats, fresh lard consumed by male rats with rancid lard consumed by male
rats, and so on. Altogether we made eight comparisons. However, there must be a
limit to the number of independent comparisons or contrasts one can make (Winer,
1971). For example, if we have three numbers-7, 5 and 3, say-the first is greater
than the second and the second is greater than the third. But this means that the
first must also be greater than the third so that we can only make two independent
comparisons.
The maximum number of independent pair-wise comparisons that we can make
is one less than the number of independent means that we are comparing. In our
example of the consumption of lard by rats, we have four independent means (Table
7.9) and we can therefore make at most three independent comparisons. The choice
of which comparisons to make depends on what we want to learn from the data.
We usually choose to compare each 'main effect' (lard freshness and sex) averaged
over the other (sex and lard freshness) and then their interaction (the effect of sex
on the change in consumption brought about by lard freshness or vice versa). If we
choose to make three comparisons, say, we should divide our significance levels by
3 as we discussed in section 7.4 for multiple range tests. In this example, both sex
and the interaction remain non-significant, as they must, while lard freshness is still
significant at the 0.1% level.

7.8 EXPERIMENTAL DESIGN

The data we collect are only as good as the design of our experiments. In her
experiment on rats, for example, Powick was careful to select the rats for uniformity
of age, weight and general health so that these factors would not affect the outcome
of her experiment. In the trials of the polio vaccine, we cannot give the placebo and
the vaccine to the same children and so we have to find children who are likely to
respond in the same way. We would not, for example, give the placebo to boys and
the vaccination to girls because the efficacy of the vaccine might depend on the sex
of the child. Indeed, we would want to give both the placebo and the vaccine to
children from each state, each city and each school in the study area. However, we
also have a moral dilemma; we would not be testing the vaccine if we did not believe
that it was beneficial, so how shall we decide who should be given it and who
should not7 One way to do this is to use a random number generator to allocate
the placebo and the vaccine within each sex in each school. This way each child has
an equal chance of having the vaccine. Finally, to ensure that there is no unconscious
bias, neither the doctors administering the vaccinations nor the doctors who eventually
are to diagnose the patients for polio should know which injections contain the
placebo and which contain the drug.
To illustrate the consequences of not carrying out well-designed trials, Rensberger
114 Comparisons
(1983) describes an operation, used for many years to treat people with angina, in
which the mammary arteries are tied off. Doctors and patients were happy with the
results. When a proper trial was eventually conducted using a control group, the
control group, whose arteries were not tied off, fared better than those whose arteries
were tied off. The operation was quietly dropped.
Another example of how not to proceed can be found in a series of trials in Kenya
in 1988 to test the use of low-dose interferon administered orally for the treatment
of AIDS patients (Pacala, 1990). The tests were carried out without a control group
but the scientists doing the tests claimed striking improvements in the treated patients.
The World Health Organization (WHO) then decided to run trials of their own.
These trials were also carried out without controls in the hope that the Kenyan
findings 'could be readily and immediately confirmed'. Eventually, the WHO conclu-
ded that though some patients appeared to improve, it was impossible to say whether
this was due to the treatment or to the patient's own raised expectations. Moreover,
the interferon used was not prepared in a consistent form. The WHO report concluded
that properly controlled clinical trials should be 'undertaken in accordance with an
appropriate experimental design in which all relevant variables are monitored' (Pacala,
1990). The attempt to cut corners simply meant that two years after the first trial
no useful evidence as to the efficacy of the drug had been obtained. Now a Canadian
team have begun a properly controlled trial to see if the issue can be decided one
way or the other (Brown, 1991).

7.8.1 Crossed and nested designs

Experiments are often much more complicated than those we have considered so far.
In Chapter 8 we will discuss an experiment designed to compare different traps for
catching tsetse fljes in which the catch depends on the trap, the site in which the
trap is placed and the day on which the flies are caught. We will discuss an experiment
to determine the amount of calcium in plants which will be seen to depend on the
individual plants, the leaves within each plant and the samples taken from each leaf.
In Chapter 9 we will consider the growth of Japanese larch trees and this will depend
on the amount of nitrogen, phosphorus, potassium and residual ash. We need to
think carefully about the design of such experiments to ensure that we can assess
the effect of each variable after allowing for the effects of the other variables and,
where possible, to assess the effect of each variable independently of all other
variables.
The design of experiments warrants an entire book in itself and Mead and Curnow
(1983) provide a good introduction. However, the most complicated experimental
designs can be analysed in terms of crossed and nested factors. This is best explained
using examples.
In the trial of the polio vaccine (section 7.2), the children were nested within the
treatment since one set of children received the vaccine while another set received
the placebo. We were unable to do a paired comparison since each child could not
be given both the vaccine and the placebo.
 Experimental design 115

When we used Student's data to compare two soporific drugs using an unpaired
t test (section 7.2.1), the patients were nested within the drugs. However, when we
considered the same data using a paired t test (section 7.2.2), the patients were crossed
with drugs since each patient received both drugs and each drug was given to each
patient. Although the difference between the mean effects of the drugs (unpaired t
test, nested factors) was the same as the mean of the differences (paired t test, crossed
factors) the latter analysis gave a more sensitive test because it separated out the
variation due to patients leaving a smaller residual error.
In Powick's experiments on rats, sex was crossed with lard freshness since male
and female rats were fed on fresh and rancid lard. We were able to assess the effects
of the lard freshness after allowing for the effects of the sex of the rats. The individual
rats, however, were nested within sex and lard freshness. The error term in our
assessment of the effect of lard freshness on weight gain was smaller when we
allowed for the effect of the sex of the rats than it was when we did not (section 7.7).
In Chapter 8 we will analyse experiments involving crossed and nested designs
in more detail. The important thing to remember is that the analysis, and in particular
our assessment of the errors, will depend on whether or not factors are crossed or
nested.

7.8.2 Sample-size determination

As we increase the number of measurements that we make, the standard deviation

of the mean decreases and we are more likely to be able to detect differences between
pairs of means. Can we decide, in advance, how many measurements we should make
in a particular experiment?
Consider again Student's data on soporific drugs (Table 7.2). Although hyoscyamine
increased the mean amount of sleep in ten patients by 0.75 hours, the increase was
not significant at the 5% level and may therefore have been due to chance. Suppose
that we decided to conduct further tests on the drug. How many patients should we
use to decide if the drug has an effect?
We need to know two things: the intrinsic variation in the data, that is, the
population standard deviation, and the smallest effect that we wish to detect. The
larger the intrinsic variation in the data and the smaller the effect we wish to detect,
the more patients we will need to use.
If we let J.l be the increase in the amount of sleep induced by hyoscyamine, our
null hypothesis is

Ho: J.l = 0 hours. 7.12

Now suppose that we are interested in the drug only if it increases the amount of
sleep by at least one hour. Then our alternative hypothesis is

HI: J.l = 1 hour. 7.13

We now let n be the number of measurements that we will make and let x be the
critical value of the increase in the amount of sleep. If the mean amount of sleep
116 Comparisons
gained is less than x we will accept Ho; if the mean amount of sleep gained is greater
than x, we will reject Ho and implicitly accept HI' To ensure that the probability of
making a Type I error is less than, say, 5% we require

x~ 1.655/,~ 7.14

where 5 is the population standard deviation and assuming that the number of measure-
ments is sufficiently large to allow us to use the normal approximation for t. But we
might also want to keep the probability of making a Type II error small. Since the
consequence of making a Type II error (deciding that an effective drug has no effect)
may be less serious than the consequences of making a Type I error (deciding that
an ineffective drug does have an effect), we might decide to keep the probability of
making a Type II error less than 10%. Then we require

I-x ~ 1.285,~ 7.15

since 1- x is the difference between the mean and the critical value if our alternative
hypothesis is true. Rearranging Equation 7.15 gives

x ~ 1-1.285/.~ 7.16

The smallest value of n that ensures that both inequalities 7.14 and 7.16 hold will
be the value for which they become equalities, so that

1.655/ ~ = x = 1- 1.285/.~ 7.17

Our best available estimate of the population standard deviation is the value we
already have, namely 1.8 hours. Solving Equation 7.17 for n we find that n must
be at least 28. In other words, to decide if the drug induces at least one hour of
extra sleep while keeping the probability of making a Type I error less than 5% and
the probability of making a Type II error less than 10%, we need to use at least 28
patients. If the effect of the drug is to increase the amount of sleep by more than one
hour, the probability of making a Type II error will be reduced so that 28 pairs of
patients gives us a conservative estimate of the number of patients that we need to
use. The critical value for our test is obtained from Equation 7.17 with 5 = 1.8 hours
and n = 28 patients which gives x = 0.56 hours or 34 minutes. If the mean amount
of sleep gained using hyoscyamine is less than 34 minutes we accept the null
hypothesis and conclude that the drug does not have a significant effect, otherwise
we reject the null hypothesis and conclude that the drug does have a significant effect.
We can apply similar arguments to our study of the polio vaccine. We want to
determine the number of children we should use. Let us assume that before the trial
we already know that the probability that an unvaccinated child will develop polio
is Po = 0.00057, the value obtained from the unvaccinated children in our clinical
trial (Table 1.1) since we will already have some knowledge of the incidence of the
disease. Then if n children receive the placebo, we expect npo ± J";;p;, children to
develop polio since the number of sick children should follow a Poisson distribution.
 Summary 117
Since the consequences of making a Type I error, and concluding that an ineffective
vaccine is effective, may be quite serious, we might decide to keep the probability
of making a Type I error below 1%. The critical value, x, for our test must therefore be

x ~ npo - 2.33~ 7.18

Under our alternative hypothesis, we expect npI ± ~ children to develop polio.

We might now decide that we are interested in the vaccine only if it reduces the
probability of a child developing polio by at least 50%, so that PI ~ 0.00029. Since
making a Type II error may be less serious than making a Type I error, we might
choose to keep the probability of making a Type II error to less than 5%. Then
the critical value, x, for our test must also satisfy

x~
npI+1.65.~ 7.19

Once again the smallest acceptable value of n will be that for which both inequalities
become equalities so that

npo - 2.33,j;;, = npI + 1.65.~ 7.20

Solving Equation 7.20 for n gives

.;;, = (1.65jp;. + 2.33vPo)/(po - PI) = 289,· 7.21

so that n is equal to 83 000. In other words, we see that while using about 200000
children in each group might seem excessive, it is in fact only slightly conservative
if we are to obtain a reliable test of the vaccine. Further discussions of sample size
determination are given by Healy (1981) and Winer (1971, p. 30).

7.9 SUMMARY

When applying statistical tests to problems in biology, we want to know if a particular

test statistic that we have calculated from our data could reasonably have come from
a certain distribution. To do this we set up an acceptance range and a rejection range.
However, the acceptance range might correspond to the upper part of the distribution
if it is large values that we treat with suspicion or it might come from the lower
part of this distribution if it is small values that we treat with suspicion. In both
cases we will use a one-tailed test but take opposite ends of the distribution for our
rejection range. It might be that we regard both large and small values with suspicion,
in which case we will use both ends of the distribution to define our rejection range.
Effectively we are choosing the rejection range in such a way as to maximize the
power of our test in relation to the alternative hypothesis that we will accept if the
null hypothesis is rejected.
To decide if two means differ significantly, we compare their difference and the
standard deviation of their difference. The standard deviation is only an estimate of
118 Comparisons
the true underlying standard deviation and it was once thought that this meant that
tests could not be perfonned on small samples for which the estimated standard
deviation is quite inaccurate. The importance of Student's discovery was that we do
not in fact need to know the true standard deviations provided we use the t distribution
rather than the nonnal distribution, for this allows for the inaccuracies associated
with our estimate of the standard deviation.
There is, however, a second trap for the unwary. The t test applies to a comparison
of two means. If we make 100 such comparisons and set a significance level of 5%
about five comparisons will appear to be significant even if all the true means are
exactly the same. We therefore need to use a more conservative acceptance range
when we make many comparisons. Unfortunately, there is no universally best way to
do this and I have outlined some of the ways that are commonly used. Of course if the
significance level of the test is much less than 0.1% the test will almost certainly be
significant however we do it. Equally, if the significance level of the test is much
greater than 5% the test will almost certainly not be significant however we do it.
You only need to tread carefully if the calculated value of the significance level is
close to the range from 5% to 0.1% or if the number of comparisons that you make
is very large.
The idea of paired and unpaired tests or crossed and nested designs is important
because a paired test or a crossed design allows us to estimate the effect of one
factor after allowing for the effects of another. Of course if we put in the wrong
biology, and assume for example that the patients are paired when they are not, we
will not have a meaningful answer. The biology is also important when deciding
between parametric and non-parameteric tests since every test that we carry out
involves a set of assumptions about the underlying distribution of the data. If we
have reason to believe that these assumptions do not hold, it will be better to use a
less powerful test that does not rely on these assumptions. As long as you understand
the biology well you should be able to decide which test is appropriate.
The concept of confidence intervals provides us with a further powerful analytical
tool because we do not only want to say a certain measurement is significant, but
also want to set limits on the range of values within which we believe the true value
lies. As long as we are clear as to how to interpret the confidence interval, there
should be no problems in using it.
This and the previous chapter contain many of the most important concepts in
the statistical analysis of data, but so far we have restricted our attention to fairly
simple experimental designs. As the design of the experiment becomes more compli-
cated, the number of comparisons that we can make, including the effects of inter-
actions among the various factors in our experiments, increases alanningly. In the
next chapter we will discuss analysis of variance, or ANOV A, which provides us
with a systematic way of analysing the results of more complicated experimental
designs.
Finally, we have seen that although the power of a test is seldom used in reporting
the results of our experiments, it provides an important tool in helping us to decide
on how big a sample we need for a particular experiment.
 Appendix I 19
7.10 APPENDIX

7.10.1 Derivation of student's t test

Suppose we have n measurements, Xi' i = 1, 2, ... , n from a normal distribution with

true mean j.l and true variance (J2 so that

Xi ,.", N(j.l, (J2). 7.22

Then
(Xi - j.l)/(J"'" N(O, 1) 7.23
and
:£(Xi - j.l)2/(J2,.", Xn2. 7.24

When we calculate the standard deviation, we use the estimated mean m, rather than
the true mean, j.l, and this reduces the number of degrees of freedom in the sum by
1 so that
:£(Xi - m)2/(J2,.", X"-l
~ 7.25
We also know that

m ,.",N(j.l, (J2/n), 7.26

so that
(m - j.l).J;z/(J"'" N(O, 1). 7.27

Dividing the left-hand side of Equation 7.27 by the square root of the left-hand side
of Equation 7.25 over n - 1, the number of degrees of freedom, we have

(m - j.l).J;z [(J2(n _1)]1/2

---- X -----,.", t 7.28
(J [:£(Xj _ m)2]l/2 n-I

since, by definition, the ratio of a normal variate to the square root ofaX2 variate
divided by the number of degrees of freedom follows a t distribution (Equation 5.33).
But, Sm' the standard deviation of the mean, is
_ :£(Xj - m)2 1/2
Sm - ---- 7.29
[ ]
n(n - 1)

so that
(m - j.l)/Sm"'" tn_I' 7.30

The importance of this result is that we do not need to know j.l, the true value of
the mean, since that will be determined by the hypothesis under test. But (J cancels
in Equation 7.28, so we do not need to know its value either.

7.10.2 t ts~t for samples of unequal variance

If the true variances or the number of measurements contributing to the means of

two independent samples are not equal, the parameter
120 Comparisons
t = (mI - m2)/s 7.31
with
S2 = s/n
~I + ~ /n2 7.32

does not strictly follow a t distribution. However, Satterthwaite (1946) has

shown that if we set Vi = s/ni'
~ the t statistic given by Equation 7.31 is approximately
distributed as tn' where the number of degrees of freedom is

n = (VI + v2)2/(v/n~ I + v/n~ 2). 7.33

Note that when VI = V2 and nI = n2, the effective number of degrees of freedom is
nI + n2 - 2, as expected. However, if VI « V2, n = n2 - 1. In this case mI is effectively
known exactly and we are comparing m2 with a fixed constant so that the appropriate
number of degrees of freedom is n2 - 1.

7.10.3 Multiple range tests

Suppose that we have a set of n observations Xi and an acceptance range R corres-

ponding to a significance level IX for each observation. Then for anyone number,
Xi' chosen at random

P(Xj is in R) = 1 - IX. 7.34

Using the law of multiplication

P(all Xi are in R) = (1 - lX)n. 7.35

Now if the biggest number is contained in R, all of the numbers must be contained
in R so that
P(xmax is in R) = (1 - IXt, 7.36
and
P(xmax is not in R) = 1- (l-lXt. 7.37

For small values of IX, the right-hand side of this equation is approximately equal to
nIX. Since P(xmaxis not in R) is the significance level IX·, we have

IX· = nIX or IX = IX·/n, 7.38

and for n comparisons we should use a significance level of IX·/ n to ensure that all
comparisons are significant at a significance level of IX·.
With k numbers we have at most k(k - 1)/2 independent pairs of differences since
we can compare the first number with (k - 1) others, the second with (k - 2) others,
since we have already compared it with the first, and so on. But

(k - 1) + (k - 2)... 1 = k(k - 1)/2, 7.39

and we see that to obtain the acceptance range at a significance level of IX for the
difference between the largest and the smallest of a set of k readings, we should use
the acceptance range for a single comparison at a significance level of IX·/n where
n = k(k - 1)/2.
 Exercises 121
7.10.4 The ratio of two variances

Consider a set of measurements from a normal distribution with true mean m and
true variance (12. As in Appendix, section 7.10.1

:E(Xj - m)2/(12 '" X_~ r 7.40

The left-hand side of Equation 7.40 is (n - 1)52/(12, so that 52/(12 is a X2 variate

divided by (n - 1). We have also seen in section 5.4.3 that the ratio of two X2
variates, each divided by their number of degrees of freedom, defines an F variate
so that

(S;/)/(S;/(1;)
~ '" Fn,-I,nz-I 7.41
where nl - 1 and n2 - 1 are the number of degrees of freedom associated with 51
and 52' respectively. Since our null hypothesis when we cQmpare variances is that
(11=(12'

S;/5;",Fnl-l,nz-l' 7.42

7.11 EXERCISES

1. In Student's study of the effects of hyoscyamine and hyoscine (Table 7.2), we

used t tests to decide if the drugs were efficacious. Would our conclusions have been
different if we had used a normal distribution rather than t distributions to determine
the critical values for each significance level?

2. Determine what the significance levels in Table 7.10 would have been if we had
used the normal distribution instead of the t distribution.

3. The scutum widths of ticks collected from four cotton-tail rabbits were measured
and the data are given in Table 7.11 (Sokal and Rohlf, 1981, p. 211). Using Satterthwaite's
rule (Appendix, section 7.10.2, Equation 7.33) to determine the appropriate number
of degrees of freedom, carry out t tests to decide which, if any, of the mean scutum
widths differ significantly. Make the comparisons with and without a correction for
the number of comparisons (section 7.4).

4. In Table 7.9 the standard deviation of the weight gained by male rats feeding on

Table 7.11 Number of ticks, mean scutum width in microns and

standard deviation of the mean width in microns, taken from each of
four rabbits

Rabbit n m 5m

1 8 372.3 6.23
2 10 354.4 2.60
3 13 355.3 3.77
4 6 361.3 2.47
122 Comparisons
fresh lard is 0.33 g while the weight gained by female rats feeding on fresh lard is
15.96 g. Our analysis assumes that the variances do differ significantly. Carry out an
F test to see if this assumption is justified. When deciding on the significance level.
remember that you are comparing the largest and the smallest of four numbers. (F2,2
at the 0.42% significance level is 237 and at the 0.083% significance level is 1200.)

5. How many patients would you need to use in order to decide if hyoscyamine
increases the amount of sleep by at least 30 minutes while keeping the probability
of making a Type I error to less than 5% and the probability of making a Type II
error to less than 10%7

6. Our analysis showed that a trial of the polio vaccine designed to detect a reduction
of at least 50% in the number of paralytic cases required 83 000 children in each
group if the probability of making a Type I error is to be less than 1% and the
probability of making a Type II error is to be less than 5%. How many children would
you need to include in your trial if you wanted to keep the probability of making a
Type II error less than 1%7
8
Analysis of variance

... Likeevery other human endeavor [logic]isjust a patchwork quilt whose patches do
not meet very well, and which are continually being tom up and restitched.
I. Hacking (1979)

In Chapters 6 and 7 we tested hypotheses about sets of data by comparing means

and using the standard deviation as a measure of their precision. For example, the
number of children contracting polio was significantly reduced when they were given
the Salk vaccine and patients given hyoscine slept for significantly longer than patients
given hyoscyamine. Rats given fresh lard gained significantly more weight than rats
given rancid lard. This is a sound and reliable way to proceed. However, we have
already seen that when we have several treatments each with several levels, such as
male and female rats feeding on fresh and rancid lard, the analysis rapidly becomes
complicated.
Let us consider an alternative approach. We have an experiment in which we apply
several treatments, each at several levels. We know that even if none of the treatments
we apply has a significant effect there will still be some variation in our data associated
with each level of each treatment due to the random nature of the sampling and
measurement process. In this chapter we will show that we can partition the variation
of all the data about the grand mean into contributions associated with each treatment
factor plus a residual term. By comparing the variation that we attribute to each
treatment factor with the residual variation we will be able to decide which factors
contribute significantly to the overall variation of the data.
In our experiment on rats, for example, we will consider the freshness of the lard
and the sex of the rats in turn and see how much of the observed variation in the
weight gain we can explain away using each factor. Instead of asking: 'Do rats gain
more weight when the lard is fresh than they do when it is rancid?' we ask instead:
'How much of the variation in weight gain can be explained on the basis of the
freshness of the lard?' If very little of the variation can be accounted for using lard
freshness, we conclude that lard freshness does not influence the weight gain of the
rats significantly. If much of the variation can be accounted for using lard freshness,
we conclude that lard freshness does influence the weight gain of the rats significantly
and we then return to the kind of analysis we carried out before and ask: 'Do male
rats gain significantly more weight than female rats?' Do rats gain more weight when
124 Analysis of variance
the lard is fresh or when it is rancid? What the analysis of variance, or ANOV A,
does for us is to test each factor and combination of factors (freshness of lard, sex
of rats, the interaction between then), enabling us to pick out the significant factors for
further consideration. If the ANOV A indicates that certain factors are significant, we
then compare the different levels within each treatment factor using the kind of
analysis developed in Chapters 6 and 7.
In this book I have not given the computational formulae needed to carry out
ANOV As. The calculations are specific to each experimental design and are given
in most of the standard books on statistics. The calculations given in this and the
next chapter can be carried out using standard statistical packages and you should
do them yourself as you read the book.

8.1 ONE-WAY AN OVA

The central role that ANOV A plays in the analysis of biological data is due to the
fact that the sums of the squares of the residuals in suitably designed experiments,
calculated without regard to the various treatments or factors involved, can be broken
down into a sum of terms, one for each factor involved. To illustrate this we will
begin by examining the simplest possible experimental design.

S.t.t Analysing the sum of squares

Consider a treatment or factor with two levels, Tl and Tz. We make two measurements
at each level, a and c, at level Tl' band d, at level Tz' as illustrated in Table 8.1. The two
levels might be two different cows, in which case a and c might be two estimates of
the number of ticks on cow 1; band d two estimates of the number of ticks on cow
2. Alternatively, the two levels might be two designs of tsetse fly traps in which
case a and c might be the number of flies caught on two days in trap 1, band d the
number of flies caught on the same two days in trap 2. Or the two levels might be
two kinds of fertilizer and the measurements two estimates of a maize harvest after
the application of each kind of fertilizer.
We proceed as follows.

• First we calculate the overall mean and the overall or total sums of squares, 55T,
of the deviations about the overall mean for all of the observations, ignoring the
fact that we have two treatments.
• We then partition the total sums of squares into a part due to differences between
treatments, 55B (arising from differences between cows, traps or fertilizers) and a
part due to differences within treatments, SSw, (arising from variability in our
sampling of ticks, flies or maize).
• For each sum of squares we calculate a corresponding variance, MSa and MSw.
• If the between-treatment variance, M5B, is significantly greater than the within-
treatment variance, MSw, we conclude that the two levels, Tl and Tz' of the
factor or treatment do have a significant effect on the measurements (the cows,
traps or fertilizer differ), but if MSa is less than or about the same as MSw, we
 One-way ANOVA 125
Table 8.1 A model of an experiment in which we apply treatment T at two levels, T! and
T2' and make two measurements of the response, a and c, band d, at each level of each treatment
The table shows the breakdown of the total sum of squares, 55r, into SSw, the sum of squares
within, and 55B, the sum of squares between, groups of measurements. M5 indicates mean
square, d.f. degrees of freedom

T] T2
a b
C d
m! m2 55B
55! 552 55T

CD a>
• 55T = (a2 + b2
+ c2 + d2) - (a + b + c + d)2/4
d.f. = 4- 1= 3 M5T = 55T/3

• 55! = (a2 + c2) - (a + c)2/2

d.f. = 2 - 1 = 1 M5! = 55} II
552 = (b2 + d2) - (b + d)2/2
d.f.=2-1=1 M52 = 552/1
CD ® @
SSw = 55] + 552 = (a2 + b2 + c2 + d2) - (a + C)2/2 - (b + d)2/2
d.f. = (2 - 1) + (2 - 1) = 2 M5w = (55! + 552)/2
® @ a>
• 55B= {m;+m;-(m! +m2)2/2} x 2=(a+c)2/2+(b+d)2/2-(a+c+b+d)2/4
d.f. = 2 - 1 = 1 M5B = 55B/l

conclude that the two levels, TI and T2' of the two factors or treatments do not
have a significant effect on the measurements (the cows, traps or fertilizers do not
differ.)
• If the ANOV A indicates that the two levels of the treatment or factors produce
significantly different results, we analyse the data further (compare individual cows,
traps or fertilizers) and decide what the differences mean.

Table 8.1 gives a step-by-step breakdown of the analysis of variance for a single
classification, or one-way, two-level experiment with two repeated measures. The
first calculation is for SST' the total sum of squares for all of the readings about the
mean of all of the readings: this is what we need to account for. To calculate this,
remember that the sum of squares of the deviations about the mean can be written
as (Equation 4'.11)

SS = x
~~ - (~
xl/n. 8.1

There are four measurements and we have used up 1 degree of freedom in calcula-
ting the mean, leaving 3 degrees of freedom. If the treatments have no effect,
126 Analysis of variance
MST, the total mean square, will be an estimate of the variance due to the intrinsic
variations in the readings.
The second calculation is for the sum of squares within each treatment about their
means m1 and mz: 551 for treatment Tl' SSz for treatment Tz and SSw for both
treatments together. Each of these has two measurements and we use up 1 degree
of freedom in calculating each mean, leaving 1degree of freedom for each treatment.
SSw therefore has 2 degrees of freedom and MSw is our best estimate of the within-
group variance.
The third calculation is for SSB' the sum of the squares between the means m1
and mz, of the two groups about their common mean. Since we have two measure-
ments within each group, we multiply by two. This time we have two means and
we calculate their overall mean, leaving 1 degree of freedom. MSB is an estimate
of the between-group variance.
The first thing to note is that

SST = SSw + SSB' 8.2

We can see this directly by substituting the expressions given in Table 8.1 into
Equation 8.2. The terms labelled CD and @ appear on both sides of the equation
and the terms labelled ® and ® cancel. In other words, it is possible to partition the
total sum of squares into a sum of squares within groups and a sum of squares
between groups. (Snedecor and Cochran, 1989, p. 225, give more general proofs of
Equation 8.2 for this and for other experimental designs.)
The second thing to note is that the number of degrees of freedom for SST' 3, is
equal to the number of degrees of freedom for SSw' 2, plus the number of degrees
of freedom for SSB' 1, so that

d.fT = d.f·w + dIB' 8.3

Dividing each sum of squares by the appropriate number of degrees of freedom

gives three variances-MST, MSw and MSB-and if there is no treatment effect,
all three are estimates of the variance due to the intrinsic variation in the data. In
this ANOV A, the hypothesis that there is no treatment effect reduces to a test of
the hypothesis that MSw = MSB.
To be clear about the calculation of the number of degrees of freedom, note that
for a one-way ANOV A with a levels and n measures, we calculate the numbers of
degrees of freedom for each sum of squares as indicated in Table 8.2.

Table 8.2 Calculation of the number of degrees of freedom in

a one-way ANOV A witn a levels each with n measurements.
Note that d.f'T = d. f.w + d. fB

S5r: an points - 1 mean -+ an - 1 d. f.

SSw: n points - 1 mean -+ n - 1 d. f.

Summed over a levels -+ a(n - 1) d. f.

SSB: a levels - 1 mean -+ a-I d. f.

 One-wayANOVA 127
8.1.2 No treatment effects
The observation that all three variances are estimates of the overall variance of the
data, provided there is no treatment effect, is critical. To illustrate this, I generated
18 normally distributed random numbers from N(O,I) and divided them into nine
pairs (Table 8.3). We can now analyse this is as a one-way ANOV A with nine levels
and two measurements per level.
Table 8.4 gives the ANOV A for the 'without treatment' data in Table 8.3 (columns
2 and 3). The sums of squares between and within groups add up to the total sum
of squares and the number of degrees of freedom between and within groups add
up to the total number of degrees of freedom. The sums of squares divided by the
appropriate numbers of degrees of freedom give the mean squares. Since there is no
treatment effect, each gives a reasonable estimate of the true variance, which we
know to be 1.

8.1.3 Treatment effects

What we really want to know is what happens when the treatment does have an
effect. To illustrate this I took the nine pairs of numbers in Table 8.3 (columns 2 and
3), subtracted 4 from the first pair of numbers, 3 from the second pair and so on,

Table 8.3 Eighteen normally distributed random numbers with mean zero and
variance 1 (columns 2 and 3). In columns 4 and 5, 4 has been subtracted from the
first pair, 3 from the second pair ... , 4 has been added to the last pair
Without treatment effect With treatment effect
Treatment 1 2 1 2

1 - 1.02 0.23 -5.02 -3.77

2 -0.24 0.01 -3.24 -2.99
3 2.18 -0.22 0.18 -2.22
4 -0.47 0.19 - 1.47 -0.81
5 -0.74 0.10 -0.74 0.10
6 0.89 1.98 1.89 2.98
7 -0.98 -1.17 1.02 0.83
8 0.23 0.31 3.23 3.31
9 0.42 0.23 4.42 4.23

Table 8.4 ANDVA table for nine pairs of random numbers chosen from N(O, 1)
Source of Sums of Degrees Mean Significance
variation squares of freedom square F-ratio level, P •
Between 6.6 8 0.82 0.65 0.72 ns
Within 11.4 9 1.27
Total 18.0 17 1.06
128 Analysis of variance
Table 8.5 ANOV A for the data in columns 4 and 5 in Table 8.3

Source of
variation
Sums of
squares
Degrees
of freedom
Mean
square F-ratio
Significance
level, P .
Between 132.2 8 16.53 13.0 0.00042 ...
Within 11.4 9 1.27
Total 143.6 17 8.45

adding 4 to the last pair. The data with the 'treatment' effect are also given.in Table 8.3
(columns 4 and 5) and the resulting ANOV A is given in Table 8.5. The between
groups sum of squares has increased dramatically, the within groups sum of squares
has not changed and the residual sum of squares has increased by the same amount
as the between groups sum of squares.
The partitioning of the sums of squares is carried out as before. The treatment
effect will not change any of the within treatments sums of squares since all of the
numbers, as well as the mean, for a given treatment level, will be changed by the
same amount and the sum of squares of the deviations about the mean will not
change. However, we can show (Snedecor and Cochran, 1989) that, on average, the
between groups sum of squares will increase by (a - l)n5; where a is the number
of levels of the treatment, n is the number of measurements made at each level and
5; is the population variance calculated for the effect of the different treatment levels
on the outcome. Since (a - 1) is the number of degrees of freedom associated with
the treatment, the between groups mean will increase by n5;.
(This is discussed further
in section 8.4.6.)
In the example given in Table 8.3 the treatment effect changes the individual levels
by - 4, - 3 ... 2,3,4. The sums of squares for the treatment effect is therefore
16 + 9 + ...4 + 9 + 16 = 60 and since two measurements are made at each level
the sum of squares for the treatment effect should increase by about 120. From Tables
8.4 and 8.5 we see that the actual increase is 125.6. The expected increase in the
between levels mean square is 120/8 = 15 and we see that the actual increase is 15.7.
Finally, we can estimate the population variance of the levels for a given treatment
by dividing the increase in the between levels mean square by two, the number of
repeats per level of treatment. This gives 7.9 which compares well with the actual
value of 60/8 = 7.5. The population variance of the treatment levels is referred to
as the added variance component per level of that treatment.

8.1.4 Testing hypotheses

It is evident that the mean square estimates of the variance in Table 8.4 are approxi-
mately equal, indicating that there is no 'treatment' effect, while in Table 8.5 the
between-groups mean square is so much greater than the within-groups mean square
that we can be confident that in this case there is a significant treatment effect. Once
again, we want to set significance levels to these statements, starting with the null
hypothesis that there is no significant treatment effect.
 One-way ANOVA 129
If there is no significant treatment effect, the between-groups mean square, MSB,
and the within-groups mean square, MSw, are both estimates of the intrinsic variance
of the data, a2, the first with a-I degrees of freedom the second with a(n - 1)
degrees of freedom (Table 8.2), so that

SSw/a2 "" X:<n-

2
SSB/a "" X:- l' 1) 804
and if the null hypothesis holds, MSB/MSw will follow an F distribution with
a-I and a(n - 1) degrees of freedom. We therefore compare MSB / MSw with critical
values for the appropriate F distribution. For our simulated experiment with random
numbers, we have 8 and 9 degrees of freedom. For the data in Table 804, the F ratio
(0.65) is well within the 5% acceptance range for F8,9 (3.23) and we write ns for 'not
significant' in the last column; for the data in Table 8.5, the F ratio (13.0) is outside
the 0.1% acceptance range for F8,9 (lOA) and we give it three stars. Tables 804 and
8.5 also give the significance level corresponding to an acceptance range that just
includes the observed point. Thus in the case of Table 804, an acceptance range from
o to 0.65, for an F distribution with 8 and 9 degrees of freedom corresponds to a
72% significance level, while in Table 8.5 the acceptance range from 0 to 13.0 cor-
responds to a significance level of 0.04%.

8.1.S Means plots

Having carried out the ANOV A and decided that something interesting is going on
in our second set of data, we have to decide what this is. Plotting the means for
each treatment, as shown in Fig. 8.1, it is clear that the mean increases from qne

Figure 8.1 Means plot for the data used in Table 8.7. The error bars indicate ± Sm'
130 Analysis of variance
treatment to the next. If this were real data, we would want to decide why the mean
increases with treatment as it does. When we carry out an ANOV A we start by
testing each factor or treatment to see if it has a significant effect on our observations;
if it does we then examine the individual levels of that treatment to decide how the
levels within that treatment factor differ in their effect. This is more efficient than
calculating the means for each pair of points at each level and then making all possible
comparisons to decide which, if any, differ.

8.1.6 Soporific drugs

Let us reconsider the data analysed by Student on the effect of hyoscyamine and
hyoscine on the amount of sleep gained by ten patients given in Table 7.2. We will
begin (as in section 7.2.1) by ignoring the fact that each patient received both drugs.
The two drugs are then two levels of treatment with ten repeats at each level. This
is a nested design with patients nested in drugs (see section 7.8.1). If we carry out
a single-classification ANOV A on the data, we obtain the results shown in Table
8.6. From the ANOV A table we see that there is no significant difference between
treatments at the 5% level, which is what we found in section 7.2.1 (Equation 7.3)
when we compared the two means using the unpaired t test. Indeed, the value of t
was 1.86, which is the square root of 3.5, the value of F given in Table 8.6, showing
that in a single-classification experiment the t test provides exactly the same answer
as the ANOV A.
Table 8.6 gives both the internal (Sj) and the pooled (sp) standard deviations of
the means. The internal standard deviation is the standard deviation calculated
separately for each of the two treatments and the values in Table 8.6 agree with the
values calculated in Table 7.2. The pooled standard deviation is the root-mean square
value of the two internal standard deviations. Since the within-groups mean square,
3.61, is an estimate of the variance of the response from one patient to the next, we
can also calculate the pooled standard deviation as (3.6/9)112 = 0.60. When comparing
responses under different levels of the same treatment, we use the pooled standard
deviation because this gives us the best estimate of the within-groups variation. (If

Table 8.6 ANOV A table for the two treatments in Student's data of Table 6.6. 5; is the
internal standard deviation and So is the pooled standard deviation

Source of Sums of Degrees Mean Significance

variation squares of freedom square F-ratio level, P •
Between 12.5 1 12.5 3.5 0.079 n5
Within 64.9 18 3.61
Total 77.4 19

Treatment Mean Sj sp 95% c.L.

Hyoscyamine 0.75 0.57 0.60 - 0.51-2.01

Hyoscine 2.33 0.63 0.60 1.06-3.59
 Unequal sample sizes 13 1

the internal standard deviations differ significantly, we have to be more careful and
use Satterthwaite's rule for the t test as recommended in section 7.2.1.)
To calculate 95% confidence limits for the amount of sleep gained, we multi-
ply the standard deviations by the two-tailed value of tIB at the 5% significance
level, 2.10, so that the amount of sleep induced by hyoscyamine, for example, is
0.75 ± 0.6 x 2.10, that is from - 0.51 to 2.01. In Table 8.6 the 95% confidence
intervals for the two means overlap, reinforcing our conclusion that they do not differ
significantly. We also note that the confidence interval for the first mean includes 0,
while that for the second does not, and this is again in agreement with our previous
observation that hyoscyamine does not significantly increase the amount of sleep
while hyoscine does.

8.2 UNEQUAL SAMPLE SIZES

When we have a one-way ANDV A with only a few levels it may be sufficient
to calculate means and standard deviations and then compare the means using a
suitable multiple range test if there is any doubt as to the significance of the various
differences. The real power of ANDV A will be apparent only when we analyse
experiments with more than one treatment or factor. However, it is important to
have a firm grasp of the basic ideas, so we will consider another example of a one-way
ANDV A but with unequal sample sizes.
An experiment was performed in which tick larvae of the species Haemaphysalis
leporispalustris were collected from four cotton-tail rabbits (Sokal and Rohlf, 1987,

Table 8.7 Scutum widths, in microns, of tick larvae taken from four
cottontail rabbits

Rabbit
1 2 3 4

380 350 354 376

376 356 360 344
360 358 362 342
368 376 352 372
372 338 366 374
366 342 372 360
374 366 362
382 350 344
344 342
364 358
351
348
348

m 372.3 354.4 355.3 361.3

sm 6.23 2.60 3.77 2.47
132 Analysis of variance

Table 8.8 ANOV A table for the data of Table 8.8. 'Among' refers to differences
among rabbits and 'within' to differences among ticks on each rabbit

Source of
variation
Sums of
squares
Degrees
of freedom
Mean
square F-ratio
Significance
level, P .
Among 1808 3
------_._-------~---_._._._-----~
603 5.26 0.0045 ..
Within 3778 33 115
Total 5586 36 102

p. 217). The width of the scutum, the dorsal shield, of each larva was measured in
microns and the data are given in Table 8.7. The mean scutum width is about the
same for rabbits 2 and 3, while for rabbit 4 it is a little greater and for rabbit 1 it is
greater still. The difference between the biggest and the smallest mean is 17.9, but
this is less than three times the standard deviation of the mean for larvae from rabbit
4, and we will need to be careful in our analysis. The analysis of variance is shown
in Table 8.8 and Sokal and Rohlf (1987, p. 168) perform the calulation step by step.
The number of degrees of freedom among rabbits is 4 - 1 = 3. Since rabbit 1
has eight ticks, the number of degrees of freedom 'within' rabbit 1 is 8 - 1 and
taking all the rabbits together the number of degrees of freedom 'within' rabbits is
(8 - 1) + (10 - 1) + (13 - 1) + (6 - 1) = 33. The total number of degrees of freedom
is 37 - 1 = 36. Both the sums of squares and the number of degrees of freedom
among and within groups shown in Table 8.8 add up to the corresponding totals. The F
ratio is significant at the 1% level and has two stars.
So far, all that we know is that the mean scutum width of the ticks differ among
the four rabbits. Now we need to consider individual rabbits. If we plot the mean
scutum widths for the ticks on each rabbit together with 95% confidence intervals,

Figure 8.2 Mean scutum width/microns for the larvae from each of four rabbits.
 Model I and model II ANOVA 133

Table 8.9 Comparison of the mean scutum widths from the data in
Table 8.7. The columns of stars and letters indicate which means differ
significantly using an LSD test (section 7.4) at the 95% significance
level

Number (scutum width)

Rabbit of larvae (microns) Groups Groups
2 10 354.4
·• a
3
4
1
13
6
8
355.3
361.3
372.3
·.• a
ab
b

as shown in Fig. 8.2, we see that the confidence limits for the scutum width of larvae
from rabbit 1 does not overlap the confidence limits for rabbits 2 and 3, but does
overlap the confidence limits for rabbit 4.
In addition to plotting the means such data are often presented as a means table
as in Table 8.9. The data are presented in such a way that rabbits sharing a star in
the same column do not differ significantly in the scutum widths of the tick larvae
taken from them. The first column of stars shows that the mean scutum widths of
larvae on rabbits 2, 3 and 4 do not differ, the second shows that the mean SClitum
widths of larvae on rabbits 4 and 1 do not differ, but since rabbit 1 does not share
stars with rabbits 2 or 3, the mean scutum width of larvae on rabbit 1 does differ
significantly from those on rabbits 2 and 3. An alternative way of .presenting this
data is shown in column 5, in which rabbits sharing the same letter do not differ
significantly in the mean scutum width of the tick larvae taken from them.
Another useful piece of information we can extract from the ANOV A table (after
we have used an F test to check that there is a significant effect) is the contribution
to the overall variance arising from a particular treatment. We have already seen
(section 8.1.3) that the mean square within groups, MSw, gives us an estimate of
(12, the intrinsic variability of the data, while the mean square between groups, MSB,
gives us an estimate of (12 + ns,~ where (1~ is the average value per level of the
added variance component due to the treatment T and n is the number of levels per
treatment, in this case the mean number of ticks per rabbit, i.e. 9.25. Therefore

(MSB - MSw)/n ~ ((12 + n(1~ - (12)/n = s~ 8.5

gives us an estimate of this added variance per level. (Sokal and Rohlf, 1981,
show how to calculate confidence limits for the added variance.) From the data given
in Table 8.8, the average value of the added variance component is 52.8 microns2
per rabbit and the best estimate of the standard deviation of the variability among
rabbits is 7.3 microns.

8.3 MODEL I AND MODEL II ANOV A

The experiment on rabbits discussed in the previous section could have been conducted
for either of two reasons. We could have been studying levels of immunity of rabbits
134 Analysis of variance
to ticks in order to identify the factors that confer such immunity with a view to
developing a vaccine against ticks. Rabbits on which ticks grow less well may have
some immunity. In this case we would be interested in the individual rabbits and we
would refer to the rabbits as a fixed factor. When we are interested in individual
levels of various treatments or factors, we talk about a Model I ANDV A. and we
choose particular treatment or factor levels (in this case rabbits) with a view to
deciding if the different levels produce significantly different effects. After calculating
the ANOV A table, we then go on to test for differences between the means or
between various combinations of the means.
We could, however, be interested in the genetic variability within various rabbit
populations of their response to ticks so that the variability itself is of interest. When
we are interested in the variability across the levels of anyone treatment or factor,
we talk about a Model II ANDV A, and we will generally choose the levels (in this
case rabbits) at random, often going to great lengths to ensure that the chosen rabbits
are representative of the population at large. We then refer to the rabbits as a random
factor. In a Model II ANOV A we tacitly assume that the effect of the variability is
not only random but is normally distributed. After calculating the ANOV A table,
we then go on to assess the added component of the variance arising from the
treatment effect. The important point to note is that for a given set of data the
precise analysis that we carry out depends on the question that we ask.
We can also have mixed models in which some factors are fixed and correspond
to Model I and others are random and correspond to Model II. In Student's study
of the effects of hyoscyamine and hyoscine on sleep, we hope to determine the
effects of these two specific drugs so that the drugs are a fixed factor. Since we are
not interested in particular patients but in these patients as representative of the
population as a whole, the different patients used in the trial are a random factor
and the variance within groups of patients gives us an estimate of the variance in
the population as a whole.

S.3.t Specifying the model

It is often useful to think explicitly about the model that underlies our analysis.
Suppose that the response variable is Yij' so that in Student's data on soporific drugs
Y indicates the number of hours of sleep gained, i indicates each of the two drugs
and j indicates each of the individual patients. Then our model is

Y 1).. = m + d + e"
l' IJ
8.6

where m is the mean number of hours of sleep gained by all the patients under
both treatments, diU = 1,2) is the change from the mean under each drug and the
residuals ejj give the residual amount of sleep gained by patient j(j = 1 to 10) not
accounted for by m and the drugs dj• In this case the grand mean tells us that m = 0.54
hours and the difference between the means for each drug and the grand mean tell
us that dl = 0.21 hours and d2 = - 0.21 hours.
 Two-way ANOV A 135
8.4 TWO-WAY AN OVA

The simplest way to extend the one-way ANOV A described in the previous sections
is to consider two treatment factors. We have already considered a problem of this
kind in the experiment on the diet of rats (section 7.7) where we examined the extent
to which the freshness of the lard and the sex of the rats affected the weight gained
by the rats.

8.4.1 Factorial designs

Suppose that we have two treatments or factors, A and B (in our experiments on
rats these would be lard and sex), each with two levels (fresh/rancid lard, male/female
rats). Ideally we would like to make several measurements with every possible
combination of A and B: Al B1, A2 B1, Al B2 and A2 B2• When we make measurements
on every combination of the relevant factors, we call it a factorial experiment. Note
that factor A is crossed with factor B since each level of B occurs with each level of
A and vice versa.
One way to approach such an experiment would be to treat each combination of
factors as though it were a separate factor. We might then use C to represent these
combined factors, so that

C1 == A1B1 C2 == A2B1 C3 == A1B2 C4 == A2B2• 8.7

We can then treat this experiment as a one-way ANOV A with C as the single
treatment or factor with four levels. In this case, C1 would be fresh lard, male rats;
C2 would be rancid lard, male rats; C3 would be fresh lard, female rats; and C4 would
be rancid lard, female rats. Our model would then be

Yij =m + + 8ij•
Cj 8.8

Table 8.10 gives the data of Table 7.8 laid out as a one-way ANOV A. The
analysis proceeds as before and the resulting ANOV A table is shown in Table 8.11
which shows that the combined treatment factor C, corresponding to sex and lard,
is highly significant.
We could now proceed as we did when we were using t tests and examine the
effect of sex allowing for the freshness of the lard, and so on. However, it is more

Table 8.10 The weight (in grams) gained by male rats fed on fresh
lard C1 and on rancid lard CZf and by female rats fed on fresh lard
C, and on rancid lard C

Treatment
C1 Cz CJ C4
171 108 153 85
172 89 109 64
172 69 160 82
136 Analysis of variance
Table 8.11 One-way ANOYA for the data of Table 8.13 for the weight gained
by male and female rats fed on fresh and on rancid lard with each combination of
sex and freshness treated as a separate factor
Source of
variation
Sums of
squares
Degrees
of freedom
Mean
square F-ratio
Significance
level, P .
Among 17779 3 5926 18.6 0.0006 ...
Within 2548 8 319
Total 20327 11

Table 8.12 The data of Table 8.10 laid out for a two-way
ANOY A with three repeated measures
AI (fresh lard) A2 (rancid lard)
171 108
BI (male rats) 172 89
172 69

153 85
B2 (female rats) 109 64
160 82

convenient to build these comparisons directly into our analysis and this is what
the two-way (or multi-way) ANOVA does for us. Table 8.12 gives us the data of
Table 8.10, but this time laid out as for a two-way ANOV A.
Our model is now

Yjjk = J-l + aj + bj + (ab)jj + Bjjk' 8.9

where aj indicates the change from the mean value when the lard is fresh or
rancid, bj indicates the change from the mean when the rats are males or females
and (ab)jj indicates the effect of the interaction so that (ab)w for example, gives the
effect of fresh lard for male rats after allowing for the overall mean, the average
effect of sex and the average effect of lard freshness.
To make the details of Model 8.9 explicit we can calculate each term separately.
Table 8.13 gives various mean values calculated from the data in Table 8.12. The
grand mean is 119.5 g so that in Equation 8.9

m = 119.5 g. 8.10

The average increase in weight for male rats is 130.2 g so that

a1 = 130.2 -119.5 = 10.7 g 8.11

and in the same way

a2 = 108.8 -119.5 = -10.7 g. 8.12

 Two-way ANOV A 13
Table 8.13 Averages calculated from the numbers in Table 8.12. The
four numbers in the comers are the means of the four sets of three
numbers in Table 8.12. The numbers in italics are the means of the
two adjacent numbers (vertically or horizontally) and the number in
bold is the grand mean of all the data

Fresh < Freshness) Rancid

Male 171. 7 130.2 88.7
<Sex) 156.0 119.5 82.8
Female 140.7 108.8 77.0

The average increase in weight when the lard is fresh is 156.0 g so that

bI = 156.0 -119.5 = 36.7 g 8.13

and similarly
b2 = 82.8 - 119.5 = - 36.7 g. 8.14

That the two coefficients for the effect of sex and the two coefficients for the effect
of lard freshness differ only in signs, reminds us that in each case we have only a
single independent number and hence only 1 degree of freedom for each factor.
To calculate the interaction term for male rats eating fresh lard we take the average
weight gain for all the rats (119.5 g), add the additional weight gain for male rats
(10.7 g), and then add the additional weight gain for fresh lard (36.7 g), to give 166.9 g.
Then since the average weight gain for male rats eating fresh lard is 171.7 g, the
interaction term is

(ab)II = 171.7 - 166.9 = 4.8 g. 8.15

We can calculate the other three interaction terms in the same way:

(abh2 = - 4.8 g (abhI = 4.8 g (abh2 = - 4.8 g. 8.16

That the four numbers differ only in their signs shows that we have 1 degree of
freedom for the interaction.
The ANOV A for the data of Table 8.12 is given in Table 8.14. Comparing Table
8.14 with Table 8.11, we see that the total and residual sums of squares are the same
in both cases and that the sums of squares for the three 'effects' -sex, lard and
sex x lard-add up to the sums of squares for the single treatment, 'sex-lard'. In
other words, the two-way ANOV A partitions the total sums of squares for the
treatments into contributions from the individual treatments and their interactions.
In Table 8.14 the effect of sex is not significant at the 5% level, as we found in
Table 7.10 using the t test. The effect of lard freshness, on the other hand, is highly
significant and has three stars, as in Table 7.10. Finally, the effect of the interaction
sex x lard is again not significant in agreement with the t test. In fact, the two ap-
proaches are entirely equivalent when we have only two levels for each treatment,
so that we are comparing means in pairs. (In Table 7.10 the t parameters are 2.07
for sex, 7.12 for lard freshness and 0.94 for the interaction term. If we square these
138 Analysis of variance
Table 8.14 Two-way ANOV A for the data of Table 8.12 for the weight gained
by male and female rats when fed on fresh and rancid lard
Source of
varialion
Sums of
squares
Degrees
of freedom
-.------.------- ..--"-----,
Mean
square F-ralio
---
Significance
level, P .
Sex
Lard
1365
16133
1
1
1365
16133
4.29
50.7
0.072
0.0001 ...
ns

Sex x Lard 280 1 280 0.88 0.39 ns

Residual 2548 8 318
Total 20326 11

numbers we have 4.28, 50.7 and 0.88, which are identical to the corresponding F
statistics given in Table 8.14.) When we have more than two levels per treatment,
the ANOV A assumes that none of them has any effect; if we find a treatment or
factor that does have a significant effect, we then proceed to examine its various
levels in pairs to discover reasons for the effect. If we were to start from the t test,
we would have to build up all possible pairs of effects and would rapidly become
bogged down in complexity.
In Table 8.14 the sum of squares for sex, lard and sex x lard add up to the sum of
squares for the combined treatment sex-lard (Table 8.11), as we would expect. For
sex we have two levels, males and females, and we use up 1degree of freedom when
we average over sex leaving us with 2 - 1 = 1 degree of freedom for sex. In the
same way we see that we are left with 2 - 1 degree of freedom for lard. For the
interaction, we have (2 - 1) = 1 degree of freedom for sex combined with (2 - 1) = 1
degree of freedom for lard, giving us 1 x 1 degree of freedom for the interaction. For
the residual term we have three repeats in each category, giving (3 - 1) = 2 degrees
of freedom for each one. We repeat this 2 x 2 = 4 times for all combinations of sex
and lard, giving 2 x 4 = 8 degrees of freedom for the residual term. These four
numbers 1, 1, 1 and 8 add up to 11 and this is the same as the number of degrees
of freedom for the total sum of squares, which is 12 - 1 = 11.
We can generalize the argument of the preceding paragraph as follows. Suppose that
in our experiment treatment A has a levels and treatment B has b levels. The relevant
numbers of degrees of freedom in a two-way factorial ANOV A would then be those
given in Table 8.15. For A we have a levels and after calculating the deviations
from the mean we are left with (a - 1) degrees of freedom. For B we have b levels
and after calculating the deviations from the mean we are left with (b - 1) degrees
of freedom. For A x B we have a x b levels, but only (a - 1) and (b - 1) independent
levels, giving us (a - 1) x (b - 1) degrees of freedom. Within each group we have
n levels and calculate 1 mean, giving us n - 1 degrees of freedom, but we repeat
this ab times~ giving us ab(n - 1) degrees of freedom for the residual sum of squares.
For the total we take the total number of points abn and subtract 1, giving us
abn - 1 degrees of freedom.
 Two-way ANOV A 139
Table 8.15 Decomposition of the number of degrees of freedom in
a two-way factorial ANDV A
Treatment No. of levels d·f
A a a-I
B b b-I
AxB ab (a - I)(b - 1)
Residual n ab(n - 1)

Total abn abn - 1

8.4.2 Multiple comparisons

Making several comparisons brings us back to the discussion of multiple comparisons

in section 7.4. If we had decided in advance that we wanted to examine only the
effect of the lard, then comparing the F ratios in Table 8.14 with the appropriate
critical values is correct. However, if we want to test all three F ratios to see if any
of them are significant at a significance level of (1.,we should use a significance level
of (1./3. The effects of sex and lard are, of course, still not significant. To decide if
the effect of lard is significant at the 0.1% level, we need to use a significance level
of 0.03% and in this case we see that it still has three stars since the significance level
corresponding to the observed F ratio is 0.01%.

8.4.3 Interactions

The roles of the two levels sex and lard in our ANOV A are clear: they tell us if the
amount of lard consumed depends on the sex of the rats (after allowing for freshness)
or the freshness of the lard (after allowing for sex). We have already discussed the
interaction term in Chapter 7 (Table 7.10), where we calculated

(F - Rim) - (F - Rlf), 8.17

that is the difference between the effect of lard freshness on the consumption
by male rats and the effect of lard freshness on the consumption by female rats.
We also calculated
(m - flF) - (m - fiR), 8.18

which is the difference between the effect of the sex of the rats on the consumption
of fresh lard and the effect of the sex of the rats on the consumption of rancid lard.
We found that the interaction term was the same whichever way we calculated it.
To understand the interaction clearly, consider the hypothetical results given sche-
matically in Fig. 8.3. In Fig. 8.3a the effect of changing A from level Al to A2 is the
same for both levels of B, and the effect of changing B from level BI to B2 is the same
for both levels of A. The effect of changing A is therefore independent of the level
of B, and vice versa, and there is no interaction. In Fig. 8.3b, on the other hand, changing
A from level Al to A2 increases the response if we are at level BI and decreases the
140 Analysis of variance

Figure 8.3 Schematic diagram of the means in a two-factor ANOV A. (AI I 81), for example,
gives the mean value under level 1 of A and level 1 of B. In (a) there is a significant effect for
A and 8 but no interaction A X B. In (b) neither treatment is significant but there is a significant
interaction A X B. In (c) treatment 8 and the interaction are significant but treatment A is not.
In (d) both treatments and the interaction are significant.

response if we are at B2' so that the effect of a change in A depends on the level
of B and there is a significant interaction. Similarly, changing B from level BI to B2
increases the response if we are at level Al and decreases the response if we are at
A2' so that the interaction is significant. Note, however, that the effect of changing
A averaged over the two levels of B is zero and the effect of B averaged over the
two levels of A is zero so that neither factor is significant when averaged over the
levels of the other factor. Fig. 8.3c shows a situation in which B and the interaction
are significant while A is not, while Fig. 8.3d shows a situation in which both treat-
ments and their interaction are significant.
Figure 8.4 is a plot of the mean weight gained by the male and female rats eating
fresh and rancid lard and we can see that there is no significant interaction because
the two lines are almost parallel.

8.4.4 Soporific drugs

In section 8.16 we analysed Student's data on the effects of two soporific drug on
20 patients treating the patients as a random factor nested in the fixed drugs factor.
 Two-way ANOV A 141

In fact, only ten patients were used, with each patient being given both drugs and
also no drug. Since each patient was given both drugs, we can treat the patients as one
factor and the drugs as a second in a factorial design. The patients are still a random
factor but they are now crossed with the drugs not nested under drugs. Since we
are not interested in the response of individual patients who have been randomly
seleded from the overall population, the patients are a Model II factor. We are,
however, interested in the effect of each drug separately, and so the drugs are a Model
I factor. The model is therefore a mixed-model, two-way ANOV A without replication.
The ten patients account for 9 degrees of freedom and the two treatments account
for 1 degree of freedom. If we tried to calculate the interaction term, this would
have 9 x 1 degrees of freedom, giving us a total of 19, but since we have only 19
degrees of freedom at our disposal, we would have none left for the error term. In
other words, we can carry out an ANOV A on a two-factor model without replication,
but only if we assume that the interaction term is zero so that we can ignore it.
Carrying out a factorial ANOV A on Student's data with the two factors crossed, we
obtain the results shown in Table 8.16. By treating patients as a separate factor, the
analysis is equivalent to a paired t test. In Table 7.2 we obtained a value of t = 4.06,
which is equal to the square root of 16.5, the value of F in Table 8.16, so that the
two-way ANOV A without replication is equivalent to the paired t test. By using
the fact that each treatment is applied to the same patients twice, we are able to
show that the effeds of the two drugs are significantly different, whereas in our
earlier ANOV A (Table 8.6), in which we did not use this knowledge, the two tests
were not significantly different. Of course, if we had used 20 patients and had given
each of them one treatment only, the analysis in Table 8.6 would have been the best
that we could do. For the two-way ANOV A our model is

Yij= m +dj + Pj + Gij' 8.19

142 Analysis of variance
Table 8.16 ANOVA table for the two drug treatments in Table 7.2. Sj is the internal standard
deviation and sp the pooled standard deviation

Source of Sums of
variation sqliares
Degrees
of freedom
Mean
square F-ratio
Significance
level, P .
Treatment
-----
.

..
--
Patients
12.5
58.1
1
9
12.5
6.46
16.5
8.53
0.0028
0.0019 ..
Residual 6.8 9 0.756
Total 77.4 19

Drug Mean Sj sp 95% c.L.

Hyoscyamine 0.75 0.57 0.27 - 0.13-1.37
Hyoscine 2.33 0.63 0.27 1.71-2.95

where m is the grand mean, di is the change from the mean due to drug i, Pi is the
change from the mean due to patient j and Bjj is the residual for patient j on drug i.
It is worth comparing Table 8.6, in which we simply treated the patients as replicates,
with Table 8.16, in which we treat the patients as a second factor. The 'treatment'
sum of squares in Table 8.16 is identical to the 'between-groups' sum of squares in
Table 8.6. However, the 'within-group' sum of squares in Table 8.6 is now
partitioned into a 'patient' sum of squares and a 'residual' sum of squares. Because
we are able to separate these out and compare the treatment effect with the residuals
after allowing for the effect of the patients, the denominator in our F test is reduced
and the test of our null hypothesis is made more sensitive. The means in Table 8.16
are exactly the same as in Table 8.6. However, the pooled standard deviation in
Table 8.16 (0.27) is less than the value of the pooled standard deviation in Table
8.6 (0.60) as it is now calculated after allowing for the variation among patients. The
reduction in the pooled standard deviation increases the sensitivity of the comparison.

8.4.5 Latin squares

We would always like to make measurements for all possible combinations of the
levels of each factor and with several repeats of each measurement so that we can
assess the effect of each factor as well as all the interactions among the factors.
Unfortunately this is not always possible. We have already seen in our study of the
polio vaccine that we cannot give the vaccine and the placebo to each child and we
will return to this again in the next section. Consider, however, the following
problem. Tsetse flies (Glossina spp.) are the vectors of trypanosome diseases called
nagana in cattle and sleeping sickness in people. Trypanosomiasis is widely regarded
as one of the most debilitating diseases in Africa and where tsetse flies are present
livestock are largely absent. This makes it impossible to keep livestock in much of
the potentially most productive areas of Africa. The lack of livestock results in a lack
of protein, milk, tractive power, fertilizer and leather goods. Recent work, however,
has shown that testse fly populations can be controlled effectively using traps
 Two-way ANOV A 143

Figure 8.5 Three traps, b- biconical, e- epsilon and n- Nguruman, are rotated around three
different sites on three different days in a Latin square experiment.

(Dransfield et al., 1991). The development of traps that are effective against tsetse
flies has involved testing many different trap shapes, colours and attractive odours.
How can we best compare different traps in field conditions? We know that the number
of flies that we catch in a trap varies significantly from day to day, for reasons
that are still not understood. We also know that the number of flies that we catch
in a trap depends on the site in which we put the trap, probably because of variations
in the visibility and degree of shading at different sites. Now we can put trap 1 into
site 1on day 1but we cannot then put another trap into the same site on the same
day, and we are faced with the problem that we cannot carry out measurements on
all possible combinations of treatments. In this situation a very useful design is the
Latin square, illustrated in Fig. 8.5. We see that we can rotate the traps between sites
on successive days so that each trap occupies each site over the 3 days. (Note that
the square refers to the design and not to the actual layout although the layout
could also be a square.)
Dransfield and Brightwell (pers. comm.) carried out an experiment to compare the
effectiveness of three traps for female Glossina pallidipes. The first and oldest trap is
called the biconical because it is made out of two cloth cones joined at their bases
(Challier et al., 1977). A more recently developed trap is the Nguruman trap (Brightwell
144 Analysis of variance
Table 8.17 Numbers of tsetse flies caught in biconical, b, epsilon, e,
and Nguruman, n, traps. Three Latin squares were used. Within each
square there were three sites and measurements were taken over 3
days so that all three traps are placed once in each site and on each
day all three traps are found in one of the three sites for that square

Day
Site 1 2 3

1 n393 e230 b158

2 b 58 n146 e392
3 e159 b 60 n392
4 e 97 b 22 n252
5 n443 e1l9 b 81
6 b 33 n166 e162
7 b 12 n103 e 38
8 e 55 b 30 n216
9 n 89 e132 b 44

et a/., 1991) and this consists of a triangular prism made of cloth with a netting cone
on the top. The third trap is the epsilon trap which is similar to the Nguruman trap
with the cone recessed into the prism. All three traps were baited with acetone, cow
urine and octenol, which are known to be attractive to tsetse flies. The results of
the experiment are recorded in Table 8.17.
The Latin square is an incomplete factorial design with all factors crossed. Before we
carry out the analysis of variance we need to consider the model that we wish to apply
to our data. We assume that the effects of sites, traps and days are multiplicative
so that if site 2 is twice as good as site 1, if there are twice as many flies present
on day 2 as there are on day 1, and if trap 2 is twice as good as trap 1, then trap 2
on day 2 in site 2 should catch 2 X 2 X 2 = 8 times as many flies as trap 1 in site 1 on
day 1. However, analysis of variance assumes that the effects are additive. We
therefore take the logarithm of the trap catches since the logarithm of a product is
the sum of the logarithms and multiplicative effects become additive after transforming
the catches to logarithms. Our model is therefore

/og(catch) = m + Sj + Ij + dk + Bjjk' 8.20

where Sj refers to the three sites, Ij to the three traps and dk to the three days.
We can now analyse each square separately and the ANOV A table for the first
square (sites 1, 2 and 3 in Table 8.17) is given in Table 8.18 from which we see that
the epsilon trap catches more flies than the biconical trap and the Nguruman trap
catches more flies that the epsilon trap. However, the only significant difference is
between the biconical trap and the Nguruman trap. If we repeat the exercise for the
other squares, then in square 2 the biconical trap differs significantly from both the
Nguruman and the epsilon while in square 3 none of the traps differs significantly
from any other. The reason for this is that in square 3 fewer flies were caught than
 Two-way ANOVA 145
Table 8.18 Analysis of variance for the first Latin square given in Table 8.17. The catches
were transformed by calculating the logarithm of the catch (base 10) before carrying out the
analysis of variance. 'Lower' and 'upper' give 95% confidence limits for the transformed mean
catches. The means are compared using Tukey's HSD at the 95% level

Source of Sums of Degrees Mean Significance

variation squares of freedom square F-ratio level, P ·
Site 0.2091 2 0.1045 21.4 0.040 •
Trap 0.4364 2 0.2182 44.7 0.022 •
Day 0.0594 2 0.0297 6.08 0.141 ns
Residual 0.0098 2 0.0049
Total 0.7147 8

Pooled standard error = 0.0403

Trap Lower Mean Upper
Biconical 1.740 1.913 2.087 a
Epsilon 2.050 2.223 2.397 ab
Nguruman 2.277 2.45 I 2.624 b
146 Analysis of variance
Table 8.19 Analysis of variance for the data in Table 8.17 treated as a Latin rectangle. The
catches were transformed by calculating the logarithm (base 10) of the catch before carrying
out the analysis of variance. 'Lower' and 'upper' give 95% confidence limits for the transformed
mean catches. The means are compared using Turkev's HSD at the 95% level

Source of
variation
Sums of
squares
Degrees
of freedom
Mean
square F ratio
Significance
level, P .
Site 1.394 8 0.174 7.76 0.0005 ......
Trap 2.165 2 1.082 48.2 0.0000
Day 0.158 2 0.079 3.52 0.058 ns
Residual 0.314 14 0.0224
Total 4.031 26

Pooled standard error = 0.0499

Trap Lower Mean Upper

Biconical 1.527 1.634 1.742 a

Epsilon 1.935 2.042 2.150 b
Nguruman 2.217 2.324 2.431 c

which to compare other traps and we can calculate an index of increase by detrans-
forming the means and then calculating the ratio of the catch in each trap to the
catch in the biconical trap and this gives us the result in Table 8.20.
To determine error limits for the indices of increase, we calculate errors for the
differences between the transformed mean catches as the square root of the sums of the
squares of the errors for each pair of traps. We then add the error term to the dif-
ference between the transformed mean and detransform the result to obtain the upper
limit and subtract the error term and proceed in the same way to obtain the lower
limit. Although the errors in the differences of the transformed means are symmetrical,
the errors in the detransformed means are not, as we can see in Table 8.20. We see
then that the epsilon trap catches, on average, 2.5 times as many flies as the biconical
with 95% confidence limits from 1.5 to 4.2, while the Nguruman trap catches, on
average, 4.8 times as many flies as the biconical with 95% confidence limits ranging
from 2.9 to 7.9.

Table 8.20 Indices of increase for the epsilon and Ngururnan. traps
over the biconical trap with 95% confidence limits

Lower Mean Upper

Biconical
Epsilon
•
1.51
1.00
2.51
•
4.16
Ngururnan 2.88 4.78 7.91
 Two-wayANOV A 147

Figure 8.6 Schematic illustration of the dependence of the calcium concentration on the plants,
the leaves and the individual calcium determinations. The data are in Table 8.21.

8.4.6 Nested designs

We have seen that in a factorial design we make measurements for all possible combina-
tions of the levels of each factor. A Latin square design is an incomplete factorial
design that we use when we cannot set up all possible combinations of the levels of
each factor. Another important class of designs arises when factors are nested within
one another. In the analysis of Student's data on soporific drugs, we noted that if we
had 20 different patients we would have performed either an unpaired t test or a
single-factor ANOV A; the patients would have been nested within the drugs. In the
analysis of scutum widths of tick larvae on rabbits the ticks were nested within rabbits
and we were able to estimate the added variance component due to rabbits.
Consider the following example of a nested design, discussed in detail by Snedecor
and Cochran (1989, p. 247), in which the calcium concentration in the leaves of turnip
plants was measured as a percentage of the dry weight of the leaves. Four turnip
plants were chosen; from each plant three leaves were chosen; from each leaf two
samples were taken and used to measure the concentration of calcium. Clearly, we
might expect the calcium concentration to vary between plants, the concentration
within each plant to vary between leaves and the concentration within each leaf to
vary between the two measurements of the calcium concentration. This hierarchical
structure is illustrated in Fig. 8.6 which shows the data given in Table 8.21.
The various factors are nested one within the other, because the same leaf cannot
be on more than one plant and each sample cannot be from more than one leaf. The
model for this design is written as follows

Yijk = m + Pi + lij + Sijk' 8.21

148 Analysis of variance
Table 8.21 Four plants were chosen (columnl). From each of these three leaves were chosen
(column 2). From each of these two samples were taken (column 3) and determinations were
made of the calcium concentration in parts per million (column 4). The sums of squares (column 5)
and the mean values (column 6) of the determinations in each leaf were calculated. The mean
values in column 6 were used to determine the sums of squares (column 7) and the mean values
(column 8) for the leaves within plants. The mean values in column 8 were use to determine
the sums of squares within plants (column 9)

1 2 3 4 5 6 7 8 9
P L S Ca% SS(S in L) <S in L) SS(L in P) <L in P) SS(P)

1 3.28
1 0.0180 3.19
2 3.10
1 3.52
1 2 0.0008 3.50 0.2180 3.18
2 3.48
1 2.88
3 0.0032 2.84
2 2.80
1 2.46
1 0.0002 2.45
2 2.44
1 1.87
2 2 0.0012 1.90 0.1542 2.18
2 1.92
1 2.19
3 0.0000 2.19
2 2.19

1 2.77
1 0.0060 2.72 1.2601
2 2.66
1 3.74
3 2 0.0450 3.59 0.6248 2.95
2 3.44
1 2.55
3 0.0000 2.55
2 2.55
1 3.78
1 0.0041 3.83
2 3.87
1 4.07
4 2 0.0012 4.10 0.3181 3.74
2 4.12
1 3.31
3 0.0000 3.31
2 3.31

where from each plant Pi three leaves were chosen so that lij is leaf j in plant i and
from each leaf two samples were taken and used to measure the concentration of
calcium so that Sijk is kth determination of concentration of calcium in leaf j from
plant i.
We can now calculate the sums of squares for samples in leaves, leaves in plants
and plants. For the samples in leaves we simply add up the individual sums of squares
given in column 5. For the leaves in plants we add up the sums of squares in' column
 Two-way ANOVA 149
Table 8.22 Sums of squares and mean squares for the data given in
Table 8.21

Source SS d.f MS
Plants 7.5603 3 2.5201
Leaves in plants 2.6302 8 0.3288
Samples in leaves 0.0799 12 0.0067

7, but since each mean is the mean of two determinations, we multiply by two to find
the total sum of squares for the leaves in plants. For the plants we take the sum of
squares in column 9 and since each mean is the mean of six determinations, we
multiply by six to find the total sum of squares for the plants. The results of these
calculations are given in Table 8.22. To determine the mean squares, we now divide
each sum of squares by the number of degrees of freedom. Since there are four
plants, there are 3 degrees of freedom for the plants. Since there are three leaves
within each of four plants, there are (3 - 1) x 4 = 8 degrees of freedom for the
leaves in plants. Since there are two samples within each of 3 x 4 leaves, there
are (2 - 1) x 12 = 12 degrees of freedom for the samples in leaves.
Now we need to think carefully about just what each mean square estimates: The
mean square for samples in leaves allows for the effects of the plants and the leaves,
since each term is calculated after allowing for their effects, and gives us an estimate
of 11';, the population variance of the samples in each leaf. If neither the plants nor
the leaves had any effect on the measurements, the mean squares for the leaves in
plants would also give us an estimate of 11';. Suppose, however, that the population
variance of the calcium concentration among the leaves is I1'f. Then we add to the
mean square for the leaves in plants ns 11'[ where we have multiplied by ns the number
of samples within each leaf. The mean square for leaves in plants is therefore an
estimate of 11'; + nsl1'f (d. section 8.1.3). The variation among plants includes the
variation among leaves so that we start with 11'; + nsl1'f, Now if the population
variance of the calcium concentration among the plants is 11'~ we must add nsnll1'~
where we multiply by nsn1 the number of samples within each plant. These results
are summarized in Table 8.23. To test the null hypothesis that the plants have no
effect, that is 11' = 0, we use the mean square for leaves within plants as our error
~
term. To test the null hypothesis that the leaves have no effect, that is I1'f = 0, we
use the mean square for samples within leaves as our error term. From Table 8.23
we see that the variation due to the plants is significant at the 5% level while the
variation due to the leaves is significant at the 0.1% level.
This experiment was performed to determine the overall variation in the amount
of calcium in turnip plants. We want to know how much of the variation is due to
plants, leaves and samples taken from each leaf. The difference between the mean
squares for plants and for leaves in plants divided by nsn, gives us 0.365 as our
ISO Analysis of variance
Table 8.23 Analysis of variance table for the determination of calcium within leaves
on turnip plants. The F ratio for plants is calculated as 2.52l'1 /0.3288 = 7.66 with 3
and 8 degrees of freedom. The F ratio for leaves is calculated as 0.3288/0.0067 = 49
with 8 and 12 degrees of freedom

Source d·f MS Parameters estimated F-ratio .

Plants
Leaves in plants
3
8
2.5201
0.3288
<1: + n,<1, + n,nl<1~
<1: + n, <1,
2
7.66
49 ...•
Samples in leaves 12 0.0067 <1,

best estimate of a;. The difference between the mean squares for leaves in plants
and samples in leaves divided by ns gives us 0.161 as our best estimate of af.
Finally,
the mean square for samples gives us our best estimate of a;
as 0.0067. The standard
deviations are then 0.60, 0040 and 0.08, for plants, leaves and samples, respectively.
The variation among leaves on a given plant is almost as great as the variation among
plants, whereas the variation among determinations on a given leaf is much less.
With a nested design and random factors we are interested in the amount of
variation within successive levels of the various factors and our analysis is aimed at
estimating the contribution to the overall variation from the variation at each level.
This is in contrast to a crossed design with fixed factors where we are interested in
comparing the different levels of each factor using a multiple range test. It is important
to remember that for random factors we tacitly assume that their effects are normally
distributed.

8.5 SUMMARY

Most of the variables that we measure have associated uncertainties that are normally
distributed or can be made so by a well-chosen transformation. Furthemore, if we
design our experiments carefully, the sum of the squares of the deviations of all of
our observations from the mean can be decomposed into contributions from each of
the contributing factors. For these reasons the analysis of variance (which should,
perhaps, be called the analysis of sums of squares) is central to the statistical analysis
of experimental data.
ANOV A allows us to take the results of a complicated experiment with many
factors, crossed and nested, fixed and random, some interacting with others, and
identify those that contribute significantly to the overall variation in the data. We
can then analyse the data further using multiple range tests for fixed factors or the
added variance components for random factors. The design of our experiments can
be as complicated as we choose although the analysis of variance will then be
correspondingly complicated.
The most complicated experimental design can be expressed as combinations of
crossed and nested factors. We say that B is nested in A if the levels of factor B
 Exercises 151
differ among the levels of factor A. In Student's data on soporific drugs (section
8.1.6), if there had been 20 patients of whom ten received one drug and ten the
other, the patients receiving the first drug would not have been the same as the
patients receiving the second drug so that patients would have been nested within
drugs. We then analyse the data using an unpaired t test or a single-factor ANOV A.
The turnip data discussed in section 8.4.6 is an example of a purely nested design, with
determinations nested in leaves, leaves nested in plants. If a factor B is nested within
a factor A, factor B must be a random factor and we are then interested in using the
ANOV A to determine the variance due to factor B.
We say that factor B is crossed with factor A if each level of factor B occurs in
combination with each level of factor A. In Student's data on soporific drugs (section
8.4.4), we actually had ten patients, each of whom received both drugs. Since each
patient was given each drug, the patients were crossed with the drugs. We were
then able to analyse the data using a paired t test or a two-factor ANOV A. Although
factor B was now crossed with factor A, factor B (patients) was still a random factor
while factor A (drugs) was a fixed factor. If we were interested only in these ten
people and never had any intention of giving the drug to anyone else, the patients
would also have been a fixed factor. We are, of course, testing the drugs on these
ten patients in the hope that if the drugs are effective we will then be able to give
them to other patients. Our ten patients are therefore a random sample from a larger
population of patients. The data on the weight gained by male and female rats fed
on fresh and on rancid lard is another example of a purely crossed design. In this
case, sex and lard freshness are both fixed factors.
The analysis of our data depends on whether the factors are nested or crossed,
random or fixed. Especially in agricultural trials, field crops and treatments may be
divided up in quite complicated ways and one needs to think carefully about the analysis.
As always, the best defence against error and confusion is to think carefully about
the biology and to think about the relationship between the particular experiment
that you have performed and the broader context in which you hope to apply your
results.
In this chapter our response variables have generally been quantitative (hours of
sleep, number of ticks, weight gained), while our predictor variables have been quali-
tative (drugs, rabbits, sex and lard freshness). In Chapter 7 we touched on non-parametric
tests, in the analysis of data for which the response variables are also qualitative.
There is no reason to restrict ourselves to qualitative predictor factors, however, and
in the next chapter we will consider the analysis of data in which both the response
and the predictor variables are measured on quantitative scales.

8.6 EXERCISES

1. In a single factor ANOV A for an experiment on four rabbits, ten ticks were removed
from each rabbit and weighed. We are interested in the variation between the average
weights of the ticks on the four rabbits. How many degrees of freedom would you
use to calculate the within-treatments sums of squares? How many degrees of freedom
152 Analysis of variance
Table 8.24 The number of male G. pallidipes tsetse flies caught by a
biconical, b, an epsilon, e, and a Nguruman, n, trap. These data were
colleded as part of the experiment described in section 8.4.5

Day
Site 1 2 3

1 n675 e 192 b292

2 b193 n114 e245
3 e395 b109 n507
4 e 132 b 39 n214
5 n506 e 94 b191
6 b 89 n105 e210

7 b 25 n 46 e 43
8 e 34 b 18 n160
9 n 49 e 40 b 55

Table 8.25 Analysis of variance table for the combined Latin redangles given in Tables 8.17
and 8.24
Source of Sums of Degrees Mean Significance
variation squares of freedom square F ratio level, P
Sex 0.050 1 0.050 2.31 0.14
Trap 2.465 2 1.232 56.3 0.0000
Day 0.698 2 0.349 15.9 0.0000
Site 4.116 8 0.514 23.48 0.0000
Sex x Trap 0.270 2 0.135 3.19 0.0060
Sex x Day 0.226 2 0.113 6.18 0.0122
Sex x Site 0.342 8 0.043 5.17 0.0912
Residual 0.613 28 0.218 1.95

would you use to calculate the beween-treatments sums of squares? How many
degrees of freedom would there be in the total sums of squares?

2. Using the data in Table 8.16 estimate the added variance component and hence
the population standard deviation in the amount of sleep gained due to the variability
among patients.

3. How many degrees of freedom are there for each factor and for the error term
in a 4 x 4 Latin square? How many measurements would you make? How many
degrees of freedom are there for each factor and the error term in a 4 X 4 factorial
experiment? How many measurements would you make?
 Exercises 153
4. Use a series of F tests to decide if any of the standard deviations given in Table
8.7 differ significantly.

5. The data for male flies given in Table 8.24 were collected at the same time as
the data for female flies given in Table 8.17. Using a computer carry out an analysis
of variance treating the three squares as one Latin rectangle.

6. The data for the number of female and male flies caught by the three traps given
in Tables 8.19 and 8.24 can be combined into a single table with sex as an additional
factor. Table 8.25 gives the results of the ANOV A on the combined data. Interpret
the significance of the interactions.
9

Regression

Nature only uses the longest threads to weave her patterns, so each smallpiece of her
fabric reveals the organization of the entire tapestry.
R. Feynman (1980)

In the previous chapter we used one or more independent variables (drugs, rabbits,
lard freshness) to explain the variation in a dependent variable (hours of sleep, scutum
width of larvae, weight gained by rats). When the independent variables are cate-
gorical, the analysis of variance provides a formal structure that enables us to assess
the contribution that each of them makes to the variation in the dependent variable.
Consider, however, the control of armyworm using Bacillus thuringiensis (B.t.) discussed
in section 1.1.4. Brownbridge (1988) was concerned to determine the biocidal effect
of B.t. on armyworms. In his experiments, the results of which are given in Table 9.1,
the control areas were not sprayed, while 0.5, 1.0 and 2.0% Rt. suspended in water
was applied for the three treatments. Estimates were made of the number of armyworms
surviving under each treatment on five successive days.
We could analyse these data using a two-factor ANOV A with days and treatments
as the factors and our model would be

njj =m + aj + bj + Bij, 9.1

where ai accounts for the effect of the ith treatment and bj for the effect of the jth
day on the number of armyworms. We would then find that the average number of
armyworms declined over time under all three treatments and that fewer armyworms
survive under treatment C than under the control, for example.
Let us try a different approach. We really want to know if the number of armyworms
declines at different rates under different treatments. We see from Fig. 9.1 that the
logarithm of the number of armyworms decreases more or less linearly with time in
all four cases. The equation for a straight line is y = a + bx and this suggests a
model of the form

Yi = In(ni) = a + bti + Bi' 9.2

where a is the intercept on the In(n) axis, b is the slope of the line, ti is the time at
which the ith observation was made and Bj is the residual term that gives us the
 Regression 155
Table 9.1 The number of armyworms counted on each of five successive days under each
of four treatments. Cont. is the unsprayed control sample, A, Band C are the areas sprayed
with 0.5%, 1.0% and 2.0% Bacillus thuringiensis suspended in water, respectively. The last four
columns give the natural logarithms of the numbers of armyworms

Day Cont. A B C In(Cont.) In (A) In (B) In(C)

0 497 320 294 295 6.209 5.768 5.684 5.687

I 463 203 213 93 6.138 5.313 5.361 4.533
2 506 155 118 33 6.227 5.043 4.771 3.497
3 487 125 124 12 6.188 4.828 4.820 2.485
4 480 101 III 4 6.174 4.615 4.710 1.386

deviation from the line for the ith observation. (For the moment we will apply
Equation 9.2 separately to each treatment.)
The model given in Equation 9.2 is attradive for several reasons. If we fit separate
lines to the data for each treatment, then in each case we have 5 degrees of freedom
and lose 2 for the model (the estimates of the slope and the intercept) leaving us with
3 degrees of freedom for the error term. We can then compare our estimates of the
slopes to see if the various treatments differ in their effed on the armyworms. Once
we have found values of a and h that give us the best fit to each set of data, we can
then choose any value for t and use Equation 9.2 to calculate the corresponding value
of In(n). The problem is to find values of a and h, the coefficients in the fit and this
is what regression does for us.
Before we find out how to determine the values of a and h, let us ask why the

Figure 9.1 The natural logarithms of the numbers of armyworms plotted against time. Under
the control no application of the biocide was made while the three treatments A, Band C
were made with 0.5, 1.0 and 2.0% B.t suspended in water, respectively.
156 Regression

data fall on a straight line when the logarithms of the numbers are plotted against
time. In the experiment on armyworms, we might expect the same proportion of
those surviving to die on each successive day. To express this idea formally, we write

dN(t)/dt = rN(t), 9.3

where dN/dt is the rate of change of N per unit time and the proportionality
constant r gives the growth or decay rate of the population. (For readers unfamiliar
with calculus, Equation 9.3 says that if we take a very small interval of time, dt, then
dN, the change in the number of armyworms in this small interval of time, will also
be very small. If we divide the one small number by the other, their ratio, which
gives the number dying per unit time, is proportional to N, the number present, and
the proportionality constant is r. Learning calculus enables you to derive Equation
9.4 from Equation 9.3.)
Integrating Equation 9.3 gives

N(t) = N(O)er~
, 9.4

where N(t) is the number of armyworms at time t, N(O) is the number at t = 0,

and r is the rate of increase or decrease, that is. the fractional change in numbers per
unit time. If r < 0, Equation 9.4 describes an exponential decline. if r > 0, an exponential
growth. (For a discussion of the dynamics of single populations, see May, 1981.) Taking
logarithms of both sides of Equation 9.4, we have

In[N(t)] = In[N(O)] + rt, 9.5

so that if we plot the logarithm of the numbers sampled on each day against time,
we expect the numbers to lie on a straight line as they do in Figure 9.1. The slope of
the line gives the mortality rate and the intercept on the vertical axis gives the number
originally present so that both have a direct biological interpretation.
The word 'regression' has an odd etymology. In 1889 Francis Galton formulated
his 'law of universal regression' (Galton, 1889), which stated that 'each peculiarity in
a man is shared by his kinsman, but on the average in less degree'. To illustrate this,
Figure 9.2 shows a plot of the heights of sons against the heights of their fathers
from records of more than a thousand families collected by Pearson and Lee (1903).
The equation of the line shown in Fig. 9.2 is

y = 37.75 + 0.457x. 9.6

The average height of the fathers is 67.0 inches. Using Equation 9.6 to calculate
the corresponding height of the sons gives 68.4 inches, so that on average the sons
were 1.4 inches taller than their fathers. However, we can use Equation 9.6 to show
that a father who is 6 inches above the average height (for fathers) produces, on
average, a son who is only 2.7 inches above average height (for sons) and a father
who is 6 inches below the average height produces, on average, a son who is only
2.7 inches below the average height (for sons). Galton therefore argued that there
was a 'regression' or 'going back' towards the average. Today we use the word 'regres-
sion' to mean fining a line to a plot of one variable against another; the use of the
 Defining the fit 157

Figure 9.2 Plot of the heights of sons against the heights of their fathers from 1078 families.
Each point is the average height of all the sons whose fathers heights are 59, 60, 61. .... inches.

word regression no longer implies Galton's law. (For a discussion of the genetics of
inheritance based on Galton's data, see Roughgarden, 1979, p. 135.)
Once we have decided on the line that gives the best fit to a set of data, we can
interpolate the data and use the fitted line to estimate values of the dependent
variable (son's height in Fig. 9.2) for any value of the independent variable (father's
height in Fig. 9.2). We can also extrapolate the fitted line beyond the extremes of
the data and make the prediction that a father who is 78 inches tall is likely to
produce sons who are 74 inches tall, even though the tallest father whose height
was actually measured was only 75' inches.

9.1 DEFINING THE FIT

When we carry out a regression we want to find the line that gives us the best fit
to our data, but we cannot do this until we decide what we mean by 'best'. We have
already seen in section 4.1 that the location of a distribution can be defined in terms
of the mean, the median or the mode, and which one is 'best' depends on what we
intend to do with our data. The mean has certain rather attractive properties, in parti-
cular, that it is the number about which the sum of the squares of the residual deviations
is a minimum. (We can also show that if the numbers come from a normal distribution
whose true mean is j.l, then the sample mean m is the estimate of j.l that has the
smallest variance.)
Since the mean has this rather desirable property, it makes sense to try a similar
approach in regression and define the 'best fit' as the line that minimizes the sum of
the squares of the residual deviations from the fit. Fig. 9.3 shows a line drawn through
four points. For each point Xi we have the observed value of Y, Yi' and the value of
158 Regr~
swn

Figure 9.3 A straight line has been drawn through four points. Each residual B; is the difference
between the observed value of the dependent variable and the value read from the fitted line.
The best fit line minimizes the sum of the squares of the residuals. a is the intercept on the
y axis when x is zero, b is the slope of the line.

y, read from the line, which we will call Yi. The

difference between these two values
of y is the residual from the fit, which we will call Bj, so that

Bi = Yi - Yi· 9.7
Since Yi is the value of y read from the fitted curve, we can write

B i = Yi - a - bxi• 9.8

We now want to find the values of a and b that minimize the sums of squares of the
residuals, 'L,B;, and it is straightforward to show (Freund, 1972, p. 455) (with the aid
of calculus) that these are

I: y I:~ - I:xyI:x
a=------ 99
nI:~ - (EX)2 .

_ nI:xy - ExI:y
b ------. 9.10
nE~ - (I:X)2
(I have dropped the subscript i from each x and y in Equations 9.9 and 9.10 to make
them easier to write; if you wish, you may put the subscripts in.) The derivation
of Equations 9.9 and 9.10 assumes that the only errors are in yand that the errors
in x are zero or at least negligibly small. The derivation also assumes that all of the Bj
are normally distributed with mean zero and a constant variance.
We can now carry out the regression of the number of armyworms against time
under treatment C. Using Equations 9.9 and 9.10, the logarithm of the number of army-
 ANOV A 159
worms surviving t hours after applying the treatment is

In[N(t)] = 5.647 -1.065t. 9.11

The best estimate of the number of larvae on day 0 is eS.647 = 284 and the larvae
are dying at a rate of 1.065/24 = 0.044, or 4.4%, per hour.

9.2 ANOV A

When we carry out a regression, we want not only to find the best-fit line but also
to know how good the fit is. There are several ways to do this. One way is to carry
out an ANOV A on the fitted line. Just as before, we begin by subtracting the mean
since we are usually concerned only with deviations from the mean and our null
hypothesis is that the 'true' values of yare constant. In Brownbridge's study of the
African armyworm, under treatment C, we start by calculating the total sums of
squares about the mean, SST' and then break this down into a sum of squares due
to the fit, SSF' and a sum of squares due to the residuals from the fit, SSR'
Using Yi for the measured points, Yi for the corresponding points on the fitted
line, after subtracting the mean, and Gj for the residuals, we have

Yi = Yi + Gj. 9.12
Squaring Equation 9.12 and summing over i

:Ey; = :Ey; + 2:EYjGj+ :EG;. 9.13

The left-hand side is the total sum of squares, S,~ the first term on· the right-hand
side is the sum of squares from the fit, SSF' and the last term on the right-hand side
is the sum of squares due to the residuals, S.~ In the Appendix (section 9.9.1) I
show that the middle term on the right-hand side of Equation 9.13 is zero, but since
the residuals Gi are as likely to be positive as negative, it seems reasonable that the
sum over YiGj should be zero, and Equation 9.13 may be written

SST = SSF + SSR' 9.14

Since we have calculated SST as the sum of squares about the mean, we have lost 1
degree of freedom, and

d.fT =n- 1. 9.15

We have then used up one more degree of freedom in setting the slope of the line,
so that for the fitted line we have

dh = 1, 9.16

leaving n - 2 degrees of freedom for the residuals, so that

d.fR = n - 2. 9.17

Now we proceed just as in a conventional ANOV A: dividing each sum of squares

by the appropriate number of degrees of freedom gives us the mean squares. If the
160 Regression

Table 9.2 ANOV A for the line corresponding to treatment C in Figure 9.1

Source
Sum of
squares
Degrees of
freedom
Mean
sq uare F-ratio
Significance
level, P .
Model 11.3402 1 11.3402
.-------
6140 0.000005 ...
Error 0.0055 3 0.00]9

Total 11.3457 4 R2 = 99.95%

mean square for the fit, MSF, is large while that for the residuals, MSR, is small, the
fit accounts for much of the total variance and is significant. If the reverse is the
case, the fit does not account for much of the variance and the fit is not significant.
The F ratio is calculated as

F=MSF/MSR 9.18

and we test our calculated F ratio against Fl.n _ 2'

The ANOV A for the fit to the armyworm data for treatment C is given in Table 9.2
and the value of F is 6140, so that the fit is highly significant. Table 9.2 also gives
a parameter, R2, which is the proportion of the variance explained by the fit: in this
case 11.3402/11.3457 = 0.9995. (Strictly, R2 is the proportion of the total sum of
squares accounted for by the model, but it is referred to as the proportion of the
total variance accounted for by the model.) Since 99.95% of the variance in the data
is explained by the fit, it is not surprising that the fit is highly significant.

9.3 ERROR ESTIMATES AND CONFIDENCE LIMITS

The analysis of variance applied to a regression enables us to determine the statistical

significance of a line fitted to a set of data. However, we would like to determine
error limits for our estimates of the parameters in our fit, especially as we hope to
interpret the parameters biologically. We would also like to place confidence limits
on our fitted line so that we can say that the true line lies within some range about
the fitted line with an appropriate degree of certainty.

9.3.1 Parameters

Equations 9.9 and 9.10 enable us to calculate the regression coefficients for a straight
line fit to a set of data. The derivation of these equations is based on the assumption
that the only errors are those associated with the dependent variable, y, and that
there are no errors associated with the determination of the independent variable, x.
If this is so, we can calculate the errors in a and b corresponding to the errors in y
(Snedecor and Cochran, 1989, p. 175). The standard deviation of the residuals is
calculated in the usual way, as

s~ = r.ef/(n-2), 9.19
 Error estimates and confidence limits 161
Table 9.3 Regression of the natural logarithm of the number of armyworm larvae against
time for treatment C. The table gives the least squares estimate of the slope and intercept
together with its standard deviation, the t statistics and the significancelevel. The slope has
units of days - 1

Parameter Value 5 t P •
Intercept 5.647 0.043 131 < 0.00001 •••
Slope - 1.065 0.014 -78 < 0.00001 •••

where the sum is trom I = 1to n and we have divided by the number ot degrees ot
freedom that remain after fitting two parameters, the intercept and the slope. The
standard deviation of the mean value of the dependent variable, y, is then

s~ = s/n,
~ 9.20

the usual expression for the standard error of a mean. The standard deviation of the
slope, b, is (Appendix, sedion 9.9.2)

s~ = s/~ 9.21
~(xi - £)2.

For the regression of the number of armyworms against time under treatment C,
Table 9.3 gives the standard deviations of the intercept and the slope and if we do
a t test on each we find that both differ significantly from zero at the 0.1% level.
Furthermore, if we multiply the standard deviation of the slope by the critical value
for a two-tailed t test at the 5% significance level, we obtain 95% confidence limits
for the slope as -1.065 ±0.014 x 3.18, so that we can assert that the rate of decline
of the number of armyworms lies between - 1.020 and -1.110/day with 95%
confidence.
In our armyworm experiments we want to know how the rate of decline varies
between treatments to see which, if any, reduces their number significantly. If we fit
straight lines to each of the four sets of data given in Table 9.1, after taking logarithms,
we obtain the rates of decline given in Table 9.4. We test each t statistic using a
one-tailed test since we are interested in the treatment only if it produces a decline
in the numbers over time. In each case there are 3 degrees of freedom since each line
is fitted to five points. The slope of the control curve is not significantly less than
zero while all three treatments give significant rates of decline, B at the 5% level, A
at the 1% level and C at the 0.1% significance level.

Table 9.4 The daily rates of decline in the number of armyworms under the four treatments
described in Table 9.1 calculated from the slopes of straight lines fitted to the data
Treatment Slope 5 t •
Control
B
-
-
0.0019
0.2489
0.0124
0.0642
0.15
3.88 .
ns a
b
A
C
-
-
0.2791
1.0649
0.0284
0.0136
9.83
78.3 ...•• b
c
162 Regression

Figure 9.4 The heavy solid line is the best fit straight line to the measured points. The thin
solid line is a 95% confidence band for the fitted line. The dashed line is a 95% confidence
interval for the individual points so that, on average, 19 out of 20 points should lie between
the dashed lines.

These experiments were conducted in order to determine the concentration of Rt.

to use in armyworm control. To make comparisons between the various treatments,
we calculate the difference between pairs of slopes divided by the pooled standard
deviation of all the slopes, after checking that the variances of the slopes do not
differ significantly. Since we do not know in advance whether the difference between
two slopes will be positive or negative, we use two-tailed tests. Since each slope has 3
degrees of freedom, each comparison has 6 degrees of freedom. Since we can make
=
4(4 - 1)/2 6 comparisons, we need to use a significance level of (5/6)% to ensure
that the significance level for all comparisons is at least 5%. When tnis is done,
treatment B does not differ significantly from treatment A, but all other pairs differ
significantly at the 5% level, as indicated in Table 9.4. In particular, treatment C is
substantially better than treatments A or B.

9.3.2 Lines

Having fitted a line to a set of data, we can also calculate a confidence band for the
line itself which will enable us to set limits within which the 'true' line lies with a
suitable degree of confidence. The uncertainty in the estimate of the line can be
regarded as arising from uncertainty in the estimate of the mean value and uncertainty
in the estimate of the slope. The expression for the standard deviation of the fitted line
at point x is (Appendix, section 9.9.3)

s=s
~ +s
~ ~ x (X-i)2. 9.22

The standard deviation of the fitted line at the mean (x = i) is simply the standard
 Residuals 163
deviation of the mean of y, since the line passes through the point (i,y), and as x
deviates from the mean in either direction, the standard deviation increases because
of the uncertainty in the slope of the line. Again we need to multiply by tn-2 for
a two-tailed test at the 5% significance leve1 to obtain the corresponding 95%
confidence band.
Figure 9.4 shows the data plotted in Fig. 9.2 with a 95% confidence band indicated
by the thin solid lines. In regression tables, such as Table 9.3, we usually specify the
standard error of the intercept on the y axis and this may be obtained by putting
x = 0 in Equation 9.22.

9.3.3 Points

We would also like to set limits on the region within which the individual points
should lie. This will enable us to detect outliers in our data. Equation 9.19 gives the
variance of the residuals from the fitted line and Equation 9.22 the variance of the
fitted line. To calculate a 95% confidence band for the individual points, we simply
add these two variances, take the square root and multiply by tn_2• For our graph
of the heights of sons against the heights of their fathers, this gives the dashed line
in Fig. 9.4. Approximately 19 out of 20 points should lie within the dashed lines.

9.4 RESIDUALS

This book began with a discussion of exponential growth and its influence on Darwin.
Exponential growth is one of the most important ideas in ecology; so let us restate
it: whenever the change in the number of objects is proportional to the number already
present, the increase or decrease in the number follows an exponential curve. If 10
plants produce 1 new plant per week (on average), 20 will produce 2, 50 will produce
5 and so on, and the change in the number per week (1, 2 or 5) is always 1/10th of
the number already present (10, 20, 50). For the same reason, if the birth rate exceeds
the death rate by a constant amount, human populations increase exponentially; if
the rate of inflation is constant, the value of your money falls exponentially; because
the number of radioactive atoms that decay in a given time is proportional to the
number that have not yet decayed, the intensity of radiation emitted by radioactive
isotopes decreases exponentially with time.
Darwin realized that exponential growth cannot go on forever (remember our
calculation on the potential growth of tsetse fly populations in Chapter 1). All natural
populations are constrained by density dependent factors that limit further growth at
a level referred to as the carrying capacity of the habitat for that population. The
convergence to the carrying capacity is itself often exponential.
To illustrate these ideas, consider the experiment of Ashby and Oxley (1935) who
observed the growth of duckweed under controlled conditions in their laboratory
and counted the number of fronds at the beginning of each day. Their data are given
in Table 9.5 and plotted in Fig. 9.5. If the growth is exponential, a plot of In[N(t)]
against t should give a straight line with a slope of r and intercept on the vertical
 164 Regression
Table 9.5 The number of duckweed fronds in a beaker kept at 24°C on each of
14 successive days

Day 1 2 3 4 5 6 7
Number 100 127 171 233 323 452 654
Day 8 9 10 11 12 13 14
Number 918 1406 2150 2800 4140 5760 8250

Figure 9.5 (a) The number of duckweed fronds and (b) the logarithm of the number of
duckweed fronds plotted against time.

axis equal to the logarithm of the number of fronds on day 0 (Equation 9.5). Figure 9.sb
shows that this is indeed the case and the parameters from the fit are given in Table 9.6.
The slope of the line is 0.349±O.OOS/day so that the growth rate is 34.9±O.S%
per day. With this value of r, eft = 2 when t = 1.99 days, so that the number of fronds
doubles every 2 days.
Although the fit in Fig. 9.sb looks good, it is always worth plotting the residuals
as a check. If we do this we obtain the result shown in Fig. 9.6, from which it is clear
that there is something odd going on at the beginning of the experiment. One possi-
bility might be that we need a certain number of fronds to start the process. If we let
the growth rate be proportional to the number of fronds above a threshold value,
say M, we can replace Equation 9.3 by

dN(t)/dt = r(N(t) - M). 9.23

Integrating Equation 9.23 gives

N(t) - M = N(O)ert 9.24

Table 9.6 Parameters for the regression line shown in Figure 9.5b. The slope
of the line is given in units of day -
.
1

Parameter Value 5 t P
Intercept 4.107 0.039 105 < 0.0001 ......
Slope 0.349 0.0046 76 < 0.0001
 Residuals 165

Figure 9.6 The residuals from the plot shown in Figure 9.sb.

so that

In [N(E) - M] = In[N(O)] - rE. 9.25

Since we do not know M, we cannot plot the data in a form that will give us a
straight line directly: what we can do is to use an iterative procedure in which we
repeatedly guess different values of M until we find the value that gives us the
best fit. We do not of course carry out the iteration blindly: if we change M and
the fit deteriorates, we know that we are changing it in the wrong direction. One
way to decide when we have fod~ the best value of M is to do an ANOV A on
each fit and find the F ratio for each value of M. The value of M that maximizes
the F ratio turns out to be 33. If we use this value and plot In (N(t) - 33) against
time, we obtain the result shown in Fig. 9.7; the parameters from the fit are given
in Table 9.7. The slope of the regression line gives a new estimate of the growth

Figure 9.7 (a) Plot of In[N(t) - 33] against time and (b) plot of the residuals from the fit
against time.
166 Regression
Table 9.7 Parameters from the regression of In[N(t} - 33] against time and the corresponding
ANOV A table

Parameter Value 5 t P •
Intercept 3.81 0.021 181 < 0.0001 •••
Slope 0.375 0.0025 151 < 0.0001 •••
Source of Sums of Degrees Mean Significance
variation squares of freedom square F-ratio level, P •
Model 31.95 1 31.95 23 059 < 0.0001 ...
Error 0.017 12 0.0014

Total 31.967 13 R
2
= 99.95%

rate of 0.375 ± 0.0025 fronds/day, slightly lower than before, but with a threshold
of 33 fronds for the process to start. Plotting the residuals in Fig. 9.7b we see that
the overall fit is considerably better than before, although the number of fronds on
day 9 is still some way off.
We now go back to the biology: why are 33 fronds needed for the growth to
start? What would happen in a smaller or a larger beaker? Eventually the numbers
must level off, but at what number of fronds? Can we deduce anything useful from
the fact that the doubling time is about 2 days? The statistics, if calculated carefully,
will raise questions, but the answers can be found only in the biology.

9.5 CORRELATION

We have seen in the previous sections how to obtain the best estimate of one variable
based on the values of another. However both variables are often subject to error
and neither is more fundamental than the other. In such situations we might be
concerned with the degree of correlation between the two variables rather than the
dependence of one on the other.
In section 4.2.1 we defined the variance of a set of numbers as

V = :E(Xj - i)2/n. 9.26

Furthermore, we have repeatedly spoken about variables being 'statistically independent'.

To make this concept of statistical independence more precise, we define a new
parameter, which we call the covariance, as

C(x, y) = :E(Xj - i)(Yi - Y)/n. 9.27

The analogy with the definition of the variance is clear since C(x, x) = V(x), but we
need to convince ourselves that the definition is sensible.
In Equation 9.27 we have, as usual, subtracted the mean of each set of numbers.
If x and yare either both positive Or both negative, after subtracting their means,
their product will contribute positively to the covariance, so that if they go up and
down together, the covariance will be positive. In the same way we can see that if
 Correlation 167
one goes up when the other goes down, or vice versa, they will contribute negatively
to the covariance. We can restate this in terms of our regression and say that if the
regression line has a positive slope, the covariance will be positive, and if the regression
line has a negative slope, the covariance will be negative. We can now write down
a more complete expression for the variance of a sum:

V(x + y) = V(x) + V(y) + 2C(x, y) 9.28

and Equation 4.8, which we have used to evaluate the variance of a sum, is correct
only when the covariance of the two variables is zero.
Just as the standard deviation is often more useful than the variance because it
has the same units as the original measurement, it turns out that a different measure
of association, the correlation coefficient, is often mote useful than the covariance.
Taking the square root of the covariance does not help us much because the units
of the two variables may be different, and if they were, say, metres and grams, we
would have something measured in metres1/2gramsl/2• Indeed, since we have already
accounted for the variance of each measurement through their separate standard devia-
tions, the most useful measure involves dividing the covariance by the standard devia-
tion of each variable so that the correlation coefficient, r, is

r= C(x,y)/[V(x) X V(y)]1I2. 9.29

If the points lie exactly along a straight line with positive slope, the correlation coeffi-
cient is + 1, and if they all lie exactly along a line with negative slope, the correlation
coefficient is - 1. If there is no correlation between the two variables, the correlation
coefficient is O.
The correlation coefficient can be illustrated using the heights of 11pairs of brothers
and sisters (Pearson and Lee, 1903)given in Table 9.8 and plotted in Fig. 9.8.Calcula-
ting the correlation coefficient using Equation 9.29 gives

r = 0.553. 9.30

Since the heights of the brothers and sisters clearly tend to go up and down together,
the correlation coefficient is positive but the relationship is not very close and the
correlation coefficient is much less than 1.
When we calculate a correlation coefficient, or perform a regression or an ANOV A
we are looking for relationships between two sets of numbers. In Fig. 9.8 the line
marked (a) is a regression of the heights of the brothers against the heights of their
sisters. However, if we carry out a regression of the heights of the sisters against the

Table 9.8 Heights, in inches, of 11 brothers and sisters, each pair belonging to a different
family

Family 1 2 3 4 5 6 7 8 9 10 11

Brother 7I 68 66 67 70 7I 70 73 72 65 66
Sister 69 64 65 63 65 62 65 64 66 59 62
168 Regression

Figure 9.8 The heights of sisters plotted against the heights of their brothers using the data
in Table 9.8. The two lines are regressions of brothers against sisters and sisters against
brothers.

heights of their brothers and plot it on the same graph, we obtain the line marked (b) in
Fig. 9.8. The reason that we can obtain two quite different regressions for the same
set of data is because each regression is based on the assumption that the errors are
only in the dependent variable and that the independent variable is measured without
error. If the variability in the data arose only from the heights of the boys, line (a) would
be correct; if variability in the data arose only from the heights of the girls, line (b)
would be correct. Consequently, line (a) gives the best prediction of the height of a
brother given the height of his sister, and line (b) gives the best prediction of the
height of a sister given the height of her brother. If we are not interested in using
heights of sisters to predict heights of brothers, or vice versa, but rather are interested
in the degree of association between the two, neither regression is then 'correct' since
both the boys' and the girls' heights are subject to random variations. In this case
the advantage of the correlation coefficient is that it is Jlot biased in favour of either.
In fact, it is easy to show that the correlation coefficient is the geometric mean of
the slopes of the two regressions lines, so that the correlation coefficient measuring
the degree of association between the height of brothers and sisters is

r=bIb2, 9.31

where bl is the slope of the regression of brothers against sisters and b2 is the slope of
the regression of sisters against brothers.
Finally, we can make a connection between the correlation coefficient and the
ANOV A of our regression. If the correlation between two variables is perfect, the
correlation coefficient should be ± 1. But the regression line should then explain all
of the variance in the data, and the parameter, which we have called R2 in Tables 9.2
 Polynomial regression 169

and 9.7, should also be 1. If there is no correlation between the two variables at all,
the correlation coefficient should be zero. But the regression should then explain none
of the variance in the data and R2 should be O. In fact, it is not difficult to show that
Rand r are one and the same, so that the square of the correlation coefficient
is the proportion of the total variance explained by the fit.

9.6 POLYNOMIAL REGRESSION

When we carry out a linear regression, we assume that the points lie along a straight
line although we sometimes need to transform the data, perhaps by taking logarithms
to achieve a linear relationship. If we cannot transform the data to lie along a straight
line, we must use a more complicated function. One way to proceed is to fit a poly-
nomial, which could be a quadratic function (a parabola), a cubic function or a still
higher order polynomial. Consider, for example, the data measured by Crooke and
Knight (1965) in a study of the relation between the age of oat plants and their content
of nickel and iron when grown in sand cultures. Like other heavy metals, nickel
supplied in excess to plants affects iron metabolism causing, in oat plants, longitudinal
white necrotic striping of the leaves accompanied by an induced iron deficiency
chlorosis.
As part of their study, Crooke and Knight measured the relative absorption of
nickel and iron from a controlled nutrient solution supplying iron at 1.2 ppm and
nickel at 2.5 ppm, with the results shown in Table 9.9 and ploUed in Fig. 9.9. The
data lie along a curve and if we decide to fit a parabola our model equation is

Yi=a+bxi+cx+e
~ i. 9.32

As in the case of linear regression, we can again find expressions for the values of
a, band c that minimize the sums of squares of the residuals about the fitted line
(Bliss, 1967, p. 37); the resulting values are given in Table 9.10 and the line is drawn
in Fig. 9.9. The constant term is not significantly different from zero while both the

Table 9.9 The ratio of the amount of nickel to the amount of iron in oat plants as a function
of age

Agel days Nil Fe Agel days Nil Fe Agel days Nil Fe

4 0.32 38 1.40 51 1.32

9 0.41 39 1.95 52 1.52
14 0.79 41 1.51 55 1.05
18 0.86 42 1.81 56 1.70
22 1.28 69 0.82 58 1.22
26 1.28 73 0.95 59 1.90
29 1.48 44 1.53 62 0.85
33 1.15 45 1.30 63 1.05
34 1.47 48 1.00 64 0.87
37 1.67 49 1.50 66 1.08
170 Regression

Figure 9.9 The data of Table 9.9 and the parabola that gives the best fit.

Table 9.10 Parameters for the regression line shown in Figure 9.9 and the ANDV A table. a
is the intercept on the y axis, band c are the coefficients of the linear and quadratic terms in the fit

Parameter Value s t P .
a
b
- 0.047
0.0735
0.197
0.01048
- 0.240
7.021
0.812
0.0000 ......
ns

c - 0.000866 0.000129 - 6.7358 0.0000

Source
Sum of
squares
Degrees of
freedom
Mean
square F-ratio
Significance
level, P .
Model 3.04 2 1.52 24.7 0.0000 ...
Error 1.66 27 0.0616

Total 4.71 29 R2 = 65%

linear and the quadratic terms are significant at the 0.1% level. From Fig. 9.9 we see that
the ratio of nickel to iron increases up to age 42 days and then decreases as the trees
age. We can again carry out an analysis of variance on the fitted line, from which
we find that the model explains 65% of the variance in the data.
In our polynomial regression, the estimate of each coefficient depends on the
inclusion of the others. If we fit a constant we will obtain an estimate of a. If we fit
a straight line we will obtain an estimate of both a and b but we will find that the
estimate of a has changed. If we fit a parabola we will obtain an estimate of a, b,
and c but we will find that the estimates of a and b have changed. Formally we say
that the constant, x, and y} are not orthogonal. The ANOV A tells us that the overall
fit using all three terms is highly significant. The regression analysis gives us the
significance level of each coefficient, after allowing for the other two. We should
 Multiple regression 171
therefore consider first the significance of the quadratic term, c, in Table 9.10. If this
is significant, we should keep all three terms. If it is not, we should drop it and repeat
the exercise for the first two terms. We should drop terms of lower power first only
if we have a sound biological reason for doing so. For example, it may be that we
know that the graph must pass through the origin of our coordinate system, in which
case we could drop the constant term.

9.7 MULTIPLE REGRESSION

Very often we would like to carry out a regression against more than one independent
variable. Leyton (1956) studied the relation between growth and the mineral composi-
tion of the foliage of Japanese larch trees. A stand of trees had been planted in hetero-
geneous soil using variable planting stock. Leyton was concerned to discover the
causes of the resulting variability in the height of the trees. Twenty-six trees were
selected, ranging in height from 65 to 351 em. From each tree one lateral shoot was
taken from the uppermost whorl immediately below the terminal shoot and after
drying the needles in an oven they were ground in a mill and analysed for nitrogen,
phosphorus, potassium and residual ash. The data are given in Table 9.11.
We now want to do a regression of height against each of the other variables so
that our model is

Yi = a + bN + CPj + d~ + eA + 8j
j j 9.33

and we can again find expressions for the coefficients of the fit that minimize the sums
of squares of the residuals (Bliss, 1967, p. 308). The results are given in Table 9.12
where only the coefficients of nitrogen and potassium are significant. We might try
omitting phosphorus and residual ash, but if we do this the other' coefficients change,
as in our polynomial regression, because the independent variables are not orthogonal.
(To test if two variables, say Nj and ,~ are orthogonal, we calculateI:.Nj~ and if
the answer is zero they are orthogonal. If we are able to choose the levels of the
independent variables we choose them to ensure orthogonality. In the case of our
larch trees, we simply have to use the measured values.)
Table 9.13 gives the correlation between each pair of variables. We see that the
level of nitrogen is positively correlated with the level of each of the other variables.
Any variable, in our case the height of the trees, that goes up and down with the
level of nitrogen will also go up and down with the level of potassium, for example.
In other words, some of the variation that is explained by variations in the level of
nitrogen can also be explained by variations in the level of potassium. Consequently,
if we first perform a regression against nitrogen alone and then perform ,a regression
against nitrogen and potassium, some of the variation explained by nitrogen in the
first regression will be explained by the potassium in the second regression and the
coefficient for nitrogen will change.
We need to proceed carefully and there are two standard approaches to carrying
out regressions when the independent variables are correlated. The first approach is
forward step-wise regression. We carry out a regression against each independent
 172 Regression
Table 9.11 The concentraHon, measured in parts million, of nitrogen, phosphorus, potassium
and residual ash in the needles of 26 Japanese larch trees with their heights in centimetres

Nitrogen Phosphorus Potassium Ash Height

2.20 0.417 1.35 1.79 351
2.10 0.354 0.90 1.08 249
1.52 0.208 0.71 0.47 171
2.88 0.335 0.90 1.48 373
2.18 0.314 1.26 1.09 321
1.87 0.271 1.15 0.99 191
1.52 0.164 0.83 0.85 225
2.37 0.302 0.89 0.94 291
2.06 0.373 0.79 0.80 284
1.84 0.265 0.72 0.77 213
1.89 0.192 0.46 0.46 138
2.45 0.221 0.76 0.95 213
1.88 0.186 0.52 0.95 151
1.93 0.207 0.60 0.92 130
1.80 0.157 0.67 0.60 93
1.81 0.195 0.47 0.57 95
1.49 0.165 0.66 0.80 147
1.53 0.226 0.68 0.66 88
1.43 0.224 0.44 0.45 65
1.54 0.271 0.51 0.95 120
1.13 0.187 0.38 0.63 72
1.63 0.200 0.62 1.10 160
1.36 0.211 0.71 0.47 72
1.76 0.283 0.96 0.96 252
2.53 0.284 0.85 1.39 310
2.59 0.303 1.02 0.95 336

Table 9.12 The regression coefficients for the model given by Equation 9.33 applied to the
data in Table 9.11

Chemical N P K A

Coefficient 97.76 256.97 126.57 40.28

s 24.57 169.91 46.43 36.61
t
• ...
3.98 1.51
ns
2.73
•
1.10
ns

variable separately and include the one that explains the greatest proportion of the
variance. We then perform regression keeping this variable and including each of the
others, one at a time, to see which variable explains the greatest proportion of the
remaining variance. In other words, we find the most significant variable and include
it. We then find the next most significant variable and include that. We proceed in
this way until none of the excluded variables contributes significantly.
 Multiple regression 173
Table 9.13 Correlation coefficients for the concentrations of nitrogen, phosphorus, potassium
and ash in Japanese larch trees

Nitrogen Phosphorus Potassium Ash

Nitrogen 1.00 0.60 0.55 0.65
Phosphorus 0.60 1.00 0.71 0.67
Potassium 0.55 0.70 1.00 0.67
Ash 0.65 0.67 0.67 1.00

The second approach is backward step-wise regression. We start by including all

the variables and then drop the least significant variable. We then recalculate the contri-
butions for the remaining variables and again drop the least significant, continuing in
this way until all of the included variables are significant.
Unfortunately, there is not usually a single 'correct' answer and forward and back-
ward step-wise regressions can leave us with different sets of variables. In practice
it is worth doing both and if they give different answers you will have to judge
between them on the basis of your knowledge and understanding of the biology of
the problem.
Table 9.14 illustrates the forward step-wise regression. We start with none of the
variables in the model and F-to-enter gives the significance level that we will obtain
if we include each variable singly. Since we have 26 points, each of these F values
should be compared with Fl.24 at the 5% level, which is 4.26, and each of them, taken
on its own, is highly significant. Since nitrogen has the biggest F ratio, we include
it first. The next part of the table shows the effect of each of the other variables,
after allowing for the effect of nitrogen. Because nitrogen has already explained much
of the variance, the contributions from each of the other variables is now much less

Table 9.14 F-to-enter, F-to-remove and the final model for the forward step-wise regression
of the height of Japanese larch trees against the amount of nitrogen, potassium, phosphorus
and residual ash. The table also gives the regression coefficients for the variables that are
included in the regression

Variables Variables
in model Coefficient F-to-remove not in model F-to-enter
Nitrogen 48.6
Phosphorus 35.9
Potassium 40.3
Ash 34.4

Nitrogen 183.4 48.6 Phosphorus 13.8

Potassium 24.1
Ash 9.4

Nitrogen 123.3 30.2 Phosphorus 3.39

Potassium 188.7 24.1 Ash 2.22
174 Regression
Table 9.15 F-to-enter, F-to-remove and the final model for the backward step-wise regression
of the height of Japanese larch trees against the amount of nitrogen, potassium, phosphorus
and residual ash

Variables Variables
in model Coefficient F-to-remove not in model F-to-enter
Nitrogen 97.8 15.8
Phosphorus 257.0 2.29
Potassium 126.6 7.43
Ash 40.3 1.21

Nitrogen 107.8 22.1 Ash 1.21

Phosphorus 304.2 3.39
Potassium 143.1 10.5

Nitrogen 123.3 30.2 Phosphorus 3.39

Potassium 188.7 24.1 Ash 2.22

than before. We now compare each of the F values with Fl.23' since we have used
up another degree of freedom, and this is 4.28. Potassium now makes the most signi-
ficant contribution so we include it also. In the last part of the table we see that the
two remaining variables are no longer significant since the variance that they would
have explained is already accounted for by nitrogen and potassium, and we omit
them from the model. We also see in Table 9.14 that when nitrogen is included on
its own, the regression coefficient is 183.4, but when we include potassium in the
regression, the coefficient for nitrogen is reduced to 123.3 because part of the increase
due to nitrogen is now explained by the increase due to potassium.
Table 9.15 shows the corresponding sequence of steps for the backward stepwise
regression and, in this case, we end by seleding the same two variables. Table 9.16
is the ANOV A for the final model.
Leyton (1956) concludes from his analysis of these data that the trees are suffering
from a lack of nitrogen and potassium so that in those areas where the concentrations
of these nutrients happen to be relatively high, the trees grow significantly taller.
This experiment illustrates the care that must be taken when drawing conclusions
from regressions. If Leyton had measured phosphorus but not nitrogen or potassium,
it would have appeared that a significant part of the variation in growth could be
explained on the basis of a phosphorus deficiency whereas the truth seems to be
that the trees are deficient in nitrogen and potassium and that the concentration of
phosphorus is correlated with both of these.

9.8 SUMMARY

ANOV A and regression are closely related and we can formulate ANOV A problems
as regression problems or analyse regression problems using ANOV A. We use ANOV A
when the independent variables are categorical. We use regression when the independent
 m
~~ I~

Table 9.16 Parameters for the regression of the height of Japanese larch trees against the
concentration of nitrogen and potassium
Parameter Value s t P •
Intercept - 181 36.9 -4.9 0.0001 ......
Nitrogen
Potassium
123
189
22.4
38.4
5.5
4.9
0.0000
0.0001 ...
Sum of Degrees of Mean Significance
Source squares freedom square F-rafio level, P •
Model 191184 2 95592 59.8 0.0000 ...
Error 36 771 23 1599
Total 227954 25 2
R = 0.84

variables are measured on a quantitative scale. Sometimes the independent variables

may be of both kinds, some qualitative, some quantitative. In this case we can carry
out a combined analysis called analysis of· covariance or ANCOV A, in which the
covariates are the quantitative variables. Brownbridge's armyworm data are an example
that we might choose to analyse using ANCOV A with the treatments as categorical
variables and time as a continuous variable against which we carry out a regression
(see Example 9.10.6). In fact, we can go further and treat ANOV A, regression and
the combination, ANCOV A, as particular cases of a wider approach called generalized
linear modelling (Dobson, 1983; McCullogh and Neider, 1983). It is important,
however, to be clear as to just what each technique does and it is useful to think of
ANOV A and regression as separate but related techniques.
The concept of interactions applies to regression just as to ANOV A. In regression
problems we simply include new variables that are products of the original variables.
In the example of Japanese larch trees we might, for example, have tried a term of
the form N x K to see if the response to the levels of nitrogen depends on the levels
of potassium and vice versa.
An important extension of these ideas involves the use of weights in ANOV A
and regression. We often have data for which the different points are measured with
differing accuracy. In this case, instead of finding the model that minimizes the sum
of the squares of the residuals, I:.e;,
we find the model that minimizes j I:.e;w
where
we choose the weights Wi so that the more precise observations are given greater
weight than the less precise observations. The best choice, in a least-squares sense,
is to set the weights equal to the reciprocal of the variance so that if you know how
the variance varies across your data you can weight the separate measurements accor-
dingly (Hays, 1988, p. 624).
The ideas discussed in this book contain many of the most important principles
of statistical analysis. However, there are a number of techniques collectively known
as multivariate methods. Multivariate analysis of variance, MANOV A deals with
problems in which we have more than one response variable. Cluster analysis allows
us to group individuals, for each of which a number of variables have been measured,
176 Regression
into hierarchical clusters starting from those that are most similar. Discriminant
analysis allows us to find the combination of a set of variables that will best
discriminate between the individuals of interest. Principle components analysis
provides a method for combining a large number of variables into a smaller number
of variables, each of which is a linear combination of the original variables. For further
reading, the book by Jeffers (1978) provides a good introduction and the books by
Manly (1986) and Green (1978) tell you how to carry out multivariate analyses.
The best advice I can leave you with comes from Max Born (Margenau, 1950):
'I believe that there is no philosophical high-road in science with epistemological
sign-posts. No, we are in a jungle and find our way by trial and error, building our
road behind us as we proceed. We do not find sign-posts at cross-roads, but our
own scouts erect them to help the rest ... My advice to those who wish to learn the
art of scientific prophecy is not to rely on abstract reason, but to decipher the langu-
age of Nature from Nature's documents, the facts of experience:

9.9 APPENDIX

9.9.1 Regression sums of squares

We wish to partition the sums of squares of the points in a linear regression into a
contribution from the fitted line and an error term. Since we are concerned with the
variation about the mean, as in the standard ANOV A, we subtract the mean value
of the dependent variable from each measured value Yi' Similarly, we can subtract
the mean value of the independent variable, from each measured value of Xi' Then
I:Yi and I:Xi are both zero and from Equation 9.9 a is also zero. The equation of the
regression line is then

Yi = bXi 9.34
and the middle term on the right-hand side of Equation 9.13 is

I:Yi6i = I:bxj(bxi -- y) = b(bI:X~ - I:XiY) 9.35

However, from Equation 9.10 with I:Yi and I:xi equal to zero
b = I:xiYJI:~ 9.36

so that the right-hand side of Equation 9.35, and hence I:Yi6i and the middle term
on the right-hand side of Equation 9.13 are zero.

9.9.2 Variance of the slope

To calculate the variance of the slope in a linear regression we will again subtract the
mean values of X and Y from each data point since this will not affect the slope. Then

b=I:~
~ 7
~
 Eu~ 177
as in Appendix, section 9.9.1. But

2
V(ay) = a V(y) 9.38
so that

V(b) = :EV~ (Y)/CE~)2 = V(y)/:E~ = Sp 2 /:E~ 9.39

as in Equation 9.21 with x equal to Xi - i.

9.9.3 Variance of the fitted line

To calculate the variance of the fitted line we again subtract the mean value of x
and y from each data point. Then

(Yi - y) = b(x j - i) 9.40

since a is zero. Then

V(YJ = V(Yj) + V(b)(xi - i)2 9.41

since we assume that the only errors are in y and V(x) = o. Therefore

V(y) = s~ + S(Xi
~ - i)2 9.42

9.10 EXERCISES

1. Use Equations 9.19 to 9.22 to calculate 95% confidence limits for the residuals
plotted in Fig. 9.7b obtained by fitting Equation 9.25 to the data given in Table 9.5.
Decide if the large residual on day 9 differs significantly from the fitted line. (Since
you are only considering the largest residual you should use a significance level of
5/14% to obtain a 95% confidence band.)

2. In studies of evolution it is important to be able to fix the rate at which evolutionary

changes occur. Given two related species, we should then be able to estimate the
time since they began to diverge. Fitch counted the mean number of nucleotide
substitutions in seven proteins in 15 related species for which it was possible to
determine the time at which they diverged from the fossil record, with the results given
in Table 9.17 (Ruse, 1982).
(a) Calculate the mean number of nucleotide substitutions per million years and
determine the time at which two species differing by an average of 40 nucleotide
substitutions diverged. (b) Calculate 95% confidence limits for the points and place
error limits on the answer to (a). [Think carefully about which variable to take as the
independent variable.]

3. Large islands generally support more species than comparable small islands and
the number of species, N, generally follows a power-law relationship with the area
of the habitat, A, so that
z
N=aA • 9.43
178 Regression
Table 9.17 The mean number of nucleotide substitutions in seven proteins
in 15 related species and the time since they diverKed
Nucleotide Time/ M years Nucleotide Time/ M years
subsfitufions substitutions
7.91 1.65 40.29 53.14
4.46 7.49 48.92 63.62
4.17 11.23 46.76 61.23
1.44 13.77 56.83 75.60
17.27 15.87 66.19 70.36
7.19 19.76 73.67 90.57
15.83 32.93 97.84 121.26
33.09 41.92

The exponent in the power law, z, typically lies between 0.24 and 0.33 (Pianka,
1983). Davis et al. (1988) studied assemblages of forest-dwelling species of small,
non-flying mammals on the mountains of New Mexico and adjacent parts of Arizona.
These montane islands of forest are surrounded by woodlands and grasslands that
isolate one from another. Their results are given in Table 9.18. Carry out a regression
of In N against In A, decide if the data show a power-law relationship and determine
the value of z. [Hint: Since the logarithm of 0 is - 00 there is a problem when there
are no species. To avoid this Davis et al. (1988) added 1 to each of the species
numbers before taking logarithms.]
The authors argue that species richness decreases with the distance from the original
source of the various species, which they assume to be in the southern extension of
the Rocky Mountains. They therefore also measured the shortest distance from there
to the individual forests while remaining within suitable habitats and this is given
as the 'isolation' in Table 9.18. Carry out a regression of the residuals from the previous

Table 9.18 The area, the number of species and the isolation of 12 montane
forest islands
Foresf Area/km2 No. species lsolafion/km

1 21813 10 241
2 2872 7 217
3 346 6 75
4 2575 6 292
5 885 4 72
6 2090 4 378
7 1279 3 173
8 750 3 281
9 173 3 271
10 205 2 269
11 57 0 475
12 8 0 392
 Eu~ 179
Table 9.19 The mean antler length and shoulder height of various cervine deer. All
measurements in inches. The last two numbers, shown in bold, are the values for the Irish FIle
- - ---- ------ -~
&&

Antler 63.6 51.1 48.7 44.9 42.4 39.5

Shoulder 65.0 50.4 55.0 54.5 45.5 50.9
Antler 40.2 36.1 34.0 31.8 28.3 26.3
Shoulder 36.8 40.9 45.7 36.3 34.8 32.5
Antler 22.9 23.8 20.9 13.4 9.7 94.4
Shoulder 39.1 27.5 34.0 27.1 27.0 72.9

fit against the isolation and see if you agree that the isolation also influences species
richness. (For discussions of the species-area relationship see Diamond and May,
1981; MacArthur and Wilson, 1963; Brown and Lomolino, 1989.)

4. The Irish Elk, which Gould (1974, 1977) points out is neither exclusively Irish nor
an elk, became extinct in Ireland 11 000 years ago. It has become famous for the
prodigious size of its antlers whose span could reach 12 feet. Early in this century
the Irish Elk was the centre of a fierce dispute between proponents and opponents of
Darwin's theory of evolution for it was hard to see how such large antlers could
confer a selective advantage on the animal and seemed more likely to lead to the
animals' extinction. To decide if the antlers really were too big in proportion to the
body size, Gould (1973) measured the spread of the antlers and the shoulder heights
of various cervine deer, the group to which the Elk belongs, with the results given
in Table 9.19. Plot a graph of the logarithms of the spread of the antlers against the
logarithms of the shoulder heights. Fit a straight line to all of the data excluding
that of the Irish Elk and decide if you agree with Gould that the antlers of the Irish
Elk are only as big as one would expect in a deer of that size.
Calculate 95% confidence limits'for the points and determine how big (or smaIl) the
average spread of the antlers would have to be in order to reject the hypothesis that
the antlers are about as big as one would expect. Do the data provide support for
Gould's argument?

5. MacFadden (1986) carried out a study to investigate the rate at which the body
size of horses has changed over evolutionary time. It is difficult to estimate the body
size of fossil horses as only fragments are usually found. MacFadden therefore began
by studying the relationship between various bodily dimensions for New World
Equidae in order to find the most useful measure of body size one could obtain from
fossil remains. In Table 9.20, for example, the head-body length is compared with
the length of the row of cheek-teeth. In general, the cheek-teeth are among the better
preserved parts of fossil skeletons. The data have been read off a graph in MacFadden's
paper. Carry out a regression of the head-body length against the length of the row
of cheek-teeth and see if the latter can be used as a good estimator of the former.
Using the data in this way MacFadden was able to estimate the body sizes of 40
fossil horses covering about 55 million years as shown in Table 9.21. Plot a graph
of the logarithm of the body mass against the age of the fossils. MacFadden concluded
180 Regression
Table 9.20 HBL is the head-body length measured from the tip of the nose to the beginning
of the first caudal vertebra and TRL is the overall length of the row of cheek-teeth from the
antero-most occlusal enamel of tooth P2 to postero-most occlusal enamel of tooth M3. All
measurements in millimetres

TRL 39.4 43.6 45.6 46.7 68.9 67.2 66.9

HBL 648 648 691 648 843 911 988
TRL 84.7 85.0 90.0 105 115 148 140
HBL 1221 1306 1349 1306 1603 1709 1752
TRL 140 158 161 173 170 173 170
HBL 1900 1900 2109 2240 2304 2304 2346
TRL 183 183 183 188 189 193 190
HBL 2648 2588 2542 2499 2448 2448 2448

Table 9.21 The mean body masses and the ages of 40 fossil horses
Number 1 2 3 4 5
Myr 57.53 52.77 52.77 50.26 50.26
Mass/kg 31.87 25.75 40.16 32.74 23.57
Number 6 7 8 9 10
Myr 45.49 45.53 31.91 30.64 25.87
Mass/kg 31.87 44.52 47.58 48.02 53.25
Number 11 12 13 14 15
Myr 23.40 21.91 20.85 18.43 18.38
Mass/kg 101. 71 115.24 93.41 78.13 47.58
Number 16 17 18 19 20
Myr 18.38 11.06 14.81 13.49 15.96
Mass/kg 132.70 412.94 199.92 268.45 69.84

Number 21 22 23 24 25
Myr 15.79 15.83 13.45 11.06 13.62
Mass/kg 85.56 53.69 78.57 122.66 139.68

Number 26 27 28 29 30
Myr 8.55 14.77 11.06 13.57 13.49
Mass/kg 61.11 108.69 153.21 170.24 161.51

Number 31 32 33 34 35
Myr 12.34 8.72 6.04 6.04 3.62
Mass/kg 137.50 235.71 244.01 313.85 84.68

Number 36 37 38 39 40
Myr 5.15 2.38 2.47 2.47 2.47
Mass/kg 434.76 436.94 501.98 351.83 457.02
 Exercises 181
Table 9.22 Numbers of armyworm larvae sampled in each plot on each day after the start
of treatment. Note that treatment C actually started one day after treatments A and B. •
indicates missing data
Row 1 1 1 2 2 2 3 3 3
Column 1 2 3 1 2 3 1 Z 3
Treatment B A C C B A A C B
0 331 425 420 287 390 290 267 214 197
No. 1 242 348 200 64 223 173 138 63 178
on 2 213 306 43 26 109 115 105 31 71
3
day
4
•
154
•
204
20 8 • • • 10 •
• • 102 106 48
• 88

that there may not have been much evolutionary change up to 35 million years ago
but that since then the body size has been increasing at a rate of 5.6%/Myr if we
exclude species 26 and 35, which seem to be especially small. These two species do
in fact belong to a different Nannipus species from the others so that it might be
reasonable to exclude them. Carry out a regression analysis to see if you agree with
his conclusions.

6. The actual data that Brownbridge (1988) collected in his experiment on the biocidal
effects of Bacillus thuringiensis on the armyworm Spodoptera exempta are given in
Table 9.22. The experiment used a Latin square design with three rows and three
columns laid out in the field under study. Considerable attention was given to the
sampling process. Within each plot 20 quadrats, each 20cm x 20cm, were chosen at
random on each day and the numbers of larvae in the quadrats were counted. The
outer 2 m of each plot were excluded from sampling to minimize effects arising from
movement of larvae into or out of each plot. As a control. 30 quadrats were chosen
at random on each day from the untreated areas around and between the treated
plots (Table 1.3).
Unfortunately, as a result of heavy rainfall at the time of spraying, only two
treatments (A and B) were applied on Day O. Treatment C was applied for the first
time on Day 1. Furthermore, counting had to be abandoned early on Day 3 because
the farm manager wished to spray the experimental area with pyrethrum as there were
still a lot of armyworms in the untreated areas and the risk of damage to the rest
of his crop was too great. On that day counts were made only for the three plots
under treatment C. After summing the larval numbers over the quadrats in each experi-
mental plot, the data given in Table 9.23 were recorded.
The data were analysed, after taking logarithms, using Mini Tab, which is able to
handle missing data points. The rows, columns and treatments were qualitative factors
and time was specified as the covariate so that this is an example of ANCOV A.
Table 9.23 gives the result of the analysis.
Because of the missing values, the factors in the analysis are not orthogonal. In
the means table, Time' gives the average slope and Time x TrX gives the deviation
from the average slope for treatment X.
182 Regression
Table 9.2.3 Analysis of Brownbridge's armyworm data using the general linear modelling
facility in MiniTab.

Source
Sums of
squares
Degrees of
freedom
Mean
square F ratio
Significance
level, P .
Rows 3.788 2 1.894 34.69 < 0.001 ....
Columns 0.375 2 0.188 3.44 0.047
Treatments
Time
0.031
17.905
2
1
0.D15
17.905
0.28
327.94
0.755
< 0.001 ......
ns

Time x Treat 7.683 2 3.842 70.36 < 0.001

Error 1.420 26 0.055

Coefficient s t P .
Constant 5.630 0.062 90.82 < 0.001 ......
Time
Time x Tr.A
- 0.533
0.255
0.029
0.040
- 18.11
6.46
< 0.001
< 0.001 ......
Time x Tr.B
Time x Tr.C
0.285
- 0.541
0.040
0.040
7.22
13.53
< 0.001
< 0.001 ...
Make sure that you understand how each term in the table is calculated. Determine
the slopes under each treatment and their confidence limits and compare these with
the values given in Table 9.4. (When I presented the data in Table 9.1 I used the fit
from Table 9.23 to determine best estimates of the missing values and then added
suitably chosen random numbers to these estimates to avoid the problem of missing
values.) Note that the sums of squares and mean squares given here are given as the
adjusted sums of squares (ASS) and adjusted mean squares (AMS) in MiniTab.
10

Tables

These tables were calculated using routines in the NAG Library provided by NAG
Ud. (For address see Preface.)

Table 10.1 Critical values for Student's f distribution. P gives the value of the c.dJ., that is
the probability that a number chosen from a f distribution is less than the value in the body
of the table for the appropriate number of degrees of freedom. At the foot of the table SL 1
and 2 give the significance levels for one- and two-tailed tests, respectively. Student's f distribu-
tion with an infinite number of degrees of freedom is identical to the standard normal distribution

Degrees of P
freedom 0.95 0.975 0.99 0.995 0.999 0.9995

1 6.31 12.71 31.82 63.66 318.31 636.59

2 2.92 4.30 6.96 9.92 22.33 31.60
3 2.35 3.18 4.54 5.84 10.21 12.92
4 2.13 2.78 3.75 4.60 7.17 8.61
5 2.02 2.57 3.36 4.03 5.89 6.87

6 1.94 (/2.45') 3.14 3.71 5.21 5.96

7 1.89 2:36" 3.00 3.50 4.79 5.41
8 1.86 2.31 2.90 3.36 4.50 5.04
9 1.83 2.26 2.82 3.25 4.30 4.78
10 1.81 2.23 2.76 3.17 4.14 4.59

15 1.75 2.13 2.60 2.95 3.73 4.07

20 1.73 2.09 2.53 2.85 3.55 3.85
30 1.70 2.04 2.46 2.75 3.39 3.65
60 1.67 2.00 2.39 2.66 3.23 3.46
120 1.66 1.98 2.36 2.62 3.16 3.37

Infinity 1.65 1.96 2.33 2.58 3.09 3.29

SL 1 tail 5% 2.5% 1.0% 0.5% 0.1% 0.05%

SL 2 tail 10% 5.0%", 2.0% 1.0% 0.2% 0.10%
184 Tabks
Table 10.2 Critical values for the X2 distribution. P gives the value of the c.d.f., that is the
probability that a number chosen from a X2 distribution is less than the value in the body of
the table for the appropriate number of degrees of freedom. SL gives the significance level
for a one-tail test. To evaluate significance levels for a critical value of X2 equal to x, say,
when the number of degrees of freedom exceeds 100, calculate y = (2x)1/2 - (2f - 1)1/2and then
look up the significance level of y for a standard normal distribution

Degrees of P
freedom 0.05 0.1 0.9 0.95 0.99 0.999

1 0.0039 0.016 2.71 3.84 6.63 10.83

2 0.103 0.211 4.61 5.99 9.21 13.82
3 0.352 0.584 6.25 7.81 11.34 16.27
4 0.711 1.06 7.78 9.49 13.28 18.47
5 1.15 1.61 9.24 11.07 15.09 20.52

6 1.64 2.20 10.64 12.59 16.81 22.46

7 2.17 2.83 12.02 14.07 18.48 24.32
8 2.73 3.49 13.36 15.51 20.09 26.12
9 3.33 4.17 14.68 16.92 21.67 27.88
10 3.94 4.87 15.99 18.31 23.21 29.59

11 4.57 5.58 17.28 19.68 24.72 31.26

12 5.23 6.30 18.55 21.03 26.22 32.91
13 5.89 7.04 19.81 22.36 27.69 34.53
14 6.57 7.79 21.06 23.68 29.14 36.12
15 7.26 8.55 22.31 25.00 30.58 37.70

16 7.96 9.31 23.54 26.30 32.00 39.25

17 8.67 10.09 24.77 27.59 33.41 40.79
18 9.39 10.86 25.99 28.87 34.81 42.31
19 10.12 11.65 27.20 30.14 36.19 43.82
20 10.85 12.44 28.41 31.41 37.57 45.31

21 11.59 13.24 29.62 32.67 38.93 46.80

22 12.34 14.04 30.81 33.92 40.29 48.27
23 13.09 14.85 32.01 35.17 41.64 49.73
24 13.85 15.66 33.20 36.42 42.98 51.18
25 14.61 16.47 34.38 37.65 44.31 52.62

26 15.38 17.29 35.56 38.89 45.64 54.05

28 16.93 18.94 37.92 41.34 48.28 56.89
30 18.49 20.60 40.26 43.77 50.89 59.70
40 26.51 29.05 51.81 55.76 63.69 73.40
50 34.76 37.69 63.17 67.50 76.15 86.66

60 43.19 46.46 74.40 79.08 88.38 99.61

70 51.74 55.33 85.53 90.53 100.43 112.32
80 60.39 64.28 96.58 101.88 112.33 124.84
90 69.13 73.29 107.5 7 113.14 124.12 137.21
100 77.93 82.36 118.50 124.34 135.81 149.45

SL 95% 90% 10% 5% 1% 0.1%

oft
Il'>
..•..
o .a. .•.••.•
0 0<')0<')'0
o<')ll'>ll'>'Oo<') 'O~0\1'-o.1l'>
'<1' 00<') •.••.•
1'-0.0<')0<')•.••.•
0<')•.••.•
'<1'•.••.•~ •.••.•
0\ Il'> 0
0 <Q 1'-0. 'Oll'>o<')~ 0
Qj
';:
~'"~';~
oIl'>•.••.• "";"";r-ir-ir-i r-ir-ir-i...<...< ~ ~ H H ...<
> ~
....•• !"oj
~
Q)
u
i::
<~Q1'-o.0\0<')
ll'>'<I'Il'>'O'<I' t:':ggg:~
<Q ~ •.••.•
o<')~"""o\<Q
Il'>~ '<I''<I'0<')0<')
t-.:oq"1~ 0<')
~
r-i "";"";"";r-ir-i r-i r-i r-i ...<...< ...<

i
0
\C 0' ~ .,.;~ M ~ """"' I"'-i
Il'>•.••.•
~
'" •.••.•
'0 ~ Il'>0 •.••.•
<Q<Q'OO 1'-0.•.••.•
0\ '<I'~ '<I''<I'Il'>Il'> '0
'<I'0<')•.••.•
0 0\
'o•"... ...,
0
oqt-.:"1
~
Oo\<Qll'>'<I'
"
~!C!t~-.:
0<')0<')0<')~ ~ r-ir-ir-ir-i...<
<Q1'-o.'OIl'>
"~"';HH
'<I'
...<
Il'>•.••.•
Il'>
0\
~
o
..•.. •.••.•
Il'>'OO'O 1'-0.'<I'Il'>'<I'1'-0. '<I'o\<Q~""" 0<')'<I'Il'>'0 1'-0.
o 0<')'<1''0 c:Qll'> <Q '<I'•.••.•0\ 1'-0. 1l'>0<')~"""0 o\<Q1'-o.'O Il'>
0
~'"~';~
o'<I'•.••.• "";"";"";r-ir-i r-i r-i r-i r-i r-i ~ H H H ...<
£
.:s
"1

~
'"
1. Q)
Il'>
•...
'<I'o<')O'O~
"! ~ t-.: ~ oq
'<1'•.••.•
'0 0\ c:Qll'> '<I' "";"";"";"";r-i
'<I'H
Il'> '~OIl'>OO\
N •....•
O\Il'>~O<Q '0 '<I'0<')~ 0
r-ir-ir-ir-ir-i
•.••.•
~ '<I'Il'> 1'-0.
00\<Q1'-o.
r-4 ,....; H H
'0
...<
,2: ~

1 '•"...
•...
"1
<Q •.••.•
'<I'•.••.•
H'<I'1'-o.0\'O
o~'"~';~
~
'<I'H
<Q g1'-o. <Q 1'-0.""" o\O<,)~<Q'O
Il'>~ 0 0\
"~";"";"";r-i
'0 Il'>'<I'~
r-ir-ir-ir-ir-i
•.••.•g;g~~
r-ir-i...<...<
Il'>
1'-0.
...<

..8
0<')1'-0.Il'>'<I'~ Il'>o<') '<I'0<')1'-0. Il'>0 0\ Il'>'<I' 1'-o.<QO~ '<I'
"
~!~
C!~ •.••.•1'-0.'<I'~ 0 <Q1'-o.1l'>'<I' 0<') ~ •....••....•
o 0\
E "':;:00 ~
....;....;....;....; r-i r-i r-i r-i r-i r-ir-ir-ir-i ...<
..g o\o\<Q'O'<I'
0<')•.••.•
Q) ~
~
..•.. •.••.•
0<')'<I''0 Il'> <Q1'-o.<Q1'-o.~ ~ '<I'Il'><Q 0
o "!")~ <~Qll'>o<')~ g:2t:':~~ ~ "! "! ~ •....•
'"gj \C '<I'O\<Q'O'<I'
0<')•.••.•
~
....;....;....;....; "";r-ir-ir-ir-i ~ ~ ~ ~ r-i
b'o ~
~ 'OOH'OIl'> 0\1'-o.0\<Qo<') 0<')1l'>1'-o.0\ •....•
~ 0<')0<')0 ~ 0 0<')0\'0'<1'0<') :~2t:::~ 1l'>'<I'0<')~ ~
'"'0
'" 00'0'-0"';
0<')H
~
....;....;....;....; "";r-ir-ir-ir-i r-ir-ir-ir-i r-i
~ ~
...:;:
-:5 1l'>1l'>~0\0\ o<')~'<I'o<')<Q '0 •.••.••.••.•
1'-0.'0 0\ •.••.•
0<')Il'> 1'-0.
1'-0.~ •.••.• 0<')•.••.•Il'>•.••.•
<Q'O '<I' ~ •.••.•
0 <Q 1'-0. 'O'OIl'>'<I' 0<')
....;....; ....; "";"";"";r-ir-i r-ir-ir-ir-i r-i
'
~ "'" o~'o'-O"';
~ H
~ ~
,9
i:: ~
~ 0\'<1''<1'00\
..0 '~O<Qo\""" '0 Il'>1'-0.'0""" '~<I''O<Q 0
~ Il'>'<I' 1'-0.0<')0 <Q 1'-0. '<I'0<')~ •.••.•
<Q •.••.• 0\ o\<Q1'-o.'O '0
] ..., "';0'0'-0"'; ....;
~ ~ ~....; "";"";"";"";r-i r-i r-i r-i r-i r-i
.:a'" ............
~
•..•..

£•... o 0 Il'>'<I'0\
1l'>01l'>0\1'-o.
'<1''<1''0'00
•.••.•
1'-0.'<I'~
0\'<1'0<')0\0\ 0~<')1l'>1'-o.
•.••.•<Q 1'-0.'0 '<I'0<') o<')~ •.••.• o 0
0
"1 0'0'0'-0"'; .,.; ~ ~ ~ ~ ....;....;....;....;....; ....;....;....;....; ....;
..8 0\ •.••.•
•....•
'" Q)

~ U") H tf') H M O\O\~'~O Il'>0 0\ Il'>'<I' 1'-o.<QO~ '<I'

> '<I'll'> H 1'-0.'0 0\1l'>0<,) •.••.•
0\ 1'-0.'0 '<I'0<')~ •.••.•
000\ <Q
•... ...<~ot-:-O .,.; .,.;.,.; .,.; ~ ~ ~
....; ....;
] '0 H •.••.•
•....•
1:
u
to)

...
Q
Q) c
•.••.• 0<')'<I'Il'> '0 1'-0.<QO\ 0
iE-< ~
~ •....• ~ '<I''0 0 Il'>
.................. ~ ~ g~~
•....• ~
Q) ,..., ,...,
> .2> coo ""'0
co II'> ,...,
"I co 11'>'0 '0 0 II'> "I t-..
,..., ,..., 0 0 co 0
~ 0;: "';0',0 ""...;
0 CO '0
,0"';
co
"" 0-...;
~ ~
""0 t-..
...;...; r-i r-i
"" r-i o CO '0
""....< 0
<:::., '0 0-"1 ,..., 0' ~ .~
....; ....<
~
I:: ..!S ""
'0
os
u
q:: t-.."I II'> CO CO ,..., ..... "I
°2 '0
""...; 0',0"';
0-""'0
0
"I
'ON
OCO ""
q 0 ""
CO
""
II'> ""'0-'0
'0
"I 0 CO "" '0'0
t-..
be 0
,..., 0' t...: .,.; II'> ""
~ ~ ...; "';r-ir-i ""r-i r-i r-i .....; .....; ""
....<
00; \() ..... 0-"1
os ""'0
•...
o t-.. ""0 CO ""0-0 II'> II'> 11'>0 CO
0 0-
t-.. "" ,..., t-.. II'>
0 0
""
0- o-II'>CO "10-"1 "I "" ""'0
0-
0 0',0"'; "" '0
~ ~ ...; ...;...; r-i r-i """I 0 CO t-..
0-
o
0
N')
'0 0-"1 ,..., 0' t...:.,.;.,.; r-i r-i r-i ....< ....<
"I
'0 '0

,..., ,...,
£
.:.0 0
"" 0-'0
0
0' 0-"1
0
0
0-
II'>
0',0 "",..., 0'
N II'> 0'0'0
""
t...:,o
,..., "" ~
.,.;
CO
""
~
'0 ""''0
CO II'> "I 0- "" 0t-..
...; ...; ...; r-i r-i
II'>t-..
11'>"""1
0

r-i r-i r-i r-i

""0 CO
CO
....<
os "I 0
...c "I
'0
8
0- ,..., 0-11'>
o-t-.. t-.. 0 "I '0 ""'''1'0'0 ""''0 0 "I II'> 0-
,...,
cu II'> CO CON t-.. II'> ""11'>0-11'> 0'0 OCO t-.. 11'>"" ""
0
£os II'>
•.... t...:o',o~
,..., 0' t...: ,o"';~
~ ~
...; "" "';r-i
...; r-i r-ir-i r-i r-i
11'>0-"1
,...,
"'3 '0

~ ,..., ""II'> t-.. 0- N t-.. t-.. ,..., ,..., II'> CO

COO
""~ 0'CO ~ t-..
'0 0
,..., CO II'> """I 0- 0-
""CO '00
'0 II'> "" ,...,
r-i r-i r-i ""
'0 t-.. "" '0
os "I
•.... ,0 0' t...: t...:,oll'i II'> ~ .~
..;...; ...; r-i r-i r-i
•... 0 0-"1 .....
,...,
..E '0

§ ""II'> 0-0 0- 0
"" ""
t-.. '0 0 0-'0 "I t-.. 0-"1'0 ,...,
"'0 "" 0' t-.. "I ,..., CO 0
~ q II'> "",..., CO II'> "" ,...,0- 00'0 II'>
cu "':;::00 ....<
"" 00
t...: ~ 0..... 00,0,0 II'> II'> ~ ~ ...; ...; ...; ...; r-i r-i r-i r-i
J! 00 0-"1"'"
0-
..•..o Il'>

"I "" ..... N..... t-.. t-.. 0- 0

"" 00
t-..
0- N
"" 0"I CO
t-..""
"I "I
t-.. 0-
•...• '0 0
'"
cu "I "" 0- .,.; '0 .....
"" t...:,o .,.; ""
CO '0
"" ...;...; 0-
00
~
be
\() o'o't...:
11'>0-
0
"I ..... .....
00 .,.; ~ ""~ ~ ...; ...; ...; r-i r-i
cu CO
"'0 II'>
....:: '00"" "I t-.. 11'>'0 ""'0 '0 0- II'> 0 ..... t-.. "I
-g 00"""1 II'> 0- t-.. ""'00'0 "" 0'0 ""
"" .....
0 CO f'..1t')~ ""
t-'4 0
"';0'00 .,.; 0 00 t...:,o,o",; .,.;~ ~ ~ ...; ...;...;...;...; ...;
"I ..... .....
os II'>
'00-
t-..
....:: II'>

-£ ..... 00 0- Il'> II'> ..... "I 0- ,..., t-..""CO "I "I

o
'0 II'>
"I t-.. 0-""
t-.. ••••• 000 0- ""
~ q t-.. "" ••••• 0 ""00'0 II'> CO
""...;
0' 00 ~ .....; ,0
"" .,.; ~ ~ ~ ~ ""
...;...; ...;
~

I::
'<jo ~
"I 0- N ..... .....
o't...:t...: II'> II'>

09 '0
II'>
~
.,g t-.. t-..'oo- CO II'> 0- 0- Il'> II'> '0 0- CO ..... ..... CO
.J:: "" ,..., "" '0
'0
0 t-.. '<I' II'> 0- II'> 0-11'> "I ""0-'0
.,.; .,.; .,.; ~ ~
"" ""..... 0-II'> t-..
II'>
.:a'" N') "';0'0',0
00-"1
r-i
••••• .....
0'00 t...:,o,o ~ ~ ~ ...; ...;
1-1.,
""
II'>
cu
-£ ,..., "I t-.. "I "I "",..., II'> t-.. o-co coo- .....
0-11'>"1 ""
II'> II'>
•...
0
11'>000 8 "'! 0-11'>'0 0
00
'0
II'>
,0,0,0
00 II'>
.,.; .,.;
"'1 .....
.,.; ~ ~
0- t-.. '0
..E "I 0'0'0 00 ..... 00'00 t...: II'> ~
0-0-"" ..... "" .....
0-

..a'"
cu ""
..... 0 '0 t-.. .....
~ '00
COli'> ..... "I "I 11'>'0'0
"" "I "I II'> 0""
t-..
""'0
""0011'> "" 0..... t-..
'0
11'>""0
00 II'>
CO ""'0
00 ~ ....< ,0 0 0'0000 00 ,0 ,0
..... ..... ..... 0..... .....
"';r-i
•.... ....< ....< t...: t...: t...: t...:
].•. II'>
0
0-"" "I .....
'c ""
U
""Q
"'" .c
*-~,.Q,...,
«I •.•.•
..... "I "" "" II'> '0 t-.. CO 0- 0
..... "I .....
..... '<1''0 0
..... "I II'>
"I g~ °2
q::
I-< 0 ~ ..... ..5
References

Ashby, E. and Oxley, T. A. (1935) 'The interadion of fadors in the growth of Lemna VI: an
analysis of the influence of light intensity and temperature on the assimilation rate and the
rate of frond multiplication', Annals of Botany, 49, 309-36.
Ayala, F. J. and Kiger, J. A Jr. (1980) Modern Genetics, Benjamin/Cummings, Menlo Park,
California.
Banic, S. (1975) Nature, 258, 153-4. Discussed by Wardlaw, A. C. (1985) Practical Statistics
for Experimental Biologists, John Wiley, New York, p. 103.
Bateson, W. (1909) Mendel's Principles of Heredity, Cambridge University Press, Cambridge.
Bliss, C. I. (1967) Statistics in Biology, McGraw-Hill, New York.
Boslough, J. (1984) Beyond the Black Hole: Stephen Hawking's Universe, Fontana, London, p. 38.
Brightwell, R., Dransfield, R. D. and Kyorku, C. (1991) 'Development of a low cost tsetse trap
and odour baits for Glossina pallidipes and G. longipennis', Medical and Veterinary Entomology,
5 153-64.
Broad, W. and Wade, N. (1985) Betrayers of the Truth: Fraud and Deceit in Science, Oxford Univer-
sity Press, Oxford, p. 31.
Brown, J. H. and Lomolino, M. V. (1989) 'Independent discovery of the equilibrium theory of
island biogeography', Ecology, 70, 1954-7.
Brown, P. (1991) 'Canada to test Kenya's disputed HIV drug', New Scientist, 12, January, p. 27.
Brownbridge, M. (1988) 'Field trials of Bacillus thurungiensis against the African armyworm
Spodoptera exempta June 8th-June 12th 1988', Internal Report, International Centre for Insed
Physiology and Ecology, P.O. Box 30772, Nairobi, Kenya.
Bulmer, M. G. (1979) Principles of Statistics, Dover, New York.
Burkhardt, R. W. (1981) 'Species', in Bynum, F., Browne, E. J. and Porter, R. (eds) (1981) Dictionary
of the History of Science, Macmillan Press, London, p. 396. See also, Ronan, C. A (1983) The
Cambridge Illustrated History of the World's Science, Cambridge University Press, Cambridge,
p.398.
Cain, A J. and Sheppard, P. M. (1952) 'The effeds of natural seledion on body colour in the
land snail Cepaea nemoralis', Heredity, 6, 217.
Cain, A J. and Sheppard, P. M. (1954) 'Natural seledion in Cepaea nemoralis', Genetics, 39,
89-116.
Challier, A, Eyraud, M., Lafaye, A and Laveissiere, C. (1977) 'Amelioration du rendement
du piege biconique pour glossines (Diptera, Glossinidae) par l'emploi d'un cone inferieur blue',
Cahiers ORSTOM, serie Entomologie medicale et Parasitologie, 15, 283-6.
Cronin, H. (1991) The Ant and the Peacock: Altruism and Sexual Selection from Darwin to Today,
Cambridge University Press, Cambridge.
Crooke, W. M. and Knight, A H. (1965) 'The relationship between nickel-toxicity symptoms
and the absorption of iron and nickel', Annals of Applied Biology, 43, 454-64. Discussed by
Bliss, C. I. (1967) in Statistics in Biology, Volume 1, McGraw-Hill, New York, p.37.
Cushny, A R. and Peebles, A R. (1905) 'The adion of optical isomers II. Hyoscines', Journal
of Physiology, 32, 501-10.
Darwin, C. (1906) Origin of Species, John Murray, London.
Darwin, C. (1958) Autobiography, edited by N. Barlow, Collins, London, p. 120, cited by A
Flew in the introdudion to An Essay on the Principle of Population, Penguin, Harmondsworth,
p.49.
Davis, R. Dunford, C. and Lomolino, M. V. (1988) 'Montane mammals of the American
188 Refe~
nc~
southwest: The possible influence of post-Pleistocene colonization', Journal of Biogeography,
15, 841-8.
Diamond, J. M. and May, R. M. (1981) 'Island biogeography and the design of natural reserves',
in May, R. M. (Ed.) (1981) Theoretical Ecology, Blackwell Scientific, Oxford, p. 228-52.
Dobson, A. J. (1983) An Introduction to Statistical Modelling, Chapman & Hall, London.
Dobzhansky, T. and Pavlovsky, O. (1957) 'An experimental study of the interaction between
genetic drift and natural selection', Evolution, 11,311-19.
Dransfield, R. D., Brightwell, R., Kyorku, C. and Williams, B. G. (1990) 'Control of tsetse
populations using traps at Nguruman, southwest Kenya', Bulletin of Entomological Research, 80,
265-76.
Dransfield, R. D., Brightwell, R. and Williams, B. G. (1991) 'Control of tsetse flies and trypano-
somiasis-Myth or Reality7', Parasitology Today, 7, 287-91.
Einstein, A. (1950) 'The fundaments of theoretical physics', Out of My Later Years, Philosophical
Library, New York, p. 98.
Einstein, A. (1954) 'Geometry and experience', in Ideas and Opinions, Crown, New York, p. 233.
Einstein, A. (1978) The Meaning of Relativity, Chapman & Hall, London, p. 1.
Farrow, R. A. (1979) 'Population dynamics of the Australian plague locust Chortoicetes terminifera
in central western New South Wales. 1. Reproduction and migration in relation to weather',
Australian Journal of Zoology, 27, 717-45.
Feynman, R. (1980) The Character of a Physical Law, MIT Press, Cambridge, Massachusetts, p. 34.
Fisher, R. A. (1936) 'Has Mendel's work been rediscovered7', reprinted from Annals of Science,
1,115-37 and in Bennet, J. H. (Ed.) (1965) Experiments in Plant Hybridization, Oliver and Boyd,
Edinburgh.
Fisher, R. A. (1965) 'Marginal comments on Mendel's paper', in Bennet, J. H. (Ed.) (1965) G.
Mendel: Experiments in Plant Hybridization, Oliver and Boyd, Edinburgh.
Freund, J. E. (1972) Mathematical Statistics, Prentice-Hall, London.
Galton, F. (1889) Natural Inheritance, Macmillan, London. Cited by Snedecor, G. W. and Cochran,
W. G. (1987) in Statistical Methods, Iowa State University Press, Ames, p. 171.
Gardiner, M. (1970) 'The fantastic combinations of John Conway's new solitaire game of life',
Scientific American, 223, 120-3.
Gardiner, M. (1971) 'On cellular automata, self-reproduction, the Garden of Eden and the game
of Life', Scientific American, 224 112-17.
Gardiner, M. (1983) Wheels, LIfe and other Milthematical Amusements, Freeman, New York.
Giessler, A. (1889) 'Beitriige zur Frage des Geslechtsverhiiltnisses der Geboren', Z. K.
Siichsischen Statistischen Bureaus, 35, 1. Discussed by M. G. Bulmer (1979) in Principles of
Statistics, Dover, New York, p. 88.
Gould, S. J. (1973) 'Positive allometry of antlers in the Irish Elk Megaloceros giganteus', Nature,
244,375-6. Discussed by Ruse, M. (1982) in Darwinism Defended, Addison-Wesley, New
York, p. 149.
Gould, S. J. (1974) 'The origin and function of "bizarre" structures: antlers and skull size in the
"Irish elk" Megaloceros giganteus', Evolution 28, 191-220.
Gould, S. J. (1977) 'The misnamed, mistreated and misunderstood Irish Elk', in Ever Since Darwin,
Norton, New York, p. 79, 85.
Gould, S. J. (1980) The Panda's Thumb, Norton, New York, p. 225.
Gould, S. J. (1981) The Mismeasure of Man, Pelican Books, London.
Gould, S. J. (1986) 'Evolution as fact and theory', Hen's Teeth and Horse's Toes, Penguin Books,
Harmondsworth, England, p. 255.
Green, P. E. (1978) Analysing Multivariate Data, Dryden Press, Hinsdale, Illinois.
Greenwood, M. and Yule, G. U. (1920) 'An inquiry into the nature of frequency distributions
of multiple happenings', Journal of the Royal Statistical Society, 83, 255-79. Discussed by
Sokal. R. R. and Rohlf, F. J. (1981) Biometry, Freeman, New York, p. 94.
Hacking, 1. (197.9) Logic of Statistical Inference, Cambridge University Press, Cambridge.
 References 189
Hays, W. L. (1988) Statistics, Holt, Rinehart and Winston, Orlando.
Healy, M. J. R. (1981) 'Choice of sample size in parasitological experiments', Parasitology Today,
3, 91-4.
Hunter, D. M. and Cosenzo, E. L. (1990) 'The origin of plagues and recent outbreaks of the
South American locust, Schistocerca cancellata (Odhoptera: Acrididae) in Argentina', Bulletin of
Entomological Research, 80, 295-300.
Jammer, M. (1974) The Philosophical FClUndationsof Quantum Mechanics, John Wiley, New York,
p. 155. Letter from Albed Einstein to Max Born, 4 December 1926.
Jeffers, J. N. R. (1978) An Introduction to Systems Analysis: With Ecological Applications, Edward
Arnold, London.
Kendall, M. and Stuart, A. (1983) The Advanced Theory of Statistics, Volume II, Charles Griffin,
London, pp. 110 and 134.
Kendall, M., Smart, A. and Ord, J. K. (1983) The Advanced Theory of Statistics, Volume 3,
Charles Griffin, London, p. 48.
Lambrecht, H. (1961) 'Die Genenkade von Pisum bei normaler Struktur der Chromosomen',
Agric. Hortique Genetica, 19,360-401, communicated to Bennet, J. H. (Ed.) (1965) G. Mendel
Experiments in Plant Hybridization, Oliver and Boyd, Edinburgh, p. 6.
Leyton, L. (1956) The relation between the growth and mineral composition of the foliage
of Japanese larch (Larix leptolepis, Murr.)', Plant and Soil, 7, 167-77. Discussed by Bliss, C. 1.
in Statistics in Biology, Volume I, McGraw-Hill, New York, p. 308.
Lorenz, K. (1967) On Aggression, Methuen, London, p. 83.
MacArthur, R. H. and Wilson, E. O. (1963) 'An equilibrium theory of insular zoogeography',
Evolution, 17,373-87. See also MacArthur, R. H. and Wilson, E. O. (1967) The Theory of
Island Biogeography, Princeton University Press, Princeton.
MacFadden, B. J. (1986) 'Fossil horses from Eohippus (Hyracotherium) to Equus: scaling Cope's
Law and the evolution of body size', Paleobiology, 12,355-69.
Malthus, T. R. (1970) An Essay on the Principle of Population, Penguin Books, Harmondswodh,
England. Published in 1798 under the title An Essay on the Principle of Population as It Affects
the Future Improvement of Society ... This is the essay referred to here. MaUhus subsequently
rewrote and considerably extended the essay, which he published under the title An Essay on
the Principle of Population; or, a View of its Past and Present Effects on Human Happiness; with an
Inquiry into our Prospects Respecting the Future Removal or Mitigation of the Evils which it
Occasions.
Manly, B. (1986) Multivariate Statistical Methods: A Primer, Chapman & Hall, London.
Margenau, H. (1950) The Nature of Physical Reality, McGraw-Hill, New York, p. 99. Attributed
to Max Born.
May, R. M. (1976) 'Simple mathematical models with very complicated dynamics', Nature,
261, 459.
May, R. M. (1981) 'Models for single populations' in Theoretical Ecology, Blackwells, Oxford,
pp. 5-29.
Mayr, E. (1942) Systematics in the Origin of Species, Columbia University Press, New York.
Discussed by Ruse, M. (1982) in Darwinism Defended, Addison Wesley, London, p. 94.
McCullogh, P. and Neider, J. A. (1983) Generalized Linear Modelling, Chapman & Halt London.
Mead, R. and Curnow, R. N. (1983) Statistical Methods in Agriculture and Experimental Biology,
Chapman & Halt London.
Mendel, G. (1866) 'Versuche iiber pflanzenhybriden', Proceedings of the Natural History Society
of Brunn. English translation reprinted as Experiments in plant Hybridization with commentary
by R. A. Fisher and biographical notes by W. Bateson in Bennet. J. H. (ed.) (1965), Oliver
and Boyd, Edinburgh. See also Peters, J. A. (ed.) (1959) Classic Papers in Genetics, Prentice Hall,
Englewood Cliffs, New Jersey, and Stem, C. and Sherwood, E. R. (eds) (1966) The Origin of
Genetics, Freeman, San Francisco. Mendel's work is discussed by Ayala, F. J. and Kiger, J. A. Jr.
(1980) in Modern Genetics, Benjamin/Cummings, Menlo Park, California.
190 Refennas
Needham, J. (1972) 'Mathematics and science in China and the West', in Sociology of Science
(ed. B. Barnes), Penguin, Harmondsworth, p. 32 citing Barbera, G. (1890) Le Opera di Galileo
Galilei, Volume 4, Florence.
Needham, J. (1988) 'Joseph Needham', Channel 4 Television, London, 13 August 1988.
Newsweek (1989) 14 August, 1989, p. 50.
Pacala, J. (1990) 'International doubts about an AIDS cure', Science, 250, 200.
Pearson, K. and Lee, A. (1903) Biometrika, 2, 357. Discussed by Snedecor, G. W. and Cochran W. G.
(1987) in Statistical Methods, Iowa State University Press, Ames.
Pianka, E. R. (1983) Evolutionary Ecology, Harper and Row, London, p.328.
Pirsig, R. M. (1980) Zen and the Art of Motorcycle Maintenance, William Morrow, New York,
p.238.
Powick, W. C. (1925) 'Inactivation of vitamin A by rancid fat', Journal of Agricultural Research,
31, 1017-27. Sokal, R. R. and Rohlf, F. J. (1981) use the amount of lard consumed by the
rats in Powick's experiment to illustrate the analysis of a two-factor experiment in Biometry,
Freeman, New York. p. 325.
Rensberger, J. (1983) 'Trial by error', RF Illustrated (Rockefeller Foundation), June, p. 2-4.
Ronan, C. A. (1983) The Cambridge Illustrated Dictionary of the World's Science, Cambridge
University Press, Cambridge, p. 429.
Ross, W. D. (Ed.) (1952) The Works of Aristole, Clarendon Press, Oxford. Translation of Aristole's
Physica, W. Ogle, quoted by S. Sambursky (1974) in Physical Thought from the Pre-Socratics to
the Quantum Physicists, Hutchinson, London, p. 65.
Roughgarden, J. (1979) Theory of Population Genetics and Evolutionary Ecology, Macmillan, New
York. p.21.
Ruse, M. (1982) Darwinism Defended, Addison-Wesley, New York, p. 150. Cites results by
Fitch, W. M., 'Molecular evolutionary clocks', in Molecular Evolution (ed. F. J. Ayala).
Satterthwaite, F. E. (1946) Biometrics Bulletin, 2, 110. Discussed by Snedecor, G. Wand
Cochran, W. G. (1989) Statistical Methods, Iowa State University Press, Ames, p. 97.
Sheppard, P. M. (1951) 'Fluctuations in the selective value of certain phenotypes in the
polymorphic land snail Cepaea nemoralis', Heredity, 5, 125.
Siegel, S. and Castellan, N. J. (1988) Nonparametric Statistics for the Behavioral Sciences,
McGraw-Hill, New York.
Snedecor, G. W. and Cochran, W. G. (1989) Statistical Methods, Iowa State University Press,
Ames.
Sokal, R. R and Hunter, P. E. (1955) 'A morphometric analysis of DDT resistant and non-resistant
housefly strains', Annals of the Entomological Society of America, 48, 499-507. Given by
R. R. Sokal and F. J. Rohlf, Introduction to Biostatistics, Freeman, New York, p. 81. In order
to simulate the results of measuring the wing lengths to one more significant figure, evenly
distributed random numbers between 0.00 and 0.10 were added to each of the numbers
given by Sokal and Hunter.
Sokal, R. R. and Rohlf, F. J. (1981) Biometry, Freeman, New York.
Sokal, R. R. and Rohlf, F. J. (1987) Introduction to Biostatistics, Freeman, New York.
Sprent, P. (1990) Applied Nonparametric Statistical Methods, Chapman &: Hall, London.
Stewart, I. (1990) Does God Play Dice?The New Mathematics of Chaos, Penguin Books, London.
Student (1907) 'On the error of counting with a haemocytometer', Biometrika, 5, 351-60.
Analysed in detail by Sokal, R. R. and Rohlf, F.J. (1981) Biometry, Freeman, New York, p. 84.
Student (1908) 'On the probable error of a mean', Biometrika, 6, 1-25.
TargeH, G. A. T. (Ed.) (1991) Malaria: Waiting for the Vaccine, John Wiley, Chichester.
Taylor, E. F. and Wheeler, J. A. (1963) Space-time Physics, Freeman, San Franciscp, p. 60. Quoted
by Harrison, D. (1979) in 'What You Get Is What You See', American Journal of Physics,
47, 576.
Tufte, E. R. (1986) The Visual Display of Quantitative Information, Graphics Press, Cheshire,
Connecticut, p. 60. Taken from Science Indicators 1974, National Science Foundation,
Washington, D. c., 1976, p. 15.
 References 191
Utida, S. (1943) 'Studies on experimental populations of the Azuki bean weevil, Callosobruchus
chinensis (L.). VIII. Statistical analysis of the frequency distribution of the emerging weevils on
beans', Mem. CoIl. Kyoto Imperial University, 54, 1-22. Discussed by Sokat R. R. and
Rohlf, F. J. (1981) Biometry, Freeman, New York, p. 92.
von Bortkewitsch, L. (1898) Das Gesetz der kleinen Zahlen, Teubner, Leipzig. Discussed by
Bulmer, M. G. (1979) in Principles of Statistics, Dover, New York, p. 92 and by Sokal, R. R.
and Rohlf, F. J. (1981) in Biometry, Freeman, New York, p. 93.
Wallace, A. R. (1905) My Life, Chapman and Halt London, Volume I, p. 232. Quoted by A.
Flew (1970) in the introduction to An Essay on the Principle of Population, Penguin,
Hannondsworth, p. 51.
Whitehead, A. N. (1928) Introduction to Mathematics, Thornton Butterworth, London, p. 9.
Winer, B. J. (1971) Statistical Principles of Experimental Design, McGraw-Hilt New York.
 Index

Page numbers appearing in bold refer to tables

* 97 computational formulae 124

• 21 factorial design 135-8
~ 60 mixed model 134
<...> 40 Modell 133-4
IL 40 ModellI 133-4
(T 44 and multiple regression 174
128 nested factors 130, 147-50
2
X distribution 53, 67--8 no treatment effect 127
Central Limit Theorem 67 one-way 124-31
definition 67,74 and polynomial regression 170
degrees of freedom 67 and regression 159
illustration 68 soporific drugs 130-1
parameters 68 nested factors 130
table 184 and t test 130, 137
X2 test 82-90, 94 testing hypotheses 128-9
degrees of freedom 84, 91 ticks on rabbits 131-3
derivation 91 treatment effect 127--8
expression for 83 tsetse flies 144-6
families of eight 83-4 turnips 147-50
Mendel's honesty 86 two-way 135-42
Mendel's peas 84 as one-way 135
weight gained by rats 135-9
Acceptance range 79,82,86,117 Antler lengths 179
one- and two-tailed tests 94-6 Aristotle 2
Accident records, data 75 Armyworms 7--8, 11, 154, 175, 181
Added variance component, see ANDV A 181-2
Variance data 8,154,181
AIDS 93, 114 Latin square 181
Alternative hypothesis model 154
one- and two-tailed tests 95 rate of decline 159, 162
sample size determination 115 regression 158-62
Analysis of covariance, seeANCDV A Ashby, E. 163
Analysis of variance, seeANDV A Ayala, F. J. 4
ANCDV A 175, 181 AZT 93
interactions 181, 182 Azuki bean weevil 75
Angina 114
ANDV A 9,43, 123-53 Bacillus thuringiensis 7,154, 181
armyworms 181 Banic, S. 92
194 Index

Bar chart 26, 30 Cervine deer 179

Basic distributions 9 data 179
Bateson, W. 22, 92 Challier, A. 143
experiments on peas 22 Chaos 25
data 22 Chinese science 77
Beans 75 Chromosomes 22, 23, 87
data 76 Cluster analysis 175
Best fit in regression 157 Cochran, W. G. 6,90, 126, 128, 147,
Bias 160
Fisher's K statistics 50 Comparisons 9,113
standard deviation 44 multiple 138
Biconical trap 143-6 pairwise 113
Binomial distribution 53-60 Conditional probability, see
conditions 54 Probability
definition 74 Confidence limits 107-9, 118, 131, 146
expression for 53 expression for 108
illustration 55 interpretation of 109
kurtosis 58 polio vaccine 108, 109
mean 57 regression 162
Mendel's peas 59-60 soporific drugs 108
non-parametric tests 102 and t statistic 108
parameters 58 Contingency tables 87-90
skewness 58 corrections for small numbers 90
variance 57 degrees of freedom 88,89
Birth registers, see Families of eight Fisher's exact test 90
Bliss, C. 1. 50, 169, 171 multi-way 88,89
Born, M. 176 Yate's correction 90,93
Boslough, J. 25 Continuous random variables, see
Brightwell, R. 143 Random variables
Broad, W. 87 Contrasts, see Comparisons
Brothers 167-8 Conway, J. 25
Brown, J. H. 179 Correlation 166-9
Brown, P. 114 multiple regression 171, 173
Brownbridge, M. 7,159,175,181 Correlation coefficient
Bulmer, M. 13,44,53,57,60,63,90, brothers and sisters 167
109 data 167
Burkhardt, R. W. 1 expression for 167
and R2 168
Cain, A. J. 93 and slope ofregression 168
Calcium in turnips 147-50 Cosenzo, E. L. 11
data 148 Covariance, expression for 166
Calculus 30 Creation 2
Callosobruchus chinensis, see Azuki bean Crick, F. 4
weevil Cronin, H. 3
Casino game 5,20 Crooke, W. M. 169
Castellan, N. J. 90, 103 Crossed factors 114-15, 118, 151
c.d.f., see Cumulative distribution Latin square 144
function soporific drugs 115
Cell division 4 Cumulative distribution function 28-9,
Central Limit Theorem 63,67 34-6,71-3
Cepea nemoralis, see Snails calculation of 35
 Index 195
families of eight 71-2 Type II 80, 81
and probability density function 73 multiple range test 104
Prussian army officers 72 sample size determination 116
Curnow, R. N. 114 Evolution, laws of 2
Cushny, A. R. 98 Evolutionary change 181
Expected frequencies 89
Darwin, C. 1,2,3,4 Experimental design 113-15
and Newton's theory 77 crossed factors 114-15
Davis, R. 178 nested factors 114-15
Degrees of freedom 44, 67 polio vaccine 113
2
X test 84,88,89 see also Design of experiments
ANDVA 125-6 Exponential growth 3, 4, 156,163
factorial 139 duckweed 164
nested 149 Extrapolation 157
Latin square 145,152
multiple range tests 105 F distribution 53, 71
paired t test 100 definition 71,74
regression 155 illustration 70
Satterthwaite's test 120 parameters 71
unequal sample sizes 132 tables 185-6
unequal variance 120 variance ratios 107
unpaired t test 99 Families of eight 55-9,107
and weight gained by rats 111 binomial fit 59
Density dependence 163 data 56
Dependent variable 168 kurtosis 58
Design of experiments 7,150-1 normal approximation 66
see also Experimental design proportion of boys 56
Diamond, J. M. 179 binomial fit 56
Discrete random variables, see standard deviation 57
Random variables skewness 58
Discriminant analysis 176 Farrow, R. A. 11
Distribution functions, using tables 73 Fathers 156
Dobson, A. J. 175 Feynman, R. 154
Dobzhansky, T. 107 Fisher, R. A. 4, 86, 104
Dominance 16, 21 Fisher's exact test 90
Dransfield, R. D. 8, 143 Fisher's K statistics 49
Duckweed 10, 163-6 expressions for 50-1
data 164 Fitch, W. M. 177
exponential growth 164--6 Fixed factors 134, 151
threshold 164 Fossil record 177
data 178
Einstein, A. 25, 39,77,78 Fossils, ages of 179
English law 80 Foundations of statistics 12
Epsilon trap 144--6 Founder effect 107
Equidae, see Horses Franklin, R. 5
Errors Frequency distributions 26-7, 30-2
and English law 80 Freund, J. E. 23, 158
theory of 63 Fruit flies 107
Type I 80, 81 data 107
multiple range test 104
sample size determination 116 Galileo5, 77
 196 Index

Galton, F. 52, 156 Hypotheses 78

Galton's Law, see Law of universal alternative 80
regression null 80, 82, 84
Gardiner, M. 25
Gauss, K. 63 Idealism 78
Generalized linear modelling Independent comparisons, see
175 Contrasts
Geometric mean, see Mean Independent variable 168
Giessler 55 Index of increase 146
Gossett, W. 69 Interactions 110,112,113,139-40
see also Student ANCOV A 181, 182
Gould, S. J. 80, 87, 179 regression 175
Greek science 77 weight gained by rats 112, 137,
Green, P. E. 176 139-40
Greenwood, M. 75 illustration 140
Guinea pigs 92 Interferon 114
Interpolation 157
Hacking, I. 12, 123 Inter-quartile range 45
Haemaphysalis leporispalustris, see Irish elk 179
Ticks data 179
Hardy-Weinberg Law 24 Iron 169-71
Hawking, S. J. 25 Isolation 178
Hays, W. L. 90, 105, 106, 175 Iteration 165
Healy, M. J. R. 117
Heterozygosity 107 Jammer, M. 25
Hindu science 77 Japanese larch 171-4
Histogram 30 correlation 173
category width 30 data 172
Homozygosity 107 Jeffers, J. N. R. 74, 176
Honestly significant difference test
105,106, 145, 146 Kendall, M. 104,109
Horses, body size of 179 Kepler, J. 77
data 180 Kiger, J. A. 4
see also Prussian army officers Knight A. H. 169
Housefly wing-lengths 30, 32-4, 38 Kurtosis 48-9
cumulative distribution 35 definition 48
data 30 standard value 48
mean 40
median 42 Lambrecht, H. 87
multiple range tests 103, 104 Lard 110-12
probability density 32 Latin rectangle 145, 153
range 45 Latin square 142-6
standard deviation 44 armyworm 181
HSD test, see Honestly significant data 182
difference test degrees of freedom 145, 152
Hsii Heng 25 missing values 181
Hunter, D. M. 11,30 model 144
Hyoscine 98-103 tsetse flies 144-6
see also Soporific drugs Law of addition, see Probability
Hyoscyamine 98-103 Law of multiplication, see Probability
see also Soporific drugs Law of universal regression 156, 157
 Index 197
Least significant difference test data 6, 17, 21
104 linkage 20-2
Lee, A. 156, 167 Michigan asylum for the insane 98
Leibniz, G. 30, 78 Mill, J. S. 13,25
Leptokurtic 48 Missing values 181
Ley ton, L. 171, 174 Mixed model, see ANOV A
Life, game of 25 Mode 42
Linkage 21, 23, 87-9 Model I, see ANOV A
Linnaeus 1, 39 Model II, seeANOV A
Locusts 10 Models
data 10 analytical 27
Logarithmic transform 144 armyworm 154
Lomolino, M. V. 179 exponential growth 156
Lorenz, K. 12 Latin square 144
LSD test, see Least significant nested design 147
difference test polynomial regression 169
regression 154
simulation 27
MacArthur, R. H. 9 soporific drugs 134
MacFadden, B. J. 179 specifying 134
Main effect 113 weight gained by rats 135, 137~
Malaria vaccine 6 Montane forests 178
Malthus, T. 2, 3 Multiple comparisons, see Multiple
Manly, B. 176 range tests
MANOV A, see Multivariate analysis of Multiple range tests 103-6, 118
variance comparison of tests 104--6
Materialism 78 degrees of freedom 104, 105
May, R. M. 25, 179 modified t test, derivation of 120
Mayr, E. 107 honestly significant difference test
McCullogh, P. 175 105,106
Mead, R. 114 Prussian army officers 106
Mean 40-1 significance levels 104
comparisons 9~ 101 Student-Newman-Keuls test 106
geometric 41 Studentized range statistic 105, 106
of a sum 41, 49-50 Type I error 104
Mean squares Type II error 104
between treatments 124--6 Multiple regression, see Regression
within treatments 124--6 Multivariate analysis of variance 175
Means plots 129-30 Mutually exclusive events, see
ticks on rabbits 132 Probability
Measures 9
of location 39-42 National Science Foundation 37
of scale 42-5 Natural selection 2-4
of shape 45-9 Needham, J. 5, 25, 77
Median 41-2 Negative binomial distribution 74
Mendel, G. 4, 10, 11, 18,22,87 Neider, J. A. 175
Mendel's honesty 8~ 7 Nested factors 114-15,118,147-51
Mendel's peas 5-6,16,17,68,94 added variance 149
X2 test 84-9 degrees of freedom 149
binomial distribution 59-60 model 147
contingency tables 87-9 null hypothesis 149
 198 Index

Nested factors cont'd Pearson, K. 98, 156, 167

polio vaccine 114 Phaseolus radiatus, see Beans
soporific drugs 115 Phosphorus 171-4
sum of squares 149 Pianka, E. R. 178
turnips 147-50 Pirsig, R. M. 94
Newton, I. 30, 77 Platykurtic 48
Nguruman trap 143-6 Poisson, S.-D. 60
Nickel 169-71 Poisson distribution 53, 60-2
Nitrogen 171-4 accident records 75
Nobel Prizes 37 Azuki bean weevil 75
data 38 definition 74
Non-parametric tests 49,118 expression for 60
binomial probabilities 102 illustration 61
comparison with parametric tests kurtosis 61
102-3 parameters 62
sign test 102 Prussian army officers62, 106
soporific drugs 101-3 sampling 63
Normal distribution 39,53,63-7 skewness 61
approximations 64-7 variance 61
binomial 65-7 yeast cells 76
Poisson 65-7 Polio vaccine 6-7, 52, 62-3, 68-9, 92, 96
definition 74 confidence limits 108,109
expression for 63 data 6
illustration 64 Poisson distribution 62
kurtosis 64 sample size 63, 116-17, 122
mathematical form 32 Polynomial, see Regression
standard 64 Pooled standard deviation, see
illustration 72 Standard deviation
selected points 73 Potassium 164-71
Normalization 64 Power of a test 78-82, 117
ns 97 one- and two-tailed tests 96
Nucleotide substitutions 177 Power-law 177-8
Null hypothesis Powick, W. C. 110, 113
nested design 149 Predation hypothesis 93
one- and two-tailed tests 95 Principle components analysis 176
Probability 12-24
Oat plants 169-71 a priori definition 12
data 169 casino definition 13-14
One-tailed test 94 concept of 12
armyworms 161 conditional 14, 17
polio vaccine 96 distributions 26-8
soporific drugs 98 classification39
Orthogonality 171 normal, see Normal distribution
polynomial regression 170 triangular 39
Outliers 42 uniform 39
Oxley, T. A. 163 law of addition 14, 53
law of multiplication 18, 19,53
Pacala, J. 114 laws of14
Pangenesis 4 long-run definition 13-14
Pavlovsky, O. 107 mutually exclusiveevents 15,16
p.d.f., see Probability density function scale 14
 Index 199

Probability density function forward step-wise 171-4

32-4, 36 F-to-enter 173, 174
Prussian army officers46-9,60-2,92, F-to-remove 173,174
106 polynomial 169
cumulative distribution function 72 ANOVA 170
data 46 model 169
kurtosis 49,62 orthogonality 170
mean 48 residuals 163--6
multiple range test 106 sum of squares 159
skewness 62 variance of the fitted line 177
standard deviation 48 variance of the slope 176
variance 48 weighted 175
Pyrethrum 7,181 Rejectionrange 79,82,86,117
one- and two-tailed tests 94-6
Quantum mechanics 25 Rensberger, J. 113
Quartiles 45 Representations 9
Residual ash 171-4
R2 160 Residuals
Rabbits 121, 131-3 in regression 158, 163--6
Rabies 92 sum of squares 157
Random factors 134, 151 Rohlf, F. J. 76, 90, 104,121, 131, 132,
Random variables 133
continuous 29 Ronan,C.A.4
discrete 26-9 Ross, W. D. 2
Range 45 Roughgarden, J. 24
Rats, see Weight gained by rats Ruse, M. 177
Ray, J. 1
Regression 9 Salk, J. 6
and ANOVA 159 Sample size determination 115-17
armyworms 158-60 alternative hypothesis 115
assumptions 158 polio vaccine 116-17
brothers and sisters 168 soporificdrugs 115
confidence band 162 Type I errors 116
degrees of freedom 155, 159 Type II errors 116
error e~ timates Satterthwaite's rule 99,120,121, 131
illustration of 162 Schwann, T. A. H. 4
mean 161 Sheppard, P. M. 93
parameters 160-2 Siegel, S. 90, 103
points 163 Sign test, see Non-parametric
residuals 160 tests
slope 161 Significancelevels 78-82
etymology 156 notation 97
expressions for coefficients 158 Sisters 167-8
F ratio 160 Skewness 46-8
fathers and sons 156 definition 47
interactions 175 Snails 93
models 154, 171 data 93
multiple 171-4 Snedecor, G. W. 6, 90, 126, 128,147,
ANOVA 174 160
backward step-wise 173-4 SNKtest, see Student-Newman-Keuls
correlation 171, 173 test
 200 Index

Sokal, R. R. 30, 76, 90, 104, 121, degrees of freedom 125-6

131, 132, 133 nested factors 149
Sons 156 partitioning 125-6, 128, 142
Soporific drugs 98-103,122, 140-2, 151 of residuals 157
ANaVA 130-1,141 total 124-6
confidence limits 108 within treatments 124-6
crossed factors 141
data 99 t distribution 53,68-70
fixed factors 140-1 definition 74,69
mixed model 134 illustration 70
model 141 parameters 69
Model I 141 tables 183
ModelII 141 t test 97-8
random factors 140-1 and ANaV A 130, 137
sample size determination 122 comparison of paired and unpaired
Speciation 4, 107 100-1
Species derivation 119
definition 1 expression for 97
origin of 3 least significant difference test 104
Species-area relationship 177-8 modified 104,105
data 179 paired 99-101, lIS, 118
Species richness 178 rationale for 97-8
Spodoptera exempta, see Armyworm unequal variance 119
Sprent, P. 103 unpaired 96-9, lIS, 118
Standard deviation 43-4 weight gained by rats 111
bias 44 Targett, G. A. T. 6
definition 44 Taylor, E. F. 10
degrees of freedom 44 Testing hypotheses 9
internal 130 ANaV A 128-9
of mean 44-5 Three, the importance of 109
in regression 161 Thrushes 93
pooled 99, 130, 142, 145 Ticks on rabbits 121, 131-3
population 45 added variance 133
of residuals in regression 160 data 131
of slope in regression 161 fixed factor 134
Standard error 45 means plot 132
see a/so Standard deviation means table 132
Standard normal distribution, see random factor 134
Normal distribution Triangular distribution 39
Stasis 4 Trypanosomiasis 142
Statistical independence 18, 19,22 Tsetse flies 3,8,13,67
Stewart, I. 25, 109 ANaV A 142-6, 153
Student's t test, see t test data 144, 152
Student 69,76,97,98 Latin square 142-6
Student-Newman-Keuls test 106 traps 143, 144
Studentized range statistics, see Tufte, E. R. 37
Multiple range tests Tukey's HSD test, see Honestly
Sum of squares significant difference test
analysing 124-6 Turnips 147-50
between treatments 124-6 Two-tailed tests 94
computational formula 125 armyworms 162
 Index 201

Two-tailed tests cont'd Vitamin A 110

polio vaccine 96 Vitamin C 92
soporific drugs 99 data 93
Type I errors, see Errors von Bortkewitsch 60, 46
Type II errors, see Errors
Wade, N. 87
Unequal sample sizes 131-3 Wallace, A. R. 2, 3
degrees of freedom 132 Watson, J. D. 4
Uniform distribution 39 Weight gained by rats 110-13,123
Utida, S. 75 ANOV A 135-9
data 110
Vaccination, see Polio interactions 137, 139-40
Variance 42-3 model 135
added component 128 Wheeler, J. A. 10
ticks on rabbits 133 Whitehead, A. 10
turnips 149-50 Wilson, E. O. 179
computational formula 43 Winer, B. J. 42,113,117
definition 43 World Health Organization 114
minimum 157
proportion explained 160,169 Yate's correction, see Contingency
ratios 107 tables
distribution of 121 Yeast cells 76
of a sum 43, 50, 167 data 76
Virchow, R. C. 4 Yule, G. U. 75
 Biostatistics
Concepts and applications for biologists
Brian Williams
•
This book is a first course in statistics for all students of biology. Most
of the examples have an ecological bias but illustrate principles which
have direct relevance for biologists involved in laboratory work. The
carefully structured approach begins with basic concepts and then
progresses towards a clear appreciation of the need for and use of
analysis of variance and regression, including the use of computer
statistical packages. The work is clearly explained with thoroughly
worked examples of real-life biological problems and is highly suitable
for undergraduate students engaged in Quantitative biological work.
Biostatistics will give students a sound grasp of the key principles of
biological statistics without overwhelming detail, and will allow
students to quickly apply techniques to their own work and data.

Brian Williams is a lecturer in the Department of Epidemiology at the

London School of Hygiene and Tropical Medicine, UK.

Also available

Resource Selection by Animals

Statistical design and analysis for field studies
Bryan F. J. Manly, Lyman L. McDonald and Dana L. Thomas
Hardback 10 41240140 11,192pllges

Mathematics Recovered for the Natural and Medical Sciences

D. Rosen
Paperback (0 412 41040 OJ. 288 pages

Distance Sampling
Estilfl8ting abundance of biological populations
S. T. Buckland, D. R. Anderson, K. P. Burnham and J. L. Laake
Hardback (0 412 42660 9) ,,,,d paperback (0 412 42670 61. about 480 pages

ISBN 0-412-46220-6

CHAPMAN & HALL

London· Glasgow· New Yorl\.· Tokyo· Melbourne· Madras
IIIIIII
9 780412462207 >

View publication stats