Professional Documents
Culture Documents
most scientists will go through their whole career without having to delve into more
complicated designs.
ms with third
have observed in an unmanipulated system is really due to the factors that we have
variables
measured andBook istoto
Title
not due carrywith
correlation out experimental
some Stage
other unmeasured variable. The only
effect that could explain this
Supplier
observation? Date
Key definitions
way to be certain of removing problems with third variables is to carry out experimental
Experimental Design for the Life Sciences Revise 1 Thomson Digital 08 Apr 2016
There is a lot of jargon associated with experimental design and
manipulations.
Reverse causation
r The second problemBook Title
of correlational studies is reverse causation. ThisStage
can occur when Supplier
Reverse causation is Date statistical analysis. We have not tried to avoid this. Indeed, we
we see a relationship between factors A and B. It means mistakenly assuming that fac mistakenly concluding that
studies is reverse causation.
Experimental Design for theThis can occur whenRevise 1Reverse
Life Sciences causation
Thomson Digital is
08variable
Apr 2016
the broad design of your
tor A influences study
factor variable A influences
B, when in fact it is change in B that drives change in A.
themselves tomistakenly
B when actually it is B that
concluding that
have deliberately tried to introduce you to as much of the jar
rs A and B. ItFormeans mistakenly
example, imagine assuming
a survey shows that those whothat fac
consider regularly
influences A.
ct it is change
use recreational drugs also consider themselves to have financial worries. It might be
intoBconclude
those
tempting that drives
conditions
that arechange
a drug stillis likely
habit into
likely to A.befinancial
cause more problems.
variable A influences variable
representative
The reversethan the conditions in a lab. gon as we can. By being exposed to this jargon, we hope that it
causation explanation is that people who have financial problems are more
Thus, itof
is usually safer to generalize
B when
from a field study than likely actually it is B that
a lab study. Generalization is
ws thatthe
lecting those
broad who
designconsider themselves
your study to regularly will become second nature to you to use—and understand—it.
influences A.
than average to turn to drugs (perhaps as a way to temporarily escape their financial
a subject that is at the heart of statistical methodology, so we explore it a little further
r themselves to have financial worries. It might be
worries). In this case, we’d consider the first explanation to be the more likely; but the
reverse
bit is likely operating
to cause
those conditions are still likely to be more representative than the conditions in a lab.
incausation
Statistics Box 3.1.
explanation
financial
Thus,
is at least plausible. Indeed, both mechanisms could be
it is usuallyproblems.
simultaneously. The from
safer to generalize reverse
a field study than a lab study. Generalization is
This should also make reading more advanced texts and scien
Sometimes, our research question means that a lab study is impractical. If we are
ple who have
the number
Let’s consider another example of reverse causation. There is a correlation between
a subject that is at the heart of statistical methodology, so we explore it a little further
financial
interested
of storksin problems
differences
nesting areofmore
between
on the chimneys
in Statistics Box 3.1.
likelyin and
the farmhouse
a Dutch sexes effort expended in feeding the young in
the number tific papers on experimental design less daunting. However, to
of children
aps as a way to intemporarily
the the family
nest using living in thefinches,
zebra house. Thisthen
escape sounds
researchtheir
bizarre until
itfinancial
would beyou think
very that larger
adifficult
lab studytoiscreate foraging
If weconditions
Sometimes, our question means that
families tend to live in larger houses with more chimneys available as nest sites for the
in an families
aviarylead
impractical. are
make negotiating the minefield of jargon more straightforward,
the first explanation towhere
storks. Large interested thethe conditions
beintodifferences
morebetween
likely;
more opportunities
were
the
but
for nesting
close
sexesthe enough
instorks
storks; effort to those naturally experienced by
expended
don’t lead to in feeding the young in
east plausible.
wild birds.
large families!
Indeed,
Reverse causation both
is
In using
the nest the wild
zebraparents
mechanisms
notwhere
a universal
will itbring
finches, then
could
problem of were
wouldback
be
correlational
a great
be very difficultdiversity
studies. Sometimes
to createof food conditions
foraging types gathered we have increased the number of definitions of key terms pro
from inaanwide
aviaryvariety the
of conditions
microhabitats close enough
sometimes tohundreds
those naturally experienced
of metres by This
distant.
reverse causation can be discounted as a plausible explanation. In our correlational
study would
wild birds. In the wild parents will bring back a great diversity of food types gathered
bestreamer
of male tail impossible to faithfully
length and recreate
mating success describedin theitlab.
earlier, wouldThe
be danger is that any aviary vided and increased the level of detail in all definitions. Each key
from a wide variety of microhabitats sometimes hundreds of metres distant. This
f reverse causation.
experiments
extremely There
difficult thatishow
to imagine a
youcorrelation
theconducted between
on
number of matingsthis would
a male beindiscounted
gets later a season by Devil’s Advocates as
would be impossible to faithfully recreate in the lab. The danger is that any aviary
could affect his tail length when we catch him at the beginning of the season. However, word or phrase is emboldened at the point where we first use it,
of being
e chimneysimagine
athat entirely
Dutch
experiments
artefactual,
farmhouse
we had instead that you
caught
due to
and the
andconducted
measured
your
the
asking
number
onmales
this at
would
birds to operate in a foraging situation far
beofdiscounted
the end the season. by Devil’s Advocates as
In thatremoved
ouse. Thishissounds
case itbeingfrom
might that
entirely
be quite which
artefactual,
plausible they
that
bizarre until you think that larger
the have
duenumber been
to your evolutionarily
of asking
matings birds
a maletogets
hormone levels. The change in hormone levels might then affect the way that the
adapted
operate
affects to cope
in a foraging with. far
situation However,
and is given a clear definition in a box nearby.
if your research
removed question
from is they
that which about how
have egg
been laying affects
evolutionarily theto
adapted distribution of reserves of
cope with. However,
with moremale’s
chimneys
calcium inavailable
different as nest
research question
parts of sites
is about howfor
females’ eggthe
tail develops during the season in such a way that his tail length comes to depend
if your bodies, laying affects
then the distribution
it would be easieroftoreserves
argue of
that this
on the number of matings that he gets. In that case we would observe the relationship
pportunities for
is less
between tail
calcium
nesting in
likelyand
length to different
storks; parts
storks
be influenced
matings that
of by
females’
don’t
we expect, the bodies,to
lead then
but itenvironmental
would be
it would be easierand
conditions,
nothing to do with
to argue
hencethat this
a laboratory
is less
the reason that likely to
we believe
unlikely to youexperiment
be influenced
(females
mightofbe
and you may know
by the
preferring long
reasonable.
no example So,
of such
environmental
tails). Such a pathwayconditions,
as ever,
a thing the most
happening.
might seem and hence a laboratory
experiment might be reasonable. So, as ever, the most appropriate approach depends
appropriate
However, you approach depends
Statistics boxes
on your research question.
rsal problem of correlational
can be sure that the Devil’s
on your studies.
Advocate
research Sometimes
will find such arguments highly compelling unless
question.
ed as a plausible explanation. In our correlational To emphasize the important link between good experimental
and mating success described
StAtiSticS
StAtiSticS
earlier,
Box
Box
it would from
Generalizing
3.1 Generalizing
3.1
be
froma study
a study design and statistics we now include a number of statistical
he number of matings a male gets later in a season
catch
An important issue in the interpretation of a study is how widely applicable the re
An important issue in the interpretation of a study is how widely applicable the re
him at the beginning sults of theof theareseason.
study However,
likely to be. That is, how safely can you generalize from the
boxes. These boxes should help you to see how thinking about
sults of the study are likely to be. That is, how safely can you generalize from the10:42 AM
03-RuxtonandColegrave-Chap03.indd 37 09/04/16
and measured the particular set of subjects used in your study to the wider world? Imagine that we
males
particular setatofthe end of
subjects usedthe
in season.
your study to the wider world? Imagine that we
measured representative samples of UK adult males and females and found that
design helps you think about statistics and vice versa. We have
le that the number of matings
measured a male
representative gets
samples affects
of UK adult
taller.males
Neitherand
we females and
elsefound
are that
males
males
in
inthen
our sample
our sample
particularly
were
were
interested
on average
on
in our
5 cm
average
specific 5the
cm taller.
nor anyone
Neither
we arewe nor anyone else are
added the boxes at points where we think that keeping the
ormone levels might affect the way thatindividuals, rather interested in what
any effect.
This is a strange set of circumstances, so let’s get this clear. The investigator h
done nothing wrong in this third experiment, the hens were all randomly chose
and the experiment was carried out correctly. The fact that the GM shells are
x Preface
Flow chart
The exact process of designing an experiment will vary considerably between studies.
Nevertheless, there are key stages in the design process that will apply to most, if not
all, studies. The flow chart at the end of this book is intended to summarize and guide
you through the main stages of designing an experiment. We have indicated at each
point in the chart the sections of the book that are most relevant.
Acknowledgements from
the fourth edition
In the 15 years we have been involved with this book, we have had explicit feedback
from countless students and colleagues; and have had our thinking on experimental
design sharpened by countless others. We don’t attempt even a partial list; but if you
spot your influence anywhere in this work be assured of our profound thanks. Even
if you can’t spot your direct influence be aware that we are both really grateful for
everyone who has forced us to think harder about experimental design and how to
explain our ideas. We are grateful for the support we have had from editorial staff at
OUP (most recently Jessica White) who have always been supportive when required
without ever feeling heavy-handed; the production staff (most recently Tomas Furby)
have always been a model of quiet competence. Julian Thomas is a copy editor who
combines thoroughness with a collegiate, open, and interactive approach. We have
both been privileged to work at institutions that have fostered the academic freedom
that has allowed us to devote so much time to this endeavour. We have both been
blessed by friends and family that have been tolerant of the demands of this book but
always there to provide compelling reasons not to let it swallow our lives entirely.
Contents
2.3 Controls23
2.3.1 Different types of control23
xiv Contents
4.2 Replication47
5 Pseudoreplication67
7.2 Randomization102
7.2.1 Randomizing study subjects102
7.2.2 Randomizing other aspects of your study103
7.2.3 Haphazard allocation104
7.2.4 Balanced and unbalanced allocation105
8.2 Interactions110
9.6 Covariates134
10 Within-subject designs141
11 Taking measurements152
11.1 Calibration153
When some life scientists see the phrase ‘experimental design’, it can either ease them
gently into a deep sleep or cause them to run away screaming. For many, experimen-
tal design conjures up unhappy memories of mathematics or statistics lessons, and is
generally thought of as something difficult that should be left to statisticians. Wrong
on both counts! Designing simple but effective experiments doesn’t require difficult
maths. Instead, experimental design is more about common sense, biological insight,
and careful planning. Having said that, it does require a certain type of common sense,
and there are some basic rules. In this book, we hope to steer you relatively painlessly
towards thinking more effectively about designing experiments.
So why are many life scientists so averse to thinking about design? Part of the reason
is probably that it is easy to think that time spent designing experiments would be
better spent actually doing experiments. After all, the argument goes, we are biologists
2 1 Why you should care about design
so let’s concentrate on the biology and leave the statisticians to worry about the design
and analysis. This attitude has given rise to two myths that you can hear from the
greenest student or the dustiest professor.
Myth 1: It does not matter how you collect your data, there will always be a
statistical ‘fix’ that will allow you to analyse it.
It would be wonderful if this was true, but it is not. There are a large number of statisti-
cal tests out there, and this can lead to the false impression that there must be one
for every situation. However, all statistical tests make assumptions about your data
that must be met before the test can be meaningfully applied. Some of these assump-
tions are very specific to the particular test. If you cannot meet these, there may be a
substitute test that assumes different characteristics of your data. But you may find
that this alternative test only allows you to use your data to answer a different scien-
tific question to the one that you originally asked. This alternative question will almost
certainly be less interesting than your original (otherwise why weren’t you asking it in
the first place?). Further, there are some basic assumptions that apply to many of the
most commonly used statistical tests, and you ignore these at your peril. For instance,
statistical tests generally assume that your data consist of what statisticians refer to
Independent data points as independent data points (more on this in Chapter 5). If your data don’t meet this
come from unconnected criterion then there is much less that statistics can do to help you. Careful design will
individuals. If the measured
allow you to avoid this fate. It’s also generally true that well-designed experiments re-
value from one individual
quire simpler statistical methods to analyse them than less-well-designed experiments.
gives no clue as to which
of the possible values the So time spent carefully designing your experiment might save you a lot of time later in
measurement of another the study trying to master more complex statistical techniques than you really needed.
individual will produce, then The group of experimental subjects that we use in our experiment (called the
the two measurements are sample) needs to be representative of the wider set of individuals in which we are
independent.
interested (called the population). One key way to achieve this is to ensure that each
individual selected is not linked in some way to another individual in the sample, i.e.
Q 1.1 If we wanted to
measure the prevalence of they are independent. For example, if we were surveying human food preferences,
both left-handedness and then gathering food-preference data from five members of the same family would not
religious practices among prison produce five independent data points. We would expect that members of the same
inmates, what population would family are more likely to share food preferences than two unconnected individuals,
we sample from?
since family members often have a long history of eating together. Similarly, gathering
Q 1.2 If we find that two data from the same person on five separate occasions certainly does not provide five
people in our sample have independent data points, since a person’s preference on one occasion is likely to be a
been sharing a prison cell for the good guide to their preferences a little while later.
last 12 months, will data from
them be independent? Myth 2: If you collect lots of data something interesting will come out, and you’ll
be able to detect even very subtle effects.
It is always reassuring to have a notebook full of data. If nothing else, it will convince
your supervisor that you have been working hard. However, quantity of data is really no
substitute for quality. A small quantity of carefully collected data, which can be easily
analysed with powerful statistics, has a good chance of allowing you to detect interest-
ing biological effects. In contrast, no matter how much data you have collected, if it is
1.2 The costs of poor design 3
of poor quality, it will be unlikely to shed much light on anything. More painfully, it will
probably have taken far longer and more resources to collect than a smaller sample of
good data.
Language: English
Produced by the
Division of Publications
National Park Service