JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms
Royal Statistical Society, Wiley are collaborating with JSTOR to digitize, preserve and
extend access to Supplement to the Journal of the Royal Statistical Society
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
SUPPLEMENT
TO THE
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
2 WISHART-Statistical Treatment of Animal Experiments [No. 1,
area, and should one particular treatment gain an ascendancy over the
others, it meant the introduction at a certain stage of an extra
animal. It was difficult, too, to fix in advance an integral number of
cows which could be depended on to graze the area effectively, while
mnixed grazing brought in other problems, and animals had a way of
falling sick and having to be replaced.
By that time satisfactory methods of crop experimentation had
been devised by Fisher, and the randomized blocks and Latin-square
lay-outs were becoming familiar, together with the essential principles
of good experimental design illustrated by these arrahgements. It
was natural, therefore, that animal experiments should be studied,
and criticised, in the light of these principles. From this point of
view the experiment I have just described was not satisfactory, since,
although live-weight gains were recorded for each cow on the plot,
there was no real replication of plots. The variation in live-weight
gain between cows on the same plot could not safely be taken as a
measure of the variation between cows in different plots; further,
the plots were bound to differ in inherent fertility, and a superiority
in pasture growth, as measured by its effect on animal growth, could
not be ascribed solely to the greater effectiveness of the particular
slag in supplying phosphatic fertiliser.
When it was suggested that the randomized block technique
should be applied to animals, various difficulties were raised. Thus,
it was stated that animals were more variable than field plots, too
variable, in fact, for small differences in growth to be detected.
Further, if this variability was to be reduced, all animals in one
experiment ought to be offspring of the same parents, and ought to
be of the same age and weight at the start of the experiment. This
limits the number of animals in a single experiment to the number of
young born at any one time, and even here it may be impossible to get
many of comparable weight. Thus the experiment can only be a
small one, and even if the standard error of a single animal's live-
weight gain is reduced by such a process of selection, the standard
error of the mean of any treatment will not, in general, be small,
beeause of the limited number of animals available. The expense of
animal experimentation is bound to limit the numbers capable of
being dealt with, thus keeping down the number of treatments to be
tested, and making an unsatisfactory experiment generally, since
while there may be several thousand unit plants of wheat in a single
plot of a cereal experiment, and several hundred unit roots on a plot
of potatoes or sugar beet, the unit with animal experimentation will
often be the single animal.
For these reasons progress in the direction of laying out animal
experiments which would permit of adequate statistical examination
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] WISHART-Statistical Treatment of Animal Expertiments 3
of the results has beell slow, and experimelnts have usually, been of a
very simple liature, comparable in fact to the duplicate plot experi-
ments that were in vogue prior to the randomized block era, except
that there was not always duplication. There have been exceptions
in the case of small animals; some of you may remember hearing
from Mr. Hartley the details of a fairly elaborate experiment on
poultry, at a recent meeting of this Section. I turn now to the series
of nutrition experiments on pigs- begun at Cambridge some two or
three years ago. These will serve as well as anything to illustrate
the methods involved, and they are experiments that I happen to
know something about. There may be others present who can deal
with cows, or sheep, or by contrast with small-scale animals such as
rabbits, or even rats and mice, while the point will not be lost sight of
that experiments on human nutrition deal with that very variable
quantity, the human animal. At an early meeting of this Section we
were given a contrast betweeni an experiment on rubber in Malaya
covering 2,ooo acres and one at East Malling, for which the total area
was only i/iOO acre. There may well be conitrasts as great on the
animal side.
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
4 WISHART-Statistical Treatment of Animal Experinments [No. 1,
Individual-feeding trial
From the litters of each of five sows were selected six pigs, three
being hogs and three gilts. The six from any one litter were run
together in a pen, one hog and one gilt being given ration A, another
pair ration B, and the third pair ration C. Mechanical methods were
adopted to segregate the pigs at feeding time into small individual
pens, so that each pig had undisturbed access to its day's ration.
The experiment consisted of five pens, or thirty pigs in all. Weighings
were recorded as in the group-feeding trial. The experiment is of
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] WISHART-Statistical Treatment of Animal Experimtents 5
the randomized block type (for the rations were randomized in the
feeding-boxes), each block consisting of the pigs from one litter.
Thus pen (or litter) differences could be eliminated by the method
of the analysis of variance, showing the possibilities of this type of
experiment in cases where more pigs are needed than are available
fromn a single litter. The analysis follows the usual lines, and it is
possible to examine the effect of food, a possible sex difference, and
the interaction of these two effects, on any one of the variables.
These were: live-weight gain, meal consumption (known for each
pig), and a great number of post-slaughter measurements on the
carcase whose examination would reveal, if any, differences in
the conformation of the bacon pigs produced in consequence of the
differences in the composition of the food.
In the case of live-weight gain the differences were small, the
figures being I5-0, I46-4 and I42-6 lb. for A, B and C respectively,
with a standard error of 3-0 lb., or 2-I per cent. One very satisfactory
feature of the experiment was the fact that the accuracy was much
greater than that of the group trial; the standard error of each pig's
live-weight gain was 9-6 lb., or 6-5 per cent. of the mean, compared
with I7 3 lb., or I2-5 per cent. in the other case. The same story
is told by the means in both cases. Small as is the standard error,
the differences between A, B and C are not, however, significant on
the analysis of variance test, even when we note that nearly all the
sum of squares for the 2 degrees of freedom for food is contained
in the single degree of freedom component of " principal effect ",
which measures the drop from A to C. There is a way, however,
in which accuracy can be gained. The growth of a pig in any one
week obviously depends to a considerable extent on its weight at
the beginning of the week. Thus differences in initial age and weight
at the beginning may, and usually will, affect the comparisons. We
have no difficulty with age in the present case, but the limited
amount of experimental material makes it impossible to have the
six pigs in any one pen of the same initial weight. But this does
not really matter, as by means of the analysis of covariance we can
examine food and sex differences in live-weight gain after correction
for initial weight-i.e., adjusted to what they would have been,
knowing the regression of live-weight gain on initial weight, had all
pigs started at the same weight. When this was done it was found
that the food differences were still insignificant on the 2 degrees of
freedom, but when the single degree of freedom " principal effect "
was isolated, this was significant at the 5 per cent. probability level.
Thus, by a process of " squeezing," the drop in live-weight gain
from A to C was declared to be significant; the drop was 5-6 per cent.,
which figure had a standard error of 2-4, compared with 2-9 before
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
6 WISHART-Statistical Treatment of Animal Experiments [No. 1,
adjustment for initial weight (without adjustment the drop was 5-7
per cent.). In neither case was there a significant difference due to
sex. It is of interest to observe that live-weight gain was signi-
ficantly correlated with initial weight (r = + 0.60), while in the
case of the group trial the correlation was not significant (r = + 0.36),
due perhaps to competition at the feeding-trough, the weakest
going, possibly literally, to the wall. This is a further point in
favour of the individual trial.
These results, together with the results of examination of the
rest of the data, were published two years ago.' The experiment
was found to be quite workable without too much trouble, once the
piggery had been built for the purpose. It was the pioneer of a
number of trials which have been conducted since, limited, of course,
to the same total number of pigs and the same number of treatments,
although the treatments themselves have varied. As far as I was
concerned, the figures then left the nutrition experimental stage and
entered on a new stage-that of the statistical study of growth.
As it is the statistical treatment of such experiments that is the
subject of our discussion (though this involves, naturally, the
question of lay-out), I shall ask you to bear with me a little longer
while I explain what I have since done.
I have always thought it a pity that more use was not made of
the regular weighings of the animals in such an experiment as this.
Ten years ago I fitted straight lines by the method of least squares
to the weights of the cows in the basic-slag trial, in the belief that
this would furnish a more accurate measure of the growth than the
difference between initial and final weights, and I turned now to the
study of the growth curves of the pigs. These appeared on graphical
inspection to be parabolic in character, with an upward curvature
-i.e., the weekly gains rose up to practically the end of the experi-
ment. Now, had the gain in weight been proportional to the weight
at the beginning of any one week, the logarithms of the weekly figures
would have been linear, but plotting revealed that this was not so,
for the graphs showed a downward curvature. I therefore began
with the fitting of second-degree parabole to the actual weights,
figures more easily interpretable by the practical man than the
logarithms, having ascertained graphically that such a parabola
appeared to be a good fit to the data. There were now two variates
for each pig, g, the average growth rate in lb. per week, and h, pro-
portional to the rate of change of growth in lb. per week per week.
Analysis of variance showed that on g the food difference was signi-
ficant on the single degree of freedom for " principal effect ", repre-
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] WISHART-Statistical Treatment of Animal Experiments 7
rgh. V ( gh 2) approxim
0-20 0 26.
V\1 - 0.65 2)-
Thus it would appear that the significance of the sex effect for h
corrected for both g and w should be less than for h corrected for g
only. This was borne out by a multiple covariance analysis, which
reduced the sex effect to complete insignificance, although the
multiple correlation was not significant (R = 0 23).
Analysis of variance of the cubic term i showed a non-significant
sex difference, but very marked differences due to food treatments,
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
8 WISHART-Statistical Treatment of Animal Experiments [No. 1,
far more significant than anything hitherto reached. The figures for
the averages of ten pigs each are:
Standard.
A B c Mean Error
Mean i ... -000959 -000610 -000248 -0O00606 0 00109
Per cent ... 158.3 100-7 40.9 1000 18.0
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] WISHART-Statistical Treatment of Animal Experiments 9
180
160
140
120
80
60
40
A 8 C A B C
HOGS GILTS
FIG. 1.-Growth curves of the "mean pig" for each treatment and sex
(actual weights).
the week after reaching 42 lb. and for the week before reaching i8o lb.
shows the same relative effects. Comparing hogs with gilts, averaging
for food treatments, we find that hogs start off with the greater
initial weekly gain and end with the smaller final weekly gain.
Hogs were, however, heavier at the beginning, and correcting for
this as before by interpolation, we find that the initial gains become
equal, though the final gains show no change. The curves showing
the two extremes in growth are those for hogs under the C treatment
and for gilts under the A treatment. These are side by side in
the diagram.
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
10 WISHART-Statistical Treatment of Animal Experiments [No. 1,
and quartic terms were not. Accordingly the first two terms, which
we may call g' and h', were fitted to all pigs, and the usual analysis
of variance was worked out in both cases. The table for g' is as
follows, g' being calculated from the logarithms taken to two
places of decimals, and treated as whole numbers:
Analysis of Variance-g'
Variation D.F. Sum of Squares Mean Square
Sex ... ... ... 1 0-2811 0*2811 1-1435
Food ... ... ... 2 0-2101 0-1050 0.6511
Interaction ... ... 2 0-0329 0.0164
Pens ... ... ... 4 0.9115 0.2279 1.0386
Error ... ... ... 20 0.5710 0*02855
The sex difference was very significant, the mean for gilts being
distinctly higher than that for hogs. Further, the food differences
were significant, the trend being in the same direction as in the case
of the g values. The following tables summarize the results, apart
from the significant pen differences:
Summary of Results-g'
Hogs Gilts M[ean Standard Error
Mean g' . ... 4-091 4-285 4.188 0.044
Per cent. 97.7 102-3 100.0 1.06
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] WISHART-Statisticat Treatment of Animal Experiments 11
12 30
c^o
~~~~~~~/ /
A B C A B C
HO8GS GILrS
weekly averages of the logarithms of the weights for each sex and
treatment. The graphs bring out the differences in average slope
as between A, B and C on the one hand, and between hogs and gults
on the other.*
It is not suggested that all the above calculations are necessary in
any similar experiment. In fact, had we started with the logarithms
instead of the actual weights, it is unlikely that the latter would have
had the intensive examination which has been given to themt. But
the figures are presented as an illustration of attempts that may be
* Further examination has shown that the cubic terms fitted to the B and
a hogs, and the C gults, are significant. The values of a3' have been added to
the table since the meeting, and the diagram has been re-drawn. It would
appear, therefore, that working on the logarithms does not necessarily save
computational labour to the extent hoped for, if a comaplete specification of
growth has to be provided.
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
12 WISHART-Statistical Treatment of Animal Experirments [No. 1,
The CHAIRMAN said that it was not the Society's custom when a
discussion had been opened to move a formal vote of thanks to the
opener, so he would take the opportunity of expressing his thanks
References.
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] Discussion on Animal Experiments 13
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
14 Discussion [No. 1,
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] on Animal Experiments 15
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
16 Discussion [No. 1,
MISS TURNER said this was her first appearance in the Society,
and she expressed her thanks for being allowed to be present. She
wished to speak on some problems which might be likened to the
long-term experiments of which Dr. Yates had spoken. Her own work
had been done in Australia, and related mainly to sheep. In
Australia the natural pastures were deteriorating, and one method
of improving them was to plant English grasses and fertilize them
annually. Some graziers had raised the objection that; although
this increased wool production, it caused coarsening of the wool,
with consequent decrease in value. It was necessary, therefore, to
find out what effect the improvement of the pasture had on the wool,
and whether such an effect was cumulative. It was also important
to determine the effect of improved nutrition, such as that provided
by improved pastures, on lambs whose growth had been checked at
various stages by feeding on poor pastures. Both these might be
described as long-term experiments.
Another point was the use of the group trial. Miss Turner
agreed that individual feeding was essential for a laboratory trial,
but thought another kind of trial was necessary in Australian
conditions. Pigs could be fed in pens either individually or in
groups, and when thus feeding them in groups one was approaching
quite closely to actual husbandry conditions. But feeding sheep
in pens was getting nowhere near actual farming conditions in
Australia, and the grazier was far more likely to be convinced by a
trial which did approach such conditions. Experimenters had to
convince both themselves and the grazier by translating any trial
from laboratory conditions to the field. She would describe one
large experiment with cattle which was being carried out by the
Australian Commonwealth Council for Scientific and Industrial
Research.
A certain cattle disease in Queensland was known to be a deficiency
disease, and preliminary experiment had shown that the feeding
of certain supplements was successful in preventing it, 13ut it was
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] on Animal Experiments 17
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
18 Discussion [No. 1,
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] on Animal Experiments 19
Note on the above Table.-Any item could have been tested exactly by the
z test. For testing the last line, see, for example, Proc. Camb. Phil. Soc., 30
(1934), 327-340 (339). For the approximate x2 test for the last line, as used
here, see Proc. Camb. Phil. Soc., 34 (1938), 33-40 (380). The other items were
chosen to make the table additive. It should be noted that (i) the x2 test is
leas accurate for these items, since the multiplying factor used to get the best
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
20 Discussion [No. 1,
x2 approximation does no
(ii) caution should be used in interpreting the significance of any partial effect
such as h.g (h adjusted for g), since the adjustment implies using covariance
with g, which is itself varying significantly between treatments. Nevertheless
the entire table still serves as a useful summary of the significance of the food
effect. It was amended to include the cubic term for the logarithmic analysis,
which was not completed by Dr. Wishart until after the meeting. The
significance level of the total x2 nearly reaches P 0-01 for the direct analysis
but is only about P = 0 02 for the logarithms. It is interesting, however, to
note the large contribution to x2 for the logarithms from the linear term. The
cubic term in this analysis was found just significant by Dr. Wishart, but does
not provide any extra significance when tested jointly with the other constants.
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
1939] on 4nimal Experiments 21
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms
22 Discussion on Animal Experiments [No. 1,
This content downloaded from 14.139.194.12 on Fri, 30 Nov 2018 06:34:30 UTC
All use subject to https://about.jstor.org/terms