Student’s review of the 1st edition of
John Aldrich,
“Student” was the pen name of William Sealy Gosset MacTutor Wilkipedia (1876-1937), a brewer of Guinness in Dublin and the pioneer in the analysis of small samples. Gosset had studied with Karl Pearson, the biometrician and leading statistician of the time. Although Pearson helped with the 1908 paper, he was used to working with large samples and doubted the value of small sample analysis. The controversy and disputes, to which Student alludes to in the review, were primarily between Pearson and Fisher and began about 10 years later. In a letter of April 3rd 1922 Gosset told Fisher “In most of your differences with Pearson I am altogether on your side and in some cases I have agreed to differ from him long ago.”
Of the established statisticians Gosset was the closest to Fisher. They were in regular touch from Fisher’s joining Rothamsted in 1919, though they met rarely and communicated by letter. Only a few of Fisher’s letters have survived but the development of his ideas can be traced through Gosset’s letters reacting to them. Student was the least “innocent” of the reviewers of the Statistical Methods for, not only had he been kept informed of the development of Fisher’s ideas, he had read the book in proof. Gosset sent Fisher a long letter with corrections and comments on October 20th 1924. Some of the comments are repeated in the review, including his opinion of the unsuitability of Example 1 and his practical objections to randomisation. On the latter he wrote, “You would want a large lunatic asylum for the operators who are apt to make mistakes enough even at present.” He continued, “If you say anything about Student in your preface you should I think make a note of his disagreement with the practical part of the thing: of course he agrees in theory.” Later in the 1930s the two would publicly disagree about randomisation.
The biographies by Egon Pearson and Box treat the relationship between Gosset and Fisher. Both draw heavily on the Letters but the Pearson volume is the more useful for indicating Student’s input into the Methods and for tracing the later disagreements between Fisher and Gosset. 2008 is the centenary of Student’s famous paper and Zabell (2008) and his commentators discuss how it influenced Fisher.
In
all there were 6 major reviews of the 1st edition of the Statistical
Methods. The other 5 are also
available on the web.
All are worth looking at as they emphasise different facets of the book
________________________________________________________________________
Student’s review (Eugenics
Review, 18, (1926), 148-150.) Now on-line.
Fisher,
Owing
to his connection with large scale experimental work it happened that the
reviewer was one of the first to draw attention of Biometricians to the fact
that a special technique is required in dealing with ‘small’ samples, by which
is meant samples so small that the statistical constants of the population
cannot be replaced by those of the samples without appreciable error.
Further, it was the incompleteness,
from the mathematical standpoint, of his solution of one of the fundamental
problems which first drew Dr. Fisher’s attention to the subject and it is
therefore with particular pleasure that he welcomes the publication of Statistical Methods for Research Workers,
the first book on Statistics which deals mainly with this special technique.
Dr. Fisher is not himself occupied
to any great extent in experimental work, but as Chief Statistician to the Rothamsted Experimental Station he is in close touch with
experiments carried out both in the field and in the various laboratories of
that institution, and it is in response to this stimulus that the book has been
written.
We must not, therefore, consider it
as a text-book but rather as a summary of the author’s own work; and the
proverbial aloofness of the text book from controversy, and often it is to be
feared from current practice, is not to be expected: in fact it must be
admitted that while no one disputes the fact that Biometrical methods must be modified,
there is still controversy as the exact nature of the most suitable
modification for the purpose.
Nevertheless, it is the reviewer’s
opinion that even if all points which are still in dispute were conceded to his
opponents, we should seldom in practice draw inferences from our data which
would be appreciably different from those which we shall draw if we follow our
author.
One of Dr. Fisher’s merits is his
instance throughout on the fact that all correct methods of treating statistics
which make use of all the available relevant information must necessarily give
the same result, and this has led to a unity of outlook which has not always
been sufficiently emphasised.
In this he has been helped by what,
in anything but mathematics, would be considered coincidences, such as the fact
that the χ2 distribution is identical in form with that of the
variance of samples drawn from a normal population.
The following tabular statement may illustrate better than anything the general arrangement of the book
|
Subject Matter |
Tables
|
Number
of numerical
examples |
I |
Introductory. |
|
1 |
II |
Diagrams |
|
|
III |
Distributions, (i) Normal, (ii) Poisson,
(iii) Binomial |
I
& II of probability integral |
6 |
IV |
Test
of goodness of fit, independence or homogeneity |
III
of χ2. |
8 |
V |
Test
of significance of means, differences of means and regression coefficients |
IV
of t. |
9 |
VI |
The
correlation coefficient, partial correlation and the z transformation |
VA of r for zero correlation. VB
r to z for any correlation. |
9 |
VII |
Intra-class
correlation and the analysis of Variance |
VI
of z for P= .05. |
8 |
VIII |
Further
applications of the analysis of Variance, experimental planning, the |
|
|
Chapter I is not easy to follow for those unfamiliar with the “jargon” of the science and such readers are advised to skip freely after Section 2, but to return and read it carefully after they have become accustomed to the use and implications of the various terms employed. This advice, however, does not apply to Example 1 which is difficult alike to those unfamiliar either with mathematics or with the higher genetics, for it may be neglected by both classes since it does not illustrate any of the methods exhibited in the later chapters of the book.
Chapters II & III, excepting sections 17 and 19 which
might have been included in Chapter IV, are mainly descriptive and should be
mastered by anyone who wishes to apply the tests in Chapters IV and V.
Chapters VI, VII & VIII trace the connection between
correlation and the analysis of variance which is dealt with in a manner
particularly convenient for those who are concerned with the planning and
interpretation of Agricultural experiments. Nevertheless the reviewer sometimes
doubts whether Dr. Fisher’s rigid adherence to the true Gospel of random, if
controlled, arrangement will find favour with those who require a regular system
as an aid to method and useful check for the avoidance of mistakes.
The tables of mathematical functions are given in skeleton
form, and I—VA are entered contrary to the usual practice by using
values of P as argument: they will be found, however, to be well suited to
their purpose and their compactness conduces to speed in use. They are not, of
course, intended to replace the full tables published elsewhere.
Table VI is admittedly less complete than the others as it
only gives z for the value P= .05 which the author uses as his ordinary limit
of significance.
All seven of the tables are reprinted at the end of the
book and can be opened out so as to be used while reading any page of the book.
They can also be cut out and mounted on cardboard, either with cloth hinges to
fold up like a map, or to form the outside faces of two triangular prisms which
stand up on the table when in use and one of which can slip inside the other
for storage purposes.
In connection with the c2
test there is a statement on p. 80 which the reviewer cannot
altogether endorse, when the author states that in cases where χ2
[sic] is unexpectedly small and P takes such a value as .999; “the hypothesis
considered is as definitely disproved as if P had been .001.”
Now in the first place there is this fundamental difference
between the values .001 and .999: —while it is generally easy to formulate any
number of likely hypotheses to account for the low value, it is apart from
blunders or cheating, exceptional to find an alternative explanation to account
for .999 which is not itself extremely improbable: in the second place χ2,
with small samples, can only take a limited number of values of which zero,
corresponding to P=1, is not unlikely to occur by chance.
For example, if twenty pairs of coins be tossed we obtain
as the “calculated” distribution on the hypothesis that heads and tails are
equally likely, five cases of two heads, ten of head and tail and five of two
tails. This will actually occur in the long run rather more than once in twenty
five trials; in such cases P=1 but the hypothesis could not be said to be in
any way weakened.
Dr. Fisher’s book will doubtless be found in the
laboratories of those who realise the necessity for statistical treatment of experimental
results, but it should not be expected that full, perhaps even in extreme cases
any, use can be made of such a book without contact either personal or by
correspondence with someone familiar with its subject matter.
|
|
|
STUDENT |
|
|
|
|
|
|
|
|
|
|
·
Information
on the technical terms used in the review, such as Fisher’s z transformation of
the correlation coefficient, is available on the Earliest Uses pages:
see Probability
& Statistics on the Earliest Uses Pages