The British Society for the Philosophy of Science
Review: Likelihood
Author(s): Ian Hacking
Review by: Ian Hacking
Source: The British Journal for the Philosophy of Science, Vol. 23, No. 2 (May, 1972), pp.
132-137
Published by: Oxford University Press on behalf of The British Society for the
Philosophy of Science
Stable URL: http://www.jstor.org/stable/686438
Accessed: 30-07-2016 16:40 UTC
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
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 [email protected].
The British Society for the Philosophy of Science, Oxford University Press are
collaborating with JSTOR to digitize, preserve and extend access to The British Journal for the
Philosophy of Science
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
132 Ian Hacking
restricted to cases in which the notion of 'chance' is involved, its domain is
narrower than that of the subjective Bayesians; at the same time it is more
explicit in its application to problems of inference in the natural sciences.
The last chapter shows signs of having been written some time after the earlier
ones, and it seems to shift the emphasis in places. For example, on page 222
Hacking comes near to discussing a 'goodness of fit' situation, and says 'My
theory of statistical support does not attempt rigorous analysis of the reasoning
here.... The theory of statistical support cannot judge the force with which
an experiment counts against a simplifying assumption.' To this extent he appears
to agree with the comment made above on his treatment of tests of significance.
Again, on page 219, Hacking appears to entertain the possibility of something
corresponding to the idea of 'prior likelihood' of 'acceptability' referred to above,
and he explicitly refers to the point made by Fraser (and also by the present
writer) that group invariance and other structural features of an experimental
set-up may be relevant to its statistical interpretation.
It is clear that in this area there is much further exploration to be done.
Hacking's book remains an invaluable guide book for anyone willing to join in
this task.
G. A. BARNARD
University of Essex
REFERENCES
ANSCOMBE, F. J. [1962]: 'Tests of Goodness of Fit', Journal of the Royal Statistical Society
(B), 25, pp. 81-94.
BARNARD, G. A. [1947]: Review of Abraham Wald: 'Sequential Analysis', Journal of the
American Statistical Association, 42, pp. 658-64.
BARNARD, G. A. [1949]: 'Statistical Inference', Journal of the Royal Statistical Society (B),
II, pp. 115-49.
BARNARD, G. A. [1950]: 'On the Fisher-Behrens Test', Biometrika, 37, pp. 203-7.
BARNARD, G. A. [1951]: 'The Theory of Information', Journal of the Royal Statistical
Society (B), 13, pp. 46-64.
BARNARD, G. A., JENKINS, G. M. and WINSTEN, C. B. [1962]: 'Likelihood Inference and
Time Series', Journal of the Royal Statistical Society (A), 125, pp. 321-72.
FISHER, R. A. [I925a]: 'Theory of Statistical Estimation', Proceedings of the Cambridge
Philosophical Society, 22, pp. 700-25.
FISHER, R. A. [1925b]: Statistical Methods for Research Workers.
RtNYI, A. [1955]: 'On a New Axiomatic Theory of Probability', Acta Mathematica
Academiae Scientiarum Hungaricae, 6, pp. 285-335.
ROBBINS, H. [1952]: 'Asymptotically Sub-Minimax Solutions of the Compound Decision
Problem' in J. Neyman (ed.): Proceedings of the Second Berkeley Symposium on
Mathematical Statistics and Probability, pp. 13 1-48.
LIKELIHOOD
The fundamental question about statistical inference is philosophical: what
primitive concepts are to be used? Only two answers are popular today. Edwards
is the first scientist to write a systematic monograph advocating a third answer.?
1 Edwards, A. F. [1972]: Likelihood. An Account of the Statistical Concept of Likelihood
and its Application to Scientific Inference. Cambridge: Cambridge University Press.
?3.80. Pp. xiii+ 235.
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
Likelihood 133
'Orthodox' statistics admits only one primitive concept, namely physical
probability, a propensity that betrays itself in stable long run relative frequencies.
The statistician devises procedures for testing hypotheses and estimating para-
meters; he chooses among possible procedures by pointing out desirable pro-
perties that show up under repeated sampling. There is much infighting about
what properties are desirable-unbiasedness, minimum variance, size, power,
significance level and all that-but all these properties have to do with long run
operating characteristics of statistical procedures. The orthodox statistician,
be he a follower of Neyman or Fisher or whoever, will not usually allow you to
measure the credibility of any particular estimate or hypothesis. He allows you
to say only that this particular estimate was made by a procedure that has
virtues that would show up under repeated sampling, or, for example, that this
hypothesis is rejected by a procedure that mistakenly rejects hypotheses only
one per cent of the time.
Bayesian statistics tells quite the opposite story. The only primitive concept is
degree of belief, which arguably ought to satisfy the probability calculus. If you
do have probabilistic degrees of belief in the propositions in some field of
interest, there is a simple model of learning from experience that Bayesians find
satisfying. Orthodox statistics is characterised by a host of locally applicable,
often ad hoc, but notably ingenious procedures. A single, simple global analysis
is the delight of the Bayesian.
Thomas Bayes's original paper was published in 1763. The use of repeated
sampling properties in inference goes back to Jacques Bernoulli, published
I7I3. Between these dates J. H. Lambert may have toyed with a third basic
concept, although he subsequently developed the theory of errors in what I
would now call the 'orthodox' way. In 1777 Daniel Bernoulli brought this
third alternative more into the open. To take his engaging example, we know that
one of two archers has been firing at a target. One is an untalented novice, the
other a master. We observe that the shots cluster around the bull. Who aimed?
Knowing the abilities of the two men, we reason: if the novice fired, then some-
thing very unlikely must have happened, but if it was the master, an event of far
greater probability has occurred. On this evidence, we strongly favour the hypo-
thesis, that the master shot the arrows at this target. If we restrict 'probability'
to the orthodox, physical, usage, we cannot say anything about the probability
that this is the target of the master, not the novice. (We could if we knew that the
two men tossed coins to decide who would shoot, but that is not part of our data.)
We can, however, follow D. Bernoulli and compare what R. A. Fisher called the
likelihoods of the two hypotheses. Likelihood is a sort of inverse of physical
probability. The likelihood of the hypotheses h, in the light of data e, is the prob-
ability of observing e if h were in fact true. D. Bernoulli apparently advises us to
prefer hypotheses of greater likelihood given the data.
Euler at once retorted that this advice is metaphysical, not mathematical.
Quite so! The choice of primitive concepts for inference is a matter of 'meta-
physics'. The orthodox statistician has made one metaphysical choice and the
Bayesian another. D. Bernoulli appears to have been proposing a third. Likeli-
hood is of course formally defined in terms of probability, but it is being offered
as a primitive concept of inference; primitive in the sense that it is supposed to
justify inference, and that its use is supposed to need no further justification.
As a primitive concept likelihood did not fare very well, although one finds it
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
134 Ian Hacking
suggested in surprising quarters. Edwards reminds us that John Venn, a good
frequentist, seems to be prepared to use likelihood as a primitive tool in assessing
statistical hypotheses, and that F. P. Ramsey, who put the subjective theory on a
solid modern footing, also, in a late paper, comes down in favour of likelihood.
But for all these hints the idea of likelihood would have lain fallow had it not been
for Fisher. And Fisher was completely ambivalent about the concept. In in-
formative asides Edwards reminds us of this ambivalence and in an historical
paper (submitted to Biometrika) he has expanded this theme into a thorough
history of likelihood.
It is important not to confuse Fisher's 'method of maximum likelihood' with
likelihood as a primitive concept. The former is a valued technique in estimation
theory, and can be justified by a large number of long run sampling properties.
Fisher himself discovered many of these properties, and made the method of
maximum likelihood an integral part of orthodox statistics. Equally, of course,
likelihood is a crucial quantity for the Bayesian, who learns from experience
by multiplying prior probabilities and likelihoods. But for neither orthodox nor
Bayesian statistician is likelihood primitive; it is only a quantity that turns out to
be central in many calculations. Fisher, from time to time, wanted likelihood
to do more.
Fisher had no use for Bayesian analysis unless one is confronted by that rare
situation in which hypotheses of interest are themselves generated randomly
from a chance set-up. So he could not, in general, speak of the probability of an
hypothesis. But he urged, from time to time, that relative likelihoods of hypo-
theses form the natural way to indicate the relation between hypotheses and
evidence. This idea is present in some of the great work of the early I920s, and
recurs somewhat nervously in the middle thirties; it comes out again in Stat-
istical Methods and Scientific Inference, his final major contribution. Fisher even
goes so far as to say that likelihood is much like Keynes's idea of logical prob-
ability, except that it is not subject to the unjustifiable addition law for prob-
abilities: we cannot add likelihoods to get the likelihood of some disjunction of
hypotheses. Harold Jeffreys once told Fisher there was nothing wrong with
postulating likelihood as a basic axiom for inference, and as Edwards puts it,
Fisher 'wistfully' contemplated that, but was never altogether sure it was the
right thing to do. Edwards, in contrast, has no doubts.
Edwards's book is more of a sermon than an attempt to provide logical founda-
tions for a new mode of argument. Himself a geneticist, he is appealing to fellow
scientists to reason in a certain way. He gives plenty of attractive examples, for
the theory is to be tested by having good consequences for science. Edwards
clearly has a strong 'intuition' that likelihood is the right tool, but unlike phil-
osophers who boringly go on about their intuitions, Edwards aims at establishing
a coherent body of method that enables the scientist to analyse his data in a
sensible way.
Concerning the philosophical question, 'What primitive concepts to use?',
Edwards is at first orthodox. There is only one kind of probability, the kind that
shows up as stable relative frequency. He grants, fleetingly, that Bayesians have
arguments showing that if one has degrees of belief in every sort of proposition,
or if one is made to bet on anything under the sun, then in coherence one's
betting rates will be probabilities. But he retorts that he simply does not have
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
Likelihood 135
degrees of belief in the genetical hypotheses he contemplates, and no one forces
him to lay bets on which hypotheses are right. Indeed the statistical models used
in genetics are not properly called true or false; they are more or less adequate
and there is no sense in betting on their truth.
Edwards tells his colleagues not to bet and so not to be Bayesian. He asks
them to assess the relation between experimental evidence and statistical models
of interest. He will not accept procedures justified merely by their long run
operating characteristics: he wants to know how this particular piece of evidence
bears on these hypotheses offered by this working model. Scientific inference, he
believes, has essentially to do with particular cases. So orthodox statistical
procedures are to be rejected. Edwards is left with likelihood.
He finds it convenient to measure relative likelihoods of hypotheses by taking
natural logarithms, and he calls (relative) log-likelihood the measure of (relative)
support for hypotheses by data. So to compare the support that e furnishes for
h1 against h2, compare the logarithm of the probability of getting e, according to
h1, with the logarithm of the probability of getting e, according to h2. Logarithms
are used because this makes support 'additive' in a certain sense. If we have two
independent pieces of data bearing on some hypotheses, the log-likelihood of the
two items taken together is the sum of the individual log-likelihoods. So although
we cannot combine the support for different hypotheses, we can combine different
pieces of support for the same hypothesis. This, says Edwards, is exactly what we
want in science, for a disjunction of hypotheses is no hypothesis at all. I may
contemplate the hypothesis that the refractive index of a crystal is r, and the
hypothesis that it is r', but there is no scientific hypotheses to the effect that the
refractive index is r-or-r'. Philosophers, however, know to their cost that there is
no good way to distinguish 'real' hypotheses from 'manufactured' ones. One can
think of real cases as well as logicians' tricks. Surely I can contemplate the hypo-
thesis that a certain quantity has a distribution from a specified part of a family
of distributions (normal or log-normal or whatever) without having much idea
about, or even interest in, specific parameter points. Can I not then ask how
well supported this 'composite' hypothesis is?
In the case of simple hypotheses there is still a question about the actual log-
likelihood numbers. Do they mean anything? The orthodox statistician will say
he does not know what to do with them. Probabilities and operating character-
istics let you do all sorts of things. Given some utilities, they allow you to com-
pute expected loss. But what can we do with likelihoods? Edwards's reply is two-
fold. First, he does not want to do any of the things orthodox statisticians can do.
He attaches no sense to a loss function over hypotheses in genetics. He wants to
report evidence and show, in brief form, how it bears on the hypothesis under
examination. All right, but what do Edwards's numbers mean? He has an
unexpected but sensible answer. Use likelihoods and you will find out from
experience what the numbers mean. It takes a while to learn, for example, what
temperatures mean, and it is notoriously hard for Fahrenheiters to take in
weather forecasts in Centigrade, but spend a summer in the South of France
and you get to know what 300C feels like. Numbers that basically record a
ranking take a lot of getting used to. Indeed we could make Edwards's remark
even about probabilities. Shortly before his death L. J. Savage was saying that
we have only just begun to get a grasp of personal probabilities, and it might
take several generations more before they were properly entrenched in our
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
136 Ian Hacking
understanding. Whether or not he is right, it is clear that numerical probabilities
have been radically extending their domain and their intelligibility. A meteoro-
logical classic from the early decades of this century tells the forecaster never to
use the word 'probable' for it would be redundant in a weather forecast, all fore-
casts being merely probable. Nowadays the U.S. Public Weather Office seems
incapable of uttering a sentence not qualified by a probability number. Savage,
incidentally, thought the Weather Office simply did not know what its own prob-
ability percentages meant. Edwards could carry his fight into the enemy camp: if
you use likelihoods in reporting experimental work, you will get to understand
them just as well as you now think you understand probabilities.
Many philosophers will resent this kind of reasoning but I do not find it
intrinsically disturbing. What does worry me is Edwards's faith that a given log-
likelihood ratio will mean the same in any circumstance whatsoever. Grant, for a
moment, that if the likelihood of h1 on e exceeds that of h2, then, lacking other
information, e supports h1 better than h2. Now suppose the actual log-likelihood
ratio between the two hypotheses is r, and suppose this is also the ratio between
two other hypotheses, in a quite different model, with some evidence altogether
unrelated to e. I know of no compelling argument that the ratio r 'means the
same' in these two contexts. Physical probability is much better off. There is
always the frequency interpretation telling us that if the probabilities of two
unrelated events, in different chance set-ups, are the same, then the two events
tend to occur equally often. That, I think, is the chief virtue of the frequency
interpretation: it shows that different probabilities in different set-ups are
commensurable. No non-Bayesian argument shows that likelihood ratios in
different situations are always commensurable, that is, measure the same levels
of evidential significance.
Indeed in artificial cases there seem to be positive counterexamples to un-
restrained use of likelihood. A classic case is the normal distribution and a single
observation. Reluctantly we will grant Edwards that the observation x is the best
supported estimate of the unknown mean. But the hypothesis about the variance,
with highest likelihood, is the assumption that there is no variance, which strikes
us as monstrous. Edwards is a practical reasoner, and is inclined to disregard this
case. If we do wish to fit it into the likelihood scheme of things, we must concede
that as prior information we take for granted that the variance is at least w. But
even this will not do, for the best supported view on the variance is then that it is
exactly w.
For a less artificial example, take the 'tram-car' or 'tank' problem. We capture
enemy tanks at random and note the serial numbers on their engines. We know
the serial numbers start at oooI. We capture a tank number 2176. How many
tanks did the enemy make? On the likelihood analysis, the best supported guess
is: 2176. Now one can defend this remarkable result by saying that it does not
follow that we should estimate the actual number as 2176, only that comparing
individual numbers, 2176 is better supported than any larger figure. My worry
is deeper. Let us compare the relative likelihood of the two hypotheses, 2176
and 3000. Now pass to a situation where we are measuring, say, widths of a
grating, in which error has a normal distribution with known variance; we can
devise data and a pair of hypotheses about the mean which will have the same
log-likelihood ratio. I have no inclination to say that the relative support in the
tank case is 'exactly the same as' that in the normal distribution case, even though
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
Likelihood I37
the likelihood ratios are the same. Hence even on those increasingly rare days
when I will rank hypotheses in order of their likelihoods, I cannot take the actual
log-likelihood number as an objective measure of anything.
Edwards's book has many technical and practical suggestions beyond the
scope of this journal. There is one 'philosophical' novelty to be remarked: the
device of imaginary experiments. This has had some small play in Bayesian
literature, and perhaps goes back to Laplace, but it is newly introduced into
likelihood. Suppose I am thinking about some hypotheses and have not, as yet,
got any new data. Still, I am not indifferent between the hypotheses; I have
background information and prejudices. I realise, perhaps, that I regard h1, in
comparison to h2, as if I had evidence e conferring a log-likelihood ratio r between
these two hypotheses. This, then, is my 'prior support'. When I do get some real
data, I can add the prior support and the actual experimental support to get a
posterior support.
Edwards's very definition of prior support reads oddly. 'The prior support for
one hypothesis against another is S if, prior to any experiment, I support the
one against the other as if I had conducted an experiment .. .' (p. 36). In most of
the book it is data that do the supporting, but here it is me! After I have done a
real experiment, I am supposed to add 'my' prior support to the support fur-
nished by the evidence; that looks dangerously like using arithmetic to add pints
of milk to pounds of apples. Edwards notes that typically prior support counts
little towards posterior support, so this may sound like the Bayesian situation
with prior and posterior probability. But the Bayesian prior is the same kind of
thing as the posterior, and moreover the prior is needed to get a posterior.
Edwards's prior support seems to me a different thing from the support that
evidence furnishes for hypotheses, and the latter kind of support does not
require the former.
Edwards's book is nicely written, has a straightforward development, plenty
of examples and instructive historical asides. It is not a book written for phil-
osophers, but it is a book that those who care about probability ought to read. It
gives a puzzling and possibly fundamental inferential concept a longer run than
anything published before now. Statistical reasoning is not well enough under-
stood by anyone, yet, and we need more fundamental concepts in the arena than
is currently the case. I do not know how Edwards's favoured concept will fare.
The only great thinker who tried it out was Fisher, and he was ambivalent.
Allan Birnbaum and myself are very favourably reported in this book for things
we have said about likelihood, but Birnbaum has given it up and I have become
pretty dubious. George Barnard is the only worker who has consistently and
persistently advocated and advanced a likelihood philosophy. I hope Edwards's
book will encourage others to enter the labyrinth and see where it goes.
IAN HACKING
Cambridge University
This content downloaded from 138.25.78.25 on Sat, 30 Jul 2016 16:40:29 UTC
All use subject to http://about.jstor.org/terms
