Bayes or Bust?

A Critical Examination of Bayesian Conjrmation Theoy

By JOHN EARMAN
MIT, 1992. xvi +
272 pp. $35.00

What Earman means by “Bayesian confirmation theory” has the following

features (pp. 1-34):

(1 ) Evidential confirmation or disconfirmation is a quantitative charac-

teristic.
(2) The fundamental quantity measured is degree of belief.
(3) Degree of belief is regimented according to the principles of the
classical probability calculus.
(4) A bettor’s degree of belief that H i s to be measured by the ratio of his
own stake to the sum of his own stake and his bookie’s.
(5) Conformity of a bettor’s degree of belief to the principles of the
classical probability calculus ensures rationality by excluding the
possibility of a Dutch Book’s being made against the bettor.
(6) A crucial theorem of the calculus (often called ‘Bayes’s theorem’,
though not articulated in these terms by Thomas Bayes himself) is the
equation of the product of the probability of E and the probability of
H o n E with the product of the probability of H and the probability of
E o n H, where H may be understood as the hypothesis at issue and E
the evidence.
( 7 ) Learning from experience is to be modelled as a conditionalisation
whereby, if it is learned for sure that E, and E is the strongest
proposition so learned, then poldand pllCw (representing respectively
degree of belief prior to, and after, acquisition of the new knowledge)
are to be related by the equation puld(H/E)= P , , ~ ~ ( H ) .

While disclaiming the intention to make a comprehensive survey

Earman describes most of the criticisms that have been levelled against
features of Bayesian confirmation theory, and also, with some considerable
reservations, proposes answers to these criticisms. Similarly he suggests
Bayesian solutions for confirmation-theoretical paradoxes like those of
Hempel and Goodman and constructs a Bayesian response to Thomas
Kuhn’s philosophy of science. He gives an overall impression of being
determined to do justice to the complexity of the issues. (For example, he
argues in careful detail for some concessions that Bayesians ought to make in
relation to Putnam’s formal learning theory.) Indeed the book as a whole is
an importantly wide-ranging contribution to the current debate about
Bayesianism and needs to be read by anyone aiming to participate seriously
in that debate.
Nevertheless even Earman is dogmatically intolerant at times. Thus, in
reply to Gruiibaum’s worry that Bayesian methods are not helpful in
evaluating causal hypotheses, like the hypothesis that drinking coffee cures
colds, Earman insists that a prelude to successful hypothesis-testing is a
precise statement of the hypothesis in non-causal language, because he is “of
the conviction that causal talk is a mare’s nest of confusions, snares and

45

(0 Basil Blac kwell Ltd. 1994

delusions” (p. 104). What is perhaps meant, he says, by the claim that
consumption of coffee cures colds is simply that the rate of remission among
people who drink coffee is higher than among those who don’t, other things
being constant. But well-known difficulties about counterfactuals need to be
faced down if we are to accept a Humean regularity-type analysis of
causation. Nor does Earman himself always avoid implicit use of causal talk,
as when he includes in his definition of Bayesian confirmation-theory the
requirement that conditionalisation relies on its having been “learned for
sure” (p. 34) that E. How can it be ‘‘learned for sure” that E unless by a
process through which something appropriate causes the learner to feeljustifiably
certain that E? If the learner’s feeling of certainty that E has been wholly or
partly caused by drugs, say, then he has not learned for sure that E.
Earman also appears to overlook altogether a very serious problem about
Bayesian conditionalisation. If p , , e w ( yis to be put equal to pold(H/E),it is
necessary - according to Earman - not only that E has been learned for sure,
but also that E is the “strongest” (p. 34) proposition so learned. This
requirement carries an echo of Carnap’s “requirement of total evidence”,
whereby a proposition of the form ‘c(h,e ) = r’ is applicable to a given
knowledge situation only if e states the total available evidence (Logical
Foundations of Probability, 1950, p. 21 I ) . But such a requirement leads to well-
known difficulties. If ‘available’ means ‘logically possible to obtain’, then no
conditionalisation can ever be legitimate, because no-one can ever learn
everything that it is logically possible to learn: correspondingly, if Earman’s
‘strongest’ means ‘containing more information than any other logically
possible proposition’, this too is not a practically fulfillable condition. But if
Carnap’s ‘available’ were to mean ‘practically possible to obtain’, and
Earman’s ‘strongest’ were to mean ‘containing the most information that it is
practicable to obtain’, then no conditionalisation would ever take place,
because of familiar problems about induction: no-one can know when it is
safe to suppose that it is in practice impossible to obtain any more evidence.
And if Carnap’s ‘available’ means ‘already learned and recorded’, while
Earman’s ’strongest’ means ‘containing a complete statement of existing
information’, then E may state far too little evidence for it to be safe to
conditionalise: for example, the patients who were coffee-drinkers may all
have also been tea-drinkers.
It looks as though the underlying error here is to suppose that conditionali-
sation is a matter of all-or-nothing, rather than of degree. That is to say, the
trouble seems to arise from Earman’s use of the adjective ‘strong’ in the
superlative rather than the comparative mode (and from Carnap’s reference
to the ‘total’ evidence). If instead conditionalisation is made a matter of
degree, then what Earman calls ‘strength’ is best understood to be the same
as what Keynes called “weight” ( A Treatise on Probability, 1921, p. 7 1 ) .
Weight, in this sense, does not conform to the mathematics of the classical
calculus ofprobability, but has a distinct logical structure ofits own which, as
I have argued elsewhere (in ‘Twelve Questions about Keynes’s Concept of
Weight’, The British Journal for the Philosophy of Science, vol. xxxvii, 1985, pp.
263-278), is closely linked with the structure of Baconian reasoning about
uniformities rooted in the causal network. So, if Earman had not been so
unwilling to attach any value to discourse about causality, he might have

46

0Basil Blarkwd L R . 1994

been more likely to recognise what is implied by his reference to the ‘strength’
of the evidential proposition on which a Bayesian conditionalisation is based.
After all, this would not involve abandoning the use of probabilistic
conditionalisation in the assessment of confirmation but rather a fuller
appreciation and understanding of the conditions under which such
conditionalisation is applicable.
There is another point about which Earman does not develop his analysis
satisfactorily. O n the one hand (pp. 203-205), he speaks slightingly of
“methodologies of science” and of philosophers who develop such methodo-
logies (such as Popper, Kuhn, Lakatos and Laudan). O n his view,
apparently, Bayesianism is not a methodology, but a thesis about validity in
scientific inference: all criteria of valid non-deductive inference derive from
the probability axioms and the rule ofconditionalisation. What Bayesianism
provides is rationality constraints on degrees of belief (p. 161). On the other
hand, Earman writes (p. 63) of “a Bayesian rationale for what are regarded
as sound methodological procedures”, such as a Bayesian explanation of the
confirmation-theoretic virtue of variety of evidence. So what is wrong with
speaking of any methodology as being Bayesian if it satisfies Bayesian criteria
of rationality? There seems to be no relevant antithesis between methodolo-
gical and non-methodological accounts of confirmation.
THE QUEEN’S COLLEGE, OXFORD L. J. COHEN

PHILOSOPHY OF LANGUAGE

T a l k About Beliefs
By MARK CRIMMINS
M I T Press, 1992. xiv + 214 pp. $25.00
A good philosophy book has an easily-fixed position in a current debate, and
uses this to make points or introduce devices which can be used by
philosophers inclined to different positions in the debate. This book satisfies
this criterion. The debate is about neo-extensionalist theories of belief.
Crimmins’ aim is to disagree with them. Belief for him is not just a relation
between a person and the things referred to in a sentence. It is in fact
(something like) a relation between a person and a proposition. But
Crimmins’ positive theory is meant to capture many of the insights of
extensionalist theories. The propositions in question are extensional entities.
Yet intensionality is preserved, to the extent that the truth values of
ascriptions of belief differing only in the presence of co-extensive terms in the
ascribing sentence may be different. The device that reconciles these at first
sight incompatible elements is that of an unarticulated constituent of a
statement. Following the outline in Crimmins and Perry’s 1989 Journal of
Philosophy article, it takes a belief ascription to refer implicitly to both the
objects named in the that-clause and various other things which are crucial
to the identity of the belief and its semantic properties.
The full theory therefore has to provide an account of propositions as
structures of complex objects and of the full content of belief-ascriptions,

47

0Basil Blackwell I d 1994