Laurence Jonathan Cohen - An Introduction To The Philosophy of Induction and Probability (1989, Oxford University Press) PDF

Download as pdf or txt
Download as pdf or txt
You are on page 1of 227

AN I N T R O D U C T I O N TO

THE PHILOSOPHY OF INDUCTION AND


PROBABILITY
AN I N T R O D U C T I O N TO
The Philosophy of Induction
and Probability
L. J O N A T H A N COHEN

CLARENDON PRESS OXFORD


1989
Oxford University Press, Walton Street, Oxford o x 2 6DP
Oxford New York Toronto
Delhi Bombay Calcutta Madras Karachi
Petaling Jaya Singapore Hong Kong Tokyo
Nairobi Dar es Salaam Cape Town
Melbourne Auckland
and associated companies in
Berlin Ibadan
Oxford is a trade mark of Oxford University Press
Published in the United States
by Oxford University Press, New York
© L. Jonathan Cohen 1989
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise, without
the prior permission of Oxford University Press
This book is sold subject to the condition that it shall not, by way
of trade or otherwise, be lent, re-sold, hired out or otherwise circulated
without the publisher's prior consent in any form of binding or cover
other than that in which it is published and without a similar condition
including this condition being imposed on the subsequent purchaser
British Library Cataloguing in Publication Data
Cohen, L. Jonathan (Laurence Jonathan)
An introduction to the philosophy of induction and probability.
1. Probabilities - Philosophical perspectives
2. Inductive inference - Philosophical perspectives
I. Title
121'.63
ISBN 0-19-875079-X
ISBN 0-19-875078-1 Pbk
Library of Congress Cataloging in Publication Data
Cohen, L. Jonathan (Laurence Jonathan)
An introduction to the philosophy of induction and probability/L. Jonathan Cohen.
1. Induction (Logic) 2. Probabilities. I. Title
BC91.C6 1988 161—del9 88-21674
ISBN 0-19-875079-X
ISBN 0-19-875078-1 (pbk.)

Set by Colset Private Limited, Singapore


Printed in Great Britain
at the University Printing House, Oxford
by David Stanford
Printer to the University
'Nostra autem ratio . . ea . . . est ut certitudinis
gradus constituamus.'
(Francis Bacon, Novum Organum, preface)
But our method is such that we establish degrees of certainty.
'Probabilitas enim est gradus certitudinis'.
(James Bernoulli, Ars Conjectandi, pt. 4, ch. 1)
For probability is degree of certainty.
PREFACE

The purpose of this book is to introduce the reader to


philosophical issues and arguments about probability and
induction, not to survey the past and present range of those
issues and arguments. The text is therefore highly selective both
in the topics and authors that it discusses and also in the points
that it makes about them. It seeks to engage the reader's interest
by presenting a coherent and readily intelligible picture of the
field, with due regard to the depth and difficulty of the funda-
mental philosophical issues, rather than to aim at a more
comprehensive and encyclopaedic type of treatment (which
might also be more superficial). And it economizes as much as
possible in the use of mathematical symbolism, the statement of
mathematical results and the derivation of statistical algorithms,
because its concern is with the philosophical issues rather than
with the mathematical ones.
In trying to show how strong or how weak a particular
philosophical theory really is I have not hesitated to include new
arguments where they seem to be needed, and to exclude
familiar ones where they seem to be of little value. No doubt
others might wish to distribute their emphases differently. But
that is inevitable in philosophy. In any case my overall purpose
is not to make converts, but to excite interest by the exploration
and comparison of alternative options. For those who then wish
to go further the footnotes include relevant bibliographical
suggestions.
I have adopted a broadly historical approach in the first
chapter, in order to show how the central problems developed.
The other five chapters seek to impose a clarificatory structure
on the consideration of those problems. In order to assist under-
standing, each section is preceded by a summary of its contents.
But, as newcomers to philosophy may need to be reminded, the
nature of philosophical argument makes it impossible for
reading such summaries to be a worthwhile substitute for read-
ing the text. Also, for the benefit of those new to the subject, I
have included in the footnotes some brief biographical notes in
Vlll Preface
order to help locate historically important writers on probability
within their cultural contexts.
In order to avoid unnecessary quotation-marks I have
normally left it to the context to determine whether symbolic
expressions are being used or mentioned. Also, unless there are
contextual reasons to suppose otherwise, the pronoun 'he' is to
be understood in the text as meaning 'he or she', the pronoun
'him' as meaning 'him or her' and the pronoun 'his' as meaning
'his or her'.
The origins of this book are in a course of lectures that I have
given for some years within the Sub-Faculty of Philosophy at
Oxford. I owe much to the questions and criticisms that have
been voiced at those lectures. But I am most indebted to
Jonathan Adler, Menachem Fisch, Greg Hunt, Jim Logue, and
Dave Schum, who read an earlier draft of the text and raised
many important points about it. None of them will be satisfied
with everything that is in the final version, and any mistakes in it
are mine, not theirs. But I have profited enormously from the
time and care which each of them devoted to reading the earlier
draft, and am deeply grateful to them all.
I am also very grateful to the Press's copy-editor, who has
saved me from many infelicities of style, and to Pat Lloyd, who
has contributed much by the care and attention with which she
has typed and processed so many pages so many times.
A portion of Section 12 was published previously in Evandro
Agazzi (ed.), Probability in the Sciences (Kluwer: 1988), 69-77.
L.J.C.
The Queen's College, Oxford
21 March 1988
CONTENTS

I. THE ORIGINS OF THE PROBLEM 1


1. An outline of the issues 1
2. The Baconian tradition in the philosophy of
induction 4
3. The rise of Pascalian probability 13
4. The combination of Baconian and
Pascalian themes 27
II. THE C O N T R O V E R S Y ABOUT THE NATURE
OF PASCALIAN PROBABILITY 40
5. Some general considerations 40
6. Indifference theories 43
7. Frequency theories 47
8. Propensity theories 53
9. Personalist theories 58
10. Multi-valued logic theories 70
11. Logical relation theories 74
III. THE FOUNDATIONS OF PLURALISM IN
THE ANALYSIS OF PROBABILITY 81
12. Some logical distinctions exploited by
differing analyses of Pascalian probability 81
13. The appropriateness of different
conceptions of Pascalian probability to
different purposes 90
14. The need to supplement Pascalian
judgements by non-Pascalian ones 94
15. How are different conceptions of
probability possible? 109
IV T H E PASCALIAN GRADATION OF
AMPLIATIVE INDUCTION 115
16. Inductive probability under a realist
construal 115
17. Inductive probability under a range-
theoretical construal 120
X Contents
18. Pascalian gradation for variative induction 130
19. Inductive probability under a personalist
construal 136
V T H E BACONIAN GRADATION OF
AMPLIATIVE I N D U C T I O N 145
20. Inductive support by the Method of
Relevant Variables 145
21. The logical syntax of the Method of
Relevant Variables 157
22. Some non-standard interpretations of
Baconian logical syntax 164
VI. FOUR PARADOXES ABOUT INDUCTION 176
23. The classical problem of induction 176
24. The paradox of the ravens 188
25. The 'grue' paradox 197
26. The lottery paradox 203
Index 213
I
The Origins of the Problem

§ 1. AN O U T L I N E O F T H E I S S U E S

Induction is termed 'ampliative' when it extrapolates beyond the


existing data. Ampliative induction may concern hypotheses
about what ought to happen as well as hypotheses about what
does happen. One main question about it concerns whether the
variety or the multiplicity of evidential instances is crucial in the
evaluation of inductive support. Another concerns whether
inductive support is a kind of probability that conforms to the
laws of the calculus of chance.
In contemporary philosophical dialogue about induction one
question is central. What kind of support can the results of
1

scientific experiments give to hypotheses about laws of nature or


about uniformities of causal connection? What kind of
knowledge can we have that energy and mass are related by the
equation e = mc or that appropriate doses of penicillin normally
2

cure septicaemia?
Such hypotheses always generate predictions that extrapolate
beyond the existing data. That is why they may be useful as a
basis for building another rocket, curing new cases of septi-
caemia, producing more food, etc. So induction about them is
conveniently labelled 'ampliative'. It involves reasoning to
conclusions that cover not only the evidential instances but
others also. It amplifies our knowledge. And it contrasts with the

1 Modern conceptions of the problem-area may be traced through H.W.B. Joseph,


An Introduction to Logic, 2nd edn. (Oxford: Clarendon Press, 1916), 378-565; W.E.
Johnson, Logic (Cambridge: Cambridge University Press), Pt. II (1922), 189-253, and
Pt. Ill (1924); M . R . Cohen and E. Nagel, An Introduction to Logic and Scientific Method,
(London: Routledge and Kegan Paul, 1934), 151-72, 197-288, 302-51; W. Kneale,
Probability and Induction (Oxford: Clarendon Press, 1949); H. E. Kyburg, Jun., Probability
and Inductive Logic (London: Macmillan, 1970—contains a good bibliography); and R,
Swinburne, An Introduction to Confirmation Theory (London: Methuen, 1973).
2 The Origins of the Problem§3
'summative' induction that, rather unproblematically, estab-
lishes a generalization on the basis of what are known to be all its
instances, as when a railway inspector, passing down the whole
train, establishes that every passenger on the train has a ticket.
Nor is all ampliative induction concerned with laws or causes.
Sometimes purely descriptive generalizations are at issue, such
as 'All crows are black' or 'All emeralds are green'. And
sometimes inductive reasoning operates in relation to generali-
zations that are not about what in fact always happens in nature
but rather about what ought properly to happen in human
society (though it may not always do so), as when in Anglo-
American jurisprudence a general statement of the law about a
particular kind of issue is supported by argument from certain
past judicial decisions within the pertinent jurisdiction. Perhaps
it will be stated that a breathalyser result ought to be accepted as
legitimate evidence of what a motorist's blood-alcohol content
was two hours previously because in one such case it was treated
as legitimate by an appropriate court.
It follows that the question about inductive support needs to
be posed more abstractly, so as to transcend such differences
of subject-matter. The fundamental problem is: how does
ampliative induction operate? And this problem raises two main
issues which combine with one another to form the framework of
discussion.
First, do generalizations derive more support from the variety
of instances that are cited as evidence for them or from the multi-
plicity of those instances? We shall see, for example, that Francis
Bacon, who founded the modern study of induction, preferred
instantial variety, while Rudolf Carnap, who has been the most
influential writer on the subject in the twentieth century,
attached primary importance to instantial multiplicity. The
former mode of reasoning is known as 'variative' or, for reasons
which will soon emerge, 'eliminative' induction, the latter as
'enumerative' induction.
Secondly, what is the relation between inductive support and
the kind of probability that conforms to the theorems of the
calculus of chance as developed originally by the seventeenth-
century mathematicians Blaise Pascal and Pierre Fermat?
Because Pascalian probabilities are measured by numbers (on
a scale that runs from 0 to 1) they hold out the prospect of
3 The Origins of the Problem§3 §3

supplying a mode of gradation for inductive support. Such


support can apparently be conceived in Pascalian terms as
existing at various levels on an ascending scale of rationality
rather than as existing only at zero level or at the level of
conclusive proof. But there have been many disputes about how
this Pascalian analysis of induction is to be carried out and not
everyone even agrees that Pascalian probability is the right basis
for inductive gradation.
An introductory account of the subject can usefully begin with
a historical sketch of how it has arrived at its present state. We
shall see (§§ 2-3) how the Baconian and Pascalian traditions
developed largely in independence of one another for about two
centuries, until (§ 4)J.S. Mill began to apply Pascalian modes
of gradation to Baconian induction. It was then possible for
other philosophers, like J . M . Keynes, to develop variations on
this theme, because more than one proposal had emerged for
measuring probability in a way that conformed to Pascalian
principles. But, whereas Keynes still stressed the importance of
instantial variety in inductive evidence, as Bacon and Mill had
done, more recently Carnap and others have proposed
Pascalian analyses that would primarily display the value of
instantial multiplicity. And yet a third Pascalian position is
possible. The most recently popular mode of Pascalian analysis
(called 'Bayesianism'), by concerning itself with the subjective
gradation of an inductive reasoner's belief, however derived,
rather than with the objective gradation of evidential strength, is
able to avoid taking sides on the issue of instantial variety versus
instantial multiplicity.
Faced with this rich diversity of philosophical positions we
need first to examine how philosophical controversy about the
nature of Pascalian probability may best be resolved (§ § 5-15).
Only when that has been done shall we be able to explore the
extent to which ampliative induction submits to a Pascalian
analysis (§§ 16-19) and the extent to which non-Pascalian prin-
ciples apply (§ § 20-22). And we shall then also have to allow
(§§ 23-26) for the impact of several important puzzles and
paradoxes that have been propounded as problems about the
rationality of induction. For example, David Hume pointed out
that, if a generalization goes beyond the observed evidence, then
even though this evidence is uniformly favourable and
4 The Origins of the Problem§3
enormously extensive, the logical possibility remains that the
generalization may be false. So for Hume in the eighteenth
century, as for Karl Popper in the twentieth, ampliative
induction can never be a rational procedure. Perhaps you may
think instead that ampliative induction can be rational just so
long as it conforms to appropriate principles of its own, whether
Baconian or Pascalian: all that is needed is to uncover the laws of
this 'inductive logic'. But in certain important respects any such
principles are difficult to state plausibly and coherently. For
example, as Carl Hempel pointed out, if any favourable
instance gives a generalization some confirmation, however
small, and if anything that confirms a proposition must also
confirm any logically equivalent proposition, then—para-
doxically—even a white handkerchief gives some confirmation
to the generalization 'All ravens are black', because it is a
favourable instance for the equivalent generalization 'All non-
black things are non-ravens'. These problems are very serious
ones. But a satisfactory philosophy of induction ought to resolve
them all. So the puzzles and paradoxes discussed in § § 23-26
may be seen as a test for the adequacy of the analyses and evalu-
ations propounded in § § 16-22.

§ 2. T H E B A C O N I A N T R A D I T I O N IN T H E P H I L O S O P H Y O F
INDUCTION

Bacon had a gradualist conception of inductive enquiry, both in


science and in jurisprudence. This gradualism was echoed by
Herschel and Whewell. Bacon and Whewell also emphasized the
role of conceptual innovation in the progress of inductive
enquiry, though Mill disagreed. Bacon did not however
distinguish, as Mill did, between induction as a method of
discovery and induction as a pattern of proof or graded
justification.
The idea that ampliative induction may issue in judgements of
comparison or gradation is one that may be traced back at least
as far as Francis Bacon's Novum Organum (published in 1620),
though it is hard to find a clear and systematic development of it
in the work of any ancient or medieval philosopher. Indeed
Bacon's gradualist epistemology was very highly regarded by
5 The Origins of the Problem§3 §3

many leading experimental scientists in the seventeenth


century, like Robert Hooke and Robert Boyle. But he has not 2

been well served by his modern interpreters. All too often they
have echoed the misrepresentation that was promoted by R.L.
Ellis in the latter half of the nineteenth century. They have held
that 'absolute certainty is . . . one of the distinguishing
characters of the Baconian induction'. Yet what Bacon wrote in
3

the preface of his Novum Organum was that his principle of


method, 'though hard to practise, is easy to state: it is such that
we set up degrees of certainty . . .'. Where degrees of certainty
are said to be admissible, it is clearly wrong to impute insistence
at every stage on absolute certainty, even if that were ultimately
what we seek.
Bacon advocated that scientists interrogate Nature in order to
be able to tabulate both the various circumstances in which
instances of the phenomenon (the 'nature') under investigation
have been found to be present and also the circumstances under
which they have been found to be absent. Thus he found heat
present in the sun's rays, in flame, and in boiling liquids, but
absent in the moon's and stars' rays, in phosphorescence, and
in natural liquids. To the extent that scientists discover a
4

circumstance which correlates uniquely with the pheno-


menon—i.e. is always present when it is present and absent when
it is absent—they have discovered its proximate explanation (or
'form') and have acquired power to reproduce it at will. Thus he
says that what constitutes finding the form for a particular
nature is that 'another nature should be discovered which is
convertible with the given nature and yet is a limitation of a

2 e.g. The Posthumous Works of Robert Hooke, (London: R. Waller, 1705), 6, and R.
Boyle, Works (London: A Millar, 1744), i. 199. Francis Bacon (1561-1626) was by
training a lawyer and was Lord Chancellor for a few years under James I. His philosophy
of science was concerned with the social as well as the intellectual conditions for the
advancement of knowledge and with the social as well as the intellectual benefits of
properly directed research. In jurisprudence he sought a systematic statement of the
principles of Common Law from the study of reported cases.
3 R.L. Ellis, 'General Preface to Bacon's Philosophical Works', in The Works of
Francis Bacon, ed.J. Spedding, R. Ellis, a n d D . N . Heath (London: Longmans, 1859), i.
23.
4 For detailed references to Bacon's writings on this subject see L. Jonathan Cohen,
'Some Historical Remarks on the Baconian Conception of Probability', Journal of the
History of Ideas, 61 (1980), 219-31.
6 The Origins of the Problem§3
more intelligible nature, like a true natural kind', and this more
fundamental nature may itself have a form that investigators can
seek to discover. For example, on Bacon's view the nature of
which heat is a limitation is motion: 'heat is a motion expansive,
restrained, and acting in its strife upon the smaller particles of
bodies'. And motion itself must be supposed to be a limitation of
a more fundamental property. So there is a hierarchy of
explanatory laws to be discovered. The investigator should
expect to make a gradual ascent to more and more compre-
hensive laws, acquiring greater and greater certainty as he
moves up the pyramid, and each law that is reached should lead
him to new experiments, that is, to experiments over and above
those that led to the discovery of the law. But in regard to the
apex of the pyramid, at which discovery of the 'summary law of
Nature' would generate absolute certainty, Bacon says that we
do not know whether this is within human reach. (Contrast 5

Ellis's misinterpretation.)
Bacon held that at each stage of the investigation greater
significance should be attached to some kinds of observable
instances than to others. Thus there are twenty-seven types of
'prerogative' instances that deserve preferential consideration,
and some of these are more valuable than others in terms of their
ability to exclude alternative explanations and thus contribute to
greater certainty. But Bacon repudiated as 'childish' the method
of induction by simple enumeration, whereby a generalization
that is as yet unfalsified is supposed to acquire support that
varies in strength with the number of known instances that
verify it. O n his view it is not the mere number of known
instances that counts but the variety of circumstances that have
been canvassed in the attempt to exclude alternative
explanations. And this is because there is only a limited number
of ultimate forms and so falsificatory evidence, by conclusively
excluding incorrect hypotheses, permits firmer progress than
verificatory evidence does towards identifying the correct
6

hypothesis. Even a well-verified hypothesis may still turn out to


be false, but the truth-value of a falsified hypothesis is known
5Of the Advancement of Learning (Works, iii. 356). Contrast, however, Bacon's more
optimistic claim in his Parasceve (Works, iv. 252).
6Novum Organum, bk. I, aphorism xlvi (Works, i. 166).
17 The Origins of the Problem§3 §3

once and for all. Moreover, on Bacon's view, at each level of 7

investigation the kinds of circumstances that vary independently


of one another are few enough to be examined, and thereby we
may eventually come to hope that every false hypothesis has
been eliminated, with an appropriate degree of certainty estab-
lished for the remaining hypothesis. Eliminative induction is
promoted by successive variations of experimental circum-
stances because, for example, if you test a hypothesis 'All A are
B' unsuccessfully in circumstances C, you have thereby elimi-
nated not only the hypothesis 'All A are B' but also the hypo-
thesis 'All AC are B', while perhaps leaving open as yet the
hypothesis 'All AD are B'.
Bacon's gradualist methodology reflects a theme that came
to dominate seventeenth-century English attitudes towards
knowledge of Nature. This theme was that scientific
achievement—of the kind that brings power over Nature—is a
matter for gradual progress in the future through open-minded,
systematic, first-hand investigation. It is not—as implicitly
categorized in much medieval and Renaissance writing—some-
thing that has already been perfected from observation in the
past and can largely be recaptured second-hand from the texts of
appropriately empiricist authorities (Aristotle, Galen, etc.).
Nor is it something to be deduced from self-evident first
principles in one glorious exercise of logical or mathematical
skill, as Descartes suggested. Moreover, one could not adopt
8

Bacon's gradualism without rejecting the view that the truth of a


scientific theory had to be absolutely certain if it were to be of
any importance at all. O n Bacon's view worthwhile results could
be achieved at many stages on the way to certainty. So Bacon's
inductivist epistemology cohered with the increasingly
prevalent assumption that different levels of credibility were
possible not only in regard to scientific contentions but also in
regard to testimony about facts of religious or forensic
importance. His philosophy was thus part of a general
movement of thought that was coming to provide an intellectual

1Valerius Terminus, ch. 13 (Works, iii. 243).


8Discourse on Method, pt. V, in The Philosophical Works of Descartes, trans. E.S.
Haldane and G . R . T . Ross (Cambridge: Cambridge University Press, 1931) i. 106-18.
8 The Origins of the Problem§3
basis for religious tolerance and for more practicable standards
of proof in the courts, as well as for scientific gradualism. But it 9

should be noted that within this movement the scale of proof on a


particular issue ran downwards from the justification of
certainty that the proposition in question was true to the absence
of any justification for belief in its truth: the lower limit of proof
was not, as a Pascalian conception of probability requires, the
presence of justification for belief in the proposition's falsity. 10

Bacon insisted that his methodology is just as applicable to


normative as to factual issues. It has a role in jurisprudence, for
example, as well as in natural science because in Common Law,
as Bacon pointed out, legal maxims, just like the axioms of
Nature in science, are reached by induction from individual
cases and then, once formulated, are applied back to determine
new particulars. Bacon was therefore as emphatic on the value of
good legal reports for jurisprudential induction as on that of
good natural history for scientific induction. By the former we
reduce uncertainty about our legal rights and duties; by the
latter we reduce uncertainty about what is the case in Nature.
And negative instances, he held, are of primary importance in
both enquiries, in order to eliminate false propositions. So too 11

in the establishment of ethical principles from moral judgements


in particular cases. Bacon was thus able to claim that the
generality of his inductive logic paralleled that of syllogistic
logic. Both of them embraced not only propositions about
12

actual facts but also propositions about what ought to happen.


Bacon's philosophy of science retained much of its reputation
into the earlier half of the nineteenth century. J . F. W. Herschel,
for example, proclaimed his Baconian allegiance by placing a
picture of Bacon's bust on the very title page of his Preliminary

See B.J. Shapiro, Probability and Certainty in Seventeenth-Century England: A Study of the
9

Relationships between Natural Science, Religion, History, Law and Literature (Princeton:
Princeton University Press, 1983). Cf. S. Shaffer's review of this book, "Making
Certain', Social Studies of Science, 14 (1984), 137-52, which is more critical in tone than in
substance. Bacon even allowed the possibility that his own methodology was capable of
improvement: Novum Organum, bk. I, aphorism cxxx {Works, i. 223).
Shapiro, Probability and Certainty, p. 33. On this question of the lower limit of proof
1 0

see further pp. 112-13 below.


11See P . H . Kocher, 'Francis Bacon on the Science of Jurisprudence', Journal of the
History of Ideas, 18 (1957), 3-26.
12Novum Organum, bk. I, aphorism cxxvii (Works, i. 220).
9 The Origins of the Problem§3 §3

Discourse on the Study ofNatural Philosophy (first published in 1831),


and his text is quite explicit about the gradualist implications of
Baconian inductivism. Experiments, he wrote, 'become more
valuable, and their results clearer, in proportion as they possess
this quality (of agreeing exactly in all their circumstances but
one), since the question put to nature becomes thereby more
pointed and its answer more decisive'. That is, as wc become
13

better able to distinguish the different variations of circumstance


that are possible in a given type of experimental situation, we
can have greater confidence about which of these variations is in
fact responsible for the phenomenon under investigation.
William Whewell republished the methodological part of his
Philosophy of the Inductive Sciences, Founded upon their History (first
published in 1840) under the title Novum Organum Renovatum, and
was another admirer of Bacon's gradualism. 'By far the most
extraordinary parts of Bacon's works are those in which, with
extreme earnestness and clearness, he insists upon a graduated
and successive induction as opposed to a hasty transit from special
facts to the highest generalizations.' And Whewell's own
14

reflections led him to add that the most important kind of step in
this progress to a higher level of generalization was what he
called 'consilience'. Consilience occurs when inductions from
altogether different kinds of facts are, quite unexpectedly, seen
to share a common explanation. Thus Newton's theory of uni-
versal gravitation, 'which had been inferred from the Perturba-
tions of the moon and planets by the sun and by each other, also
accounted for the fact, apparently altogether dissimilar and
remote, of the Precession of the equinoxes' , In such a case, wrote
15

Whewell, when the evidence in favour of our induction enables


us to explain and determine facts of a kind different from those
which were contemplated in the formation of our hypothesis,

13 A Preliminary Discourse on the Study of Natural Philosophy (London: Longman, Rees,


1833), 155. John Frederick William Herschel (1792-1871) made important
contributions to astronomy and chemistry, and wrote about the philosophy of science on
the basis of his own experience as a scientist.
1 4 The Philosophy of the Inductive Sciences (London: J . W . Parker, 1847), ii. 232
(WheweH's italics). William Whewell (1794-1866) wrote a history of science in
preparation for his philosophy of the subject. He worked also in mineralogy, and was
influential in reforming the teaching of mathematics at Cambridge University.
15 Ibid. 66 (WheweH's italics).
10 The Origins of the Problem§3
'this evidence is of a much higher and more forcible character'
than when the facts explained are of the same kind as those
initially observed. Or, as Bacon would have put it, our
16

certainty increases as we progress up the pyramid of laws,


embracing more and more 'natures' within the scope of our
explanations. And Whewell was able to illustrate his epistemo-
logy with intuitively convincing examples drawn from the
history of science since Bacon's day, whereas most of Bacon's
own examples were necessarily speculative because his philo-
sophy of science, formulated as it was in the early seventeenth
century, had to be largely prospective and programmatic rather
than retrospectively analytical.
Whewell saw also that Bacon had recognized, as Aristotle had
not, how important it was in inductive reasoning to operate
17

with the right concepts as well as to perform the right


experiments. Indeed, Whewell himself thought that in every
18

new inductive inference a new conception is invented and


'superinduced' upon the facts. For example, according to
19

Whewell, when Kepler discovered the laws of planetery motion


he bound together particular observations of planets' positions
by the conception of an ellipse 'which was supplied by his own
mind'. And on this point Whewell's quasi-Kantian idealism
20

has naturally been opposed by realists like J . S. Mill (writing in


the Lockean tradition) who insist that 'the ellipse was in the facts
before Kepler recognised it'. 21

But the issue is a complex one. Even if we set aside the


metaphysical dispute between realism and idealism, we can still
distinguish two rather different methodological claims, of
unequal merit, which might be taken to be implicit in Whewell's
remarks. First, there is the claim that conceptual inno-
vation—the invention of a new concept—has always been, and
1 6 Whewell, The Philosophy of the Inductive Sciences, 65.
17 Ibid. 50. 1 8 Ibid. 237-8. 1 9Ibid. 50-1.
20 w Whewell, Of Induction, with especial reference to Mr J. Stuart Mill's System of Logic
(London: J. W. Parker, 1849), 28.
21 J.S. Mill, A System of Logic Ratiocinative and Inductive (London: Longmans, Green,
1896, 193. John Stuart Mill (1806-73) sought to found his theory of induction on the
practice of the natural sciences and then to apply it within his philosophy of the social
sciences. He argued that the course of human history is subject to the operation of
general laws, knowledge of which would assist in furthering the central objective of utili-
tarian ethics—human happiness.
11 The Origins of the Problem§3 §3

is always likely to be, central to any major theoretical advance in


natural science. Whewell produced plenty of historical data that
illustrate this thesis, and Mill produced none that refute it. It is
indeed integral to any understanding of what Whewell called
'consilience', as we shall see later (§ 20). (The importance of
conceptual innovation for innovative inductive reasoning has
also been noticeable recently in knowledge engineering,
22

whereby particular kinds of expertise are extracted from human


practice and articulated in so-called expert systems for
implementation on appropriate computers.) Secondly, there is
the claim that conceptual innovation in this sense is necessary
for any kind of inductive reasoning whatever, however humble.
But such a claim would draw too tight a boundary around our
notion of induction. Too much of the quite respectable causal
reasoning that Mill discusses would then have to be rejected
because it contained no conceptual innovation.
One other preliminary issue about induction needs some
discussion: is it to be understood as a process in time? Bacon
remarked that knowledge is not communicated in the same
order as that in which it was discovered. But he drew no sharp
23

distinction between the concept of induction as a process of


research or gradual discovery and the concept of induction as a
pattern of proof or of graded justification, though the former is
concerned with temporal sequences of events and the latter with
timeless relations between propositions. Mill had approached
this distinction by 1851, when he claimed, in the third edition of
his System of Logic, that even if his inductive methods were not
methods of discovery they were at least methods of proof. And 24

since Mill's day most writers on induction have concentrated on


the logic of inductive proof rather than the methodology of
inductive discovery. (Perhaps the only systematic exceptions to
this tendency today are computer scientists writing about the
construction of expert systems.) The present book, as an
introduction to the contemporary state of the subject, will

2 2See e.g. N . M . Cooke, 'Modelling Human Expertise in Expert Systems', New


Mexico State University Memoranda in Computer and Cognitive Science, 12 (1985).
2 3Valerius Terminus, ch. 18 {Works, iii. 248).
68 System of Logic, p.25569Ibid.214.7 Ibid.293.
0
12 The Origins of the Problem§3
necessarily follow the dominant trend. It will therefore be 25

mainly concerned with induction as a pattern of proof or graded


justification. And in any case we need to have an adequate
criterion of what is to count as well-supported before we can
profitably begin to discuss how to discover well-supported
generalizations.
Bacon has in fact often been criticized for urging scientists to
collect facts before they form hypotheses, rather than vice versa.
But we need not concern ourselves at all with this issue of
temporal priority, if our concern is with the logic, not the heuri-
stics, of induction. In our sense of the term induction is not in
any kind of competition with the so-called hypothetico-
deductive method, for instance, whereby a generalization is first
propounded and then evaluated by deducing its consequences in
a particular situation and checking whether they obtain.
Rather, the hypothetico-deductive method may be said to be a
heuristic that exploits our beliefs about the relationships of
inductive support which obtain between some propositions and
not between others, since it is those beliefs that justify the
direction of enquiry which the method takes on any particular
occasion of its employment. Certain consequences of a general-
ization are checked because the results of such checks reveal the
extent of inductive support that exists for the generalization.

2 5 Even a genetic epistemology (see R.F. Kitchener, 'Is Genetic Epistemology


Possible?', British Journal for the Philosophy of Science, 38 (1987), 283-99) needs to postulate
criteria for evaluating any particular stage of cognitive development. More detailed
expositions and evaluations of Bacon's and Whewell's views are to be found in M.
Hesse, 'Francis Bacon', in D.J. O'Connor (ed.), A Critical History of Western Philosophy
(New York: Free Press ofGlencoe, 1964), 141-52; L. Jardine, Francis Bacon: Discovery and
the Art of Discourse (Cambridge: Cambridge University Press, 1974); P, Urbach, Francis
Bacon's Philosophy ofScience (La Salle: Open Court, 1987); G. Buchdahl, 'Inductive Versus
Deductivist Approaches in the Philosophy of Science as Illustrated by some Contro-
versies between Whewell and Mill', TheMonist, 55 (1971) 343-67; M. Fisch, 'Whewell's
Consilience of Inductions—An Evaluation', Philosophy ofScience, 52 (1985), 239-55, and
'Necessary and Contingent Truth in William Whewell's Antithetical Theory of
Knowledge', Studies in History and Philosophy ofScience, 16 (1985), 275-314; and the papers
in M. Fisch and S. Schaffer (eds.), William Whewell: A Composite Portrait—Studies of his
Life, Work and Influence (Oxford: Clarendon Press (forthcoming) ). For some medieval
anticipations see J . R . Weinberg, Abstraction, Relation and Induction (Madison: University
of Wisconsin Press, 1965), 121-53.
13 The Origins of the Problem§3 §3

Not that it is easy to operate in this area with a vocabulary that


tidily abstracts from all temporal or methodological consi-
derations. Even the term 'hypothesis' is potentially misleading,
since in ordinary speech it denotes not just a proposition or unit
of logical analysis, but a person's working assumption. Indeed,
the ordinary concept of a hypothesis may have an important role
to play in the history and heuristics of natural science. But it is
not intrinsic to the logic of inductive support. Similarly, though
the term 'confirmation' is often used in place of 'inductive
support', it has a temporal implication that may also be
misleading. In ordinary usage a prediction needs to have been
advanced, or a hypothesis entertained, before it can be said to be
confirmed. What confirms normally comes later than what is
confirmed. But it is only so far as inductive confirmation is
conceived as a timeless relation that the question arises whether
it is properly categorized as a form of probabilification that is
subject to the principles of the mathematical calculus of chance.
We must now move further towards understanding the
problem which that question formulates. As a start in that
direction the present section has outlined the classical, Baconian
tradition about inductive support. The other necessary
preliminary is to sketch the fundamental features of Pascalian
probability. And this task will also be approached initially, in the
next section, from a historical point of view.

§ 3. T H E R I S E O F P A S C A L I A N P R O B A B I L I T Y

The mathematics of Pascalian probability originated in the study


of games of chance, and in the latter half of the seventeenth
century it began to be applied to measurements of the credibility
of a proposition. But, though aleatory probability always
requires a complementational principle for negation and a multi-
plicative principle for conjunction, there are contexts in which
credibility conforms to non-Pascalian principles. The mathe-
matics of Pascalian probability has subsequently been enriched
in many ways, not least by Bernoulli's Limit Theorem, by
Bayes's Law, and by modern axiomatizations.
We have seen that in the Baconian tradition certainty was
treated as a matter of degree. At least, the grade of certainty of
14 The Origins of the Problem§3
one generalization about the cause of a particular physical
property could be compared with that of another generalization
about the same property, on the evidence to hand, or a general-
ization's grade of certainty on the basis of one set of evidential
facts could be compared with its grade of certainty on the basis of
another such set. Perhaps comparisons could also be drawn
between generalizations about different properties, so far as
they had been subjected to checks against the same kinds of
evidence or could be assigned to particular levels in the pyramid
of laws; and such comparisons might affect the choice of means
to given ends. But that was as far as gradation could go.
Certainty was not treated as a quantitative property that could
be measured in terms of standard units, as length can be
measured in terms of metres or yards, or time in terms of days or
years. Nor was there any logical syntax for judgements of
certainty, enabling anyone to infer, for example, the level of a
proposition's certainty from the level of its negation's certainty,
or the level of a conjunctive proposition's certainty from the
level of its conjuncts' certainty (i.e. the certainty-level of H &H, t

from the certainty-levels of H and H ). In fact J . S. Mill was the


i 2

first to propose a more refined mode of gradation within a


systematic treatment of variative induction. His proposal
exploited what had by his day come to be known as 'the theory of
probability': Pascal's legacy at last converged with Bacon's. But
before we traverse Mill's ideas on the subject it will be useful to
sketch, very briefly, the history of Pascalian probability.
The mathematics of probability originated in the discussion of
rules for calculating the chances of specified kinds of outcomes or
combinations of outcomes in dice-throwing and similar games of
chance. Thus the great astronomer Galileo took an interest in
the problem of calculating the number of different permutations
in which three dice could be thrown so that the numerals on their
topmost sides sum to a particular number. Similarly the
Chevalier de Mere's famous questions put two problems to
Pascal: how should the stakes be divided when two players
26

separate without finishing their game and how many tosses are
needed, in throwing two dice, to have at least an even chance of

2 6Blaise Pascal (1623-62) was well known as a strong supporter ofJansen's theology
as well as being an important mathematician.
15 The Origins of the Problem§3 §3

getting a double six? And the discussion of such issues was


further promoted by Pascal's 1654 correspondence on them with
Fermat, by Christiaan Huygens's 1657 book on reckoning in
games of chance and by the much-delayed publication of the
mathematician and physician Cardano's mid-sixteenth-century
writings on this subject in 1663.
But the study of games of chance has an epistemological
ambiguity that is found in other branches of elementary
mathematics also. On the one hand its problems are soluble
a priori, by theorem-licensed calculation from the figures
supplied, just as, if we are given the length of a square's side in
Euclidean space, we can calculate the length of its diagonal on
the basis of Pythagoras's theorem. O n the other hand, to the
extent that the dice used are unbiased and thrown at randon we
expect the relevant calculations to be empirically confirmed by
relative frequencies in the actual outcomes, just as we expect
actual diagrams on paper or actual squares of wood to confirm
our Euclidean calculations. The a priori calculations represent
ideal cases, to which we expect the real world to be seen to
conform within appropriate limits of approximation. If this
conformity is not found, then either the calculations are
incorrect or the dice-throws are biased. Not surprisingly,
therefore, the observational experience of generations of dice-
throwers seems sometimes to have acted as a historical check on
the mathematics, as in the case of Galileo's claim that the
outcomes of three dice-throws had been observed to have a
greater chance of summing to 11 than to 12. To establish this
result mathematically it was necessary to make sure that one
assumed all the different permutations of outcomes, not the
different combinations, to have equal chances of occurrence.
(A 'combination' of possible outcomes specifies just which
particular numbers are thrown and how often, irrespective of
which throw produces which number, while a 'permutation'
assigns a number to each particular throw. So the only
combination of three throws that sums to 4, for example, is two
l's and a 2, but this combination has three different permu-
tations: 1-1-2, 1-2-1 and 2-1-1.) 27

2 7I. Hacking, The Emergence of Probability: A Philosophical Study of Early Ideas about
Probability, Induction and Statistical Inference (Cambridge: Cambridge University Press,
1975), 51-3.
16 The Origins of the Problem§3
In this seminal period it was not just the mathematics that was
developing, but also its range of applications. The crucially
important step that had clearly been taken by 1662 was to apply
the calculus of chance to other domains than the outcomes of
aleatory games. Thus the agnostic, in Pascal's famous wager, 28

is supposed to treat the existence and non-existence of God as


having equal chances and then to choose the religious life
because of its infinitely superior expectations, much as in a fair
toss we should put our money on heads if it had a higher pay-off
than tails. In any case there are problems about the force of
Pascal's arguments. For example, so far as it applies at all, it
applies equally well to decisions about the existence of any other
benevolent deity than the Judaeo-Christian one. But what is
relevant in the present context is that Pascal's argument extends
the concept of chance in an important way. The existence and
non-existence of God are not types of repeatable event or
alternative types of outcome in any kind of game, like a coin's
falling heads or tails. So no relative frequencies among actual
events can operate as a check on the assignment of equal chance
to the two propositions, nor can we usefully characterize the
particular issue at stake—the existence or non-existence of
God—as an instance of a familiar kind of issue, in the way that
the outcome of the next toss can be characterized as just another
instance of the issue of a fair toss. Indeed, on this issue we seem
not to be dealing at all with a measurable property of classes or
kinds of events either in an ideal world or in the real one, but
rather with a measurable property of propositions. It is a matter
of the probabilities we assign in our reasonings, on the basis of
the information that we have in our minds, rather than a matter
of the chances that are objectively there to be counted.
Moreover, in the Port-Royal Logic, first published in 1662,
29

the fact that a certain type of event has happened more often
than not when the circumstances are of such-and-such a kind is

2 8 B. Pascal, Pensees, trans. H.F. Stewart (London: Routledge and Kegan Paul,
1950), 120-1.
2 9 A. Arnauld, The Art of Thinking: Port-Royal Logic, trans. J . Dickoff and P. Jones
(Indianapolis: Bobbs-Merrill, 1964), 350-1. Antoine Arnauld (1612-94) supported
Jansen's theology against that of thejesuits. He is generally credited with the authorship
of the book La Logigue ou I'Art de Penser, which was published by Jansenists from Port-
Royal in 1662.
17 The Origins of the Problem§3 §3

said to warrant belief with a high degree of probability, on a


particular occasion in which these circumstances are present,
that an event of the kind in question occurred. For example, if
999 contracts out of 1,000, that have been signed by two
notaries, have not been post-dated, it is very highly probable
that the contract I see (which has been signed by two notaries)
has not been post-dated. The validity of such an argument needs
further discussion—see § 14 below. But it is clear that the
measurement of a proposition's credibility is now being treated
as having the same mathematical structure as the measurement
of an aleatory chance and that the term 'probability' is now
being used in both connections. 30

Nor was this identification of structure as much of a truism as


it may seem. The chance of a fair die's falling six uppermost (f)
is obviously equal to one minus the chance of its falling with
some other number uppermost (i.e. to 1 — f). And in general, if
the chance or relative frequency of is n, the chance or relative
frequency of not-^4 is the complement of n. However, credibility
differs here from chance or relative frequency. It is not
necessarily governed by a complementational principle for
negation. Such a principle is certainly in operation if the
negation we have in mind in connection with the credibility of A
is the incredibility o f A . As the credibility of A goes up, the incre-
dibility of A goes down, and vice versa. But a complement-
ational principle will not apply if the negation we have in mind
in connection with the credibility (belief-worthiness) of A is not
the incredibility of A, but the credibility (belief-worthiness) of
not-A. On the basis of a rather jejune statement of supposed
evidence the credibility of A and the credibility of not-,4—that is,
the justification for believing^ and the justification for believing
not-^4—may both be rather low. For example, the report that a
certain document has not been notarized at all may provide as
little warranty to believe that it has not been post-dated as to
believe that it has. In this sense of'probability' the probability of
its having been post-dated will not vary inversely with the
probability that it has not been. And the everyday concept of
certainty is similarly ambivalent. As the certainty of A goes up,
the uncertainty of A goes down, and vice versa. But there may
3 0 The Art of Thinking, pp. 351, 355.
18 The Origins of the Problem§3
well be very little certainty that A and also very little certainty
that not-y4.
So the quantitative gradation of belief, credibility, etc. took
one road initially with the Pascalian theory of probability and its
aleatory model, although another mathematical road was also
open. Later in the seventeenth century Jacob Bernoulli took
some exploratory steps down that other road, and so did J . H .
Lambert in the latter half of the eighteenth century. And more 31

recently this important issue has been reopened (see § 14 and


§§ 22-24 below). Certainly the dominant mathematical
doctrine has been that—in accordance with the aleatory
model—probabilities are complementational or additive, as
canonically declared by Abraham de Moivre in 1718. De
Moivre wrote t h a t ' If the Probability of Happening and Failing
are added together, the Sum will always be equal to unity': in 32

modern terms
p(/4) + p(not-^4) = 1.
But it would be dangerously misleadng to regard this thesis as a
truism.
Compare the Pascalian law for the probability of a
conjunction of outcomes, which is easily established from an
aleatory model. When the outcomes are independent of one
another, as when each ball drawn at random from an urn is
immediately replaced, we must just multiply together the
probabilities of the separate outcomes. For example, if there are
two white and two black balls in the urn, the chance of drawing a
white ball on any one drawing is 2 out of 4, and, if there are two
drawings, then, since each of the four possible outcomes of the
first drawing may be followed by any of the four possible
outcomes of the second drawing, the chances of drawing a white
ball twice is j . Of course, the ball drawn first may not be
replaced. If so, the second outcome will not be independent of

3 1 See G. Shafer, 'Non- Additive Probabilities in the work of Bernoulli and Lambert',
Archive for the History of Exact Sciences, 19 (1978), 309-70.
3 2 A. de Moivre, The Doctrine of Chances, or a Method of Calculating the Probability of Events
in Play (London: de Moivre, 1718), 1. Abraham de Moivre (1667-1754), a French
Huguenot who emigrated to England, was one of the commissioners appointed by the
Royal Society to arbitrate on the dispute between Newton and Leibniz about which of
them invented the differential calculus.
19 The Origins of the Problem§3 §3

the first, and we shall be concerned in its case with only the three
outcomes left possible by the outcome of the first drawing. The
chance of drawing a white ball twice will then be
Symbolically, let A be the event of drawing a white ball first, and
B the event of drawing a white ball second; and let the dyadic
function p(5|yl) evaluate the probability of B given^l. Then both
cases will be covered by the same multiplication principle for
conjunction,
P(A8lB) = p(A) x {B\A),
V

which says that the chance of drawing two white balls is the
product of the chance of drawing a white ball first and the chance
of drawing a white ball second, given that a white ball was also
drawn first. According to that principle, where the ball is
replaced after being drawn p(^4&5) equals because the two
outcomes are independent and p(5|^4) equals p(-B) and thus
equals 7; and, where the ball is not replaced after the first
drawing so that the two outcomes are not independent, p(.B|^4)
equals j and so equals (We say that p(^4) is a
'monadic' function of A, because its value depends on just one
issue—the issue of whether A occurs or not. Similarly p(^4&5) is
a monadic function of^4&fi. But p(SW) is a 'dyadic' function of
A and B because its value depends on the two issues that it
relates.)
All this is straightforward and incontrovertible if our model is
an aleatory one. But suppose an art historian declares two
pictures to be genuine Vermeers. He seems to have given you a
warranty—the warranty of his expertise—to believe that the
first is a Vermeer, a warranty to believe that the second is, and a
warranty of just the same nature to believe that they both are.
The warranty for the conjunction seems no less reputable, and
no less thoroughly researched, than for either of the conjuncts.
But this conflicts with Pascalian principles because the above-
mentioned multiplicative law for the probability of a
conjunction ensures that, except in limiting cases, the
conjunction is less probable than either conjunct. If p(A) > 0
and p(B\A) < 1, p(/l&5) must always be less than p(^4), since in
accordance with the multiplication principle for conjunction
p(^4&5), as we have just seen, is equal to a proper fraction of
p(^4), namely, p(^4) X p(B\A). Of course, the chance of both
20 The Origins of the Problem§3
pictures' being genuine may well be a lot less than the chance of
just one's being genuine. But is the credibility of their genuine-
ness to be judged in terms of such chances (which are
presumably to be viewed metaphorically as chances in the
lottery of life) or in terms of the reputation of the author of the
warranties that have been given you? The dominant theory has
treated such warranted credibilities as sharing the same mathe-
matical structure as aleatory chances. But this assignment of
mathematical structure is by no means as incontrovertible in the
case of warranted credibilities as in that of aleatory chances.
So both in dealing with conjunction and in dealing with
negation more than one structure is conceivable. And if there is
more than one way of dealing with conjunction there must also
be more than one way of relating dyadic to monadic probabili-
ties. For it is a matter of elementary algebra, where p(^4) > 0,
that in accordance with whether p(/l&.B) is, or is not, equal to
p(^4) X p ( £ | i ) , correspondingly p(fi|^4) is, or is not, equal to
p(/l&£)

Moreover, when using Pascalian functions, we have to


distinguish carefully between judgements of the form 'p(5|^4) =
n', judgements of the form 'If A, then p(B) = n', and judge-
ments of the form 'p(If^4, then B) = n'. Pascalian principles are
demonstrably insufficient to ensure that a judgement of one of
these three forms entails a judgement of one or other of the other
two forms. For example, if the monadic Pascalian probability
33

of a conditional statement, i.e. if p(If A, then B), is n, then the


monadic Pascalian probability of its contrapositive, i.e. p(If
not-5, then not-^4), should also be n since a Pascalian function
admits the replacement of an expression by a logically equi-
valent one. But in a similar context, if the dyadic—often called
34

3 3 See L. Jonathan Cohen, The Probable and the Provable (Oxford: Clarendon Press,
1977), 30 n. 21.
3 4 The contraposability of conditional statements has been called into question by
E.W. Adams, The Logic of Conditionals (Dordrecht: Reidel, 1975). But Adam's argument
rests on the counter-intuitive, or at least controversial, assumption that the probability of
a conditional 'If A, then B' has to be equated with the conditional probability p(SU),
and that conditionals cannot be assigned dichotomous truth-values in a way that will
always and necessarily equate the probability of a conditional with the probability of its
being true. Indeed, we should be quite unnecessarily impoverishing our conceptual
resources if we excluded ourselves from ever employing the expression 'if . . ., then
21 The Origins of the Problem§3 §3

'conditional'—probability p(BIA) is n, it does not follow that the


contrapositive of this probability, namely p(not-^41 not-Z?), is also
n. Though there is only a rather low probability that a person
lives in Oxford given that he lives in England, there is quite a
high probability that he does not live in England given that he
does not live in Oxford. So when we want to use a kind of dyadic
probability function that is invariant under contraposition, we
must either apply some appropriate non-Pascalian principles or
some appropriate refinement of Pascalian ones. And such a
function does sometimes seem appropriate. It may be tempting,
for example, always to equate the probability, in a particular
case, that the accused is guilty, given that such-and-such
evidence is before the court, with the improbability of that
evidence's being before the court, given that the accused is not
guilty. For this purpose we seem to need a way of judging
probability according to which the dyadic probability p(5|/4) is
necessarily equal to the monadic probability p(If A, then B).
Since its beginnings in the years around 1660 the Pascalian
theory of probability has made enormous progress, both in the
enrichment of its mathematics and in the extension of its range
of application. Indeed, these two types of progress have
constantly reinforced one another. But it will suffice to cite here
only three major developments in the mathematics of
probability that will be mentioned in later discussions of the
philosophy of induction.
In 1692 Jacob Bernoulli proved an important theorem about 35

any countably infinite set of chance outcomes. Let n be the


number of outcomes in a finite subset of this set; let r be the
relative frequency with which members of that subset have a
certain feature; and let p be the probability of that feature's
occurring in a member of the superset. Then Bernoulli proved
that as n increases without bound, the probability that r is within
. . .', or an equivalent, in a necessarily contraposable sense or in a sense in which the
probability of the conditional 'If A, then B' is not necessarily equivalent to the condi-
tional probability p(B\A). See also D. Lewis, 'Probabilities of Conditionals and Condi-
tional Probabilities', Philosophical Review, 85 (1976), 297-315.
3 5 Later published in pt. IV, ch. 5 of his Ars Conjectandi (Basle: Thurnisii Brothers,
1713). For an exposition of the proof see W. Kneale. Probability and Induction (Oxford:
Clarendon Press, 1949), 136-9, or J R. Lucas, The Concept of Probability (Oxford:
Clarendon Press, 1970), 73-83. Bernoulli (1654-1705) opened a public seminary for
teaching experimental physics in Basle after returning in 1682 from travels in England,
France, and Holland. In 1687 he became Professor of Mathematics at the University of
Basle and later its Rector.
22 The Origins of the Problem§3
a specifiable, small interval ofp, increases towards 1 That is to
say, the larger we make the subset, the higher the credibility that
the relative frequency of the feature in the subset approximates
p, given that/; is the true probability of the feature's occurring in
a member of the superset. (As a corollary, for any specified level
of credibility, the interval within which r may be said, with that
level of credibility, to approximate p, will decrease as n
increases.) This result, known as 'Bernoulli's limit theorem' or,
more colloquially, as 'the law of large numbers', has numerous
applications. For example, where the superset consists of
random drawings—with replacement—from an urn containing
9 black and 3 white balls, we can determine a priori a value for
the true probability of drawing a black ball, namely, f.
Bernoulli's theorem then allows us to predict, with a credibility
not too far from 1, that in a long sample-sequence of actual
drawings the empirically ascertainable relative frequency of
black balls' being drawn should approximate this figure. Again,
if a large insurance company knew that the true probability with
which a man of certain age, life-style, and medical condition will
die within a year is j, it could use Bernoulli's theorem to derive,
with credibility not too far from 1, the approximate proportion
of its own clients in that category who will die within the year.
But how can the company determine the true probability? In
the case of the urn, we obtained this probability a priori, on the
assumption that the drawings were random ones But mortality
rates are facts of nature and need to be discovered empirically. It
is tempting, therefore, to invert the inference and to derive (with
sufficient credibility) an approximate figure for the true
probability from a year's relative frequency of deaths in some
sufficiently large sample of the appropriate category of men. But
this inverse inference was not established by Bernoulli's math-
ematics. And we can easily see that the credibility (i.e. Pascalian
probability) of B, given A, is not necessarily identical with the
credibility (i.e. Pascalian probability) of A, given B, if we
compare the probability (namely, j) of a dice-throw's being
even, given that it is greater than 2, with the probability
(namely, f) of a dice-throw's being greater than 2, given that it is
even.
So how can Bernoulli's theorem be usefully applied when the
true probability is an empirical fact that needs to be estimated?
Bernoulli himself claimed that even though an approximate
23 The Origins of the Problem§3 §3

value for the true probability does not follow inversely, with high
credibility, from the size of the relative frequency in a large
sample, at least the existence of that value for the true
probability is the simplest possible explanation of the size of the
relative frequency. A different manoeuvre is adopted by the
36

various statistical procedures that derive from the work of R. A.


Fisher, E. Pearson, or J . Neyman in the earlier part of the
present century. There the underlying idea is that the investi-
37

gator is invited to agree some conventional value for a specified


parameter—for example, a 'significance level'—such that if that
value is attained, in respect of a sample of specified size, then the
true probability may be taken not to differ from the observed
relative frequency by more than some stated interval. For
example, it might be unacceptably improbable that a relatively
large sample with that relative frequency in the sample had been
observed unless the true probability was reasonably close to the
relative frequency. It might be unacceptably improbable that 80
per cent of a 10,000-membered sample of tuberculous patients
recovered after treatment with a certain drug unless the long-
run expectancy was reasonably close to that percentage. This
conventionalist idea emerges in a variety of sophisticated forms
in many modern statistical textbooks, and very many
probabilities have been estimated on its basis in various fields of
scientific research.
But some statisticians, often called 'Bayesians', prefer instead
to exploit the mathematical law that governs the inversion of a
probability and relates the probability of B given A to the
probability of A given B. This law is widely known as Bayes's
law,Letter
3 6
because
to Leibnizawareness
of 20 April 1704,of initLeibnizens
is implicit—although
mathematische Schriften, ed. not
C.I.
Gerhardt, iii
explicit—in (Halle: H
a paper. W . Schmidt,
of Thomas 1855), 87-8.
Ba^es (1763).
Workers, 7thInedn.
its (Edinburgh:
simplest
38
See esp. R. A.
3 7 Fisher, Statistical Methods for Research
Oliver and Boyd, 1938), 120 ff.; J. Neyman, 'Outline of a Theory of Statistical
Estimates Based on the Classical Theory of Probability', Philosophical Transactions of the
Royal Society of London, ser. A, 236 (1937), 333 ff.; E. Pearson, 'On the Criteria that a
given System of Deviation from the Probable in the case of a Correlated System of
Variables is such that it can be reasonably supposed to have arisen from Random
Sampling', The London, Edinburgh, and Dublin Philosophical Magazine andJournal of Science,
5th ser. 50 (1900), 157-75, and 'The Fundamental Problem of Practical Statistics',
Biometrika, 13 (1920), 1-16.
3 8 T. Bayes, 'Essay towards Solving a Problem in the Doctrine of Chances',
Philosophical Transactions of the Royal Society of London, 53 (1763), 371-418.
24 The Origins of the Problem §3
form the law states that, where p(B) > 0,

And in that form the law is very easily derived within the
calculus of Pascalian chance. It results, by elementary algebra,
from putting equal to each other the two basic ways of spelling
out (in accordance with the multiplication principle for
conjunction) the probability of the conjunction of two outcomes,
A and B, namely, p(A IB) X p(5) and v(B\A) X p(^4). But
Bayes's law obviously does not solve all the problems here.
When you know the value of p(A!B), Bayes's law can be used to
derive p(5M) only if you also know the values of p(A) and p(B)
or accept some conventional procedure for determining them.
So whether you think of Bayes's law as a supplement to
Bernoulli's theorem that is useful in relation to the inverse
inference, or wish to exploit it in some other way, you must be
able to evaluate these monadic probabilities (see further § 9 and
§ 19).
In the present century the mathematics of Pascalian
probability have been studied systematically by means of
axiomatization, with varying thoroughness of formalization.
Thus A. Kolmogorov in 1933 produced six axioms for the
39

monadic probability function that are demonstrably consistent


and independent, plus a definition of the dyadic function in
terms of the monadic, one that, where p(£) > 0, puts p(A\B)
equal to
p{AtkE)
P (B)
Of course, except in certain limiting cases, such as p(A&.not-A)
= 0, this axiomatization gives us, as theorems, only laws
governing relations between probability-judgements. It does
not enable us to assign determinate probability-values to
specified outcomes. Indeed, axiomatization has made it possible

See A. Kolmogorov, Foundations of the Theory of Probability, trans. N. Morrison (New


3 9

York: Chelsea Publishing Company, 1950).


25 The Origins of the Problem§3 §3

thus to distinguish between problems about the syntax of


probability-functions and problems about their semantics, in a
way that was impossible in previous centuries.
Kolmogorov's treatment assumes probabilities to be real-
number-valued functions of sets. That is, the expressions that
are to occupy the argument-places in his functors 'p(. . .)' or
'p(. . .1. . .)' are expressions of a Boolean algebra that
designate sets, and when the functors are thus filled out they
designate real numbers. So, though his axiomatization has
several mathematical advantages, it is not as neutral as it might
be between different interpretations of the Pascalian structure.
Since truth-functional operations on propositions, such as
negation, conjunction, and disjunction, obey the same formal
principles as Boolean operations on sets, such as complementa-
tion, intersection, and union, respectively, an interpretation of
Boolean algebra in terms of propositions is as legitimate as an
interpretation in terms of sets. But Kolmogorov's treatment
excludes probabilities from being functions of propositions,
unless we are willing to identify each proposition with some
corresponding set, such as the set of possible worlds in which it is
true. And this creates an unnecessary, technical obstacle to any
analysis of inductive support in terms of mathematical
probability. If inductive support is a relation between one
proposition (stating evidence) and another (stating a
hypothesis), Kolmogorov's theory of probability seems to rule
out any Pascalian analysis of inductive support ab initio, unless
you are willing to introduce controversial assumptions about the
nature of propositions.
One should note, therefore, that it is also possible to
axiomatize the structure, or logical syntax, of Pascalian
probability in a way that imposes less restriction on the applica-
tion of its principles. This more thoroughly formalized axioma-
tization of the Pascalian theory was achieved by K. R. Popper in
1938. The expressions that occupy the argument-places in his
40

monadic functors may designate any elements of a Boolean


algebra. Moreover, Kolmogorov's definition of the dyadic
probability functor in terms of the monadic one was restricted to
cases where one of the monadic functions involved had a non-
4 0K . R . Popper, 'A Set of Independent Axioms for Probability', Mind, 47 (1938),
275 ff.
26 The Origins of the Problem§3
zero value (see above), and this too was unnecessarily
restrictive. Popper later replaced his 1938 axiomatization by a
41

system of demonstrably independent and consistent axioms in


which the dyadic functor is taken as primitive in place of the
monadic one, and the latter is defined in terms of the former. In
this system p(^4) is put equal to p(A\E) if for every C p(A\A) is
equal to p(5| C). (The point here is that, since p(A\A) must equal
1, the requirement that, whatever C may be, p(fi| C) should
equal p(A\A) compels B to have the probability 1 on any
condition whatever, and if B is thus necessary there can be no
difference between p(^4) and p(A | B).
Another important advance in the mathematics of Pascalian
probability has been the development of interval-valued
functions. Probabilities are then conceived as intervals between
upper and lower limits, instead of as points, and it becomes
possible to reflect the intuition that, especially in relation to
single outcomes, precise estimates of probability often seem
artificial and unrealistic. Alternatively we can say in such cases
that point-valued judgements are idealizations of interval-
valued ones. 42

4 1 The Logic of Scientific Discovery (London-. Hutchinson, 1959), 326-48.


4 2 For the history of the earlier period in the mathematics of probability see I.
Todhunter, A History of the Mathematical Theory of Probabilityfrom the Time of Pascal to that of
Laplace (Cambridge: Macmillan, 1865). For a variety of views about the interrelations
between the origin and development of probability theory and its social context see L. E.
Maistrow, Probability Theory: A Historical Sketch, trans. S. Kotz (New York: Academic
Press, 1974); I. Hacking, The Emergence of Probability, D. Garber and S. Zabell, 'On the
Emergence of Probability', Archivefor the History ofExact Sciences, 21 (1979), 33-53; a n d l .
Schneider, 'Why Do We Find the Origin of a Calculus of Probabilities in the
Seventeenth Century?', i n j . Hintikka, D. Gruender, and E. Agazzi (eds.), Probabilistic
Thinking, Thermodynamics and the Interaction of the History and Philosophy of Science: Proceedings
of the 1978 Pisa Conference on the History and Philosophy of Science (Dordrecht: Reidel, 1981),
ii. 1-24. There are also some interesting historical remarks in J . M. Keynes, A Treatiseon
Probability (London: Macmillan, 1921), 79-91, and in addition a useful bibliography,
431-58. R. Jeffrey, 'De Finetti's Probabilism', Synthese, 60 (1984), 73-90 emphasizes,
against Hacking, the influence of the Sceptic philosopher Carneades (214-129 BC), via
Sextus Empiricus and Montaigne, on the development of the concept of probability. But
this is poor historiography. It crucially underrates the importance of later mathematical
developments (Cardano, Galileo, Pascal, Fermat, Huyghens) which were essential to
give a gradational structure to the concept of probability. No doubt some concept of
probability existed long before Carneades even. What are historically important are the
events that triggered off a rich new development in its structure. These events took place
mainly in the seventeenth century. Carneades and the other Sceptics merely encouraged
the use of the concept, whatever its structure, in contexts where certainty would
otherwise be taken for granted. For the history of probability in nineteenth-century
science see L. Kriiger et al. (eds.), The Probabilistic Revolution (Cambridge, Mass: M I T
Press, 1987) i and ii. On the contribution of legal notions of fairplay to the aleatory
27 The Origins of the Problem§3 §3

But it would be out of place to comment at greater length on


the mathematics of Pascalian probability. Enough has now been
said about its principles to give a fairly clear sense to the philo-
sophical question: does such a probability afford an appropriate
measure for Baconian induction?
§ 4. T H E C O M B I N A T I O N O F B A C O N I A N A N D P A S C A L I A N
THEMES
Leibniz, Bernoulli, Hume, and Herschel made no attempt to
grade the strength of variative induction by Pascalian
probability. But, when Mill developed a theory of variative
induction in terms of his Methods of Agreement, Difference,
Residues, and Concomitant Variations, he implicitly or
explicitly sanctioned the use of Pascalian probability in the
evaluation of some relevant parameters.
What we have seen in § § 2-3 is that in the seventeenth century
the gradation of certainty began to develop along two apparently
different lines. O n the one hand, Bacon explored the role of
evidential variety as a basis for ampliative induction, and gave
no indication of any kind that he would have accepted an
aleatory model for the gradation of inductive support. He does
not even seem to have conceived a generalization's degree of
certainty as a ratio of any kind. On the other hand, Pascal,
Fermat, Huygens, and others explored the mathematics of
chance as a syntax for judgements of probability, but did not
apply it to the evaluation of variative inductions. Nor did these
two lines of enquiry begin to converge for about a couple of
centuries. Thus Leibniz called one scientific hypothesis more
probable ('probabilior') than another if it was simpler and yet it
explained a larger variety of phenomena from a smaller number
of postulates, and especially if it allowed the prediction of as yet
unknown phenomena. But he apparently never tried to
43

represent or elucidate these thoroughly Baconian criteria within


the mathematics of chance, despite his readiness to apply the
calculus of chance to other kinds of probability-judgement.
model see L. Daston, Classical Probability in the Enlightenment (Princeton: Princeton
University Press, 1988) 3-48.
4 3 Letter to Conring, 19 March 1678, in Diephilosophische Schriften von Gottfried Wilhelm
Leibniz, ed. C.I. Gerhardt. (Berlin: Weidmannsche Buchhandlung, 1865), i. 195-6.
Leibniz (1646-1716) originated many important ideas in logic, philosophy, and
mathematics, and was the first president of the Berlin Academy of Sciences.
28 The Origins of the Problem§3
Bernoulli certainly thought that his own theories about math-
ematical probability had important implications for the
evaluation of scientific hypotheses. But he spoke of the number,
not of the variety, of evidential instances as being relevant for
the purposes of this evaluation. 44

H u m e explicitly recognized the issue. Probabilities arising


from chance, as in a game of dice, he evaluated by the ratio of
favourable outcomes among a class of outcomes that are all
indifferently possible. And probabilities arising from a causal
45

tendency that does not operate in every instance Hume


evaluated by the same rule as that applied in the Port-Royal
Logic (see § 3 above) to the authenticity of notarized contracts.
In Hume's example, if I have found by long observation that of
twenty ships which go to sea only nineteen return, then -55-
measures the strength of my belief that a ship which goes to sea
in the future will return. Indeed, Hume did not just assume the
46

appropriateness of an aleatory model for such judgements: he


argued it. O n his view
Every past experiment may be consider'd as a kind of chance; it being
uncertain to us, whetheK the object will exist conformable to one
experiment or another: And for this reason every thing that has been
said on the one subject is applicable to both. 47

But this kind of probability is concerned with the numbers of


evidential instances, not their variety. And H u m e treated
probabilities 'arising from analogy' as a third, quite distinct
category, which he never suggested measuring by a numerical
ratio. He maintained that reasoning from causes or effects
requires not only the conjunction of two objects in all past
experience but also the resemblance of a present object to any
one of them. He inferred that
as this resemblance admits of many different degrees, the reasoning
becomes proportionably more or less firm and certain. An experiment
loses of its force, when transfer'd to instances which are not exactly
resembling; 'tis evident it may still retain as much as may be the foun-
dation of probability, as long as there is any resemblance remaining. 48

4 4Letter to Leibniz of 20 April, 1704, in Leibnizens mathematische Schri/ten, ed.


Gerhardt, iii. 87-8.
4 5A Treatise of Human Nature, ed. L.A. Selby-Bigge (Oxford: Clarendon Press, 1888),
125 ff. David Hume (1711-76) was born in Edinburgh and eventually became librarian
to the Faculty of Advocates there. His sceptical epistemology has constituted an
important challenge to many subsequent philosophers.
4 6Ibid. 134-5. 4 7Ibid. 135 . Ibid. 142.
4 8
29 The Origins of the Problem§3 §3

So, since some closer resemblance between selected evidential


and predictive instances may be expected when a generalization
is confirmed by a greater variety of evidential instances, it is clear
that what Hume called 'probability arising from analogy' is
closely connected with the gradation of variative induction. But,
though he uses the term 'probability' here, he makes no
suggestion that it is a probability subject to the mathematics of
chance.
Similarly Herschel was well aware of the theory of Pascalian
probabilities and applied it in the resolution of problems about
accuracy of measurement in the physical sciences. But he made 49

no reference to this theory when he used the term 'probability'


elsewhere in talking about the 'probability of success' in
discovering correct explanations of natural phenomena or when
he enunciated a gradualist approach, like Bacon's, to variative
induction. 50

In fact, it seems to have been J . S. Mill who first attempted to


combine a Baconian emphasis on the importance of variative
induction with Pascalian convictions about the mathematics of
probability, while withdrawing in the second edition of his
System of Logic (1846) the somewhat sceptical view about the
overall foundations of the calculus of chance that he had taken in
the first edition of that book. 51

For the most part, and despite criticism of Bacon in respect of


some details, Mill's inductive logic is an elaboration and
extension of Bacon's ideas on the subject. He too assigns an
inferior value to induction by simple enumeration. 'The
tendency of unscientific enquirers', he says, 'is to rely too much
on number, without analysing the instances; without looking
closely enough into their nature, to ascertain what circum-
stances are not eliminated by means of them.' Like Bacon, Mill
insists on the superior value of a single falsifying instance that
excludes some feature from being causally relevant to the
occurrence of a particular phenomenon, as against any
multitude of verifying instances that are evaluated purely by
their number. And, as in Bacon's inferences from his tables of
52

presence and absence, Mill acknowledges two main principles to


4 9 Herschel, The Philosophy of the Inductive Science, p. 217.
5 0 Ibid. 148, 155.
5 1 System of Logic, p. 351. Compare the preface to the second edition, pp. vii-viii.
68 System of Logic, p.25569Ibid.214.7 Ibid.293.
0
30 The Origins of the Problem§3
be at work. One, called 'the Method of Agreement' compares
together different instances in which the phenomenon under
investigation occurs. By this method, on its own, we may
establish only what Mill calls 'empirical laws', that is correla-
tions or connections between phenomena that are not the
ultimate laws of causation. The other method, however, is
53

capable, on Mill's view, of leading us to ultimate causal laws. It


is called 'the Method of Difference' and compares instances in
which the phenomenon does occur with instances in other
respects similar in which it does not.
The principle regulating the Method of Agreement,
according to Mill, was that
If two or more instances of the phenomenon under investigation have
only one circumstance in common, the circumstance in which alone all
the instances agree is the cause (or effect) of the given phenomenon. 54

Or, schematically, if circumstances A&B&C issue ina&6&c, and


circumstances A&not-B&not-C issue in a&not-^&not-c, then a is
the effect of A because its occurrence is unaffected by the
presence or absence of B or of C. For example, if an alkaline
substance is combined with an oil in several otherwise different
varieties of circumstance, and in each case a soap results, then
'the combination of an oil and an alkali causes the production of
a soap'. So the Method of Agreement may be said to stand on
55

the ground that whatever can be eliminated, because the


phenomenon under investigation occurs whether or not it is
present, is not connected with the phenomenon by any causal
law. And an experimenter's object will normally be to eliminate
a number of such factors by 'controlling' for them, as it is now
called. (The 'control' situation is the one where supposedly just
the suspect factor is absent.)
Clearly there is also room for a complementary method
whereby whatever can not be eliminated, because the
phenomenon under investigation does not occur when it is
absent, must be connected with the phenomenon by some causal
law. Mill calls this complementary method the Method of
Difference, and regards its regulative principle as being that
If an instance in which the phenomenon under investigation occurs,
and an instance in which it does not occur, have every circumstance in
5 3 System ofLogic, pp. 338-43. 5 4 Ibid. 255. 5 5 Ibid. 254.
31 41 The Origins of the Problem§3 §3

common save one, that one occurring only in the former; the
circumstance in which alone the two instances differ is the effect, or the
cause, or an indispensable part of the cause, of the phenomenon. 56

Schematically, if circumstances A&B&C issue in a&b&c and


circumstances not-A&B&C—the control—in not-a&cb&c, then a
is the effect of A because its presence is adversely affected by the
absence of A. For example, says Mill, 'when a man is shot
57

through the heart, it is by this method we know that it was the


gunshot which killed him: for he was in the fulness of life imme-
diately before, all circumstances being the same except the
wound'.
Thus both methods are eliminative, in that they guide invest-
igators as to which factors can or cannot be eliminated from the
enquiry. And neither is enumerative, because neither exploits
the mere accumulation of instances irrespective of those
instances' circumstances. Moreover, both methods assume that
there is always a uniformity to be found. As Mill puts it, in every
induction we assume that 'the universe, so far as known to us, is
so constituted, that whatever is true in any one case, is true in all
cases of a certain description; the only difficulty is, to find what
description'. Or, more precisely, we assume that this type of
58

regularity pervades 'the particular class of phenomena to which


the induction relates'. So, for example, an induction
59

concerning the motions of the planets would not be vitiated if we


supposed that wind and weather are the sport of chance,
provided it be assumed that astronomical phenomena are under
the dominion of general laws.
But what ground can there be for supposing the truth of this
assumption about a universal 'law of causation', as Mill called
it? Mill's answer to that question is to claim that here induction
by simple enumeration has a legitimate field of application. 60

The assumption has been found true in a very large number of


cases and has never been found false. It is therefore justifiable to
treat apparent counter-instances (cases where apparently no
causal law operates) as due merely to human ignorance. Mill is
careful to qualify his thesis in relation to circumstances that are
unknown to us and beyond the possible range of our experience.
The law of causation, on his view, must not be received as a law
5 6 Ibid. 256. 5 7 Ibid. 5 8 Ibid. 201.
5 9 Ibid. 203. 6 0 Ibid. 372-6.
32 The Origins of the Problem§3
of the universe but as a law 'of that portion of it only which is
within the range of our means of sure observation, with a
reasonable degree of extension to adjacent cases'. But he does
61

not seem to have appreciated that even when formulated in these


modest terms his law of causation is not easily substantiated. T o
say that it is supported by enumerative induction is just to raise
the further question: what is the rational warranty for assuming
that enumerative induction is a justifiable procedure?
Mill pointed out that induction by the Method of Agreement
is subject to a special kind of uncertainty because a particular
kind of phenomenon may be produced by more than one kind of
cause. For example, two great philosophers might both have
been educated at the same school and be otherwise different in
background: yet their philosophical abilities might be due to
different causes. Mill saw two ways of reducing such an
62

uncertainty. One was by the observation of very numerous and


sufficiently various instances, so that coincidental similarities
between all observed instances are gradually eliminated, and it
then becomes clear whether evidence for a unique cause persists
in being present in the data or whether the apparent uniqueness
of cause was just an illusion created by insufficient variation of
circumstances. And the other way of reducing uncertainty, in
Mill's view, was by attempting to confirm the results obtained
from the Method of Agreement, either by using the Method of
Difference in combination with it or by connecting its results
deductively with some law or laws ascertained by the latter
method. But Mill apparently did not see that even the Method of
Difference may not reveal a unique cause. For example, even
though in some cases where there is no bacterial infection the
presence or absence of an ulcer is responsible for the presence or
absence of a stomach-ache, there may be other, ulcer-free, cases
in which the presence or absence of bacterial infection is
responsible for the presence or absence of a stomach-ache.
Scientists often need to build on previously acquired
knowledge. So Mill proposed also a Method of Residues, and
enunciated its principle as follows:

68 System ofLogic, p.25569Ibid.214. 7 Ibid. 293.


0
33 The Origins of the Problem §3

Subduct from any phenomenon such part as is known by previous


inductions to be the effect of certain antecedents and the residue of the
phenomenon is the effect of the remaining antecedents. 63

Schematically, if circumstances A&B&C issue in a&b&LC, and if


B is already known to cause b and C to cause c, then A must be
the cause of a. For example, said Mill (following J . F . W .
Herschel), if the movements of a comet cannot be wholly
64

accounted for by its gravitation towards the sun and planets, the
residual feature must be explained by the resistance of the
medium through which it moves. But in such a case Mill
recognized that in practice we could not be certain that A is the
only antecedent to which the residual phenomenon a may be
referred. So any induction by the Method of Residues needs to
be confirmed by obtaining A artificially and trying it separately,
or by deriving its operation from otherwise known laws. 65

, Finally, Mill drew attention to the fact that some kinds of


circumstance cannot be varied experimentally. For example, we
cannot remove the seas from the presence of the moon. Never-
theless we can discover the influence of the moon on tidal flows
by the Method of Concomitant Variations. This method
operates in accordance with the principle that
Whatever phenomenon varies in any manner whenever another
phenomenon varies in some particular manner, is either a cause or an
effect of that phenomenon, or is connected with it through some fact of
causation. 66

Mill acknowledged that it may be dangerous to predict the same


numerical rates of variation for quantitative phenomena that are
outside the limits within which our own observations have been
conducted. But he claimed—ignoring altogether the possibility
of sustained coincidences—that the Method of Concomitant
Variations could always establish at least the existence of some
causal connection between the two kinds of phenomena
investigated. 67

Mill's writings on induction can easily give the impression


that he oversimplified the issues, because he tended to introduce
his theses initially in a rather dogmatic and unqualified form.
6 3 Ibid. 260. 6 4 Ibid. 280. 6 5 Ibid. 261.
6 6 Ibid. 263. 6 7 Ibid. 26b.
34 The Origins of the Problem§3
But the requisite qualifications are often found elsewhere in the
text. Thus after giving examples in which the Method of
Agreement seems to work smoothly and uncomplicatedly he
points out that 'it is hardly ever possible to ascertain all the
antecedents' of a phenomenon and that even when we can
produce the phenomenon artificially this difficulty is merely
lightened and not removed. In other words we have to
68

distinguish between the highly tangled actual situations that we


encounter when we apply his inductive methods in practice and
the ideally simplified situations in terms of which Mill
formulated the principles that underlie those methods.
Inevitably, therefore, to the extent that we have not ascertained
all the antecedents of a phenomenon under investigation, some
uncertainty must remain attached to our conclusion about
which of those antecedents is causally connected with the
phenomenon. And that source of uncertainty must affect the
Method of Difference, though Mill apparently failed to see this,
just as much as it affects the Method of Agreement. We cannot
be absolutely sure, when we operate the Method of Difference,
that the two canonical instances have every circumstance in
common save one. We therefore cannot be sure that some
unnoticed difference is not the true cause.
Again, Mill initially speaks of causes as invariably followed by
their effects. But later he insists that all causal laws 'require to
69

be stated in words affirmative of tendencies only', because he70

recognizes that any one causal factor may be counteracted in its


operation by another. For example, a body may be kept in
equilibrium by two equal and opposing forces, each of which on
its own tends to move the body in a certain direction and each of
which would suffice to move it if circumstances were otherwise
favourable. Now obviously this doctrine does not cohere well
with Mill's formulation of the principle underlying the Method
of Difference. According to that principle what would cause the
body in equilibrium to move would be the absence of one of the
two mutually counteracting forces, whereas, if causal processes
are viewed as tendencies, what causes it to move is the presence
of a force. But on the latter view there will clearly be room for
uncertainty on any particular occasion about whether such a
tendency will actually be realized.
6 8 System ofLogic, p. 255 69 Ibid. 214. 7 0 Ibid. 293.
35 The Origins of the Problem §3

How is inductive uncertainty to be measured? Mill dealt with


this issue in at least three connections. And in all three he came
to evaluate uncertainty in broad conformity with the math-
ematical calculus of chance or in terms of a ratio or relative
frequency that could be treated as a Pascalian probability.
First, he argued that inductive reasoning by the Method of
Agreement is to be evaluated in accordance with the math-
ematical probability that the observed correlation is due to
causation rather than to chance. But in order to have the data for
such an evaluation 'it would be necessary to know what
proportion of all the individual sequences or co-existences
occurring in nature are the result of law, and what proportion
are mere casual coincidences'. And Mill thought it evident that
71

we cannot form even a plausible conjecture as to the size of these


ratios. That is why we have to fall back, in any particular case,
on the assurance afforded by observing the correlation in a large
number of appropriately varied instances. But such assurance
may itself be a matter of degree and Mill made no suggestion
how it might be measured.
Secondly, Mill, like Bacon, held that as empirical laws rise in
their degree of generality they become more certain. But Mill,
unlike Bacon, sought an explanation of this fact. O n his view,
when an empirical law derives somehow from causal laws, this is
because relatively few changes in the combination of causal
factors could counteract i t ' since the greater number of possible
combinations must have already existed in some one or other of
the instances' in which the law has been found to operate. So
apparently a hitherto unrefuted empirical law's degree of
generality should be measurable in terms of the proportion of
possible combinations of potentially counteractive causal factors
that have already existed in its instances. And even if the law is
just 'an ultimate co-existence', and not a causal connection, the
more general it is the greater is the probability that if exceptions
had existed some would already have presented themselves. For
example, we are readier to accept the possibility of an error in
our beliefs about the properties shared by all crows than in our
beliefs about the properties shared by all birds. However, 72

Mill's argument here assumes that different combinations of

71 Ibid. 359. 72 Ibid. 384.


36 The Origins of the Problem§3
potentially counteractive caused factors tend to occur uniformly
and homogeneously throughout human experience. And this
assumption is scarcely tenable in the context of scientific
enquiry. The design of an experiment is often motivated by the
73

desire to produce combinations of circumstances that do not


occur naturally in observable human experience. So the actual
probability that exceptions to a generalization have already
occurred will in practice depend on the decisions of research-
workers and those who fund their researches. It is certainly
important to gauge the impact of increased generality on the
inductive standing of a hypothesis. But Mill's proposal for a
method of measuring this is not successful.
Thirdly, Mill insisted that there is no intrinsic difference
between inductive and analogical reasoning. By ordinary 74

induction we infer that anything resembling certain known


items in being A resembles them also in being B. By analogy,
according to Mill, we can infer that anything resembling a
known item in a certain proportion of its hitherto observed
properties will have a probability, equal to that proportion, of
resembling the item in any other of its properties.
If we discover, for example, an unknown animal or plant, resembling
closely some known one in the greater number of the properties we
observe in it, but differing in some few, we may reasonably expect to
find in the unobserved remainder of its properties a general agreement
with those of the former, but also a difference corresponding propor-
tionately to the amount of the observed diversity. 75

So analogical reasoning is another form of induction that, on


Mill's view, has gradations of certainty that conform to the
mathematics of chance. Mill clearly assumes here, as elsewhere,
that there is only a finite number of mutually independent
observable properties for a thing to have or not to have. But,
unlike Bacon, he never acknowledged this assumption. Also,
unlike Bacon, he made the rather implausible assumption that
every feature in which one item may or may not resemble

See also § 16 below.


7 3

An Examination of Sir William Hamilton's Philosophy and of the Principal Philosophical


7 4

Questions Discussed in his Writings (London: Longmans, Green, 1865), 402.


68 System of Logic, p.25569Ibid.214.7 Ibid.293.
0
37 The Origins of the Problem§3 §3

another is of equal importance in relation to any kind of


prediction about it. Yet the fact that two patients resemble one
another in literary interests, for example, would normally be
much less important for medical prognosis than that they have
the same viral infection in their bloodstreams.
Thus Mill's philosophy of induction, though it agreed with
Bacon's in the importance it attached to eliminative procedures
as distinct from enumerative ones, took at least one important
new step within the eliminativist tradition. Whenever Mill
thought that criteria for the gradation of inductive certainty are
called for, he applied criteria that, however crudely, exploited a
conception of probability which conforms with the mathematics
of chance.
Bernard Bolzano, like Mill, assigned a central role in science
to the discovery of causes, and in a book published in 1837 he 76

had already advocated methods of ampliative induction that


were very similar to those that Mill advocated independently in
1843. He differed from Mill, however, in holding that, when
one investigates the reliability of a hypothesis, one should take
multiplicity of confirmatory evidential instances to be as
'important a ground for confidence as variety'. Moreover, 77

variative induction, he thought, could sometimes be rendered


particularly difficult because the number of circumstances that
accompany every event is not finite (as Bacon and Mill 78

apparently supposed it to be). And, though Bolzano thought,


like Mill, that the gradation of inductive certainty could be
accomplished in terms of Pascalian probability, he relates this
measure to the number of evidential instances rather than to 79

their variety. 80

7 6B. Bolzano, Wissenschqftslehre(Sulzhach: J . E . von Seidelchen, 1837), trans, and ed.


R. George, under the title, The Theory of Science (Oxford: Blackwell, 1972). Bernard
Bolzano (1781-1848) was appointed professor of religious instruction in Prague
Univeristy in 1805 and was dismissed for heterodoxy in 1819.
77 The Theory of Science, p. 378.
7 8Ibid.
7 9Ibid. 380-1.
8 0Useful discussions of J.S. Mill's ideas about induction may be found in L.S.
Stebbing, A Modern Introduction to Logic, 7th edn. (London: Methuen, 1950), 331-43;
G. H. von Wright, A Treatise on Induction and Probability (London: Routledge and Kegan
Paul, 1951), 84-166; R.P. Anschutz, The Philosophy ofJ.S. Mill (Oxford: Clarendon
Press, 1953), 78-123; K. Britton, John Stuart Mill (Harmondsworth: Penguin Books,
38 The Origins of the Problem§3
Severed twentieth-century philosophers, like J . M . Keynes
and Mary Hesse, have followed Mill's example here, though
with a considerably more sophisticated appreciation of the
problems that affect such an application of probability theory to
variative induction. However, though the historical fact of
Mill's innovatory move is clear enough, it is quite another
question whether he was making a move in the right direction.
Three other moves are co-ordinate possibilities that have been
much discussed in the analysis of ampliative induction. One
such move is to retain a Pascalian framework but treat multi-
plicity, not variety, of evidential instances as crucial, as
Bernoulli did. Another is to retain a Baconian emphasis on
evidential variety, but develop an intrinsically non-Pascalian
system of gradation. And a fourth move is to construct a
Pascalian system of gradation that amalagamates support from
evidential multiplicity with support from evidential variety.
These four philosophical options will be referenced, explored,
and evaluated in § § 16-24 below.
But as a foundation for that exploration and evaluation we
need first to enquire rather more deeply into the nature of
Pascalian probability. So far we have merely noted a few facts
about the history of its mathematics, which have introduced the
reader to some basic principles of logical syntax governing the
constraints that one judgement of Pascalian probability imposes
on another: the complementational principle for negation, the
multiplicative principle for conjunction, and so on. It is clear
that an axiomatization of those syntactic principles serves to
define the concept of a Pascalian function, in the sense that any
functor which satisfies the axioms is to be considered capable of
expressing such a function. So what makes a judgement of
probability Pascalian is its logical syntax. But what is the
semantics of such a judgement? What makes it a judgement of
probability? What are we measuring when we measure a
Pascalian probability? And how is that measurement actually
1953), 147-85; a n d J . L . Mackie, The Cement of the Universe (Oxford: Clarendon Press,
1974), 297-321. Bolzano's views are discussed in L. Jonathan Cohen, 'Bolzano's
Theory of Induction', in Bernard Bolzano (1781-1848), special issue 13 of Acta Hisioriae
Rerum Naturalium nec non Technicarum (Prague: Czechoslovak Academy of Sciences,
1982), 443-57, repr. in M . H . Salmon (ed.), The Philosophy of Logical Mechanism
(Dordrecht: Reidel, forthcoming).
39 The Origins of the Problem§3 §3

carried out? Only when these issues have been resolved shall we
be in a position to investigate whether ampliative induction is
best graded in terms of Pascalian probability. And the topic is in
any case an important one since, whether or not Pascalian
probability has a central role to play in the gradation of
induction, it certainly has many other valuable uses in
contemporary culture.
II
The Controversy about the Nature of
Pascalian Probability

§ 5. S O M E G E N E R A L C O N S I D E R A T I O N S

Widely different theories have been proposed for the semantics of


probability judgements that have a Pascalian syntax. We can
distinguish roughly, but only roughly, between theories
concerned to argue how such probabilities are actually
conceived and theories concerned to argue how they ought to be
conceived, and also between theories that adopt a realist point
of view and theories that adopt an idealist one.
Pascalian mathematics, as we saw in § 3, provides a system of
syntactic constraints that judgements of probability impose on
one another—the complementational law for negation, and so
on. It thus allows us to formulate an axiomatic definition for the
concept of a Pascalian function. A function may then be said to
determine a Pascalian probability only if it satisfies the axioms of
that mathematical system. But this does not tell us, except in
certain limiting cases (e.g. p(^4&not-^4) = 0), how to evaluate
that function in particular cases, or what the truth-conditions
are for such evaluations. Nor does it even provide an adequate
definition of a probability-function, since it does not exclude
non-standard—i.e. non-probabilistic—interpretations of the
Pascalian system, such as in terms of area-ratios.
M a n y philosophers have therefore produced theories about
the epistemology or semantics of Pascalian probability, and
there is an extensive polemical literature on the subject. In order
to show the dimensions of the problem, the present chapter seeks
very briefly to sketch the main types of theory that have been
produced and the main considerations that weigh for or against
them. No attempt is made here, however, to argue that any one
such theory is superior to the rest. Instead, it will be argued later
( § § 12-15) that this eristic approach to the subject may
profitably be replaced by a more eirenical strategy in which we
§6the Nature of Pascalian Probability 41
explain and justify the existence of a variety of legitimate types of
probability judgement by exploring the wide range of syntactic
and semantic factors that are relevant to the taxonomy and
employment of such judgements. Only when this task has been
adequately carried out will we be in a proper position to discuss
(§§ 16-21) what types of probability-judgement are suitable
—and under what conditions—for employment as a mode of
inductive evaluation.
No one achieves even a one-sided philosophy of probability,
however, by making a few dogmatic, off-the-cuff statements
about the concept in the course of developing some system of
mathematical calculations or statistical procedures. At the least
arguments have to be given—arguments that support or oppose
some analysis of the concept. And even well-argued analytical
theories about the nature of probability are not always targeted
at exactly the same objective. One polar aim for such a theory is
to describe, analyse, and explain how probability is actually
conceived in human judgements. The alternative polar aim is to
prescribe how probability should be conceived and, in
particular, how the formalism of the mathematical calculus
should be interpreted. But those two objectives, though quite
different from one another in principle, are often difficult to hold
apart in practice. O n the one hand, it is natural for philosophers
of science to prefer to describe what they take ideally to be
implicit in the best usage that they can find among reputable
scientists and mathematicians, rather than whatever is actually
implicit in the imprecise and inexpert thoughts of the man-in-
the-street. O n the other hand, anyone who prescribes how
probability should be conceived presumably thinks of himself as
understanding and answering a question that relates to the
familiar existing practice ofjudging probabilities: the question is
about how this practice is to be executed, and so any answer to it
must presuppose at least some description of the presumed aims
of the practice and of the factors that constrain it. Accordingly,
though it is important to keep the descriptive-prescriptive
polarity in mind here, I shall not try to identify in every case
exactly which type of objective has been dominant in the minds
of those proposing a particular kind of theory. (It will be evident
enough, however—see pp. 64-5 below—that the 'guarded
assertion' theory, for example, is content to be descriptive in
42 The Controversy about
purpose, while Bayesian personalism, with its criterion for
rationality, is determinedly prescriptive. So at least those two
theories are not in serious competition with one another. They
answer different questions—though their adherents may well
disagree about which question is the more important one.)
Another distinction that cannot in practice always be drawn
very sharply is that between realist and idealist theories. In the
context of the philosophy of probability a theory may be
classified as a realist one if it treats probability as assessable
primarily in relation to events, objects, classes of events, classes
of objects, natural kinds, physical properties, or other similar
elements, groupings, or features of reality. A theory may be
classified as an idealist one if it treats probabilities as assessable
primarily in relation to arguments, propositions, beliefs,
assertions, or other similar elements, groupings or features of
our thoughts about reality. Here again the difference is evident
enough in some cases. To identify probabilities with the relative
frequencies of events is clearly a realist move, while to identify it
with strength of belief is clearly an idealist one. Such theories are
necessarily in conflict with one another. But the theory that is
based on the so-called 'principle of indifference', as we shall see,
has both realist and idealist versions. 1

1 A useful survey of different mathematical ideas about probability, within the


Pascalian mainstream, is to be found in T. L. Fine, Theories of Probability: An Examination
of Foundations (New York: Academic Press, 1973). See also, for a comprehensive
textbook on the mathematics of Pascalian probability, W. Feller, An Introduction to
Probability Theory and its Applications, 3rd edn. (New York: J . Wiley, 1968). A pluralist
approach to the semantics of probability, within Pascalian limits, is to be found in E.
Nagel, 'Principles of the Theory of Probability', in O. Neurath, R. Carnap and C.
Morris (eds.), Foundations of the Unity of Science, (Chicago: Chicago University Press,
1939), i. 341-422;J.J. Mehlberg, 'Is a Unitary Approach to Foundations of Probability
Possible?' i n H . Feigl and G.Maxwell (eds.), Current Issues in the Philosophy of Science (New
York: Holt, Rinehart and Winston, 1961), 287-301; H. Freudenthal, 'Realistic Models
in Probability', in I. Lakatos (ed.), The Problem of Inductive Logic (Amsterdam: North-
Holland, 1968), 1-23; and J . L . Mackie, Truth, Probability, and Paradox (Oxford:
Clarendon Press, 1973), 154-236.
§6 the Nature of Pascalian Probability 43
§ 6. I N D I F F E R E N C E T H E O R I E S

Indifference theories define the probability of an outcome as the


ratio of the number of favourable cases to the total number of
equally possible cases. They vary in their implications according
to whether equal possibility is understood in realist or idealist
terms. They avoid circularity only where this equal possibility is
supposed to be established a priori or to be demonstrable within a
well-supported theory.
What an indifference theory claims is that the probability of an
outcome may be defined as the ratio of the number of favourable
cases to the total number of equally possible cases. Such a
theory, strictly speaking, quantifies over a domain of outcome-
types, not of individual outcomes. So the probability of
throwing a four in a game of dice is j, since a dice-throw has six
possible types of outcome, of which only one is favourable for the
throw in question. Indeed, though an indifference theory
obviously cannot allow probability-functions to include irra-
tional numbers among their values, it seems particularly well
suited to the aleatory models with which the earliest mathem-
atics of probability was mainly occupied. Also we can easily
see why, if some form of indifference theory is correct, the
mathematics of probability must be what it is: the calculus of
chance is just the arithmetic of rational numbers.
But by what principle are equally possible cases to be
enumerated? The principle of indifference, as it has come to be
called, states that any cases are equally possible if there is no
reason for one such case rather than for another. And this
principle admits of both realist and idealist interpretations.
O n a realist interpretation, it means, if a monadic
probability—i.e. a one-place function p ^ ) — i s at issue, that
outcomes are equally possible if and only if there is nothing to
cause one type of outcome rather than another (as perhaps with
certain movements of a sub-atomic particle). Or, if a dyadic
probability—i.e. a two-place function p(Z? |^4)—is at issue it
means, as William Kneale proposed, that outcomes are equally
2

possible in relation to such-or-such a specified property if and


2 W. Kneale, Probability and Induction (Oxford: Clarendon Press, 1949), 116.
44 The Controversy about
only if nothing is caused by that property to have one type of
outcome rather than another.
O n an idealist interpretation, however, the principle of indif-
ference is concerned not with gaps in the causal determination of
reality, but with gaps in our knowledge about it. The principle
then means, in regard to monadic probabilities, either that cases
are equally possible if and only if there is no humanly available
knowledge why any judge of the probability at issue should
expect one type of outcome rather than another, or that, for the
present judge of the probability at issue, cases are equally
possible if and only if he or she knows no reason to expect one
type of outcome rather than another. The former of these two
idealist interpretations for the monadic function, apparently
favoured by Laplace, still implies, like a realist interpretation,
3

that on each issue there is a uniquely true probability to be


discovered. The latter interpretation, defended by Jevons and 4

Mill, has no such implication. It implies that even on the same


5

issue different judgements of probability may be correct for


different people. It is thus a subjective rather than an objective
idealism. And in regard to dyadic probabilities the tension
between a subjective and an objective idealism is preserved. An
objective idealist version of the indifference principle for the
dyadic function would state that, within such-or-such specified
evidence, there is nothing to justify expecting one type of
outcome rather than another, whereas a subjective idealist
version would state that within such-or-such specified evidence
the present judge finds nothing to justify expecting one type of
outcome rather than another.
The idealist form of indifference theory has long been accused
of 'professing to evolve knowledge out of ignorance', and the 6

realist version might with equal justice be accused of professing


to discern something evolved out of nothing. But some may

3 P S. de Laplace, A Philosophical Essay on Probabilities, trans. F. W. Truscott and F. L.


Emory (New York: Dover, 1951), 4 ff. Laplace (1749-1827) is famous for his work on
gravitational mechanisms as well as for his work on probability.
4 W.S.Jevons, The Principle of Science: A Treatise on Logic and Scientific Method (London:
Macmillan, 1883), 198 ff. Jevons (1835-82) was an economist as well as a logician.
5 System of Logic, p. 351: Mill took this view in the second edition (1846), having held a
more realist view in the first (1843).
6 See Jevons, Principle of Science p. 199.
§6the Nature of Pascalian Probability 45
think it interesting if we could evolve knowledge out of ignorance
or discern something evolved out of nothing. A more powerful
complaint against any indifference theory is therefore that,
while professing to let us do this, the theory provides no
unambiguous instructions on how we are to do it and in practice
licenses mutually contradictory conclusions.
For example, such a contradiction arises when an attempt is
made to answer the question: what is the probability of getting
heads in both of two tosses of the same coin? One line of
inference here from the indifference theory would conclude that
the answer is \ because there are four equally possible outcomes:
two heads, two tails, first heads and then tails, and first tails and
then heads. Another line of inference would conclude instead
that the answer is j because there are three equally possible
outcomes; two heads, two tails, and the combination of heads
and tails. But this kind of objection can be met by insisting that
the circumstances are not indifferent, because there is in fact a
reason to expect the heads and tails combination rather than the
two heads (or two tails) combination. The reason is that the
heads and tails combination can occur in two different ways,
depending on whether heads or tails comes first. So, if we break
down the set of possibilities into all the ultimate permutations
(see p. 15 above), a uniquely correct answer of 7 seems to be
inferable.
Even so, that answer might still be challenged. We have
established only that there is no a priori reason to expect any one
of the four possibilities rather than any other. The possibility
remains, in the case of any particular coin or tossing procedure,
that there is an empirically detectable reason to expect one of the
four ultimate permutations rather than another. On the sub-
jective idealist version of the indifference theory for monadic
probabilities—the Jevons-Mill version—this possibility does
not matter at all to a judgement of probability so long as the
author of the judgement does not know any such reason. Only
his or her actual knowledge at the time is relevant. But on any
other version of the theory for monadic probabilities, the
possibility does matter: the existence of such a reason has to be
excluded. So, to exclude the existence of such a reason, we need
to be able to judge that no factor is present—like, for example,
the existence of more wear on the heads side—that makes one of
46 The Controversy about
the four permutations more probable than the others. And
clearly, if we do need to be able to judge that, the indifference
theory would be circular. With respect to any one issue, it would
define the probability of a particular type of outcome in terms
that presuppose an appropriate set of comparative judgements
of probability, on available evidence, over all the possible types
of outcome.
At best, therefore, an indifference theory applies only where it
is either assumed or demonstrable that no such set of indi-
vidually researched judgements of comparative probability is
needed. For example, in a game like coin-tossing or dice-
throwing this is assumed by calling it a game of chance. Every
game has rules that regulate how it is played, and the rules of a
game of chance must therefore determine a priori its range of
mutually exclusive and mutually co-ordinate outcome-types. In
a normal game of dice-throwing, for example, the rules
determine a priori six possible types of outcome for any one
throw depending on which side is uppermost: a die's being
poised on an edge or a corner is not a recognized type of outcome
and voids the throw. Also in certain areas of natural science it
may be demonstrable within a well-supported theory that ideally
there is nothing to cause one type of outcome (within a certain
set of possible outcome-types) rather than another. This is the
situation in, for example, Boltzmann's interpretation for the law
of entropy (the second law of thermodynamics ). 7

But there will be a wide variety of judgements to which the


indifference theory is inapplicable. Sometimes this will be
because the various possible outcomes are not equally probable,
as with a biased coin. Sometimes it will be because the number of
possible outcomes is infinite and therefore the ratio of favourable
outcomes cannot be calculated. And sometimes, indeed often, it
will be because there is no relevant constraint on the range of
alternative possibilities. In judging the probability of heads on
the toss of an unbiased coin to be 7 a person assumes that the coin
is restricted to falling either heads or tails and will not fall on its
edge. But in judging the probability that the main course of any
student you pick out at random on the campus is in the history of
the bassoon you need not assume anything about the number of
7 See A. Pap, An Introduction to the Philosophy of Science (London: Eyre and
Spottiswoode, 1963), 210 ff.
§6the Nature of Pascalian Probability 47
main courses available. The relevant ratio seems to be that of the
number of bassoon-historians to the total number of students,
rather than that of the number of bassoon-history courses to the
total number of courses. The probability has to be established by
an empirical enumeration of students' actual choices, not by
relatively a priori research into the curriculum catalogues. It is a
ratio within a domain of individual entities, not within a domain
of types of individual entities. 8

§ 7. F R E Q U E N C Y T H E O R I E S

If the probability p(A \ B) is understood as the relative frequency


of /Is among Bs, it becomes impossible to give an account of the
probabilities of individual events or of the probability of As where
there are infinitely many Bs. Attempts have been made to deal
with the problem of infinite reference-classes by equating
probabilities with mathematical limits (in a sequence of
cumulative relative frequencies). But this makes probabilities
depend on an ordering of outcomes that is arbitrary or
irrelevant. It may also be objected that some sequences of
outcomes are accidentally cut short, before a significant relative
frequency establishes interval-valued conception of probability.

Because the indifference theory fails to cope with judgements of


probability that are supported by empirical enumerations, it is
tempting to suppose instead that probabilities are empirically
estimatable relative frequencies. The supposition then is not
merely that observed relative frequencies normally constitute
the evidence on the basis of which we estimate probabilities, but
that there is nothing more to the probability of a B's being an A
than the overall relative frequency of ^4s among Bs. If the
probability of any randomly selected student's being a bassoon-
historian, for example, is evaluated at this means just that the

8 The principle of indifference is discussed in J . M . Keynes, A Treatise on Probability


(London: Macmillan, 1921), 41-64 (where the name of the principle originates) and
81-3; B. Russell, Human Knowledge: Its Scope and Limits (London: Allen and Unwin,
1948), 391-7; K . R . Popper, The Logic ofScientific Discovery (London: Hutchinson, 1959),
168-9; G . H . von Wright, A Treatise on Induction and Probability (London: Routledge and
Kegan Paul, 1951), 228-34; and S. Blackburn, Reason and Prediction (Cambridge:
Cambridge University Press, 1973), 116-35.
48 The Controversy about
relative frequency of bassoon-historians among the students is 1
in 500. Probability-judgements thus count individual outcomes,
not outcome-types. And on such a view probabilities are
essentially dyadic, since they will always be concerned with the
frequency of some specified feature (such as that of being a
bassoon-historian) in some independently specified or specifi-
able reference-class (such as the class of students enrolled on our
local campus in 1986). They will also be essentially objective,
constituting a touchstone against which human thoughts and
reasoning can be impartially evaluated. Indeed, a frequency
conception of probability was used by Galileo when he cited the
observed results of long runs of dice-throws as a check on a priori
calculations of aleatory probabilities. It is also a great 9

convenience that for finite reference-classes, according to a


relative frequency analysis, the mathematics of probability
reduces to the arithmetic of rational numbers.
But the frequency approach has at least three weak points.
The first one is that it can offer no account of the probabilities of
individual events. In a game of chance the indifference theory
authorizes a value of 7 for the probability that heads is the next
toss's outcome-type, because in the case of any toss it authorizes
a value of 2 f ° the probability that the toss's outcome-type will
r

be heads. But the frequency analysis cannot authorize a value


of, say, 555 for the dyadic probability that George is a bassoon-
historian, given that he is a student. One reason for this is
that such a purported judgement of probability can have no
literal meaning on that analysis. It is features, not 10

propositions—whether about individuals or about anything


else—that have frequencies of occurrence relative to a reference-
class. And another reason is that the probability of George's being
a bassoon-historian, given that he is a student, may not be at all
the same as the general probability that anyone is a bassoon-
historian, given that he or she is a student. George may be very
much a special case, since his father is a well-known bassoon-
player. So that the probability of his being a bassoon-historian, if

I. Hacking, The Emergence of Probability. A Philosophical Study of Early Ideas about


9

Probability, Introduction and Statistical Inference (Cambridge: Cambridge University Press,


1975), 53.
1 0As admitted by R. von Mises, Probability, Statistics and Truth, 2nd edn. (London:
Allen and Unwin, 1957), 11.
§6the Nature of Pascalian Probability 49
it is to be equated with any relative frequency at all, can be
equated only with the relative frequency of bassoon-historians
among students whose fathers are well-known bassoon-players.
And then, since George is the only such student, we can
apparently not learn the probability that he is a bassoon-
historian without learning whether or not he is in fact a bassoon-
historian. But this would be unacceptably paradoxical, because
we should obtain a reliable probability for George's being a
bassoon-historian if and only if we are 100 per cent certain that
he is one or that he is not one. In short, a frequency analysis not
only allows no literal meaning to judgements of individual
probability but also creates difficulties for any attempt to treat
them systematically as covertly general judgements.
Nor is it open to a frequency theorist to claim that all import-
ant probabilities are indeed general, not singular. It often
seems very important to be able to calculate the probability of
success for your own child's appendectomy, the probability of a
bull market in equities next spring, or the probability of rain
tomorrow.
The frequency approach's second weak point is that it
encounters serious problems in dealing with infinite reference-
classes. If a coin were tossed for ever, and the tosses remained
fair throughout, then not only would the number of tosses be
infinite but also the number of tosses that landed heads. (If,
instead, in an infinite sequence of tosses the number that landed
heads were finite, then, however large that number was, the
probability of heads would be infinitely small—which conflicts
with our assumption of fair tosses.) So the probability of a toss's
landing heads, which is presumably j, cannot be equated with
the relative frequency of heads simpliciter, i.e. with And
analogously, if student enrolments go on for ever, while the
usual variety of courses are favoured, the probability of a
student's being a bassoon-historian cannot be equated with the
relative frequency of bassoon-historians. There is a risk that this
fact may be obscured by the use of ambiguous expressions like
'in the long run'. In one sense this means 'in a very large, but
finite, set of cases', in another it means 'in an infinite set of
cases'. In the former sense, a relative frequency in the long run
may exist, in the latter it cannot.
In the case of endlessly repeated trials at a game of chance,
50 The Controversy about
where such repetition is assumed to have no effect on the
physical state of the coin, die, etc., there is a plausible way out of
the difficulty, adopted by S.D. Poisson and others. If the 11

reference-class consists of the outcomes of such trials, it is well-


ordered by their succession in time: each member of the
reference-class has a unique position in the temporal ordering.
So there is an infinite succession of outcomes, and the
probability of a certain kind of outcome can apparently be
equated, as the number of outcomes goes to infinity, with the
mathematical limit of the sequence of relative frequencies of that
kind of outcome. That is, the probability p(A \B) can be equated
with the limit of the sequence of cumulative relative frequencies
of As among Bs that consists of the frequency (which will be 0 or
1) in the first outcome, the frequency (which will be 0, 7, or 1) in
the first two outcomes, the frequency in the first three outcomes,
and so on. (An infinite sequence of functions r r , r . . . has a lt 2 3

limit r if and only if for as small a positive real number as you


L

please, e, there is for each member of the sequence r a later (

member r such that r does not differ by more than e from r .)


s ] L

However, as Bertrand Russell pointed out, the same infinite 12

set of fractions can be arranged into different infinite sequences


that converge to different limits. Russell took as his example the
probablity that an integer chosen at random will be a prime. If
we take the integers in their natural order, then, as we go to
infinity, the limit of the sequence of cumulative relative
frequencies of primes is zero. But suppose we rearrange the
integers by putting first the first 9 primes, then the first number
that is not a prime, then the next 9 primes, then the second
number that is not a prime, and so on indefinitely. When the
integers are arranged in this order, the limit of the sequences of
cumulative relative frequencies of primes is And Russell also

1 1 S.D. Poisson, Recherches sur la probability des jugements en matiere criminelle el en matiere
civile precedees des regies generates du calcul des probability (Paris: Bachelier, 1837). Poisson
(1781-1840) was a friend of Laplace. His most important work was in the application of
mathematics to physics. See also H. Reichenbach, The Theory of Probability, trans. E. H.
Hutten and M. Reichenbach, 2nd edn. (Berkeley: University of California Press, 1971),
68-9.
B. Russell, Human Knowledge pp. 384-5 A. Pap, Introduction to the Philosophy of
1 2

Science, pp. 180-1, contradicted Russell, but his argument is not cogent.
§ 6 the Nature of Pascalian Probability 51
showed how by yet another arrangement of the integers the limit
in question would be 1. So it is clear that, if a probability is
defined in terms of the limit to a sequence of relative frequencies,
it might differ under certain rearrangements of the underlying
set of outcomes. But the probability that a B is an A should be
unique. So Poisson's procedure makes the probability p {A\B)
depend on there being only one relevant ordering for the class of
Bs, which is the reference-class for the probability.
Now the trouble is that the members of some reference-classes
do not all obviously belong to some set that is well-ordered by
temporal succession or by some other uniquely appropriate
relation. Thus, if the class of student enrolments is an infinite
one, some of its members may be simultaneous with one
another, so that it is certainly not well-ordered by temporal
succession. Moreover, even when the reference-class is well-
ordered by some supposedly appropriate relation, the sequence
of cumulative relative frequencies may not in fact converge to a
limit. Or—and this is the commonest situation—there may be
infinitely many other equally appropriate well-orderings of the
class, each of which determines a different limiting value for the
sequence of relative frequencies in question.
A possible way of dealing with those difficulties would be to
identify the probability with the limit to which the relative
frequencies converge within an infinite-membered sub-class of
the reference-class that satisfies appropriate conditions for
randomness and convergence, as proposed by von Mises. But 13

this proposal still raises the question why any ordering should be
relevant to the value of the probability, within a set of outcomes
that are independent of one another. If the outcome of a toss is
supposed to be unaffected by its position in a sequence of tosses,
because repeated tosses are assumed to have no effect on the
state of the coin or of the tossing mechanism, why should the
ordering of the sequence be relevant to the probability of heads
as against tails? Moreover, a statement evaluating the limit of a
sequence of relative frequencies hypothetically continued to
infinity implies nothing whatever that is empirically testable
about any initial segment of the sequence, however large,
13 See R. von Mises, Probability, Statistics and Truth, 2nd edn. (New York: Dover,
1957), 24-5.
52 The Controversy about
because such a segment can be replaced by any arbitrary
sequence of the same length without affecting the limit of the
sequence as an infinite whole.
It might be thought that these problems about infinite
reference-classes at least leave the frequency analysis safely
available in the case of finite reference-classes. And, after all,
very many of the probabilities with which we are concerned in
the natural or social sciences are probabilities in supposedly
finite classes, although such classes are often indeterminately
large. But a third weakness of the frequency analysis seems to be
clearly displayed in certain cases where the reference-class is
certainly finite in size. One may well believe that a nicely
produced coin, fresh from the mint, has a probability of falling
heads that is correctly evaluated at j. But suppose it is tossed
three times, landing heads twice and tails once, and is then
melted down for scrap. A crude relative frequency theory would
apparently evaluate the probability of its falling heads at f,
which cannot be correct. Perhaps a frequency theorist has
therefore to say instead in such a case that the probability of
heads is not to be identified with the actual relative frequency of
heads, but with what that relative frequency would have been in
the long run. But then either that means 'in an infinite set of
tosses', in which case the difficulties already mentioned apply;
or it means 'in a very large, but finite, set of actual or potential
tosses', in which case the probability would still be altered a bit if
the coin were tossed just one more time, since the relative
frequency of heads would inevitably increase or decrease,
depending on the actual outcome of that additional toss. So
where the reference-class is finite the frequency analysis seems to
make the exact value of a probability depend on some quite
accidental fact about the precise number of instances that are
supposed to belong to the reference-class, just as where it is
infinite the probability is made to depend on some quite
accidental fact of ordering.
Of course, where the reference-class is finite we can escape
introducing such a dependence on irrelevant contingencies if we
are willing to take the probability-statement in question to be
saying that the actual or potential relative frequency falls within
a specified interval (e.g. 0.5 plus or minus 0.16). But the
§ 6 the Nature of Pascalian Probability 53
difficulties about individual events and about infinite reference-
classes cannot be so easily dodged. 14

§ 8. P R O P E N S I T Y T H E O R I E S

If p(A I2J) is to be thought of as a propensity, this cannot in every


case be the propensity of B to cause A. But a propensity
interpretation of the probability calculus does allow probabilities
to be assigned to individual outcomes. The main weakness of
such an analysis is that it does not supply any intrinsic guidance
in regard to the actual evaluation of probabilities or any obvious
rationale for the principles derivable within the calculus of
chance.
If probabilities are indeed objective features of reality, but are
not to be identified with some collective property of the
reference-class, like the property of having a certain relative
frequency of members that have such-or-such a characteristic,
perhaps instead they are related distributively to each member
of this class. The j probability that a coin-toss will fall heads is
then a physical propensity that is operative in each coin-toss
rather than a relative frequency that characterizes the totality of
coin-tosses. O n this view probabilities are displayed in, or
15

evidenced by, the relative frequencies observed in appropriate


samples, but are not to be identified with them or with relative
frequencies 'in the long run'. Relative frequency analyses are to
be looked on as the outcome of empiricist prejudice, confusing
the meaning of a statement with its evidential warranty.
Indeed, unless we think of a coin's objective probability of
falling heads as a measurable propensity rather than as a relative
frequency, it would not seem reasonable to assign a point-value
1 4Analyses of probability in terms of relative frequency are expounded or discussed in
J. Venn, The Logic of Chance, 2ndedn. (London: Macmillan, 1876); Reichenbach, Theory
of Probability, Pap, Introduction to the Philosophy of Science; W . C . Salmon, The Foundations of
Scientific Inference Pittsburgh: University of Pittsburgh Press, 1966), 83-108; R.B
Braithwaite, Scientific Explanation (Cambridge; Cambridge University Press, 1953),
115-95; and D. Gillies, An Objective Theory of Probability (London: Methuen, 1973).
15 D . H . Mellor, The Matter of Chance (Cambridge: Cambridge University Press,
1971), 63-82.
54 The Controversy about§11
(as distinct from an interval-value) to such a probability in the
case of a coin that is melted down very soon after manufacture,
perhaps after only three tosses. Nor would it seem possible to
assign an irrational number as the value of a probability within a
finite reference-class, unless we think of that probability as a
propensity rather than a relative frequency. Yet irrational
numbers enter very easily into our reasonings about
probabilities. For example, when we judge the joint occurrence
of two independent and equiprobable events to have a
probability of 7, we imply that the probability of each event on
its own is 1/V2, because of the multiplicative principle for
conjunction.
Sometimes, where dyadic probabilities are at issue, it may be
tempting to treat such propensities as measurable degrees of
causal power. For example, this looks appropriate in the case of
the probability that a match will light, given that it is struck.
Striking the match may be thought of as having a certain degree
of efficacy in regard to the match's lighting. But there are also
many kinds of case where probabilities cannot be conceived in
this way. For example, when we speak of the probability that a
person elected to Parliament was educated at a university, or of
the probability that the accused is guilty given his demeanour in
the witness-box, we are certainly not in either case measuring
the power of one event to cause another, since, if we were, we
should be supposing that a later event can cause an earlier one.
So what are propensities? Popper has argued that they are a
kind of physical property. He explicitly compares a field of
propensities, as 'a dispositional property of singular physical
experimental arrangements', with a field of forces. He thinks 16

of his account of propensities as 'a new physical hypothesis (or


perhaps a metaphysical hypothesis), analogous to the hypothesis
of Newtonian forces'. Such a conception might be used, for
17

example, in measuring the probability of a radioactive particle's


decaying within some specified period. And it associates
naturally with a rejection of determinism. If physical
dispositions of this kind exist in nature, some events are
inherently unpredictable.

1 6 K . R . Popper, Realism and the Aim of Science (London: Hutchinson, 1983),


351.
1 7 Ibid. 360.
§6the Nature of Pascalian Probability 55
But, if this were the only appropriate way to conceive
propensities, there would be very many kinds of probability-
judgement to which the propensity account would not apply.
For example, we could not conceive thus the backward-looking
probability that a person elected to Parliament was educated at a
university or even the forward-looking probability that a fifty-
year-old politician will survive to the age of seventy. Many
issues in probabilistic reasoning are not concerned with
fundamental physical properties or their inherently unpre-
dictable manifestations, but with statistically evidenced tend-
encies that are taken to result from the operation of just some
out of the many different causal factors that combine to deter-
mine the actual outcome. In such cases to declare the probability
of A given B is merely to evaluate the contribution or sig-
nificance towards A of some potentially operative factor or set of
factors, B.
Indeed, we have here an analogy between the indifference
theory, in its realist presentation (see pp. 43-4), and the
propensity theory. O n causal issues both kinds of theory have a
strong, monadic version that takes the occurrence of one
outcome rather than another, with a probability less than 1, as a
manifestation of indeterminism, and both theories have a
weaker, dyadic version that is compatible with treating the
actual outcome as a fully determined one.
All versions of the propensity theory have two important
features that no relative frequency account can have. They
apply just as well to infinite populations as to finite ones. And
they all allow probabilities to be assigned to specified individual
events, since propensities may be treated as properties either of
what Popper calls 'singular physical experimental arrange-
ments' or of individual set-ups of some other kind. More
strictly, perhaps, we can say that, as a conditional probability,
propensity is a second-order relation—a relation of probab-
ilification that links one first-order property to another. In virtue
of that fact we can express a general probability—a probability
that does not refer to any individual object, event, or set-up—as
a propensity, like the propensity of a fifty-year old politician to
survive to the age of seventy. But, because individuals have first-
order properties, we can also express many singular judgements
of probability in terms of propensities, such as the propensity
that Smith will survive to the age of seventy in virtue of his being
56 The Controversy about§11
at present fifty years old. The propensity then measures the
strength of the link between one first-order property that Smith
might have and another. And singular judgements of uncondi-
tional probability lend themselves to a similar form of analysis.
If there is a 0.8 probability of such a kind that Smith will survive
to the age of seventy, we must suppose a 0.8 propensity in
Nature for this to happen. Propensities are then being treated
either as properties of properties or as properties of events.
The main weakness of a propensity analysis is that it does not
intrinsically carry with it any distinctive type of guidance in
regard to the actual evaluation of probabilities. The indifference
theory tells us to calculate the ratio of the favourable to the
equally possible cases. It may sometimes be difficult to
determine the implications of this requirement coherently and
unambiguously. But at least in standard games of chance the
theory can be made to work as a basis for assigning probability-
values, and it also provides an obvious rationale for the
principles that are derivable within the calculus of chance—the
multiplication law for conjunction, the complementational law
for negation, and all the other principles that fit an aleatory
model. Again, the relative frequency theory states a ratio that we
need to estimate. It may sometimes be difficult to do this reliably
because of the large size of the reference-class and the small size
of the available sample. But at least the theory tells us in
numerical terms what we ought to find out (where the reference-
class is finite), and it too provides us with an obvious rationale
for the mathematics of probability. The propensity analysis, on
the other hand, speaks the language of causality or of physics,
not of ratios. Correspondingly, since the talk about propensities
has no distinctive numerical implications, it provides no
inherent basis for the assignment of actual probability-values.
And it also provides no wholesale rationale for the principles
derivable within the calculus of chance. 18

1 8 Propensity analyses of Pascalian probability are further defended or discussed in


C.S. Peirce, Collected Papers, ed. C. Hartshome and P. Weiss (Cambridge, Mass.:
Harvard University Press, 1932), 409-14 (paras. 2. 662-8); Popper, Realism and the Aim
of Science, pp. 281-401; I. Hacking, Logic of Statistical Inference (Cambridge: Cambridge
University Press, 1965), 1-26; A.R. White, 'The Propensity Theory of Probability',
British Journal for the Philosophy ofScience, 23(1972), 35-43; T. Settle, 'Propensity Theories
of Probability Unscathed: A Reply to White', British Journal for the Philosophy of Science,
§6the Nature of Pascalian Probability 57
Indeed, a propensity theory may even be seen, in certain
contexts, to suggest one or two non-Pascalian principles. For
example, it is not at all obvious that propensities have to obey the
complementational law for negation, if we just think of them
intuitively and forget for the moment what we know about the
mathematics of Pascalian probability. Drinking a cup of tea a
day may have only a small propensity (or a slight tendency or a
weak disposition) to make a person well, but it certainly does not
have a correspondingly large propensity (or a big tendency, or a
strong disposition) to make him or her unwell: it just has very
little effect on health in either direction. No doubt we can
regiment our thoughts about propensities, so that they always
conform to the mathematics of probability. But in its naive and
unreconstructed form the concept of a propensity will not
necessarily lead us to draw inferences that are isomorphic with
those which the Pascalian concept of probability leads us to
draw. Conversely, the crucial issue is whether in comparing
p(^4|5) with p(not-^4|jB) we take ourselves to be comparing the
extent of B's propensity for A with the extent to which B has no
propensity for A or with the extent to which B has a propensity
for not-^4. In the former case our logic of propensity needs an
additive, complementational principle for negation, in the latter
it needs to avoid such a principle. Compare what was said
19

earlier about credibility (pp. 17-18).


This important issue will be discussed more fully in § § 14 and
15 below. But one possible objection to what has just been said
needs to be considered here. It might be claimed that there is a
mathematically consistent way of describing two weak, yet
opposing propensities in terms that are both Pascalian and non-
additive: we can just say that y>{A\B) does not differ much from

23 (1972), 331-5; L. Sklar, 'Is Probability a Dispositional Probability?', Journal of


Philosophy, 67 (1970), 355-66; H.E. Kyburg, 'Propensities and Probabilities', British
Journal for the Philosophy of Science, 25 (1974), 358-75; T. Settle, 'Induction and
Probability Unfused', in P. A. Schilp (ed.), The Philosophy of Karl Popper (La Salle: Open
Court, 1974), 722-49; P. Suppes, 'Popper's Analysis of Probability in Quantum
Mechanics', ibid. 760-74; P. Clark, 'Determinism and Probability in Physics',
Proceedings of the Aristotelian Society, supplementary vol. 61 (1987), 185-210; and J.
Butterfield, 'Probability and Disturbing Measurement', ibid. 211-34.
19 See also P, Humphreys, 'Why Propensities Cannot be Probabilities', The
Philosophical Review, 94 (1985), 557-70.
58 The Controversy about§11
p(J) and that p(not-^4| B) does not differ much from p(not-/l).
J

The strength of the propensity-type connection between B and A


is thus viewed as the difference between the conditional and
unconditional probabilities of A—often called the 'relevance' of
B to A—and the strength of the propensity-type connection
between B and not-^4 is viewed analogously as the difference
between the conditional and unconditional probabilities of
not-A. And obviously these two differences can both be quite
small without any transgression of Pascalian mathematics. (For
example, let p(AIS) = 0.55 and p(not-y4|fi) = 0.45, while p(^4)
= p(not-/4) = 0.5.) However, the price paid for this manoeuvre
is that, since propensities are defined by it only in terms of the
difference between dyadic and monadic probabilities, no sense
is given to statements that are just about monadic probabilities
or just about dyadic ones, and not about the difference between
the former and the latter. And an analogous argument applies to
any other attempt to treat propensities as complex functions of
probabilities rather than as probabilities simpliciter. The original
purpose of the propensity theory—to provide an adequate
interpretation for all well-formed formulae of the Pascalian
calculus of probability—is not attainable when propensities are
treated in that kind of way.

§ 9. P E R S O N A L I S T T H E O R I E S

Ramsey held that a person's judgement of probability measured


his degree of rational belief, where degree of belief is manifested
by the lowest betting-odds on A that he would accept within a
coherent system of wagers. But Ramsey was mistaken in
thinking that p(A IB) can then be interpreted as expressing the
lowest odds at which the person would now bet on A, with the bet
only being valid if B is true. Moreover, someone who holds that a
person's judgement of probability measures his degree of
rational belief may hold one or other of several different views
about the linguistic function of sentences that evaluate
probabilities. A Ramsey-type analysis of probability-judge-
ments is in practice often combined with endorsement of a
Bayesian strategy for evaluating them. But problems arise about
the extent to which different persons' evaluations can then be
expected to converge.
§6the Nature of Pascalian Probability 59
On a realist account probabilities concern events, classes of
events, properties, or other features of reality. Yet in the
ordinary speech of both scientists and laymen there is often talk
about the probability with which a certain proposition or belief is
true or a certain inference valid. A realist may wish to claim that
all probability-judgements that are formulated in this way could
just as well be formulated in terms of the corresponding features
of reality that underlie them. But there are idealist theories that
offer a more idiosyncratic set of truth-conditions for statements
of probability-assignments—that is, a set of truth-conditions
that has no equivalent in realist terms.
Thus Ramsey held that a person's judgement of probability
20

measures his or her degree of rational belief, where the


measurement relates not to any kind of introspected feeling of
belief-intensity but to the kind of action that would in suitable
circumstances result from the belief. Roughly, a person's belief
that/) is to be measured by the lowest odds he will accept on a bet
that p. If the lowest odds he will accept are n to m, his degree of
belief is to be measured by the relative size of what he is prepared
to lose if he is wrong, that is, by the ratio of his own stake to the
sum of both stakes, m/(n + m). Thus if George will accept the
offer of a bet at 3 to 1 that it will rain tomorrow, but not at 2 to 1
or at any other odds lower than 3 to 1, then he must have a \
degree of belief that it will rain then. Or, to put it another way,
George is so doubtful about there being any rain tomorrow that
he will risk losing £1 in a bet on it only if he stands to gain at least
£3 if there is no rain. But Ramsey recognized that such a bet
should not be assumed to be made in monetary units, partly
because of the diminishing marginal utility of money (an extra
£3 is not worth as much to a millionaire as to a pauper) and
partly because a particular person may be abnormally attracted,
or abnormally averse, to monetary gambles. He therefore
specified that the bet should be supposed to be made in terms of
whatever measurable values or utilities the bettor ultimately
desires, and he was willing to accept that it might be an
idealization to suppose the actual existence of such ultimate
values.
Ramsey also recognized that degrees of belief, measured thus
F.P. Ramsey, The Foundations of Mathematics and other Logical Essays (London:
2 0

Routledge and Kegan Paul, 1931), 156-98.


60 The Controversy about§11
in terms of a policy of conduct, are not necessarily degrees of
rational belief. For example, perhaps someone will try to get me
to take bets both on its raining tomorrow and on its not raining. I
am to take a bet at odds of 1 to 2 that it will rain and at the same
odds that it will not rain. Then, if it rains, I shall gain only one
unit on the first bet and lose two on the second, and, if it does not
rain, I shall lose two on the first and gain only one on the second.
What is called—in race-track jargon—a 'Dutch book' will have
been made against me. That is, a set of bets is registered on
which, whatever happens, I am bound to make an overall loss.
So because a rational person would not wish to behave in such a
way as to ensure a loss in relation to his or her ultimate values,
Ramsey inferred that a person's system of beliefs at any one
moment is rational only if it is such that a Dutch book cannot be
made against it. And he showed that a set of degrees of belief that
satisfied this condition would also satisfy the laws of the
mathematical calculus of probability—a theorem that was later
independently proved by Bruno de Finetti (who described such a
set of probability-judgements as 'coherent'). 21

Ramsey's analysis was the archetype of a group of theories


that are sometimes called 'personalist', sometimes 'subject-
ivist', and sometimes (for reasons that will emerge) Bayesian.
Such an approach applies most naturally to the interpretation of
monadic probability-functions. Those functions can readily be
supposed to evaluate degrees of belief unconditionally by
reference to lowest acceptable betting odds. But what about
dyadic functions? Ramsey himself at one point interprets a
person's dyadic probability p{A\B) as expressing 'the odds at
which he would now bet on A, the bet only to be valid if B is
true'. And an obvious trouble, unnoticed by Ramsey, is that
22

this interpretation in terms of a conditional bet would allow a


Dutch book to be made against certain otherwise unexception-
able assignments of equal values to p{A\B) and p(not-^4|C), if B
and C both turn out to be true. Consider, for example, a case in
which A is 'It will rain tomorrow', B is 'The barometer is fall-
2 1See B. de Finetti, 'Foresight: Its Logical Laws, Its Subjective Sources (1937)', in
H. E. Kyburg and H. E. Smokier (eds.), Studies in Subjective Probability (New York: John
Wiley and Sons, 1964), 99-158.
2 2Ramsey, Foundations of Mathematics, p. 180. Ramsey's full definition here is too
complex to be enlightening.
§6the Nature of Pascalian Probability 61
ing', and C is 'The wind is in the east', where the bets would be 1
to 2 on A, if B is true, and 1 to 2 on not-^4, if C is true. And there
would also be trouble if, as might well be the case, p(^4|5) were
said to be not equal to 1C). For then what would be asserted
on Ramsey's interpretation would be, if B and C were both true,
a self-contradictory statement about what are the lowest accept-
able odds in a bet on the truth of A. Those odds on A would be
said to be not equal to those odds on A.
We must therefore in any case reject Ramsey's interpretation
for dyadic probability functions. It has devastatingly counter-
intuitive consequences. Indeed Ramsey here seems to have been
confounding statements of the form 'p(^4|£) = n with
statements of the form 'If and only if B is true, then p(A) = n',
and to have been interpreting the former kind of statement in a
way suited only to the latter. Nor have other personalist
philosophers of probability, working independently of Ramsey,
always been at their most clear-headed on this issue. Parts of de
Finetti's writings, for example, are hit by the same paradoxes as
Ramsey's analysis, though L.J. Savage was more careful.
23 24

Someone might try to defend the Ramsey-de Finetti


interpretation of dyadic probability-functions by pointing out
that it will work quite satisfactorily in practice if the proposition
upon which the validity of the bet is made conditional is required
not only to be true, but also to state the only additional fact that
must become known, if the condition's fulfilment is to be
25

warranted to be assertable. This manoeuvre will certainly


preclude generating the kind of incoherence or contradiction to
which Ramsey's analysis was shown above to be exposed. You
cannot have B and C stating different things and yet both, as
premisses, stating the only additional facts to become known. It
may still seem pradoxical, however, that the determination of
betting-odds relating to a conditional probability is made
dependent on such an unrealistic assumption as that in ordinary
life items of knowledge can come one at a time in neatly
packaged separation from one another, so as to allow only B to
be known additionally, or only C. The world in which we
De Finetti, 'Foresight', p. 108.
2 3

L.J. Savage, The Foundations of Statistics (New York: Dover, 1972), 44.
2 4

This is the analysis adopted by de Finetti in his Theory of Probability, trans. A. Machi
2 5

and A. Smith, i (London: J . Wiley, 1974), 134.


62 The Controversy about § 11
actually make bets is not like that and, if it were, so much might
be different that it is very difficult to predict how we should be
inclined to bet on a particular issue. What perhaps one wants to
be able to say here, therefore, is that p{A\B) is the probability
assigned to A where B states the only relevant additional fact to
become known. But any reference to relevant evidence makes
the interpretation circular, since relevance must itself be defined
in this context in terms of dyadic or conditional probabilities. A
true proposition B states all the evidence relevant to another
proposition A only if there is no true proposition C, such that C
does not logically entail A and p(A \B) is not equal to p(A | BlkC).
However, there is an obviously safe way out of these
difficulties. We have in any case, on a personalist view, to treat
p(A | B) as measuring the ratio of the lowest odds acceptable
on A&.B to the lowest odds acceptable on B alone, since (as we
have seen in § 3) the mathematics of the Pascalian calculus gives

where p(B) > O. We can, if we like, treat this ratio as evaluating


the strength of a person's belief in the truth of A, given just B as
evidence, that is, the strength of belief in the truth of A that is
attributed to the truth of B, and the ratio would thus constitute a
measure of conditional belief. Both monadic and dyadic
judgements would then express belief-intensity—unconditional
and conditional, respectively. But, whether conditional or
unconditional, belief-intensity would on this interpretation
always be measured in terms of unconditional bets.
We should note here an important shift in the question that is
being answered, when we compare a subjective idealist theory,
like Ramsey's or de Finetti's, with a realist theory, like those
discussed in § § 6-8. Both types of theory purport to analyse the
concept of probability. But, whereas the realist theories offer
analyses that might help to answer the question 'How am I to
reason about probabilities?', a Ramsey-de Finetti type of
personalist theory is directed rather towards the question 'How
am I to reason about a person's probability-judgements?'
Realist theories aim to clarify the truth-conditions for assigning
§6the Nature of Pascalian Probability 63
particular values to the probability-function in relation to
particular events or kinds of events. But a Ramsey-de Finetti
theory tells us the truth-conditions for statements about what
values a particular person actually assigns to his probability-
function. If you want to know instead what a man's probability-
assignment means, Ramsey and de Finetti can tell you only that,
if sincere and prudent, it expresses the man's degree of rational
belief, and that answer supplies no reason at all why two
different people ought properly to have the same degree of
rational belief about the same issue if given the same evidence.
So far as the Ramsey-de Finetti theory goes, two people could be
equally rational and yet assign very different probabilities to the
same event, on the same evidence, provided that each person's
system of probability-judgements was internally coherent (in
the sense that a Dutch book could not be made against him).
Correspondingly, there can be no valid arguments that one
assignment of this type is correct and others incorrect, whereas
the objectivist theories do—each in its own way—allow such
arguments.
It follows that a belief-theoretical account of probability-
judgements could be combined with a realist one, although this
possibility was not envisaged by Ramsey himself and was
explicitly rejected by de Finetti. The realist theory could be
26

taken to disclose the meaning of a person's probability-


statements, while the belief theory could be taken to show how
we may test the sincerity and prudence with which those
statements are asserted. Of course, a person may pretend to
have beliefs or partial beliefs that he does not have, by accepting
wagers in public that run counter to his own private judgement.
But, if he is not so pretending, and is sufficiently clear-headed,
then, if the lowest odds he will accept on the outcome A (in a
coherent system of bets) are 3 to 1, this could be taken by belief-
theorists as showing the strength of his belief in A when he says,
according to a realist interpretation, that the (probabilistic)
propensity for A, for example, is j.
There are, however, at least three possibilities open to
someone who rejects a realist analysis of probability-judgements
and wishes to treat Ramsey's type of theory as providing an
2 6 De Finetti, 'Foresight', p. 154
64 The Controversy about§11
account not just of partial belief but also of the way in which
sentences about probability function in human discourse.
One possibility is to treat the utterance of these sentences not
just as an indication of what the speaker believes, and of how
strongly he believes it, but as purporting to describe this. To say
'The probability of A is 7' is then equivalent to saying ' M y own
degree of belief that A is 7'. Such a view is rather similar to
Hume's analysis of moral judgements:
When you pronounce any action or character to be vicious, you mean
nothing, but that from the constitution of your nature you have a
feeling or sentiment of blame from the contemplation of it. 27

But this type of analysis, analogously to Hume's, suffers from an


obvious disadvantage. It implies that in disputing a person's
spoken judgement of a probability one must be disputing the
accuracy or correctness with which he reports his own state of
belief. And that cannot be what advocates are arguing about in
court, for example, when they differ about the probability of the
defendant's having fired the fatal shot, any more than a dispute
about the alleged viciousness of blood sports is a dispute about
the actual feelings or sentiments of the disputants.
A second possibility is to suppose that mental states of partial
belief are matched by speech-acts of partial commitment. The
analysis is now a contribution to the pragmatics, rather than the
semantics, of word-use. T o say that there is a f probability of
rain on any day in April is now just to hedge one's prediction, as
Toulmin claims, and describes neither the speaker's internal
28

state of mind nor the relative frequency with which rain occurs
in April, though willingness to bet at the corresponding odds
would again be a test of sincerity. And such a view too has a
counterpart in the analysis of moral discourse, when it is urged,
as by Hare, that words of moral evaluation, like 'good' and
29

'bad', should be understood in terms of the characteristic


speech-acts that they perform. 'Good', for example, is to be
analysed as an adjective of commendation. But it is easily
objected that this kind of analysis does not work satisfactorily
2 7 D. Hume, A Treatise of Human Nature (London: J. Noon, 1739-40), bk. Ill, pt. 1,
sect. 1, ed. L. A. Selby-Bigge (Oxford: Clarendon Press, 1888), 469.
2 8 S. Toulmin, The Uses ofArgument (Cambridge: Cambridge University Press, 1958),
44-93.
2 9 See R . M . Hare, The Language ofMurals (Oxford: Clarendon Press, 1952), 79-150.
§6the Nature of Pascalian Probability 65
when the words in question, like 'probable' or 'good', occur in
the antecedents of conditional sentences or in any other
sentential context in the utterance of which no speech-act of the
relevant kind is normally performed. Nor can the speech-act be
30

replaced in such contexts by its own description, since the


statement 'If I were to assert guardedly that it will rain this
afternoon, then I should be excessively cautious in my weather
predictions' may well be true when the statement 'If it were
probable that it will rain this afternoon, then I should be
excessively cautious in my weather predictions' was false.
There is also a third way in which judgements of probability
may be treated as a measure of partial belief. Instead of treating
a person's willingness to accept certain betting-odds as an
indication of his belief-strength, or as what he is asserting, or as
the expectable associate of a guarded commitment, one might
suppose that judgements of probability are intended to measure
the lowest odds that a rational bettor should accept. Sentences
about probability are then characteristically normative or
prescriptive in relation to partial belief. Indeed, this version of
31

a Ramsey-type theory is the most defensible one, though


perhaps it should not be called a 'personalist' analysis, as
Ramsey-type theories are often called. It is an objective
idealism, rather than a subjective idealism. It avoids the
difficulties that face the self-descriptive and speech-act analyses,
without reverting to a realist semantics. But it still has at least
three important limitations which it shares with all Ramsey-type
theories.
First, it cannot apply to that aspect of belief in which (as we
saw on pp. 17-18 above) a person's degree of belief that not-^4 is
not necessarily the complement of his degree of belief that A.
The commonest kind of situation in which this non-
complementationality seems to operate is when a person thinks
himself to be rather short of relevant evidence and is therefore
inclined to have only a weak belief that A along with only a weak
belief that not-A. The immediately available evidence may even
be strongly in favour of one conclusion rather than another, but
also contain so small a number of the relevant facts that no
3 0C o m p a r e J . R . Searle, 'Meaning and Speech Acts', Philosophical Review, 71 (1962),
423-32.
3 1See e.g. P. Horwich, Probability and Evidence (Cambridge: Cambridge University
Press, 1982), 32-5.
66 The Controversy about§11
strong belief is justified. In that kind of situation a rational
person may well be reluctant to accept any bet at all that A or that
not-^4. It is not just that he would not, or should not, accept a bet
of £1 to £1 on a coin's falling heads if he does not have reason to
believe that the coin would be unbiased and the toss would be
fair. In such circumstances he would not, or should not accept a
bet on heads at any other odds either. So, where the weakness of
justified belief is due to the overall inadequacy of the given
evidence rather than to the fact that some of it favours one
conclusion and some another, one cannot measure degree of
justified belief by the lowest acceptable odds. (Note that this
limitation on the scope of a Ramsey-type theory operates both in
regard to monadic probability-functions, which the theory takes
to evaluate unconditional degrees of belief, and also in regard to
dyadic ones, which the theory takes to evaluate conditional
degrees of belief.)
The second limitation affects the strength of beliefs that may
be evaluated by reference to the lowest acceptable odds on their
truth. Clearly betting odds cannot be used to measure either of
the two limiting cases. Where a person supposes a probability of
1, or a probability ofO, that such-or-such is the case, the strength
of his justified belief cannot be represented by appropriate
betting odds, because both sides to a wager need to have
something at stake. And it is no good objecting that a person's
certainty about an outcome is adequately indicated by his
unwillingness to risk any wager on it, since as we have just seen,
such unwillingness is also appropriate where relevant evidence
is in short supply.
The third limitation affects the context of beliefs. A rational
person does not accept bets on issues that can never be settled.
So beliefs about the past present a special problem for Ramsey-
type theories. Knowing that literary or archaeological evidence
would eventually be available to settle the issue, a rational
person might bet on whether or not a Roman legion once
marched along Fleet Street. But suppose that all the relevant
evidence which there will ever be is already available. Then one
could never settle abet that was conditional on this evidence. We
have no time-machine to take us back to look at the actual event.
Yet there does seem to be a determinate probability, on the
evidence, that such an event actually occurred. And a similar
§6the Nature of Pascalian Probability 67
difficulty arises in regard to any beliefs that such-or-such
a generalization states a law of nature. An open-ended
generalization of that kind might be conclusively falsified
(though even this is doubtful: see pp. 181-2 below). But it could
certainly never be conclusively verified. So a bet on its truth
could never be won (and perhaps never lost either). Of course,
you could imagine the existence of an omniscient umpire who
would kindly arbitrate between human bettors. But would it be
rational to undertake any commitment on the assumption that
such a being exists? If not, it would be irrational to bet about
issues that humans cannot settle. Yet it does not seem irrational
to have partial beliefs about at least some such matters.
The term 'Bayesianism' is often used as a name for some
version or other of what I have here been calling a 'Ramsey-
type' theory. Ramsey himself did not introduce any reference to
Bayes's law (p. 23 above) into his writings on probability. But
de Finetti and Savage have exploited Bayes's law in the attempt
to develop a methodology for any coherent pattern of
probabilistic reasoning that employs a personalist analysis of
probability-judgements. Their thesis is, roughly, that if the
recommended procedure is followed people's personal
probabilities, even if very different at the outset, will converge
towards identity, because differences in personal probability
must be largely due to differences of knowledge and experience.
The claim is not that there is an objectively true value, or an
objectively valid system of measurement, for the probability of a
particular outcome, but that, if people are constrained by their
awareness of more and more of the evidence and by the need to
avoid a Dutch book's being constructable against them, they
will gradually develop a consensus. For a pure, unmixed
Bayesian this kind of convergence replaces the estimation of
propensities or relative frequencies from statistical samples, and
constitutes a research strategy of wide application. The end-
product thus has at best the status of social convention rather
than of knowledge or discovery. But Bayesians can point out
that standard Fisher or Neyman-Pearson procedures for
estimating probabilities (p. 23) also involve an irreducibly
conventional element about the acceptability of results with
certain significance-levels, confidence-intervals, and so on.
The term 'Bayesianism' thus comes to name championship of
68 The Controversy about§11
a certain procedure for reasoning about probabilities, rather
than just the acceptance of a personalist analysis of probability-
judgements. In its simplest form the basic Bayesian procedure is
as follows. Suppose the question at issue to be the probability, on
all available evidence, of the hypothesis H—for example, the
proposition that it will rain within an hour. The investigator
begins by assigning a prior value to the probability of //(that is,
he decides his lowest acceptable betting-odds) in advance of any
consideration of the evidence, and he also asigns a value to the
conditional probability of E (his first piece of evidence) given
l

the truth of H. For example, if E reports the presence of dark


x

clouds, p(E \H) might be quite high. If a probability is also


1

assigned to p(£i), or to its equivalent p ^ IH) + p(£, | not-H), it is


then possible to calculate p(H\E ) by means of Bayes's theorem,
1

since that theorem puts p{A\B) equal to


y{B\A) x Q4) P

P(B)
where p (B) > 0. The investigator next updates his value for the
monadic probability p ( H ) by putting it equal to p(//|/?i), and he
takes the next piece of evidence available into account— perhaps
E reports a rising barometer—by evaluating p(Z? 1H) and p(£' ).
2 2 2

It is then possible to calculate p(//|/? ) by a fresh application of


2

Bayes's theorem, and the whole procedure is repeated again and


again until all available items of evidence have been taken into
account. At each stage—that is with each E, on which H is
conditionalized—the probability-function employed is properly
regarded as a different one, expressing the investigator's state of
mind at just that stage of the investigation. After the last piece of
evidence, E , has been taken into account and a corresponding
n

value obtained for a penultimate dyadic function p(H\E„), a


final monadic probability-function emerges, which expresses
the investigator's state of mind in the light of all the available
evidence, and p(H) according to this function is put equal to
p(H\E ) according to the penultimate function.
n

In practice, of course, the investigator may often come to


know the truth of E- before he has had an advance opportunity to
decide his lowest acceptable betting-odds on that proposition. In
such a case he presumably has to imagine what these odds would
§6the Nature of Pascalian Probability 69
be in the ideal situation in which the truth ofE is not yet known.
x

But there is a more serious problem about the procedure of


successive re-conditionalization. Does it really tend to produce
consensus? This looks at first sight like being a question that can
be settled by psychological experiment. But it has certainly not
yet been so settled. On the one hand, sharp controversy exists as
to whether people who are untutored in probability-theory are
inclined to make certain systematic kinds of mistake when
confronted with Bayesian tasks. For example, do they tend to
32

overestimate, or to underestimate, the importance of prior


probabilities? O n the other hand, if instead the experiment is
carried out on instructed and informed investigators, it may
reveal nothing more than the results of imparting such
instruction and information in a particular way. Moreover, any
apparent failure of Bayesian reasoners to converge towards
consensus can always be put down to their having stopped too
soon in the process of seeking out, and assimilating, evidential
facts. So on a wider view the convergence thesis looks as though
it may be unfalsifiable.
It is sometimes claimed that the eventual convergence of
Bayesian reasoners needs no empirical confirmation. The
argument is that de Finetti has proved the applicability of
Bernoulli's limit theorem (the so-called 'law of large
numbers'—see pp. 21-2) to sequences of what de Finetti calls
'exchangeable' events. So that in the case of such sequences,
whatever be the opinion with which a person starts out, he must,
if he is to be coherent in his beliefs and has made a sufficient
number of observations, come to assign a probability to the type
of event in question that is sufficiently close to its observed
relative frequency. But the trouble is that what makes a
33

sequence of events—such as a sequence of coin-tosses


—exchangeable, in de Finetti's sense, is that their order of
3 2 For the view that they do make such mistakes, see R. Nisbett and L. Ross, Human
Inference: Strategies and Shortcomings of Social Judgement (Englewood Cliffs: Prentice-Hall,
1980), and D. Kahneman, P. Slovic, and A. Tversky (eds.), Judgment under Uncertainty:
Heuristics and Biases (Cambridge: Cambridge University Press, 1982). For the view that
they do not, see L. Jonathan Cohen, 'Can Human Irrationality be Experimentally
Demonstrated?', The Behavioral and Brain Sciences, 4 (1981), 317-31, 359-70, and The
Dialogue of Reason (Oxford: Clarendon Press, 1986), 157-92.
3 3 See Kyburg and Smokier, 'Introduction', in Kyburg and Smokier, (eds.) Studies in
Subjective Probability, pp. 12-15.
70 The Controversy about§11
occurrence does not affect the probabilities in which we are
interested. So even exchangeability depends, in the end, on
subjective judgements of probability, and if two people differ in
such judgements, or on what to count as evidence, the events in
a particular sequence may seem exchangeable to one and not to
the other. As de Finetti himself has said 'the condition of
"exchangeability" itself has, from the beginning, only a
subjective value'. Inevitably, therefore, the possibility remains
34

open, where there is no ad hoc agreement on the exchangeability


of a particular sequence, that two people who are fully rational
by Bayesian standards may differ in their initial judgements of
probability and never converge to a consensus, despite the
internal coherence of both sets of beliefs and the extensive
observations that give rise to them. 35

§ 10. M U L T I - V A L U E D L O G I C T H E O R I E S

The conception of probability as a truth-value in a multi-valued


logic has three disadvantages. It fails to imply any guidelines for
the measurement of probability, it seems inapplicable to dyadic
probability, and it requires our multi-valued logic to be non-
truth-functional.
We have seen that the Ramsey-de Finetti-Savage philosophy of
probability is primarily directed towards the question 'How
3 4 De Finetti, 'Foresight', p. 153.
3 5 Personalist analyses are discussed further in D. Gillies, 'The Subjective Theory of
Probability', British Journal for the Philosophy of Science, 23 (1972), 138-57; G . H . von
Wright, 'Remarks on the Epistemology of Subjective Probability', in E. Nagel, P.
Suppes, and L. Tarski (eds.), Logic, Methodology and Philosophy of Science: Proceedings of the
1960 Congress (Stanford: Stanford University Press, 1962), 330-39; R . C . Jeffrey, The
Logic of Decision (New York: McGraw-Hill, 1965); E, Eells, Rational Decision and Causality
(Cambridge: Cambridge University Press, 1982), 1-86; I. Hacking, 'Subjective
Probability', British Journal for the Philosophy of Science, 16 (1966), 334-9; and K. Heilig,
'Carnap and de Finetti on Bets and the Probability of Singular Events: The Dutch Book
Argument Reconsidered', British Journal for the Philosophy ofScience, 29 (1978), 325-46. In
regard to exchangeability see also B. de Finetti, 'On the Condition of Partial
Exchangeability', in R . C . Jeffrey (ed ), Studies in Inductive Logic and Probability ii
(Berkeley: University of California Press,. 1980), 193-205. In regard to convergence see
also G. Shafer, 'A Subjective Interpretation of Conditional Probability', Journal of
Philosophical Logic, 12(1983), 453-66, B. Skyrms, Pragmatics and Empiricism (New Haven:
Yale University Press, 1984), 20-92, a n d K . Lehrer, 'Consensus', in R.J. Bogdan(ed.),
Keith Lehrer (Dordrecht: Reidel, 1981)63-75.
§6the Nature of Pascalian Probability 71
does a person assign values to his probabilities?' rather than
towards the question 'What are the correct values to assign to
one's probabilities?' According to such a personalist theory a
judgement of ^4's probability consists in its author's degree of
belief that A is true. But many people take their own degree of
belief to be the inter-subjectively correct one, as in the
normativist version of a Ramsey-type theory. So for such
people, if a probability is attached to a proposition, it can be
regarded as the inter-subjective truth-value of that proposition
rather than as the degree of justifiable mental commitment to it.
Accordingly it is possible to interpret the Pascalian calculus of
chance as a multi-valued logic. To judge that the proposition A
has zero probability is to judge that it is false; to judge that it has
probability 1 is to judge that it is true; and to judge that it has
some intermediate degree of probability is to assign it one from
an infinite set of intermediate truth-values.
At first sight this conception of probability seems to have the
great advantage of simplifying our conceptual system. A
familiar branch of mathematics turns out to instantiate a
familiar species of logic. Also, as Lucas points out, the 36

conception of probability as a partial truth fits particularly well


with Toulmin's idea (p. 64) that in voicing a judgement of
probability one utters a speech-act of partial commitment. If
anybody who asserts a proposition unguardedly is thereby
claiming its truth, anybody who asserts it guardedly may be
regarded as thereby claiming an intermediate truth-value for it.
But, even apart from the objections to Toulmin's idea that
have already been noted (pp. 64-5), there are at least three
reasons why a multi-valued logic analysis is less useful than it
may seem.
The first reason is that, considered on its own, the analysis
gives us no hint as to how we should in practice measure
probabilities. In that respect it is similar to a propensity analysis.
Its distinctive feature is that it introduces gradations into the
evaluation of propositions, rather than of natural tendencies.
But, whether you say that the calculus should be interpreted in
objective-idealist style as a multi-valued logic or in a realist style
as a theory of natural tendencies, you imply no guide-lines for
3 6 J . R . Lucas, The Concept of Probability (Oxford: Clarendon Press, 1970), 1-22.
72 The Controversy about§11
carrying out conscious measurements of probability. Such
measurements are normally rooted in something quite
different—most often in an examination of the evidence
provided by observed relative frequencies. And in fact
Reichenbach showed how the conception of probabilities as
truth-values could be mapped on to a Pascalian-structured
calculus of relative frequencies in the special case in which the
reference-class is always identical with the universe. Roughly,
the truth-value of 'This is an A' is equated with the relative
frequency of v4s in the contextually assumed universe. Even 37

then we are left with the problem of how to measure the truth-
value of 'This is an A' where we know the relative frequency of
As only within a reference-class that is smaller than the relevant
universe as a whole.
Secondly, though a propensity analysis can interpret dyadic
judgements of probability as statements of conditional
tendencies (and a personalist analysis can relativize strength of
belief to awareness of particular items of evidence), a multi-
valued logic account sits easily only with monadic judgements,
because each proposition is naturally assumed to have just one
truth-value, whether or not we know it. The proposition that it
will rain soon may be probable on the evidence of dark clouds,
and improbable on the evidence of increasing atmospheric
pressure. But this proposition presumably has only one truth-
value, which is determined by the actual event, when it happens,
and is not relative to some other, quite different events. No
doubt we could restrict ourselves to a monadic judgement here
by identifying the conditional probability with the probability of
the corresponding conditional—that is, we could identify p(A \B)
with p(If^4, then B). But there would be a price to pay for this. It
would reduce the variety of different facts that our judgements of
probability are capable of expressing, as we have already seen
(pp. 20-1). The better course would be just to define p(A\B) as
p(A&B)
P (B)~
where p(5) > 0.

Reichenbach, Theory of Probability, pp. 395-7


§6the Nature of Pascalian Probability 73
Thirdly—and this is perhaps the most important point
—standard multi-valued logics are truth-functional, in the
sense that the truth-value of a logically compounded pro-
position, like a conjunction or a disjunction, always depends
solely on the truth-values of the propositions compounded. But
the probability of a conjunction depends solely on the
probabilities of the conjuncts only in the special case where these
probabilities are independent of one another. For example, if we
replace the ball after each drawing (see pp. 18-19 above) we
can calculate in this way the probability of drawing a white ball
twice from an urn containing two black balls and two white ones;
but, if the first ball drawn is not replaced, we cannot do this.
Indeed, the general law for conjunction—p(ASzB) = p(.4) X
p(B\A)—shows clearly that, if the Pascalian calculus is
interpreted as a multi-valued logic, that logic must be non-truth-
functional, in the sense that the probability of a compound
proposition, like p(^4&S), is not necessarily a function of the
probabilities of the propositions compounded, p(A) and p(S). So
instead of being able to treat the mathematics of the calculus as
just another instance of a familiar kind of logic we have in fact to
extend our conception of such a logic in order to embrace a new,
non-truth-functional type of system. Such an extension is more
acceptable to some logicians than to others, and this has
inevitably affected the extent to which it has been thought
helpful or illuminating to conceive of probability as an
intermediate truth-value. So in one important respect the
38

conception of probability as a truth-value within a multi-valued


logic is like the conception of probability as a propensity. It is
compatible with a Pascalian structure but does not demand one,
as does any conception that takes probabilities to be ratios,
whether these be ratios of the permutations of chances (as in
§ 6), of the cardinalities of sets (as in § 7), of the sizes of betting
stakes (as in § 9), or of quantities of logically possible worlds (as
in § 11). Indeed the normal principle for conjunction in a
39

multi-valued logic (equating the value of a conjunction with its


least-valued conjunct) is in fact shared by the non-Pascalian

3 8 N. Rescher, Many-valued Logic (New York: McGraw-Hill, 1969), 186-8.


3 9 Ibid. 131.
74 The Controversy about § 11
system of inductive and probabilistic reasoning that is discussed
in § 21 below. 40

§ 11. L O G I C A L R E L A T I O N T H E O R I E S

If probability is thought of as a logical relation, we can take the


ratio of the quantity of logically possible worlds in which A&B is
true to the quantity of logically possible worlds in which B is true
as determining the extent to which A is a logical consequence of B
and thus as defining the probability of A given B. Carnap con-
structed a systematic version of this analysis, but pointed out in
effect that there is an infinite number of different ways of meas-
uring quantities of logically possible worlds.
If the mathematics of Pascalian probability is to be treated as a
kind of generalized logic, the conception of probabilities as the
truth-values of a multi-valued logic is not the only option. That
analysis in any case, as we have already seen, sits easily only
with monadic probabilities. And there is another logical option
that sits more naturally with dyadic probabilities, namely the
conception of p(A \ B) as the degree to which B entails A. O n this
view, as Keynes put it in 1921, no proposition is probable, or
improbable, in itself but only relative to certain premisses, and
this relation is a logical one. So we shall have p(A | B) = 1 if and
41

only if B8cnot-A is self-contradictory, and p(A\B) = 0 if and only


if B&A is self-contradictory. The argument from B to A is
certain—has maximum probability—when A is a logical con-
sequence of B. But how are the intervening values to be
determined?
Keynes thought that some probabilities are just not
comparable at all with one another. Most others, in his view, are
not measurable, despite being intuitively comparable in terms of
'greater than', 'lesser than', or 'equal to'. And he thought that
only in a very special type of case (suitably constrained and
qualified inferences from statistical data) are probabilities
4 0On the multi-valued logic analysis see further A. A. Zinov'ev, Philosophical Problems
ofMany-valued Logic (Dordrecht: Reidel, 1963), 107-9; and A.R. White, Modal Thinking
(Oxford: Basil Blackwell, 1974), 59-74.
J . M . Keynes, A Treatise on Probability (London: Macmillan, 1921), 7. Keynes
4 1

(1883-1946) is perhaps better known for his work in economics than for his philosophy of
probability.
§6the Nature of Pascalian Probability 75
numerically measurable. But Carnap showed, in 1950, that it
42

is possible to develop numerical measures for Pascalian prob-


ability that apply whenever this probability is conceived as a
logical relation between sentences in a particular kind of artifi-
cial language.
Unlike Keynes, Carnap did not claim a monopoly of truth
for the logical relation theory. Rather, he argued that there
were two primitive or presystematic concepts of probability,
which he called 'probability', and 'probability' , respectively. 2

Probabilityi is a relation between conclusion and evidence,


broadly describable as degree of confirmation. A function that
assesses probability, takes sentences as fillers of its argument-
places. Moreover, judgements of probability,, assigning
numbers as values of such a function for particular fillers of its
argument-places, are either analytically true or analytically
false: their truth-value depends on the rules of the language in
which they are formulated. Probability , on the other hand, is a
2

relation of relative frequency between properties, classes, or


kinds of events. Judgements of probability are therefore factual2

and empirical. Carnap fully acknowledged the utility of


probability for many purposes. He accepted that in statistical
2

enquiry judgements of probability characteristically relate


properties, classes, or kinds of events, and should be analysed as
empirically assessable statements about relative frequencies.
Later he also worked with a Ramsey-type analysis of
probability. But his 1950 theory of probability as a relation
43

between propositions was concerned only with a priori


judgements of the extent to which a particular statement of
evidence would logically confirm a particular hypothesis at
issue. And Carnap acknowledged that the germ of his theory of
44

logical probability derived from Waismann's elaboration of an 45

idea originally formulated in Wittgenstein's Tractatus* Indeed, 6

4 2Ibid. 34.
4 3R. Carnap, 'Inductive Logic and Rational Decisions', in R. Carnap and R.C.
Jeffrey (eds.), Studies in Inductive Logic and Probability, i (Berkeley: University of California
press, 1971), 5-31.
4 4R. Carnap, Logical Foundations of Probability (Chicago: Chicago University Press,
1950), 299
4 5F. Waismann, 'Logische Analyse des WahrscheinlichkeitsbegrifP, Erkenntnis, 1
(1930-1), 228-48, repr. in English as 'A Logical Analysis of the Concept of Probability',
in F. Waismann, Philosophical Papers (Dordrecht: Reidel, 1977), 4-21.
4 6 L. Wittgenstein, Tractatus Logico-Philosophicus (London: Kegan Paul, Trench,
76 The Controversy about§11
there are also some anticipations of Carnap's approach to the
problem in the writings of Leibniz. 47

If A is a logical consequence of B, then A&B is true in every


logically possible world in which B is true. But if A is neither a
logical consequence of B nor inconsistent with it, then A&.B is
true only in some, not all, logically possible worlds in which B is
true (namely, those in which A also happens to be true). If,
therefore, we have a measure for possible worlds, we can take
the ratio of the quantity of logically possible worlds in which
A&B is true to the quantity of logically possible worlds in which B
is true as determining the extent to which A is a logical
consequence of B and thus as defining the probability of A given
B (if we conceive degree of probabilification as degree of logical
consequence).
That is the nub of the Wittgenstein-Waismann-Carnap
analysis. But in keeping with the prevailing nominalism of the
period, Carnap formulated it in terms of expressions in an
artificial language rather than in terms of possible worlds.
Specifically, Carnap began by supposing a standard type of
artificially composed logical language, with a finite number of
monadic, first-order predicates F, G, H, . . . (like 'is red', 'is
square', 'is male', etc.) and a finite number of individual
constants a,b,c, . . . (like uniquely assigned proper names). An
atomic sentence is then defined as an assignment of an
individual constant to a predicate, e.g. Fa (like ' T o m is male').
A state description is defined as a conjunction of sentences
containing every atomic sentence or its negation but not both.
(So a state description describes a distinct possible world, within
the limits of the language's descriptive potential.) The range of a
sentence may then be defined, roughly, as the class of state-
descriptions in which that sentence holds. More precisely, the
range of a sentence is the class of state-descriptions such that, for
each of these state descriptions, the sentence would be true if
individuals had just those properties attributed to them by the
state description.
Trubner, 1922), 111 (para. 5.15). Some sketchy remarks survive which suggest that
Wittgenstein later changed his mind on this issue. See e.g. his Philosophical Remarks, ed.
R. Rhees, trans. R. Hargreaves and R. White (Oxford: Blackwell, 1975), 227-47, and
Wittgenstein and the Vienna Circle, ed. B. McGuinness, trans. J . Schulte and B.
McGuinness (Oxford: Blackwell, 1979), 93-6, 98-9.
4 7S e e l . Hacking, The Emergence of Probability, pp. 134-142.
§6the Nature of Pascalian Probability 77
Each state-description is then assumed to have a measure
which is a positive real number such that the sum of all the state-
descriptions' measures is equal to 1; and the measure of a
sentence's range is said to be the sum of all the measures of the
state-descriptions in which that sentence holds. Carnap
abbreviates 'the measure of the range of h' to 'm (h)' and defines
the extent to which e confirms h (abbreviated to c (h,e) as the
ratio
m (e&h)
m(e)
—in effect, as the ratio of the quantity of logically possible
worlds in which both e and h are true to the quantity of logically
possible worlds in which e is true. Finally, for languages with an
infinite number of individual constants m(A) is defined as the
limit of its values in a sequence oflanguages with ever increasing
finite numbers of individual constants, and c(h,e) is defined ana-
logously. And Carnap proves that any confirmation-function,
thus defined, conforms to the standard mathematics of condi-
tional probability, while any such measure-function conforms to
the standard mathematics of unconditional (i.e. monadic)
probability.
Carnap then concentrates his attention on a particular type of
confirmation-function which he calls 'symmetrical'. A symmet-
rical confirmation-function is one that treats all individuals alike
and thus allows any uniform replacement of one individual
constant within its scope by another, provided that this
preserves existing relations of sameness and difference between
occurrences of those constants. For example, if c is a
symmetrical confirmation-function c(Fa,Fb) will have the same
value as both c(Fa,Fc) and c(Fc,Fb), but not the same value as
c(Fa,Fa). The theory of symmetrical confirmation-functions
derives from assuming, not unreasonably, that two state
descriptions have the same measure if they differ only in respect
of which individuals have the particular set of properties
described. Indeed Carnap believed that there was general
agreement among writers on probability! that any analysis of
logical probability should attribute to it this property of
symmetry, just as all individuals have equal rights before the
tribunal of deductive logic. When we infer from 'All men are
mortal' and 'Socrates is a man' to 'Socrates is mortal', the
78 The Controversy about§11
validity of our inference is unaffected if we replace each
occurrence of 'Socrates' by an occurrence of 'John Doe' or of
any other proper name. (Compare the account of de Finetti's
idea of 'exchangeability, on pp. 69-70 above.)
But no unique measures emerge in this way. For example,
one possible option would be to assign the same real number to
each state-description of any kind (with all these numbers
summing to 1). After all, there seems to be no reason, prior to
actual experience, why one state-description should be regarded
as less probable than another. The principle of indifference
seems to be applicable here. And Carnap assigned the name ' c t '
to the symmetrical confirmation-function that is evaluated in
this way in any language. Yet it would also be possible to apply
the principle of indifference in another way. Specifically, we
might suppose that what were ultimately to be put on a level with
one another were patterns of state-description, rather than
individual state-descriptions, on the ground that what mattered
were the particular predicates instantiated or not instantiated,
rather than which particular individuals instantiated them.
Carnap called such a pattern a 'structure-description', defining
it as a'disjunction of isomorphic state-descriptions. If each
structure-description is assigned—for its measure—the same
real number as every other structure-description (with all these
numbers summing to 1), a measure can then be assigned to each
of the n disjoined state-descriptions within a particular
structure-description by dividing the measure of that structure-
description by n. And Carnap designates a confirmation-
function that is evaluated in this way by the symbol c*. In terms
of the distinction between permutations and combinations that
was noted on p. 15 above, (compare p. 45) c t treats the
different permutations of predicates and individual constants as
the fundamental units of account whereas c* treats the different
combinations of them as the fundamental units.
Waismann had already made the general point that no
uniquely correct range-measure is possible. But it was Carnap
who first took this problem seriously. He not only drew the
explicit distinction between c t and c*, but also emphasized that
they have importantly different properties, as a simple example
will illustrate. Consider a language with only one atomic
predicate H (which could mean 'falls heads uppermost') and
only three individual constants a, b, and c (which could designate
a particular coin's first, second, and third tosses, respectively).
§6the Nature of Pascalian Probability 79
There are then eight different state-descriptions of which one is
Ha&Hb&Hc, two include Ha and Hb, two include Ha and He,
and four include Ha. It follows that, using Carnap's system of
measurement, we shall have c t (He,Ha) = 7/7 = 7 and also ct
(Hc,Ha&.Hb) = 7/7 = 7. But there are four different structure-
descriptions (all heads, two heads, one heads, and no heads). So
m*(Ha&Hc) = 7 + ,7 = because Ha&aHc holds good in one
state-description that is the only member of its structure-
description (the structure-description where all tosses are said to
fall heads) and in one state-description that is a third of the
membership of its structure-description (the structure-
description where only two tosses are said to fall heads).
Similarly m* (Ha) = 7 + jf + ,7 = 7. So c*(He,Ha) = f; and,
when calculated analogously, c*(Hc,Ha&cHb) = f. In other
words, when measured by ct the logical probability of He, on
the evidence of Ha, is not increased by the additional evidence
that Hb provides. But, when measured by c*, the probability is
thus increased. Clearly c t might be a suitable measure when
successive outcomes are assumed a priori to be independent of
one another and empirical evidence about past outcomes is
taken td be irrelevant to the probabilities of future ones, as in an
acknowledged game of chance; whereas c* might be suitable
when no such assumptions are made and empirical evidence
about past outcomes is allowed to affect the probabilities of
future ones.
In both cases the sentence evaluating the function—for
example, the equation c t (He,Ha) = 7 or c *(Hc,Ha) = f—is to
be regarded as an a priori truth, since it is derivable from the
principles of the system and does not require looking and seeing,
as does the measurement of a relative frequency. But in the case
of c t it is an a priori truth about the value of a probability that is
not conditional on the evidence's consisting of any particular
number of resembling items, whereas in the case of c* it is an a
priori truth about the value of a probability that is conditional on
such evidence. In the latter case, if anyone gives us both the
hypothesis and the evidence, we can work out a priori the
probability of the one on the other: in the former case, if anyone
gives us just the hypothesis we can work out a priori its
probability on any number—from zero upwards—of distinct
but resembling evidential items. And it is clear that, in addition
to c t and c*, an infinite number of other symmetrical
c-functions could be designed, so as to provide other bases for
80 The Controversy about§11
the a priori measurement of probability when conceived as a
logical relation. As different kinds of inequality are introduced
into the measure of state-descriptions, according to the weight-
ing attached to the number of state-descriptions isomorphic with
a given state-description, probabilities are made sensitive to
evidence to varying extents. Also, c-functions are possible that
share with c* its sensitivity to additional evidence but avoid
making the degree of confirmation for any sentence depend in
part, as c* depends, on the choice of primitive predicates within
each family of primitive predicates in the language, which 48

seems to be a purely conventional issue. But any true statement


assigning a specified value to a c-function for specified fillers of
its argument-places—like 'c(Fa,Fb) = j'—is always analytic. In
any treatment of probability as a logical relation intermediate
uses of that relation are just as much an issue for a priori
assessments as are the limiting cases of entailment and
contradiction.
Carnap's theory of logical probability was intended to
provide a system of gradation for inductive inference, and its
adequacy for this purpose will be discussed in § 17 below. But,
within fhe constraints of an artificial language where primitive
predicates are monadic in structure and finite in number,
Carnap's theory certainly provided a rigorously developed
method of interpreting the Pascalian calculus along idealist
lines. And it inspired a number of other philosphers, in
particular Hintikka (see § 17 below), to seek to improve on
Carnap's formulations. But, though Carnap's analysis
generated different probability-functions as the size or content
of the language is altered or a different measure-function
adopted, he explicitly denied that the choice of one measure-
function rather than another could ever be justified by empirical
evidence. So judgements of probability, inevitably contain, on
49

Carnap's view, an element of intuitive or conventional


arbitrariness, when compared with judgements of probability . 2
50

4 8 I. Hacking, 'Linguistically Invariant Inductive Logic', Synthese, 20 (1969), 25-47.


4 9 Logical Foundations of Probability, p. 299.
5 0 On the logical relation analysis see further H. Jeffreys, Theory of Probability (Oxford:
Clarendon Press, 1948); R. Swinburne, 'The Probability of Particular Events',
Philosophy of Science, 38 (1971), 327-43; and the references given in nn.6-15 for ch. IV
below
Ill
The Foundations of Pluralism in the Analysis of
Probability

§ 12 S O M E L O G I C A L D I S T I N C T I O N S E X P L O I T E D BY D I F F E R I N G
ANALYSES OF PASCALIAN PROBABILITY

Different analyses of Pascalian probability have different impli-


cations in regard to certain important issues about the logic of
probability-judgements. One such issue concerns whether a
probability-judgement is at best necessarily true or contingently
true. A second concerns whether the judgement is sentence-
related or predicate-related. A third concerns the substitutivity
conditions for A and B in a judgement of probability p (A E) = n.
A fourth concerns whether or not a given judgement of pro-
bability is counterfactualizable.
T h e previous chapter distinguished at least six different kinds of
theory about the semantics and epistemology of Pascalian pro-
bability. O n e kind of theory evaluated probabilities with the
help of a principle of indifference, a second by reference to
relative frequencies, a third by natural propensities, a fourth by
actual or appropriate strength of belief, a fifth by multi-valued
logic, and a sixth by ratios of logical ranges. What are we to
make of this embarrassingly rich diversity of philosophical
analyses? None of the analyses considered seemed altogether
without a rationale. But neither did any seem universally
applicable.
Of course, not all philosophical analysts of probability have
claimed that their own analysis is the only correct one. For
example, though indifference theorists and Bayesians have 1 2

1 e.g. P.S. de Laplace, A Philosophical Essay on Probabilities (1814), trans. F.W.


Truscott and F.L. Emry (New York: Dover, 1951), 4 ff.
2 e.g. B. de Finetti, 'Foresight: Its Logical Laws, Its Subjective Sources (1937)', in
H. E. Kyburg and H. E. Smokier (eds.), Studies in Subjective Probability (New York: John
Wiley and Sons, 1964), esp. 111-18.
82 The Foundations of Pluralism in§15
often done this, Carnap originally thought that probability
3

could be conceived either as a logical relation or as a relative


frequency and others, like Nagel and Mackie, have been even
4 5

more open-minded and have acknowledged the existence of four


or five different concepts of Pascalian probability. But a tolerant
pluralism leaves many questions unanswered. In particular, we
need to understand how it is that several very different concepts
of probability are possible; we need to know which concepts are
most appropriate to which contexts of employment; and we also
need to discover whether there is reason to suppose that the six
kinds of theory examined in § § 6-11 are the only philosophical
analyses of probability that are possible, or whether the under-
lying problem-situation admits a wider space of possible
solutions.
It will not do to tell us here, as Mackie does, quoting with
approval from J . S. Mill, that 'names creep on from subject to
subject until all traces of a common meaning sometimes
disappear', unless there is historical evidence that some such
6

gradual process of linguistic expansion accounts for all the


differences under examination. But Mackie offers no historical
evidence to support his thesis. Indeed, we have seen ( § 3 ) that
some diversity of conception was present at the outset. Even in
the earliest years of work on the mathematical principles of
Pascalian probability those principles were already being
applied both to judgements of aleatory chance and also to judge-
ments of evidenced credibility. Compare judgements of
goodness. Most philosophers would agree with Aristotle that 7

the various uses of the word 'good' in 'good man', 'good


health', 'good pen', etc. are not due to chance homonymy. No
more should we simply assume that the various available senses
of the word 'probable' (and of its equivalents in other languages)
are due to accidents of linguistic history.
3 Logical Foundations of Probability (Chicago: Chicago University Press, 1950), 23-36.
4 E. Nagel, Principles of the Theory of Probability, in O. Neurath, R. Carnap, and C.
Morris (eds.), Foundations of the Unity of Science (Chicago: Chicago University Press,
1938), i. 342-422.
5J . L . Mackie, Truth, Probability and Paradox (Oxford: Clarendon Press, 1973),
154-236.
6 Ibid. 155, quoting from J.S. Mill, A System of Logic (London: Longmans, Green,
1896), 24.
Nicomachaean Ethics 1096 , 26-27.
7 b
§15the Analysis of Probability 83
Nor is any light cast on the problem by Mackie's claim that 8

probability should be seen as a 'family-resemblance' concept in


Wittgenstein's sense, since that approach to the understanding
9

of conceptual pluralism is inherently unrewarding. For


example, men ride petrol-tankers, horses, and tractors, but not
oxen; oxen, horses, and tractors, but not petrol-tankers, are
used on farms for pulling things; petrol-tankers, tractors, arid
oxen stay on the ground, while horses sometimes jump; and all
petrol-tankers, oxen, and horses have their front and rear means
of locomotion approximately the same height, while most
tractors do not. But no one supposes it appropriate to mark this
particular nexus of family resemblance by carrying over the
name of one of the four sorts of objects to describe the other three
sorts. The fact is that some groups of four or more sorts have
common names, and some do not, even when the sorts exhibit
family resemblance to one another. So unless the exponent of a
family-resemblance approach to probability tells us why his
supposed nexus of family-resemblance generates a common
name, in contrast with others that do not, he has not explained
anything. But if he does do this, and does it adequately, he has to
go beybnd a merely family-resemblance account. It will
certainly not be adequate to say that a common name is gener-
ated when the resemblances are sufficiently close, if the only
criterion of sufficient closeness is the use of a common name.
A deeper level of philosophical understanding is clearly
necessary here. To achieve it we must begin by paying some
attention to the fact that the different kinds of analysis of
Pascalian probability that philosophers have proposed differ not
only in the substance of their central message, but also in what
they imply about certain important logical issues. Not all of
these differences of implication have been adequately
recognized in discussions of the relevant analyses. But when
they are recognized they help us to see that behind the different
kinds of philosophical analysis lie quite a variety of differ-
ent conceptions of Pascalian probability that are available
for human reasoning. For example, let us consider first the
modality of probability-judgements, then whether they are
8 Mackie, Truth, Probability and Paradox p. 155.
9 L. Wittgenstein, Philosophical Investigations (Oxford: Blackwell, 1953), 32-3.
84 The Foundations of Pluralism in§15
sentence-related or predicate-related, then their substitutivity
conditions, and then what may usefully be termed their
'counterfactualizability'.
Thus, if you adopt an indifference analysis for a particular
judgement of probability, and assume the value of the pro-
bability to be determined ultimately by the rules of a game, that
judgement is made out to be either necessarily true or neces-
sarily false. If in a valid throw, a die can fall only on one or other
of its six sides and if there is assumed to be an equal chance of its
falling on any one side, there is no possible world in which, when
such a game is played, the licit probability of its falling twice with
the same side uppermost is anything other than • Similarly
Carnap's logical range theory imposes necessary truth or
necessary falsehood on any evaluation of a particular
probability-function for a particular pair of sentences, such as
c*(h,e) = } or ct(h,e) = j. But relative frequencies, natural
propensities, or strengths of belief (or empirically established
indifference-dependent chances) could be different in other
logically possible worlds than the actual one. So judgements of
probability in those terms can at best be contingently—not
necessarily—true.
Again, we commonly take the overall structure of a
probability-judgement to be represented by the symbolism p(^4)
= n or p(/4|£) = n. It is understood here that the letter n stands
in for a real number, or for a rational one, such that 1 ^ n ^ 0.
But what do the letters A and B stand in for—whole sentences,
expressing fully determinate propositions, or just sentence-
schemata or parts of sentences? Carnap's analysis requires the
letters A and B to stand in for sentences, since it is concerned
with probability as a confirmatory relationship between pro-
positions, and the multi-valued logic analysis also is obviously
sentence-related rather than predicate-related. But the relative
frequency theory requires A and B to stand in for terms desig-
nating classes, since it is concerned with the relative frequency
with which members of one class are also members of another.
And the indifference, propensity, and personalist kinds of
theory seem in principle to be compatible with both options,
though particular versions of these analyses may not be. For
example, you can speak in general about any game of drawing a
card at random from a well-shuffled pack, and then A in p(A) =
§ 15 the Analysis of Probability 85
-i-will signify a type of outcome, namely, the drawn card's being
an ace; sov4 will stand in for a sentence-part. Or you can speak
specifically of the very next round of playing that game and say
that there is a probability of that the next card drawn here this
afternoon will be an ace, in which case A will stand in for a whole
sentence. Again, you could measure both the strength of my
belief, for any person, that he will survive to the age of seventy,
given that he is a lorry-driver, and also the strength of my belief
that our friend John Doe, specifically, will survive to the age of
seventy given that he is a lorry-driver. In the former case A
would stand in for a sentence-schema, in the latter for a fully
determinate sentence.
Another important issue is that of substitutivity. Under what
conditions of relationship between A and A' and between B and
B' is the truth of p(A | B) = n sure to be unaffected if we replace A
by A' and/or B by B'?
We have already seen (pp. 72-3) that, when the mathematics
of Pascalian probability is treated as a multi-valued logic, the
logic is not truth-functional. So it is clear that propositions with
the same truth-value cannot always be safely substituted for one
another within a judgement of probability.
In certain other analyses of probability further restrictions on
substitutivity also operate. For example, Carnap's analysis
makes p(A\B) depend on the ratio of the range of A&B to the
range of B (§ 11). This ratio is 1 where B logically implies A
(because then the range of A&cB is identical with the range of B).
Now consider the Carnapian representation of the case where B
is a conjunction of atomic sentences (p. 76) that logically implies
A, and B' differs from B only in so far as one of the atomic
conjuncts in B' has a predicate that differs from the corres-
ponding predicate in B but is co-extensive with that predicate in
its application within the actual world. Under these conditions
the ratio of the range of A&.B' to the range of B' will not be 1,
because A&cB' and B' will not be true in just the same possible
worlds. So c(A\B) = 1 will be true, but c(A IS') = 1 will be false.
For example, though it is certain that this man is George's
sibling, given that he is George's brother, it may be only
probable that the man is George's sibling, given that he is
George's partner, even though it is in fact the case that all
and only George's brothers are his partners. Thus only those
86 The Foundations of Pluralism in§15
predicative expressions that are analytically equivalent can be
safely substituted for one another within Carnapian c-functions.
And some restriction on substitutivity, albeit a lesser one,
applies also to a propensity theory, because accidentally co-
extensive predicates do not attribute the same property or
natural tendency. For example, there may be a strong natural
tendency—equalling a probability of, say, 0.8—for persons
elected to Parliament to have been educated at a university. But
even if it were accidentally the case, through a surprising set of
coincidences, that all and only Parliamentarians turn out to
have names with eight vowels we should not suppose this feature
of people' s names to depend to any extent on their having been to
a university: the probability that such a person had been to a
university would be a relative frequency, not a propensity. For
predicative expressions to be safely substitutable for one another
within a statement of a propensity-type probability, their equi-
valence must be guaranteed by such laws as those of nature,
logic, mathematics, or language.
Other kinds of substitution are less restricted.
For example, on a relative frequency account it obviously
makes no difference to the value of a probability if we use a
different expression to designate the same class. If the
passengers on a particular train have return tickets if and only if
they intend to return the same day, then the relative frequency
of passengers with return tickets is just the same as that of
passengers intending to return the same day. So on a relative
frequency account co-extensive predicates—predicates satisfied
by all and only the same individuals—can be substituted for one
another within a judgement of probability without any risk of
altering the judgement's truth-value. And the same holds good
for the realist form of indifferentist account (§ 6), since the
actual ratio of favourable to equally possible cases should be
quite unaffected by a change in the expression used to designate
the outcome of which the probability is at issue.
Even on a personalist analysis probability-judgements are by
no means as resistant to substitution as personalist talk about
probability's measuring strength of belief might lead one to
suppose. In everyday descriptions of people's beliefs not even
logically or mathematically equivalent expressions may safely
§15the Analysis of Probability 87
be substituted for one another, let alone expressions that refer to
the same individuals or predicate the same properties. And the 10

same holds good for probability-judgements if we accept a per-


sonalist analysis that does not insist on what de Finetti (p. 60)
called coherence. But, where coherence is required, the restric-
tion on substitutivity has to be discarded: the context needs
to be referentiaJly transparent. Anyone whose probability-
judgements are to be taken as declarations of the lowest odds
that he is prepared to accept will guard himself adequately
against a Dutch book only if he treats all his ordinary
probability-judgements as being open to the mutual substitution
of terms that have the same reference or meaning as one
another. You cannot allow yourself to lose a bet just because the
person you bet with uses a different designation for the same
outcome: a horse runs just as fast (like a rose smells just as sweet)
under another name. The only wagers in which such substi-
tutivity is restricted are those on the truth of propositions that
are resistant to substitution even when they are not the subject of
a wager: for example, certain propositions about a person's
fears, hopes, desires, beliefs, or other mental states.
Important questions of substitutivity may also arise even
where the two expressions that are candidates for inter-
substitution do not designate the same thing. Consider a
monadic or dyadic judgement about a specified individual, such
as John Doe (as distinct from an explicity general judgement
about 'any' person). We might say, for example, that the pro-
bability ofJohn Doe's surviving to the age of seventy, given that
he is a lorry-driver, is 0.8. Now does this statement refer to John
Doe as just one individual among many others, so that if the pro-
bability holds for John Doe it must hold also for Richard Roe or
for anyone else? Or does it refer to John Doe in all his singularity
of circumstances (healthy exerciser, cautious eater, non-
smoker, etc.), so that the probability of Richard Roe's survival
to the age of seventy, given that he is a lorry-driver, may be quite
different? In the former case the name 'John Doe' may be
replaced by any other expression that designates a particular
See R. Carnap, Meaning and Necessity (Chicago: Chicago University Press, 1947),
1 0

59-64, and W . V . O . Quine, Word and Object (Cambridge: Massachusetts Institute of


Technology Press, 1960), 143-7.
88 The Foundations of Pluralism in§15
individual, in the latter it may not. So let us use the term
'implicitly general' to describe those grammatically singular
judgements of probability about a specified individual which
necessarily remain true whenever expressions denoting that
individual are uniformly replaced by expressions denoting any
other individual within the same domain of discourse. And let us
use 'implicitly singular' to describe a grammatically singular
judgement of probability in which such an expression is not
necessarily replaceable without affecting the judgement's truth-
value. Then an indifference theory will impute implicit gene-
rality to grammatically singular judgements of probability. Also
singular judgements made with the use of any of Carnap's
symmetrical confirmation-functions (pp. 77-8) will be impli-
citly general, though singular judgements made with the use of
non-symmetrical confirmation-functions will be implicitly
singular. Again, a propensity theory will require grammatically
singular judgements of probability to be implicitly general, in so
far as the characteristic evidence for claiming the existence of a
conditional propensity is an observed relative frequency. The
observed ratio of ^4s among Bs is just as good evidence for the
strength of the tendency for any one particular B to be an A as it
is for the strength of any other B's tendency to be an A. But a per-
sonalist theory need not require grammatically singular judge-
ments of probability to be implicitly general unless the theory is
confined to dealing with what de Finetti calls 'exchangeable'
events (pp. 69-70). Certainly my belief that John Doe will
survive to the age of seventy may well be stronger than my belief
that Richard Roe will, and the lowest odds that I would accept
on the one outcome might well be lower than the lowest that I
would accept on the other, just as I might expect often to bet at
different odds on different horses in the same race. So on a per-
sonalist analysis grammatically singular judgements of pro-
bability are implicitly singular also, unless otherwise stated."
Yet another important logical issue concerns what is best
called 'counterfactualizability'. This issue arises only in regard
to those judgements of probability that are both dyadic and
11 Pace Carnap, who claimed that 'all authors on probabi]ity|, both classical and
modern' must be taken to have accepted the implicit generality of grammatically
singular judgements of probability: Logical Foundations of Probability (Chicago: Chicago
University Press, 1950), 488.
§15the Analysis of Probability 89
implicitly or explicitly general. Such judgements typically
concern the probability that an entity is A given that it is B. But
does such a judgement apply even to entities that are not actually
B, in which case it may be said to be 'counterfactualizable', or
only to entities that are actually B, in which case it may be said to
be 'non-counterfactualizable'? For example, to accept that, on
the assumption that he is a forty-year-old asbestos worker, a
man has a 0.8 probability of death before the age of sixty, is to
accept an explicitly general conditional probability that applies
not only to any person in the history of the universe who is
actually a forty-year-old asbestos worker but also to any indivi-
dual who is not. That is, the same judgement of probability
would presumably hold good even if the number of asbestos
workers were larger than it in fact is. Only thus might someone
legitimately infer a reason for not going to work in a asbestos
factory. But to accept a 0.6 probability, for any person picked
out at random at a certain conference, that he is staying in the
Hotel Excelsior is not to accept a probability that could be relied
on to have held good if more people had been at the conference.
Perhaps the additional participants would all have had to stay
elsewhere because the Hotel Excelsior had no more unoccupied
rooms. Such a probability-judgement states an accidental fact
and is non-counterfactualizable, whereas the judgement about
mortality among asbestos workers reflects a causal connection
and is therefore counterfactualizable.
Just as some universal truths are counterfactualizable and
some are not—compare 'Everyone who drinks cyanide is
poisoned by it', with 'Everyone at this session is staying in the
Hotel Excelsior'—so too some probability-judgements are
counterfactualizable and some are not. But (in regard to
probability-judgements that are both dyadic and general) a pro-
pensity analysis seems applicable only to counterfactualizable
judgements. If a natural tendency exists for a certain kind of
event to occur in such-and-such circumstances, it would
normally still have existed even if those circumstances had
occurred more often than they actually occurred. Similarly a
relative frequency analysis imputes counterfactualizability
where the reference-class has an infinite number of members,
since any relevant convergence to a limit might be expected to
maintain itself even if additional members joined the reference
90 The Foundations of Pluralism in§15
class. And an interval-valued relative frequency analysis (p. 52)
also imputes counterfactualizability where the reference-class
has an indeterminately large number of members, since the
impact of any additional members might be expected to be
accommodated within the specified interval. But where pro-
bability is conceived of as point-valued and the reference-class is
of finite size, a relative frequency analysis does not impute
counterfactualizability since in that case the probability-
judgement is normally supposed just to report actual ratios of
occurrence. Similarly the indifference theory does not impute
counterfactualizability, because it supposes a probability-
judgement to state the ratio of the number of outcome-types of a
certain specified kind to the total number of mutually co-
ordinate outcome-types and to say nothing about what this ratio
might have been if there had been an additional co-ordinate
outcome-type over and above the actual totality of them.
Thus (though the point is often unnoticed) different theories
about the analysis of probability may in effect attribute different
logical properties to probability-judgements. And we shall have
to bear this in mind when we consider the relevance of different
analyses to the various different purposes for which probability-
judgements are required.

§ 13 T H E A P P R O P R I A T E N E S S O F D I F F E R E N T C O N C E P T I O N S O F
PASCALIAN PROBABILITY TO DIFFERENT PURPOSES
The Pascalian calculus permits a wide range of variation in the
logical properties of probability-judgements, which are then
useful for correspondingly different purposes. For example, to
judge a purely aleatory probability we require a conception of
probability, like the indifference-theory's one, that makes
correct judgements of probability necessary, extensional, impli-
citly general, and non-counterfactualizable. But where out-
comes cannot be assumed to be determined by pure chance, we
need probability-judgements that are contingent, extensional,
implicitly general, and counterfactualizable or ones that are
contingent, non-extensional, implicitly general, and counter-
factualizable. There are good reasons for supposing that a
variety of different conceptions of probability are operative here,
rather than just a variety of different ways of measuring the same
underlying parameter.
§15the Analysis of Probability 91
Different conceptions of Pascalian probability are suitable for
different purposes, depending partly on the epistemology and
partly on the logic of the judgements employing those con-
ceptions. Indeed, by taking into account all the possible com-
binations of logical micro-features, we might conceivably obtain
a greater potential diversity of probability-judgements than the
catalogue (§ § 6-11) of six standard Pascalian analyses suggests.
Different analyses of probability, it has already been shown
(§ 12), have different implications about whether judgements
of probability are necessary or contingent, sentence-related or
predicate-related, extensional or non-extensional, implicitly
general or implicitly singular, counterfactualizable or non-
counterfactualizable. Yet the Pascalian mathematics of prob-
ability is quite neutral in regard to preferring a particular
alternative on any of these five logical issues. The classical
calculus admits of interpretation in any of the numerous dif-
ferent senses that can be generated by combining the various
alternatives in the different ways that are possible. Clearly, if a
probability-judgement had to be contingent, extensional, and
predicate-related, for example, some kind of relative frequency
interpretation would have to be accepted, while if it had to be
necessary, non-extensional, and sentence-related, some kind of
interpretation in terms of logical relations would be required.
But there is no mathematical reason—no reason inherent in the
uninterpreted calculus—to argue on such grounds in favour of
any particular interpretation or semantical analysis.
Nor could one argue thus on the basis of ordinary linguistic
usage or of current scientific practice. Even if in ordinary usage,
or current practice, only one particular interpretation were
assigned to the Pascalian calculus, it would still be possible to
query whether that interpretation was the only useful one or
even whether it was as useful as this or that available alternative.
Moreover, what is useful for one purpose may be useless for
another. It follows that the appropriate question is not 'What is
the correct semantical analysis of probability-judgements?'
Rather we should ask 'Which semantical analysis of probability-
judgements is best suited to which kind of context?'
For example, in a context in which the probabilities being
judged relate to a game that is assumed a priori to be one of pure
chance, an objective idealist (p. 44) version of the indifference
92 The Foundations of Pluralism in§15
theory is appropriate. In specifying the nature of the game we
have to postulate a domain of statutorily basic outcome-types,
such as the two sides on which a presumably perfect coin can fall
on any one randomly executed toss or the six sides on which a
presumably perfect die can fall on any one randomly executed
throw. And from that postulate the principle of indifference
allows us to derive, as a necessary truth, the probability of any
particular basic outcome-type or the probability of a particular
compound outcome-type (such as the probability of two tosses'
both landing heads). Correspondingly we can say that the
establishment of these probabilities is an a priori, not an empir-
ical task. In this context any claim to be measuring propensities
of the gaming-device would appear to be misrepresenting an a
priori issue as an empirical one.
So, since the probability, in this sense, of a particular
outcome-type is determined by the rules that legitimate the
range ofbasic outcome-types, it risks being affected if in our own
judgements we choose to name the outcome-type in some other
way than the rules do. If the rules mention a side named '5', but
not a side named 'my favourite side', then the probability of
throwing my favourite side is not derivable a priori from the
rules and certainly cannot be equated on their authority with the
probability of '5'. In short, because true judgements of pro-
bability in the relevant indifferentist sense are necessarily true,
they are also non-extensional—i.e. resistant to the inter-
substitution of accidentally co-extensive terms. Moreover, so
long as the set-up considered continues to be one of pure chance,
any unconditional evaluation for the probability of, say, a
throw's being '5' will be implicitly general and apply also to the
probability of a throw's being '6', and so on. But the probability
of an outcome-type's being '5', given that it is nameable by an
odd number, would not be the same if a more-than-six-sided
die were in play. So judgements of aleatory probability are
non-counterfactualizable. In sum, a purely aleatory context
requires a conception of probability that satisfies an indif-
ferentist analysis and makes correct judgements of probability
necessary, non-extensional, implicitly general, and non-
counterfactualizable .
Suppose, however, the game is not assumed a priori to be one
of pure chance, but the question whether the game is one of pure
§15the Analysis of Probability 93
chance is treated as an open, empirical issue, perhaps because
there are reasons to think the coin or dice might be biased
(although the bias may be a stable one). Clearly that issue may
be put in the form: how closely does the empirically estimated
probability of a particular outcome approximate to the pro-
bability of that outcome which is calculated a priori in indif-
ferentist terms? So in order to take the issue seriously we need to
operate here with another, «o«-indifferentist conception. We
want to know how closely the value of the one function
approaches the value of the other in regard to the same
outcomes. More specifically, alongside the indifferentist con-
ception we need here also a conception of probability that allows
judgements of probability to be contingent and based on
observed trials or samples, as in the calculation of life-
expectancies for an annuity. These judgements must still be
implicitly general, but, as they now quantify over a domain of
outcomes, not of outcome-types, they would now be counter-
factualizable (so far as any bias or absence of bias is stable). Such
a conception fits either a relative frequency or a propensity
account, depending on whether or not there is a requirement for
extensionality and finitude.
Perhaps someone will be tempted to object that what is at
stake here, in the issue whether the game is one of pure chance, is
not the closeness with which the value of one probability-
function approaches the value of another but rather the close-
ness with which results from one method of measuring the value
of a particular probability-function approximate results from
another method of measuring the value of the same function.
Indeed, that would presumably be the gloss that a personalist,
belief-theoretical account would wish to put on the situation.
But what is this single parameter, then, that is allegedly being
measured in two different ways? Clearly it is not the actual
strength of a person's belief that is being measured in such cases:
no actual betting behaviour is being observed (p. 64). At best a
personalist might claim that what is being measured—in two
different ways—is the strength which his belief ought to have.
And then a realist could still rejoin that the propriety of a par-
ticular belief-strength must depend here, as elsewhere, on its
correspondence with some objective correlate like a ratio of
chances or a relative frequency, so that, if two different measure-
94 The Foundations of Pluralism in§15
ments are both correct, two different parameters are being
measured.
Moreover, in yet another situation we might be dealing
neither with an assumed game of pure chance nor with the acci-
dental vagaries of a man-made gambling device, but with the
physical properties of a particular example of this or that natural
kind. For example, we might need to judge the probability that a
neutron will decay in six minutes, where the assertion of a con-
nection between physical property and natural kind imposes a
requirement of non-extensionality. Since such a judgement will
also be contingent, grammatically singular, implicitly general,
and counterfactualizable, it operates with a concept of pro-
bability that satisfies a propensity analysis. Indeed, the physical
tendencies of natural kinds are paradigm cases for that analysis.
If you want to warn a non-smoker against the dangers of
smoking, you need to be able to cite counterfactualizable pro-
babilities. And that brings out another reason why the apparent
plurality of probability-concepts cannot be reduced to a
plurality of modes of measurement that are used in connection
with a single probability-concept. We have to think of
probability-judgements not only as conclusions that may be
reached by different methods of calculation or measurement,
but also as premisses from which different kinds of conclusion
may be drawn in accordance with differences in their logical
structure. Issues about exterisionality, implicit generality,
counterfactualizability, and so on are crucial here. For example,
the conception of probability on which we can base a warning
about smoking or about working in an asbestos factory is not the
same as that which relates only to accidental issues like the pro-
bability that a member of a certain conference is staying in a
specified hotel (pp. 88-9).

§ 14 T H E N E E D T O S U P P L E M E N T P A S C A L I A N J U D G E M E N T S BY
NON-PASCALIAN ONES

Various unsatisfactory proposals have been made, such as those


by H u m e and Laplace, for inferring the probability that a parti-
cular individual of a specified kind has a certain characteristic
from evidence about the relative frequency with which indivi-
duals of that kind have the characteristic. We need here to attend
§ 15 the Analysis of Probability 95
to the factors affecting inferability from an implicitly general
conditional probability to a singular unconditional probability.
This inferabilty is a matter of degree, which Keynes called
'weight'. The weight of an argument is quite independent of its
probability and has a non-Pascalian structure. There is a role for
the concept of weight on more than one analysis of Pascalian pro-
bability, but weights can at best be ranked, not measured, within
a particular category of judgement.
It is unnecessary, and perhaps impossible, to give a plausibly
complete list of the ways in which different analyses of Pascalian
probability are appropriate in different contexts. Different
possible combinations of underlying logical distinctions (such as
the distinctions examined in § 12) may well generate a wider
spectrum of possible analyses than the six familiar ones that were
discussed in § § 6-11. But all that is required her.e is to have
shown (in § 13) how different analyses, out of those discussed in
§ § 6-11, may be appropriate in different contexts: the need for
pluralism is deeply rooted. After all, we think of quantities of
apples as volumes for the purpose of storage, as weights for the
purpose of cooking, or as numbers for the purpose of distri-
bution among children, and we evaluate people differently if we
are considering them as athletes, as philosophers, or as friends.
Why should we not also conceive probabilities differently for
different purposes? Just as my harvest of apples may be smaller
than yours by one criterion and larger by another, so too the pro-
bability of this coin's falling heads on the next two tosses may be
0.25 as a calculation in a game of chance but 0.5 as a measure of
the tosser's subjective belief.
It will also be worth while enquiring whether there are even
some contexts of probabilistic reasoning to which no analysis of
probability is appropriate that imputes a Pascalian structure.
We have seen already (pp. 17-21) that some logically possible
ways of assessing evidentially warranted credibility have a non-
Pascalian structure. They do not conform to an additive
principle for negation. O r they do not conform to a multi-
plicative principle for conjunction. So the question arises
whether those logically possible ways of assessing credibility
have any real value. For, though they certainly occur sometimes
in ordinary speech, it does not follow from this that their use is
necessary or valuable for any important purposes.
96 The Foundations of Pluralism in§15
It turns out that there is at least one important use for a non-
Pascalian criterion of evidential strength—Keynes's concept of
'weight'—that is ancillary to the use of any Pascalian judgement
for one particular type of purpose. You cannot make the
Pascalian judgement without raising the non-Pascalian issue. So
in this context the argument for a non-Pascalian analysis is not
aimed at recommending the use of a non-Pascalian judgement
in place of a Pascalian one in certain inferences but at elucid-
ating what further kind of evaluation needs to be made when
we do use a Pascalian judgement in such an inference. We shall
come later (§§ 20-26) to consider the case for using a non-
Pascalian conception of probability in place of a Pascalian one in
certain contexts. And that rather bolder enterprise may seem
less shocking if we first consider the use of a non-Pascalian con-
ception in alliance with a Pascalian one.
Consider first some of the difficulties that arise where the pro-
bability that a particular individual of a specified kind has a
certain characteristic is to be judged on the basis of evidence
about the relative frequency with which individuals of that kind
have the characteristic.
One of the earliest ideas about this—discussed by Hume, for
example —is the so-called 'straight rule'. Where the evidence
12 13

is that in an observed sample of Bs the relative frequency of^4s is


r, the straight rule authorizes a conclusion that, on the evidence,
the probability of the next B's being A is also r. If 19 out of the
previous 20 ships to put to sea have returned safely, there is
according to the straight rule a probability on that evidence
that the next ship will do so also.
But there are two strong objections to allowing any general
validity for this rule. The first is that it attaches no importance to
the absolute size of the evidential sample, so that an observed
relative frequency of, say, in a sample of 20 individuals
justifies a predictive probability of -55-just as much as does an
observed relative frequency of in a sample of 2,000 indivi-
duals. And the second objection to the straight rule is that it
attaches no importance to the representativeness of the evi-
dential sample, so that an observed relative frequency of, say,
1 2D. Hume, A Treatise of Human Nature, ed. L.A. Selby-Bigge (Oxford: Clarendon
Press, 1888), 134 ff.: see § 4 above.
13 The name is due to Carnap, Logical Foundations of Probability, p. 227.
§15the Analysis of Probability 97
in a sample of ships that were all in sound condition can
apparently provide evidence for a prediction about a ship that,
as it turns out, has a rotten hull. These two objections are parti-
cularly strong where the observed relative frequency is 1 and the
straight rule, by therefore authorizing a prediction with prob-
ability 1, makes no allowance for the uncertainty of the future
or the risk of error, even when the evidence consists of a very
small and unrepresentative sample.
Laplace's 'rule of succession' seems at first sight to be much
14

more suitable. According to this rule, where E reports evidence


that in a sample of n Bs the number of /4s is m, we may infer that,
on the evidence E, the probability of the next B's being A is
m + 1
n+ 2
Hence, even when m/n—the relative frequency of/4s among the
5s—is equal to 1, the probability of the prediction is still less
than 1. And this probability will also then be sensitive to the
absolute size of the sample and approach asymptotically to 1 as
the size of the sample increases without limit. Thus the rule of
succession seems to be an intuitively plausible way of grading
the importance of sample-size in relation to certainty of pre-
diction by enumerative induction, where the evidence is
uniformly favourable.
But, although it is an improvement on the straight rule in this
respect, the rule of succession attaches no more importance than
the straight rule does to the representativeness of the sample.
Even a very large sample may in practice be heavily biased. And
the rule of succession generates some obvious paradoxes when
applied to empirical issues. Thus it implies that observably rare
events are less rare than, rather than as rare as or rarer than,
they are observed to be. For example, if only 1 out of 99 ships has
already come to grief the rule of succession tells us that on this
evidence, the probability that the next ship to sail will come to
grief is . Again, if there is no evidence, we have m = n = 0. So
according to the rule of succession there is then a probability of j

1 4 Laplace, A Philosophical Essay on Probabilities, p. 19. The rule is discussed by, among
others, J . Venn, The Logic of Chance, 2ndedn. (London: Macmillan 1876) 171-83, and
J . M . Keynes, A Treatise on Probability (London: Macmillan, 1921), 372-83.
98 The Foundations of Pluralism in§15
that the next B will be A. But, since this probability derives from
the absence of any evidence, and holds good whatever A and B
are, it can hardly constitute an empirically learned fact about the
actual world. Indeed there is analogous trouble even when
evidence is not entirely missing. An inherent fault in the rule of
succession is that it gives implausibly precise results on the basis
of very small samples.
The rule of succession is therefore at best applicable only
where we are reasoning about a supposedly chance mechanism.
And, even where the rule of succession is applied to such a sup-
posedly chance mechanism, there must not be more than two
possible outcomes for any one trial of the mechanism. Otherwise
contradictions may result. For example, suppose that an urn
contains only white, black, and yellow balls, and that we draw
first a white ball from it, next a black one, and next a yellow
one—replacing each ball after it has been drawn. The
probability that the fourth ball to be drawn will be white is,
according to the rule of succession,
1+ 1 2

By parity of reasoning, this is also the probability for the fourth


drawing to be black, and the probability for the fourth drawing
to be yellow. But such a set of probabilities contravenes the
Pascalian principle of additivity. According to that principle the
probabilities of the three possible outcomes should sum to 1, and
in fact they sum to f.
Moreover, Laplace's rule of succession turns out not even to
be a necessary consequence of Pascalian axioms. Rather the
probability-function embodying it is just a special case of c*,
which itself is just one out of the infinitely many different c-
functions—all conforming to Pascalian axioms—that can be
based, as Carnap showed (§ 11), on a choice of measure for
logical range. Even where the rule of succfession is applicable it
15

carries no greater inherent authority than does any of the other


1 5 See R. Carnap, Logical Foundations of Probability, pp. 227-8; A, Pap, An Introduction
to the Philosophy of Science (London: Eyre and Spottiswoode, 1963), 201-2; and I.
Niiniluoto, 'Inductive Logic as a Methodological Research Programme', in M . L . D .
Chiara et at., Logic in the 20th Century (special vol. of Scientia\ Milan, 1983), 83.
§15the Analysis of Probability 99
functions that are co-ordinate with it as a priori measures of
logical probability.
How are we to do better than this?
Let us assume that the evidence about the relative frequency
in the sample grounds a corresponding judgement of probability
for the whole population concerned. This judgement can be
understood either in terms of a relative frequency analysis or in
terms of a propensity one. The problem is then how we may
determine our empirically based entitlement to infer from that
judgement and any other relevant premisses to a probabilistic
conclusion about a particular individual. If, for example, our
premiss is that the general probability of a ship's returning safely
is -53-, under what circumstances is it safe to infer that the pro-
bability of the next ship's returning safely is The key to the
solution of the problem lies in distinguishing the assessment of
this inference's credibility from the assessment of the premissed
general probability—a distinction that is excluded by the way in
which the straight rule and the rule of succession approach the
problem.
Let us consider an example in detail. Suppose that, on the
basis of a randomly selected sample, the estimated probability
that a person will survive to the age of sixty-five, on the evidence
that he is a lorry-driver, is 0.8. No doubt the worth of the estim-
ate may be rated by some approved statistical parameter, such
as a Neymanian coefficient of confidence, that depends on the
size of the sample (pp. 118-19). But the judgement itself is an
explicitly general judgement of conditional probability, in the
sense that it does not assert anything about a particular person.
From it we can derive its instantiation for a particular person,
say, Smith, only in the form of an implicitly general
judgement—that is, only on the assumption that our reference
to Smith does not add to or subtract from the evidence on which
the probability is conditional. Now, on the unconditional issue
no such assumption is possible. If we want to infer Smith's
unconditional survival prospects—that is, not the probability that
he will survive to sixty-five on the evidence known about him,
but just the probability that he will survive to sixty-five—we
cannot avoid the need to allow for the fact that he, Smith, may be
specially circumstanced in some relevant way: lorry-driving
may be his work, but hang-gliding may be his hobby. The
100 The Foundations of Pluralism in§15
statement of the unconditional probability is implicitly singular,
not implicitly general. Survival to the age of sixty-five may be
quite common among lorry-drivers in general but much rarer
among hang-gliding lorry-drivers. So inferability to the
estimate that the unconditional probability of Smith's surviving
to sixty-five is 0.8 must depend on how big a spread of the
relevant facts about Smith is included in the premisses along
with the general judgement of probability that pertains to it.
And this is true however large the size of the sample on which the
estimate is based.
Schematically we can write the propositional function or
predicate-schema Hx here for anyone's surviving to the age of
sixty-five, Ex for anyone's being a lorry-driver and a hang-glider
and . . ., Hs for Smith's surviving to the age of sixty-five, and Es
for Smith's being a lorry-driver and a hang-glider and . . .
Then we can say that, if the premisses are the estimated evalua-
tion p(Hx|-Ex) = n and the observed fact Es, the derivability of a
correspondingly estimated evaluation p ( H s ) = n depends upon
how big a spread of the relevant facts is predicated by E. Indeed,
it is just because that derivability varies in this way that (as we
have already seen—p. 20 and pp. 60-1) we cannot identify
statements of conditional probability like p(^4|£) = n with
conditionalized statements of probability like 'If B, then p(^4) =
n' that permit inferences of uniform strength, by modus ponendo
ponens, from B to p(^4) = n.
Note too, by the way, that p(Hs) need not equal p ( H x ) here,
since p ( H x ) is explicitly general and is just the general pro-
bability of a person's surviving to the age of sixty-five, while
p(//i-) is the implicitly singular probability that Smith, in all his
special circumstances, will do this. So we cannot use Bayes's
theorem about the relation between monadic and dyadic pro-
babilities (see p. 23) to help the derivation. That theorem relates
p(Hx\Ex) to p ( H x ) , not to p(Hs). Similarly, the term 'statistical
syllogism', which is sometimes used to describe such an
inference from p(Hx\Ex) = n and Es to p ( H s ) = n, is peculiarly
inappropriate for that purpose, because the kind of probability
talked about in the major premiss differs from the kind of pro-
bability talked about in the conclusion (since the former is expli-
citly general and the latter is implicitly singular).
Philosophers sometimes invoke here Carnap's requirement of
§15the Analysis of Probability 101
total evidence. For a conditional probability-judgement to be
applied as the major premiss of a prediction or explanation of an
individual fact, they say, it must be based on all the available
evidence that is relevant to the conclusion. Or, if they prefer a
relative-frequency theory of probability to a logical-relation
one, they may invoke Hempel's requirement of maximal
specificity for the reference class or Salmon's requirement that
the reference class be the broadest homogeneous class of which
the individual concerned is a member. But in practice such16

requirements can rarely, if ever, be satisfied. Even to know just


the available evidence (physical, meteorological, geological,
astro physical, epidemiological, economic, socio-political, etc.)
bearing on a person's survival to sixty-five one would have to
work away indefinitely, since what is not learned today might,
with sufficient effort, be learned tomorrow. And almost
certainly, if one were thinking in terms of relative frequencies,
one would soon be reduced to a reference-class of one
member—the person himself—so that no statistical data could
be compiled. Accordingly inferability here is normally a matter
of degree, not of all or nothing. What is important in practice is
for the evidence to have as big a spread as is permitted by the
methodology of estimation, by pertinent economic constraints,
by the current state of relevant scientific knowledge, by the
nature of the subject-matter, and by any other limits to enquiry.
It follows that at least comparative judgements of evidential
spread have to be made, and it may be worth while investigating
whether different quantities of evidential spread can also be
ranked or measured. Even where some theory about economic
rationality assumes perfectly informed agents and therefore
treats all judgements of probability as being on the totality of
possible evidence, we still need to be able to evaluate the less
perfect states that exist in real life in order to make due
allowances in applying the theory to actual situations. In the real
world people are occupied with what is better and what is worse,
not with what is perfect. But obviously important questions will
arise about the structure of those judgements that assess
1 6Carnap, Logical Foundations of Probability, p. 211; C.G. Hem pel, Aspects of Scientific
Explanation and Other Essays in the Philosophy of Science (New York: The Free Press, 1965),
397; and W . C . Salmon, The Foundations of Scientific Inference (Pittsburgh: University of
Pittsburgh Press, 1966), 91.
102 The Foundations of Pluralism in§15
evidential spread. In particular we need to ask: how are those
judgements evaluated and do they themselves have a Pascalian
structure?
Keynes was the first to broach this issue seriously, but only in
a rather tentative and guarded way. He held a logical relation
17

theory of probability (as remarked in § 11) and treated the pro-


bability o f i o n B a s a property of the argument from B to A. The
probability of the argument depends for its value, he thought, on
the balance between the favourableness and unfavourableness
of the evidence that B states in relation to A. But he considered
that there may be another respect in which some kind of quant-
itative comparison between arguments is possible.
He wrote:
This comparison turns upon a balance, not between the favourable
and the unfavourable evidence, but between the absolute amounts of
relevant knowledge and relevant ignorance. As the relevant evidence
at our disposal increases, the magnitude of the probability of the
argument may either decrease or increase, according as the new
knowledge strengthens the unfavourable or the favourable evidence;
but something seems to have increased in either case,—we have a more
substantial basis upon which to rest our conclusion. 18

Keynes gave a name to this basis by saying that an accession of


relevant evidence increases what he called the 'weight' of an
argument. Thus the weight of an argument is independent of the
correctness or incorrectness with which such-or-such a pro-
bability is assigned to the argument. And it is also not neces-
sarily determined by the probable error of the argument's
conclusion (where that conclusion assigns a value to a mag-
nitude). So weight does not depend on mere numbers of evid-
ential instances. Keynes remarked that, metaphorically, the
weight of the argument from B to A measures the sum of the
favourable and unfavourable evidence that B states for A, and
the probability measures the difference. But he did not suggest
any method by which weight might be measured and in fact
admitted that often one cannot even compare the weights of
different arguments. He thought that, 'in deciding on a course
of action, it seems plausible to suppose that we ought to take
17 J . M. Keynes, A Treatise on Probability, pp. 71-8.
1 8 Ibid. 71.
§15the Analysis of Probability 103
account of the weight as well as the probability of different
expectations'. But he found it difficult to think of any clear
example of this, and he did not feel sure that the theory of evid-
ential weight has much practical significance.
Because Keynes held that probability should be thought of as
a logical relation (and because he apparently had no conception
of argument, or deducibility, from zero premisses), he could
hardly allow a place for judgements of unconditional probability
other than as mere ellipses of 'ordinary speech'. A fortiori he
19

could hardly appreciate the existence of the problem of how to


grade our entitlement to detach an unconditional probability
from a conditional one. It is scarcely surprising, therefore, that
Keynes admitted to doubting whether the theory of weight had
any practical significance, since by his very treatment of pro-
bability as a function of ordered pairs of propositions he had cut
himself off from the possibility of articulating the nature of this
significance.
But Keynes was right to insist that the weight and the pro-
bability of an argument are independent of one another.
Though both are properties of ordered pairs of propositions (if
we adopt a logical-relation account of probability), we can have
arguments where both weight and probability are high,
arguments where both are low, and arguments in which one of
the two is high and the other low. We can assess the weight of an
argument without knowing the value of its probability, and we
can assess its probability without knowing its weight. Moreover,
we can assess the weight of an argument without knowing the
size of the sample from which its probability has been estimated;
and we can grade that estimate by some appropriate statistical
criterion (such as a Neymanian coefficient of confidence:
pp. 118-19) without knowing the argument's weight.
I have been assuming here that, when talking about weight,
what we are interested in is the degree of derivability of an impli-
citly singular monadic probability judgement, such as the deriv-
ability of p ( H s ) = n from premisses p(Hx\Ex) = n and Es. If,
instead, what is to have its degree of acceptability evaluated on
the evidence is the statement Hs itself, and if' weight' is the name
given to this degree of acceptability, weight will not then be
1 9 Ibid. 7
104 The Foundations of Pluralism in§15
altogether independent of the probability stated in the explicitly
general judgement p(Hx\Ex) = n. Indeed, under those
conditions weight will not measure the absolute amount of
relevant evidence stated by Es, but rather the extent to which the
spread of evidence available is sufficient to justify accepting Hs
on the ground that its probability on that evidence is sufficiently
higher than that of any rival hypothesis. And Levi accuses 20

Keynes of begging the question because, according to Keynes,


weight does measure the absolute amount of relevant evidence.
But there is nothing question-begging in Keynes's discussion of
the subject. He and Levi are just using the same term to talk
about different parameters.
It is also worth nothing that judgements of weight have a
different logical structure from judgements of Pascalian pro-
bability. This may be shown both in regard to negation and also
in regard to conjunction.
Consider negation first. Because Pascalian probability is
additive, p(Hx I Ex) = n is necessarily equivalent to p(not-Hx\Ex)
= 1 — n and p ( H s ) = n is necessarily equivalent to p(not-//f) =
1 — n. From this it follows that the derivability of p(not-//f) = 1
— n from premisses p(not-//*|.E*) = 1 — n and Es must be just
the same as the derivability of p ( H s ) = n from premisses
p(Hx\Ex) = n and Es, even when both derivabilities are very
low, or very high, and therefore not mutually complementary.
Correspondingly the weights of the arguments from Es to Hs
and from Es to not-i/f must be reckoned the same. Hence the
additivity of Pascalian probability in effect ensures the non-
additivity of weight. Weight is therefore a non-Pascalian para-
meter of evidential strength. Indeed, on Keynes's definition,
though not on Levi's (see above), one can evaluate the weight of
the argument from Es to Hs without coming to any implicit con-
clusion whether, on balance, Es favours Hs or not-Hs.
Keynesian weight is logically independent of Pascalian
probability.
Again, if the only evidence we have about Smith is that he is a
twenty-five year-old lorry-driver, the number of relevant facts
that we know about him is the same whether we are considering
the probability that he will survive to the age of sixty-five or the
2 0 I. Levi, Gambling with Truth (New York: Knopf, 1967), 142.
§15the Analysis of Probability 105
probability that he will marry before the age of thirty or the pro-
bability of the conjunctive prediction that he will do both. And,
even if we also know that Smith is engaged (which is relevant to
his date of marriage but scarcely relevant to his survival to the
age of sixty-five), the weight of evidence for the conjunctive con-
clusion is just as strong as, but no stronger than, the weight of
evidence for the less weightily supported conjunct. So the weight
of an argument for a conjunctive conclusion need not normally
be less than the weight of the least weighty argument for one of
its conjuncts, even though the probability of the argument for the
conjunction—because of the standard multiplicative principle
for conjunction (§ 3)—is normally less than the probability of
the argument for either conjunct.
Nor is the need to assess weight alongside Pascalian pro-
bability confined to those probabilities that are conceived of as
logical relations. Keynes himself recognized that, if instead we
think of a probability as the frequency with which members of
one specified class are also members of another, the weight of
such a relative frequency can be increased by a relevant parti-
tioning of its reference class. The relative frequency of survivors
to the age of sixty-five among those lorry-drivers who are also
hang-gliders has a greater weight than their relative frequency
among lorry-drivers in general, even though the relative
frequency itself is lower. Of course, because the relative
frequency analysis of probability cannot (see pp. 48-9) apply to
implicitly or explicitly singular judgements, we must again
envisage that the conception of probability used in the
conclusion about Smith's survival prospects will differ from the
conception of probability used in the major premiss on which the
derivation of that conclusion is based. But the strength of our
entitlement to make the derivation will still depend on the
relevant weight, though here—in the context of a major premiss
about a relative frequency—weight has to be regarded, like pro-
bability, as a property of an ordered pair of classes rather than as
a property of an ordered pair of propositions.
Again, assessments of weight also have a function to perform
in association with personalist conceptions of probability.
Coherence alone is an insufficient condition for the rationality of
a betting-system. Even if no one may make a Dutch book
against you (see p. 60), your bets may still be irrational because
106 The Foundations of Pluralism in§15
you have taken too little of the available information into
account. Perhaps you should have checked at least on how your
favourite horse performed previously on a race-track as muddy
as today's. And the Bayesian procedure of conditionalization
(pp. 68-9) is a personalist's strategy for increasing weight. As
the Bayesian takes into account more and more beliefs about
relevant items of evidence, his probabilistically graded strength
of belief in a certain hypothesis may fluctuate. It may go down
when some items are taken into account and up when others are.
But, whatever its level at the later of two successive stages, that
level will be the result of more considerations than was the level
at the earlier stage. So the weight of the belief, on the accu-
mulated evidence, will be greater. Of course, if all relevant
evidence could always be identified and exploited, every
exercise of Bayesian method could be pushed through to
completion. But in practice, as already remarked, the require-
ment of total evidence is a rather Utopian ideal, and there is,
therefore, undeniable room for comparisons of weight to be
associated with declarations of personal probability.
Nor is there any difficulty in seeing how assessments of weight
may also be associated with judgements of propensity-type
Pascalian probabilities, since these may be either singular or
general. But, where an indifference-type analysis is appropriate,
assessments of weight have no role to play. Once the alternative
outcomes in a supposed game of chance have been specified, any
a priori calculations that ensue depend on just that specification
alone. Weight of evidence comes into the picture only if we try to
determine the probabilities of those outcomes empirically. And
so, because writers about probability have so often derived their
primary stimulus to reflection from a priori problems about
aleatory models, it is scarcely surprising that many of those
writers have found it easy to overlook the issues that arise about
weight. Similarly, an excessive concentration on examples
involving given, not derived, monadic probabilities has tended
to blind decision-theorists to the fact that in deliberating under
uncertainty we often need to assess weights as well as
probabilities. 21

But how exactly is weight to be assessed? As we have seen it is


2 1 See e.g. R.C.Jeffrey, The Logic of Decision (New York: McGraw-Hill, 1965).
§15the Analysis of Probability 107
independent of the size of the sample from which a probability is
estimated. So how can it be gauged?
It cannot be right to say, though it is often said, that, because a
personalist analysis ties probability to strength of belief, such an
analysis allows a probability judgement to reflect awareness of
the weight of the evidence as well as awareness of the degree to
which that evidence tends in a specified direction. For, as you
learn more about a particular horse, say, your personal odds on
its winning the race may either shorten or lengthen. Personal
odds are a poor indicator of felt weight if, with the same increase
of weight, they can become either shorter or longer.
It is sometimes suggested instead that, if the probability that a
person assigns to his belief on given evidence may be quantified
in terms of the lowest odds at which, given the evidence, he
would accept a bet on the truth of his belief, then the weight of
the given evidence in relation to that belief may be taken to be
reflected in the amount that he is prepared to bet. A person
might be expected to put a larger sum at risk when there is more
evidence from which to estimate appropriate odds. But other
considerations also may affect our attitude towards the size of a
bet. Suppose that there is a great deal of evidence, and that this
evidence suggests the appropriateness of very long odds. Would
you really be willing to risk losing just as large a part of your
fortune then as you would risk losing if the odds, on the same
amount of evidence, were much shorter?
It is also sometimes suggested that the weight of a probability-
judgement may be taken to vary inversely with the mathe-
matical expectation of gain from the acquisition of further
relevant evidence.. But this suggestion is open to at least two
cogent objections. First, one may well need to estimate the
weight of already acquired evidence independently of the
expectation of gain from the acquisition of further evidence, in
order to be able to compare the greater utility of having an
evidentially even weightier answer to one question than of
having it to another. It may be much more important to have
such an answer to the question 'What is the probability that
Smith will survive to the age of seventy?' than to the question
'What is the probability that Smith's cat will survive to the age of
fifteen?' Secondly, what are we to say when, for example, a vital
eye-witness has died without ever disclosing what he saw, or
108 The Foundations of Pluralism in§15
when no eye-witness at till was present? The expectation of any
kind of gain from further research in that direction may then be
zero, but the weight of evidence about what actually happened
would not normally be thought to have been increased because
of the missing data. This is because the weight of the evidence
already obtained is being assessed by comparison with some
supposed list of relevant kinds of evidential fact, not of actually
discoverable relevant fact. So, even ifwe had all the available evid-
ence, our argument might still not have maximal weight. Some-
times, for example, the prosecution cannot find enough evidence
to prove guilt beyond reasonable doubt, even though someone
must undoubtedly have committed the crime in question.
Nor will it do to equate superiority in weight with superiority
in Pascalian probabilistic relevance. If p(A\B &.B^) is not equal
1

to p(A | Z?,), then B is said to be relevant to A, given B ; and, if the


2 1

difference between p(A IS,&B ) and p(A\B^) is greater than that


2

between p(^4|S,&S ) and p(A\B^), then B may be said to be


3 2

more relevant than B to A, given B .- For example, in assessing


3 t

the probability of a thirty-year-old's survival to the age of


seventy, his previous medical history may be more relevant than
his sex. But, when a relevant premiss is thus added to the
evidence on which £ probability is assessed, the extent of
increase in weight does not depend for its size on the extent of the
new premiss's relevance, as can easily be shown. Suppose a set
of evidential items B B , . . . B in regard to a hypothesized
lt 2 100

conclusion A. Suppose too that quite a lot of these items, taken


on their own, ground low probabilities in favour of A, quite a lot
ground high probabilities, and quite a lot ground intermediate
probabilities at various levels. One way of ordering these items
would be to begin with those highly in favour of A, then proceed
with those slightly less in favour, and so on down, ending up
with those highly in favour of not-^4. In such a carefully
graduated order the extent of the relevance of each new piece of
evidence after the first, would normally tend to be small. So if
the weight of the probability were to be affected by the extent of
the relevance of each incremental piece of evidence, the addi-
tional effect on the overall weight would be minimal. But, if
instead the evidential premisses were ordered so as to alternate
as violently as possible between favourable and unfavourable
items, the overall effect on the weight would be very different, if
extent of relevance was allowed to affect the issue at each incre-
mental step. Hence, if we allow the extent of an added premiss's
§15the Analysis of Probability 109
relevance to affect the cumulative weight of an argument, we
could end up with different weights for arguments to the same
conclusion from logically equivalent conjunctions of premisses,
just because we calculated these weights on the basis of different
orderings for the conjoined premisses. Of course, the order in
which different items of evidence are stated may have a different
psychological effect on some people. But it ought not to affect the
logical cogency of the evidential statement as a whole. So the
weight added by a new piece of evidence is not to be measured by
the magnitude of its relevance.
Indeed, it looks as though weight cannot be measured at all, but
only compared or, at best, ranked within a fairly narrow field of
comparison. Roughly, we need to identify, for a particular
category of hypothesis, those families of mutually exclusive pre-
dicates that contain at least one predicate the ascription of which
is relevant to at least one hypothesis within the category, in the
way that, for example, the predicate 'has already had a coronary
incident' is relevant to any hypothesis about thirty-year-old
person's expectation of survival to the age of seventy. We need
then to order those families of predicates, perhaps exploiting for
the purpose any differences in extent of relevance, And where 22

we concern ourselves only with arguments from premisses that


contain just predicates belonging to the first family, or just those
predicates plus predicates belonging to the second family, or just
predicates from each of the first three families, we shall be able to
rank the weight of the argument in accordance with the number
of different predicate-families instantiated in the premisses.
So it seems that the Keynesian weight (i.e. evidential spread)
of an argument is a (sometimes rankable) characteristic that has
an important function to perform in our reasonings about
Pascalian probability but has a logical structure that defies any
attempt to map it on to the Pascalian calculus.

§ 15 H O W A R E D I F F E R E N T C O N C E P T I O N S O F P R O B A B I L I T Y
POSSIBLE?

Both realist and idealist interpretations are possible for non-


Pascalian principles of probability. Different types of pro-
2 2 See L.Jonathan Cohen, 'Twelve Questions about Keynes's Concept of Weight',
BritishJournal for the Philosophy ofScience, 37(1985), 263-78, for further details. Foraclaim
that the weight of a probability may be equated with its degree of counterfactualizability
see L. Jonathan Cohen, The Dialogue of Reason (Oxford: Clarendon Press, 1986), 182.
110 The Foundations of Pluralism in§15
bability, whether Pascalian or non-Pascalian, are generated by
differences within the gradation of inferability that are analogous
to differences between different systems of deducibility, as when
we compare the additivity issue in the theory of probability with
the completeness issue in the theory of deductive systems.
In § 14 1 showed how we need a certain type of non-Pascalian
assessment of the strength of the evidence for or against a
specified proposition in order to complement one use to which
Pascalian conditional probabilities are often put—namely, use
as premisses for the derivation of the corresponding singular and
unconditional probabilities. We need this type of non-Pascalian
assessment in order to deal with that dimension of credibility
which rests on the weight of the evidence as distinct from that
which rests on the extent to which the evidence is favourable
or unfavourable. Note, however, that we have so far been
considering evaluations of weight only in association with
evaluations of Pascalian probability. For a weight-dependent
evaluation of evidential support that was operative on its own we
should need to regard the weight of one proposition, B, for
another, A, as greater than zero only where B is on balance in
favour of A: p(A\B) would have to be greater than 0.5 but it
would not matter by how much. We should thus be interested
only in directed weight—weight that on balance pushes our
conclusion in a specified direction—and such a weight vector
would obviously have a different negation principle from pure
weight. If B states any positive weight of evidence for A, it will
now state none for not-^4. But the conjunction principle for
directed weight will be the same as that already determined for
pure weight (§ 14).
Now, if either directed or undirected weight is a useful mode
of non-Pascalian assessment, there is no reason to suppose that
its underlying, non-Pascalian structure may not be shared by
other modes of assessment also. That is to say, just as there are
various probabilistically useful interpretations of the Pascalian
calculus of chance, so too there may be various useful inter-
pretations of at least one non-Pascalian calculus. Indeed, it is
already clear ( § 9 ) that, by following a non-additive principle
for negation, judgements of propensity can take on a non-
Pascalian logic instead of a Pascalian one, so that we have here a
realist version of non-additive probability alongside the idealist
§ 15 the Analysis of Probability 111
version that is constituted by judgements of directed Keynesian
weight. Correspondingly, alongside the non-multiplicative con-
junction principle for Keynesian weight (§ 14), we have
standard systems of many-valued logic in which the truth-value
of a conjunction is no lower than the lowest truth-value of a
conjunct —rather than being determined by a multiplicative
23

principle as in Reichenbach's Pascalian system. Of course, I


have not yet proposed axioms for a suitable non-Pascalian
calculus (see § 21 below). And we should not expect that every
useful interpretation of the Pascalian calculus is matched by a
useful interpretation of such a non-Pascalian one. Neither the
indifference analysis nor the relative frequency analysis are
likely to have useful non-Pascalian counterparts, because they
depend essentially on the existence of appropriate given ratios.
But from what has already been said it is clear that we may legit-
imately expect such counterparts for Pascalian analyses in terms
of propensities, logically characterizable relations, multiple
truth-values, or belief-strengths.
Nor should anyone be surprised that there are so many
different types of Pascalian probability-judgement, if he is
prepared to think of probability as a gradation—whether by
subjective or objectiye criteria—of inferability. For, if we
consider the situation where inferability is not a matter of
degree, but of all or nothing, a wide variety of differences is
familiar. There are systems of deductive inferability that claim
to depend on necessary truths, such as laws of logic, and systems
that claim to depend on contingent truths such as supposed laws
of nature. There are logics of terms and logics of propositions.
There are systems that are open to the substitutivity of co-
extensive predicates; and systems that are not. There are
systems that allow no inferences about an individual that cannot
be made also about any other individual, and systems that do
not. There are systems that are counterfactualizable, in the
sense that they license inferences irrespective of the premisses'
truth-values, and systems that are not counterfactualizable, in
the sense that they license inferences only from true premisses.
Accordingly, on the assumption that judgements of probability
grade inferability, we must expect that these judgements will
exhibit a corresponding diversity. And we have, indeed, already
2 3 SeeN. Reseller, Many-valued Logic (New York: McGraw-Hill, 1969), 131.
112 The Foundations of Pluralism in § 15
found (in § 12) that just such a diversity exists and forms the
basis for the widely differing analyses of probability that philo-
sophers have sometimes produced. Probability-judgements
may be either necessary or contingent. Probability may be
viewed as a function either of terms or of propositions. State-
ments of probability may be either extensional or non-
extensional. They may be either general or singular, and either
counterfactualizable or non-counterfactualizable. Differing
analyses of probability are made possible by the variety of
possible combinations of such options that they can exploit. So
all this pluralism may be viewed as operating within the frame-
work that is constituted by the unifying conception of pro-
bability as a gradation of inferability.
Moreover, that framework is capable of generating non-
Pascalian options as well as Pascalian ones. Consider again the
situation where inferability is not a matter of degree, but of all or
nothing. Then, assuming that negation is always expressable in
such a deductive system, one can distinguish between systems of
inferability that are complete and systems that are incomplete. A
system is complete, in the relevant sense, just so long as any
sentence of the system is inferable from the postulates of the
system if and only if its negation is not inferable. Now, if pro-
bability is to be conceived as degree of inferability, the deductive
inferability of A from B is that limiting case of probability where
\J(A\B) = 1 and the non-inferability of not-^4 from B is stated
correspondingly as p(not-/41B) = 0. So, ifp(*4|.B) = 1 if and only
if p(not-^4|5) = 0, it looks as though we are dealing with a
complete system and that we may expect that in general the pro-
bability or inferability of A, given B will vary inversely with the
probability or inferability of not-^4, given B. What emerges is
thus the standard Pascalian principle for negation: Y>(A\B) =
1 — p(not-/l \B). Accordingly completeness as a property of
certain deductive systems, may be viewed as a limiting case of
probabilistic additivity (or complementationality), and pro-
babilistic additivity (or complementationality) may be viewed as
a generalization of completeness. And it is mathematically
provable that if degrees of inferability are to be measured over a
Boolean domain, with a complementational principle for
negation, all the other principles of the Pascalian calculus must
§15the Analysis of Probability 113
also be obeyed. In a consistent but incomplete deductive
24

system, however, there must be at least one sentence such that


neither it nor its negation is inferable. That is to say, when infer-
ability is described in terms of probability, there must be an A
and B such that both p(/l|S) = 0 and also p(not-^4|£) = 0
because neither A nor not-/l is inferable. It follows that any
generalization of inferability in an incomplete system will have
a non-Pascalian structure, in virtue of excluding an additive,
i.e. complementational, principle for negation. Indeed, the
additivity issue is about as important for differentiating between
Pascalian and non-Pascalian probabilities, as the parallel lines
issue is for differentiating between Euclidean and non-
Euclidean spaces. 25

No doubt many other concepts of probability are also


possible. The aim of the present chapter has not been to
construct an exhaustive semantic and syntactic taxonomy for
such concepts, let alone an epistemological taxonomy for their
applications (empirical/a priori, idealist/realist, on balance of
evidence/on amount of evidence, etc.). The aim here has been
merely to illustrate the nature of the pluralism that seems to be
justified in our philosophy of probability, to argue that this
pluralism is quite compatible with accepting that there are
special roles for particular concepts of probability in particular
contexts, and to elucidate why the existence of such a rich
plurality of probability-concepts is only to be expected. So, if we
now revert to the earlier problem about the gradation of ampli-
ative induction, we can approach it with a relatively open mind.
There is no single concept of probability that we can know in
advance must somehow be fundamental to the evaluation of
inductive reasoning. Rather, we shall have to investigate quite a
variety of concepts, both Pascalian and non-Pascalian, in regard
to their suitability for this particular purpose.
Note, however, that in investigating which Pascalian or
non-Pascalian concepts of probability are most suitable for this
2 4R . T . Cox, The Algebra of Probable Inference (Baltimore: Johns Hopkins Press, 1961),
4-5, and J . R . Lucas, The Concept of Probability (Oxford: Clarendon Press, 1970), 33 ff.
2 5For a study of some early work on non-additive probability see G. Shafer, 'Non-
Additive Probabilities'in the Work of Bernoulli and Lambert', Archive for the History of
Exact Sciences, 19 (1978), 309-70.
114 The Foundations of Pluralism in§15
or that purpose we leave on one side any psychological questions
about which concepts are actually used by people in this or that
context of reasoning. Philosophy here, as elsewhere, pursues a
normative rather than a descriptive end. During the last twenty
years experimental psychologists have conducted a lot of
research into the ways in which people without any relevent
training measure or compare probabilities. But unfortunately
this research has sometimes been marred by an a priori assump-
tion, on the part of the investigators, that use of the term 'pro-
bability' (and its equivalents in other languages) admits of only
one syntax and only one semantics. By imputing their own
favoured interpretation to every use of the term the investigators
have then put themselves in a position to claim the existence of
systematic, non-accidental errors in the judgements of prob-
ability that are made by the subjects of their experiments. But,
when due regard is paid to the variety of available conceptions of
probability that may be used in appropriate contexts, the case
for supposing the existence of such errors seems much less well
supported. M a n y of the subjects may have been trying to answer
different questions from those that the investigators took them-
selves to be asking. If one quantity of apples is larger than
another by weight, a person who says that it is smaller has not
necessarily made a mistake if he is comparing the two quantities
by number or by volume rather than by weight. So too in 26

regard to probabilities.
2 6 Relevant references are given in n. 32 to ch. II.
IV
The Pascalian Gradation of Ampliative Induction

§ 16. I N D U C T I V E P R O B A B I L I T Y U N D E R A R E A L I S T C O N S T R U A L

Though an indifferentist conception of probability may be used


to grade some kinds of generalization, it is inapplicable to
ampliative induction in science. Nor can we measure the latter in
terms of the relative frequency with which a hypothesis that is
supported by such-or-such evidence turns out to be true. And if
we try to replace the generalization that all ,4s are Bs by the asser-
tion of a very high propensity-type probability of A, given B
(where estimates of this probability can be graded by standard
statistical procedures), we find that the applicability of the
generalization to an individual case has to be underwritten by a
judgement of Keynesian weight, which introduces a non-
Pascalian element into the situation.
W e have seen elsewhere (§ 13) how different conceptions of
Pascalian probability may be appropriate for different kinds of
task, and also (§ 14) how for at least one purpose a non-
Pascalian conception is indispensable. Clearly we need some
system of gradation for ampliative induction, so as to be able to
assess the extent to which a proposed generalization about all
members of a certain class is supported by this or that body of
evidence about some of its members. But is that need best served
by a Pascalian or a non-Pascalian system? T h e present chapter
will explore various Pascalian options and evaluate their vi-
ability. And it will begin with what were earlier ( § 5 ) called
'realist' interpretations of the Pascalian calculus.
In certain circumstances an indifference-type conception of
probability (see § 6) seems applicable to the task. Consider, for
example, a game of chance where the colour of a playing card is
to be predicted. T h e card in question is to be drawn at random
from a well-shuffled standard pack of fifty-two cards and
restored to the pack when its colour has been noted, and this is to
be done five times, with the pack reshuffled on each occasion.
116 The Pascalian Gradation of§17
Then consider how Pascalian probability might be used to grade
progressively the validity of the generalization 'Every card
drawn will be a red one', on the basis of uniformly confirmatory
evidence from each of its instances. When only one confirm-
atory instance is known and the outcomes of four card-
drawings remain as yet unknown, the probability that each of
the four remaining instances is red will be (that is, j X y X ' j
X 2), since we can apply the indifference theory here, as in any
game of pure chance, and also treat each outcome as being quite
independent of every other. Similarly, after two favourable
draws the probability will be after three favourable draw j,
after four favourable draws 7, and after five draws 1. In each
case we just need to calculate the ratio of chances. So in this kind
of situation the indifference theory provides us, in the absence of
disconfirmatory instances, with a Pascalian measure of induct-
ive support that has the property essential to enumerative
induction: it increases as the number of confirmatory instances
increases. Or, more exactly, and more generally, we can say
that, if m is the number of alternative outcomes that are possible
at each draw in such a game and i is the number of draws that are
still to be carried out, then the probability that the generalization
is true, given that all.the evidence so far is favourable to it, is
1I mi. (This result is reached by the following step. Let 'A' be 'the
outcome of draw 1 is red and the outcome of draw 2 is red and
. . . and the outcome of draw n is red', where there are n
drawings in all. L e t ' B ' be 'The outcome of draw 1 is red and the
outcome of draw 2 is red and . . . and the outcome of draw n — i
is red'. Then in virtue of its being a priori given that there are
just n drawings we replace 'A' in Y>(A\B) = 1 /m'' by 'The out-
1

comes of all the drawings are red'.)


It follows that if human scientists could be supposed to play a
system of analogous games of chance with natural processes (for
example, predicting the outcome of a particular type of medica-
tion), the evidential support available for successful scientific
hypotheses could be measured by a Pascalian probability-func-
tion that was interpreted in accordance with the indifference
theory. But unfortunately the analogy breaks down at several
points. The number of co-ordinate alternative outcomes that are
possible in any one trial of the issue investigated may be infinite,
indeterminate, or at least unknowable. The total number of
§22Ampliative Induction 117
possible trials may also be infinite, indeterminate, or at least
unknowable. And, even more importantly, the trial outcomes
may not be independent of one another: where causal laws are
operative the same cause is, other things being equal, expected
to produce the same effect. In short, science is not a game of
chance with Nature, and we can grade enumerative induction
by an indifference-type Pascalian probability only when we are
generalizing about outcomes over a selected finite set of trials in
a supposedly genuine game of chance. 1

Turning next to the relative frequency conception of probab-


ility, we can see that it could be applied to the gradation of
ampliative induction, with a cogency appropriate to the circum-
stances, if we could estimate the relative frequency with which a
hypothesis that is supported by such-or-such a kind of evidence
is true. But where (as often in natural science) a hypothesis is a
2

generalization about an infinitely, or indeterminately, large


domain, we can never know for certain (see § 23 below) that it is
true, though we can know the evidence supporting or opposing
it at any one time. So the relative frequency at issue would
instead have to concern the number of hypotheses that remained
unfalsified after some canonical interval. For example, the
probability of the particular hypothesis that all swans are white,
on the evidence that all the thousand observed swans are white,
would be, with a rather low weight (see § 14 above), the relative
frequency with which generalizations about plumage-colour
remain unfalsified for, say, a hundred years, after a thousand
favourable, and no unfavourable, instances have been
observed. And we might obtain a rather higher weight (see § 14
above) for the probability of our hypothesis if the evidential
instances were suitably varied in regard to features that are
sometimes relevant to plumage-colour (season, sex, climate,
and so on).
Nevertheless, any such gradation of a particular exercise in
ampliative induction depends inevitably on contingent histor-
ical facts about the collection of relevant evidence. If research
funds have not been made available for thorough ethological

1 See also A.J. Ayer, Probability and Evidence (London: Macmillan, 1972), 30-3.
2 See e.g. H. Reichenbach, Experience and Prediction (Chicago: Chicago University
Press, 1961), 397 ff.
118 The Pascalian Gradation of§17
and experimental enquiries into plumage-colour, for example,
we cannot expect many generalizations on that topic to have
been even proposed, let alone falsified. But, if there has been a
lot of funding over a long period, we may expect not only that
many such generalizations have been proposed, but also that
enough research has been done in the area for many relevant
generalizations, whether explicitly considered or not, to have
been falsified. In short, other things being equal, the probability
assignable in this way to a particular hypothesis, on the evidence
available at the time, will tend to vary inversely with the size of
past funding for research in that field. So this method of assign-
ing probabilities to particular hypotheses introduces a quite
paradoxical type of dependence on the direction of previous
research. One might well have supposed instead that the more
research goes into a subject, the more reliable become the hypo-
theses that eventually issue: they are supported, as it were, by
piles of eliminated conjectures. But the method of assessment
under consideration measures the reliability of a hypothesis by a
probability that is equal to the relative frequency of unfalsified
hypotheses in the field of enquiry.
It may therefore seem plausible to suggest instead that
enumerative induction involves establishing some very high
propensity-type probability (see § 8) on the basis of a very large
sample. Thus, if the generalization to be inductively supported
is that all Bs are As, this generalization would be treated as being
equivalent to the estimate p(A | E ) — 1 — e where e is very small,
and the evidence would be constituted by a large sample of Bs, in
which the observed relative frequency of ^4s was at least 1 — e.
The relation between the evidential sample and the estimated
propensity would then be evaluated by some conventional statis-
tical procedure which relies ultimately for its validity on
Bernoulli's limit theorem (see pp. 21-2 above). For example,
we might aim at having an estimate that could be said to be
correct within 95 per cent confidence limits. Roughly, the 3

observed relative frequency (within a stated degree of approx-


imation) would then be forthcoming in appropriately large
samples on 95 per cent of occasions, given the correctness of the
3 SeeJ. Neyman, 'Outline of a Theory of Statistical Estimation Based on the Classical
Theory of Probability', Philosophical Transactions of the Royal Society of London, Ser. 236
(1937), 333 ff., esp. 348.
§22Ampliative Induction 119
estimate. And this would be fine if we were interested in the
properties of such samples, as in industrial quality-control,
where no more than a certain percentage of imperfect batches
can be tolerated. Or Fisher's method of significance-levels can 4

be used, whereby we can seek to avoid both the error of rejecting


the null-hypothesis—the hypothesis that, in respect of the rele-
vant relative frequency, the evidential sample is non-representa-
tive of the population as a whole—when we should accept it, and
also the error of preferring the null-hypothesis when we should
reject it. In the present kind of context, by determining conven-
tionally the level of improbability at which we shall regard an
evidential sample of a certain size as significant we can charac-
terize the type of composition that, if it occurs in the sample,
should make us reject the null-hypothesis. That is, our sample
should be treated as representative if its occurrence would other-
wise be highly improbable (for example, occurring on only 5 per
cent of occasions). And this again would be fine, within any
community of scientists that shared the same conventions about
significance-levels.
But in transforming the problem about inductive support for
a generalization into a problem about evidential samples for
estimating a very high propensity we must not lose anything
important from the characterization of the problem.
We have in any case to accept that, when this transformation
is made, contraposability will no longer be logically guaranteed.
'All not-vl are not-i?' is logically equivalent to 'All B are A' but
'p(not-S|not-^4) = «' is not necessarily equivalent to 'p(A | B) =
n'. Similarly, while 'AlI.Bare;4' logically implies 'All SCare A',
'p(y4|S) = n' does not necessarily imply p(A\BC) = n'. l

Sacrifice of these, and connected, implications is a price that has


to be paid by anyone who wishes to replace the problem about
generalizations by a problem about high Pascalian probability. 5

Perhaps the price is too high.


Moreover, there is another kind of implication that we
certainly cannot afford to lose, because it is integral to one of the
main purposes that the discovery of well-supported generaliza-
4 R.A. Fisher, Statistical Methodsfor Research Workers, 7th edn. (Edinburgh: Oliver and
Boyo, 1938), 120 ff.
5 As is recognized by, for example, H. Reichenbach, The Theory of Probability 2nd edn.
(Berkeley: University of California Press, 1971), 435-6.
120 The Pascalian Gradation of§17
tions can accomplish. From a reliable law we can draw reliable
conclusions in particular cases. That is how safe bridges are built
and ships are navigated successfully. And we can draw these
reliable conclusions in particular cases because the uniformity
stated by the generalization is known to hold good in the kinds of
circumstances prevailing in those cases. So the generalization
has to be well-supported by variative induction if it is to consti-
tute a reliable premiss for a sufficiently wide range of predictions
in particular cases. There might be many instances that would
support it by enumerative induction, but they might all have
occurred in other circumstances than those for which the par-
ticular conclusion is desired—in bridges not exposed to high
winds, for example. Analogously, if a judgement of high
probability is to constitute such a premiss, it must (as we saw in
§ 14) have high Keynesian weight. In other words its reliability
for the purpose is not established just by enumerative induction
from the size of the evidential sample but also by partitioning the
reference class in a way that ensures the applicability of the
generalized statement of probability to the circumstances of the
particular case. It follows that if we treat generalizations as high
propensity-type probabilities we cannot regard sample-size as
the only evidential parameter that should affect our employment
of the generalization. We have to look also at the range of other
relevant circumstances present in the evidential instances and to
use the non-Pascalian mode of appraisal that the assessment of
weight requires (§ 14).

§ 17. I N D U C T I V E P R O B A B I L I T Y U N D E R A
RANGE-THEORETICAL CONSTRUAL

Carnap's range-based analysis of inductive support does not


apply satisfactorily to support for open-ended generalizations.
Hintikka's system avoids Carnap's difficulty here. But both
systems tend not to differentiate adequately between enumer-
ative and variative induction and fail to represent the type of
reasoning which determines our choice of properties to control as
a basis for variative induction. It looks as though such a system
should refer ultimately to overlaps between ranges of physically
possible worlds, not of logically possible worlds. But any attempt
to define confirmation thus would be circular.
§22Ampliative Induction 121
Carnap's range-based theory oflogical probability seems at first
sight to provide us with a more promising approach to the
gradation of ampliative induction. His preferred confirmation-
functions certainly guarantee that new predictions of a par-
ticular kind get more and more probable as more and more old
ones of that same kind turn out to be true. For example, as we
have already seen (in § 11 above), it is a feature of his 1950
system that c*(Hc,Ha&Hb) > c*(Hc,Ha). And in Carnap's final
version of this system it was proved more generally by Humburg
that the degree of confirmation of a singular statement h is {

always increased when in addition to original evidence stated in


e, which is consistent with h,, we accept a new evidential state-
ment h which reports another instance of the same property as h
2 l

reports: that is c(h,,e&k ) > c(h,,e). This theorem is known as


2
6

the principle of instantial relevance.


But induction is concerned with evidential support for
generalizations, and unfortunately Carnap's system does not
produce satisfactory confirmation values for that kind of state-
ment. Even his favoured confirmation-function c* assigns a
very small degree of confirmation to such a statement, whatever
the evidence, when the supposed number of individuals in the
universe is very large, and assigns zero when this number is
infinite. (That is because in Carnap's system a generalization
that everything has a certain property is equivalent to a conjunc-
tion of singular statements which say, in turn, of each individual
in the universe that it has that property, and the larger the
number of singular statements thus conjoined the lower the
measure of the conjunction's range—since there are bound to be
fewer possible worlds in which it is true.) But in actual science
the number of individuals in a given domain of investigation is
normally unknown, so that it cannot enter into calculations of
the extent to which the evidence supports a given generaliza-
tion. Nor is it normally assumed that generalizations over an
infinite-membered domain are incapable of acquiring any pos-
itive degree of evidential confirmation, and that any one such
generalization is thus as worthless as any other.
6 J. Humburg, 'The Principle of Instantial Relevance', in R. Carnap and R.C.
Jeffrey (eds.), Studies in Inductive Logic and Probability, i(Berkeley: University ofCalifornia
Press, 1971), 225-33.
122 The Pascalian Gradation of§17
Carnap argued that in practice we do not need confirmation
values for generalizations over large numbers of instances. This
is because an engineer who is building a bridge, and wishes to
apply certain physical laws to the task, is chiefly interested in
that bridge's not being a counter-instance to any of the laws on
which he is relying, or in there being no such counter-instance
among all the bridges that he will construct during his lifetime.
He is not so interested in the immense, perhaps infinite, number
of other instances 'dispersed through all time and space'. So
Carnap held that 'what is vaguely called the reliability of a law is
measured not by the degree of confirmation of the law itself but
by that of one or several instances'. And he therefore sought to
capture this idea by defining a new concept of 'instance
confirmation'. I f g is a generalization over a certain domain of
individuals, the instance confirmation of g on the evidence
stated by e is defined as being equal to c*(h,e), where h states in
effect that a specified individual not mentioned in e fulfils the
generalization g. 7

Instance confirmation, however, as defined by Carnap, is


very far from measuring what is often understood to be the level
of evidential support for a generalization. For example, there
may be quite strong evidential confirmation for the prediction
that the weather will be sunny in Oxford tomorrow, but strong
evidential disconfirmation for the generalization that the
weather is sunny in Oxford every day. Nor can we easily dis-
pense with assessments of inductive support for at least some
propositions that are genuine open-ended generalizations. It is
quite true that engineers, navigators, physicians, and others
employed in day-to-day technology can for the most part restrict
their horizons of intellectual interest. But, important as such
technological purposes are, they are not the only purposes
served by the advancement of scientific knowledge. Intellectual
understanding, and the explanation of why things happen as
they do, are also a legitimate objective. And that purpose
requires us to see individual phenomena or particular regu-
larities in the light of much more widely pervasive laws. T o
explain the positions of the planets, for example, on the
occasions on which these have been observed, we need to be able
7 Carnap, Logical Foundations of Probability (Chicago: Chicago University Press, 1950),
571-4.
§22Ampliative Induction 123
to cite laws of planetary motion like Kepler's or general laws of
motion like Newton's—not just knowledge of where the planets
will be tomorrow or next week or of where they were a minute
before they were observed. So questions are bound to arise
about the strength of evidential support that exists for hypo-
theses stating such laws. And these questions do not remain
unanswered. Scientists frequently draw inductive comparisons
between one law-stating hypothesis and another. Philosophers
of induction are therefore obliged to try and elicit the principles
of gradation that are implicit in those comparisons. They cannot
legitimately escape this obligation by confining their attention to
the needs of engineers and other technologists.
Carnap's theory of logical probability gets into difficulties
about inductive support for generalizations because it makes the
measures of a universal proposition's range depend in part on
the number of individuals in the domain. Accordingly Hintikka
proposed a type of range-measure that was, in the case of univer-
sal propositions, independent of this number. Suppose that 8

we are given k primitive monadic predicates. By means of


these predicates and the usual propositional connectives we can
partition our domain of individuals, whatever its size, into 2*
different kinds (empty or non-empty), distinguished by the
various complex predicates that describe them. For example,
with two primitive predicates F and G, we should have four
kind-describing complex predicates, namely, F,8LG, F&not-G,
not-F&G, and not-.F&not-G. Then we can describe a possible
world by a sentence describing all the non-empty kinds of indivi-
duals in it. Such a sentence, which Hintikka calls a ' constituent',
will be a consistent conjunction of sentences each of which
asserts the existence of at least one individual belonging to such-
or-such a kind, conjoined with a sentence asserting that every
individual belongs to one or other of those kinds. A constituent
thus determines a species of possible world, namely, that species
of possible world in which just such-and-such kinds are instan-
tiated, and because it is formulated in general terms—by means
of existential and universal quantification—a constituent may
be made true by each of many different state-descriptions (in
8 J . Hintikka, 'Towards a Theory of Inductive Generalisation', in Y. BarHilIel(ed.),
Proceedings of the 1964 Congress for Logic, Methodology and Philosophy of Science (Amsterdam:
North-Holland, 1965), 274-88.
124 The Pascalian Gradation of § 17
Carnap's sense of that term: § 11). Hintikka supposes that an
equal value should be assigned to each constituent, just as
Carnap's c* was based on assigning an equal value to each struc-
ture-description, and that the value of each constituent may be
divided among all the state-descriptions that make it true. A
sentence's range is then to be measured, as Carnap measured it,
by the sum of the values of the various state-descriptions in
which that sentence holds, and degree of confirmation analog-
ous to c* is to be defined in terms of the same ratio of range-
measures that Carnap used for the purpose:

' m*(e) •
[ J

In this system the measure of a generalization's range will be


quite independent of the number of state descriptions that make
any particular constituent true, and therefore quite independent
of the numbers of individuals of various kinds in the universe or
of the total number of individuals, because every generalization
has an equivalent that is just a disjunction of constituents with
the same complex predicates.
Hintikka's system thus avoids Carnap's difficulties about
inductive support for generalizations over large numbers of
individuals. It provides a realistic basis for comparing different
generalizations in relation to the degree of support available for
them from evidential instances. But it seems to fit elementary
generalizations like 'All swans are white' rather more easily than
it fits serious scientific hypotheses, such as those asserting one
physical parameter to be such-or-such a function of certain
others, like e = mc . Indeed, though it can apparently be
2

extended to languages containing polyadic predicates as well as


monadic ones, its extendability to languages containing higher-
9

order predicates as well as first-order ones has not been demon-


strated. Nevertheless, if the system provided a correct measure
of inductive support even for first-order generalizations about
monadic properties, it would be a substantial achievement. So it
calls for a careful assessment.

9 See I. Niiniluoto, 'Inductive Logic as a Methodological Research Programme', in


M . L . D . Chiararfa/., Logic in the 20th Century (special vol. of Sciential Milan, 1983), 93.
§22Ampliative Induction 125
Philosophers sometimes object to theories like Carnap's or
Hintikka's that, quite apart from questions about the number of
individuals presupposed, they make a proposition's degree of
confirmation, for given evidence, depend on choice of language
in another respect also, namely in regard to the number of
primitive predicates. T o this objection there are three plausible
answers. One is that, as we have already seen (p. 80), it is
possible to construct systems in which degree of confirmation is
invariant under changes in the choice of primitive predicates
within a particular family of primitive predicates. The second
is that the impact of introducing new predicates into range-
theoretical systems can be studied systematically. And the 10

third is that even within a single language it is possible to draw


useful comparisons on the assumption of the language's
adequacy.
But the fact remains that, so far as degree of confirmation is to
depend on choice of language, we need to know not only what
consequences flow from a change of language but also what kind
of considerations justify such changes. An adequate analysis of
inductive support that is based on a measure of range must treat
the choice of relevant terminology as a procedure implicit in
any chain of inductive reasoning. Such a choice is just as much
in need of analysis, or rational reconstruction, as are any sub-
sequent inferences that are drawn from the structure of the
terminology chosen. For example, we may suppose that the
term 'pregnant' was absent from the relevant terminology of
those who obtained the original experimented results on the tox-
icity of thalidomide, since they failed to control (p. 30) those
experiments for pregnancy. So what would, rationally, compel
the addition of this term to the relevant terminology? We need to
suppose some discernible incoherence between newly observed
empirical facts and the assessment of degree of confirmation
relative to previously observed facts. And the point is not just
that on the new evidence the degree of confirmation is different
but that even on the old evidence it was wrongly assessed. The
experimenters were too confident that, on the evidence they
had, thalidomide was non-toxic. Hence supposedly true assess-
ments of the form c(h,e) = n, ought to be regarded as empirically
1 0 See I. Niiniluoto and R. Tuomela, Theoretical Concepts and Hypothetico-Deductive
Inference (Dordrecht: Reidel, 1973), 165-95.
126 The Pascalian Gradation of§17
falsifiable and their method of possible falsification needs to be
analysed by inductive logicians, whereas in fact they are
represented by Carnap and Hintikka as a priori or true by
convention. In short, an adequate confirmation-metric for
natural science ought to be empirically self-correcting. And
neither Carnap's nor Hintikka's system has this property.
There is another feature of their systems that also needs to be
discussed. Carnap's emphasis on the importance of the principle
of instantial relevance (see above) has sometimes misled his
critics into complaining that his preferred confirmation-
function c* offers an analysis of enumerative, but not of varia-
tive, induction. But this is not quite the right complaint to make.
Carnap in fact pointed out" that c* takes on a higher value not
only when the total number of new instances is greater and other
features are the same, but also when, with equal numbers of new
instances, the number of different kinds from which the new
instances are taken is greater or the instances are distributed
more evenly among the kinds. So, superficially at least,
Carnap's preferred confirmation-function does seem to reflect
the requirements of variative induction, and Hintikka's does
also. But there are two underlying faults in the way in which a
12

range-theory achieves this.


First, it amalgamates enumerative and variative support
under a single measure, whereas in practice these are normally
held apart and may be cited for different purposes. Thus, if the
same scientific experiment is carried out a second time and has
the same outcome, that is taken as helping to establish the
genuineness of the experimental result. The result is then better
entitled to count as a datum that constitutes support, or refuta-
tion, for a proposed hypothesis. But, if on the second occasion
the experimenter fails to replicate the previous outcome, suspi-
cion naturally arises that the description of the first occasion's
relevant circumstances may have been incomplete or inaccurate
in one or more respects. Of course, the flaw might be in the
description of the second occasion's relevant circumstances. But

11 In R. Carnap, ' O n Inductive Logic', Philosophy of Science, 12 (1945), 72-97.


12 See J . Hintikka, 'Induction by Enumeration and Induction by Elimination', in
I. Lakatos (ed.), The Problem of Inductive Logic (Amsterdam: North-Holland, 1968),
191-216.
§22Ampliative Induction 127
one way or the other the relevant circumstances of the two
experimental trials must have been different if substantially
different outcomes were produced. And in this way enumerative
induction seems to be at work in our reasoning about the validity
or authority of experimental reports. On the other hand, vari-
ative induction seems to be at work in our reasoning about the
thoroughness with which such experiments test hypotheses. The
thoroughness mounts as a greater variety of supposedly relevant
factors are controlled in the tests, and support mounts at the
same time if the hypothesis that is being tested remains unfalsi-
fied by any authoritative experimental data. But this important
difference between the roles of enumerative and variative induc-
tion in experimental reasoning is totally obscured if a single
confirmation-function is made to amalgamate both parameters.
What is thus ignored is that we may wish to characterize some
experimental data as well replicated but as the outcome of a
rather weak test on a particular hypothesis, and other data as the
outcome of a test that is rather strong but poorly replicated. The
replication of a given experimental outcome, and the complexity
with which the experimental test is structured, are mutually
independent issues.
The second underlying fault in both Carnap's and Hintikka's
treatments of variative induction is that they assign equal impor-
tance to each kind of primitive predicate that may occur in the
description of evidential instances. That equality stems in turn
from the fact that in Carnap's preferred system each structure-
description is assigned an equal measure, and in Hintikka's
each constituent is assigned this. But in the world of real science
different variations of experimental conditions may well be of
very different value. For example, in an experiment to test the
toxicity of a pharmaceutical product the previous medical
history of a patient may be much more important than his or her
normal diet. Moreover, even if two experimental conditions
may be of equal value when occurring on their own, they may be
of very unequal value when each is combined with another parti-
cular condition. A new drug may be equally safe when taken in
combination with standard doses of aspirin or of barbiturate,
but not when taken in combination with both aspirin and alcohol
or with both barbiturate and alcohol. Ultimately indeed what
these facts call into question is the whole strategy of making
128 The Pascalian Gradation of§17
degree of confirmation depend on a measure of logical range,
whereby c(h,e) is defined as
m (h&e)
m 0)
What this definition does, in effect, is to make the degree to
which e confirms h depend on a measure of the logically possible
worlds in which h and e are jointly or severally true. But, if differ-
ent experimental conditions are to count differently in certain
cases, what is at issue is the physical effects of different conditions
in different circumstances. So in the end every kind of logically
possible world that is described by a structure-description or a
constituent would have to be evaluated for its potential of ocur-
rence. We should be measuring physical range, as it were, not
logical range.
T o take an extreme case, let us suppose that it is physiolo-
gically impossible for a man or a five-year-old girl to become
pregnant. T h e n in a language that included terminology for sex
and age differences and for pregnancy it seems just as wrong for
degree of confirmation to take into account the physiologically
impossible worlds in which such events occur as to take into
account logically impossible ones. Why should it matter at all, in
relation to a particular description of evidence, that it is logically
possible for the description to hold good in certain physiolo-
gically impossible worlds? Again, the idea behind relating
confirmation to the ratio of m(h&te) to m(«) is that e's confirma-
tion of h may be seen as a measure of the overlap between those
possible worlds in which both h and e are true and those in which
e is true. Degree of confirmation thus measures, as it were, the
extent to which the evidence narrows down the logical possibil-
ities to those in which h is true. But what scientists ought to be
interested in instead is the extent to which the evidence narrows
down the relevant physical possibilities. And here again, as
when we considered problems about the introduction of new
terminology, it looks as though any adequate assessments of
degree of confirmation by a range-theoretical method must be
empirical judgements based on what we think that we know
about physical possibilities and impossibilities, rather than a
priori judgements based on calculations from the structure of
our language.
§22Ampliative Induction 129
It may seem, therefore, that there is nothing irremediably
unsound in the range-theoretical approach to the analysis of
ampliative induction. In order to avoid Carnap's difficulties
about inductive support for law-stating generalizations, we need
to base our measurements on the assignment of values to con-
stituents rather than to structure descriptions. And in order
to do justice to the realities of variative induction we need to
assign these values in the light of what is physically possible
rather than of what is logically possible. But, unfortunately,
to require such an assignment is to make the whole procedure
circular. O n the one hand we need to derive our preferred
confirmation-function's assessments of inductive support from
premisses about the differing degrees of physical possibility that
it is appropriate to assign to different combinations of natural
conditions: on the other hand we need to base those assignments
on our preferred confirmation-function's assessments, in each
case, of the support that the presence of one such condition gives
to the presence or absence of another. The difficulty here is of the
same kind as that remarked earlier in the case of indifference
theories (pp. 45-6). Just as an indifference theory is circular if it
has to presuppose an appropriate set of comparative judgements
of probability over all the possible outcomes, so too a range-
theoretical definition of one's preferred confirmation-function is
circular if the range of a proposition has to be made an empirical
issue and based on assessments of degree of confirmation.
This does not mean that Carnap's type of confirmation
system has no legitimate domain of application. He himself
constructed an example —his only fully worked-out
13

example—in which the universe consists of players in a chess


tournament and their properties are being local or non-local,
being junior or senior, being male or female, and being the
winner or a loser. The evidence states the numbers of males and
females among local juniors, local seniors, non-local juniors,
and non-local seniors, and it is assumed that each of the ten
players has an equal chance of becoming the winner. It is then
possible to calculate a priori the degree to which the evidence
range-theoretically confirms a hypothesis like 'a man wins',
without even having to exploit the consequences of choosing one
1 3 Logical Foundations of Probability, pp. 382 ff.
130 The Pascalian Gradation of§17
confirmation-function rather than another. But this is like 14

calculating probabilities in a game of chance that has ten co-


ordinate possible outcomes altogether, with each of these out-
comes falling into one or more other categories. And indeed any
more complex example, in which structure descriptions (as for
Carnap's c*) or constituents (as in Hintikka's system) are
assigned equal values a priori, would also resemble a game Of
chance, where the basic equalities are stipulated a priori in the
rules. But we know no such set of rules for Nature. So the process
of Nature cannot be treated as one big game of chance, and we
have to do a lot of empirical work to discover the extent to which
it admits of this or that combination of circumstances. 15

§ 18. P A S C A L I A N G R A D A T I O N F O R V A R I A T I V E I N D U C T I O N

Variative induction is important enough for us to ask whether a


distinctively variative mode of inductive reasoning can be given
a Pascalian representation. Keynes's system of gradation for
inductive support rests on a principle of limited independent
variety. That principle establishes a non-zero probability for the
hypothesis under investigation, and the probability increases
with each evidential instance that differs in some respect from all
previous ones. But there are at least four objections that may be
made to this analysis of inductive support.
1 4 Carnap, Logical Foundations of Probability, p. 347.
15 Useful discussions of Carnap's inductive logic by J . G . Kemeny, A. W. Burks, H.
Putnam, and E. Nagel, together with Carnap's replies to each of these, are to be found in
P. A. Schilpp (ed.), The Philosophy of R. Carnap (La Salle: Open Court, 1963), 711-825,
966-95. Carnap's later ideas about inductive logic are to be found in 'Inductive Logic
and Rational Decisions' and 'A Basic System of Inductive Logic', in R. Carnap and
R . C . Jeffrey (eds.), Studies in Inductive Logic and Probability, i (Berkeley: University of
California Press, 1971), 5-171, and R.C.Jeffrey (ed.), Studies in Inductive Logic and Proba-
bility, ii (Berkeley: University of California Press, 1980), 7-155. The shift in Carnap's
ideas is discussed from a Popperian point of view by I. Lakatos; 'Changes in the Problem
of Inductive Logic', in Lakatos (ed.), The Problem of Inductive Logic, pp. 315-417. For an
attempt to show how Hintikka's kind of analysis can be applied under conditions of
conceptual change, see I. Niiniluoto and R. Tuomela, Theoretical Concepts. A reformula-
tion of Camap-Hintikka inductive logic is to be found in T.A.F. Kuipers, A Survey of
Inductive Systems (Dordrecht: Reidel, 1978). The variety of available confirmation-
functions is discussed in R. Carnap, The Continuum of Inductive Methods (Chicago:
Chicago University Press. 1952); and J . Hintikka, 'A Two-Dimensional Continuum of
Inductive Methods', in J . Hintikka and P. Suppes (eds.), Aspects of Inductive Logic
(Amsterdam: North-Holland, 1966), 113-32.
§22Ampliative Induction 131
So far we have been considering Pascalian modes of gradation
for those forms of ampliative induction in which a superior
number of favourable instances (with no, or hardly any,
unfavourable ones) establishes superior merit whether or not
these instances are thought to vary in any relevant respects. But,
if we distinguish effectively, as suggested in § 17, between ques-
tions about the strength with which certain test-results, if rep-
licable, support a particular generalization and questions about
the replicability of those test results, we can see that, at least in
natural science, mere multiplicity of instances is significant only
for the replicability of test-results and not for the strength of
support that they give.
The traditional argument for the superiority of variative
induction is that it avoids the possibility of biased sampling
which is inherent in enumerative induction. You may have seen
a large number of hares that were grey and none that were
not grey, but your observations were all of common hares
in summer so that you never saw a mountain hare in
winter—which would have been white. Hence perhaps you mis-
takenly concluded, by enumerative induction, that all hares are
grey, whereas variation of the species observed or of the season
of observation would have generated a more correct conclusion.
But this argument for the superiority of variative induction is
of somewhat limited scope. The more homogeneous the domain
over which you generalize (and elements of which you observe),
the less needful variative induction seems to be.
A more powerful argument rests on the assumption that the
generalization at issue is conceived to state a causal connection,
a law of nature, or a consequence of one or more of these, as
distinct from some accidentally operative uniformity. Charac-
teristically such a generalization would be supposed to apply to
counterfactual (possible but non-actual) instances as well as to
actual ones. If penicillin cures septicaemia, then an adequate
dosage of penicillin would have cured your cat's septicaemia last
year. Now enumerative induction seems incapable of support-
ing any implication about counterfactual instances. However
many evidential instances are observed they seem to have no
describable connection with non-actual instances. But variative
induction does establish such a connection. The wider the
variety of circumstances to which a generalization is seen to
132 The Pascalian Gradation of§17
apply, the wider the variety of counterfactual as well as of actual
instances that it may be said to cover. In variative induction the
potential hazards against which generalizations are checked are
properties, characteristics, or conditions, not individuals. So a
generalization that is supported by variative induction may be
conceived to embrace counterfactual instances within its
domain, just so far as these too exemplify the relevant proper-
ties, characteristics, or conditions.
Whether or not these arguments for variative induction are
acceptable, it is at least worth examining whether some Pascal-
ian conception of probability can be applied to an exclusively
variative form of ampliative induction, in which additional
favourable instances have evidential value only if they differ
from previous instances and from one another in some relevant
respects. Mill's attempt to carry out such a project suffered from
various defects that were noted in § 4 above. Other attempts
need therefore to be examined.
Keynes tried to construct a solution of the problem in terms of
his conception of probability as a logical relation. He postulated
that in any domain of inductive investigation a principle of
limited independent variety operates. He believed that Bacon
16

had explicitly recognized the importance of this principle and


that Mill too had used it though without explicit acknowledge-
ment. According to the principle all the properties or qualities of
the domain's individual elements are bound together in a
limited number of groups. Every characteristic belongs to at
least one group, within which the characteristics are invariably
connected with one another through the operation of a distinct
generative mechanism. But the same characteristic may belong
to more than one group, just so long as the generative
mechanism that produces all the characteristics within any one
group can operate independently of each generative mechanism
that produces the characteristics in another group. For example,
there can be a plurality of causes for any one type of thing, as
Mill also held (see p. 32). Moreover, there is some definite finite
number of distinct groups, although there need be no such
number of distinct individuals. And Keynes thought that events
in Nature could be treated as constituting one such domain of
1 6 J . M. Keynes, A Treatise on Probability (London: Macmillan, 1957), 258-61
§22Ampliative Induction 133
inductive investigation, and numbers another. Like Bacon
(p. 8), but unlike Mill or Carnap, he conceived of ampliative
induction as a logical procedure that was by no means confined
to the investigation of Nature.
Keynes then builds up his analysis of inductive support on the
basis of his assumption about limited independent variety.
His first step is to establish a non-zero prior probability for a
typical hypothesis. Suppose, for example, that we are investigat-
ing in some domain whether everything that has the charac-
teristic B also has the characteristic^. Since independent variety
is limited A has to belong to at least one out of a finite number of
groups of characteristics, though in advance of any specific
information A has an equal probability of belonging to any one
of them. And the same is true of B. Hence, if there are n distinct
groups, and each is of antecedently unknown composition, the
probability that A belongs to a particular group Gis at least 1 In,
the probability that B belongs to group G is also at least 1 In, and
accordingly the probability that they both belong to G is at least
1 In . It follows that there is antecedently some finite probability
2

that A belongs to all the groups to which B belongs and thus that
everything which has B has A. So Keynes's theory is at any rate
not hit by the difficulty that Carnap encountered, whereby any
generalization over an infinite domain had zero prior
probability.
Keynes next points out that each evidential instance of B that
differs in some respect from any previously noted instance of B
can help to eliminate the possibility of B's belonging to any
group to which A does not belong. Hence, as each such possib-
ility is eliminated we may treat the probability that everything
which has B has A as being correspondingly increased, since
each step reduces the number of distinct groups about which it is
not known whether B and A are both members.
Of course, Keynes recognized that by this method a numer-
ically definite probability could be obtained for an induction
only if we are able to make definite assumptions about the
number of independent equiprobable causal mechanisms at
work. But he pointed out that commOn-sense judgements of
inductive probability were normally qualitative or comparative
rather than numerical. People remark that certain inductive
arguments are stronger than others, or that some are very
134 The Pascalian Gradation of§17
strong. So Keynes was content that within his system inductive
evaluations would normally just bear relations of greater or
lesser to numerical probabilities according to the approximate
limits within which our assumptions place the number of
independent generative mechanisms. That is, his assessments of
inductive support were, as we should now say, interval-valued.
Another notable feature of Keynes's system was that
enumerative induction plays no part in it whatever. An additi-
onal evidential instance is of no inductive value unless, by
exhibiting some hitherto unremarked combination of charac-
teristics, it furthers the eliminative process. Nor does a parti-
cular location in space or time count as an inductively relevant
characteristic in the investigation of natural processes. So, if an
additional evidential instance is to have any value there, it must
differ from previously noted instances in respect of some charac-
teristic other than spatial or temporal location. Indeed in this
connection Keynes invokes what he calls 'the Uniformity of
Nature', according to which (as most investigators of Nature are
prepared to assume) mere position in time and space cannot
possibly affect, as a determining cause, any other characteristic.
It is therefore no objection to his system that at any one time of
investigation all the evidential instances are past or present,
none future. Pastness, presentness, and futurity must also be
inductively irrelevant if spatio-temporal location is. (A similar
result was achieved in Carnap's system, by his preference for
symmetrical c-functions—pp. 77-8).
Keynes acknowledged that the principle of limited indepen-
dent variety was crucial to his system. But he argued that this
principle did not have to be proved true, in relation to any parti-
cular domain for inductive investigation. It was sufficient, in his
view, for the principle to have a finite probability a priori, which
our experience has considerably increased. Indeed, he says, 'it is
because there has been so much repetition and uniformity in our
experience that we place great confidence in it'. But there are at
least four points that can be made against Keynes here, and the
approval that one accords to his theory must vary inversely with
the strength that one attributes to these points.
First, it is by no means obvious that actual human experience
has tended to increase the probability that the principle of
limited independent variety is valid for Nature. As natural
science progresses to new levels of understanding new properties
§22Ampliative Induction 135
of matter tend to be discovered and perhaps this is because
Nature's independent variety is unlimited.
Secondly, even if the principle's probability has undergone
some increase, this is quite compatible with its still being more
probably false than true. But, if the principle on which our
inductive reasonings rely is more probably false than true, then
on the balance of probability our inductive reasonings are with-
out any rational foundation. So Keynes has given us insufficient
reason to suppose that our inductive reasonings do have a
rational foundation rather than that they do not.
Thirdly, as well as these difficulties about the epistemology of
the principle, there are problems about its implications. Keynes
did not explain how exactly his method of inductive appraisal
assigns antecedent probabilities to hypotheses about dyadic or
relational characteristics as distinct from monadic ones, or to
hypotheses about quantitative characteristics as distinct from
qualitative ones. For example, if there is a non-denumerable
infinity of points on a scale of physical magnitude—a scale of
heat, say, or velocity—must there not be a corresponding
infinity of generative mechanisms? And if you prefer to identify
a single generative mechanism, such as molecular motion, as the
cause of heat, you will be grouping parameters (of temperature,
velocity, etc.), rather than single characteristics, into your
supposedly finite number of groups, and this will interfere with
the assignment of antecedent probabilities to hypotheses about
monadic, qualitative characteristics.
Fourthly, Keynes assumes that within any domain of induct-
ive investigation all mutually independent generative mechan-
isms are of equal inductive importance in relation to any one
hypothesis, just as J . S . Mill (see pp. 36-7 above) assumed that
in analogical reasoning every feature of similarity or difference
between the two entities concerned is equally important. Yet (as
already remarked in connection with Mill's account) the fact
that two patients resemble one another in literary interests, for
example, would normally be much less important for medical
prognosis than that they have the same viral infection in their
bloodstreams. 17

" Keynes's views on induction are criticized from a different point of view in
J. Nicod, Geometry and Induction, trans. M. Woods (London: Routledge and Kegan Paul,
1969), 173-242, where the importance of enumerative induction is defended.
136 The Pascalian Gradation of § 17
§ 19. I N D U C T I V E P R O B A B I L I T Y U N D E R A P E R S O N A L I S T
CONSTRUAL

The treatment of inductive support as a Pascalian probability


under a personalist interpretation has both advantages and dis-
advantages connected with its consignment of the calculations
that underlie the assessment of probability-values to a black-box
mechanism. A personalist analysis in terms of positive relevance
is more revealing. But while the former analysis presupposes the
Special Consequence Condition, the latter presupposes the
Converse Entailment Condition. These two conditions are not
co-tenable, and the positive relevance analysis has to be given
up. Moreover no Pascalian analysis can allow a theory to be well
supported when the evidence includes minor anomalies.

If we examine the applicability of a personalist conception of


Pascalian probability to the assessment of ampliative induction,
we can see that it has some advantages for this purpose over a
logical relation theory, while retaining whatever merit goes
along with treating degree of confirmation as having a Pascalian
structure.
It is clear that there is some such merit. For example, if the
results of testing our hypothesis turn out favourable to it but
were not otherwise to be expected, we normally think that our
hypothesis is better confirmed than if those favourable results
were in any case predictable from previously established
theories. A novel but successful prediction is always a feather in
the cap for a new theory. And this is reflected in any Pascalian
analysis of ampliative induction because of Bayes's law (p. 23).
For, according to that law, if the values of p ( H ) and p(E\H)
remain unaltered (but non-zero), the value of p(HIE) gets
bigger as that of p(£) gets smaller without reaching zero. The
lower the prior probability of the evidence (provided it is non-
zero), the greater the probability of the hypothesis upon that
evidence (other things being equal). Moreover, because of this
Bayesian relationship, a hypothesis that predicts or explains two
mutually independent items of evidence will have a higher
probability on that combined evidence than on either evidential
item alone, since p(E,&E ) will normally be less than p(£',).
2

But a personalist analysis does have some special advantages


also. It escapes several of the difficulties that beset a logical
§22Ampliative Induction 137
relation analysis when applied to the task of assessing ampliative
inductions. In particular, a personalist analysis says nothing
about how a person does or should determine his lowest accept-
able betting odds on a particular proposition. It treats the mind
of an inductive reasoner as containing a black box that takes in
all kinds of information and outputs acceptances or rejections
of wagers for the purpose of Bayesian conditionalization
(pp. 67-9). Consequently the strength of a person's belief in the
truth of a hypothesis is not necessarily tied either to enumerative
or to variative evidence, but his mind may be supposed to be fed
with either type of information or both. Correspondingly a
personalist analysis avoids having to face up to any special
problems about calculating Pascalian probabilities for general-
izations that are not restricted to the use of monadic first-order
predicates. Carnap and Hintikka had to wrestle with those
problems (see §§ 11 and 17) and Keynes's analysis suffered
from ignoring them. But the personalist's black box takes care of
all such difficulties for him. At the same time a personalist
analysis does allow judgements of inductive probability to be
sensitive to the different degrees of importance that are felt to
characterize different kinds of evidential variation. No a priori
principles of equality between different structure-descriptions,
or between different constituents, are written into the person-
alist's black-box mechanism. Again, Keynes's relatively
unobjectionable principle of the uniformity of Nature, whereby
differences of time or place do not count as inductively relevant
characteristics (p. 134) has a counterpart in de Finetti's
principle of exchangeability (pp. 69-70). But the more
controversial principle of limited independent variety has no
counterpart in an application of a personalist analysis, or
Bayesian methodology, to the assessment of ampliative induc-
tion. It is true that a rational man can hardly wager seriously on
the truth of a generalization over an infinite or indeterminately
large domain, since it would be possible to settle the bet only if
the generalization were falsified and the bet was lost. But the
generalization could, in principle, be over a domain that is as
large as you please, just so long as its size is finite and deter-
minate. So perhaps such a generalization could be adequate for
most, if not all, of the explanatory purposes that Carnap's
concept of instance confirmation (pp. 122-3) failed to serve.
A personalist analysis of ampliative induction does, however,
138 The Pascalian Gradation of § 17
have an important weakness which follows necessarily from its
consigning the calculations that underlie the assessment of
probability-values to a black-box mechanism. It tells us nothing
about how a person does or should determine his lowest accept-
able betting-odds other than that a rational person does so coher-
ently on the basis of whatever evidence is available to him. So the
analysis offers us no guidance on how to resolve controversies
between those rational people who disagree with one another
about the probability of a particular hypothesis. There may be
detectable flaws in the analyses offered by logical relation
theorists like Keynes, Carnap, or Hintikka. But at least they
addressed this problem, whereas the personalist analysis seems
just to duck the real difficulty. It does not take us behind the
scenes, as it were, and provide us with a breakdown on how
degree of confirmation is, or ought, actually to be determined.
Carnap even came to think that personalism articulates the
rational framework within which any choice of axioms for induc-
tive logic has to be made. But he could not accept it as a satis-
factory substitute for a range-theoretical approach to the analysis
of induction. 18

The situation is not quite so bad in this respect if the personal-


ist analysis for degree of confirmation is supposed not to be the
one that treats degree of confirmation as degree of probability
simpliciter, but the more sophisticated one that treats degree of
confirmation as the difference between prior and posterior
probabilities, as in Hesse' s account. Thus on the former view E
19

is said to give some positive confirmation to H if and only if


p(H\E) is greater than some chosen constant, such as But on
the latter view E is said to give some positive confirmation to H ii
and only if E is positively relevant to H—that is, if and only if
p(H\E) is greater than p ( H ) . And this thesis does purport to
provide at least a minimal insight into the calculations deter-
mining degree of confirmation. It does not tell us how the values
of p ( H ) and p(H\E) are, or ought to be, established. But it does
at least tell us to measure the difference between the one and the

1 8See R. Carnap, 'Inductive Logic and Rational Decisions', in Carnap and


R. Jeffrey (eds.), Studies in Inductive Logic i. 25-7.
1 9M. Hesse, The Structure of Scientific Inference (London-. Macmillan, 1974), 133-4.
§22Ampliative Induction 139
other, while retaining the feature (referred to above) that, other
things being equal, degree of confirmation is greater as p(-E) gets
smaller.
Another important difference between the implications of the
simple probability analysis and the implications of the more
sophisticated probabilistic relevance analysis has to be noted.
Consider a case where H logically implies H', and E gives
some positive confirmation or inductive support to H. One
intuitively plausible claim here is made by what is commonly
called the 'Special Consequence Condition'. According to this
principle E must give at least as much support to H' as it gives to
H. And that allows us to exploit whatever support exists for a
scientific theory whenever we apply a generalization which is
derivative from that theory. Another intuitively plausible claim
is made by what is called the 'Converse Entailment Condition'.
According to this principle, if H logically implies H', H'
supports H. And this fits well with the commonly held belief that
a theory acquires some support from each independently
established generalization that it entails.
But the Special Consequence Condition and the Converse
Entailment Condition do not belong well together. The simple
probability analysis implies the validity of the Special Con-
sequence Condition and does not imply that of the Converse
Entailment Condition. This is because the Pascalian calculus
gives us p(H'\E) > p(H\E) where p(H'\H) = 1, but it does not
ensure p(H\H') > \ where p(H'\H) = 1. O n the other hand the
positive relevance analysis implies the Converse Entailment
Condition and not the Special Consequence Condition. That is
because the Pascalian calculus gives us p(H\H') > p ( H ) , where
p(H'IH) = 1 and neither p(H) nor p(H') is zero or one, but it
does not ensure p(H'\E) > p(H') where p(H'\H) = 1 and
p(H\E) > p(H).
Indeed, not only do the simple probability analysis and the
positive relevance analysis have these importantly different
implications for the logical structure of inductive support. It is
also impossible to conceive of any serious analysis that could
accept both the Special Consequence Condition and the Con-
verse Entailment Condition. That is because, if both principles
are accepted, any proposition B may be shown to support any
other proposition A, as follows. Consider the hypothesis A&B.
140 The Pascalian Gradation of § 17
By the Converse Entailment Condition B supports this hypo-
thesis. But if B supports the hypothesis A&B, then, by the
Special Consequence Condition, B supports^.
It follows that, whatever our actual analysis of ampliative
induction, we have to give up one or both of these two principles.
Nor should that be surprising if we notice that the two principles
reflect inherently different conception of support. According to
the Special Consequence Condition the degree of support that
the evidence gives to a hypothesis as a whole flows down to all the
logical consequenes of the hypothesis. So this degree of support
is conceived as permeating throughout the content of the hypo-
thesis. Such a conception of support is typical of ampliative
induction, where the observation of certain evidential instances
is supposed to give a corresponding degree of support to a
generalization as a whole. Thus the existing evidence comes to
support predictions about new cases that are derivable from the
generalization. But no such permeation of support is implicit in
the Converse Entailment Condition. The support that H' gives
H, when H entails H', may apply just to that part of the content
of / / w h i c h enables it to entail / / ' . Indeed, the greater the addi-
tional content that //possesses, the less the support that H' gives
to H as a whole. And, in any case, in ampliative induction the
typical relationship between generalization and the reports of
evidential instances that support it is certainly not a relation of
entailment. 'All swans are white' does not entail's is a swan and
white'. So the Converse Entailment Condition does not get a
grip there.
Accordingly, in any analysis of inductive support we need to
ensure conformity with the Special Consequence Condition, not
the Converse Entailment Condition. It does not follow from this
that the simple probability analysis has to be accepted, since an
analysis in terms of Pascalian probability may not be the only
one that allows deduction of the Special Consequence Condi-
tion. But the positive relevance account has certainly to be dis-
carded, as an analysis for ampliative induction, since it implies
the validity of the Converse Entailment Condition, which we
have seen to be inappropriate for ampliative induction.
Moreover, there are two other good reasons why the positive
relevance analysis can be questioned.
The first is that the positive relevance analysis allows evidence
§22Ampliative Induction 141
which disconfirms or undermines a hypothesis to support a
conjunction of that hypothesis with any other hypothesis you
please. For example, imagine a situation where H and H' are
independent of one another and p(H\E) = 0.9, p(//'|.£) = 0.2,
and p ( H ) = p(//') = 0.3. W e shall then have p(H\E) > p ( / / ) ,
p(H'\E) < p(H'), and p(H&H'\E) > p(H&H'). So that in such
a case, according to the positive relevance analysis, E dis-
confirms or undermines / / ' and yet still supports the conjunc-
tion / / & / / ' . Maybe this might be acceptable if H' made only a
small and insignificant claim about the world, when compared
with H. W e might then feel inclined to concede that E on balance
supported the conjunction / / & / / ' , just because it supported the
major conjunct H. But unfortunately we could just as well be
faced with the opposite situation, in which E c a m e to support the
whole conjunction just because it supported the minor conjunct.
T h e second additional argument against the positive rel-
evance analysis is that, as Popper and Miller have shown, it 20

seems not to account for the ampliative element in ampliative


induction. Let us suppose a (logically contingent) scientific
hypothesis H, a statement of evidence E, and background know-
ledge B. Then, according to the positive relevance analysis, if
the conjunction H&CB entails E, we shall have p(H\E8CB) >
p(H\B) where 1 > p(£'IS) > 0. But H can be factored into two
elements HW — E and H\/E, such that, though E&LB supports
HWE (since p(H\/E\E&.B) = 1 while p ( H V E ) < 1), E under-
mines or disconfirms H\J — E (since p(H\/— E\E&B) <
p(//V - E\B). And yet //V — E, though thus undermined by E,
seems to be the ampliative element within the factorization of H
into the two elements //V — E and HVE, since //V — E is equi-
valent to the (truth-functional) conditional E H. It is this
conditional which represents the inferential move from the
evidence E to the hypothesis / / t h a t goes beyond the evidence.
So it ought not to be tfoconfirmed by E.
Thus, considerable difficulties confront the positive relevance
analysis of ampliative induction. It certainly offers a somewhat
2 0 K . R . Popper and D. Miller, 'A Proof of the Impossibility of Inductive Proba-
bility', Nature, 302 (1983), 687-8, and ibid. 310 (1984), 434. But see A.R. Rodriguez,
'On Popper-Miller's proof of the Impossibility of Inductive Probability', Erkenntnis, 27
(1987), 353-7, and E. Eells, 'On the Alleged Impossibility of Inductive Probability',
Brit. J. Phil. Sci., 39(1988), 111-16.
142 The Pascalian Gradation of§17
more structured account of the situation than does the simple
probability analysis. So, if ampliative induction is to be analysed
in terms of a personalist concept of probability, the positive rel-
evance analysis is potentially more illuminating than the simple
probability analysis. But unfortunately the positive relevance
analysis also generates undesirable paradoxes.
Is it possible to construct some less paradoxical analysis of
ampliative induction in terms of a personalist conception of
probability? Such an analysis would have to avoid assessing
degree of inductive support by a pure posterior probability or by
the difference between a posterior probability and a prior one. It
would need to exploit some more complex function of the
relevant probabilities. But there is at least one task that no such
function could accomplish. It could not represent the attitude
normal among scientists towards important generalizations, or
systems of theory, that have been confronted by minor
anomalies.
As an example of this attitude, consider how Newton's theory
of gravitation was thought at first, even by Newton himself, to
give a markedly incorrect value for the forward movement of the
apse of the moon. It was not till 1752 that Clairaut showed how
the theory could be made to produce results that agreed with the
observed movements. Yet no serious thinker in the meantime
either qualified or rejected Newton's theory because of its
apparent failure to accord with known facts. Similarly no 21

serious thinker rejected Newton's theory because it failed to


explain the movement of the perihelion of Mercury, although
this movement was in the end explained only by Einstein's
general relativity theory. So the history of science makes it clear
that a theory may remain widely acceptable even when there is
evidence that falsifies it, just so long as the theory has sufficient
explanatory power elsewhere. It follows that, so far as a theory's
level of inductive support is to constitute the overriding ground
for its acceptability, a theory must be capable of being attributed

2 1 See the remarks by Florian Cajori in the appendix to his edition of Newton's Prin-
cipia Malhematica (Berkeley: University of California Press, 1962), 648 ff. Other
examples are given by R.G. Swinburne, 'Falsifiability of Scientific Theories', Mind, 73
(1964), 434 ff., and I. Lakatos, 'Falsificationism and the Methodology of Scientific
Research Programmes', in I. Lakatos and A. E. Musgrave (eds.). Criticism and the Growth
of Knowlege (Cambridge: Cambridge University Press, 1970), 138 ff.
§22Ampliative Induction 143
a substantial level of inductive support on the basis of evidence
that includes minor anomalies, just so long as the favourable
evidence is sufficiently strong. Schematically, therefore, we can
say that the existence of an inconsistency between E and H
should be compatible with E's giving a higher level of inductive
support to H than to some comparable theory with which E is
also inconsistent.
It is evident, however, that any such compatibility is ruled out
not only if E's degree of inductive support for H is to be a simple
probability but also if it is to be any function of Pascalian
probabilities involving E and/or H. Because Pascalian probab-
ility functions are functions of the elements of a Boolean algebra,
it can easily be shown —as a consequence of Bayes's theorem
22

—that, in assessing how much inductive support E gives to H,


the only probabilities that we could need to take into account are
those that are functions of p(E\H), P(//), and p(£). And what-
ever the function we construct in terms of those probabilities we
shall be unable to distinguish between the level of inductive sup-
port that E gives H, and the level that E gives H ' , in any case in
which H and H' have the same prior probabilities and E contains
anomalies for both / / a n d H ' but much stronger positive support
for H than for H'. That will be because in such a case the values
of p(E\H) and p(E\H') will both be 0 and therefore equal to one
another, p(//) and p(//') will also be equal to one another, and
there will be just one statement of evidence, E. Accordingly, on
a Pascalian analysis of inductive support, if we want to regard
some counter-instances to a particular hypothesis as unimport-
ant anomalies that do not serve to give the hypothesis zero value,
we have no alternative but either to revise this hypothesis so that
it is no longer falsified or to claim that it is not intended to apply
the domain of entities within which the counter- instances are
found.
The overall position then is not that ampliative induction
altogether resists gradation in terms of Pascalian probability,
but that each method of securing such a gradation encounters
problems that restrict the range of its validity or is forced into
trade-offs that restrict its analytical value. Some methods are
suited only to enumerative induction, some sacrifice the
L. Jonathan Cohen, The Probable and the Provable (Oxford: Clarendon Press, 1977),
2 2

189-90.
144 The Pascalian Gradation of§17
connection between support for a generalization and support for
particular predictions that derive from it, some cannot represent
the existence of inductive support for generalizations over
indeterminately or infinitely large domains, some rely on calcula-
tions from the structure of language where calculations from the
structure of reality would be more appropriate, some stand or
fall with the truth of some questionable metaphysical principle,
some resist objections by sacrificing informativeness, and some
can only be sustained if one is prepared to turn a blind eye to the
paradoxes that they entail. It is obviously worth considering,
therefore, whether any adequate analysis can be achieved in
non-Pascalian terms. The next chapter will be concerned to
explore the possibility of such an analysis and to evaluate what
gains and losses result from adopting it. 23

2 3 For personalist analyses of induction see also J . Dorling, 'A Personalist's Analysis
of Statistical Hypotheses and Some Other Rejoinders to Giere's Anti-Positivist Meta-
physics', in L.J. Cohen and M. Hesse (eds.), Applications of Inductive Logic (Oxford:
Clarendon Press, 1980), 271-81; P. Horwich, Probability and Evidence (Cambridge:
Cambridge University Press, 1982); and R . D . Rosenkrantz, Inference, Method and
Decision: Towards a Bayesian Philosophy of Science (Dordrecht.: Reidel, 1977). Foracritique
of the personalist approach to induction see C. Glymour, Theory and Evidence (Princeton:
Princeton University Press, 1980), 63-93. Further arguments against any Pascalian
analysis of induction are to be found in R. Harre, The Principles of Scientific Thinking
(London: Macmillan, 1970), 157-77 and K . R . Popper, The Logic of Scientific Discovery
(London: Hutchinson, 1959), 251-81
V
The Baconian Gradation of Ampliative
Induction

§ 20. I N D U C T I V E S U P P O R T BY T H E M E T H O D O F R E L E V A N T
VARIABLES

In order to grade inductive support on the basis of evidential


variety, we need to set up a list of relevant variables. By
manipulating cumulatively increasing combinations of these
variables we can, ideally, devise a series of canonical tests, such
that the more complex the text a hypothesis has passed, the
higher its inductive support. This Method of Relevant Variables
elucidates 'consilience' and other ideas in the Baconian tradi-
tion. But it is always logically possible for newly discovered data
to refute any assumptions about the list of relevant variables by
reference to which we assess how much certain experimental
results support a given hypothesis.
Let us look more closely into the structure and implications of
respect for evidential variety in inductive reasoning.
It has already ( § § 17-18) become apparent that instantial
variety is valueless unless it is potentially relevant to the kind of
hypothesis under consideration. For example, von Frisch's
hypothesis that bees discriminate between blue and other
colours received more support from bees returning to a blue-
coloured source of food that was moved around several different
locations than it would have received from the same number of
bees returning to a blue-coloured source of food that remained
in the same place on an equal number of different occasions. 1

That was because relative location was known to be a potentially


relevant factor in studies of bees' recognition-capacities: bees
1 See K. von Frisch, Bees, their Vision, Chemical Sense and Language (Ithaca: Cornell
University Press, 1950). This experiment is analysed in greater detail in
L. Jonathan Cohen, The Probable and the Provable; (Oxford: Clarendon Press, 1977)
130-3.
146 The Baconian Gradation of
have a good memory for places. And in the experiments on
colour-recognition variation of the food's position served to
eliminate the not wholly implausible hypothesis that memory of
place, rather than of colour, was operative. So we have here a
clear example of a case in which instantial variety is evidentially
superior to instantial multiplicity. But take away relevance from
the variety and the situation would be quite different. Suppose,
for example, the blue-coloured source of food had been placed
the same number of times in the same position but on different
days of the week. That kind of variety in the evidence would
have had no greater value than mere instantial multiplicity,
since there has never been any reason to suppose that bees are
influenced in their behaviour by the day of the week.
What is potentially relevant for hypotheses of one category,
however, may be quite irrelevant for hypotheses of another and
vice versa. Assessments of potential relevance can at best have
only a local, not global, validity. Day of the week may be
relevant to how certain kinds of people—especially religious
people—behave under certain conditions, though quite irrel-
evant to how bees do. It follows that, where we conceive degree
of inductive support to depend on evidential variety, we pre-
suppose some division of hypotheses into distinct categories of
enquiry that are defined by reference to their terminology and
logical structure. This division may be adjusted from time to
time as different lines of enquiry begin to diverge or coalesce.
But cross-category comparisons of evidential support are inher-
ently difficult. A hypothesis that has a great deal of variative
support within its own category is obviously superior to one that
has no support at all, or only a slight amount, in another cat-
egory. But exact comparisons at intermediate levels are scarcely
possible, because they would depend on the possibility of exact
comparisons between the degree of potential relevance attrib-
utable to a certain characteristic or circumstance in one category
of enquiry and the degree attributable to some other charac-
teristic or circumstance in the other category. Indeed, compar-
isons of potential relevance are difficult enough even within a
particular category. How can one tell whether a hypothesis
about bees' response to colour should be judged to be better
supported if it is apparently instantiated under change of loca-
tion for a distinctively coloured source of food than if it is
§22Ampliative Induction 147
apparently instantiated under change of shape for this? And
does greater inductive support come from instantiation under
two changes of location—one vertical, say, and the other
horizontal—than from instantiation under one change of loca-
tion and one change of shape?
It is easy enough to see how a characteristic or circumstance
qualifies for potential relevance within a given category of hypo-
theses: at least one such hypothesis must be falsified by the
presence of this feature. Of course, in practice, the belief that it
has been so falsified is itself open to falsification. The original
result may not be replicated and another factor may turn out to
have been operative instead. But, if we are entitled to presume
the efficacy of the original feature in causing the falsification, we
are entitled to treat that feature as potentially relevant. What is
rather more speculative is to assume a measure of relevance, so as
to permit comparisons between the inductive values of different
kinds of instantial variety. One might go a little way towards
achieving such a measure at any one time for a particular feature
by counting the number of importantly distinct hypotheses,
within the given category, that are currently believed to be fals-
ified by the presence of the feature. But appropriate criteria of
distinctness may be hard to establish here, and distinctness may
itself be a matter of degree. The hypothesis that bees recognize
red seems closer to the hypothesis that they recognize blue than
to the hypothesis that they discriminate square shapes from
circular ones.
The best way to sidestep all these problems is to give up any
attempt to measure inductive support from evidential variety
and to aim at ranking or comparing it. This aim can be
achieved, within a particular category of hypotheses, by suppos-
ing a single idealized sequence of cumulatively more and more
thorough tests, any one—or more—of which each hypothesis
may in principle undergo. Any hypothesis may then be ranked
in accordance with the thoroughness of the test that it is known
to have passed, and any two hypotheses may be compared in
terms of their rankings.
No doubt there are various ways of constructing such a
sequence of tests. But the way that best reflects normal patterns
of reasoning in experimental science requires us to introduce the
concept of a relevant variable. A relevant variable, within a
148 The Baconian Gradation of
particular category of hypotheses, is a set of one or more features
that are potentially relevant within that category and which are
(if more than one) mutually inconsistent with one another. An
example, for hypotheses about insects' recognition capacities,
might be features of location-change, like being vertical or
horizontal, or features of shape-change, like being rounded off
or squared off, in the circumstances of the experiment. In each
case these features are different from some non-relevant circum-
stances or characteristic, such as remaining in the same place, or
retaining the same shape, respectively. We can think of these
non-relevant circumstances or characteristics as being the relat-
ively norma] ones, divergences from which are introduced by
the manipulation of relevant variables. If under some such
manipulation a counter-instance to the hypothesis occurs, this is
presumably caused by the presence of that relevant feature just
in so far as no counter-instance occurred in the corresponding
relatively normal situation. The relatively normal situation is
commonly said then to function as a 'control'.
In practice, within a particular field of enquiry, experimental
scientists generally have a rough scale of importance for relevant
variables, depending partly on the accepted falsificatory efficacy
of their variants and partly on convention. When time or money
is in short supply, only the most important relevant variables are
manipulated in a test. But, given an ordering for the list of
relevant variables, it is possible to conceive of a canonical
sequence of more and more thorough tests. Suppose that our
hypothesis is of the form 'Anything that is A is B'. Then a
favourable result for canonical test t in an ideal case, would be
u

constituted by an instance of A that is also B, in circumstances


that are normal in relation to all relevant variables: the bees'
food-source would be unmoved, and its shape unchanged, when
their ability to recognize it is tested. What has been eliminated
thus is the possibility that normal circumstances suffice to falsify
the hypothesis that is being tested. The evidence furnished by
the test-result gives the hypothesis lst-grade support, whereas
an unfavourable test-result here, or evidence that is not the out-
come of any canonical test, gives Oth grade support. A favour-
able result for test t , in an ideal case, would be constituted by a
2

favourable result for test t , together with instances of A that are


x

also B under each variant of the first relevant variable (for


§22Ampliative Induction 149
example, change of location) in circumstances that are normal
in relation to all the other relevant variables (such as shape of the
food-source). What has been eliminated thus is the possibility
that, in normal circumstances, manipulation of the first relevant
variable suffices to falsify the hypothesis that is being tested: the
bees' recognition-capacity is not defeated by vertical or horizon-
tal shifts in the location of their food-source, and the evidence
now gives the hypothesis 2nd-grade support. A favourable result
for test t , in an ideal case, would then be constituted by a favour-
3

able result for test t together with instances of A that are also B
u

under each possible combination of a variant of the first or


second relevant variables with zero or more variants of the
second or first relevant variable, respectively, in circumstances
that are normal in relation to all the other relevant variables.
That is, alterations of location are now tried out in the various
possible combinations with alterations of shape. This test would
have an analogous eliminative power giving 3rd-grade support,
and subsequently numbered tests would be correspondingly
more and more thorough in virtue of their cumulative com-
plexity, with favourable results providing higher and higher
levels of inductive support for the hypothesis in question in
virtue of their superior eliminative power.
A hypothesis that fails a particular test would also fail any
more complex test, because of the cumulative complexity built
into the canonical sequence of tests: where^' > i > 0, each test ^
always includes each test t . But the converse would not hold. A
{

hypothesis might fail a more thorough test, t but pass a cruder


p

one, t , so that the evidence's support for the hypothesis is at least


{

Jth-grade but less thanjth-grade. Consequently it is possible for


a hypothesis to have greater-than-zero inductive support even
on the basis of evidence that includes counter-instances. The
Method of Relevant Variables, as we may call the system of
assessment that has just been described, is therefore not hit by
the problem about anomalies that confronts a Pascalian analysis
(p. 142), as when Newton's theory of universal gravitation was
agreed to have strong evidential support even though it was
known not to explain the movement of the perihelion of
Mercury.
Where a hypothesis has instead the form 'Anything that is/I is
thereby caused to be B', a similar sequence of tests would apply,
150 The Baconian Gradation of
except that in each trial a successful result would require not just
an instance of A that is also B but, in addition, as what is called a
'control', an instance of not-^4 (some normal situation upon
which A has not intervened) that is also not-fi. And correspond-
ingly appropriate adjustments serve to adapt this method of
comparing or ranking inductive support to the task of evaluating
scientific generalizations that correlate different qualitative or
quantitative variables with one another or to the task of evaluat-
ing more comprehensive theories. Thus Newton found the
2

connection between spectral colours and different degrees of


refrangibility of light rays to be unaffected by the position of the
glass, by unevenness of the glass's thickness, by the size of the
aperture through which the sun's rays enter, by the angle of
these rays' incidence, and so on. 3

But it is obviously not an essential feature of a canonical test


on a hypothesis, at any level of complexity or thoroughness, that
the circumstances of the test have to be artificially created. If the
hypothesis can be observed in operation over an appropriate
variety of naturally occurring circumstances, there is no need for
experimental contrivance. What shows the grade of inductive
support that exists for the hypothesis is the failure of the hypo-
thesis to be falsified in such-or-such circumstances, irrespective
of how those circumstances came into being. The Method
of Relevant Variables is in principle applicable in non-
experimental sciences (astronomy, ethology, etc.), as well as in
experimental ones, even though there may be many practical
problems about finding or constructing just the right combina-
tions of circumstances for a particular type of test. Equally it
does not matter to the Method of Relevant Variables whether
facts constituting a canonical test-result are observed before or
after the appropriate hypothesis has been thought up: inductive
support is a timeless, quasi-logical relation (see pp. 11-13). Nor
does it matter if the variations supporting one theory are
described in the language of another. The assessment of induct-
ive support is not tied to any dichotomy between theoretical and
observational terminology.

2Details are given in Cohen, The Probable and the Provable, pp. 144-57.
3See I.B. Cohen (ed.), Isaac Newton's Papers and Letters on Natural Philosophy
(Cambridge: Cambridge University Press, 1958), 47-52.
§22Ampliative Induction 151
One point of special interest arises in regard to the application
of the Method of Relevant Variables to inductive reasoning
about wide-ranging explanatory theories. If the explanatory
range of one such theory not only embraces, but also extends
beyond, the range of another, it is normally considered induct-
ively superior to the latter. But in assenting to this intuitively
cogent principle we must exclude those cases in which superior
explanatory power is achieved trivially, as happens when two
mutually independent explanatory theories are tacked into one
by conjunction. So, as a sign that no such trivial extension lies at
the root of a particular theory's superior explanatory power, we
normally welcome the ability of a theory to lead us to some new
kind of knowledge. That is, we look to the possibility of deriving
novel kinds of prediction that are then experimentally
confirmed, as Bacon, Leibniz, Lakatos, and many others have
emphasized. Alternatively, it may be that two surprisingly
4

separate sets of phenomena have each independently led us to


propose the same theory as an explanation for them. As already
remarked (in § 3) Whewell pointed out the importance of such a
'consilience' of inductions, and cited as an example the way in
which Newton's theory of universal gravitation 'which had been
inferred from the Perturbations of the moon and planets by the sun
and by each other, also accounted for the fact, apparently
altogether dissimilar and remote, of the Precession of the
equinoxes' .1 5

How does the Method of Relevant Variables represent these


facts? First, the different kinds of phenomena that a relatively
comprehensive theory ought to explain, in a particular area of
enquiry, may be regarded as the relevant variables for that
theory, even if no variants of these different variables are in
practice ever combinable. So the greater, and therefore the
6

more surprising, the variety of phenomena that a theory


4 F. Bacon, Novum Organum, bk. I, aphorisms ciii, cvi, cxvii, in The Works of Francis
Bacon, ed. J. Spedding, R. Ellis and D.N. Heath (London: Longmans, 1859), i. 204,
206, and 212; G.W. Leibniz, Die philosophischen Schriften von G. W. Leibniz, ed. C.J.
Gerhardt (1865), i. 196; I. Lakatos, 'Falsificationism and the Methodology of Scientific
Research Programmes', in I. Lakatos and A. E. Musgrave (eds.), Criticism and the Growth
of Knowledge (Cambridge: Cambridge University Press, 1970), 124.
5 W. Whewell, The Philosophy of The Inductive Sciences, 2nd edn. (London:J.W. Parker.,
1847), ii. 66 (Whewell's italics).
6 See Cohen The Probable and the Provable, pp. 152-7.
152 The Baconian Gradation of
explains, the more thorough the test that it passes and therefore
the greater its inductive support from the evidence. Secondly,
trivial increases in support are in any case barred by the
conjunction principle that is shown below (p. 158) to follow
from the Method of Relevant Variables: no conjunction of two
propositions can have higher inductive support on given
evidence than the less-well-supported conjunct. Thirdly, a
theory that generates novel and successful kinds of prediction is
in effect expanding our knowledge of relevant variables and is
correspondingly more valuable, other things being equal, than a
theory that just satisfies our existing expectations.
There is an interesting difference here between the implica-
tions of a Pascalian mode of assessment for inductive support
and the implications of the Method of Relevant Variables. Any
Pascalian mode of assessment implies, as we have seen
(pp. 136-7), that the greater the improbability of the evidence,
the greater the probability of the hypothesis upon that evidence
(other things being equal). And, of course, a high Pascalian
improbability that a statement is true is equivalent to a high
probability that it is false. So, on a Pascalian analysis of induct-
ive reasoning, a theory gains support from the sheer novelty of a
piece of evidence in its favour only if there is a prior reason or
tendency to regard the statement of that evidence as false. But
according to the Method of Relevant Variables what is required
in such cases is not necessarily the presence of a prior reason or
tendency to regard that statement as false: it suffices to note the
absence of a prior reason or tendency to regard the statement as
true. Thus the Method of Relevant Variables admits a wider
spread of cases in which the novelty consideration operates.
Indeed, while the Method of Relevant Variables treats the
novelty of awareness of the fact as sufficient to enhance its
confirmatory value, the Pascalian analysis requires that fact also
to be a paradoxical one. But it is doubtful whether the actual
history of science reflects so strong a requirement in the matter,
nor are there any obvious reasons why it ought to do so.
This point also affects the analysis of what Whewell called
'consilience' (see above). As was remarked in an earlier section
(p. 137), a hypothesis that predicts or explains two mutually
independent items of evidence will, other things being equal",
have a higher Pascalian probability on the combined evidence
§22Ampliative Induction 153
than on either evidential item alone. So a Pascalian analysis does
appear to represent the value of consilience. But it achieves this
only at the cost of treating the pertinent evidential items as
paradoxical rather than just novel.
The Method of Relevant Variables is an idealization or
rational reconstruction, not a description, of what actually goes
on. It generalizes and systematizes the intuitions underlying
Bacon's tables of presence and absence and Mill's Methods of
Agreement and Difference ( § 4 ) . The manipulation of a
relevant variable, so as to diverge from a relatively normal
situation, echoes Mill's Method of Difference, as does the treat-
ment of causal generalizations. The canonical proliferation of
trials, in each of which a different combination of variants of
relevant variables is present, echoes Mill's Method of Agree-
ment. But, in order to give explicit recognition to the various
ways in which Mill's statement of his Methods represents a
somewhat crude and oversimplified system of gradation for
inductive reasoning from instantial variety, it is necessary to
show what a system of gradation for this would ideally be like.
Actual scientific reasoning no doubt falls somewhere inter-
mediate between what Mill's Methods license and what the
Method of Relevant Variables does. And for the most part we
can live with such an intermediate position, given the
constraints on research that are imposed by shortage of available
resources, by social pressures for useful results, and so on.
Nevertheless it is useful here to set up cognitive ideals as canon-
ical standards, even if only to be able to tell, when reviewing any
working judgement of inductive support, how far intellectual
values have been subordinated to practical ones. Moreover 7

arguments based on the Method of Relevant Variables are also


available for use by scientists in retrospective assessments of
what they or their colleagues are justified in claiming to have
achieved. Nor is the Method of Relevant Variables the only
philosophical idealization that has been proposed in this area.
As Ramsey himself was prepared to concede, for example, the
personalist theory of probability rests on an underlying idealiza-
tion about human decision-making (p. 59). Idealization is often
7 A fuller account of the Method of Relevant Variables is given in Cohen, The Probable
and the Provable, pp. 135-44.
154 The Baconian Gradation of
as useful for theory construction in philosophy as it is in natural
science.
The Method of Relevant Variables also echoes Keynes's prin-
ciple of limited independent variety (pp. 133-6), in so far as the
assumed list of relevant variables at any one time is a finite one
and each variant of any one variable is logically independent of
any variant of another variable. The Method of Relevant
Variables licenses the construction of a plurality of higher-level
empirically corrigible theories that state which variables are
relevant, in what order of importance, to which categories of
ordinary-level hypotheses.
In the progress of human knowledge we always enter into
some pre-existing state of the art: if a solitary infant survives to a
solitary maturity he will not contribute to the development of
science. So there is always some empirical information already
available to suggest how we might begin to formulate an ordered
list of relevant variables for a given category of hypotheses. And,
while we use the list as an (often highly tentative) basis for our
inductive evaluations, experience will show, from time to time,
how the list should be corrected, extended, or pruned, or how
our hypotheses should be recategorized. For example, when
pre-clinical tests did not disclose the toxicity of thalidomide, the
tragic discovery of its teratogenic properties revealed later that
the pregnancy issue should be included in the list of relevant
variables for testing hypotheses about the non-toxicity of such
drugs. That is, the results of the tests were "not replicated in
actual practice, with the result that the operation of a hidden
variable was diagnosed: some critical factor was absent in the
tests and present on some real-life occasions. But at any one time
we can test a hypothesis only up to the limit of what we then
believe we know about relevant variables. Accordingly, if a
hypothesis survives even the most thorough canonical test that
we then think to be appropriate, it is reasonable for us then to
believe that we know it to be true. Correspondingly scientists
often claim to know such-and-such a law or to have proved such-
or-such a hypothesis. But others may later nevertheless discover
8

8 For example, G. Thomson, 'Matter and Radiation', in R. Harre (ed.), Scientific


Thought 1900-1960: A Selective Survey (Oxford: Clarendon Press, 1969), 52 ('proof of the
nature of radioactivity), and N . W . Pirie, 'The Viruses', ibid. 237 ('knowledge' of
effects of virus infection). Compare A. Einstein and L. Infeld, The Evolution of Physics
(Cambridge: Cambridge University Press, 1938), 156 ('proof of the existence of elec-
tromagnetic waves).
§22Ampliative Induction 155
that hidden and uncontrolled variables were at work in produc-
ing the results of earlier tests and that their predecessors were
wrong to believe that they knew the truth of their hypotheses.
Equally, someone may discover later that his predecessors were
wrong to believe that they did not have fully adequate evidence
for the truth of a particular hypothesis (which they nevertheless
believed to be true) just because at the time they thought their
list of relevant variables was incomplete: they knew the truth of
the hypothesis but did not know that they knew it. And it follows
from this latter point that, in regard to knowledge that is based
on ampliative induction in accordance with the Method of
Relevant Variables, we cannot accept the so called KK thesis, 9

whereby a person's knowing thatp entails that he knows that he


knows that p.
This incompatibility with the KK thesis is due ultimately to
the fact that any judgement of inductive support—that is, any
judgement that a particular hypothesis has such-or-such a level
of support on given evidence—must, if it is made in accordance
with the Method of Relevant Variables, be itself an empirically
corrigible judgement, just because, as we have seen, our list of
relevant variables for that category of hypotheses is always open
to revision on the basis of new experiences. Any claim to induct-
ive knowledge is then indefinitely defeasible. On the other hand,
where evaluations of inductive support on given evidence are
demonstrable a priori, as on Carnap's theory (§ 17), it will
follow that, if the evidential support for a hypothesis were so
strong that its truth could be said to be known, then that too
could be said to be known. Where judgements of inductive
support are a priori, the KK thesis applies.
Nor is it only the principle of limited independent variety that
fragments, in the Method of Relevant Variables, into a plurality
of empirically corrigible theses. The same holds for the principle
that Mill called 'the law of causation' (§ 4) and Keynes 'the
Uniformity of Nature' (§ 18). To assume that only such-and-
such characteristics or circumstances are potentially relevant
within a particular category of generalizations is to assume that
only the presence of one or more of those features ever falsifies a
generalization which holds good in at least one instance when
none of those features are present. Moreover, a feature that is
9 See e.g. J. Hintikka, Knowledge and Belief (Ithaca: Cornell University Press, 1962),
28-9, 103-25.
156 The Baconian Gradation of
assumed to be potentially relevant because it falsifies a
generalization on one occasion must also be supposed to falsify it
on every other occasion that is similar in relevant respects.
Otherwise there would have to be some additional reason why it
falsified the generalization on that one occasion. And, if a
generalization holds good on at least one occasion despite the
presence of this or that combination of relevant variants, it must
also be supposed to hold good on any other occasion when the
only potentially relevant variants present are the same ones.
Otherwise one is not assuming that only the presence of one or
more features out of the appropriate list of potentially relevant
variants is capable of falsifying the generalization. In short, the
thesis that certain variables are all and only those that are poten-
tially relevant within a given category of generalizations implies
that (for any such generalization) the results of any test
constructed in the canonical way, on the basis of those variables,
should always be replicable. Failure to replicate a test-result is
always a ground for supposing that there is some mistake in
current assumptions about what variables are potentially
relevant or about how the various different categories of
generalizations are constituted. So it is not a global and
apparently unfalsifiable law of causation that is involved in
every attempt to replicate a test-result in a particular field of
enquiry, as Keynes (though not Mill—see pp. 31-2) seems to
have supposed, but rather an empirically corrigible thesis which
is only about the kinds of uniformities that exist within that field
of enquiry. Such a localized thesis has itself a certain degree of
10

higher-level inductive support, which at any one time we may,


or may not, judge accurately from the evidence of previous
research."

1 0 Localization is a strategy that has been pursued by quite a number of writers in the
philosophy of induction: see the papers and bibliography in R.J. Bogdan (ed.), Local
Induction (Dordrecht: Reidel, 1976).
11 Some discussions of the Method of Relevant Variables (by I. Levi, A. W. Burks, S.
Blackburn, R. Hilpinen, A. Margalit, J . L . Mackie, a n d M . Hesse, with replies by L.J.
Cohen) are to be found in L.J. Cohen and M. Hesse (eds.), Applications of Inductive Logic
(Oxford: Clarendon Press, 1980), 26-7, 64-7, 172-201, 207-14, 245-50. See also I.
Levi, 'Support and Surprise: L.J. Cohen's view of Inductive Probability', British Journal
for the Philosophy of Science, 30 (1979), 279-92.
§22Ampliative Induction 157
§ 21. T H E L O G I C A L S Y N T A X O F T H E M E T H O D O F R E L E V A N T
VARIABLES

Baconian gradations of inductive support have a logical syntax


(for conjunction, etc.) that cannot be mapped on to the Pascalian
calculus. This system can instead be systematized within a
generalized modal logic. The Carnapian and Baconian systems
of inductive reasoning can then be compared with one another
on a common basis.
In § 20 we sketched how inductive support may be graded by
the Method of Relevant Variables. This method of gradation
incorporates and develops certain key ideas of those philo-
sophers (in particular Bacon, Mill, and Keynes) who thought
that evidential variety, rather than instantial multiplicity,
deserves paramount respect in ampliative induction. But Mill
and Keynes thought also (as we saw earlier in § 4 and § 19) that
the gradation of inductive support from evidential variety has a
logical syntax that is Pascalian. On their view, any function that
measures one proposition's support for another on this basis
must conform to the familiar principles of the classical calculus
of chance: the complementational law for negation, the
multiplicative law for conjunction, and so on. So the question
that arises now is whether the Method of Relevant Variables
generates evaluations of inductive support that are also subject
to these familiar principles. Are those evaluations related to one
another by a Pascalian logical syntax? And it will be convenient
henceforth to refer to those evaluations as 'Baconian', because
they are a systematic development of Bacon's seminal ideas, just
as the modern mathematics of probability is a systematic
development of Pascal's seminal ideas.
Of course, as we saw in § 20, the Method of Relevant
Variables supplies at best a locally valid method for ranking, not
measuring, inductive support. If the number of relevant
variables is finite, that method maps pairs of propositions (typ-
ically a generalization and a description of test-results) on to the
first n + 1 integers ^ 0, where n is the number of cumulatively
more and more complex tests that are constructable for the given
category of generalizations. But this restriction to a ranking
158 The Baconian Gradation of
function does not suffice to exclude the possibility of an
underlying, though unknown, measure-function. One can at
least conceive of a field of intellectual judgement in which only
rankings and comparisons are practically possible though what
are ranked are Pascalian probabilities of which the absolute
values are unknown. In such a field of study the ranking for a
conjunction of two hypotheses, for example, would always be
lower than that for either conjunct, given that neither ranking
for a conjunct was maximal or minimal, because of the under-
lying operation of the multiplicative principle for the Pascalian
probability of a conjunction. The product of two numbers m and
n, where 1 > m > 0 and 1 > n > 0, must always be less than
either m or n itself.
However, though the mere fact of restriction to ordinal values
does not exclude an underlying Pascalian syntax, the specific
pattern of rankings that is generated by the Method of Relevant
Variables does indeed suffice to exclude this. Consider what
happens in the case of conjunction, as in the case of von Hess's
hypothesis that both fishes and invertebrates are colour-blind. 12

Let E state evidence that a generalization H passes test t„ in the


appropriate series of more and more thorough canonical tests,
and also that another hypothesis H' passes test t where j is
p

greater than, or equal to, i. Then E must state evidence that the
conjunctive hypothesis H&H' passes at least test t„ since that
hypothesis must resist falsification under all possible combina-
tions of variants of relevant variables manipulated in test t In
r

other words, if the support that E gives H' by Baconian methods


is greater than, or equal to, that which it gives H, the support
that it gives the conjunction H&.H' is equal to, not less than, that
which it gives H. So the Pascalian, multiplicative principle for
conjunction is not operative here.
Perhaps someone will object that the proposed non-Pascalian
principle cannot be correct because it seems to allow the possib-
ility of a positive level of support for a proposition that conjoins
two conflicting hypotheses. E might give some support to H and
some to H', we shall be told, but if Hand H' conflict they cannot
possibly both be true and consequently their conjunction cannot
deserve to be attributed any positive support. If the Method of
Relevant Variables entails this attribution in certain cases, it
must obviously be rejected.
1 2 K. von Frisch, Bees, p. 4.
§22Ampliative Induction 159
But the Method of Relevant Variables entails no such thing.
Suppose the conflicting hypotheses are 'All A are B' and 'All A
are not-5', respectively. Then any test that the one passes will be
failed by the other, so that, by the Method of Relevant
Variables, it is impossible for both hypotheses to have a positive
level of support: at least one of the two hypotheses must fail test
. And, if therefore at least one of the two must have a zero level
of support, so too must the conjunction. That is, a proposition
that conjoins two conflicting hypotheses cannot have a positive
level of support from any evidence.
Someone may object: 'What about a case in which two hypo-
theses are not logically incompatible, but are highly unlikely, on
the basis of other evidence, to be both true?' Well, what this
other evidence then shows, if we rely on it, is the presence of a
serious flaw in any series of tests by which the conjunction of the
two hypotheses is reckoned to have greater than zero support.
Though the conjunction principle for the Method of Relevant
Variables is a priori, the appropriateness of a particular series of
tests is an empirical issue.
Moreover, this conjunction principle for the Method of
Relevant Variables not only excludes any underlying concep-
tion of inductive support as a Pascalian probability. It can also
be shown to exclude inductive support from being thought of as
some complex function of any of the related Pascalian probab-
ilities. Indeed, a similar conclusion results from the fact that
13

the Method of Relevant Variables allows the possibility of situa-


tions in which a hypothesis has some level of positive support
from a statement of evidence that also falsifies it. The statement
could report evidence to the effect that the hypothesis passes
a less thorough test and fails a more thorough one. Such a
hypothesis would cope adequately with the more highly relevant
variables, but fail to cope when some less important ones are
introduced into the tests. And we have already seen (pp. 142-3)
that no simple, or complex, Pascalian analysis of inductive
support can accommodate treatment of minor anomalies in this
way.
How then can we systematize the logical syntax of Baconian
support-gradings? Attention to that question is vital. For,
if a philosopher (like Mill, Keynes, or Carnap) succeeds in

1 3 The proof is given in Cohen, The Probable and the Provable, pp. 190-9.
160 The Baconian Gradation of
demonstrating that his analysis of inductive support-gradings
has a Pascalian structure, what the demonstration achieves is not
merely a clarification of that analysis but also—to some
extent—a confirmation of it. Inductive support, so analysed,
has been shown to instantiate a pattern of logical syntax that has
several familiar and established manifestations in other kinds of
gradation of certainty. A somewhat controversial thesis has been
strengthened, it may be hoped, by the acquisition of respectable
associations, much as Pascal and the Port-Royal Logic originally
( § 3 ) brought intuitive beliefs about certain non-aleatory issues
within the scope of the familiar mathematics of chance, in order
that they should be consolidated thereby. So, if the logical
syntax of inductive support-gradings generated by the Method
of Relevant Variables were wholly idiosyncratic, an analysis in
terms of that method would be at a corresponding disadvantage
in comparison with any Pascalian analysis.
As it turns out, however, though Baconian evaluations of
inductive support do not conform to Pascalian principles, they
do conform to the principles of a generalized modal logic, and
this modal-logical connection serves in turn to corroborate the
legitimacy of such an analysis. The key idea here is that a
generalization which holds good (or would hold good) under all
manipulations of relevant variables must state a law of nature or
a consequence of one, while generalizations that hold good (or
would hold good) only under less thorough tests have a status
analogous to, but falling suitably short from being, that of such a
statement. So a generalization of the latter kind may be said to
have a certain degree of legisimilitude, or similarity to law,
depending on the thoroughness of the test that it is thought
capable of passing. If the evidence is, in effect, that a generaliza-
tion has been instantiated in all the various combinations of
circumstances that constitute test t then background assump-
n

tions, conventions or beliefs about relevant variables constitute


laws that authorize us to infer from this evidence that the
generalization has z'th degree legisimilitude. And that is what it
means, according to the Method of Relevant Variables, to say
that the generalization has Jth grade inductive support on the
evidence. An inductive support-grading on this view is typically
a law-like conditional statement (supported by background
evidence) of which the antecedent is a statement of the current
§22Ampliative Induction 161
evidence and the consequent is a statement that a certain
generalization has a specified degree of legisimilitude.
It follows that there is an important difference here from the
position on a Pascalian analysis, like Keynes's or Carnap's. A
judgement of inductive support here has the structure of a law-
like conditional statement, not of a conditional probability, so
that the grading of the generalization itself is detachable.
Symbolically, where E is the statement of evidence and G the
generalization, what the judgement in effect normally asserts is
'It is in accordance with an accepted law that if E is true then G
has such-or-such a level of legisimilitude'. So, given E, we can
normally detach the specified level of legisimilitude for G,
whereas from the Pascalian probability p(G\E) = n, we cannot
given E, detach p(G) = n except in special circumstances (see
also § § 3 , 14, and 26).
In an ordinary modal logic there is either an operator to
express 'necessarily', in terms of which 'possibly' is defined as
'not necessarily not', or an operator to express 'possibly', in
terms of which 'necessarily' is defined as 'not possibly not'. If we
generalize on the former type of logic appropriately there will be
a sequence of such operators, which we may symbolize (in rela-
tion to a category of generalizations for which there are n
relevant variables and therefore a series of n canonical tests as 14

necy, nec , etc. up to and including an operator nec„ . The


2 +2

operator nec„ is to express logical necessity. Just below the


+2

latter operator is one that expresses the possession of at least


natural necessity—the quality of being a law of nature (capable
of withstanding all canonical tests) or a consequence of one. So
the top two operators are to imply truth. That is to say, 'If
nec„ A, then A' is to be a principle of the system, and so is 'If
+2

nec„ A, then A'. Modal operators numbered less than or equal


+2

to n are not to imply truth, and thus can be used to imply levels of
legisimilitude at which there is still a risk that a more thorough
test may reveal falsity. But a higher-numbered modal operator
will always imply a lower one: that is 'If nec/l, then nec/l, where
j > i' is a principle. So 'nec/l' expresses the fact that A has at
least «th degree legisimilitude, while 'not-(nec A)' expresses the
i +l

1 4A more complex system of numbering modal operators is needed where the


number of relevant variables is assumed to be denumerably infinite. Such a system is
described in Cohen, The Probable and the Provable, pp. 222-33.
162 The Baconian Gradation of
fact that A has at most 2th degree legisimilitude. Familiar modal
principles from the Lewis-Barcan calculus S4 may then be 15

generalized, so that we have, for example, 'If nec, (if A, then B),
then, if nec^A, then nec,B' as another principle.
This kind of generalized modal logic can be rigorously forma-
lized and a large number of theorems derived. Some of the 16

theorems are analogous to those derivable within the Pascalian


calculus. Thus a theorem is derivable which corresponds to the
Special Consequence Condition for inductive support (see
pp. 139-40). Such a theorem echoes the fact that any logical
consequence of what is unfalsified by certain potentially relevant
circumstances will certainly also fail to be falsified by them. But
other theorems are radically different from those derivable
within the Pascalian calculus. Thus we obtain 'If nec/l and nec.B,
then nec^A&B)' where 7 ^ i. This accurately reflects the way in
which the Method of Relevant Variables treats conjunction, as
we have already seen (§ 21). But it is radically different from the
multiplicative principle for conjunction that is integral to the
Pascalian calculus.
The 'possible worlds' metaphor, which has been used to
provide a systematic formal semantics for ordinary modal
logic, may also be used not only to do this for generalized
17

modal logic but also to elucidate the way in which generalized


modal logic provides a logical syntax that fits the Method of
Relevant Variables. Roughly, a logically possible world is said
18

to be physically possible in relation to a particular category of


generalizations if it contains nothing to falsify any such
generalization that holds good in the actual world. A physically
possible world JV is said to be a t, world for a particular category
of generalizations, if every physically possible combination of
characteristics and circumstances that can occur in test t, for that
category, is instantiated in W, and no other features are instan-
1 5 See C.I. Lewis a n d C . H . Langford, Symbolic Logic (New York: Dover, 1932), and
R . C . Barcan, 'A Functional Calculus of First Order Based on Strict Implication',
Journal of Symbolic Logic, 11 (1946), 1-16.
1 6 Cohen, The Probable and the Provable, pp. 229-40.
1 7 See G.E. Hughes and M . J . Cresswell, An Introduction to Modal Logic (London:
Methuen, 1968), 75 fT.
1 8 A fuller account is given in Cohen The Probable and the Provable, pp. 240-2, where it is
also shown why the appropriate system generalizes the Lewis-Barcan calculus S4, and
not S5.
§22Ampliative Induction 163
tiated in it. So the actual world, and physically possible worlds
that are as rich in instantiated features as it is, each contain a
plenitude of relevantly different events. Moreover, each t, world
is said to contain nothing that falsifies any generalization that
holds good in a t world, where7 ^ i. And a generalization may
s

be said to have at least ?th degree legisimilitude if it holds good in


any (and therefore in every) t, world. The execution of an experi-
mental test t, can thus be thought of as being ideally a simulation
of a minimal t• world, though in practice it is often very difficult
to ensure that no hidden variable is operative (no relevant
impurities in the materials, no extraneous forces affecting the
apparatus, no unregistered alterations of procedure between
trials, and so on).
We are also now in a position to epitomize the difference
between a Carnapian system of inductive reasoning and a
Baconian one. The former is based on measuring the logical
range of a proposition by reference to the sum of the values sev-
erally assigned to each logically possible world in which the
proposition holds good. The latter is based on ranking the legi-
similitude of a generalization by reference to the richest kind of
physically possible world—highest grade of relevantly eventful
world—in which the generalization always holds good. In the
former case we calculate a priori from a quantitative summation
of possible worlds: in the latter we base ourselves empirically On
a qualitatively determined ordering of them. The Carnapian
system, mapping pairs of propositions on to real numbers, can
give us a single, globally applicable, mathematically convenient
measure. But unfortunately it is a measure for which there is no
mode of empirical validation. The Baconian system, mapping
pairs of propositions on to ordinal numbers, can give us only a
plurality of locally applicable and mathematically unsophist-
icated criteria. But the moulding or remoulding of these criteria
is inherently sensitive to the changing state of empirical know-
ledge about their subject-matter.' 9

1 9That is why the Method of Relevant Variables escapes the anti-inductivist threat of
historically oriented philosophers of science like Hanson, Toulmin, etc.: see L.Jonathan
Cohen, The Dialogue of Reason (Oxford-. Clarendon Press, 1986), 118-28. A more extens-
ive overview of the situation is afforded by J . E . Adler, 'Criteria for Good Inductive
Logic', in L.J. Cohen and M. Hesse (eds.), Applications of Inductive Logic, pp. 379-405.
164 The Baconian Gradation of
5 22. S O M E N O N - S T A N D A R D I N T E R P R E T A T I O N S O F B A C O N I A N
LOGICAL SYNTAX

The level of a generalization's inductive support from available


evidence may be affected by introducing appropriate modifica-
tions into its antecedent or into the characterisation of its
domain. When the implications of this are explored it becomes
clear that Keynesian weight is closely interconnected with
Baconian legisimilitude. Light is also shed thus on the relation-
ship between scientific idealizations and the evidence that
supports them. And the same underlying structure is found in
other types of reasoning, such as in the jurisprudence of common
law.
The Pascalian calculus is open, as we have seen (§ § 5-13), to a
variety of different interpretations or applications for the grada-
tion of certainty. So too, it turns out, is the generalized modal
logic that articulates the syntax of Baconian judgements.
Indeed, because of these two facts, if a logical syntax that is
suggested for the gradation of certainty is not open to a variety of
fruitful uses for the purpose, it can hardly have a strong claim on
our interest or attention.
Consider first Keynes's concept of weight, which was dis-
cussed above in § 14. Roughly, the weight to be associated with
the Pascalian probability p(^41B) is the spread o{relevant evidence
that B affords in regard to A. More exactly, we can say that, for
any n, the weight of an ordered pair of characteristics or circum-
stances A and B, where p(A\B) = n, is to be understood as the
reliability with which p(-4j) = n may be inferred from Bs (where
As and Bs attribute A and B, respectively, to a particular entity
s). And this reliability is to be graded by the extent to which B
embraces all the characteristics or circumstances that are poten-
tially relevant to A. Smith's probability of survival to the age of
seventy is more reliably inferred from premisses about his
present age and his present state of health than from just a
premiss about his present age. Thus we can imagine B to be a
sequence of appropriately selected characteristics or circum-
For philosophies of induction that are neither Pascalian nor Baconian see S.F. Barker,
Induction and Hypothesis (Ithaca: Cornell University Press, 1957) and G. Schlesinger,
Confirmation and Confirmability (Oxford: Clarendon Press, 1974).
§22Ampliative Induction 165
stances C,, C . . . that are conjunctively attributed to j; and, as
2

the sequence gets longer, so the conditional 'If C s&.C s8i . . .,


t 2

then p(.4.j) = n' is less exposed to falsification. Moreover, this is


true whatever the particular entity s may be. So we are dealing
here implicitly with a generalized conditional proposition that is
supposed to become less and less exposed to falsification as its
antecedent gets more and more heavily qualified in relevant
respects.
A close connection with the Method of Relevant Variables
now begins to emerge. In the Method of Relevant Variables, we
standardly hold the content of a generalization constant and
grade its legisimilitude. But, if at some level of test-thorough-
ness, the hypothesis fails an appropriate test, an alternative,
non-standard procedure is available. Instead of grading the legi-
similitude of the generalization at less than maximal because of
this test-result, it is possible to qualify its antecedent in such a
way as to avoid the kind of falsification that produces the
unfavourable test-result. Whatever the variants of relevant
variables that are responsible for such falsifications, the hypo-
thesis can be modified in a way that will exclude it from purpor-
ting to cover situations in which those variants are present. But,
in order for the legisimilitude of a hypothesis to be thus increased
by the insertion of qualifications into its antecedent, those
qualifications have to be necessary for the avoidance of falsifica-
tion. That is, the hypothesis in its unmodified form needs to be
one that would be falsified by the presence of some charac-
teristics or circumstances that are excluded by the modification.
For example, if bees fail to re-identify a food-container where
the container is red, perhaps the hypothesis about their ability to
re-identify food-containers should be restricted to those food-
containers that are yellow, blue-green, blue, and ultra-violet. In
this way any desired level of legisimilitude can be obtained for a
hypothesis in the face of adverse evidence, though at the cost of
introducing a corresponding level of reduction in the scope of the
claim that it makes. And, when we increase legisimilitude by
this means in the case of generalizations that evaluate a monadic
probability in their consequent, we are implicitly increasing
what Keynes called the weight of the corresponding dyadic
probability. That is, when we increase the legisimilitude of a
generalization that has the form 'For any x, if C,x&C x& . . . 2
166 The Baconian Gradation of
SLC,X, then p(AX) = n', by qualifying its antecedent in further
favourably relevant ways, we thereby change the underlying
conditional probability into one that has greater weight. This
probability is now p(Ax\C x&C x& t . . . &C x), say, not
2 i+t

p(Ax\C 8CC X8L . . . &C,x). So, because of this correspondence in


1 2

value, gradings of weight must have, mutatis mutandis, the same 20

logical syntax as statements about legisimilitude. For example,


the conjunction principle, as we have already seen (§ § 14 and
21), is the same.
Admittedly the concept of relevance that is employed in
determining whether the presence of a certain condition or
circumstance increases weight is different from the concept of
relevance that is employed in determining whether it increases
legisimilitude. T h e former depends on an inequality between
Pascalian probabilities (§ 14), the latter on avoiding the
falsification of some generalization within the category (§ 20).
But in any area of enquiry in which it is assumed that, whether
or not they are ever actually discovered, uniformities rather than
probabilities lie at the roots of the data, the two concepts of rel-
evance must be co-extensive in regard to any particular grading
of weight or legisimilitude.
Moreover, our ability to increase the legisimilitude of a
generalization by qualifying its antecedent is not limited to
generalizations that have evaluations of Pascalian probabilities
in their consequents. W e can do this also, as shown above, for
ordinary generalizations that have the form 'For any x, if
C X&LC X&L . . . 8IC,x, then Ax'. Hence, so far as the reliability of
1 2

any inference authorized by such a generalization—an


inference from premisses C^s,C s, . . . C,s to conclusion As—is
2

graded by the legisimilitude of the generalization, this reliability


will increase with the n u m b e r of appropriately chosen
premisses. W e thus have a way of grading the credibility of a
singular proposition, on favourable evidence, which depends on
the extent of that evidence. T h e credibility of As given C,s&C s& 2

. . . fkCjS is equal to the legisimilitude of the generalization that


the conditional statement 'If . . . C s, then Ar' instan-
t

tiates. And such a way of grading credibility m a y conveniently


The Special Consequence Condition, in particular, requires some modification:
2 0

Cohen, The Probable and the Provable, 186-7


§22Ampliative Induction 167
be called 'Baconian probability', because it obviously contrasts
with the use of Pascalian probability for this purpose. The
Pascalian probability of a proposition on stated evidence grades
the extent to which that evidence favours rather than disfavours
the proposition. The Baconian probability grades instead the
extent of the relevant evidence that is stated, when that evidence
in any case favours the proposition rather than disfavours it.
Note how this contrast corresponds with the difference that
was remarked earlier (pp. 17-8, p. 57, p. 104) between the
intuitive concept of probability that is complementa-
tional—because it treats the credibility of a proposition as being
equal to the incredibility of its negation—and the intuitive
concept that is non-complementational—because under appro-
priate conditions it allows the credibility of a specified proposi-
tion and the credibility of its negation to be both rather low.
Correspondingly the Pascalian calculus affords us the theorem
p(^4|S = 1 — p(not-^41B),whereas the generalized modal logic,
within which the logical syntax of Baconian probability may be
developed, supplies instead the theorem 'If p(^4|fi) ^ i and
p(not-^4|£) ^ i, then p(not-fi) ^ i'. That is, the negation prin-
21

ciple for Baconian probability represents a kind of generaliza-


tion of the idea of proof by reductio ad absurdum. If on a stated
evidential premiss B both A and its negation have some degree of
credibility, and so the conjunction of A and its negation is
implied to have some degree of credibility, then a corresponding
degree of credibility must attach to the denial of that premiss. So
in the logic of Baconian probability a relative limit is imposed on
the extent to which a proposition and its negation may both be
equally credible on the evidence stated, since as the level of this
credibility mounts, so too must credibility of the proposition that
there is something false in the statement of evidence. There is
thus a relative limit on the height to which the Baconian
probabilities, on given evidence, of two contradictory proposi-
tions can rise, when these probabilities are equal to one another,
though there is no limit short of zero to which they can fall. The
ultimate reason why both can fall to zero is that they both stem
from the legisimilitude-levels of appropriate covering
generalizations, and it is quite possible for two conflicting
2 1 Ibid. pp. 217-44, esp. p. 223.
168 The Baconian Gradation of
generalizations, like 'Anything if j5is^4' 'Anything i f £ is not A',
both to have Oth grade legisimilitude, if neither has passed even
the simplest canonical test. But the principle of complementa-
tionality limits the depth to which the Pascalian probabilities of
two contradictory, and equiprobable, propositions can fall at the
same time as it limits the height to which those probabilities can
rise: specifically, they can neither both fall lower than 0.5 nor
both rise higher than that same point.
Similarly, Baconian probability (like legisimilitude) conforms
to a non-multiplicative principle for conjunction. We have
already remarked (pp. 19-20) on the suitability of such a prin-
ciple for judgements of credibility in certain contexts. And it is
arguably needed in any court of law that requires proofs of fact
to be established on the balance of probability, as for civil cases
in the Anglo-American system. The point is that two or more
22

independent issues, may be at stake in a case and then a multi-


plicative principle for conjunction generates paradox. For, even
if a plaintiff proved each of two such issues with a satisfactory
Pascalian probability of 0.6, say, he would have established only
a 0.36 Pascalian probability of being in the right overall,
whereas separate proofs of each issue are normally taken by the
courts to prove the case as a whole.
Again, we remarked earlier (pp. 20-1), on the need, in
certain forensic contexts, for judgements of probability to be
necessarily contraposable. We may want our logic or mathe-
matics to ensure an equality between the probability that a parti-
cular accused is guilty, given that such-and-such evidence is
before the court, and the improbability of that evidence's being
before the court, given that the accused is not guilty. And,
whereas Pascalian conditional probabilities are not necessarily
contraposable and so cannot capture this intuition, a dyadic
judgement of Baconian probability is necessarily contraposable
in the required way because it stems from the legisimilitude of its
covering generalization.
Moreover, when we develop the logical syntax of Baconian
probability we come across yet other features that, in certain
2 2 Cohen, The Probable and the Provable, pp. 58-67, 265-7. An analogous problem
arises about the standard of proof in criminal cases. For another point of view see Sir R.
Eggleston, Evidence, Proof and Probability, 2nd edn. (London: Weidenfeld and Nicolson,
1983), 34-49.
§22Ampliative Induction 169
contexts, may be preferable to the syntactic principles of
Pascalian probability. Thus it may sometimes seem paradoxical
that, though a very minute value for the monadic Pascalian
probability p(^4) is in principle compatible with a very high value
for the dyadic Pascalian probability of the same proposition
p(A\B), a zero value for the Pascalian p(^4) requires a zero value
for p[A\B). Cannot belief in a particular outcome sometimes
seem quite groundless on prior reflection, until evidence favou-
ring it begins to come in? We obviously need to distinguish here
the non-inferability of A on prior grounds from the inferability of
not-yl (see pp. 112-13). The Baconian p(4) = 0 expresses the
former, while the Pascalian p(A) = 0 expresses the latter.
Correspondingly, a zero value for the monadic Baconian
probability p(^4) is in principle compatible with any value what-
ever for the dyadic Baconian probability p(^4|5).
Another pattern of reasoning that has an important isomor-
phism with the Method of Relevant Variables is the establish-
ment of idealizations, as when Boyle's law is argued to be true of
perfect gases though not of actual ones, or the velocity of a falling
body is correlated with its period of fall on the assumption that it
is falling in a frictionless medium. Some philosophers suppose 23

that this pattern of reasoning is inherently different from induc-


tion, on the ground that it does not aim to register and syste-
matize observed phenomena but rather to establish models that
oversimplify reality. However, to look at the matter thus you
have to conceive induction as a rather crude and unstructured
procedure in which people build solely on what they have
passively observed. Once the importance of experimental inter-
ference with Nature is accepted and the Method of Relevant
Variables is properly understood, it is apparent instead that
idealization is a form of variative induction. It is a way of
ensuring maximal legisimilitude if the generalization in ques-
tion would be falsified not only where any variant of a particular
relevant variable is present but also in the normal situation (see
pp. 148-50 above) where no such variant is present. For
example, the standard equation for the velocity of a falling body
fails to hold not only in the special case of a stone falling in water
2 3For example, L. Nowak, 'Idealization and Rationalization: An Analysis of the
Anti-Naturalist Programme', Epistemologia, 2 (1979), 287-8.
170 The Baconian Gradation of
or treacle—on which many substances float—but also in the
normal case of a stone falling in the earth's atmosphere. So
idealization has some similarity to the procedure of qualifying a
generalization that underlies judgements of weight or of
Baconian probability. Like that procedure idealization aims at
the modification of a hypothesis, rather than its rejection, in the
face of falsifying factors. But in the case of weight or Baconian
probability the modification is achieved by complicating the
antecedent of the hypothesized generalization, whereas in the
case of idealization it is achieved by simplifying the domain to
-which the generalization applies. The particles of a perfect gas,
for example, have positions but no dimensions or interactions.
Perhaps someone will object that if an idealized generaliza-
tion does not have to be true of the actual world, but only of some
much more jejunely furnished non-actual world, there can be no
inductive grounds for choosing between one such idealization
and another. No observable fact, he may say, is predictable
from, or explainable by, any idealization, and therefore no
24

inductive confirmation or disconfirmation of it is possible. So


the Method of Relevant Variables tolerates any idealization
whatever, because no experience in the actual world could ever
falsify a generalization that purports to apply only under condi-
tions that do not occur in the actual world.
But the objection misses a crucial feature of the procedure
whereby the appropriate modification of a generalization can
raise its legisimilitude. The point is that the modification has to
be one that is necessary if falsification by the evidential facts is to
be avoided: if the modification is not made, those facts suffice to
falsify the generalization. This requirement operates, as we saw,
when such a modification takes the form of introducing an
appropriate qualification into the generalization's antecedent.
And it operates also when the modification changes instead the
generalization's domain of application. When we want such a
generalization to have high legisimilitude, we seek a generaliza-
tion that has no other modifications than those forced on it by the
evidential facts. So every difference between the values of a

2 4For the claim that idealized generalizations cannot generate explanations in accor-
dance with the covering law model see N. Cartwright, How the Laws of Physics Lie
(Oxford: Clarendon Press, 1983), 44-53.
§ 22 Ampliative Induction 171
particular physical variable that are predicted to hold in certain
circumstances under ideal conditions, and the values that are
observed to hold in those circumstances under conditions occur-
ring in the actual world, has to result from features of the actual
world that are excluded by the idealization: if it were not for the
idealization, the value predicted would be the one that is actually
found. Correspondingly, any idealization implying unexplain-
able differences from values in the actual world has to be rejected
or treated with reserve. Of course, the processes of approxima-
tion, by which observable evidence from the actual world is
matched to generalizations applying only to an ideally simplified
world, are in practice often delicate and difficult, allowing room
for considerations of theoretical elegance to enter into the
precisification of the ultimate solution. But in principle idealiza-
tions are to be graded inductively, by the Method of Relevant
Variables, in terms of those criteria of legisimilitude that are
appropriate to their category of subject-matter.
So far, in the present section, we have discussed various non-
standard ways in which the Method of Relevant Variables
impinges on the gradation of legisimilitude. The situations
envisaged were those where, in place of assuming a fixed
formulation for a hypothesis about some factual issue and
examining the conditions under which this or that grade of legi-
similitude would be assignable to it, we instead assume a desired
degree of legisimilitude for the hypothesis and examine the
conditions under which this or that reformulation would be
appropriate to it. Indeed, if we consider only the appropriate
types of modification in the antecedents of our hypothesis, the
underlying modal calculus can actually be interpreted as a
logical syntax for gradings of evidentially permissible simplifica-
tion. That is, if H is an unmodified hypothesis in a particular
category, then, for every support assessment stating that on the
evidence of E, H has zth grade support, there is an equivalent
statement of evidentially permissible simplification stating that,
on the evidence of E, the simplest fully supported version,
or versions, of H has, or have, zth grade simplicity, accord-
ing to appropriate criteria of simplicity. But in all this the
25

See L. Jonathan Cohen, The Implications of Induction (London: Methuen, 1970),


2 5

142-55, for the details of this.


172 The Baconian Gradation of
subject-matter of the hypotheses considered is standard enough:
namely, the properties of observable objects, causal relations
between observable events, correlations between physical para-
meters, and so on. What is non-standard is the mode of induct-
ive reasoning.
It is also possible, however, to reason inductively about non-
factual issues. For example, as we saw earlier ( § 2 ) Bacon
thought that when people extract general legal maxims from
judicial decisions in particular cases they use the same patterns
of reasoning as are used in experimental science. Jurists op-
posing an inductivist analysis of common law reasoning, like
Bacon's, have sometimes pointed out that a great deal of this
reasoning is concerned to establish the correct meanings of legal
rules rather than their truth or validity. They have argued
thence that the determination of vagueness is a function of legal
reasoning which sharply differentiates it from any kind of deduc-
tion or induction. But we have already seen how the appro-
26

priate reformulation of a hypothesis in the light of relevant


evidence may be seen as an exercise in ampliative induction.
And the Method of Relevant Variables can be shown to consti-
tute a systematic basis for the analysis of reasoning in common
law jurisprudence. T o regard a hitherto ignored type of
27

circumstance as material to a decision in a particular trial is


comparable with discovering the operation of a hidden variable
in a particular natural-scientific experiment. Again, Keynes
and Polya, .developing the hints of some older mathematicians,
such as Euler and Laplace, thought that ampliative induction
was a mode of justification for mathematical conjectures that is
useful where a demonstration is not available. (This inductive 28

procedure is essentially non-demonstrative and so is not to be


confused with what is often called 'recursive induction',
whereby what is mathematically demonstrable about the first
member of a series, and also about the successor of any member
about which it is demonstrable, is held to be demonstrable about
2 6 For example, O . C . J e n s e n The Nature of Legal Argument (Oxford: Blackwell, 1957),
28, and E . H . Levi, 'An Introduction to Legal Reasoning', University of Chicago Law
Review, 15 (1948), 502.
2 7See Cohen, The Implications of Induction 155-71.
2 8J . M . Keynes, A Treatise on Probability (London: Macmillan, 1921), 242-3, and G.
Polya, Patterns of Plausible Inference (Princeton: Princeton University Press, 1954), 109 IT.
§22Ampliative Induction 173
every member of the series.) And, though Keynes and Polya
thought that any ampliative inductive justification of a math-
ematical conjecture should be assessable by an appropriate
Pascalian credibility-function, Lakatos developed some of
Polya's ideas in a way that, in effect, exhibited a non-standard
use of the Method of Relevant Variables. He was particularly
concerned to explore the various ways in which a mathematical
conjecture might be revised, rather than rejected, in the face of
counter-examples. Moreover, several other kinds of non-
29

factual issues are open to inductive reasoning in accordance with


the Method of Relevant Variables. For example, we can grade
the acceptability of ethical, grammatical, or philosophical
principles, or of computer-implementable expert systems
programs, in this way. 30

In general a subject-matter is implied to admit of such in-


ductive reasoning if non-accidental uniformities are thought
capable of existing in it. But it is important to be careful about
what exactly are to count as the evidential data in each case.
Harman has argued, for example, that scientific theories have
an explanatory role in relation to particular observings that
ethical theories do not have in relation to particular exercises of
our moral sense. He points out that neither the relevant moral
principles nor the wrongness of the act contribute anything
towards explaining why you have the moral feelings that you do
when you see some children setting fire to a cat. But what kind 31

of explanation is at issue here? If the explanation in question is


of some matter of fact—that is, if it is an explanation that a
causal or statistical hypothesis or a scientific theory might
provide—then of course one should not expect any normative
judgement to provide that type of explanation. Indeed, when it
is claimed that norms can constitute a topic for inductively
supported generalization as well as facts can, then an intrinsic
part of the claim is that explanatory inductive generalization

2 9I. Lakatos, 'Proofs and Refutations', British Journal for the Philosophy of Science, 14
(1963-4), 1-25, 120-39, 221-45, 296-342.
3 0See Cohen, The Implications of Induction, 172-82 and, for inductive reasoning in
philosophy, L.Jonathan Cohen, The Dialogue ofReason (Oxford: Clarendon Press, 1986),
63-148.
3 1G. Harman, The Nature of Morality (New York: Oxford University Press, 1977),
1-9.
174 The Baconian Gradation of
need not be tied up exclusively with factual explanation in terms
of causes, probabilities, and so on. More specifically, if the
explanation sought is not of why you have the moral feelings that
you do but rather of why the moral feelings that you have are the
right ones to have—that is, if the explanation sought is not an
explanation of your attitude's de facto psychological properties
but of its dejure normative status—then the relevant moral prin-
ciple may provide a perfectly good explanation. You are right to
be outraged by what the children are doing to the act, people
might say, because it is an act of wanton cruelty. And the
relevant datum in such a case is the wrongness of the act, which
normal people feel, rather than the fact that they feel this
wrongness.
We need to note here, however, that the general theories,
principles, and so on that are cited in our explanations serve this
explanatory purpose best if they also have inductive support that
is independent of the items currently requiring explanation.
Moreover, it is by no means the rase that every inductively
supported generalization provides the best explanation of its
own evidential support. The blackness of the birds that we now
see on the summer fields is perhaps best explained by the opera-
tion of their genes, or by the operation of certain environmental
factors in the evolution of their species, but not at all well
explained by the fact that they are crows and the generalization
that all crows are black. Yet they constitute inductive evidence
for that generalization. So it is certainly not safe to suppose that
induction is always an inference to the best explanation. 32

Finally, it should be noted that, while the Method of Relevant


Variables imposes a realist interpretation on the underlying
Baconian syntax, an idealist interpretation is also possible. That
is, we can discern this generalized modal structure not only in
judgements of legisimilitude—that is, in judgements about the
objective falsificatory powers of relevant factors—but also in
certain kinds of judgement about the subjective acceptability of
hypotheses. Levi's logic of acceptability, for example, which
was developed on the basis of earlier work by G.L.S. Shackle,
has a non-multiplicative principle for conjunction that is
isomorphic with the principle implicit in the Method of Relevant
Variables. 33

3 2 PaceG. Harman, Thought (Princeton: Princeton University Press, 1973), 130-41.


3 3 See I. Levi, Gambling with Truth (New York: Knopf, 1967), and 'Potential Surprise:
its Role in Inference and Decision-Making', in Cohen and Hesse (eds.), Applications of
Inductive Logic, pp. 1-27.
§22Ampliative Induction 175
In sum, then, what has been shown in the present chapter is
that the Method of Relevant Variables presents us with a mode
of ampliative induction which is based on relevant evidential
variety. It systematizes some key intuitions in the Baconian
tradition, but has a non-Pascalian structure which may be
articulated in a generalized modal logic. This logic is applicable
over a wide field of reasoning-patterns, and we must therefore
now pass on to examine some further considerations that bear on
the relative merits of Pascalian and Baconian analyses of
ampliative induction.
VI
Four Paradoxes about Induction

§ 23. T H E C L A S S I C A L P R O B L E M O F I N D U C T I O N

Hume's problem about induction cannot be resolved by


reference to the ideas of Bernoulli, of Kant, or of Harre and
Madden. Nor is it satisfactory to accept Hume's own resolution
of the paradox by reference to the operation of our natural
inclination. Popper's earlier treatment of the issue tended to cut
science off from any rationally defensible utilization, while any
serious exploitation of the concept of verisimilitude, as in his later
writings on the subject, tends to topple over into unacknow-
ledged inductivism. Since Hume attacked the rationality of
induction by stressing those features in which it apparently
differs from deduction, we may answer his attack by stressing
those features in virtue of which it resembles deduction.
Deducibility must be seen as a limiting-case of inducibility,
whether the latter be conceived in Pascalian or in Baconian
terms. But we can never have conclusive certainty about any fact
that is as .yet unobserved.
Paradoxes have always been a stimulus to philosophical
analysis. W h e r e irresistible considerations seem to lead to
opposed conclusions, philosophical intervention is needed. But
the guiding motivation of such intervention is to maintain the
consistency of the whole intellectual framework within which we
reason, not just to solve a particular puzzle in its own terms. So
in examining various paradoxes about induction we should treat
them as providing a battery of tests that any adequate account of
induction should pass, rather than as constituting a set of
unrelated problems that admit of mutually independent
solutions.
T h e classical paradox about induction arose as a by-product
of empiricist epistemology in the late seventeenth and early
§23Four Paradoxes about Induction 177
eighteenth centuries. It was hinted at in Locke's Essay and else-
where' and explicitly formulated for the first time in Hume's
Treatise though neither Locke nor Hume used the term 'induc-
2

tion' at all in this context. The paradox was generated by the


3

fact that, according to an empiricist, the premisses of our factual


knowledge were all particular, distinct perceptual experiences,
not general principles of any kind, whereas the content of our
factual knowledge seemed to include many causal laws and other
generalizations about natural processes. A rationalist could
claim that such general conclusions were deducible, in accord-
ance with standard logical rules, from even more general
premisses that Providence had kindly made self-evident. But
there was no way that these general conclusions, with their
implications about the as yet unperceived, could be so deduced
from premisses about particular past events. Indeed, there was
nothing logically impossible, nothing self-contradictory, in
supposing such a commonly accepted conclusion to be false even
though the premisses were true.
So just at the very period at which the scientific revolution
seemed to have produced its greatest theoretical achieve-
ment—Newtonian mechanics—the triumph of empiricist over
rationalist epistemology seemed to imply, if Hume was right,
that this achievement had no rational foundation. No valid
inference was possible from the past to the future, or from what
is in sight to what is out of sight. Nor was it any use trying to
rescue the situation by an appeal to some overarching principle
about the uniformity of nature, whereby, in relevant respects,
the future might be guaranteed to be like the past and unseen
instances of a property like seen ones. With such a principle as a
major premiss, H u m e thought, the requisite deductions would
be forthcoming. But, he pointed out, the truth of that major

1 See J . Locke, Essay Concerning Human Understanding (London; T. Basset, 1690), bk.
IV, ch. 3, 5 14 adfin., andch. 6 § 16; and J. R. Milton, 'Induction before Hume', British
Journal for the Philosophy of Science, 38 (1987), 49-74.
2 A Treatise of Human Nature, (1739), ed. L. A. Selby-Bigge (Oxford: Clarendon Press,
1888) bk. I, pt 3, Sect. 14.
3 Neither Locke's Essay nor Hume's Treatise pays any attention, whether critical or
commendatory, to Francis Bacon's inductive methodology, though Bacon's writings
were Well known and widely read in the century after his death.
§ 23 §23Four Paradoxes about Induction 178
premiss would itself be at least as impossible to substantiate
by formallogically sanctioned deduction from the ultimate
premisses of experience.
Of course, where induction is not about factual matters, but
about a system of ethical principles, say, or of common law
maxims (see p. 172), no such paradox need arise, since a general
rule authorizing inductive inference may be supposed to be
implicitly stipulated with the imposition of the system. Justice
requires that like cases be treated alike, and legislators can
decide to impose justice on a human community. But we
humans cannot decide to impose uniformity on the processes of
Nature.
One tempting line of approach to Hume's problem is via
Bernoulli's limit theorem (see pp. 118-19). If under appro-
priate assumptions the examination of a sufficiently large
sample can let us learn something, with a corresponding degree
of confidence, about the probability with which a certain charac-
teristic exists in a specified population, then it looks as though a
confidence-function can operate here as a measure of predict-
ability by enumerative induction. But the value of such a func-
tion for this purpose is restricted by our having to assume that
the sample is not biased by some lack of homogeneity in the
population. If we want to calculate the survival prospects of a
twenty-year-old student, we should be wary of drawing our
sample solely from the members of our university's rock-
climbing club. Of course, as the weight of the evidence (see
§ 16) goes up, it would be normal to discount this kind of worry.
But assessments of weight rely on the assumption that existing
judgements of evidential relevance will remain true for the
future. So there is no escape here from the jaws of Hume's
problem: we are merely shifting the point at which the bite is felt.
Some philosophers have tried to resolve Hume's problem by
proposing a different view of the typical premisses from which
reasoning about causal connections proceeds. Thus Kant
argued that the objects of our perceptual judgements must them-
selves be structured by the category of cause and effect, if we are
to have the kind of experience that we do have. And every parti-
cular cause must presuppose a uniformity of causal connection.
So the uniformity of nature pervades the whole of the per-
ceivable world. But the details of Kant's intricate argument are
§ 23 Four Paradoxes about Induction 179
notoriously open to criticism. Nor is it at all clear from Kant's
4

text why the actual procedures of enumerative or variative


induction should have any rational cogency. His theory does not
elucidate why superiority in the number or variety of evidential
instances can justify grading the support for one hypothesis
higher than that for another. There is, for example, no attempt
at a transcendental argument for a principle of limited inde-
pendent variety (see § 18). 5

Harre and Madden have claimed that Hume was wrong to


6

regard the objects of our perception as distinct existences.


Instead, Harre and Madden argue, objects intrinsically possess
causal powers which they exercise on one another, and the
rationality of inductive inferences is founded on these links of
natural necessity. But unfortunately all the old difficulties about
proving the uniformity of nature arise again about proving the
reality of causal powers. One can hardly rely on arguing that
such powers are to be seen all round us, since so many philo-
sophers, over so many centuries, have evidently failed to see
them anywhere. And one can hardly argue that their underlying
reality is established by the observable uniformities that they
produce, since it is just those observable uniformities about
which, according to Hume, we have no justifiable certainty.
Hume's own way of handling the problem was to accept the
sceptical conclusion. O n his view human beings, like animals,
have a natural inclination to believe in the existence of a causal
connection between one kind of event and another, whenever
they have observed events of the one kind uniformly followed in
their experience by events of the other. And the inclination
becomes proportionately stronger, according as the events in
question resemble the hitherto observed ones more closely. This
natural inclination suffices for all the practical purposes of life.

4 I. Kant, Critique ofPure Reason (1781), trans. N. Kemp Smith (London: Macmillan,
1929), 218-27. For an example of the criticism see S. Korner, Kant (Harmondsworth:
Pelican Books, 1955), 87.
5 Indeed, Kant had at one time held that natural variety is unlimited: W . R . Shea,
'Filled with Wonder: Kant's Cosmological Essay, The Universal Natural History and Theory
of the Heavens', in R.E. Butts(ed.), Kant's Philosophy of Science (Dordrecht: Reidel, 1986),
110-14.
6 R. Harre and E.H. Madden, Causal Powers: A Theory of Natural Necessity (Oxford:
Blackwell, 1975).
§ 23 §23Four Paradoxes about Induction 180
So the absence of any rational justification for the belief is a
matter of concern only to philosophers, and the latter can dispel
their worries at any moment by leaving their studies and in-
dulging their natural inclinations.
Such a resolution of the paradox has at least three weaknesses.
In increasing order of importance they are as follows.
First, it is in any case an inaccurate psychological observation
that when people believe in a causal connection they have
previously perceived a constant conjunction, since the causal
connections that we discover in everyday life are not to be identi-
fied with the underlying uniformities that scientists aim at dis-
covering. We all believe that a fire emits sensible heat although
we also know that heat insulation often prevents this emission.
The causal connections that we talk about in everyday life are
sequences of events that hold good only in normal circum-
stances, not universally. But H u m e could easily have corrected
himself on this point, without sacrificing anything integral to his
sceptical thesis.
Secondly, and more importantly, Hume seems to be arguing
with his readers and trying to persuade them of the existence of a
psychological law. This law is to the effect that people develop
the natural inclination which H u m e describes whenever they
have observed the corresponding conjunction of events to be
uniformly present in their experience. But belief in the existence
of such a psychological law can itself have no rational founda-
tion, if H u m e is right, since no causal law then has a rational
foundation. So, like many other sceptics, H u m e can be hoisted
on his own petard. If what he says is true, we should already
believe it before reading what he says, and if somehow we do not
already believe it he can certainly give us no reason for doing so.
Either way his remarks are pointless.
Thirdly, just as Hume's resolution of the paradox makes no
allowance for his own apparent ability to argue about psycho-
logical laws and the strength of evidence for their existence, so
too it makes no allowance for the many ways in which scientists
reason with one another about what the laws of nature are and
what the strength of evidence is for their existence. Since these
scientists apparently think that they are appealing to rationally
cogent considerations, some account of those considerations is
needed along with a brief explanation of their plausibility.
§23Four Paradoxes about Induction 181
Indeed Hume himself does give an account of them in another
section of his Treatise where he speaks of' rules by which to judge
of causes and effects', of 'the LOGIC that I think proper to employ
in my reasoning' on the subject, and of a constant conjunction
that 'proves' a causal connection. But this way of speaking is
7

not consistent with the sceptical thesis that is Hume's dominant


doctrine, since it concedes the existence of an inductive logic,
however jejune, as distinct from the deductive logic which is all
that Hume's sceptical thesis admits. Nor does Hume ever offer
any explanation of why it is that scientists and others come to
believe that they are appealing to rational considerations, which
could stand beside his explanation of why they come to believe in
the existence of this or that natural uniformity.
Popper's way of handling the problem is to accept Hume's
8

conception of logical reasoning as being purely deductive, but to


reject any inference from this that certain characteristic patterns
of scientific thinking are inherently non-rational. On Popper's
view, though ampliative induction has no logical validity, that
does not detract from the rationality of science since valid sci-
entific thought is always deductive, never inductive. And valid
scientific thought is deductive because scientists properly seek to
falsify their conjectures, not to verify them. For falsification the
standard logical procedure of modus tollendo tollens suffices, in
that, if a prediction deducible from a generalization is falsified,
so too is the generalization itself. Admittedly scientists require
any conclusive falsification of a hypothesis to be replicable in
order to guarantee inter-personal objectivity. So that it might
seem as though Popper has to let inductive reasoning come in
again by the back door in order to establish the replicability of
any test-result that is supposed to falsify a hypothesis conclus-
ively. But Popper treats the claim that such-or-such a test-result
is replicable as being subject in its turn to the falsificatory pro-
cedures of deductive logic. There is thus on his view an inevit-
able regress by which systems of theory are tested by deducing
from them statements of a lesser level of universality, which are
then in turn themselves testable, and so on. At any one time the
7 Bk. I, pt 3, Sect. 15.
8 K . R . Popper, The Logic of Scientific Discovery (London: Hutchinson, 1959), 29-30.
The main text of this book was originally published in German as Logik der Forschung
(Vienna: Julius Springer, 1934).
§ 23 §23Four Paradoxes about Induction 182
regress is just interrupted in practice at some point by scientists'
decision or convention to accept a testable proposition without
(as yet) actually testing it further.
Of course, falsification will not do the whole job, because at
any one time, in a particular field of enquiry, more than one
hypothesis may survive the toughest tests currently available.
How, then, does one choose between one unfalsified hypothesis
and another? In such a situation, Popper said, scientists prefer 9

the hypothesis that has most content, as measured by its prior


improbability (which is to be judged in Pascalian terms). And
Popper's philosophy of science thus faces in a direction starkly
opposite to the philosophy of those who, like Carnap (see § 17
above), consider that the proper aim of scientists is to maximize
the Pascalian probability of their hypotheses on the available
evidence, since, other things being equal, the probability on the
evidence will, according to Bayes's Law (p. 23), be lower if the
prior probability is lower.
Later Popper came to think that his conception of scientific
reasoning needed supplementation, for two main reasons. 10

First, it did not do justice, he thought, to the fact that theories


which are known to be false may nevertheless be very highly
regarded by reputable scientists. Minor anomalies are not
treated as grounds for rejecting a powerful theory that has no
serious competitor in its field (see pp. 142-3 above). So one
needs to find a way of distinguishing between false theories that
are more valuable and false theories that are less valuable.
Secondly, there is a deeply engrained idea among scientists that
they are engaged in a search for truth, albeit for relevant and
interesting truth, and this idea is not at first sight compatible
with the doctrine that they characteristically seek to falsify their
hypotheses rather than to verify them.
Popper therefore developed a measure of verisimilitude
whereby even false hypotheses might be compared with one
another in respect of their truth-content. In terms of such a
measure an appropriate value was to be attached to powerful
theories that encountered minor anomalies, and at the same
time due regard could be paid to the intuitive idea that scientists
are engaged in a search for truth. As it turned out there were
9K . R . Popper, The Logic of Scientific Discovery, pp. 113-19.
K . R . Popper, Conjectures and Refutations: The Growth of Scientific Knowledge (London:
1 0

Routledge and Kegan Paul, 1963), 215-48.


§23Four Paradoxes about Induction 183
serious flaws in Popper's own measure for verisimilitude, and 11

other measures have been proposed to take its place, sometimes


directed towards assessing the quantity of true content and
sometimes towards assessing closeness to truth. But any admis- 12

sion of the importance of verisimilitude tends to jeopardize the


integrity of Popper's anti-inductivist stance. This becomes clear
when we consider whether verisimilitude can be measured.
If there is never any effective way even of estimating, let alone
of accurately measuring, the relative verisimilitude of two
falsified theories in the light of the available evidence, then the
concept of verisimilitude is idle. Under those conditions it would
have no discoverable application and nobody could be said to be
aiming at superior verisimilitude because nobody could hope to
recognize when or where this had been achieved. But, if the
concept of verisimilitude is not idle and there is sometimes an
effective way of estimating relative verisimilitude in the light of
the available evidence, then it looks as though induction is being
allowed back under another name. It is just that instead of trying
in such cases to establish the absolute and unqualified truth of
his hypothesis, a scientist is said to be trying, more modestly, to
establish its superior share of, or closeness to, the truth. And
Popper himself suggests that, where two theories have both been
falsified but one has withstood some tests that the other has
failed, this would count as evidence that the former has greater
verisimilitude than the latter. We should then 'accept' the
former theory, he says, 'because of experiments which were
crucial between' the two theories. But there is little essential
13

difference between this procedure and what the Method of


Relevant Variables entails in a similar context (see § 20). What
is really being sought is legisimilitude, not just verisimilitude. 14

11 P. Tichy, 'On Popper's Definition of Verisimilitude', British Journal for the Philo-
sophy of Science, 25 (1974), 155-60, and D. Miller, 'Popper's Qualitative Theory of Veri-
similitude', ibid. 166-77.
1 2 See e.g. G. Oddie, Likeness to Truth (Dordrecht: Reidel, 1986); I. Niiniluoto, Truth-
likeness (Dordrecht: Reidel, 1987); and the collection of papers by various authors:
T.A.F. Kuipers (ed.), What is Closer-to-the-Truth? (Amsterdam: Rodopi, 1987).
13 Popper, Conjectures and Refutations, p. 235.
1 4 On resemblances between Popper's philosophy of science and Bacon's see P.
Urbach, 'Francis Bacon as a Precursor to Popper', British Journal for the Philosophy of
Science, 33 (1982), 113-32. On resemblances between Popper's philosophy of science and
the philosophies of Whewell and of Peirce see I. Niiniluoto, 'Notes on Popper as
Follower of Whewell and Peirce', Ajatus, 37(1978), 272-327
§ 23 §23Four Paradoxes about Induction 184
Nor should this be surprising, since only legisimilitude, not veri-
similitude, can guarantee the applicability to counterfactual
instances that is often needed when knowledge of general laws or
causal connections is exploited (see pp. 131-2).
Indeed, it seems a general weakness in Popper's position that
when attempts are made to buttress it against obvious implaus-
ibilities it tends to topple over into some kind of unacknow-
ledged inductivism. Lakatos, for example, wanted to emend it
in a different direction. Instead of considering the rationality
issue in relation to single hypotheses, Lakatos thought rather
about research programmes, in which a series of hypotheses
replaced one another as the programme either succeeded or
petered out. But (though his objections to the qualification of
15

hypotheses in the light of new evidence were too sweeping ) the 16

criteria that Lakatos proposed for telling a 'progressive' from a


'degenerating' research programme were largely the same as
those familiar in the inductivist tradition for choosing between
better and worse theories. What makes the move from theory to
theory a progressive research programme is that it achieves
some kind of consilience and also leads to novel predictions that
are subsequently confirmed (compare pp. 151-3 above). More-
over, even a Popperian emphasis on falsification is quite com-
patible with the Baconian conception of induction by elimina-
tion. Bacon himself wrote that 'the foundations of true induction
are laid in elimination'. 17

O n the other hand, if we stay with the hard core of Popperian


anti-inductivism we are left with a conception of pure science as
a mode of enquiry that reacts excessively to minor anomalies
and is totally cut off from any rationally defensible utilization for
technological purposes. In Popperian science, unfortified by the
doctrine of verisimilitude, we aim at maximizing the boldness of
the theories that we entertain and at rejecting all falsified
theories if a competing theory survives available tests. But what
1 5 I. Lakatos, 'Falsification and the Methodology of Scientific Research Pro-
grammes', in I. Lakatos and A. Musgrave (eds.), Criticism and the Growth of Knowledge
(Cambridge: Cambridge University Press, 1970), 91-195.
1 6 For a relevant example (Mendelian genetics) see R.D. Rosenkrantz, Inference,
Method and Decision (Dordrecht: Reidel, 1977), 127-8.
17 Novum Organum, bk. II, aphorism xix: The Works of Francis Bacon, ed. J. Spelding, R.
Ellis, and D . N . Health (London: Longman, 1859) i. 260 see also bk. I, aphorism xlvi
(Works, i. 166).
§23Four Paradoxes about Induction 185
reason has an engineer or physician to believe that such sur-
viving claims are also true? According to anti-inductivist
Popperianism he can have none. The alleged justification for
building an aircraft in a certain way or for prescribing a certain
pill is not the evidential support that exists for saying that the
theory endorsing such action is true but rather the fact that this
theory is the boldest to escape falsification from available
evidence. In other words Popperian technology encourages
engineers and physicians to take as big a risk as the circum-
stances allow, since it does not burden them with the duty to
possess even a partial justification for the truth of any predic-
tions that they make. But fortunately our engineers and physi-
cians have so far failed to respond to this encouragement: many
of us would not be still alive if they had done so. 18

Accordingly it looks altogether too paradoxical to suppose, as


Hume and Popper do, that only deductive inferences are
rational. Not surprisingly, therefore, some philosophers, like
Strawson, come to assume, in one way or another, that the
19

whole problem is resolved just by rejecting this supposition.


Some inferences, they say, are inductively valid, in that it is
rational, or reasonable, to proportion one's convictions in such
cases to the multiplicity or variety of the evidential data. That is
exactly what such a term as 'rational' means there. So it is wrong
to suppose, we are told, that any further justification of induc-
tion is either necessary or possible. To ask whether inductive
procedures are rational is like asking whether the law is legal.
But this is no answer to Hume. It begs the question, since the
rationality of induction was just what Hume was attacking by
stressing the differences between induction and deduction. The

1 8 For an attempt to develop a form of Popperianism that is not open to this criticism
see J . Watkins, Science and Scepticism (London: Hutchinson, 1984), 337-48.
1 9 For this see P. F. Strawson, Introduction to Logical Theory (London-. Methuen, 1952),
248-50. See also A.J. Ayer, The Problem of Knowledge (Harmondsworth: Penguin Books,
1956), 74-5, and P. Edwards, 'Bertrand Russell's Doubts about Induction', in A. Flew
(ed.), Logic and Language (Oxford: Blackwell, 1951), 68-70. Further recent analyses of the
problem are to be found in G. H. von Wright, The Logical Problem of Induction (Oxford:
Basil Blackwell, 1957); N. Rescher, Induction (Oxford: Basil Blackwell, 1980); D.C.
Stove, The Rationality ofInduction (Oxford: Clarendon Press, 1986); and in a useful collec-
tion of papers by various authors, R. Swinburne (ed.), The Justification of Induction
(Oxford: Oxford University Press, 1974).
§ 23 §23Four Paradoxes about Induction 186
issue that has to be settled may be put in these terms: what
entitles us to use the same evaluative term 'rational', or
'reasonable', in both contexts? Hume pointed out a pervasive
lack of analogy between, on the one hand, reasoning from
premisses to logically implied conclusions, and, on the other,
reasoning—vulgarly so called—from premisses about the
already observed to conclusions that embrace the as yet
unobserved. The former is a product of thought, he argued, the
latter of custom; the former justifies certainty, the latter not; the
former cannot be rejected without self-contradiction, the latter
can; and so on. It does not matter here whether we speak of
'reasoning' or of 'rationality' or of 'validity'. The question that
arises is the same: how can we establish the existence of a suffi-
ciently strong analogy between deduction and induction so as to
undermine H u m e ' s scepticism?
The clue to an answer for this important question lies in the
fact that inductive reasoning is a matter of degree, not of all or
nothing. So deducibility must come to be seen as a limiting case
to which inductively based inferability may be seen to exhibit
gradable degrees of approximation, whether we analyse such
gradations in terms of Pascalian probability, Baconian legi-
similitude, or some other kind of parameter. We are not then left
with two different kinds of inferential rationality, a deductive
kind and an inductive kind, but rather with one or more scales of
rationality or reasonableness, on each of which deductive
inferability.figures as a limiting case. Indeed, we have already
seen (§ 15) that in order to understand why there is such a
variety of different conceptions of Pascalian probability we can
usefully regard probability as a gradation of inferability. The
diversity of available interpretations for the Pascalian calculus
can then be seen to correspond with the familiar diversity of
systems for deductive inferability. Similarly, the Method of
Relevant Variables, in each of its varied applications, was seen
(§ 21) to conform to the rules of a generalized modal logic, with
degrees of legisimilitude mounting towards necessity. There is,
therefore, no case now for supposing, as Hume supposed, that
deductive and inductive thinking are so different from one
another that the former possesses some honorific status to which
the latter does not even approximate. Rather there is an obvious
analogy between judgements of the form 'The inference to A
§23Four Paradoxes about Induction 187
from B is logically certain' and judgements of the form 'The
inference to A from B has this-or-that level of probability', since
both kinds of judgement conform to the same Pascalian axioms.
And there is also an obvious analogy between judgements of the
form 'The proposition that if A, then B, is necessary' and 'The
proposition that if A, then B, has this-or-that level of legisimili-
tude' since both kinds ofjudgement conform to the same modal-
logical axioms.
Of course, there is another aspect of Hume's problem which
these logico-syntactic rapprochements between deduction and
induction can do nothing to resolve. Hume insisted that no
conviction about the future, or the unseen, could ever be justi-
fied on the basis of the evidence that we at present have. And, if
we read 'justified' here as meaning 'conclusively justified', we
must surely agree with him. The validity of the inductive
inferences that we make from available evidence must depend
on the reliability of the assumptions that we adopt about such
inferences in each recognized category of factual enquiry. How-
ever various the evidence that we at present have, there must
always remain the possibility, as we saw (§ 20), that our
assumed list of relevant variables, or our assumed categoriza-
tion of hypotheses, will at some time be held to be faulty. Our
evaluation of the evidence may then have been incorrect either
because the results of some past experiment were affected by the
undetected operation of an unlisted variable or because a new
experiment, can be designed in which such a variable is deliber-
ately manipulated. In other words any inductive judgement is
itself empirically corrigible. Conclusive certainty is never
inductively justifiable.
But that is a deprivation that we can live with. To keep our
minds open to the possibility of new evidence is all that it
requires of us, and while we do this we are still entitled to claim
an appropriate degree of justification for the theories that have
survived the most thorough tests which we at present believe
ourselves capable of devising. H u m a n rationality requires us at
any one time to do the best we then can, not to do the best that
could ever be possible.
§ 23 §23Four Paradoxes about Induction 188
% 24. T H E P A R A D O X O F T H E R A V E N S

Hempel's paradox of the ravens arises from the fact that three
propositions about inductive confirmation that are each
independently plausible are not co-tenable. Scheffler sought,
unsatisfactorily, to resolve the paradox by defining confirmation
in such a way that logically equivalent hypotheses may not be
equally confirmed by given evidence. Hempel's kind of solution
is more satisfactory. It requires a Pascalian measure for
confirmation and relies on our already knowing that black ravens
are much rarer than non-black non-ravens. Alternatively, if we
grade confirmation by the Method of Relevant Variables, our
resolution of the paradox must point out that known relevant
variables operate causally on plumage-colour rather than on
species-membership.
Hume's argument against the rationality of induction rested on
his assumption of a deductivist paradigm. He looked at induc-
tion from the outside, as it were, and could see in it only a
varying strength of mental process, not a gradable pattern of
evidential justification. But, even when we look at induction
from the standpoint of its own apparent rules and principles and
acknowledge the analogies that rebut Hume's scepticism, we
may still find ourselves confronted with antinomy and paradox.
One such antinomy is commonly called 'the paradox of the
ravens', and is best known from Hempel's publication of it in
1945. Here we may formulate the core of the problem in terms
20

of three propositions that are not co-tenable despite the fact that
each has considerable intuitive appeal:
(1) Any object that is both an A and a B confirms the hypo-
thesis that everything which is an A is a B.
(2) Any object that confirms a hypothesis confirms also any
proposition that is logically equivalent to that hypothesis.
(3) A white handkerchief does not confirm the hypothesis
that all ravens are black.
2 0 C . G . Hempel, 'Studies in the Logic of Confirmation', Mind, 54 (1945), 1-26,
97-121, repr. with some changes, in C . G . Hempel, Aspects of Scientific Explanation and
Other Essays in the Philosophy of Science (New York: Free Press, 1965), 1-51. The germ of
the problem originally appeared in C . G . Hempel, 'Le Probleme de la verite', Theoria, 3
(1937), 206-46.
§23Four Paradoxes about Induction 189
These three propositions can easily be seen to generate an
antinomy. A white handkerchief is a non-black thing that is a
non-raven. So, according to (1), it confirms the hypothesis that
everything which is non-black is a non-raven. And therefore,
according to (2), it confirms the logically equivalent hypothesis
that all ravens are black. But that is inconsistent with (3) which
asserts, plausibly enough, that a white handkerchief does not
confirm the hypothesis that all ravens are black. It follows that, if
a coherent account of inductive reasoning is to be presented, one
or more of the premisses on which this antinomy rests has to be
rejected as false or inapplicable.
Thus some philosophers have held, in effect, that (2) is the
source of the trouble. For example, Scheffler has argued for the 21

importance of the notion of 'selective confirmation' as distinct


from the notion of 'confirmation' defined in (1). He stipulates
that for a piece of evidence to be said to confirm a hypothesized
generalization selectively it must not only satisfy that generaliza-
tion but also refute its contrary. So, on the one hand, an object
that is both an A and a B selectively confirms the hypothesis that
everything which is an A is a B, as in (1), because it not only
satisfies that generalization but also refutes its contrary, namely
the hypothesis that everything which is an A is not a B. On the
other hand, selective confirmation is not automatically trans-
mitted to equivalent propositions, as in (2) because not every
object that selectively confirms 'All /Is are Bs' will also selec-
tively confirm 'All non-Ss are non-^s'. The contrary of 'All
non-ZJs are non-^4s' is 'All non-fis are .4s' and the latter
generalization is not refuted by an object that is both an A and a
B, even though such an object does satisfy the former generaliza-
tion. In other words, if we interpret 'confirms' in (1) and (2) as
'selectively confirms', we have to accept that (2) is false: the
confirmation attributable to a specified hypothesis may not be
attributable to a logically equivalent hypothesis. And the para-
dox then dissolves because the confirmation that a white
handkerchief gives to the hypothesis 'All non-black things are
non-ravens' is not now automatically passed on to its equivalent
'All ravens are black' and thus no conclusion is generated that is
incompatible with (3).
I. Scheffler, The Anatomy of Inquiry: Philosophical Studies in the Theory of Science (New
2 1

York: Knopf, 1963), 283-9.


§ 23 §23Four Paradoxes about Induction 190
Scheffler's solution of the paradox is clearly inconsistent with
any analysis of confirmation in terms either of simple Pascalian
probability or of positive Pascalian relevance. That is because
such an analysis is compelled by the logical syntax of Pascalian
judgements to accept (2): logically equivalent propositions are
interchangeable with one another within any judgement of
Pascalian probability, since they give each other the probability
1 (see pp. 85-7). So it may seem that, if Scheffler's preferred
criterion of confirmation is to be provided with a rationale, this
is best found in the Baconian idea of eliminative induction,
whereby support for one hypothesis is built up by excluding its
rivals. Because Scheffler's solution assumes that every object or
event that confirms a hypothesis must eliminate a contrary
hypothesis, it seems to go along with the Baconian method that
was described in § § 20 to 21. But in fact there is an important
difference, because, so far as 'confirms' means the same as
'inductively supports', the Method of Relevant Variables is just
as much committed to the truth of (2)—which Scheffler
rejects-—as is any Pascalian analysis. As we saw (pp. 139-40)
any judgements of inductive support that are based on the
Method of Relevant Variables must accord with the Special
Consequence Condition. And it is easy to see that the Special
Consequence Condition entails (2). For that principle gives us
both that, if //logically implies H', then E's support for H' is
greater than or equal to its support for H, and also that if H'
logically implies H, then E's support for H is greater than or
equal to its support for H' \ and when these two corollaries of the
Special Consequence Condition are conjoined the conjunction
is equivalent to (2).
But, even apart from considerations germane to the Method
of Relevant Variables, it is arguable that a solution which is
bought at the price of rejecting (2) is too costly to be satisfactory.
If it is logically necessary for H to be true if and only if H' is true,
it hardly seems conceivable that there should be evidential
support for H which is not also evidential support for H'.
Accordingly it may seem tempting to argue not that (2) is false
but just that it does not apply in the context in which the paradox
is supposed to be generated. The reason why (2) does not apply
here would be that the hypothesized generalizations with which
principles of inductive logic, like (1), (2), and (3), are concerned,
§23Four Paradoxes about Induction 191
carry an existential implication. On this view a hypothesis of the
form 'Everything which is an A is a 5 ' or 'All /Is are Bs', asserts
not only that if there is anything that is an A it is also a B, but also
that there is something which is an^4. Consequently 'All ravens
are black' is not to be regarded as logically equivalent to 'All
non-black things are non-ravens', since the former generaliza-
tion logically implies that there is something which is a raven but
does not imply that there is something which is a non-black
thing, while the latter generalization has the opposite sense.
Hence the white handkerchief that, because it is a non-black
non-raven, confirms 'All non-black things are non-ravens',
does not confirm 'All ravens are black' and so (3) is not
contradicted.
But, as Hempel himself pointed out, such an argument rests
on a rather shaky foundation. The representation of every
general hypothesis by a conjunction of a universal conditional
and an existential sentence would invalidate many deductive
inferences that are generally accepted as logically permissible in
scientific reasoning. Thus, for example, the assertions that all
sodium salts burn yellow, and that whatever does not burn
yellow is no sodium salt, are customarily treated as logically
equivalent formulatiQns. Moreover, there are many kinds of
generalization in natural science for which it would be very diffi-
cult to determine the exact location of an existential implication.
Consider the hypothesis that, if a person after receiving an injec-
tion of a certain test substance has a positive skin reaction, he has
diphtheria. Should we construe the existential clause here as
referring to persons, to persons receiving the injection, or to
persons who, upon receiving the injection, show a positive skin
reaction? T o the extent that the decision seems arbitrary, the
alleged existential implication seems unattached. One could
add, too, that so far as the assertion of 'All .4s are 5s' commits a
22

speaker to the existence of /Is, it also normally commits him to


the existence of non-/4s (else he would have said 'Everything is a
5'), to the existence of 5s (since he is committed both to the
existence of /Is and to the view that all /Is are 5s), and to the
existence of non-5s (since if everything were a 5 there would be

2 2M. Fisch, 'The Paradoxes of Confirmation and their Solutions' (M. A. Disserta-
tion at the University of Tel Aviv: May, 1981).
§ 23 §23Four Paradoxes about Induction 192
no point in asserting that all /Is are Bs). So the assertion of 'All
non-/ls are non-5s' has just the same existential commitment as
has the assertion of 'All /4s are Bs'. And in any case there are
many uses in science for generalizations that definitely lack any
existential implications, because they relate to ideal or extreme
conditions which are either not known to exist or even known
not to exist. These generalizations are certainly a target for
inductive reasoning (as we saw in § 22), and the paradox will
still arise about them even if it fails to arise in more homely fields
of enquiry, like ornithology.
Accordingly Hempel was ready to accept the applicability of
(2) and, with it, the argument that any white object confirms the
hypothesis that all ravens are black. He thus treated (1) as
stating a sufficient, but not a necessary, condition for confirma-
tion. But he gave two reasons for rejecting proposition (3).
The first reason was that any generalization of the form
'Everything which is an A is a B' or 'All As are Bs' asserts some-
thing about, and imposes a corresponding restriction on, all
entities of the appropriate category (for example, on all physical
objects). It is not just about all /Is, and so, specifically, 'All
ravens are black' is not just about all ravens. Hence there is no
entity of the appropriate category that is not caught up in the
generalization, and any such entity, even a non-raven, is there-
fore in a position to constitute confirmatory or disconfirmatory
evidence.
But Hempel's argument here is not very persuasive. If there is
a concept of confirmatory evidence such that any entity within
the domain of a generalization constitutes either confirmatory or
disconfirmatory evidence for that generalization, then that
unlocalized (see p. 156) concept of confirmatory evidence is not
a very interesting or important one. It excludes any such entities
from being regarded as irrelevant to the confirmation or dis-
confirmation of the generalization. If the generalization that all
ravens are black is supposed to be 'about' physical objects, then
the colours of shoes and ships and sealing-wax are all to be
regarded as relevant to its truth. Thus the field of relevant
investigation is made unacceptably large, and recognition of this
provides a motive for taking the scope or domain of the
generalization to be co-extensive with the category of entities
denoted by its antecedent term. But, if we therefore fall back on
§23Four Paradoxes about Induction 193
taking the domain of'All ravens are black' to be just ravens, we
have done nothing to eliminate the paradox.
Hempel's other reason for rejecting proposition (3) was that
when we judge the extent to which some object confirms a stated
hypothesis we tend tacitly to introduce a comparison of the
hypothesis with a body of evidence which includes not only that
particular object but also other items of information with which
we happen to be acquainted. For example, suppose that in
support of the hypothesis 'All sodium salts burn yellow' some-
body was to adduce an experiment in which a piece of pure ice
was held into a colourless flame and did not turn the flame
yellow. That result might be held to support the hypothesis in
question because it confirms the generalization 'Whatever does
not burn yellow is no sodium salt'. And yet at the same time this
appears paradoxical, because we happen to know anyhow that
ice contains no sodium salt and thus our experiment seems
irrelevant to the hypothesis that it was designed to test. Dis-
regard the background knowledge, argues Hempel, and the
paradox disappears.
But this argument is not very persuasive either. The sodium
salt example introduces a special feature that is not present in the
way in which the original paradox of the ravens may be
presented. It involves a piece of chemical knowledge that had at
one time to be discovered by empirical scientific enquiry,
namely the knowledge that ice contains no sodium salt. But the
knowledge that handkerchiefs are not ravens is part of the a
priori linguistic competence with which we approach any
intellectual problem. It cannot so easily be disregarded in order
to resolve a particular paradox.
The spirit of Hempel's solution (though not of his arguments
for it) is better maintained if confirmation is treated quantita-
tively rather than qualitatively and degrees of confirmation are
measured in Pascalian terms. This approach to the problem was
originally taken by Hosiasson-Lindenbaum and was followed 23

by Mackie among others. We are still to reject (3) but to hold


24

that the degree to which a white handkerchief confirms 'All


2 3J . Hosiasson-Lindenbaum, 'On Confirmation', Journal of Symbolic Logic, 5(1940),
133-48.
2 4J . L. Mackie, 'The Paradox of Confirmation', British Journal for the Philosophy of
Science, 13 (1963), 265-77.
§ 23 §23Four Paradoxes about Induction 194
ravens are black' is very small—so small, indeed, that we can
easily understand why we might suppose it to be altogether non-
existent. Mackie claims that to develop such a solution we have
to appeal to what he calls the 'Inverse Principle', which states
that a hypothesis H is confirmed by an observation-report E in
relation to background knowledge K if and only if the observa-
tion-report is made more probable by adding the hypothesis to
the background knowledge. That is, p(E\H&.K) must be greater
than p ^ l i f ) , and E confirms H better the more the adding of H
to K raises the probability of E. He then argues that, if we
assume this principle and assign probabilities according to the
relative sizes of the several classes constituted by all ravens, all
non-ravens and so on, observation-reports of the form 'This is a
black raven' will confirm the hypothesis 'All ravens are black'
far more than do observation-reports of the form 'This is a non-
black non-raven'.
Much the same result is achieved if instead of Mackie's
Inverse Principle we adopt either the simple probability analysis
of ampliative induction or the positive relevance analysis (§ 19).
This is because of the relationship between p(E\H&K) and
p(H\ElkK) that is determined by Bayes's law, in the form
P ( M . p .

Indeed one can easily see from this law that, other things being
equal, p(H\E&.K) will get larger as P(E\K) gets smaller, and vice
versa. So that because non-black non-ravens are common
objects, with a relatively high probability of occurrence, they
can do little to raise the probability of the hypothesis 'All ravens
are black', whereas black ravens are relatively rare objects and
will raise that probability much more (see pp. 136-7).
This way of resolving Hempel's paradox will not satisfy any-
one who is intuitively convinced that any observations of non-
black non-ravens are totally irrelevant to any judgement of the
evidential support or confirmation that exists for the hypothesis
'All ravens are black'. All that such a resolution of the paradox
can do to explain away those intuitions is to suggest that they
arise from mistaking a very, very low degree of relevance for no
relevance at all.
Philosophers who find this explanation implausible, and also
wish to retain (2), have no alternative but to reject (1), and the
§23Four Paradoxes about Induction 195
Method of Relevant Variables provides a rationale for so doing.
According to that method there are just two ways in which a
hypothesis may be shown to possess inductive support. Either an
appropriately structured test on the hypothesis turns out to have
satisfactory results, or the hypothesis is shown to be the logical
consequence of one or more propositions that have been satis-
factorily so tested. It follows that the evidence which supports a
hypothesis, according to the Method of Relevant Variables, is
constituted by the results of canonical tests on that hypothesis or
on logically entailing ones, and not by objects that just happen to
satisfy the antecedent and consequent of the hypothesized
generalization. Observation of a black raven, tout court, does not
provide any inductive support for the hypothesis that all ravens
are black, nor does a non-black non-raven provide any inductive
support for the hypothesis that all non-black things are non-
ravens. Thus (1) is rejected and the paradox cannot be
generated in Hempel's terms.
But that is not enough to dispose of the underlying problem, if
the paradox can be restated in terms of a concept of confirmation
that conforms to the Method of Relevant Variables. And indeed
it might seem that the paradox reappears if (1) is replaced by
(1') Any set of objects that are both ^4s and Bs in each of an
appropriately selected variety of circumstances, con-
firms the hypothesis that everything which is an A is a B
and (3) is replaced by
(3') A set of white handkerchiefs in each of the circumstances
referred to in (1') does not confirm the hypothesis that all
ravens are black.
For in the light of (1') and (2) it might seem that a set of white
objects that have been observed in an appropriately wide variety
of circumstances confirms the hypothesis that all ravens are
black, whereas (3') denies that such a set of objects does confirm
the hypothesis.
However, when we use the Method of Relevant Variables we
are always in the position of having to exploit background know-
ledge or belief or assumptions about what variables are indeed
relevant to the hypothesis under examination. If we have no
such knowledge or belief or assumptions, we are not in a position
to test the hypothesis appropriately. Now, we may well know of
§ 23 §23Four Paradoxes about Induction 196
circumstances that are relevant to hypotheses of the category to
which 'All ravens are black' belongs. There are circumstances
like season, climate, diet, and so on, that may cause a bird to
change its plumage colour. But how could we have the requisite
kind of knowledge or belief about the hypothesis 'All non-black
things are non-ravens'? In order to have it we should need a list
of variables relevant to hypotheses of this category—that is, a list
of types of circumstances that have sometimes falsified such
hypotheses by causing objects or events that satisfy their subject-
terms to fail to satisfy their predicate-terms. But in fact we have
no information about circumstances that can thus turn a non-
raven into a raven or a non-swan into a swan. Such metamorph-
oses, or changes of species-membership, do not occur within our
experience and are contrary to what is thought genetically
possible. Consequently, even if a set of white objects were
observed in a wide variety of circumstances, it would give no
support thereby to the hypothesis that all non-black things are
non-ravens, if we are to judge support by the Method of
Relevant Variables. And unless there is some support for that
hypothesis the paradox cannot get off the ground. Or, in other
words, though 'All non-black things are non-ravens' must have
the same grade of inductive support, on given evidence, as 'All
ravens are black', such evidence must be the outcome of tests on
the latter proposition, not the former, if we adopt the Method of
Relevant Variables.
In sum, so far from its being the case that the paradox arises
because we take background knowledge into account inappro-
priately, as Hempel held, we actually have to take background
knowledge into account, quite properly, in order to dissolve the
paradox. While a Pascalian analysis of inductive reasoning
requires us to invoke contingent beliefs about the numbers of the
objects concerned so as to reject (3), a Baconian analysis in terms
of the Method of Relevant Variables requires us to reject (1) and
then to invoke contingent beliefs about the direction in which the
relevant variables operate in canonical tests. 25

2 5 Many other discussions of Hempel's paradox are to be found, such as M. Black,


'Notes on the "Paradoxes of Confirmation" ', in J . Hintikka and P. Suppes (eds.),
Aspects of Inductive Logic (Amsterdam: North-Holland, 1966), 175-97; P. Suppes, 'A
Bayesian Approach to the Paradoxes of Confirmation', ibid. 198-207; G . H . von
Wright, 'The Paradoxes of Confirmation', ibid. 208-18; R. Swinburne, An Introduction
to Confirmation Theory (London: Methuen, 1972), 152-73, which includes a useful
§23Four Paradoxes about Induction 197
§ 25. T H E ' G R U E ' P A R A D O X

Goodman's 'grue' paradox cannot be resolved by restricting the


terminology of our hypotheses to purely qualitative terms that
imply nothing about a particular time or place. And Goodman's
own solution, in terms of a preference for what he calls
'entrenched' predicates, would force science into an excessively
conservative posture. Instead we need to invoke either Bayes's
theorem or the Method of Relevant Variables.
The definition of confirmation that is invoked in the paradox of
the ravens is also involved in another paradox, propounded by
Goodman. Suppose that all emeralds examined before a
26

certain time t are green. At time t, then, all our relevant observa-
tions confirm the hypothesis that all emeralds are green. But
consider the predicate 'grue' which applies to all things
examined before t just in case they are green and to other things
just in case they are blue. Obviously at time t, for each statement
of evidence asserting that a given emerald is green, we have a
parallel evidence-statement asserting that that emerald is grue.
And each evidence-statement that a given emerald is grue will
confirm the general hypothesis that all emeralds are grue. Hence
a hypothesis implying that all emeralds subsequently examined
will be green and a hypothesis implying that they will all be blue
are both confirmed at t by evidence statements describing the
same observations. Two mutually conflicting hypotheses are
equally well confirmed by the same evidence. Moreover, since t
may be whenever you please, the evidence collected prior to t
may contain very, very many green emeralds but the incom-
patible predictions will still be equally well confirmed at t. And
by choosing an appropriate predicate instead of 'grue' we can
clearly obtain equal confirmation for any prediction whatever
about other emeralds, or indeed for any prediction whatever
about any other kind of thing. For example, suppose 'grue'
applies to all things examined before t if and only if they are

pre-1972 bibliography; H. Gaifman, 'Subjective Probability, Natural Predicates and


Hempel's Ravens'. Erkenntnis, 14(1979), 105-47; a n d M . Fisch, 'Hempel's Ravens, the
Natural Classification of Hypotheses and the Growth of Knowledge', Erkenntnis, 21
(1984), 45-62.
2 6N. Goodman, Fact, Fiction and Forecast (London: Athlone Press, 1954), 73-120.
§ 23 §23Four Paradoxes about Induction 198
green and to other things if and only if they exist in a world in
which pigs have wings. Then, if emeralds examined before t are
green, we can predict that after t pigs will have wings. Nor is this
type of paradox confined to temporal extrapolations from
observable evidence. If 'grue' is redefined so as to apply to all
things examined in the observer's own laboratory if and only if
they have one characteristic, and to other things if and only if
they have a different one, then an analogous argument seems to
lead to the absurd conclusion that no experimenter is ever
entitled to draw universal conclusions about the world outside
his laboratory from what goes on inside it.
There are essentially two assumptions that combine to
generate this paradox. One is that a report of an A which is B
confirms a hypothesis of the form 'All/4s are B'. The other is that
a term like 'grue' can occur as the predicate of a well-confirmed
hypothesis. And in so far as the latter assumption carries little
intuitive conviction, because 'grue' is such a bizarre term, it
obviously needs to be discussed first. T o reject that assumption
we apparently need to impose some terminological constraint on
the confirmation of our hypotheses, so that the occurrence of a
predicate like 'grue' within a hypothesis comes to count against
the well-confirmedness of that hypothesis. But what are the
crucial features that a predicate has to have in order to be like
'grue' in relevant respects?
The most obvious suggestion to make is that the predicates
which are likely to be troublesome are those that imply some
spatial or temporal restriction or a reference to some particular
individual. Such predicates occur typically in statements of what
is accidentally true, while hypotheses that state natural laws
characteristically concern all ravens, say, or all emeralds, or all
pieces of copper. If, as we commonly suppose, Nature is
uniform throughout space and time, a description of its laws
need make no reference to particular places or times. Now to
predicate 'grue' of emeralds, it may be said, is not to say the
same thing about all emeralds, but to say one thing about all
those examined before a particular point in time and another
thing about the remainder. So 'grue' does not belong in the
terminology of confirmable hypotheses about natural laws.
But there is no guarantee that we could not replace 'grue' by
an equivalent disjunction of complex descriptions, built up from
§23Four Paradoxes about Induction 199
purely qualitative terms that imply nothing about particular
times or places. Each group of emeralds examined before t
would then consist of the only emeralds to satisfy such-or-such a
description (about size, shape, dents, scratches, flaws, etc.) and
the disjunction of those descriptions would pick out all and only
the emeralds examined before t. So the paradox could then be
stated without the introduction of any bizarre terms like 'grue'
into our vocabulary.
Goodman's own defence of the paradox here is rather differ-
ent. He argues that the concept of what is to count as a purely
qualitative term is unclear. The temptation is just to take it for
granted that 'green' and 'blue' are purely qualitative because
they are basic colour-terms and involve no reference to a parti-
cular time or place, while 'grue' is not purely qualitative because
it does involve such a reference: similarly 'bleen' is not purely
qualitative if it applies to emeralds examined before time t just in
case they are blue and to other emeralds just in case they are
green. But, as Goodman points out, if we treat 'grue' and
'bleen' as basic and define 'green' and 'blue' appropriately, we
shall find that 'green' and 'blue' come to involve a reference to a
particular time or place. For example 'green' will apply to
emeralds examined before time t just in case they are grue and to
other emeralds just in case they are bleen. And similar redefini-
tions are available for any other supposedly qualitative terms.
Perhaps it will be suggested instead that comparative simpli-
city, not qualitativeness, is the distinguishing mark of the pre-
dicate that will be acceptable in such a context. An analogous
problem arises, it may be said, where the values of some
quantitative physical variable have been plotted on a graph,
against certain values of some apparently correlated variable
(such as the velocity of a falling body against the time for which it
has been falling). Any number of different curves can be drawn
through the whole set of plotted positions, and no finite number
of additional plottable positions will determine one curve
uniquely. Which of these curves is to be preferred? Presumably
the one described by the simplest mathematical equation. So is
not comparative simplicity also the reason why, in the situation
described, we prefer 'All emeralds are green' to 'All emeralds
are grue'? Wherever our choice of hypothesis is under-
determined by the evidential data, it may be said, we have to
§ 23 §23Four Paradoxes about Induction 200
supplement our criterion of confirmation by an appropriate
principle of terminological selectivity.
No doubt some accepted set of criteria for comparative simpli-
city have to be invoked when fitting a description to a set of data,
especially when those data are quantitative. Fewer arithmetical
operations, lower numerical exponents, shorter symbolic
expressions, etc. are all prima facie desirable features in any
mathematical formula involved. But any such set of criteria
have to be considered as relative to an assumed list of primitive,
undefined symbols. O u r ultimate comparison of simplicity must
be between expressions in this primitive, undefined form.
Otherwise it is all too easy to invent an appropriate definition
that will—quite trivially and superficially—reduce a highly
complex expression to a much simpler form. So Goodman could
make the point that, if one starts with the 'grue' and 'bleen'
terminology as basic, 'green' and 'blue' must be regarded as
relatively more complex expressions, and therefore the paradox
cannot be resolved by an appeal to the greater simplicity of the
familiar colour-vocabulary when compared with the 'grue' and
'bleen' one. What apparently needs to be shown instead is that
use of the latter vocabulary inherently disadvantages a
hypothesis.
Goodman tries to show this via an account of what he calls
'entrenchment' and 'projectibility'. He begins by defining a
hypothesis as being 'projected' at a particular time if it is
adopted at that time after some of its instances have been
examined and found to be true and before the rest of its instances
have been examined. A hypothesis may be said to be 'project-
ible' if and only if it is validly projected. So a hypothesis that is
falsified thereby becomes unprojectible. But not all hypotheses
that avoid falsification are thereby rendered projectible. In
order to introduce a further criterion of projectibility Goodman
defines the concept of 'entrenchment'. Entrenchment is what
accrues to a predicate in virtue of its occurrence, or of the occur-
rence of any predicate co-extensive with it, within a projected
hypothesis. So some predicates become better entrenched than
others. For example 'green' is much better entrenched than
'grue', because it has occurred much more often than 'grue' in
projected hypotheses. Then relative entrenchment is a further
criterion for separating the projectible from the unprojectible.
§23Four Paradoxes about Induction 201
That is, if two conflicting hypotheses have the same number of
evidential instances in their favour, and none against them, and
there is just one difference in entrenchment between a predicate
in the one and a predicate in the other, then the hypothesis with
the better entrenched predicate is projectible and the other is
not. So 'All emeralds are green' would be projectible and 'All
emeralds are grue' would not.
Goodman insists that entrenchment is not the same as famili-
arity. 'Any wholesale elimination of unfamiliar predicates', he
says, 'would result in an intolerable stultification of language.
New and useful predicates like "conducts electricity" and "is
radioactive" are always being introduced and must not be
excluded simply because of their novelty.' And he points out
that his criteria of projectibility legislate against novel or unfa-
miliar predicates 'only to the extent of eliminating projections of
them that conflict with projections of much better predicates'.
But is not that bad enough? If scientists followed Goodman's
prescription they could easily come to eliminate one of two
conflicting hypotheses because of its poorly entrenched pre-
dicate, even though, when more evidence became available, it
turned out to be better supported. A more prudent policy might
often be not to eliminate in advance the hypothesis with the
novel or unfamiliar predicate, but to retain it for further
consideration at a later date when more evidence becomes avail-
able. Indeed, this was how the pathogenic role of vitamin
deficiencies came eventually to be understood, even where, as in
the case of beri-beri, the presence of toxic micro-organisms was
for long the favoured aetiology. In sum, Goodman's proposal
for solving the paradox makes far too drastic an inroad on our
terminological open-mindedness, and forces science into an
excessively conservative posture. Nor does it take account of the
gradualness with which fundamentally new ideas tend to spread
within the scientific community.
It is not possible to avoid this difficulty by allowing predicates
to inherit entrenchment from what Goodman calls 'parent'
predicates. Goodman calls a predicate P a parent of a given
predicate Q if P applies only to certain mutually disjoint classes
and one of these classes is the extension of Q. For example, he
says, the predicate 'army division' is a parent of the predicate
'soldier in the 26th division'. Similarly, no doubt, 'class of
§ 23 §23Four Paradoxes about Induction 202
mutually similar sub-atomic particles' would be a parent of the
predicate 'proton'. But there is no reason to suppose that in the
relevant period of conceptual innovation the predicate 'class of
mutually similar sub-atomic particles' was any better
entrenched than the predicate 'proton'. So the latter would not
inherit any entrenchments from the former.
Accordingly we must seek to resolve the paradox in some
other way than by excluding terminology of the trouble-making
kind. Obviously we have to invoke background knowledge in
some way, in order to eliminate the paradoxical hypotheses. But
it looks as though the background knowledge that we rely on
must be factual rather than terminological. And there are
perhaps two main ways of trying to put this knowledge to work
in order to achieve the desired result. One uses Bayes's law, the
other the Method of Relevant Variables.
If we employ a Pascalian measure for degree of confirmation
we must presumably infer that the prior probability of 'All
emeralds are grue' is much less than that of 'All emeralds are
green', since in the past hypotheses about objects' changing
their colour if not already examined have all proved false. Then
by Bayes's law (p. 23) the hypothesis 'All emeralds are green'
must also be much rpore probable on the evidence than 'All
emeralds are grue' is, since in both cases the evidence is the same
and so is the probability of the evidence given the hypothesis. At
first sight this type of solution is admittedly open to the objection
that it begs the question at issue. It assumes that there is induct-
ive evidence in favour of the hypothesis that grue-type hypo-
theses are always subsequently falsified. And such evidence,
Goodman might say, is just as much in favour of the hypothesis
that grue-type hypotheses are always falsivered, where
'falsivered' applies to all hypotheses propounded before / just in
case they have been falsified and to others just in case they are
verified. So there is apparently a case for saying that background
knowledge allows just as much prior probability to be assigned
to 'All emeralds are grue' as to 'All emeralds are green'. How-
ever if our past experience is such as to assign only a low prior
probability to the hypothesis about falsivering, Goodman's
potential objection would lapse.
The Method of Relevant Variables can also exploit back-
ground knowledge in order to achieve a solution of the paradox.
§23Four Paradoxes about Induction 203
It must treat grue-type hypotheses as being subject to an appro-
priate series of canonical tests, and presumably the first of these
tests must manipulate the before-//after-< variable, whatever
time 't' denotes. After all, our past experience must be that
'grue-type' hypotheses are generally falsified by the manipula-
tion of this variable: otherwise we should not find it so para-
doxical to have an apparently well-confirmed prediction that
emeralds are going to change their colour forthwith. Con-
sequently, if all confirmatory evidence has to come from the
results of canonical tests, there will never be such evidence, for
any value of 't', unless the hypothesis under examination does
indeed describe, for a specified value o f t ' , a genuine turning-
point in Nature. Moreover, the argument that was used to
answer a possible objection to a Bayesian resolution of the
problem has an analogue in the Method of Relevant Variables.
If we consider the hypothesis about tests and relevant variables
that is assumed by the present treatment of the problem, we can
see that, in any contest between it and a 'falsivere'-type hypo-
thesis, it has as good a prospect of winning as common sense
would think it to deserve. It too would just need to be tested over
the crucial before-</after-/ variable.
In short, from a Baconian point of view Goodman's paradox,
like Hempel's, may be seen as arising from the use of too coarse
and global a concept of confirmation. Both paradoxes arise
because this concept is neither graduated nor variative and is
therefore insufficiently sensitive to background knowledge
about locally relevant factors. Once we replace this concept by
one that corresponds more closely to the actual practice of
natural science the paradoxes can no longer be constructed. 27

§ 26. T H E L O T T E R Y P A R A D O X

Kyburg's lottery paradox discloses an apparent conflict between


deductive cogency and any Pascalian criterion for the acceptance
2 7Many other discussions of Goodman's paradox are to be found, such as S.
Blackburn, Reason and Prediction (Cambridge: Cambridge University Press, 1973),
61-96; H. Fain, 'The Very Thought of Grue', Philosophical Review, 76 (1967), 61-73;
and M. Hesse, 'Ramifications of " G r u e " ', British Journal for the Philosophy of Science 20
(1969), 13-25 (which includes a useful bibliography).
§ 23 §23Four Paradoxes about Induction 204
of a hypothesis. Levi proposes to resolve the paradox by insisting
that acceptance must always be relative to some ultimate parti-
tion of the hypotheses under consideration. But the paradox also
disappears if we base acceptance on a Baconian criterion.
Twentieth-century philosophers have used quite a variety of
terms in their references to inductive reasoning as the topic of
their analyses. They have spoken about a relationship of
'confirmation', 'justification', 'validation', or 'support', for
example—with or without assistance from the adjective
'evidential'. Such terms have many differences in everyday
nuance or implication that are irrelevant for present philo-
sophical purposes. For example, as already remarked (§ 2),
confirming items normally occur later than confirmed ones: a
hotel reservation has to be made before it can be confirmed. O n
the other hand, supporting items are not normally either later or
earlier than supported ones: the relationship is spatial rather
than temporal. But this difference in ordinary language is of no
importance in modern philosophical discussions of inductive
reasoning, which treat temporal priority and spatial subjacence
as equivalent metaphors for justification. The analysandum is
the same, whether or not the proposed analyses are different.
What is ultimately under discussion in each case is the nature of
a certain epistemic relationship that can exist between one
proposition and another—a relationship that, since the early
seventeenth century, most philosophers have regarded as a
matter of degree.
Nevertheless there are several purposes for which this rela-
tionist stance is inadequate, and what may conveniently be
called 'the need for inductive detachment' arises. We want to be
able not only to assert that E gives such-or-such a level of induct-
ive support to / / b u t also, in appropriate cases, because E is true,
to detach as a conclusion the truth of H. Certainly engineers,
physicians, navigators, law-courts, and other decision-taking
agents often need in practice to detach conclusions from pre-
misses which state both what the evidence is and how well that
evidence supports such-or-such a conclusion. So do the writers
of scientific text-books, though Carnap was inclined to deny that
there was ever any need for it. Indeed most scientists feel the
28

2 8 R. Carnap, 'Inductive Logic and Rational Decisions', in R. Carnap and R . C .


Jeffrey (eds.), Studies in Inductive Logic and Probability (Berkeley: University of California
§23Four Paradoxes about Induction 205
need at some stage to translate the relational inductive grada-
tions of their more cautious colleagues into the common-sense
categorical dichotomies of knowledge versus ignorance, or proof
versus disproof (see p. 154). This happens particularly when
some previous discovery, in the role of 'background know-
ledge', constitutes an indispensable premiss for the evaluation of
a new hypothesis, as when any measuring instrument is used in
a laboratory experiment. But it also happens when the effects of
a particular medical drug have to be predicted before it is pre-
scribed, or a causal law has to be derived before it can be
exploited, or a theory has to be established before it can be cited
as the explanation of anything, or juries have to bring in an
unconditional verdict of 'Guilty' before the accused can be
sentenced.
So it is natural to ask: what level of support for a proposition,
in the light of available evidence, justifies belief in its truth or
acceptance of it as being true? Obviously this would be justified
by a Pascalian probability of 1 or by maximal legisimilitude in
accordance with the Method of Relevant Variables. But in the
typical inductive situation we never have a probability of 1 and
we are never entitled to be sure that our list of relevant variables
is complete. So in practice we are often content with a lower
threshold, which we treat as the level at which, for the purpose in
hand (with whatever utilities it involves), further doubt or
hesitation would be unreasonable and an appropriate conclu-
sion may be detached. But what should this threshold be? Any
attempt to fix it in terms of a stipulated level of Pascalian
probability seems at first sight to be exposed to the derivation of
a paradox which is often called 'the paradox of the Lottery' and
was first enunciated under that name by Kyburg. 29

The core of the paradox may be formulated as follows.


Suppose three conditions:

Press, 1971), 29-30. See also Y. Bar-Hillel's comment in H.E. Kyburg and E. Nagel
(eds.), Induction: Some Current Issues (Middletown: Wesleyan University Press, 1963), 46.
2 9 H. Kyburg, jun., 'Probability, rationality and a rule of detachment', in Y. Bar-
Hillel(ed.), Proceedings of the 1964 Congressfor Logic, Methodology and the Philosophy of Science
(Amsterdam: North-Holland, 1965), 301-10. The first statement of the paradox is to be
found in H.E. Kyburg, jun., Probability and the Logic of Rational Belief (Middletown:
Wesleyan University Press, 1961), 197. A similar paradox is discussed by C.G.
Hempel, 'Deductive-Nomological vs. Statistical Explanation', in H. Feigl and G.
Maxwell (eds.), Minnesota Studies in the Philosophy of Science (1962), iii. 144-7.
§ 23 §23Four Paradoxes about Induction 206
(1) If E states all the available evidence, and the Pascalian
probability of H on E is within some suitably small
interval from 1, it is rational to believe H (i.e. justifiable
to accept H).
(2) If it is rational to believe H rational to believe H , . . .
u 2

and rational to believe H , then it is rational to believe any


n

logical consequence of H H . . . and H .


iy 2 n

(3) It is not rational to believe an inconsistent proposition.


These three conditions all have a certain intuitive appeal. But
they are demonstrably not co-tenable. Consider a lottery with a
million (or more) tickets, that is assumed to be administered
fairly. There is a very high probability on this evidence that
ticket no. 1 will not win, and that ticket no. 2 will not win, . . .
and that ticket no. n will not win. So, according to (1), it is
rational to believe each of these propositions. Then, according
to (2), in the light of the information that there are just n tickets it
is rational to believe that no ticket will win. But also, according
to (1), it is rational to believe that some ticket will win, since the
lottery is assumed to be administered fairly. So, according to (2),
it is rational to believe the conjunctive proposition that no ticket
will win and some ticket will win. But that conjunctive proposi-
tion is inconsistent and, according to (3), it is not rational to
believe an inconsistent proposition. Accordingly, we have
demonstrated from (1), (2), and (3) both that it is, and that it is
not, rational to believe the proposition that no ticket will win and
some ticket will win. At least one of the three conditions must
therefore be rejected.
Condition (3) is hard to reject. The intuitive connection
between inconsistency and irrationality is in any case a strong
one. And there is nothing much to be gained by objecting that it
might be rational for a non-logician to believe a complex, non-
obviously inconsistent proposition that was asserted by a compe-
tent authority. For it is possible to replace (3) by
(3') It is not rational to believe an obviously inconsistent
proposition.
and the paradox will still arise.
So those who, to meet the need for inductive detachment,
wish to set up a Pascalian criterion, like (1), for the rationality of
inductively based belief are naturally inclined to interfere with
§23Four Paradoxes about Induction 207
(2). The idea that underlies (2) is often referred to as 'deductive
cogency' or 'deductive closure'. It is the idea that rationality
requires a commitment to the logical consequences of one's
beliefs. But is this commitment to be understood as applying to
one's beliefs collectively or distributively? The collective
application, as in (2), helps to generate the paradox. But, if (2) is
replaced by a distributive version, as in
(2') If it is rationed to believe H, then it is rational to believe
any logical consequence of H,
the paradox cannot be constructed. And it is along these lines
that Kyburg himself proposed to solve the problem. After all, if
it is rational to believe that one may sometimes err, then even if
one believes all and only the several propositions H H , . . . U 2

and H it seems quite rational not to believe at the same time


N

either that the conjunctive propositions H^SiH^&i . . . &H is n

true or that it is false.


But here some rather subtle differences between rationality of
belief and justifiability of acceptance are in play. We need to 30

check our intuitions rather carefully. It may well be rational to


believe—in the everyday sense of that expression—that one may
sometimes err. But if it is justifiable to accept H justifiable to u

accept H , . . . and justifiable to accept H , it is certainly justi-


2 N

fiable to accept both the conjunction H,&H &L . . . &CH and also
2 N

any logical consequence of that conjunction. Indeed, unless we


could assume the truth of (2) in this sense, we should be intoler-
ably hampered in compounding the premisses for our infer-
ences. We should never be entitled, as the phrase goes, 'to put
two and two together'.
Levi, however, has an argument against (2) even when this
31

is its intended meaning. He claims that acceptance is always to


be thought of as being relative to some ultimate partition of the
hypotheses that are candidates for acceptance, where the
evidence (including background evidence) entails that at least,

3 0For further discussion of the difference between belief and acceptance see
L. Jonathan Cohen, The Dialogue of Reason (Oxford: Clarendon Press, 1986), 92-7.
3 1I. Levi, Gambling with Truth: An Essay on Induction and the Aims of Science (New York:
Knopf, 1967), 38-42. See also I. Levi, 'Deductive Cogency in Inductive Inference',
Journal of Philosophy, 57 (1965), 68-77, repr. inl. Levi, Decisions and Revisions: Philosophical
Essays on Knowledge and Value (Cambridge: Cambridge University Press, 1984), 42-50.
§ 23 §23Four Paradoxes about Induction 208
and at most, one such hypothesis is true. (Compare Scheffler's
approach to the paradox of the ravens in § 24.) For example,
acceptance of the hypothesis that ticket no. 1 will not win the
lottery is relative to a partition into the two possibilities 'Ticket
no. 1 will win' and 'Ticket no. 1 will not win'. And, analog-
ously, acceptance of the hypothesis that ticket no. 2 will not win
is relative to a different ultimate partition, according to which
the hypothesis that ticket no. 1 will not win is not even a candid-
ate for consideration. So, if it is therefore required that the cons-
traints of deductive cogency apply only in a way that is relative
to the same ultimate partition, we must put this requirement in
place of (2). We thus have no warrant, according to Levi, for
inferring that all the tickets in the imaginary lottery will fail to
win. And without such a warrant the paradox cannot get off the
ground. Of course, in order to resolve the paradox in this way
Levi has not only to reformulate the constraints of deductive
cogency but also to modify (1). O n his view a high probability
for H, on E, does not provide sufficient grounds for accepting H
outright, even if E is all the available evidence. It may provide
grounds only for accepting H as against other hypotheses in the
ultimate partition to which H belongs. Nor is a high probability
necessary for H, if there are to be grounds for so accepting that
hypothesis. If each rival hypothesis has a lower probability than
H, H might still deserve victory over them in the struggle for
acceptance, on Levi's view, even if the probability of H itself
were in fact quite low.
Levi's thesis that a high probability is not a necessary
condition for the acceptance of H, as against its rivals, has some
plausibility in the context of inductive reasoning. If H and its
rivals are generalizations that make conflicting predictions over
an indeterminately large domain, we might expect that even the
winning hypothesis has a low probability on the evidence
because of the very low prior probability that is to be attached to
any generalization of this kind. So, if any such hypothesis is to be
accepted at all in science when its rivals have been elimin-
ated—and many are in fact so accepted—it is no use insisting
that a high probability is necessary, as in (1). But in order to
resolve the lottery paradox this point does not need to be
pressed. All Levi needs for the resolution of that paradox is his
insistence on making acceptance relative to an ultimate parti-
§23Four Paradoxes about Induction 209
tion. Or, at any rate, that is the situation if the criterion for
inductive detachment is to be stated in Pascalian terms.
But there are some further points to clarify here. On Levi's
view there can be no authority for a clear-cut, unconditional,
non-relative acceptance or rejection that is parallel to belief or
disbelief. Yet the paradox originates because of the felt need for
a criterion to control detachments that are indeed of just that
kind. Levi's resolution of the paradox has the disadvantage of
presupposing that no such criterion is ever possible.
In practice the situation varies according to the nature of the
issue and the evidence. Where there is apparently nothing to
choose between several different ultimate partitions, there is no
reason to regard the hypothesis that is more acceptable than any
other within one particular partition as somehow superior to any
hypothesis that is more acceptable than the others within any
other partition. So, where more than one ultimate partition
seems equally respectable, no single hypothesis can be the over-
all winner and therefore declared to be the one that it is justi-
fiable to accept tout court. For example, the hypothesis that in the
above-mentioned lottery ticket no. 1 will not win is no more
acceptable, in an absolute sense, than the hypothesis that ticket
no. 2 will not win or than any other such hypothesis. But the
situation is often quite different, especially in the natural
sciences. Often just one set of rival hypotheses may be
considered to contain all the serious candidates for acceptance,
because of what we know already about the subject. And the best
of these hypotheses, if the evidence is strong enough, may then
be regarded as acceptable in an absolute rather than just a
comparative sense. It is better than any rival within the only ulti-
mate partition that is itself acceptable on the available evidence.
Thus Galileo, in his famous Dialogue Concerning the Two Chief
World Systems, discussed only whether the Ptolemaic or the
Copernican theory was preferable, and did not entertain also
some quite different partition of the fundamental possibilities.
Nevertheless, the lottery paradox cannot be reproduced in this
kind of situation, so it does not constitute an objection to Levi's
resolution of that paradox. Levi can still maintain his claim that
acceptance is always relative to an ultimate partition, even if in
many cases only one ultimate partition needs to be considered.
Indeed, the lottery situation is unrepresentative of acceptance
§ 23 §23Four Paradoxes about Induction 210
issues in more than one way. Not only does it allow the equal
legitimacy of a large number of different ultimate partitions,
with the consequences that have just been under examination. It
is also insulated from any practical question about whether all
the relevant facts are included in the available evidence. Since
the lottery is explicitly assumed to be administered fairly we can
treat it as a perfect game of chance and calculate the probabilities
of the various outcomes according to an indifference analysis
(§ 6). But where we are not dealing with aleatory probabilities
questions about the extent or spread of relevant evidence may be
highly pertinent. If the issue is, for example, whether it will rain
tomorrow or not, then what Keynes called the weight (§ 14) of
the available evidence is just as important as the probability on
that evidence. It is justifiable to accept that it will rain only if the
weight and the probability are both sufficiently high. So in this
respect any Pascalian criterion for detachment needs some
appropriate supplementation if it is to cover an important
feature of many non-aleatory situations (as Levi recognizes: see
p. 174).
Finally, it is worth noting that the lottery paradox arises only
if the criterion for inductive detachment is stated in Pascalian
terms. Suppose that we grade the evidence's support by the
Method of Relevant Variables or in accordance with some
derived scale of Baconian probability (pp. 167-8), and that,
instead of adopting (1), we select some suitably high Baconian
grade as the appropriate basis for acceptance. Then all we can at
best say in these terms about ticket no. 1, or about any of the
other tickets, is that the evidence is sufficient for us to accept that
its chances of not winning are 999,999 in 1,000,000. That is to
say, each ultimate hypothesis under consideration, and open to
acceptance, is a proposition about the aleatory probability of
winning, or of not winning, not a hypothesis about winning, or
about not winning. And in this context (2) is quite innocuous:
the hypotheses that we are justified in accepting have no para-
doxical consequences, even when considered collectively.
Moreover, this solution fits well with the fact that in normal
circumstances it would seem queer for someone to buy a ticket to
a lottery which he feels justified in accepting that he will not win.
Since many people do buy lottery tickets, many people presum-
ably do not feel justified in accepting that they will not win and
§23Four Paradoxes about Induction 211
would find the relevant implications of (1) counter-intuitive.
Correspondingly the Method of Relevant Variables resolves the
prima-facie conflict between the need for inductive detachment
and the idea of deductive cogency—the conflict that underlies
the lottery paradox—by retaining (2) but reformulating (1) in
Baconian terms. It could well be argued, however, that, like
Levi's analysis, the Method of Relevant Variables still makes
inductively based acceptance relative to an ultimate partition of
hypotheses, since any eliminativist analysis of induction pre-
supposes such partitions (see § § 2, 18, and 20).
So, just as (see § § 23-5) the classical problem of induction,
the paradox of the ravens, and the 'grue' paradox all admit of
resolution both in terms of a Pascalian gradation for ampliative
induction and in terms of a Baconian one, so too the lottery
paradox can be dissolved both in Pascalian terms and in
Baconian ones. Indeed, even on the general issue of Pascalian-
32

ism versus Baconianism, it may be best to refrain from suppos-


ing that this is an issue where in the end one side has to be
supposed correct and the other incorrect. What we have are
alternative kinds of proposal for the systematic reconstruction of
inductive reasoning, each of which has both advantages and dis-
advantages (see also § § 16-22). In such philosophical dilemmas
it is normally worth while to carry forward both options.
The underlying problem about detachment in inductive reasoning is discussed by
3 2

H.E. Kyburgjun., Y. Bar-Hillel, P. Suppes, K. Popper, W.C. Salmon, J. Hintikka


and R. Carnap, in I. Lakatos (ed.), The Problem of Inductive Logic (Amsterdam: North-
Holland, 1968), 98-165.
INDEX

Adams, E.W. 20 Cajori, F. 142


additivity, see complementational Cardano, G. 15, 26
principle for negation Carnap, R. 2-3, 74-80, 84-8, 96, 98,
Adler, J.E. 163 100-1, 120, 122-30, 133, 137-8, 155,
agreement, method of 29-30, 32, 34-5, 182, 204, 211
153 Carneades 26
aleatory model 18-20,56 Cartwright, N. 170
ampliative induction 1-3, 27, 39, 113, causal tendency 28
115, 117, 129, 132, 138, 140-3, 157, causation, law of 31-2, 155
181 chance, game of 14-15, 28, 48-9, 56,
analogy, probability from 28, 36, 135 79, 82, 91-2, 93, 94, 98, 106, 116-17,
anomalies, falsification by 142,149, 130
182, 184 Clairaut, A.C. 142
Anschutz, R.P. 37 .Clark, P. 57
Aristotle 7, 10, 82 Cohen, L.J. 5, 20, 38, 69, 109, 143,
Arnauld, A. 16 145, 151, 156, 159, 161-3, 166, 171-3
axiomatic definition of probability 38, Cohen, M . R . 1
40 coherence 60, 69, 87, 105
axiomatization 24-5, 38 commitment, partial 64-5
Ayer, A.J. 117,185 complementational principle for
negation 17-18, 38, 56-7, 95, 112,
background knowledge 155, 195-6, 202, 157, 168
209 completeness 112
Bacon, F. 2-12, 14, 27, 29, 35-7, 132, comprehensiveness of induction 6
151, 157, 177 concomitant variations, method of 33
Bar-Hillel, Y. 204, 211 conditional probability 20, 58, 61, 72,
Barcan, R.C. 162 77, 89, 100
Barker, S.F. 164 conditional, probability of 20, 72, 161
Bayes, T. 23 conditionalization 68, 105
Bayesianism 3, 23, 42, 60, 67, 69-70, conditionalized probability 100
81, 105, 137, 203 confidence, coefficient of 99, 178
Bayes's law 23, 68, 136, 192, 202 confidence-interval 67, 118
belief, degree of 59, 62, 64-5, 71,81, 'confirmation' 13
84, 93, 111 conjunction, multiplicative principle
Bernoulli, J . 18, 21-2, 27, 38, 69, 173 for 18-20, 38, 54, 56, 95, 105, 157,
bet, conditional 60 162, 168
betting odds, 59, 65-6, 68, 87, 107, 138 conjunction, non-multiplicative
Black, M. 196 principle for 19-20, 73, 111, 152,
Blackburn, S. 47, 156, 203 162, 168, 174
Boltzmann, L. 46 consilience 9, 11, 151, 184
Bolzano, B. 37-8 constituent 123, 127, 129, 137
Boyle, R. 5, 169 contingency 91-4, 111
Braithwaite, R. 53 contraposability 20-1, 119, 168
Britton, K. 37 controls 30, 148
Buchdahl, G. 12 convergence of probability
Burks, A.W. 130, 156 judgements 67, 69-70
Butterfield,J. 57 converse entailment condition 139-40
214 Index
Cooke, N . M . 11 Freudenthal, H. 42
counterfactualizability 84, 88-94, Frisch, K. von 145
111-12
covering law model 170, 173-4 Gaifman, H. 196
Cox, R . T . 113 Galen 7
credibility 17-20, 22, 57, 82, 95, 99, Galileo 14-15, 26, 48
110, 167 Garber, D. 26
Cresswell, M.J. 162 generality, explicit 88
generality, implicit 87-8, 91, 93-4, 99
decision theory 106 genetic epistemology 12
deductive closure 206-8 Gillies, D. 53, 70
Descartes, R. 7 Glymour, C. 144
descriptive-prescriptive polarity 41, 114 Goodman, N. 197-203
detachment, inductive 204-5, 210 gradualism 7-9
determinism 54 'grue' paradox 197-203
difference, method of 30, 32, 34, 153 guarded assertion 41, 64-5
Dorling, J . 144
Dutch book 60, 67, 87, 105 Hacking, I. 15, 26, 48, 56, 70, 76, 80
dyadic probability-functions 20-1, 44, Hanson, N. 163
48, 54, 58, 61-2, 66, 68, 72, 74, 87-9 Hare, R . M . 64
Harman, G. 173
Edwards, P. 185 Harre, R. 144, 179
Eells, E. 70, 141 Hempel, C. 4, 101, 188-96, 205
Eggleston, Sir R. 168 Herschel, Sir J . F . W . 8-9, 29, 33
Einstein, A. 142, 154 Hesse, M. 12, 38, 138, 156, 203
eliminative induction 2, 7, 31, 37-8, Hilpinen, R, 156
149, 184, 190 Hintikka, J. 80, 123-7, 137-8, 155, 211
Ellis, R.L. 5-6 Hooke, R. 5
empiricism 176-7 Horwich, P. 65, 144
entrenchment 200-2 Hosiasson-Lindenbaum, J . 193
enumerative induction 2, 6, 29, 31-2, Hughes, G.E. 162
37, 97, 116-18, 120, 126-7, 131, 134, Humburg, J . 121
137, 143, 146, 178-9 Hume, D. 3, 28-9, 64, 96, 177-81,
ethics, induction in 178 185-8
Euler, L. 17 Humphreys, P. 57
exchangeability 69-70, 78, 88, 137 Huygens, C. 15, 26-7
expectation, mathematical 107 hypothetico-deductive method 12
expert systems 11 'hypothesis' 13
explanatory power 123, 151, 173-4
explicit generality 88 idealism 42, 44, 59, 62, 65, 71, 80, 92,
extensionality 87, 91, 93-4, 112 111, 174
idealization 59, 153, 169-71
Fain, H. 203 implicit generality 87-8, 91-4, 99
falsificatory efficacy 148-9, 184 implicit singularity 89, 91, 99-100, 103,
family resemblance 83 105
Feller, W. 42 incompleteness 112
Fermat, P. 2, 15, 26-7 indeterminism 54-5
Fine, T.L. 42 indifference analysis 43-7, 55-6, 81, 84,
Finetti, B. de 60-3, 67, 69-70, 78, 81, 88, 92, 106, 115-17, 129
87-8, 137 indifference, principle of 42, 44, 47, 78,
finite reference-class 48, 52, 54-5, 90 80
Fisch, M. 12, 19, 196 individual outcome, probability of 48-9
Fisher, R.A. 23, 67, 119 53, 55
forensic proof 21, 168 Infeld, L. 154
Index 215
infinite reference-class 49-53, 55, 89 Maistrow, L.E. 26
instance confirmation 122, 138 Margalit, A. 156
instantial relevance 121, 126 mathematics, induction in 172-3
interval-valued function 26, 54, 90, 134 maximal specificity, requirement of 101
inverse principle 193-4 Mehlberg, J.J. 42
Mellor, D.H. 53
Jardine, L. 12 Mere, Chevalier de 14
Jeffrey, R.C. 26, 70 Mill, J.S. 3, 10-11, 14, 29-38, 44-5,
Jeffreys, H. 80 83, 132-3, 135, 153, 155-7, 159
Jensen, O.C. 172 Miller, D. 141,183
Jevons, W.S. 44-5 Milton, J . R . 177
Johnson, W.E. 1 Mises, R. von 48, 51
Joseph, H.W.B. 1 modal logic 160-4, 167, 175, 186
jurisprudence, induction in 2, 8, 172, modality 83
178 Moivre, A. de 18
Kahneman, D. 69 monadic probability-functions 20-1,
Kant, I. 178-9 44-5, 55, 58, 60, 62, 66, 68, 72, 77,
Kemeny, J . G . 130 87, 103, 106
Kepler, J. 10,123 Montaigne 26
Keynes, J . M . 3, 26, 38, 47, 74-5, 95, multi-valued logic 71-4, 81, 84-5, 111
97, 101-5, 132-8, 154-7, 159, 164-5, multiplicative principle for
172 conjunction 18-19, 38, 54, 56, 95,
105, 157, 162, 168
Kitchener, R.F. 12
Kneale, W. 1, 21, 43 Nagel, E. 1, 42, 82, 130
Kocher, P H. 8 necessity 91-2, 112, 161
Kolmogorov, A. 24-5 negation, complementational principle
Kuipers, T.A.F. 130 for 17-18, 38, 56-7, 95, 112, 157,
Kyburg, H.E. 1, 57, 69, 203, 205, 207, 168
211 negation, non-complementational
principle for 17-18, 57, 65, 104, 110,
Lakatos, I. 130, 142, 151, 173, 184 113, 167
Lambert, J . H . 18 Newton, I. 9, 18, 123, 142-51
Langford, C.H. 162 Neyman, J. 23, 67, 118
language, choice of 125 Niiniluoto, I. 98, 124, 130, 183
Laplace, P.S. de 44, 50, 97, 172 Nisbett, R. 69
legisimilitude 160-71, 174, 183-4, 186, nominalism 76
205 non-additivity see non-complementa-
Lehrer, K. 70 tional principle for negation
Leibniz, G.W. 23, 27, 76, 151 non-complementational principle for
Levi, E.H. 172 negation 17-18, 57, 65, 104, 110,
Levi, I. 104, 156, 174, 207-9, 211 113, 167
Lewis, C.I. 162 non-extensionality 91-2, 94, 112
limit of infinite sequence 50 non-multiplicative principle for conjunc-
localization strategy 156,203 tion 19-20, 73, 111, 162, 168, 174
Locke, J . 176-7 non-truth-functionality 73, 85
logical relation analysis 74-80, 82, 91, normal circumstances 148
102, 105 novelty of prediction 136, 152, 184
logical syntax of probability 25, 38, 109, Nowak, L. 169
157-68
lottery paradox 203-11 O'Connor, D.J. 12
Lucas, J . R . 21, 71, 113 objective idealism 44, 65, 71, 92
Oddie, G. 183
Mackie, J . L . 38, 42, 82-3, 156, 193-4
Madden, E.H. 179 Pap, A. 46, 53
216 Index
paradoxes, inductive 3-4, 177-211 relevant variables, method of 147-63,
partition, ultimate 207-11 165, 169-75, 183, 186, 190, 194-6,
Pascal, B. 2, 14, 16, 26-7, 157, 160 202, 205, 210
Pearson, E. 23, 67 replicability of test-results 127, 131,
Peirce, C.S. 56, 183 156, 181-2
personalism 42, 58-72, 84, 86-8, 93, Rescher, N. 73, 111, 185
105-6, 136-8 research programme 184
Pirie, N.W. 154 residues, method of 32-3
pluralism 82-3, 113 Rodriguez, A.R. 141
point-valued function 54, 90 Rosenkrantz, R.D. 144, 184
Poisson, S.D. 50 Ross, L. 69
Polya, G. 172-3 Russell, B. 47, 50
Popper, Sir K. 4, 25, 47, 54-6, 141,
144, 181-5, 211 Salmon, W.C. 53, 101, 211
Port-Royal logic 16, 28, 160 sample, representativeness of 96-7
possible worlds, measure for 76-7, sample, size of 23, 96-8, 100, 118
128-30,162-3 Savage. L.J. 61, 67, 70
pragmatics of probability-statements 64 scepticism 179-81, 186
predicate-relatedness 84, 91 Schaffer, S. 8
prerogative instances 6 Scheffler, I. 189, 207
projectibility 200-2 Schlesinger, G. 164
propensity 54, 56, 81, 84, 110-11, 119 Schneider, J. 26
propensity analysis 53-8, 71-3, 84, 86, Searle, J . R . 65
88, 93-4, 99, 106, 118-19 selective confirmation 189
psychology of induction 180 semantics of probability 38, 91
psychology of probability 69, 114 sentence-relatedness 83-4, 91
Putnam, H. 130 Settle, T. 56-7
Sextus Empiricus 26
£>uine, W . V . O . 87 Shackle, G.L.S. 174
Shafer, G. 18,70,113
Ramsey, F. 59-67, 153 Shapiro, B.J. 8
range-measure 77-8, 81, 85, 98, 123-4, Shea, W.R. 179
128, 163 significance-level 23, 67, 119
rationality 3, 59-60, 105, 177, 180-1, simplicity 171, 199-200
185-7, 206-7 singularity, implicit 88, 91, 99-100,
ravens, paradox of the 188-96 103, 105
realism 42-3, 55, 59, 62-3, 71, 86, Sklar, L. 57
110-11, 115, 174 Slovic, P. 69
recursive induction 172 Smokier, H.E. 69
reductio, generalisation of 167 special consequence condition 139-40,
reference-class, finite 48, 52, 54-5, 90 162, 166, 190
reference-class, infinite 49-53, 55, 89 speech-acts 64-5
referential transparency 87, 91, 94 state-description 76-80, 124
Reichenbach, H. 50, 53, 72, 111, 117, 'statistical syllogism' 100
119 Stebbing, L.S. 37
relative frequency 16-17, 35, 47, 49, straight rule 96-7
51, 53, 72, 75, 79-80, 84, 86, 88, 91, Strawson, Sir P.F. 185
96-7, 99, 101, 105, 118 structure-description 78-9, 124, 127,
relative-frequency analysis 47-53, 55-6, 129-30,137
75, 82, 84, 89-90, 98, 101, 117 subjective idealism 44, 62, 65
relevance 58, 62, 108, 146-47, 156, subjectivism see personalism
166, 178, 190 substitutivity 85-7, 111
relevance analysis 139-42 succession, rule of 97-8
Index 217
Suppes, P. 57, 196, 211 variable, hidden 154, 163
Swinburne, R. 1, 80, 142, 196 variables, method of relevant see relevant
symmetrical confirmation-function 77-9, variables, method of
88, 134 variative induction 2, 27, 29, 32, 37-8,
120, 126-7, 131-7, 145-63, 169, 179
test, canonical 148-56, 158, 161, 168 variety, limited independent 36, 132-6,
test, thoroughness of 127, 150, 165 154-5, 179
Thomson, G. 154 Venn, J. 53,97
Tichy, P. 183 verisimilitude 182-4
time, relations in 11, 13, 150, 204
Todhunter, I. 26 Waismann, F. 75, 78
total evidence, requirement of 100, 106 Watkins, J . 185
Toulmin, S. 64, 71, 163 weight 95, 102-11, 120, 164-6, 178
truth-functionality 73, 85 Weinberg, J. R. 12
Tuomela, R. 130 Whewell, W. 9-12, 151, 183
Tversky, A. 69 White, A.R. 56, 73
unconditional probability 58, 77, 103 Wittgenstein, L. 75, 83
uniformity of nature 31, 134, 137, 155, Wright, G.H. von 37, 47, 70, 185,
177-8, 198 196
Urbach, P. 12, 183
utilitarianism 10 Zabell, S. 26
utility 59, 205 Zinov'ev, A.A. 73

You might also like