On the Argument by Analogy

Author(s): P. R. Wilson
Source: Philosophy of Science, Vol. 31, No. 1 (Jan., 1964), pp. 34-39
Published by: The University of Chicago Press on behalf of the Philosophy of Science Association
Stable URL: http://www.jstor.org/stable/186744 .
Accessed: 13/12/2014 09:10

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

The University of Chicago Press and Philosophy of Science Association are collaborating with JSTOR to
digitize, preserve and extend access to Philosophy of Science.

http://www.jstor.org

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
 ON THE ARGUMENT BY ANALOGY*

P. R. WILSON
University of Sydney

Conditions are stated under which the "argument by analogy" is consistent with
the principle of inverse probability. It is contended that the argument by analogy,
in conjunction with a crucial test, has a legitimate place in scientific logic. As an
example the astrophysical problem of solar granulation is discussed in detail and
other examples are mentioned more briefly.

1. Introduction. The notion of the "argumentby analogy" is generallyunaccept-

able to the formal logician. Many writers on the philosophy of science ignore it
altogetheror dismiss it in a few words ([1], p. 262), while those who do discuss the
problem generally create the impression that the argument by analogy is not only
logicallyunsound but almost intellectuallydishonest, since it appealsto the emotions
rather than the mind. Thus Hospers ([3], p. 355) writes "arguments from analogy
although psychologicallycompelling are logically very weak".
This type of argumentis howeverwidely used in many branchesof scientificresearch
and the phrase "by analogy with ..." occurs frequently in discussions and in texts,
particularlyin association with the design of future experiments. Any practising
scientist could quote examples of this but it will suffice here to mention briefly three
well known cases. The analogy between the elementarypropertiesof magnetic poles
and electrostaticcharges led to many speculations concerningthese phenomenaand
ultimatelyto Maxwell'selectromagneticfield theory. Again, by analogywith the solar
system, Bohr suggested his atomic model of a compact centralnucleus surroundedby
orbitalelectrons.A less successfulanalogyis that between the field of a bar magnet and
the earth'smagneticfield.
In view of these and many other examples,it is fair to ask whether analogyhas any
legitimate part to play in the logic of scientific development.If the answer is in the
negative, it appears that much important progress, based on analogy, can only be
attributed to its heuristic or suggestive nature. It is implied that those who allow
themselvesto be influencedby analogyaredoing somethingwhich is ratherunscientific
and, if the analogyis unsuccessful,they deservelittle sympathy.In this paperI attempt
to show that the answer is in the affirmative,and to state precisely the conditions
under which analogymay be used legitimatelyas part of inductive logic.
It is often stated that "analogycan never prove a hypothesis",and in the deductive
sense this is perfectly true. However it is equally true (though less frequently men-
tioned) that no other single experiment or test can prove a hypothesis deductively
(although some may disprove it). In terms of the principle of inverse probability
(as stated by Jeffreys,[4]), any successfulstep in the testing of a particularhypothesis
servesonly to increaseits probabilitytowardssome limit value. Becauseof the inductive
nature of hypotheses, this limit is in practice unattainable,but any observation or
experiment which increases the probabilityof a hypothesis has a valid place in in-
ductive logic. The argumentof this paper is thus aimed at showing that, under care-
fully definedconditions,analogycan be used to increasethe probabilityof one of a set

* Received August, 1961.

34

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
 ON THE ARGUMENT BY ANALOGY 35

of initiallyunlikelyhypotheses.The discussionis illustratedby referenceto the problem

of the interpretationof solar granulationwhich is currentlyof some interest.

2. Inverse probability. The principle of inverse probabilitymay be stated briefly.

If on given data H, the probabilityof a particularhypothesisqis P(q/H), and if Grepre-
sents some additionaldata, the posteriorprobabilityof hypothesisq is given by

P(q10*H) = P(q/H) * P(/Oq*H)

P(O/H)
where P(q/6.H) representsthe probabilityof q given 0 and H.
Jeffreys[4] discusses the problem of "initial probabilities",i.e. the probabilityof a
hypothesis on the minimum possible data. He suggests that "ab initio"the hypotheses
availableto explain a given phenomenon form an enumerableset (e.g. the set of all
differential equations of finite order and degree forms an enumerable set which
certainly contains all the laws of classical physics). In order to assign the initial
probabilitiesof these hypotheseshe suggests the adoptionof the "simplicitypostulate"
which requiresthat the orderof decreasinginitial probabilitiesis the order of increas-
ing complexity.
Kemeny [5] suggests that, although theoreticallysatisfactory,the postulate would
entail considerableunnecessarylabour in the practicaltesting of hypotheses. It is the
legitimate function of the argumentby analogy to increase the probabilityof one of
these hypotheses and thus eliminateunnecessarytesting. This step is not complete in
itself but must be followed up by the design of crucialtests (experimentalor observa-
tional) which may greatly increase the probabilityof this hypothesis if successful or
may lead to its rejection.It is not claimed that analogyprovides a complete argument
but it is suggestedthat, when used in this way, it is a legitimateargument.
It now remainsto demonstratethat the probabilityof a hypothesis may be increased
by analogy.

3. Application to "Argument by Analogy". Jeffreys [4] has shown that if p is

a hypothesis which on data H entails consequencesq1, q2, ..., qn, ..., and if the initial
probabilityof p on data H, P(p/H) is non-zero, then

q2...qn.1. -H) P(p/H)

P(p/ql
P(q1/H) - P(q2/ql*H)... P(q+jl/q1...qn, H)
Thus, since P(p/ql...q.+1 *H) < 1, it follows that
limr P(q.+/lq1 q2...q- -H) 1 (1)

although lim P(p/ql.q2...q.n1.H)= where aomay be less than 1. This result may be
applied to the argumentby analogyin the following way.
A well known phenomenonX is to be comparedwith a less known phenomenon Y,
each of which have n observablepropertiesxn, Yn,which areidentical.It is assumedthat
the propertiesof X and Y which are obviously dissimilar are finite in number or at
least enumerable.Furtherit is assumedthat X has observedpropertyxn+lwhich may
or may not also belong to Y, but which is currently unobservable. This may be due
to the remoteness of Y or to other observational difficulties. Consider the hypothesis p
that, after the obvious dissimilarities between X and Y have been enumerated, the
remaining properties of X and Y are identical. It should be noted that although in

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
36 P. R. WILSON

many cases this hypothesis is unlikely, we require only that on initial data H the
probabilityP(p/H) is non-zero. If qj is the result that propertyxi is observedidentical
to propertyyi, then hypothesisp entails qi for all i. Further, from our assumptions
regardingX and Y it follows that tests ql, q2 ... qnhave been carriedout successfully.
The result qB1l is that the observed propertyx1+ is identical with the unobservable
propertyYn+l, and from (1)
lim P(q,1+jq1.q2...q,.H)= 1.

Thus it follows that as the number of observed identical properties of X and Y

approachesinfinity, the probability that an unobservableproperty of y is similar to
an observed propertyof x approaches1.

4. Discussion. The fallacies which occur when an analogy is incorrectly applied

usually arise because in most practicalcases the numbern of successful comparisonsis
small, perhapsonly three or four. Thus it is incorrectto say that the success of com-
parisonq4indicatesthat the probability
P5 = ' 1
P(q5/q1.q2.q3.q4.H) (2)
since the probability
P4 = P(q4/qj.q2.q3.H) (3)
of the success of q4given ql, q2, q3 and H may not be high. It cannot even be affirmed
that P5 > P4, althoughit would be extremelydifficultto arguethat P5 < P4. All that
may be inferredis that for two phenomenaX and Y there exists a numberN such that
for n > N
Pn+l > Pn.

This follows directlyfrom (1) since Pn ?: 1 by definition.

Thus before the scientist may make a meaningfulstatementregardingP,+1 he must
make some estimate of the probablevalue of N for the particularphenomenaX and
Y. Having assessed N, he is in a position to state whether or not the probabilityof a
hypothesis is increasedby appeal to the argumentby analogy.
I now discuss a problem in solar physics in which the argument by analogy has
increased the probability of a particular hypothesis, and show how the complete
argument(i.e. analogywith crucial test) has led to some useful progress in this field.

5. Application to solar granulation. High resolution photographs of the solar

photosphere indicate that it is not uniformly bright, but consists of a large number
of bright granules against a dark background.This was first observed by Janssen
(1896) and subsequently by many other observers. Perhaps the best photographs
availabletoday are those of M. Schwarzschild[8] obtainedfroman unmannedballoon
at a height of 25 km. These have been intensively analysedby Edmonds [2] and his
resultsrepresentthe best currentlyavailableevidence.Owing to the obvious difficulties
of direct investigation, the astrophysicistwishing to interpret this phenomenon is
faced with many hypotheses, each of which has a low initial probability.The task of
designing a crucial test for each of such hypotheses, differentiatedonly by the sim-
plicity postulate, would be indeed formidable.However, many observershave noted

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
 ON THE ARGUMENT BY ANALOGY 37

the similarityin appearancebetween the granulationpattern and convection cells of

the Benard type which may be set up and photographedin the laboratory(see [7]).
The resemblance between these and some of Schwarzschild'sbest photographsis
particularlystriking.
In order to make progress towards an understanding of the phenomena which
manifest themselves as granulation,the "argumentby analogy" is now used, in the
strict sense as defined in paragraph3 and paragraph4 above. Given only the granula-
tion photographsand a knowledgeof Benardcells as dataH, we considerthe hypothesis
p on data H that, apart from obvious dissimilarities,the Benard cells and the pho-
tosphericgranulationare parallelphenomena.Although perhapsunlikely, it cannot be
ruled out a priori as having zero initial probability,and this is all that the argument
requires. The obvious dissimilaritiesmay be classified and therefore enumeratedin
terms of differencesin size, temperatureand pressure of the phenomena. Two other
distinctionshave been noted however;(i) The Benardcells are almostregularpolygons
while the granules,althoughpolygonal, are less regularin shape; (ii) The Benardcell
patternis stationarywhile the granulationpatternvaries.
The hypothesis p now entails that, apart from these features, comparisontests qi
between the two phenomenashould be successful. Thus the observed similaritiesin
the general polygonal shape of the granules, the dark background,the dark central
region in some granules,the generalstatisticaldistributionof the patternetc. represent
successful tests ql, q2 ... q,. If the test qnl+ comparesthe mode of formation of the
phenomena,its probabilityof success P4+1 P(q,+l/ql.q2...qn. H) is equivalentto the
probabilityof the hypothesis r that granulationis essentially a manifestationof con-
vection in the upper part of the photosphere.This is but one of the many hypotheses
concerningthe natureof the granulation,all of which have a low initial probability.In
paragraph3 it was shown that
lim Pn+1 = 1,
n--coo

which entails that tor n > N, where N is specific to a particularpair of phenomena

Pn+1 > Pn > Pi,

where P1 represents the probability of success of a random comparison test between

the granulation and the Benard cells. Given only data H, the initial probabilities of the
tests qi are equal, or at least of the same order, i.e. P(q1/H) = P(q,+1/H). Thus we
may equate P1 to the initial probability of the hypothesis r, P(r/H). The only difficulty
at this stage lies in the decision whether n, the number of successful tests, exceeds the
hypothetical number N and such a decision cannot be made with certainty. However,
the probability that n exceeds N may be assessed by the astrophysicist and, if sufficiently
large, he may feel justified in concluding that

Pn+1 > Pi,

i.e. the probability of r given the success of the n comparison tests P(r/ql.q2...qn.H)
is greater than the initial probability of r, P(r/H). In this event a crucial test should
be designed on the basis of the hypothesis r.
Recently considerable progress has been made in this direction. Edmonds' detailed
observations provided hitherto unavailable opportunities for testing hypotheses and
coincided with the development of a new method of interpretation using radiative
transfer theory in non-uniform media [9]. Using this I have been able to show that

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
38 P. R. WILSON

an essential featureof Edmonds' results (a peak in his intensity contrastdistribution)

may be accounted for by a simple model of a convective cell system [10]. Although
this result has not, as yet, been explained on any other hypothesis, this possibility
cannot be ruled out at this stage. However, it can be claimed that, as a result of this
test in conjunctionwith the analogywhich suggested it, the probabilitythat the granu-
lation is a manifestationof non-stationarycellularconvectionis considerablyincreased.
Of courseother analogiesbetweengranulationand well knownterrestrialphenomena
have been observed and widely supported. It is not the purpose of this paper to
discuss all such theories,but ratherto show how the argumentby analogysupportsthe
serious considerationgiven to one of them.

6. Conclusion. By studying the recent literatureon granulationit is possible to see

how the variousobservershave estimatedthe probabilitiesinvolved in the convection
analogy, and to follow their resulting progress.It is all the more interestingbecause
the chapter is not yet closed and further developmentsmay confirmor deny current
hypotheses. In this respect the recent observations of photospheric velocities by
Leighton et al. [6] are of considerableinterest.
On the other hand analogieswhich were briefly mentioned in the introductionare
almostpartof scientifichistory and thus providean opportunityto assess the complete
applicationof the argument.Although Bohr's atomic model is now consideredrather
naive and has been replacedby the more sophisticatedtheoriesof quantummechanics,
it is a striking example of the correct and logical use of an analogy which led to a
majorscientificadvance.In orderto test this hypothesis,many importantexperiments
were designed and it was found possible to explain or predict such phenomena as
atomic spectra, valency and radioactivityusing Bohr's model. However when the
model was found to be inconsistentwith variousnuclearphenomenaand the statistical
nature of atomic phenomena,it was correctlyrejected in favour of the quantum me-
chanicalmodel.
A less successful analogy was that between the earth's magnetic field and a bar
magnet, which was used to support the hypothesis that the earth's core is solid. Its
main weaknesswas its failure (at the time) to suggest any confirmatoryobservations,
and without these, as I have suggestedabove, the argumentis of little value. Estimates
of temperaturesand pressuresthought to exist in the core (if these may be considered
as tests) were inconsistent with ferro-magnetismas observed in the laboratory,and
the argumentwas dropped. Only comparativelyrecently has seismology provided the
possibility of furthertesting and the theory of a solid inner core is now under serious
discussion.
This second example has been mentioned to show that although the argumentby
analogy is not always spectacularlysuccessful, it should not lead to false conclusions
provided it is correctlyused. It is the contention of this writer that the argumentby
analogy is a legitimate part of scientific logic and therefore deserves a less humble
place in the esteem of scientificphilosophersthan it holds at present.

7. Acknowledgement. This projectwas originallysuggestedto me by Dr. Bruce A.

Bolt of the Universityof Sydney.'It is a pleasureto acknowledgehis helpful suggestions
and constructive criticism.

1 Now at the University of California, Berkeley, U.S.A.

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions
 ON THE ARGUMENTBY ANALOGY 39

REFERENCES

[1] BUNGE, Mario, Phil. of Sci., 27, 262, 1960.

[2] EDMONDS, F. N., Astrophys. J., Supp. Series No. 50, 6, 357, 1962.
[3] HOSPERS, J., An Introduction to Philosophical Analysis (Routledge and Kegan Paul), 1960.
[4] JEFFREYS, H., Scientific Inference, (2nd Edit.) (Cambridge University Press), 1957.
[5] KEMENY, J. G., Phil. Rev. 62, 391, 1953.
[6] LEIGHTON, R. B., NOYES, R. W., and SIMON, G. W., Astrophys. J., 135, 474, 1962.
[7] MINNAERT, M., The Sun, ed. G. Kniper. (University of Chicago Press), 1, 173, 1953.
[8] SCHWARZSCHILD, M., Astrophys. J. 130, 345, 1959.
[9] WILSON, P. R., Monthly Notices R.A.S., 123, 287, 1962.
[10] WILSON, P. R., (b) Astrophys. J., 137, 606, 1963.

This content downloaded from 128.235.251.160 on Sat, 13 Dec 2014 09:10:22 AM

All use subject to JSTOR Terms and Conditions