R. A.

Fisher on the History of Inverse Probability

Author(s): Sandy Zabell
Source: Statistical Science, Vol. 4, No. 3 (Aug., 1989), pp. 247-256
Published by: Institute of Mathematical Statistics
Stable URL: http://www.jstor.org/stable/2245634 .
Accessed: 02/10/2013 23:47

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.

Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and extend access to
Statistical Science.

http://www.jstor.org

This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM

All use subject to JSTOR Terms and Conditions
 StatisticalScience
1989, Vol. 4, No. 3, 247-263

R. A. Fisheron the History

of Inverse Probability
SandyZabell

Abstract.R. A. Fisher'saccountofthedeclineofinverseprobabilitymeth-
ods duringthe latterhalfof the 19thcenturyidentifiesBoole,Venn and
Chrystalas the keyfiguresin this change.Carefulexaminationof these
and otherwritingsof the period,however,revealsa different and much
morecomplexpicture.Contrary to Fisher'saccount,inversemethods-at
leastin modifiedform-remained theoretically untilthe1920's,
respectable
whentheworkofFisherand thenNeymancausedtheireclipseforthenext
quartercentury.
Keywordsandphrases:R. A. Fisher,inverseprobability, of
history
statistics.

R. A. Fisherwas a lifelongcriticofinverseproba- of Fisher'sown evidenceforhis claims revealsan

bility.In thesecondchapterofhislastbook,Statistical
Methodsand ScientificInference(1956), Fisher traced
thehistory ofwhathe saw as theincreasing disaffec-
tion withBayesian methodsthat arose duringthe
secondhalfof the 19thcentury.Fisher'saccountis
oneofthefewthatcoversthisneglected periodin the
historyof probability,in effecttaking up where
Todhunter(1865) leftoff,and has oftenbeen cited
(see,e.g.,Passmore,1968,page550,n. 7 andpage551,
n. 15; de Finetti,1972,page 159; Shafer,1976,page
25). The pictureportrayed is one ofgradualprogress,
thelogicallacunaeand misconceptions oftheinverse
methodsbeingsteadilyrecognized and eventuallydis-
credited.
But on reflection Fisher'sportraitdoes notappear
entirelyplausible.Edgeworth and Pearson,twoofthe
mostdistinguished statisticiansofthegeneration im-
mediately priorto Fisher's,werebothsympathetic to
inversemethods; andindeed,as willbe discussedlater,
Bayesianmethodswerewidelytaughtand employed
inEnglandand elsewhere untilthe1930's.It was only
thenthatFisherand Neymansimultaneously admin-
istereda nearlylethalblowto Bayesianstatistics, one
fromwhichitwas notto recoveruntilthepublication,
nearlya quarterofa century later,ofSavage'sFoun-
dationsofStatisticsin 1954.
How was sucha disparity betweenFisher'saccount
and historicalrealitypossible?Carefulexamination
Sandy Zabell is Associate Professorof Mathematics
and Statisticsat NorthwesternUniversity.His mailing
address is: Department of Statistics, Northwestern
University,2006 Sheridan Road, Evanston, Illinois
60208.
interestingstory,tellingus perhapsin somewaysas
muchabout Fisher as it does about the periodhe
discusses.
1. FISHER'S ACCOUNT
Fishercitesthreemajorauthorities forthe decline
in the prestigeof inversemethods:Boole, Vennand
Chrystal.He had doneso repeatedlyin earlierpapers
(Fisher,1922, pages 311 and 326; 1930, page 531;
1936a,page 248; 1951,page 49), and his accountin
Statistical
Methodsand Scientific
Inference (SMSI) is
an elaborationon these earlier,fragmentary com-
ments.The following passagesgivethe flavorof his
argument:
The firstseriouscriticism
was developedbyBoole
in his "Laws of Thought"in 1854.... Boole's
criticismworkedits effectonly slowly.In the
latterhalfof the nineteenthcenturythe theory
ofinverseprobabilitywasrejectedmoredecisively
by Venn and by Chrystal.... [Fisher,1936a,
page 2481
[Venn's criticismsof the Rule of Succession],
froma writerof his weightand dignity,had an
undoubtedeffectin shakingthe confidenceof
mathematicians in its mathematicalfoundation.
[SMSI, page 251
Perhapsthemostimportant resultofVenn'scrit-
icismwas the departuremade by ProfessorG.
Chrystalin eliminatingfromhis celebratedtext-
bookofAlgebrathewholeofthetraditional ma-
terial usuallypresentedunderthe headingsof
InverseProbability
and oftheTheoryofEvidence.
[SMSI, page 29]

247

This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM

All use subject to JSTOR Terms and Conditions
248
Fisherdidnottrytooverstate theimmediate impact
of these He
criticisms. noted "the slowness withwhich
theopinionsofBoole,Vennand Chrystal wereappre-
ciated,"and drewattentionto thedefensesofinverse
probabilitymounted by Edgeworth (1908) and
Pearson(1920). Fisherwas not alwaysconsistenton
this point,however.Writinga fewyears later,he
describesthesupposedrejectionofinverseprobability
in Englandas occurring "abruptly and dramatically"
(Fisher,1958,page 273),and uses thephrase"as late
as 1908"in referring to Edgeworth's paper.Neverthe-
less, Fisher'searlierreference to the "decisivecriti-
cismsto which[themethodsof inverseprobability]
had been exposedat the handsof Boole, Venn,and
Chrystal"(1922, page 326), and his assertionthat
"[t]hesecriticismsappeartobe unanswerable, andthe
theoryofinverseprobability . . . is nowalmostuniver-
sally abandoned"(1951, page 49) capturethe basic
pointsof his moreextendedaccountin SMSI: these
werethekeycritics, theircriticisms werewell-founded
and theywerelargelyresponsible forthe declineand
fallofinverseprobability.
The reader,however, whoturnsto Boole,Vennand
Chrystalto see whattheyactuallywrote-howaccu-
ratelyFisherrepresents theirviewsandtowhatextent
theyactuallysupportFisher'sposition-willfindthe
resultsurprising.
2. BOOLE
Boole,Fishersays,was thefirstto seriously criticize
"Bayes'doctrine"(Fisher,1936a,page 249;cf.Fisher,
1951,page 49). This was onlypartiallytrue.Robert
LeslieEllis had a decadeearlierformulated a frequen-
tisttheoryof probability (Ellis, 1844) and criticized
the Laplacian approachto inference on a numberof
grounds includingex nihilonihil(outofnothing, noth-
ing)-i.e., no inferenceat all is warranted in a situa-
tionofcompleteignorance. JohnStuartMill had also
been,albeitbriefly, a critic.In addition,bothJakob
FriedrichFries in Germanyand AntoineAugustin
Cournotin Francehad earlierdiscussedobjectiveor
frequentist theoriesof probability and attackedun-
criticalapplicationsof inverseprobability. (Cournot
was less stridentthan an earlierFrenchtradition
represented byDestuttde Tracy,PoinsotandAuguste
Comte;see generallyPorter(1986,pages 77-88) and
Stigler(1986,pages 194-200).Fishersometimesap-
pearsto havebeen surprisingly unfamiliar with19th
centurydevelopments outsideof England,and this
oftengiveshis historicaldiscussionsa somewhatin-
sular flavor.Thus, he also makes no mentionof
Bertrand,althoughBertrand'sCalculdesprobabilites
(lst edition,1889) sharplycriticizedinversemethods
and was withoutquestionhighlyinfluential.)
Boole's criticismswerea naturaloutgrowth of his
philosophical viewthatprobability is a logicalrelation
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
S. ZABELL
betweenpropositions.In this he was veryclose to
De Morgan;bothDe Morgan'sFormalLogic (1847)
and Boole's Investigation of the Laws of Thought
(1854) treatedprobability as a branchof logic.But
whileDe Morganand othersbelievedthatany event
possessed-at leastin principle-a definitenumerical
probability relativeto a givenbody of information
(see, e.g.,De Morgan,1847,page 178; Donkin,1851,
pages 354-355),Boole arguedthat,lackingsufficient
information, the probabilitiesof some eventswere
indeterminate.
This was an important point,becausea majorde-
fenseofuniform priorsin Boole's daywas a challenge
to doubtersto producea moreplausiblealternative:
"Apersonwhoshoulddisputethepropriety ofdividing
ourbeliefequallyamongsthypotheses aboutwhichwe
are equallyignorant,oughtto be refutedby asking
himto statewhichis to be preferred. He musteither
admittheproposedlaw,or maintainthatthereis no
law at all" (Donkin,1851,page 355). The latteris
preciselywhatBoole did. As a result,he was able to
criticizeprevioustreatmentswhichattempted to side-
stepindeterminacy byhypothesis:
It has been said, that the principleinvolvedin
the above and in similarapplicationsis that of
theequal distributionofourknowledge, orrather
ofourignorance-theassigning todifferentstates
of thingsof whichwe knownothing,and upon
the verygroundthat we know nothing,equal
degreesofprobability.I apprehend,however, that
thisis an arbitrarymethodofprocedure.[Boole,
1854,page 370]
Boole supported thiscriticism bymakingthesimple
buttellingpointthatin somecasestheprinciple could
be appliedin morethanonewayto thesameproblem,
resultingintwoormoreconflicting assign-
probability
ments.For example,Bayes had arguedthat "in the
case of an eventconcerning theprobabilityofwhich
we absolutelyknownothing to anytrials
antecedently
madeconcerning it ... I haveno reasonto thinkthat,
in a certainnumberoftrials,it shouldratherhappen
anyone possiblenumberoftimesthananother;"i.e.,
that
P[Sn = k] = 1/(n+ 1), k = O, 1,.. n
(whereSn denotesthenumberofsuccessesin n trials).
But,as Boolepointedout,onecouldequallywellargue
thatall sequencesof outcomesin n trialsshouldbe
in an entirelydif-
viewedas equallylikely,resulting
ferentprobabilityassignment.(Bertrand'sparadox
(involvingrandomchoiceof a chord)madethe same
pointfora continuousvariate(Bertrand,1907,pages
4-5). Alongthesamelines,Fisherwasfondofpointing
out that uniformpriors on continuousparameter
 FISHER'S HISTORY OF INVERSE PROBABILITY 249

spaceswerenotinvariantunderall continuous trans- positionofEF betweenAB and CD is equallylikely;

formations (e.g.,Fisher,1956,page 16).)
This was an important observation, but it did not
compelabandonment oftheprincipleofindifference.
It didprovidea warningthatnaiveapplicationofthe
principlecouldlead to paradoxesand inconsistencies,
and duringthe next centurymanyphilosophers-
notablyvon Kries,Keynes,Jeffreys, and Carnap-
undertookto refineit in an attemptto avoid them
(von Kries, 1886; Keynes, 1921; Jeffreys, 1939;
Carnap,1950).
Nor didBoole himself advocateabandonment ofthe
principle.This mightnothavebeenapparentto some-
one readingonly The Laws of Thought,for there
mentionoftheprincipleis indeedlimitedto a discus-
sion of its improper usage.But Boole repeatedly re-
turned to the foundations of probability in
his subsequentpapers, and Fisher would scarcely
have foundhimselfin agreement withBoole's later
opinions.
In his last, perhapsmostconsideredthoughtson
thesubject,Boole wrotethat:
All theprocedure ofthetheoryofprobabilities is
foundedon the mentalconstruction oftheprob-
lemfromsomehypothesis, either,first,ofevents
known to be independent;or secondly,of
eventsof the connexionof whichwe are totally
ignorant;so thatupon the groundof this igno-
rance,we can againconstruct a schemeofalter-
nativesall equallyprobable,and distinguished
merelyas favouring or notfavouringtheeventof
whichtheprobability is sought.In doingthiswe
are not at libertyto proceedarbitrarily. We are
subject,first,to the formalLaws of Thought,
whichdetermine thepossibleconceivablecombi-
nations;secondly,to thatprinciple,moreeasily
conceivedthanexplained,whichhas beendiffer-
entlyexpressedas the"principle ofnon-sufficient
reason,"the "principleof the equal distribution
ofknowledge or ignorance,"and the"principleof
order."We do not knowthatthe distribution of
propertiesin the actual urn is the same as it is
conceivedto be in theidealurnoffreeballs,but
the hypothesisthat it is so involvesan equal
distributionofouractualknowledge, and enables
us to constructthe problemfromultimatehy-
potheseswhichreduceit to a calculationofcom-
binations.[Boole, 1862,pages 389-390 of 1952
edition]
or at least we knowwhat amountof assumptionis
involvedinthissupposition. which
In theapplications
have been made of Bayes's theorem,and of such
resultsas that whichwe have taken fromLaplace
in Art. 551, therehas howeveroftenbeen no ade-
quate groundfor such knowledgeor assumption"
(Todhunter,1865, pages 299-300). Fisher praised
Todhunter'semphasison the necessityfor a fac-
tual rather than an axiomatic basis for prior
probabilities.Nevertheless,because of Todhunter's
use ofthequalifyingphrase"orat leastwe knowwhat
amountofassumption is involvedinthissupposition,"
Fisherconcludedthat"Nearas he cameto clarifying
the situation,Todhunter'sname cannotproperlybe
added to thosewho finallysucceededin extricating
themathematical thought ofthemid-nineteenth cen-
turyfromits bewildering This suggests
difficulties."
thatFisherwouldhavebeenhighlycriticalofBoole's
laterremarks.)
Had Boole changedhismind?He claimsnot,forhe
addedin a footnote:
.. . I take this opportunityofexplaininga passage
in the Laws of Thought,page 370, relatingto
certainapplicationsoftheprinciple.Validobjec-
tion lies not against the principleitself,but
against its applicationthrougharbitraryhy-
potheses,coupledwiththe assumptionthatany
resultthusobtainedis necessarily the trueone.
The applicationoftheprincipleemployedin the
text and foundedupon the generaltheoremof
development in Logic,I holdto be notarbitrary.
The distinction thatBooleintendspitstheso-called
"principleof insufficient reason,"againstwhat was
laterdescribedas the"principle ofcogentreason,"i.e.,
thattheprobabilities assignedto alternativesshould
be takento be equal if the information aboutthose
alternatives equallyfavorseach (as Boole puts it, if
thereis "an equal distribution of our actualknowl-
edge").In anycase, it is clearthatBoole was notan
opponentoftheuse ofsomeformoftheprinciple, and
was opposedinsteadto whathe considered itsuncrit-
ical application.(As Keynes (1921, page 167) and
manyothershave noted,Boole's writingson proba-
bilityare also marredby a systematic confusionbe-
tween two differentmeanings of independence.
Hailperin(1976)providesa helpfulguidethrough the
thicket.)

ObviouslyFisher could neverhave acceptedthis

viewofthenatureofprobability,ortheimprimatur it
bestowsupon the use of theprincipleof insufficient
reason.(In thethirdeditionofSMSI, Fisheraddeda
subsectionon Todhunter, who had emphasizedthat
"in Bayes's ownproblem,we knowthata prioriany
3. VENN
JohnVennwas a Cambridgelogician,best known
todayforhispopularization of"Venndiagrams," and
in his own day forhis influential
textbookSymbolic
Logic (1st edition,1881; 2nd edition,1894). Yet in

This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM

All use subject to JSTOR Terms and Conditions
250 S. ZABELL

termsof originality and long-termimpact,Venn's This was not a minormatter,inasmuchas Venn

mostimportant work is his LogicofChance(lst edi- made substantialrevisionsin both the second and
tion,1866),which gave thefirstdetaileddiscussionin thirdeditionsofThe LogicofChance.(A comparative
English of a frequentist theoryofprobability,as well studyof the threeeditionsof the Logic,tracingthe
as a carefulcritiqueof the earlierLaplacean position, evolutionofVenn'sthought, wouldbe ofconsiderable
includingbothits use ofuniform priorsand the con- interest.Salmon (1980) discussessome differences,
sequences that followfrom such an assumption.It was butlargelyconfineshis attention to thefirstedition.)
Venn'sdiscussion of one of these consequences that In thisinstance,betweenthesecondand thirdedi-
Fisher examined in SMSI. tionsVennmademajorchangesin thechapteron the
RuleofSuccession,takingoutprecisely theexamples
thatFisher so vehemently objectedto. It is naturalto
3.1 The Rule of Succession
assume that betweeneditionsa colleagueor corre-
Laplace's"RuleofSuccession"states(in brief)that spondent-verylikelyEdgeworth, whosehelp is ac-
an eventwhichhas occurred n timesin successionwill knowledged in the preface-voiced criticismsvery
recurthe next time with probability (n + 1)/(n+ 2). similarto Fisher's;indeed, Venn's revision addresses
Vennridiculed the Rule of Succession, pointingouta preciselythe pointsraisedby Fisher:the mathematical
variety of cases where it contradicted common sense assumptions underlying thederivationoftherule,and
(rainon three successive days; death caused by admin- theirpossibleempiricalvalidity.
isteredstrychnine on threeseparateoccasions;people Anotherpuzzleis the tenorof Fisher'sdiscussion.
answering a falsecall offireon threedifferent occa- Fisherwas in a certainsense very"political"in his
sions). WhileFishercitedVennwithgeneralappro- writings;oftenquick to attack the opposition,he
bation,hetookissuewithhimon thisparticular point. seldomexpressedin printreservations he mightex-
As Fisherwas quickto pointout,"sucha rulecan be press to close friendsand allies. That he should
basedon Bayes'theoremonlyon certainconditions." sharplycriticizean ally like Venn seems strangely
In particular, the successivetrialsmustbe indepen- inconsistent withhisusualpractice.In thiscase,how-
dent, which is certainly notthecase in twoofVenn's ever,a simpleexplanationsuggestsitself.
examples. Venn's criticisms werenot of the inverseruleper
Fisherwas in fact highlycriticalof Venn: Venn se, but its mathematicalconsequence,the Rule of
"perhapswas not awarethatit [theRule of Succes- Succession.Thus, the exampleshe adduces,to the
sion] had a mathematicalbasis demonstrated by extentthattheydiscredit theRuleofSuccession,also
Laplace;""thereis no doubt that Venn in thischapter discreditany form of inferencethatgivesriseto the
uses arguments of a qualitywhichhe wouldscarcely Rule of Succession.
have employedhad he regardedthe matteras one And that wouldincludefiducialinference. For in
open to rationaldebate;"Venn's examples"seemto thenextchapterofSMSI, duringa discussionofthe
be littlemorethanrhetorical salliesintendedto over- applicationofthefiducialargument to discontinuous
whelman opponent with ridicule;"and "byhis eager- data,Fisher notes that:
ness to dispose of [the Rule of Succession] ... he
becameuncriticalofthe qualityofthe arguments he An odd consequenceof the analysisdeveloped
used." above is that the Rule of Successionderivable
In ordertojudgethevalidityand persuasiveness of fromthe particulardistribution a
of probability
Venn'streatment, in the lightof Fisher's comments, priori
it is naturalto turnto Venn'soriginaldiscussion,in
orderto read his arguments in context.The reader dp
who turnsto the reprinted edition of The Logic of
Chance,however, will find to his surprise that al-
thoughVenndoes indeed devote an entire chapter to ofsuccessin thenext
namelythattheprobability
theRuleofSuccession, the passagesthat Fisher quotes trialis
are nowhereto be found!
The solutiontothispuzzle,however, is notdifficult. a+ 1/2
The Logicof Chance went through three editions a+ b+ 1
1866,1876,and 1888, the currently available Chelsea
reprint being the last of these. Although Fisherdoes is justifiable,at least to a remarkably highap-
not indicate in SMSI which edition he consulted, a proximation, in the absenceof any knowledge a
comparison ofeditions revealsthat Fisher was quoting priori;and thisalthoughthecorresponding com-
fromthe 2nd edition,a copyof whichhe mayhave pletedistribution is notso justifiable.
a posteriori
owned(thiseditionis citedin Fisher,1955). [Fisher,1956,page68]

This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM

All use subject to JSTOR Terms and Conditions
 FISHER'S HISTORY OF INVERSE PROBABILITY
Thusan attackon theRuleofSuccessionwas actually
an indirectattackon the fiducialargumentas well
and,as such,had to be met.But Fisherwas curiously
coy aboutthe matter.In his discussionof Venn,no
mentionis madeof the factthatthe Rule can be so
justified,onlythat Venn's criticismswerespecious.
AndwhenFisherderivestheRule as an approximate
consequenceofthe fiducialargument, no mentionis
madeofVenn'scriticisms.
Thereis noclearevidencewhether Fisherwasaware
of the thirdeditionof Venn's Logicof Chance.Cer-
tainly,had he seen it,he wouldhaveapprovedofthe
changesVenn made in the chapteron the Rule of
Succession.But Venn madea numberof otherrevi-
sionsas well,oneofwhichFisherwouldmostcertainly
nothaveapproved.
3.2 Probabilityand Listerism
In 1879,Dr. Donald MacAlister
posedthefollowing
questionin thepagesoftheEducationalTimes:
Of 10 cases treatedbyLister'smethod,7 didwell
and3 suffered fromblood-poisoning:
of14treated
withordinarydressings,9 did well and 5 had
bloodpoisoning;whatare the oddsthatthe suc-
cess ofLister'smethodwas due to chance?
Due to the smallsizes ofthe samplesinvolved, the
large-samplemethodsthen available for analyzing
suchdifferences wereinapplicable, and the Bayesian
solutionadvocatedby MacAlisterinvolvedassigning
independent uniform priorsto the twounknownbi-
nomialproportions (see generally
Winsor,1947).
In the 3rd editionof the Logic of Chance,Venn
includeda discussionof MacAlister'squestion.Con-
sistencyrequiredthat Venn rejectMacAlister'sap-
proach,yetVennwas obviously uncomfortable witha
positionthatno inference couldbe drawn.The result
was a surprising reversal.Venndescribestheexample
as illustratingthosecases whichafforded "[t]henear-
estapproachto anypracticaljustification for[inverse]
judgments," and approvesof MacAlister'streatment
of it as a 'bag and balls' problem;being "the only
reasonablewayoftreatingthe problem,ifit is to be
considered capableofnumerical solutionat all" (Venn,
1888,pages186-187).Thus farFishermightstillhave
had no difficulty.But thenVennwenton to add:
Of coursethe inevitableassumptionhas to be
madehereaboutthe equal prevalenceofthedif-
ferentpossiblekindsofbag-or, as the support-
ers ofthejusticeofthe calculationwouldput it,
of the obligationto assume the equal a priori
likelihoodofeach kind-but I thinkthatin this
particularexamplethe arbitrariness of the as-
sumptionis less thanusual. This is becausethe
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
251
problemdiscussessimplya balancebetweentwo
extremely similarcases,andthereis a certainset-
offagainsteach otherof the objectionableas-
sumptionson each side. Had one set of experi-
mentsonlybeenproposed, andhadwebeenasked
toevaluatetheprobability ofcontinued repetition
of themconfirming theirverdict,I shouldhave
feltall thescruplesI havealreadymentioned. But
herewe havegottwosetsofexperiments carried
on underalmostexactlysimilarcircumstances,
and thereis therefore in assum-
lessarbitrariness
ingthattheirunknownconditionsare tolerably
equallyprevalent.
Venn's logicis difficult
to follow;the last three
sentencesseemmorea rationalization thana carefully
thought-out argument.(This is hardlysurprising,
sincethe positionVenn nowtakes is totallyincom-
patiblewiththeonehe hadpreviously adopted.)What
is clearis thatFisherwouldhave rejectedit entirely.
Todhunterhad been excludedfromthe pantheonof
clarificationfordefendingBayes'spostulatewhen"we
knowwhatamountof assumptionis involvedin this
supposition." Fisher'sreactionto Venn'sapostasycan
onlybe conjectured.
4. CHRYSTAL
Chrystal, Fishersays,"does not discussthe objec-
tions to this material[inverseprobabilityand the
theoryof evidence]."This was onlypartlytrue.Al-
thoughChrystaldid not elaboratein his Algebraon
his reasonsforomittinginverseprobability, he did
returnto the subject5 yearslater and presenthis
objectionsin detail.It was easyto overlookthispaper
of Chrystal's,forit appearedin the Transactions of
theActuarialSocietyofEdinburgh(1891), a journal
not widelyavailable, as anyone who attemptsto
consult Chrystal'spaper will readilyfind. In his
1891paper,Chrystalspelledout his viewson proba-
bility,viewsthatFisherwouldhave founda serious
embarrassment.
Fisherhad alwaysbeen at painsto emphasizethat
he had no objectionto the use of Bayes's theorem,
only to its unwarrantedapplicationin situations
whereinformation justifying the use of a priorwas
in
unavailable; particular, Fisherobjectedto theprin-
cipleof insufficientreasonto assignpriors(see, e.g.,
SMSI, page20). Chrystal's objections, were
ironically,
exactlythe opposite:he did not objectto the use of
ignorancepriors,but thoughtthat given a prior,
Bayes'stheoremcouldgeneratean incorrect answer!
He writes:
Perhapsthefollowing... willmaketheabsurdity
of the supposedconclusionof the InverseRule
stillclearer.
925
A bag contains threeballs, each ofwhichis either
whiteor black, all possible numbersofwhitebeing
equally likely.Two at once are drawn at random
and prove to be white;what is the chance that all
the balls are white?
Any one who knows the definitionof mathemat-
ical probability,and who considersthis question
apart fromthe Inverse Rule, will not hesitate for
a momentto say that the chance is 1/2; that is to
say,that the thirdball is just as likelyto be white
as black. For thereare fourpossible constitutions
of the bag:
10 20 30 40
W 3 2 1 0
B 0 1 2 3 two most prominentEnglish statisticiansof the gen-
each of which,we are told, occurs equally often
in the long run,and among those cases there are
two (10 and 20) in whichthereare two whiteballs,
and amongthese the case in whichthereare three
white occurs in the long-runjust as oftenas the
case in whichthere are onlytwo.
Chrystalthen goes on to correctlycalculate that,in
contrast,the "application of the Inverse Rules" leads
to posteriorodds of 3 to 1 in favorof the third ball
being white,and concludes:
No one would say that ifyou simplyput twowhite
balls intoa bag containingone ofunknowncolour,
equally likelyto be black or white,that this action
raised the odds that the unknown ball is white
fromeven to 3 to 1. It appears, however,fromthe
InverseRule that ifwe findout that the twowhite
balls are in the bag, not by puttingthem in, but
by takingthem out, it makes all the difference.
Indeed it does. Chrystal'serroris exactlythe point of
the closely related Bertrand box paradox (Bertrand,
1907, pages 2-3).
In the lightof this fundamentalmisunderstanding,
Chrystal's objections to inverse probability can
scarcelybe describedas intellectuallydevastating.He
was merelyone of many (e.g., D'Alembert and Mill)
whose intellectualattainmentsin otherareas led him
to uncriticallyaccept his own untutoredprobabilistic
intuitions.As Jevons once noted, "It is curious how
oftenthe most acute and powerfulintellectshave gone
astray in the calculation of probabilities" (Jevons,
1877, page 213). (In 1893, shortlyafterChrystalread
his paper beforethe Actuarial Society of Edinburgh,
JohnGovan read a paper beforethe same body,point-
ing out the errorsand confusionsin Chrystal'spaper.
It went unpublished,however,until 1920, when the
eminentmathematicianE. T. Whittakerread a simi-
lar expose before the London Faculty of Actuaries
(Whittaker,1920).)
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
S. ZABELL
The conclusion to this episode in the historyof the
historyof statisticsis somewhatbizarre.Of his trinity
of authorities-Boole, Venn and Chrystal-Fisher
thoughtBoole was an opponent of inverse methods,
but Boole was not; Venn was an opponent,but only
in part; and Chrystal was an unqualified opponent,
but on grounds Fisher would have found repugnant,
had he known of them.
5. INVERSE PROBABILITY FROM 1880 TO 1930
What was the actual impact of these critics? Con-
traryto what Fisher suggests,they did not eliminate
inversemethods.Edgeworthand Pearson, perhaps the
eration immediately preceding Fisher's, both re-
mained sympatheticto Bayesian methods.Moreover,
we have the testimonyof Fisher himselfthat he had
"learned it at school as an integralpart ofthe subject,
and for some years saw no reason to question its
validity"(Fisher, 1936a, page 248). Indeed, he had to
"plead guiltyin my originalstatementof the Method
of Maximum Likelihood [Fisher, 1912] to having
based my argument upon the principle of inverse
probability.. ." (Fisher, 1922, page 326).
The real effectof Boole, Venn, and Chrystal and
other critics appears rather to have been to cause
the exponents of inverse methods to hedge their
claims for the theory. For example, William Allen
Whitworth,the author of a popular 19th century
textbookChoice and Chance, dealt with objections to
the rule of succession by conceding that expressions
such as "entirelyunknown" in its formulationwere
"vague." He proposedthat theybe replaced in the rule
by the explicithypothesisthat "all possible probabil-
ities [are] equally likely,"and noted that:
Though the cases are very rare in which the
radical assumption of the Rule of Succession is
strictlyjustified,the rule may be taken to afford
a rough and ready estimate in many cases in
which the assumption is approximatelyjustified.
[Whitworth,1901, page 193]
This defenseessentiallyoriginateswithEdgeworth,
who was an importantdefenderof inverse methods
throughoutthis period (see Stigler, 1978, page 296;
1986,page 310). In 1884,at the beginningofhis career,
Edgeworthwrote a review of Venn's Logic, entitled
"The Philosophy of Chance," which appeared in the
English philosophicaljournal Mind. (Nearly 40 years
later, in the twilightof his career, Edgeworthwould
returnto the same subject withan articleof the same
titlein the same journal, this time reviewingKeynes's
Treatise.) Edgeworthtook an empiricaland pragmatic
view of the subject, and, as noted earlier, may well
 FISHER'S HISTORY OF INVERSE PROBABILITY
havebeen responsible
formanyofthe changesVenn
made in the thirdeditionof The Logic of Chance.
The defensesmountedbyEdgeworthand othersfell
intothreebroadcategories.
Theywere:(1) TheBayes-
Laplace postulate of equiprobabilitycorresponds,at
least approximately,
to experience.Karl Pearson found
thisargument particularly
persuasive,and adoptedit
in his influentialGrammarof Science (lst edition,
1892)and laterarticles(Pearson,1907;Pearson,1920,
page4). (2) Otherpriors.Anothermovewasto concede
that experiencemightindeedpointto otherpriors.
Boththe actuaryG. F. Hardy(1889) and the mathe-
maticianWhitworth (1897,pages 224-225)proposed
the class of beta priorsas suitableforthispurpose.
Others,such as Gosset (1908) and Bachelier(1912),
suggestedthe use of polynomialpriors. (3) The
suppressionofa prioriprobabilities(Edgeworth,1922,
page 264). A thirdand finaldefensewas thatwhen
largesampleswereinvolvedthe particularpriorem-
ployeddid not matter.This had been notedas early
as 1843bybothCournot(1843,Section95,page 170)
and Mill (1843,Book 3, Chapter18, Section6), and
had beenextendedbyEdgeworth to parameters other
than binomial proportions (Edgeworth, 1884b,
page 204). A related developmentwas Poincare's
methodof arbitrary functions; see, e.g.,Borel (1965,
Chapter9).
These were creditableargumentsand, giventhe
imprimatur ofEdgeworth and Pearson,it is not sur-
prisingtofindacceptanceofpriorprobabilities at least
initiallyevenamongstatisticians ofFisher'sowngen-
eration.Gosset's ["Student"]discussionof the issue
in his classic 1908paperon the "Probableerrorof a
correlation coefficient"
is a goodexample.Gossetde-
scribesthe estimationproblemfor the correlation
coefficientas thatofdetermining "theprobability that
R [thepopulationcorrelation coefficient]forthepop-
ulation fromwhichthe sample is drawnshall lie
betweenany givenlimits"(Gosset,1908,page 302).
He thenadds:
It is clearthatin orderto solvethisproblemwe
must knowtwo things:(1) the distribution of
values of r [the sample correlationcoefficient]
derivedfromsamplesof a populationwhichhas
a givenR, and (2) the a prioriprobability thatR
forthepopulationlies betweenanygivenlimits.
Now (2) can hardlyeverbe known,so thatsome
arbitrary assumption mustin generalbe made...
I maysuggesttwomoreor less obviousdistribu-
tions.The firstis thatanyvalueis equallylikely
between+1 and -1, and the second that the
probability thatx is the value is proportionalto
1 - x2: this I thinkis morein accordancewith
ordinaryexperience:the distribution of a priori
probability wouldthenbe expressedbytheequa-
tiony = (?/) (1- x2).
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
253
Gosset's discussion clearlyreflectsa change in cli-
mate; "some arbitraryassumptionmust in generalbe
made;" and a nonuniformprior seems "more in ac-
cordancewithordinaryexperience."Nevertheless,his
basic view of estimationis clearlyBayesian. Nor were
the referencesto priorprobabilitiesin the statistical
literatureof this period mere lip-service:Edgeworth's
important1908 papers on maximumlikelihood were
based in part on them, and Neyman himself later
employed prior probabilities in some of his earlier
papers (Neyman and Pearson, 1928; Neyman, 1929).
(Neymanhad originallyhoped to have Pearson's name
appear as a co-authoron the second paper,but by this
time Pearson was unwillingto have his name associ-
ated in print with prior probabilities (Reid, 1982,
pages 82-85).)
Acceptance of inverse methods continued into the
1920's, when theyreceiveda powerfulassist fromthe
work of Frank Ramsey (1926). Indeed, Fisher would
appear to be the firstBritishstatisticianof any stand-
ingto publiclyattack Bayesian methods.The remark-
ably hostile reaction to his 1935 JRSS discussion
paper (Fisher, 1935) may reflectin large part the
antagonismof the Bayesian old-guardto the nouvelle
statistique.Writing as late as 1934, Neyman could
state that "until recently"it had been assumed that
the problemof statisticalestimationin samplingfrom
a population required "knowledge of probabilities a
priori"(Neyman, 1934).
Nearly half a centuryelapsed between the appear-
ance of the firstedition of Chrystal'sAlgebra(1886)
and Fisher's attacks on inverse probability.During
that period inverse methodswere debated, claims for
the theoryqualified, and caution in its use advised,
but the theoryitselfwas nevertotallyabandoned, and
there is no evidence whatever for what Fisher de-
scribed on one occasion as an abrupt and dramatic
change.Textbookscontinuedto coverthe subject(e.g.,
Coolidge, 1925; Burnside, 1928; Fry, 1928), questions
on it continued to appear on actuarial examinations
(A. Fisher, 1915,page 56), respectedstatisticianscon-
tinued to employ it (Bowley, 1926). Fisher suggests
that the mostimportantresultofVenn's criticismhad
been Chrystal's omission of inverseprobabilityfrom
his Algebra.Surely moreto the point is that virtually
everytextbookin probabilitywrittenin English dur-
ing the period 1886-1930 includesthe topic,as well as
most texts in French and German. Indeed, it is diffi-
cultto findexceptions-apart fromBertrand-at least
amongtextsofthe firstrank.Writingin 1921, Keynes
could state that "the reaction against the traditional
teaching duringthe past hundredyears has not pos-
sessed sufficientforceto displace the establisheddoc-
trine, and the Principle of Indifferenceis still very
widely accepted in an unqualified form" (Keynes,
1921, page 84).
254
Fisher was, in fact,being too modest when he as-
cribedthe demiseofinverseprobabilityto Boole, Venn
and Chrystal.The two mostimportantpersons in that
undertakingwere none otherthan Fisher himselfand
Neyman. (Thus forEgon Pearson, the inverseproba-
bilityapproach "had been foreverdiscreditedbyFisher
in his 1922 paper .. ." (Reid, 1982, page 79).) Human
nature being what it is, no matter how cogent or
convincingthe argumentsofthe opponentsof inverse
probabilitywere, until a credible alternativeto the
Bayesian methodologywas provided,any attemptto
demolishthe edificeofinverseprobabilitywas doomed
to failure(see, e.g., Pearson, 1920, page 3).
The HarvardmathematicianJulianLowell Coolidge
was perhaps merely being more candid than most
when he wrote (1925, page 100):
Why not, then, reject the formula outright?
Because, defectiveas it is, Bayes' formulais the
only thingwe have to answer certain important
questions which do arise in the calculus of prob-
ability.... Thereforewe use Bayes' formulawith
a sigh, as the only thing available under the
circumstances:
'Steyning tuk him for the reason the thief tuk
the hot stove-bekaze there was nothing else
that season.' [Kipling, Captains Courageous,
Chapter 6]
6. DISCUSSION
Paradoxically,the historyof science when written
byscientiststhemselvesis sometimesseriouslyflawed.
A typologyof possible reasons forthis suggeststwo
generalcategories,involvingsins of omissionand sins
of commission.
First and foremost,there may be simplya lack of
interest,resources,time or training.A common man-
ifestationof this is the uncriticalcopyingof earlier,
secondary,oftenhighlyflawedaccounts withoutcon-
sulting original sources. Everyone "knows," for ex-
ample, that during the Middle Ages the Ptolemaic
modelofthe solar systemwas modifiedbythe addition
ofepicycleupon epicycleto artificiallyforceagreement
with increasinglyaccurate experimentaldata. But in
reality, nothing of the kind occurred: the original
Ptolemaic model of one deferentand one epicycle
provided a remarkablygood fit to the observational
data available prior to the time of Tycho Brahe;
indeed, given the mathematical sophistication of
Ptolemy's original system, more simplified models
weretypicallyemployedthroughoutthe Middle Ages,
not more complex ones (see, e.g., Gingerich,1973,
page 95). But this misconceptionfitspopular preju-
dices about the science of the Middle Ages (see, e.g.,
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
S. ZABELL
ArthurKoestler's The Sleepwalkers,1959) and so is
repeated fromone misinformedsource to another. It
does not occur to someone to check the authenticity
of such a story,any more than it would occur to him
to check whether Einstein was responsible for the
special theoryof relativity,or whetherWatson and
Crick discoveredthe structureof DNA.
Even when a person has first-handknowledgeof
the events about which he is writing,the passage of
time may lead to a subtle erosionin the accuracywith
which those events are remembered.A notable ex-
ample is Karl Pearson's historicalaccount of correla-
tion (Seal, 1967; Plackett, 1983). As Stigler notes,
Pearson's commentary"reflects well neither upon
Pearson nor the generaltrustworthiness of the latter
recollections of great scientists" (Stigler, 1986,
page 344, n. 11).
Under the rubric of sins of commission may be
placed an interrelatedcomplex of causes including
subconsciousbias, dogmatism,sensationalismand de-
liberate distortion.Everyone "knows," for example,
that the nightbeforehe was fatallywoundedin a duel,
the unfortunateEvariste Galois stayed up feverishly
writingdown a sketch of his theoryof equations so
that it would not be lost to posterity.In realityGalois
had publishedan outlineof his resultsmonthsearlier,
and although he did write furtherdetails down the
nightbeforethe fatal duel, therewas not the urgency
oftendepicted. Reality does not make nearlyas good
a story as the piquant version in circulation. As
Rothman (1982) discusses, this is not an isolated
incident in Galois's biography: several of the best
known accounts of Galois's life (those of Bell, Hoyle
and Infeld) are marredby serious inaccuracies which
occur because of-rather than in spite of-the ability
of their authors to appreciate the technical achieve-
ments of Galois; "the misfortuneis that the biogra-
phers have been scientists" (Rothman, 1982,
page 104). Similarly,Stigler (1982) argues that many
accounts of Bayes's originalpaper are seriouslyinac-
curate;here foundationalbiases oftenled statisticians
of the statureof Pearson, Fisher and Jeffreys to mis-
read into Bayes their own viewpoints.
Fisher's account ofthe historyofinverseprobability
is marredforreasons fallingintoboth ofthese general
categories. Due perhaps in part to poor eyesight,
Fisher was never veryscholarlyin documentingpre-
vious work; this was to prove vexatious years later
when Neyman and otherswould criticizehim fornot
adequately acknowledgingEdgeworth'searliercontri-
butions to maximum likelihood (Savage, 1976,
pages 447-448; Pratt, 1976).
Nevertheless,throughouthis lifeFisher had a seri-
ous interest in historical matters. Leafing through
Todhunter, he was quick to note the Bernoulli-
 FISHER'S HISTORY OF INVERSE PROBABILITY
Montmortcorrespondenceabout the optimalstrategy
in the game of "le Her," and realized (a decade before
the workof von Neumann and Morgensternon game
theory) that a randomized strategywas appropriate
(Fisher,1934). (On the otherhand,had Fisherreferred
to Montmort's book he would have discovered an
extract of a letter from Waldegrave to Montmort
discussing the possibility of randomized strategies!
(Montmort,1713,pages 409-412).) He was oftenfond
of using an historicaldata set as the perfectpedagog-
ical foil;the entirethirdchapterof Fisher's Design of
Experiments,forexample,is centeredabout an analy-
sis ofDarwin's data on cross and self-fertilized plants.
Occasionally, the result mighteven suggesta radical
historicalreassessment,as in his article on whether
Mendel fudged his data (Fisher, 1936a; Root-
Bernstein,1983).
And what Fisher was acquainted with, he often
knew very well indeed. As Savage (1976, page 447)
notes,Fisher "was well read in the statisticalliterature
of his past," and Fisher's writingsdisplay a detailed
knowledge of Bayes, Boole, Venn, Todhunter and
Keynes. But it is a common failingto read into the
words of the past the thoughtsof the present,and to
view the evolution of history as the progressive
triumphof one's own viewpoint.This Fisher appears
to have done.
ACKNOWLEDGMENTS
The author expresses his thanks to Persi Diaconis
and Paul Meier fora numberof helpfulcommentsand
suggestionsduring the preparation of the paper, to
Elisabeth Vodola for supplyinga copy of Chrystal's
1891 paper, and to an anonymousrefereefora careful
readingof the manuscript.
REFERENCES
BACHELIER, L. (1912). Calcul des probabilites1. Gauthier-Villars,
Paris.
BERTRAND, J. (1889). Calcul des probabilites. Gauthier-Villars,
Paris. (2nd ed., 1907, reprintedby Chelsea, New York.)
BOOLE, G. (1854). An InvestigationoftheLaws of Thought.Walton
and Maberly,London. (Reprintedby Dover, New York, 1976.)
BOOLE, G. (1862). On the theoryof probabilities.Philos. Trans.
Roy. Soc. London 152 225-252. (Reprintedin G. BOOLE, Col-
lected Logical Works 1. Studies in Logic and Probability
(R. Rhees, ed.) 386-424. Open Court PublishingCo., La Salle,
Ill., 1952.)
BOREL, E. (1965). Elements of the Theoryof Probability.Prentice-
Hall, Englewood Cliffs,N.J.
BOWLEY, A. L. (1926). Measurement of the precision of index-
numbersattained in sampling.Bull. Internat.Statist. Inst. 22
6-62.
Box, J. F. (1978). R. A. Fisher: The Life ofa Scientist.Wiley, New
York.
BURNSIDE, W. (1928). The TheoryofProbability.CambridgeUniv.
Press, New York. (Reprintedby Dover, New York, 1959.)
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
255
CARNAP, R. (1950). Logical Foundations of Probability. Univ.
Chicago Press, Chicago. (2nd ed., 1962).
CHRYSTAL, G. (1886). Algebra.Adam and Charles Black, London.
CHRYSTAL, G. (1891). On some fundamentalprinciplesin thetheory
of probability. Trans. Actuarial Soc. Edinburgh (N. S.) 2
421-439.
COOLIDGE, J. L. (1925). An Introductionto MathematicalProbabil-
ity.OxfordUniv. Press. (Reprintedby Dover, New York, 1962.)
COURNOT, A. A. (1843). Exposition de la theoriedes chances et des
probabilites.Librairie de L. Hachette, Paris.
DE FINETTI, B. (1972). Probability,Induction,and Statistics: The
Art ofGuessing.Wiley,New York.
DE MORGAN, A. (1847). Formal Logic: or, the Calculus ofInference,
Necessary and Probable. Taylor and Walton, London. (Re-
printedby The Open Court Co., London, 1926.)
DONKIN, W. F. (1851). On certain questions relatingto the theory
of probabilities.Philos. Mag. (4) 1 353-368, 458-466.
EDGEWORTH, F. Y. (1884a). The philosophy of chance. Mind 9
222-235.
EDGEWORTH, F. Y. (1884b). A prioriprobabilities.Philos. Mag. (5)
18 204-210.
EDGEWORTH, F. Y. (1908). On the probable errors of frequency
constants.J. Roy. Statist. Soc. 71 381-397, 499-512, 651-678.
Addendum72 (1909), 81-90.
EDGEWORTH, F. Y. (1922). The philosophyof chance. Mind 31
257-283.
ELLIS, R. L. (1844). On the foundationsof the theoryof probabili-
ties. Trans. CambridgePhilos. Soc. 8 1-6. (Reprinted in The
Mathematicaland Other Writingsof RobertLeslie Ellis M. A.
(W. Walton, ed.). Deighton and Bell, Cambridge,1863.)
FIENBERG, S. E. and HINKLEY, D. V. (eds.) (1980). R. A. Fisher:
An Appreciation.Lecture Notes in Statist 1. Springer,New
York.
FISHER, A. (1915). The MathematicalTheoryofProbabilitiesand its
Application to Frequency Curves and Statistical Methods 1.
MathematicalProbabilitiesand Homograde Statistics,2nd ed.
Macmillan, New York, 1923.
FISHER, R. A. (1912). On an absolute criterionfor fittingfre-
quency curves. Messenger Math. 41 155-160. (Collected
Papers 1.)
FISHER, R. A. (1921). On the "probable error"or a coefficientof
correlationdeduced froma small sample. Metron 1 3-32. (Col-
lected Papers 14; contains Fisher's firstcritical comment on
inverseprobability.)
FISHER, R. A. (1922). On the mathematicalfoundationsof theoret-
ical statistics. Philos. Trans. Roy. Soc. London Ser. A 222
309-368. (CollectedPapers 18.)
FISHER, R. A. (1930). Inverse probability.Proc. CambridgePhilos.
Soc. 26 528-535. (CollectedPapers 84.)
FISHER, R. A. (1934). Randomisation, and an old enigma of card
play. Math. Gaz. 18 294-297. (CollectedPapers 111.)
FISHER, R. A. (1935). The logic of inductive inference.J. Roy.
Statist. Soc. 98 39-54. (CollectedPapers 124.)
FISHER, R. A. (1936a). Uncertaininference.Proc. Amer.Acad. Arts
Sci. 71 245-258. (CollectedPapers 137.)
FISHER, R. A. (1936b). Has Mendel's workbeen rediscovered?Ann.
Science 1 115-137. (CollectedPapers 144.)
FISHER, R. A. (1951). Statistics. In Scientific Thought in the
TwentiethCentury(A. E. Heath, ed.) 31-55. Watts, London.
(CollectedPapers 242.)
FISHER, R. A. (1955). Statistical methodsand scientificinduction.
J. Roy. Statist. Soc. Ser. B 17 69-78. (CollectedPapers 261.)
FISHER, R. A. (1956). StatisticalMethods and ScientificInference.
Hafner,New York. (2nd ed., 1959; 3rded., 1973; page references
are to the 3rd ed.)
FISHER, R. A. (1958). The natureof probability.CentennialReview
2 261-274. (CollectedPapers 272.)
256 S. ZABELL
FISHER, R. A. (1971-74). Collected Papers of R. A. Fisher 1-5
(J. H. Bennett,ed.). Univ. Adelaide.
FRY, T. C. (1928). Probabilityand Its EngineeringApplications.van
Nostrand,New York.
GINGERICH, 0. (1973). Copernicus and Tycho. ScientificAmerican
229 86-101.
GosSET, W. S. (1908). Probable errorof a correlationcoefficient.
Biometrika6 302-310.
HAILPERIN, T. (1976). Boole's Logic and Probability.North-Holland,
Amsterdam.
HARDY, G. F. (1889). Letter. Insurance Record 457. (Reprinted,
Trans. FacultyActuaries8 180-181, 1920.)
JEFFREYS, H. (1939). Theoryof Probability.Clarendon Press, Ox-
ford.(2nd ed., 1948; 3rd ed., 1967.)
JEVONS, W. S. (1877). The PrinciplesofScience, 2nd ed. Macmillan,
London.
KEYNES, J. M. (1921). A Treatise on Probability.Macmillan,
London.
KOESTLER, A. (1959). The Sleepwalkers.Macmillan, New York.
MILL, J. S. (1843). A SystemofLogic, Ratiocinativeand Inductive,
Being a Connected View of the Principlesof Evidence and the
Methods of ScientificInvestigation.John W. Parker, London.
(Many later editions.)
MONTMORT, P. R. (1713). Essai d'analyse sur les jeux de hazards,
2nd ed. Jacques Quillan, Paris. (lst ed., 1708.)
NEYMAN, J. (1929). Contribution to the theory of certain test
criteria.Bull. Internat.Statist.Inst. 24 3-48.
NEYMAN, J. (1934). On the two differentaspects of the represent-
ative method: The method of stratifiedsampling and the
method of purposive selection. J. Roy. Statist. Soc. 97
558-625.
NEYMAN, J. and PEARSON, E. S. (1928). On theuse ofinterpretation
of certain test criteria for purposes of statistical inference.
Biometrika20 175-240, 263-294.
PASSMORE, J. (1968). A Hundred Years of Philosophy,2nd ed.
Penguin, New York.
PEARSON, K. (1892). The Grammarof Science. Walter Scott, Lon-
don. (2nd ed., 1900; 3rd ed., 1911.)
PEARSON, K. (1907). On the influenceof past experienceon future
expectation.Philos. Mag. (6) 13 365-378.
PEARSON, K. (1920). The fundamentalproblemof practical statis-
tics. Biometrika13 1-16.
PLACKETT, R. L. (1983). Karl Pearson and the chi-squared test.
Internat.Statist. Rev. 51 59-72.
PORTER, T. M. (1986). The Rise ofStatisticalThinking:1820-1900.
PrincetonUniv. Press, Princeton,N.J.
PRATT, J. W. (1976). F. Y. Edgeworth and R. A. Fisher on the
Corment
Robin L. Plackett
Sandy Zabell deservesour thanksfordiscovering
furtherdetails of what Boole, Venn and Chrystal
forexplain-
wroteonthesubjectofinverseprobability,
Robin L. Plackett is Emeritus Professorof Statistics,
Universityof Newcastle upon Tyne. His mailing ad-
dressis: 57 Highbury,Newcastleupon TyneNE2 3LN,
UnitedKingdom.
This content downloaded from 129.173.72.87 on Wed, 2 Oct 2013 23:47:10 PM
All use subject to JSTOR Terms and Conditions
efficiencyof maximum likelihood estimation.Ann. Statist. 4
501-514.
RAMSEY, F. P. (1926). Truth and Probability.In The Foundations
of Mathematics and Other Logical Essays (R. B. Braithwaite,
ed.) 156-198. Routledge and Kegan Paul, London (1931).
REID, C. (1982). Neyman-From Life. Springer,New York.
RoOT-BERNSTEIN, R. S. (1983). Mendel and methodology.History
ofScience 21 275-295.
ROTHMAN, T. (1982). Genius and biographers:the fictionalization
of Evariste Galois. Amer.Math. Monthly89 84-106.
SALMON,W. C. (1981). JohnVenn's LogicofChance. In Probabilistic
Thinking,Thermodynamicsand the Interactionof the History
and Philosophy(J. Hintikka,D. Gruenderand E. Agazzi, eds.)
2 125-138. Reidel, Dordrecht.
SAVAGE, L. J. (1976). On re-readingR. A. Fisher (withdiscussion).
Ann. Statist. 3 441-500.
SEAL, H. L. (1967). The historicaldevelopmentof the Gauss linear
model. Biometrika54 1-24.
SHAFER, G. (1976). A Mathematical TheoryofEvidence. Princeton
Univ. Press, Princeton,N.J.
STIGLER, S. M. (1978). Francis Ysidro Edgeworth,statistician(with
discussion). J. Roy. Statist. Soc. Ser. A 141 287-322.
STIGLER, S. M. (1982). Thomas Bayes's Bayesian inference.J. Roy.
Statist. Soc. Ser. A 145 250-258.
STIGLER, S. M. (1986). The HistoryofStatistics:The Measurement
of UncertaintyBefore 1900. Harvard Univ. Press, Cambridge,
Mass.
TODHUNTER, I. (1865). A Historyof the Mathematical Theoryof
Probability.Macmillan, London. (Reprinted by Chelsea, New
York, 1949.)
VENN, J. (1866). The Logic of Chance. Macmillan, London. (2nd
ed., 1876; 3rd ed., 1888; reprintedby Chelsea, New York,
1962.)
VON KRIES, J. (1886). Die Prinzipien der Wahrscheinlichkeits-
rechnung.Eine Logische Untersuchung.Freiburg. (2nd ed.,
Tubingen, 1927.)
VON WRIGHT, G. H. (1941). The Logical Problem of Induction.
Finnish LiterarySoc., Helsinki. (2nd rev. ed. Macmillan, New
York, 1957.)
WHITTAKER, E. T. (1920). On some disputedquestions ofprobabil-
ity (with discussion). Trans. FacultyActuaries77 163-206.
WHITWORTH, W. A. (1897). DCC Exercises in Choice and Chance.
(Reprintedby Hafner,New York, 1965.)
WHITWORTH, W. A. (1901). Choice and Chance, 5th ed. GeorgeBell
and Sons, London.
WINSOR, C. P. (1947). Probability and listerism.Human Biology
19 161-169.
ing why Fisher could not have relied on them to
provide consistent arguments against this form of
statistical inferenceand for an analysis of how far
Fisher's claims concerningthe eclipse of inverseprob-
ability are justified. Like everythingelse connected
withFisher,mattersare indeed complex,and Zabell's
paper providesa good topic fordiscussion.
At the height of his career, Fisher was certainly
familiar with what mattered in developments of