Mind Association

On Conditionals
Author(s): Dorothy Edgington
Source: Mind, New Series, Vol. 104, No. 414 (Apr., 1995), pp. 235-329
Published by: Oxford University Press on behalf of the Mind Association
Stable URL: http://www.jstor.org/stable/2254793 .
Accessed: 27/01/2015 18:12

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 [email protected].

Oxford University Press and Mind Association are collaborating with JSTOR to digitize, preserve and extend
access to Mind.

http://www.jstor.org

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals,
DOROTHYEDGINGTON

The ability to think conditionalthoughtsis a basic partof our mental

equipment.A view of the worldwouldbe an idle, ineffectualaffairwith-
out them.There'snot muchpointin recognisingthatthere'sa predatorin
yourpathunlessyou also realisethatif you don'tchangedirectionpretty
quicklyyou will be eaten.
Happily,we handleifs with ease. Naturally,we sometimesmisjudge
them, and sometimes don't know what to think. But we know what it
wouldtaketo be in a positionto thinkor say thatB if A, whatwouldcount
for or againstsuchjudgements,how they affect what we shoulddo and
whatelse we shouldthink.They causeus no unduepracticaldifficulty.
The theoryof this practiceis anotherstory.Judgedby the qualityand
intensity of the work, theorising about conditionalshas flourishedin
recent years-bold, fertile ideas developed with ingenuityand rigour,
hithertounnoticedphenomenaobservedandexplained,surprisingresults
proved.But consensushas not emerged.Not just aboutdetails,but about
fundamentals,almosteverythingis at issue. Is a unifiedtheorypossible,
or are thereirreduciblydifferentkindsof "if"? If the latter,whatmarks
thedistinctionbetweenkinds,andwhichexamplesbelongtogether?Is the
core of a theorya thesis aboutwhatmakesa conditionalstatementtrue?
Thosewho supposeso disputeaboutthekindof truthconditionsinvolved;
othersthinkit is a mistakenpresumptionthatconditionalsarepartof fact-
statingdiscourse,evaluablein termsof truth.Given these disputes,it is
unsurprisingthattherearedisagreementsaboutwhich inferencepatterns
involvingconditionalsare valid. Thereis even dissentaboutthe logical
formof conditionals:we arealreadytheorisingwhenwe representa con-
ditionalas a particularmode of combiningtwo simplerpropositionsinto
one, andthis representationhas been questioned.
'I am extremelygratefulto the BritishAcademyfor the awardof a Research
Readership,duringthe tenureof whichmostof my workon this paperwas done.
I amalso indebted,for commentsanddiscussion,to KeithHossack,RautlOrayen,
DavidPapineauandScottSturgeon.Otherhelpfulinfluences,whenI was writing,
weredraftsof JonathanBennett'srecentwork(1995), MichaelFirestone'sthesis,
anda manuscriptby the late MichaelWoods.

Mind, Vol. 104. 414. April 1995 ? Oxford University Press 1995

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
236 DorothyEdgington

1. One theory or two?

L.L.Somethingmustbe said at the outsetaboutthe classificationof con-

ditionalsinto kinds, for some theoriesaddressone kind, some another.
Traditionally,"indicativeconditionals"have been distinguishedfrom
"subjunctiveconditionals"or "counterfactuals" (these lattertermsbeing
used interchangeably). Some worksconcernconditionalsof these forms:
(1a) If the gardenerdidn'tdo it, the butlerdid;
(lb) If the gardenerdoesn'tdo it, the butlerwill.
Forinstance,W.V. Quine,in Methodsof Logic, writes"thecontrafactual
conditionalis best dissociatedfromthe ordinaryconditionalin the indic-
ative mood ... We shall not recurto it here"(1952, p. 21). Otherworks
concernthose like
(1c) If the gardenerhadnot done it, the butlerwouldhave;
(Id) If the gardenerwere not to do it, the butlerwoulddo it.2
Forinstance,DavidLewis'sCounterfactuals. "Icannotclaimto be giving
a theoryof conditionalsin general",he says (1973, p. 3).
"Thereare differentkindsof conditionals"can be takenas an innocu-
ous remark,inevitablytrue.But the traditionaldistinctionis less between
two speciesof a genus,thanbetweentwo genera,requiringseparatetreat-
ment.This can surprise,for, it would seem, the samplesentencesabove
could each be used to expressthe same conditionalthoughton different
occasions.Changingthe example:we are arguingaboutwhether,if you
eat this apple,you will be ill. Youthrowit awayin disgust.Ourargument
continuesunabated-about whetheryou would have been ill if you had
eatenit. Wedo not appearto havechangedthe topicof debate.Justbefore
throwing it away, you say "If I were to eat it, ..."; someone who left our
companyearliersays lateron "I'mconvincedthatif he ate the apple,he
was ill".The bipartiteapproachneeds some explanation.
Partof the explanationis, I think,historical."Contrary-to-fact"or "sub-
junctive"conditionalsfirst surfacedas a problemin the philosophyof
science, for the attemptby logical empiricists to regiment scientific
languageusingFrege'spowerfulnew logic -to do for sciencewhatFrege
and Russell'had done for mathematics.At the heart of this logic is a
treatmentof the conditionalof remarkablesimplicityandclarity:a condi-
tionalis trueif and only if it is not the case thatit has a trueantecedent
anda falseconsequent.3 Whenit cameto analysingdispositionalpredicates
2Throughoutthis section the letters a-d, following the numbers, indicate the
form of the conditional.
3See Frege's Begriffsschrift, ?5 (1960, pp. 5-6); his letter to Husserl translated
in Frege (l980,p. 69); and his "Introductionto Logic" in Frege (1979, p. 186). (In

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 237

like "soluble"and "fragile",a different kind of conditional made its

presencefelt. Being fragileis being such as to breakif dropped.Frege's
analysiscannotbe used here;for it cannotexplainwhy, if the vase is not
droppedat a particulartime,thatdoesn'tsettlewhetherthe vase is fragile
at thattime-whether it wouldbreakif it weredropped.RudolfCarnap's
"TestabilityandMeaning"(1936) was a valiantattemptto deal with this
problemin termsof Fregeanlogic. Domesticatingthe non-Fregeancon-
ditionalbecamea majorproblem.
If a theoryfits some butnot all of the data;the lesson mightbe thatthe
dataare not amenableto uniformtreatment;it might equallybe thatwe
need a bettertheory.But there is more to be said in favourof dualism
about conditionals, independently of prior theoretical commitment.
ErnestAdams(1970) madethe pointwiththis strikingpairof examples:
(2a) If Oswalddidn'tkill Kennedy,someoneelse did;
(2c) If Oswaldhadn'tkilledKennedy,someoneelse wouldhave.
Everyonewho knowsof Kennedy'sassassinationagreeswith (2a);many
such people dissentfrom (2c). Takesomeonewho thinksOswalddid it,
actingalone. "Butwhat is the case if he didn'tdo it?"gets one answer.
"Butwhatwouldhavebeenthe case if he hadn'tdoneit?"gets a different
answer.(Like any good philosophicalexample,this is no isolatedcase.
Once you graspits structure,you have a recipe for constructingindefi-
nitely many such pairs,which I shall call "OKcases".)"Thereforethere
really are two differentsorts of conditional",says Lewis (1973, p. 3),
commenting on this phenomenon,"not a single conditional that can
appearas indicativeor as counterfactualdependingon the speaker'sopin-
ion aboutthe truthof the antecedent".Hereis how one mightfill out the
argumentfor this conclusion.Considerthe two past-tensesentences:
0: Oswalddidn'tkill Kennedy
S: Someoneelse killed Kennedy
andconsiderthe sentenceframes:
If it is the case that ..., it is the case that ....
If it were the case that ..., it would be the case that ...

the lattertwo passages,whicharecommentson the first,Fregeexplainsthe con-

ditionalas I did above. In the Begriffsschrifthe has "affirmed" and "denied"in
placeof "true"and"false"(p. 5). It is difficultto interpretthis plausibly.On the
followingpagehe has "tobe affirmed"and"tobe denied",which,in the context,
can moreplausiblybe interpretedas "true"and"false".)
4 Lewis (1973, pp. 2-3) explainsthe counterfactual connectivein termsof this
secondsentenceframe.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
238 DorothyEdgington

Substitute the two sentences in the two sentence frames, and you have
regimentations of (2a) and (2c). Replace each sentence frame by a symbol
to be written between the sentences, say "-", and "o-+",respectively. So
we have "O->S", "O o-*S". One may accept "->S", yet reject "0 o-+S".
So " -" and" o-+"donYmeanthe same.QED.
This argument for two meanings of "If ..." is resistible. Our regimen-
tations may have misrepresented the syntactic structure of the two sen-
tences. Even when a single sentence has a true and a false reading, it does
not follow that one of its semantic components is ambiguous. Consider
The Prime Minister has never been a woman.
That has a true and a false reading, but it is a case of syntactic, ratherthan
semantic, ambiguity: the sentence may be read as structuredin different
ways, though each of its components has a uniform meaning. Or consider
I could have been in New York today; but I can't, now, be in New
York today.
This is not an example of an ambiguity in a modal term: something was
possible, which is no longer possible.
This last example is instructive. V. H. Dudman5 has convinced many
that (2c): "If Oswald hadn't killed Kennedy, someone else would have" is
simply the past tense of
(2b) If Oswald doesn't kill Kennedy, someone else will.
""Would have" is the past tense of "will", as "could have" is the past tense
of "can"; the verb forms in the antecedents typically indicate that they
concern a time earlier than the consequents.6
The analysis of the counterfactual as a past-tense indicative could be a
step in the direction of monism. If we can explain how the evaluation of
a conditional depends on time, we can explain the OK cases without mul-
tiplying senses of "if". But this is not the moral drawn by Dudman and
others. They remain dualists, and retain the view that (2a) and (2c) are dif-
ferent kinds, but maintain that (2b) has been wrongly classified: it is of a
kind with (2c), not with (2a).7
The OK phenomenon does not supportthis new line, however, for it can
be used to drive a wedge between future indicatives and counterfactuals
as well as past ones. You think that such-and-such will happen. You can
distinguish the questions: "But what if it doesn't?" (i.e., what if you're
wrong in thinking it will?); and "But what if it were not going to?" (retain-
I See e.g. Dudman(1983, 1984a, 1988, 1989).Adams(1975), Ayers(1965)
andEllis (1984) also treatthe "counterfactual"
as a pasttenseconditional.
6This is not how Dudmanwouldputit. He does not carefor the terms"ante-
cedent"and"consequent" (1986, 1988).
7 See Dudman(1984), Smiley (1984), Bennett(1988), Mellor(1993).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 239

ing yourbelief thatit will). For instance,thereare two prisoners,Smith

andJones.Wehavepowerfulevidencethatone of themwill tryto escape
tonight.Smithis a docile,unadventurous chap,Jonesjusttheopposite,and
very persistent.We are inclinedto thinkthat it is Jones who will try to
escape.We have no reasonto accept:
(3c) If Joneswerenot to try to escapetonight,Smithwould.
However,we couldbe wrongin thinkingthatit is Joneswho will escape:
(3b) If Jonesdoesn'ttry to escapetonight,Smithwill.
Anotherexample:I'm beingchasedthroughenemyterritory,anda warn-
ing light on my (eccentric)car indicatesthateitherI am aboutto runout
of fuel, or the radiatoris aboutto boil over. I'm prettysureit's the fuel.
Bother!If I hadn'tbeen going to run out of fuel, I would get away. Of
course,I couldbe wrongaboutthe fuel. Butthen,if I don'trunoutof fuel,
the radiatorwill boil over.
The differencemarkedby the OKcases seemsto be the traditionalone.
Butit maybe morelikethedifferencebetweenmaturecheddarandfreshly-
madecheddarthanthedifferencebetweenchalkandcheese.As timepasses
but relevant information stays the same, "If he eats the apple, ...", "If he
were to eat it, ...", "If he ate it ..." and "If he had eaten it, ..." may all
expressthe sameconditionalthought.But the passingof time may bring
new relevantinformation:"If he ate it, it did him no harm;but if he had
eatenit, he wouldhave been ill".8Furtherargumentwill have to wait on
whetherthis differencecan be explainedwithina unifiedaccountof "if".
1.2. The terminologyfor the traditionaldistinctionis less thansatisfac-
tory.Forthosewho acceptthe distinction,thisis a minorirritant;forthose
who don't, it is a symptomof confusion. Lewis says "Youmay justly
complainthatmy title 'Counterfactuals' is too narrowfor my subject.I
agree, but I know no better .... The title 'Subjunctive Conditionals'
wouldnot have delineatedmy subjectproperly[either]"(1973, pp. 3-4).
Longbefore,RoderickChisholmannouncedthathe woulduse "subjunc-
tive"and "contrary-to-fact"interchangeably, althoughthey were not co-
extensive."Neithertermis adequate"(1946, p. 482). MichaelAyerscom-
plainedthatit was "as if he had said that some mammalsare not carni-
vores and some carnivoresarenot mammals,but he wishedto talkof an
importantclass of animalto which ... he wouldreferindiscriminatelyas
mammalsand as carnivores"(1965, p. 348). JonathanBennett's"Fare-
8Although thecase of SmithandJonesconcernswhatwill happen,thereis still
a temporaldifferencebetweenwhatis neededto establishthatSmithwill if Jones
doesn't,andwhatis neededto establishthatSmithwouldif Joneswerenotto. The
latterrequiresthat before the evidence arose, the judgementthat Smithwill if
Jonesdoesn'twouldhavebeencorrect.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
240 DorothyEdgington

well to the PhlogistonTheoryof Conditionals"(1988) also arguesthatthe

terminologicalinadequacyis a sign thatourtheoriesarein bad shape.
A true counterfactualmay have a true antecedentand consequent,
accordingto acceptedusage. Consider"If you had droppedit, it would
havebroken"."You'reright-I did dropit, andit broke,but I did such a
marvellousrepairjob, you never could tell". Still, the idea behindthe
name is that counterfactualsarefor talking aboutunrealisedpossibili-
ties -we use themwhenwe thinkthe antecedentis false. But thereis one
importantuse of the "counterfactual"formwhichdoes not fit this pattern.
AlanRoss Anderson(1951) gave the exampleof a doctor'ssaying"Ifhe
had taken arsenic, he would have shown just these symptoms [those
which he in fact shows]".The doctorcould not convey the same thing
with "If he took arsenic,he is showingjust these symptoms".This is no
one-off example:"Abus is coming.""Howdo you know?"(for we can't
see the oncomingtraffic)."Peoplein line are pickingup theirbags and
inching forward -and that's what they would be doing if a bus were com-
ing."(Not:that'swhatthey aredoingif a bus is coming.)Thesecases are
importantas ingredientsin "inferenceto the best explanation"and in
Bayesianreasoning:whichhypothesisH is suchthatwhatwe do observe
is whatwe wouldexpectto observe,if H were true?
(Conversely,it is sometimesthe indicativewhichis neededto express
disbeliefin the antecedent:"If he took arsenic,he's showingno signs".
Not: "Ifhe hadtakenarsenic,he wouldbe showingno signs".)
Would"SubjunctiveConditionals"have been a bettertitle for Lewis's
book?Dudman(1988) andBennett(1988) arguethatthe "hadbeen"and
"would"are a matterof tense, not mood. They quote grammarians who
pourscornon the ideathatthe subjunctivehas any serioususe in English.
Grammariansare no more proneto unanimitythanphilosophers,how-
ever: Fowler's Modern English Usage gives the examples
If he heard, he gave no sign (heard and gave past time); and If he
heard, how angry he would be! (heard and would be, not past
time,bututopia,the realmof non-factor the imaginary);the first
heardis indicative,the secondsubjunctive.(1965, p. 597)
Evenif thisis right,we lacka good explanationof why someconditionals
requirethis mood andothersforbidit. Furtherilluminationis unlikelyin
advanceof some theorising.It will cause less confusionand no greater
offenceif I stickto the labels"indicative"and"counterfactual"whendis-
cussingtheoriesaddressingone or the otherside of the traditionaldivide.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 241

2. Truthconditions of thefirst kind

2.1. Thereareconditionalquestions,commands,expressionsof wish, etc.,

as well as conditional statements;but we follow the methodology of
mainstreamphilosophyof languageif we assumethatan understanding
of fact-stating discourse is our first task. Put counterfactualsaside.
Assumethatthe conditionalis a devicefor constructinga proposition,apt
for truth,out of two componentpropositions,aptfor truth.And it is a sys-
tematicdevice:if you understandany conditional,you understandevery
conditionalwhose componentsyou understand.Still followingthe main-
stream,assumethatunderstanding a sentenceis knowingunderwhatcir-
cumstancesit wouldbe true.Understandinga sub-sententialmeaningful
componentis knowingwhatcontributionit makesto the truthconditions
of the sentencesin which it occurs. Some such componentsare used to
constructcomplex sentencesout of simplersentences.Let M be: Mary
went to Paris.Let J be: Johnwentto Paris.Consider
(la) It is not the case thatM;
(lb) It is possiblethat,probablethat,importantthat,relevantthatM;
(2a) M andJ, M or J;
(2b) M beforeJ, M becauseJ.
(la) and(2a) areoperatorswith a peculiarlysimpleproperty:in any pos-
siblecircumstance,the truthvalueof the complexsentenceis fixedby the
truthvalue(s)of the simplesentence(s).Thuswe writetruthtables,show-
ing the truthvalueof the whole for differentpossiblecombinationsof the
truthvaluesof the parts;they arethe truth-functional sentenceoperators.
(lb) and (2b) lack this simpleproperty.The truthvalues of the partsare
not alwayssufficientto determinethe truthvalue of the whole. They are
non-truth-functional sentenceoperators.We need to examinethe thesis
that"if" is truth-functional.
2.2. Thereare signs of it in ancienttimes, andit is sometimescalled the
PhilonianconditionalafterPhilo of Megaraof the 4th centuryBC, but it
is to Fregethatwe owe its role in currentthinkingaboutconditionals.It
is a cornerstoneof his systemof logic, takenup enthusiasticallyby Rus-
sell (who called it "materialimplication"),Wittgensteinand the logical
positivists,and is now foundin every logic book. It is the firsttheoryof
the conditionalthat studentsof philosophyencounter.And it has many
defenders.Wehavealreadyseen the one-liner:"IfA thenB" is trueif and
only if it's not the case thatA is trueandB is false. It is thusequivalentto
-(A&-B) andto -AvB. "ADB"has,by stipulation,thesetruthconditions.
The substantivequestionis whetherthis is an adequaterenderingof "IfA,
B"

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
242 DorothyEdgington

It is easyto see thatif"if" is truth-functional,

thisis therighttruth-func-
tion to assign it. For no one doubts9thata conditionalis sometimes true
whenthe truthvaluesof its componentsare(true,true),or (false,true),or
(false, false). Giventruth-functionality, it follows thatit is always truein
these circumstances-for the truth-valuesof the componentsfix the truth
value of the whole. Takea conditionalwhichis truecome whatmay,for
example"IfMaryandJohnarebothin Paris,thenMaryis in Paris".The
componentsaresuchthatit is impossiblethatit has a trueantecedentand
false consequent.But the other three combinationsare possible, and
whicheverobtains,the conditionalis true. Given truth-functionality, it
follows thatwheneverone of these threecombinationsobtains,a condi-
tionalis true.
2.3. But is "if" truth-functional? Thereare powerfulargumentsthat it
mustbe. No one deniesthat"IfA, B" entails-(A&-B), whichis equiva-
lent to -A vB. If the converse entailmentholds, the truth-functional
accountis right.Gettingthe negationsigns in moredigestibleplaces,the
issue is equivalentto whether(i) AvB entails"If -A, B";or (ii) -(A&B)
entails"IfA, -B". I0But surelythey do! Knowingjust thatat least one of
the propositions,A, B, is true,is enoughto inferthatif A is not true,B is
true;and(ii) knowingjust thatA andB arenotbothtrueis enoughto infer
thatif A is true,B is not. For example:(i) havingeliminatedall but two
suspects,I'm surethateitherthe gardeneror the butlerdid it. So, if the
gardenerdidn't do it, the butlerdid (waterthe aspidistra,that is); (ii)
knowingthatMaryandJohnwerenot boththere,I inferthatif Marywas
there,Johnwas not.
Puttingthe matterthe otherway round,supposeAvB did not entail"If
-A, B" (butthe propositionsarecompatible).Thenthese aretwo distinct
possibilities:
AvB IfA, B
1. T T
2. T F
Supposeyou are certainthatone of these two possibilitiesobtains-but
minimallyso: you have eliminated-A&--AB,nothingmore.This would
not be enoughfor certaintythatif -A, B, becausethe possibilityat line 2
would be compatiblewith your information.But, we have seen above,
minimalcertaintythatAvB is enoughfor certaintythatif -A, B. Onlythe
9 No one who speaksof truthfor conditionalsat all, thatis.
0O(i) Let A = -C. Then AvB entails "If - A, B" iff - CvB entails "If -_-_ C,
B", i.e. "If C, B" (given doublenegationelimination);(ii) Let B = --D; then _
(A&B) entails "If A, - B" iff - (A&-,D) entails "If A, -__-nD",i.e. "If A, D". (In-
tuitionistworriesaboutdoublenegationeliminationcan be waivedby assuming
thatthe propositionsaredecidable.)

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 243

truth-functional truthconditionsget this right:any strongertruthcondi-

tions get this wrong.
Again,hereis a littleproofof one of the crucialentailments.Wemake
threeassumptions:(i) -(A&B); (ii) A; (iii) B. We derivea contradiction.
So, keeping assumptions (i) and (ii), we derive -B. So, by Conditional
Proof,keepingassumption(i), we derive"IfA, -B".
2.4. So what'sthe snag?Well, it seems strangeto say thatthe falsity of
"Sheate the apple",is sufficientfor the truthof "If she ate the apple,she
was ill", as it is on this account.(-A entails -(A&-B) for any B; let B
be: she was ill). And this kindof exampleis the sourceof a catalogueof
oddities.But perhapsit seems strangefor the followingreason.Whenwe
consultourintuitionsaboutthe inferencefrom"Shedidn'teat the apple",
we imagine ourselves certainof that premiss.Then we don't have any
serioususe for a conditionalthatbegins"Ifshe ate the apple".If a theory
which servesus well most of the time has the consequencethatall such
uninterestingconditionalsare true,perhapswe can and shouldlive with
thatconsequence.It is too much-or maybetoo little-to expectourthe-
ories to matchordinaryusage perfectly.Perhaps,in the interestsof sim-
plicity andclarity,we shouldreplace"if' with "D".
We shouldnot. The unacceptabilityof the inferencefrom-A to "IfA,
B"emergesmostclearlyin thecontextof beliefswhichareless thancertain.
Theproblemwasinvisibleto FregeandRussell(amongmanyothers):their
maintargetwas mathematicalreasoning;holdingbeliefs on less-than-cer-
tain groundswas not in theirmainline of business.The worstdefectsof
the truth-functional conditionaldon't show up in mathematics.
I shalluse "thinkthat","believe"and"disbelieve"in sucha way as not
to implycertainty.If youbelieveP, anddisbelieveC, andthereis a simple,
decidable,valid argumentfromP to C, yourbeliefs areirrational.I have
in mind things like: believing that somethingis squarebut disbelieving
thatit has4 sides;believingthatJohnandMaryarein Parisbutdisbeliev-
ing thatJohnis in Paris.If P entailsC, thereis no way thatP can be true
withoutC beingtrue.If the entailmentis obvious,you shouldnotbe more
confidentthatP is truethanyou arethatC is true.
WhenI think,but am not certain,that-nP, it is not at all uninteresting
or unimportant to contemplatewhatis trueif P. For example,(i) I think
thatmy husbandisn't homeyet. But if he is, he'll be worriedaboutwhere
I am. So I shouldtry to phone.Compare(ii): I thinkthatthe Queenisn't
homeyet (at BuckinghamPalace,thatis). Butif she is, she'll be worrying
aboutwhere I am. So I shouldtry to phone.The firstthoughtsare sane
enough, the second a sign of madness. Not so on the truth-functional
account. Suppose, having read in the newspaperof her day's engage-
ments,I'm about90%certainthatthe Queenisn't athomeyet (-,Q); then

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
244 DorothyEdgington

I mustbe at least 90%certainthatat least one of the propositions{_Q,

W} is true,i.e. at least 90%certainthat-_QvW,i.e. at least 90%certain
thatif she is at home,she is worryingaboutmy whereabouts(on thetruth-
functionalreadingof thatthought).Someonewho believes-Q butdisbe-
lieves "If Q, W' (on this reading)is makingan IncrediblyGrossLogical
Error.Forto disbelieveQDW,i.e. -,(Q&-_W),is to believe its negation,
Q&-_W.How can anyonebe so stupidas to believe Q&-_Wyet disbelieve
Q, i.e. believe -_Q?
Contraryto this account, any sane ordinarysubjectnot on intimate
termswith royalty,who thinksthe Queenisn't homeyet, rejectsthe con-
ditional"Butif she is, she'll be worriedaboutwhereI am".Wedo notuse
conditionalsas this accountwouldhaveit. But thatempiricalobservation
is not the mainpoint,which is this: we would be intellectuallydisabled
withoutthe abilityto discriminatebetweenbelievableand unbelievable
conditionalswhose antecedentswe think are unlikely to be true. The
truth-functional accountdeprivesus of this ability:to judgeA unlikelyis
to commitoneself to the probabletruthof ADB.
2.5. In his WilliamJameslectures,"Logicand Conversation", delivered
in 1967, H. P. Gricedefendedthe truth-functional account,emphasising
the importanceof distinguishingthe false from the misleading-but-true
(see Grice,1989).Therearemanyways of speakingthe truthyet mislead-
ing youraudience,given the standardsto whichyou areexpectedto con-
form in conversation.One way is to say somethingweakerthan some
otherrelevantthingyou are in a positionto say. Considerdisjunctions.I
amaskedwhereJohnis. I'm surehe's in thebar,andI knowhe nevergoes
nearlibraries.Inclinedto be unhelpfulbutnot wishingto lie, I say "He's
eitherin the baror in the library".My hearernaturallyconcludesthatthis
is the most preciseinformationI am in a positionto give, and also con-
cludesfromthe truth(let us assume)thatI told him "Ifhe's not in the bar
he's in the library".The conditional,like the disjunction,accordingto
Grice,is trueprovidedhe is in the bar,but misleadinglyassertedon that
ground.
Again:"Youwon't eat those and live", I say of some wholesomeand
delicious mushrooms-knowing that you will now leave them alone,
deferringto my expertise.I told no lie -for indeedyou don't eat them-
butof courseI misledyou. (Lewis 1976, p. 143, uses this example.)
Gricedrewattention,then,to situationsin whicha personis justifiedin
believing a proposition,which would neverthelessbe an unreasonable
thingfor the personto say, in normalcircumstances.His lesson was salu-
taryin manyareasof philosophy:the oddityof remarkingundernormal
conditionsof observation(e.g.) "Thepillar-boxseems red to me", does
not showthatthatsentenceis false, or truth-value-less,or meaninglessin

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 245

thatcontext. Theremarkis true,butmisleadingunlessyou havea reason

for doubtingthatit is red. Gricealso explainscorrectlythe behaviourof
disjunctionsandnegatedconjunctions.BelievingthatJohnis in the bar,I
can't consistentlydisbelieve the proposition"He'seitherin the baror in
the library";if I have any epistemicattitudeto thatproposition,it should
be one of belief, howeverinappropriate it is for me to assertit. Similarly
for "Youwon'teat thoseandlive"whenI believeyou won'teatthem.But
the difficultieswith the truth-functional conditionalcannotbe explained
away in termsof what is an inappropriate conversationalremark.They
ariseat the level of belief. BelievingthatJohnis in the bardoes not make
it logically impermissibleto disbelieve"if he's not in the barhe's in the
library".Believingyou won't eat them,I may withoutirrationalitydisbe-
lieve "if you eat them you will die". Believing thatthe Queen is not at
home,I may withoutirrationalityrejectthe claimthatif she's home, she
will be worriedaboutmy whereabouts.As factsaboutthe normsto which
peopledefer,theseclaimscanbe tested.12But,to reiterate,the mainpoint
is not the empiricalone. We need to be able to discriminatebelievable
from unbelievableconditionalswhose antecedentwe think false. The
truth-functional accountdoes not allow us to do this.
P.F. Strawson(1986) arguesthatif Griceis rightaboutindicativecon-
ditionals,his thesis shouldbe, andcouldbe, extendedto counterfactuals.
He gives the examples:
Remarkmade in the summerof 1964: "If Goldwateris elected,
thenthe liberalswill be dismayed".
Remarkmade in the winter of 1964: "If Goldwaterhad been
elected,thenthe liberalswouldhave been dismayed";
andcommentsthat"theleast attractivethingthatone could say aboutthe
difference between these two remarks is that ... 'if ... then...' has a dif-
ferentmeaningin one remarkfromthe meaningwhichit has in the other"
(p. 230).
Strawsonsuggeststhatthe Griceanstorycanbe extendedto the second
remark.For,if it is madein a contextin whichit is knownthatthe ante-
cedentis false(orequally,if theformof theremarkconventionallysuggests
thatthe antecedentis false), then,on the hypothesisthatit is truth-func-
" Austin, Ryle, Wittgenstein and others were prone to argue in this way about
various important philosophical concepts. The first chapter, "Prolegomena", of
Grice (1989) discusses many examples.
12 I am not talking about cases where you are certain that the antecedent is
false, which are difficult to assess; but about cases where you think, but are not
completely certain, that the antecedent is false.
A good enough test is to take a co-operative subject, who understandsthat you
are merely interested in what she believes, as opposed to what would be a reason-
able remarkto make; and note which conditionals she assents to.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
246 DorothyEdgington

tional,theheareris boundto look for someotherpointto its utterance,and

will concludethatthespeakermusthavejustthesortof groundsforit which
wouldhavemadethefirstremarkreasonablein a contextin whichthetruth-
valueof the antecedentis not known.All counterfactuals with false ante-
cedentsaretrue;butGricecan explainwhy some arereasonablethingsto
say, some arenot, in termsof principlesof good conversation.
Strawsongives no hintthathe expectsthe readerto contrapose:to find
the storyunbelievableforcounterfactuals, andso to weakenits credibility
for indicativestoo,'3 but this readingis compatiblewith his conclusion.
The truthandmeaningof a conditionalhave now becomequitedivorced
fromwhatmattersaboutit.
Here is a story.English*is identicalto Englishexcept in one respect
which will become clear.It has the word "dog",and names for various
breedsof dog. Onebreed,however,lacksa name.Speakershabituallycall
dogs of this breed"Labradors". But "Labrador"
reallymeansthe sameas
"dog".If you called a poodle a Labrador,this would not be false, but it
wouldbe misleading.Foryou couldhave said"poodle";or, if less specif-
icity is called for, thereis the word"dog",whichis shorterandeasierto
say than"Labrador". Hence,speakerstendnot to call otherbreeds"Lab-
radors".So the wordis quiteuseful,for when it is used, youraudienceis
likelyto cottonon andrealisethatyou aren'tspeakingof a dog of another
breed,butof this namelessone.
The story is incredible.Wordsmean what people use them to mean,
given the distinctionsthey need to make.Even if "Labrador" onrginally
meant"dog",nothingcanpreventit comingto performa moreusefulrole
in the language,the namefor the namelessbreed.Somethinglike this is
whatStrawsonhas in mindwhenhe concludes"Onlyin the speciallypro-
tectedenvironmentof a treatiseon logic can 'D' keep its meaningpure"
(p. 242).14
2.6. Frank Jackson has a different defence of the truth-functional
account.'5He claims that there is a special conventiongoverningthe
assertabilityof an indicative conditional: it is not enough simply to
believe that its truthconditionsare satisfied;this belief must be robust
'3 Some post-Dudman readershave alreadybeen convertedto the view that
these forward-looking "indicatives"behavelike "counterfactuals" (see fn. 5 for
references).Then change Strawson'sexample.X: "If he ate the apple, he was
ill". Y:"Hedidn'teat it".X: "Wellthen, if he had eaten it, he would have been
ill".
14My shaggydog storyis a little unfair.Truthconditionsfor conditionalsare
problematic,in a way thatnamingbreedsof dog is not.Wecouldbe drivento the
Griceanmanoeuvreas the alternativeleastatoddswiththefacts.Butit does show
thata consistentGriceanstoryis not necessarilya believableone.
'5 See Jackson(1979, 1981, 1987);LewisadoptsJackson'sdefence(1986, pp.
152-6), havingpreviously(1976, pp. 142-5) supportedGrice.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 247

withrespectto the antecedent,thatis, it mustbe thatyou wouldnot aban-

don belief in the conditionalif you were to discoverthe antecedentto be
true.This ensuresthatan assertableconditionalis fit for modusponens.
This conditionis not satisfiedif you believe ADB solely on the grounds
that -A. If you discoveredthatA, you would abandonyour belief that
AzB ratherthanconcludethatB.
The detailsof this defencerequirethe notionof conditionalprobability
to be discussedbelow (?5), andI shallreturnto Jacksonlater(?9.1). On
the face of it, the shift from questionsof assertabilitybased on general
conversationalpropriety,to questionsof assertabilitybasedon a specific
convention governing conditionals, leaves the objection to Grice
untouched. Jackson speaks of "the need to facilitate conversational
exchanges"(1980, p. 133). But this doesn'tappearto be wherethe prob-
lem with "v" lies: there is no evidence that one believes a conditional
wheneverone believes the correspondingmaterialimplication,andthen
is preparedto assert it only if some furtherconditionis satisfied.
2.7. Ourinvestigationof the truth-functionalconditionalleaves us with
a conundrum.In ?2.3 we arguedthat only truth-functional truthcondi-
tions could explainwhy knowingjust AvB was enoughto concludethat
if -A, B; any strongertruthconditions would demandmore, and so
would not license this inference.In ?2.4 we arguedthatthe truth-func-
tional account had intolerableconsequences, and we have not seen a
way to makethemtolerable.Thereis a solutionto this conundrum,but it
lies ahead.

3. Earlytheoriesof counterfactuals

3.1. These deservea mention,for the problemsthey raiselive on. Coun-

terfactualsappearedto be connectednot only with dispositionalproper-
ties but with laws of nature. Laws, it seemed, have counterfactual
implications,accidentallytrue generalizationsdon't. If we understood
counterfactuals,this mightilluminatethe notionof law.And conversely.
Leavingthe problem"Whatis a law?"for anotherday,perhapscounter-
factualscanbe explainedas law-governedconditionals.Thiswas triedby
Chisholm(1946), andNelson Goodman(1947), reprintedas chapterI of
Fact, Fiction and Forecast. TakeGoodman'sexample,
If the matchhadbeen struck,it wouldhave lit.
Its truth,it seemed,requiresthereto be a law, andfacts aboutthe match
andits situation(it was dry,therewas oxygen,etc.), fromwhich,together

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
248 DorothyEdgington

withthe assumptionthatit was struck,we can deducethatit lit. A theory

of the followingshapeemerges:
A counterfactualconditional"A-<C" is trueif andonly if there
is a conjunctionof truthsTwhichincludea law of nature[andsat-
isfy conditionX] suchthatA&T entailsC.
X is a place-holderfor the difficultbit. In fact, the matchwas not struck,
and did not light. Assumingthat it was struckinvolves supposingthat
somethingswhichareactuallytruewerenottrue.Forinstance,the match
remainedmotionlessanduntouchedon the table.True,but this wouldn't
havebeen trueif it hadbeen struck,so we need to forgetaboutthatfact,
in consideringwhat would have happenedif it hadbeen struck.But we
need to rely upon otherthingswhich are actuallytrue,remainingtrueif
the matchhadbeen struck,for instance,the fact thatit was dry.Whatdis-
tinguishesthose facts we may rely upon,fromthose which we may not,
whenwe makea counterfactual supposition?Using Goodman'snamefor
the problem,whichfacts arecotenablewith the assumptionthatthe ante-
cedent is true?(Thenthe squarebracketreads"andare cotenablewith
A".) Goodmandefinescotenabilitythus:
B is cotenablewithA iff it is not the case thatif A hadbeen true,
B wouldnot havebeen true.(1955, p. 15)
But now circularitylooms:we needcotenabilityto definecounterfactuals
andcounterfactuals to definecotenability.Consider:
(1) If the matchhadbeen struck,it wouldhave lit (S->L).
(2) If the match had been struck, it would not have been dry
(S-->-D).
Suppose (1) is true and (2) is false. How does the theory deliver this
result?With(1), thereis a derivationfromthe assumptionthatS, together
with a law, andfacts such as it was dry (D), to the conclusionthatL. But
with(2), thereis a derivationfromthe assumptionthatS, togetherwiththe
same law, and facts such as it didn'tlight (L), to the conclusionthatit
was not dry.The asymmetrymustlie in whichfactsarecotenablewiththe
assumptionthatit was struck.(1) is truebecausethe matchwas dry (D),
and this is cotenable with the assumptionthat S; (2) is false because,
althoughthe match did not light (-L), this is not cotenable with the
assumptionthat it was struck.Applying the definitionof cotenability,
theseclaimsamountto:
It's not the case that if the matchhad been struckit would not
have been dry [-(S ->-,D)].
It is the case thatif the matchhad been struckit would have lit
[S->L].
Now the circularityis blatant.Why is "S->L"truebut "S->-1D"false?
Because"S->--,D"is false and"S-L" is true.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 249

Goodmandecidedthathe hadreacheda deadend.Therehavebeensome

attemptsatescaperoutes,butnogeneral solutiontotheproblem. 16 Ofcourse,
wehaveanintuitivegraspof whatis cotenablewithacounterfactual assump-
tion.Butthen,we haveanintuitivegraspof counterfactuals. Goodmanwas
afteran explicationof thatintuitivegraspwhichdid not presupposeit.
3.2. I turnto a differentcriticismof this approachto counterfactuals(so
assumethatthe problemof cotenabilityhas been solved-or understand
thatnotionintuitively).Is the connectionwith laws of natureas tightas it
requires?Consider:
If I'd knownyou werecoming,I'd have bakeda cake;
If the LabourPartyhadwon, the poundwouldhave fallen;
If you had askedme yesterday,I wouldhave accepted.
Confidencein counterfactualsabout our own or others'behaviour,for
instance,does not requireus to settlethe difficultphilosophicalquestion
whetherthere are laws of naturefrom which, togetherwith cotenable
facts, the consequent is deducible from the antecedent. But on this
account,the counterfactuals standor fall withthe answerto this question.
Hereis a perfectlyordinaryuse of a counterfactual: "They'renot athome;
for the lights are off, and if they had been at home, the lights would have
been on".Youmightbelieve this counterfactual even if you are surethat
theirsittingin the darkis not inconsistentwiththe laws of natureplusrel-
evantfacts.But on this theory,if you aresurethatthe consequentdoesn't
follow fromlaws etc., you shouldbe surethatthe counterfactual is false:
it deserveszero credibility.
Manyof the counterfactuals we accept,aboutmatches,humanbehav-
iour,etc., may be ontologicallyon a parwith
(3) If you hadtossedthis (fair)coin ten (ora hundred)times,it would
have landedheadsat leastonce.
First, assumeindeterminism.Then, on the law-governedaccount(3) is
plainlyfalse:thereis no way of deducingconsequentfromantecedentby
law. If you knowthe facts, you knowit is certainlyfalse. It deserveszero
credibility,andin thisrespectis indistinguishable from"Ifyou hadtossed
the coin ten (or a hundred)times,it wouldhavelandedheadseverytime".
Indeed,it is no morebelievablethan
(4) If you hadtossedthe coin, it wouldhave turnedinto a giraffe.
Second,assumedeterminism.Now, you didn'ttoss the coin. Nor was it
possible for you to do so, given the laws and the past, underdetermin-
ism. But assumeyou hadtossed it. The antecedentof (3) specifiesa type
of event thatcan be realisedin manydifferentways, andthe consequent
16 Bennett(1984) surveysthe laterliteratureon the notionof cotenability,pp.
85-8.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
250 DorothyEdgington

will be sensitive to exactly how you had tossed it. Not all instancesof
tossing will bring the consequentout true, by the deterministiclaws.
Again, (3) is plainlyfalse, on this account,and deservesno more credi-
bility than(4).17 A theoryof counterfactuals shouldexplainwhy, though
(3) is not certain,it is plausibleandcrediblewhile (4) is not.
The,coin is merely an illustrationof a generaldifficulty.For many of
the thingsthathappen,the disjunction,indeterminism or fine-tuneddeter-
minism, is the safest of bets. If we accept it, and the "law-governed"
accountof counterfactuals, thereis some riskthatall contingentcounter-
factualswhose consequentsare at all specific,whose antecedentsarenot
unutterablylong and whose consequentsare not formulatedspecifically
in termsof chances,turnoutfalse. The explanatoryandinferentialuse we
makeof suchcounterfactuals as "IfMaryhad askedJohnto do the shop-
ping, he would have done it", "If I had climbedover the wall, the dog
would have attackedme", "If Bill had been in London,he would have
been in touch"wouldbe vitiated.

4. Possible worlds semantics

4.1. WithSaulKripke'ssemanticsformodallogic (1963) cametherevival

of the philosopher'sdream,a possibleworld.It is a promisingtool forthe
elucidationof non-truth-functional sententialconnectives.It is certainly
usefulintheformulationandclarificationof modalthought.Andit is natural
to turnto it for an elucidationof conditionals,which,on the face of it, are
aboutpossiblesituations.In the late 1960sDavidLewis,RobertStalnaker
andRichmondThomasondevelopedcloselyrelatedtheories,Stalnakerand
Thomasonfor conditionalsin general,with only pragmaticdifferences
betweenindicativesand counterfactuals,Lewis just for counterfactuals.
The openingsentenceof Lewis's Counterfactuals gives the gist:
'If kangaroos had no tails, they would topple over' seems to me
to mean somethinglike this: in any possible state of affairsin
which kangarooshave no tails, and which resemblesour actual
1' On one interpretation of Davidson's anomalousmonism, (see "Mental
Events"in Davidson,1980) counterfactuals with mentalcontentwill be like (3)
underdeterminism."If you had invitedme, I would have accepted".Eventsof
these mentalkindscan be instantiatedin variousphysicalways, and it is under
physicaldescriptionsthatthey instantiatelaws. It maybe thatthe laws guarantee
a physicalrealisationof the consequentfor manybut not for all physicalrealisa-
tionsof the antecedent.(Thereareno strictlaws statablein mentalterms.)In such
a case we wouldwantthe counterfactual to come out as probablebutnotcertain,
while on Goodman'saccountit comes out as certainlyfalse.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 251

stateof affairsas muchas kangarooshavingno tails permitsit to,

the kangaroostoppleover. (Lewis 1973, p. 1)

Stalnakersays:
Considera possibleworldin whichA is true,andwhichotherwise
differs minimallyfrom the actualworld. "IfA, then B" is true
(false) just in case B is true (false) in that possible world. (Stal-
naker1968, pp. 33-4)
Between Stalnakerand Lewis, thereare differencesin formulation,and
some substantivedifferences,but also a differencein aim. Stalnaker's
projectis less ambitious.He does not expect thereto be an informative
analysis of "A-worldwhich differs minimally from the actual world"
whichcouldbe specifiedindependentlyof judgementsaboutwhatwould
havebeen trueif A weretrue.Lewis seeks a genuineanalysisof counter-
factualsin termswhichdo not presupposethem.18
Similarityto the actualworld plays the role in these theorieswhich
cotenabilityplaysin Goodman's.Goodman'struthconditions,in possible-
worldjargon,have the form:"A-4C"is trueiff in any possibleworldin
whichA is trueandX is satisfied,C is true.For Lewis and Stalnakerthe
problemof specifyingX is theproblemof decidingwhichworldsareclos-
est to actuality.
Similarityis, of course,vague. Comparingcities, or faces, or worlds,
theremay be no determinateanswerto the question:is A moresimilarto
B thanC is? But equally,theremaybe no determinateanswerto the ques-
tion:whatwouldhavehappenedif A hadbeentrue?Lewis'saimis to ana-
lyze one vaguenotionin termsof another.Onthe otherhand,similarityis
not so vagueas to be useless. Often,clearjudgementscan be madeabout
the comparativeoverall similarityof cities, people, etc., or of how life-
like as opposedto fantasticalis a novel or a film.
areas follows:
4.2. Lewis's truthconditionsfor counterfactuals
(i) If A is truein no possibleworld, Ao-->Cis vacuouslytrue.(ii)
An->Cis non-vacuouslytrueif and only if some A&C-worldis
closer to the actual world than any A&-_C-world."In other
is, for Lewis,partof a largerpicture:causa-
i 8The analysisof counterfactuals
tion is to be analyzedin termsof counterfactuals, mentalstatesdefinedas occu-
pantsof causalroles,semanticfactsobtainin virtueof mentalstates.... His name
forthe projectis "HumeanSupervenience", "allthereis to the worldis a vastmo-
saic of local mattersof particularfact, just one little thing and then another"
(1986, p. ix). By assumingthatthereare otherpossibleworldsbesidesthis one,
he hopesto be ableto reconcilemostof whatwe believe in with an austereview
of the fundamentalnatureof ourworld.
Lewis'stheoryof counterfactuals is muchmorewidelyacceptedthanhis theory
of the natureof possibleworlds.I shall say nothingaboutthe latter.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
252 DorothyEdgington

words,a counterfactualis non-vacuouslytrueiff it takes less of

a departurefrom actualityto make the consequenttrue along
with the antecedentthan it does to make the antecedenttrue
withoutthe consequent".(Lewis 1979, p. 164)
If thereis a uniqueclosestA-world,An->Cis trueif f C is trueattheclosest
A-world.But theremay not be: (a) theremay be no A-worldsat all, in
whichcase the counterfactual is vacuouslytrue;(b) theremay be ties for
firstplace,andC may be truein some butnot all of the closest. The liter-
atureaboundswith exampleslike:
If Bizet andVerdiwere compatriots,Bizet wouldbe Italian;
If Bizet andVerdiwere compatriots,Verdiwouldbe French.
If the closest Bizet-and-Verdi-compatriot worldscontainsome in which
Bizet was Italianand some in whichVerdiwas French,then, on Lewis's
account,boththese counterfactualsare false. (Herehe differsfrom Stal-
naker,for whomthey have no determinatetruthvalue.)(c) Perhapsthere
is no closestA-worldbecausefor anyA-worldthereis a closerone. Con-
sidera conditionalof the form"IfI weretallerthanI am, C". Considera
worldin whichI am an inch taller;thenthereis a closerworldin whichI
am half an inchtaller;andso on, ad infinitum.ForLewis, the conditional
is trueiff sometaller-&C-world is closerthananytaller-&-C-world.This
hasthemildlyembarrassing consequencethat,giventhatsomedifferences
in heightaretoo smallto be detectable,"IfI weretallerthanI am, no one
would know the difference"comes out as incontrovertiblytrue.Again,
supposeyou are a little tallerthanme, say half an inch taller.Then "If I
were taller,I would still be shorterthanyou" also comes out absolutely
certainlyandobviouslytrue;whereas,"Well,maybe,butnot necessarily"
is a commonresponseto thisthought.Thisexampleis notveryinteresting
in itself (one like it is mentionedby Lewis in his case againstthe assump-
tionthattheremustbe a closestA-world);butit servesto illustratea ques-
tionto whichwe shallreturn:why putall youreggs in theclosestbaskets?
4.3. Lewis calls the counterfactuala "variably-strictconditional".There
is thematerialconditional,ADB;thereis the strictconditional,O(ADB)-
in all possible worlds,ADB;we could define weakerstrictconditionals
with referenceto some subsetof possible worlds,e.g. all those with our
laws of nature;butforthe counterfactual, the degreeof strictnessdepends
we
on the antecedent: departfrom the actual world enough to include
some A-world; throughout some "A-permittingsphere" of possible
worlds,ADB is true. This explains some curious logical propertiesof
counterfactuals. Forexample,a piece of masonryfalls fromthe corniceof
a building,narrowlymissing a worker.The foremansays: "If you had
beenstandinga foot to the left, you wouldhavebeenkilled;butif you had

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 253

(also)been wearingyourhardhat,you wouldhavebeen all right";i.e. he

says
Srj-*K;but (S&H)o-*-K.
Strengtheningof the antecedentfails for counterfactuals:the nearestS-
worldsareK-worlds;butthe nearestS&H-worldsare-nK-worlds.
Failuresof strengtheningare failuresof transitivity;for (S&H)o-*Sis
obviouslytrue;yet we have S o-K trueand(S&H)o-rK false. Otherfail-
uresof transitivitycan be constructed,for instance:
(2) If Brownhadbeen appointed,Joneswouldhaveresignedim-
mediatelyafterwards;
(1) If Jones had died beforethe appointmentwas made,Brown
wouldhave been appointed;butnot:
(3) If Jones had died beforethe appointment,Jones would have
resignedimmediatelyafterthe appointment.
Departingfromrealityenoughto get Brownappointedhas Jonesresign-
ing. Departingfromit further,to get Jonesdead,hasBrownappointed.On
this reading,(3) does not follow.
It helped to get you to read (2) before (1); if (1) had come first,you
might have said, after reading(2), "But not if he was dead!".Crispin
Wright(1983) has arguedthatthe samepossibleworldsshouldbe in play
throughouta single piece of reasoningor discourse(see also Lowe 1990).
Whentheyare,transitivityholds.Wright'sintuitionis mirroredin Lewis's
semanticsby the validityof
Ao-4B;(A&B)o-4C;so Ao-->C.
This restrictedtransitivitypreventsthe firstpremissfrombeing "further
out" than the second. Wrightholds that the "A&"is, as it were, silent,
always contextuallyimplied,in the secondpremiss.AgainstWright,the
buildingforeman'sremarksabove,violatingtransitivityas they do, con-
stitute a single, pointful piece of discourse;and one can believe both
premissesaboutBrownandJones.Naturally,if one says somethingof the
form"IfA thenB andif B thenC", thereis presumedto be some pointin
this utterance,and the most naturalone (otherthanthat of producinga
philosophicalcounterexample)is that the heareris being askedto con-
cludethatif A then C. But Lewis need not deny that.
It is not as thoughwe shouldhave a badconscienceaboutall the times
we have used or acceptedtransitivereasoning.First,if the conditionals
involved are necessaryor a priori,as in maths,logic and sometimesin
philosophy,the reasoningdoes not fail. In othercases the test is whether
the secondpremiss,Bo-4C,survivesthe additionof A to the antecedent-
survivesconversioninto "If B [still assumingA], thenC". If it does, the
conclusion follows. And it usually does. Wrighthypothesisesthat we
alwaysreadthe secondpremissthatway. So the samesentenceas second

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
254 DorothyEdgington

premisswill have a differentcontentin differentarguments.I don'tthink

thereis a deepissue here:we could go Wright'sway andsave transitivity
at the price of increasing ambiguity or context-dependence.But the
strengtheningcase suggestswe need not.
Therearealso failuresof contraposition.Stalnaker'sexample(1968, p.
39):
If the US hadhaltedthe bombing,NorthVietnamwouldnothave
agreedto negotiate;
butnot:If NorthVietnamhadagreedto negotiate,the US would
not have haltedthe bombing.
And ConditionalProoffails. "-1(A&B);A; therefore-B" is a valid argu-
mentform;but" -,(A&B);thereforeAo&--B"is invalid.Let A be "Shewas
hit by a bomb yesterday"andB be "Shewas injuredyesterday";it does
not follow fromthe falsityof A&Bthatif she hadbeen hit by a bomb,she
would not have been injured,i.e. that in the closest possible world in
which she was hit by a bomb,she was not injured.
4.4. Lawsof naturearenot mentionedin Lewis'struthconditions.But he
can explain why they loom large in judgementsaboutcounterfactuals.
Laws of natureare importanttruthswhich say muchaboutthe character
of the world.In general,the differencebetweentwo worldswiththe same
laws will be less thanthe differencebetweentwo worldswith different
laws. If, in assessingcounterfactuals,
we stickas close to the actualworld
as the specified difference allows, it follows that we tend to consider
worldswith the same laws as ours.ThusLewis explainsthe connection
Goodmantook as primitive.And, primafacie, he will have no difficulty
with examples like "If I'd known you were coming I'd have baked a
cake".Thereis no requirementthatthe consequentbe derivablefromthe
antecedentfrompremissesincludinglaws.
"Thoughsimilaritiesand differencesin laws have some tendencyto
outweighdifferencesor similaritiesin particularfacts,I do not thinkthey
invariablydo so", says Lewis (1973, p. 75). His reasonis as follows. A
tree blows over, destroyingthe roof of a house. Suppose our world is
deterministic,at least withrespectto the causalchainsconnectedto these
events.Consider
If the treehadn'tblown over,the roof wouldbe intact.
Now considertwo possibleworldsin whichthe treedidn'tblow over.In
w,, thelaws areexactlythe sameas in the actualworld.By hypothesis,the
relevantlaws aredeterministic.Then,as the treedidn'tblow over,some-
thingearliermusthavebeen different,and somethingearlierstill, andso
on, backto the beginningof time. This differenthistoryhas furtherfor-
wardconsequences,andw1, at t, the time in question,is staggeringlydif-
ferentfromthe actualworld.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 255

The historyof w2is just like the actualworlduntiljust beforet. Then

thereis a small, inconspicuous,violationof law-a "tinymiracle"(rela-
tive to the laws of the actualworld),andthe tree staysupright.
Whichof these two worldsis most like the actualworld?Surelyw2,
the one which does not obey exactly our laws. Lewis writes "Lawsare
very important,but greatmasses of particularfact count for something
too ... . I therefore proceed on the assumption that the preeminence of
laws ... is a matter only of degree" (1973, p. 75).
Thisis thebeginningof anapparentdifficultyforLewis'saccount.'9His
single guidingprinciplebehindcounterfactual judgementsis overallsim-
ilarityto the actualworld.In optingfor w2ratherthanw,, we keepthepast
in line, at the priceof a "smallmiracle".Butthe futureof the actualworld
is very differentfromthe futureof w2.Whynot purchasefuturesimilarity
at the priceof anothersmall miraclewhich destroysthe roof despitethe
tree remainingupright?ConsiderW3:like w2,its historyis just like the
actualworldto just beforet. A smallmiraclepreventsthetreefromfalling
over.Butin W3anothersmallmiracle-e.g., a lightningbolt-destroys the
roof.Its futureis very similarto the actualworld's-some inhabitantsare
killed, the family is homelessand impoverished,andfurtherdire conse-
quencesensue. Back in w2,the family continuesits peacefulexistence,
quiteunlikewhathappensin the actualworld.w2,notW3, is the way things
wouldhavebeenif thetreehadn'tblownover.ButW3,notw2,is (arguably)
the more similarto the actualworld-for a reasonapparentlysymmetric
with Lewis'sreasonto preferw2to w,.
The difficultyis general.It is oftenthe case thatif somethinghadhap-
penedwhich didn't,the worldwould have been very different.Suppose
Hitlerhad died in infancy.Then thingswould have been quite different
in the 1930s and 1940s.But considerthe worldmost similarto the actual
worldin which Hitlerdied in infancy.(Here,if you prefer,just focus on
the time betweenantecedentandconsequent.)Thatmay be one in which
some otherchild grew up to occupy a virtuallyidenticalHitler-likerole.
Not that that would have happened,mind you. Imagine two films in
which Hitlerdied in infancy.One of themhas a non-Hitlerdoing all the
kinds of thing Hitler did. It strikes you as remarkablylike the actual
world, almost indistinguishablefrom the newsreels. The other strikes
you as a very plausibleaccountof how the worldwouldhave been with-
out Hitler-rather different.Judgementsof similaritygo one way,judge-
mentsaboutcounterfactuals, the other.
19Lewis mentionedthe problem(1973, p. 76). It was pursuedin two reviews
of Counterfactuals, Bennett (1974) and Fine (1975).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
256 DorothyEdgington

Lewis repliedto this objection,by specifyingwhichaspectsof similar-

ity mattermost,on the "standard resolutionof vagueness"for counterfac-
tuals.Thesearethe criteria:
(1) It is of the firstimportanceto avoidbig, widespread,diverse
violationsof law.
(2) It is of the secondimportanceto maximizethe spatio-tempo-
ral regionthroughoutwhichperfectmatchof particularfact pre-
vails.
(3) It is of the-thirdimportanceto avoid even small, localized,
simpleviolationof law.
(4) It is of littleor no importanceto secureapproximatesimilarity
of particularfact, even in matters which concern us greatly.
(Lewis 1979, pp. 47-8)
To see how these work,returnto the tree.We arestill operatingunderthe
assumptionof determinism.There was not perfect symmetrybetween
Lewis's case for w2 over w, andmy case for w3 over w2.In w2 the pastis
exactlythe same as in the actualworld.In W3, the futureis approximately
the same as in the actualworld;but I did not imagine that the second
miraculousdisasterwould makethe worldexactlyas it actuallyis, with
the tree blown over. It would take a massive miracleto secure perfect
reconvergenceto the actualworld,and(1) rulesout similarfuturesat that
price.By (2), we preferw2to w,, as Lewis requires.W3has two tiny mira-
cles, w2only one, so, by (3), w2is to be preferredto W3, despitethe greater
approximatesimilarityof particularfact in w3, which,by (4), countsfor
"littleor nothing".We get the right answer:the most similarworld,by
these criteria,is the one thatwouldhave happened.
Lewis is not a determinist,andin a Postscriptto this article(1986, pp.
58-65) he discusseswhathappenswhenwe dropthatassumption.A theory
of counterfactualsshould not requiredeterminism.Supposethere was
some chancethat the tree would not blow over. So no small miracleis
requiredto keepthe pastin line in worldsin whichit did not. Lewis puts
mosteffortintoarguingthatwe shoulddiscountworldsin which,although
the tree doesn't blow over, a "quasi-miracle"securesperfect reconver-
genceto the actualworld.Evenif thishasa non-zerochanceof happening,
such peculiarthings20don't happenin worldssimilarto ours,he claims.
Let us granthim this. But he is too cavalieraboutthe possibility,which
also has a non-zerochanceof occurring,of gettingthe worldsapproxi-
matelybackin line again.He says:
20 A quasi-miracle is not just a very improbable event. Very imNprobable
events
happen in this world and those like it. "What makes a quasi-miracle is... the re-
markable way in which the chance outcomes seem to conspire to produce a pat-
tern [like]... the monkey at the typewriter [producing] a 950-page dissertation
on... anti-realism" (1986, p. 60).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 257

The thing to say about approximateconvergence remainsthe

same.Even if approximateconvergenceis cheap-and even if it
is cheaperstill whenit can be hadwithouteven a little miracle-
still we can say thatit countsfor little or nothing,so it is not the
case thatif Nixon hadpressedthe button,therewouldhave been
approximateconvergence to our world, and no holocaust.21
(Lewis 1986, p. 59)
Supposetherewas a tiny chanceat t (butno laterthant) of (e.g.) a light-
ning bolt destroyingthe roof. It didn't happen:if the tree hadn'tblown
over,the roof wouldhavebeen intact.But now w2andW3 minustheirmir-
acles, andwiththe samestretchof identicalpasts,areequallysuitableby
criteria(1)-(3). If approximatesimilaritycountsfor nothing,we have a
tie, and have been given no guidance on choosing between them. If
approximatesimilaritycountsfor something,albeitlittleratherthannoth-
ing, then, arguably,the wrongworld (W3) wins. Similarlyfor the Nixon
example.
Consider Kennedy's assassination. Suppose Oswald did it, acting
alone, andthatif he hadn't,no one else wouldhave. Considersome pos-
sible worldsin whichOswaldhadlast-minutefrightanddid not shoot.Ex
hypothesi,no one else even thoughtof shootingKennedy.But the crowd
containedpeoplecarryinggunsandnotaverseto usingthemwho couldn't
standthe man:assume,whatmay well be true,thatit was consistentwith
the laws of natureandthe pastthatsomeoneelse act on a suddenimpulse
to shoot.Again we have two possibleworldsin whichOswalddidn'tdo
it, not distinguishedby Lewis's criteria(1)-(3), one of which is what
wouldhave happened;the other,in whichsomeoneelse shoots,the more
approximatelysimilarto ours.If this countsfor nothing,a tie; if it counts
a little, similaritytakesus in the wrongdirection.
Shouldit be "little"or "nothing"? Lewis (1986, p. 48) isn't sure.Many
examplessuggestthatapproximatesimilaritycountsfor something:"If I
had bet on heads,I would have won";"If I hadboughtthese shareslast
year,I wouldbe richtoday";"IfI hadleft 5 minutesearlier,I wouldhave
avoidedtheaccident";thesedo relyon approximate similarityto theactual
worldafterthedivergencefromperfectmatchneededto get theantecedent
true.Therearecountlessexampleslike these.To say "nothing"is to deny
thetruthof anycounterfactual like "IfI hadgot outof bedone minuteear-
lier,the resultof the Swedishelectionwouldhavebeenno different".The
example which tempts Lewis to say "nothing"is due to Pavel Tichy
(1976):whenFredgoes out,if the weatheris bad,he alwayswearshis hat;
if the weatheris fine,it's a random50-50 whetherhe wearshis hat.In fact
the weatheris bad,andhe wearshis hat.Consider"Iftheweatherhadbeen
21This is the example, first used by Fine (1975), in terms of which Lewis con-
ducts this discussion.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
258 DorothyEdgington

fine, he would have worn his hat".The fine-weatherworldin which he

does so is morelike the actualworldthanthe fine-weatherworldin which
he does not, butthe counterfactual is not clearlytrue.
It is notdifficultto spotthe differencebetweenTichy'sexampleandthe
earlierones. The weatheris not causally independentof whetherFred
wears his hat: fine weatherreducesthe chancefrom 100%to 50%. By
contrast,my getting out of bed is causally independentof the Swedish
election,my buyingshareshas a negligibleeffect on theirprice,etc. But
Lewis does not allow himselfaccess to the notionof causationin analyz-
ing counterfactuals, for they are to be used-toanalyzecausation(1973a,
pp. 159-72). His difficultyheregeneratesfurtherdoubtaboutwhetherthe
notion of similarityalone, howevertailored,will yield the rightjudge-
mentsaboutwhatwouldhavebeentrueif A hadbeentrue.Anotherdoubt
aboutLewis'scriteriawas raisedby JohnPollock(mentionedby Bennett,
1984, p. 68): I leave my coat in a restaurantat noon, andreturnfor it at
midnight.A steadystreamof potentialcoat-thieveshave passedit by, but
it is still there.By Lewis's criterion(2) (p. 256 above),"If my coat had
been stolen this p.m. it would have been stolen very close to midnight"
comes out true.
4.5. Lewis's elaboratedtheoryhas the effect thatwe stick to the laws of
the actualworldat times laterthanthe antecedent-time,t, when we eval-
uatecounterfactuals, andactualfacts at timeslaterthant areunimportant
at best. (We exclude worldswith "quasi-miracles" as well: this doesn't
concernme.) Recallthatunlessthe consequentis truein all closest ante-
cedent-worlds,the counterfactualis false. Now Lewis's theoryis in the
same positionas Goodman's(see above, ?3.4). If the consequentis true
in almost all close antecedent-worlds,the counterfactualis false, and
deserveszero credibility.Again,
(3) If you had tossed the coin ten times, it would have landed
headsat least once,
is no moreworthyof belief than
(4) If you hadtossedthe coin ten times,it wouldhaveturnedinto
a giraffe.
If indeterminism is rife,almostall counterfactuals aboutwhatwouldhap-
pen if you hadstruckthe match,invitedme to dinner,etc., turnout false.
And if determinismis truebutfine-grained,while thereis no way thatthe
antecedentcould have come about,given the laws andthe past,the laws
won'tguaranteetheconsequentfor anyold smallmiraclegettingthe ante-
cedenttrue-maybe the vast majorityof "close"ways of instantiatingthe
antecedentwill guaranteethe consequentbut the odd one won't. Then
again, all such counterfactualsare false. Contingentcounterfactuals,
except those with very unspecificconsequents("If you had tossed the

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 259

coin, it would have landed",perhaps),or consequentsaboutchances,or

unutterablylong antecedents,will come out false.
Supposewe wantthe resultthatsomeonewho knowsthe relevantfacts
(be they indeterministicor fine-graineddeterministic)shouldbe almost,
butnotquite,certainthat(3). Thenwe shouldwantsomethingalongthese
lines: a measureof the credibilityof a counterfactualis the proportionof
close A-worldsin which C is true.But it is not clear whatthe truthof a
counterfactual like (3) wouldconsistin.
Wherethereis a tie for closenessandtheconsequentis truein somebut
notall closestantecedentworlds,Stalnakermakestheconditionalindeter-
minate-neither determinatelytrue nor determinatelyfalse (Stalnaker,
1981,p. 87). Thisis morepromisingfromthepointof view of theprevious
paragraph, like (3) being"almost
for it is compatiblewitha counterfactual
true";whereas,for Lewis, it is "flatly,determinately
false"(1981, p. 331).
4.6. Thereis a relateddifficultyfor StalnakerandLewis, mentionedon p.
252 above(why putall youreggs in the closest basket?).Suppose,on the
rightaccountof closeness,a B-worldwins amongantecedent-worlds, but
-B-worlds areonly a hair'sbreadthbehind(as it were).A wins the elec-
tion.If he hadn't,it wouldtakeminimallyless departurefromactualityfor
B to win thanfor C to win. ForLewis and Stalnaker,-A o-+Bis clearly
true.It is not even truethat C might have won if A hadn't,on eitherof
Lewis's readingsof "mighthave"(Lewis, 1986, pp. 63-4). If we find it
moreacceptableto say thatit is only probablethatB wouldhavewon if A
hadn't,we are taking a probabilitydistributionover close -A -worlds,
whichis consonantwithhow,I suggest,we shouldreactto (3). I turnto a
theoryof uncertainconditionaljudgements.

5. Conditional uncertainty

5.1. Youmay be surethatB if A, butoften you will be less thansure(e.g.

thatthe patientwill recoverif he has the operation).Thereare different
degreesof uncertainty. Youmay be nearlysure,fairlysure,thinkit more
likely than not, less likely thannot, downto beingcertainthatit's not the
case thatB if A. The samegoes for propositionsin general.Ourcapacity
for a spectrumof epistemicattitudestowardsa propositionis important
in our deliberationsaboutwhatto do, and whatelse to think.Oftencer-
taintyis unachievable,andnear-certainty is nearlyas good. One applica-
tion of the theoryof probabilityis to providea "logicof partialbelief'
as Ramseycalled it in his 1926 paper"Truthand Probability"(Ramsey,
1931, p. 166). The theory has its own way of expressing conditional

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
260 Dorothy Edgington

uncertainty-you don't get far in the studyof probabilitywithoutmeet-

ing the "conditionalprobabilityof B given A". In his 1929 paper"Gen-
eral Laws and Causality"Ramsey suggested that deliberationabout
whetherif A, B fitsthe probabilisticmodel(1931 pp. 246-7). In the 1960s
and 1970s several philosopherssought illuminationaboutconditionals
fromthis source.22
Any theory of conditionals has consequences for less-than-certain
judgements.Somethingis proposedof the form:"IfA, B"is trueiffA*B.23
If a clear-headedperson,free from confusionsof a logical, linguisticor
referentialsort,can be nearlysurethatA*B yet farfromsurethatif A, B,
or vice versa,then this is strongevidenceagainstthe proposal.I am not
suggestingthata competentuserof "if"has figuredout the correcttheory
of conditionals.But if a theorystates an equivalencebetween items of
belief to which competentusers stably,incorrigiblyand unhesitatingly
takedifferentattitudes(andtheirpracticeservesthemwell), then,on the
face of it, the theoryis wrong.At best, it has a lot of explainingto do of
massiveerror;andit is hardto see whatwouldconvinceus, in a case like
"if", thatthe peoplearewrongandthe theoryis right.
We have alreadyseen this patternof argumentat work.
(1) Proposal:"IfA, B" is trueiffADB, i.e. -AvB. Objection:suppose
I'm 90%certainthat-A, hence90%certainthata sufficientconditionfor
the truthof the righthandside is satisfied,yet 0%certainthatif A, B. (Let
A be "TheQueenis at home"andB be "She'sworryingaboutme". See
pp. 243-4 above.)If the proposalwerecorrect,I wouldbe guiltyof gross
irrationality.But I am not, so the proposalis incorrect.
(2) Proposal:a counterfactual "A--B" is trueiff B is deduciblefromA
+ laws of nature+ suitablefacts. Objection:(a) I may be highlysceptical
aboutwhether"Thelights areon"is deduciblefrom"Theyareat home"
+ laws of natureetc.;yet close to certainthatif theyhadbeenathome,the
lights wouldhave been on. (b) I am surethatit is not the case that"The
coin landedheadsat least once"is deduciblefrom"Thecoin was tossed
ten times"+ laws of natureetc. YetI am close to certainthatif it hadbeen
tossedten times,it wouldhavelandedheadsat leastonce. (See above,pp.
249-50.)
(3) Proposal:a counterfactual "Ao-B" is trueiff B is trueat all closest
A-worlds.Objection:(a) take the coin example.I am sure that (even if
determinismis true)it's notthecasethattheconsequentis trueatall closest
antecedent-worlds (butonly in the vast majorityof them).YetI am close
22 Richard
Jeffrey(1964) publishedan abstlract
of a paperon thistheme.Ernest
Adams,(1965, 1966, 1975)hasdonemostto developthislineof thought.See also
BrianEllis (1973) andRobertStalnaker(1970) for earlyworkon this idea.
23Noteverytheorymusttakethisform.Ourpresentinterestis inthosewhichdo.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals261

to certainthatif the coin hadbeen tossedten times, it wouldhave landed

headsat least once. (b) I am certainthata worldin whichI am tallerthan
I actuallyam and still shorterthanyou is closer to the actualworldthan
any worldin whichI amtallerthanI actuallyambutnot shorterthanyou.
But I am less thancertainthatif I hadbeen taller,I wouldstill havebeen
shorterthanyou.
Some degreeof uncertaintyis the normfor the contingentconditionals
we meet everyday.It is not a peripheralobjectionto a theorythatit gets
uncertainjudgementswrong.Canwe finda substitutionforA*B whichis
immuneto this objection?To find out, we must firstturnto the logic of
less-than-certainjudgements.
5.2. Extendthe term"belief' to includepartialbelief. If we idealiseby
expressing a person's degree of belief in a proposition as a number
between1 (forcertaintythatit's true)and0 (forcertaintythatit's false),24
thenit can be shownthatsatisfyingthe principlesof probabilitytheoryis
a requirementof consistency25upona person'sdegreesof belief. Call "a
partition"a set of propositionswhich aremutuallyexclusive andjointly
exhaustive-not morethanone of them can be true,and it mustbe that
one of themis true.The claimthatdegreesof belief behavelike probabil-
ities is the claimthatyourdegreesof belief in the membersof a partition
shouldsum to 1; Some consequencesof this claim:26(1) Yourdegreeof
belief in -A shouldbe one minusyourdegreeof belief in A (because{A,
-A } is a partition).(2) If you recognisethatA and B are incompatible,
then your degreeof belief in AvB shouldbe the sum of yourdegreesof
belief in A andin B (because{A, B, -A &-B } and {A vB, -A &-B } are
bothpartitions).(3) Yourdegreeof belief in A shouldbe the sum of your
degrees of belief in A&B and A&-B (because {A, -A} and {A&B,
A&-B, -A } arebothpartitions).(4) If you recognisethatA entailsB, then
yourdegreeof belief in A shouldnotbe greaterthanyourdegreeof belief
in B. (Write"b(-)"for your degree of belief in (-). RecognisingthatA
entails B, you have b(A&-B) = 0. So b(A) = b(A&B), by (3). b(B) =
b(A&B) + b(-A&B) ? b(A).) (5) If you recognise that A and B are logi-
24
For many things that come in degrees, it is a useful idealisation to represent
degrees numerically. For then we can use relations between numbers (arithmetic)
in our theory of relations between degrees. This-often in physics as well as in
philosophy -represents the phenomena as more precise than they really are. Such
is the natureof idealisations-but their utility is beyond doubt. The test of the ad-
equacy of an idealisation is that it deliver results of the right order of magnitude.
25 Consistency is not the only virtue of a set of partial beliefs. But it is all that
concerns us here, in investigating the logical relations between degrees of belief.
26
Here I assume that A and B are bivalent propositions. It will be important
later that we need not assume this. We could allow for truth-value gaps, indeter-
minate or intermediatevalues, provided we read "-I" as "It is not true that"rather
than "It is false that".

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
262 DorothyEdgington

cally equivalent,you shouldhavethe samedegreeof belief in each(from

(4)).
A partitionslices up a space of possibilities.It is convenientto repre-
sentapplicationsof the PartitionPrincipleby pictures:

A A

'I A A B A&B

1 LS LMB -A

I A -&-, -A -A&--B -A&-B3

b(A)+b(-A)=1 A andB in- b(A)=b(A&B) A entailsB A andB
compatible +b(A&-B) b(B) ?b(A) equivalent
b(AvB)= b(A)=b(B)
b(A)+b(B)

The rectanglesareof height1. The internalhorizontallines representhow

you divide yourbelief betweenthe possibilities,in accordancewith the
PartitionPrinciple.
The other"fundamental law of probablebelief' introducesthe idea of
a conditionalprobability:
(CB) Degreeof belief in (p andq) = degreeof belief in p x degreeof
belief in q givenp (Ramsey,1931,p. 181).
Thiswas not an innovation.ThomasBayes, in an essay publishedposthu-
mouslyin 1763, has as Proposition3 "Theprobabilitythattwo ... events
will bothhappenis ... the probabilityof the first,[multipliedby] the prob-
ability of the second on the supposition that the first happens" (1940, p.
378, myitalics).TheFourthPrincipleof Laplace'sEssaiphilosophiquesur
les probabilites(1795, in 1951,p. 14) is the same.27The notionof a con-
ditionalprobability-the probabilitythatB on the suppositionthatA-
playsa big rolein manyapplicationsof thetheoryof probability.Now "on
the suppositionthatA" and "givenA" would appearto be mere stylistic
variationson "ifA".So, it wouldseem, CB statesa logicalrelationwhich
shouldholdbetweenyourdegreesof belief in B if A, A&B,andA. Again
writing"b"for "yourdegreeof belief in",it may be rewritten
to thinkof probabilitiesas degreesof epistemicun-
27 Norwas it an innovation

certainty.Thetitleof Bernoulli'sfamouswork(1713) is ArsConjectandi -the art

of conjecturing.Whatwas new was the argumentthatdegreesof belief have the
structureof probability.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals263

b(B if A) = b(A&B)lb(A).
Callthis "TheThesis".
Examplesof CB at work:yourdegreeof belief in A&Bis notin general
determinedby your degrees of belief in A and in B. Supposeyou have
degreesof belief 1/2 in each of the following:headson toss 1 (H,); tails
on toss 1 (T,); not tails on toss 1 (-1T,); and heads on toss 2 (H2);yet
b(H,&T,) = 0, b(H,&-,T,) = 1/2, b(H,&H2) = 1/4. The difference lies not in
yourdegreesof belief in the conjuncts,but in the facts thatb(T, if HI) =
0, b(-,Tl if H,) = 1, b(H2 if H,) = '/2. Each case is an instance of b(A&B)
= b(A) x b(B if A).
Anotherillustration:the examineris to selectatrandomone of five top-
ics for the exam.You are around90%certainthatJim will pass if one of
the threeNice Topicsis selected,but only about30%certainthathe will
pass if a NastyTopic(conditionalsor probability)is selected.How confi-
dent should you be that he will pass? Well, the 60% chance of a Nice
Topic divides into:Nice Topic and Pass (90%of 60%);Nice Topic and
Not Pass (10%of 60%).The 40% chanceof a Nasty Topicdivides into
Nasty and Pass (30%of 40%);Nasty and Not Pass (70%of 40%).The
probabilitythathe will pass is the probabilityof (Nice Topic and Pass)
plus the probabilityof (NastyTopicandPass) which is (90%of 60%)+
(30% of 40%) = 54% + 12%= 66%.

60% P 54%

-,p 6%
P 12%
40% Q _P 28%

Theseexamplesusedthe word"if",naturallyenough,in accordancewith

the Thesis. Whenwe come to comparethe Thesis with rivalaccountsof
"if",we cannothijackthe word.The standardnotationfor"theprobability
of B given A", understoodaccordingto CB, is "p(BIA)". Stressingthe
of probabilityas degreeof belief, we may write"b(BIA)".
interpretation
The standardnotation(which I will use) is potentiallymisleading,we
shallsee, andwouldbe betterrendered""cp(BI A)", "cb(BIA)". Moreper-
spicuous still, perhaps, would be "pA(B)", or "bA(B)". For your present
cb(BIA) is yourdegreeof belief in B, not in yourpresentbelief distribu-
tion, b, but in a hypothetical-beliefdistribution,bA,derivedfrom your
actualdistribution,b, by assumingthatA-eliminating the -A -possibili-

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
264 DorothyEdgington

ties-and keeping the relative probabilitiesof all the A-possibilities

unchanged.

{ AB 1 A B

b bA

In the pictures, let b(A) = 0.5, b(A&B) = 0.4 and b(A&-,B) = 0.1. bA(B)=
b(A&B)lb(A) = 0.8. (Note, the Partition Principle applies to bA,as much
as to b: bA(B)+ bA(-B) = 1; if B and C are incompatible, bA(BvC)= bA(B)
+ bA(C),etc.)
So we have a substantiveThesis aboutwhatit is for you to be moreor
less confidentthatB if A. YouassumeA. Underthatassumption,youjudge
it more or less likely that B. And this judgementis equivalentto your
judgementof the relativelikelihoodof A&BandA. Yourdegreeof belief
in anunconditionalproposition,thatit will raintomorrow,is proportional
to yourrelativeconfidencein rainas opposedto no rain:if you thinkit 9
timesmorelikelythatit will rainthanthatit will not,yourdegreeof belief
in rainis 0.9. Yourconditionaldegreeof belief thatthe partywill be can-
celled (C) if it rains (R), is proportionalto your relativeconfidencein
R&Cas opposedto R&-,C:if you thinkit 9 timesmorelikely thatit will
rainandthe partywill be cancelled,thanit is thatit will rainandthe party
won'tbe cancelled,yourdegreeof beliefthatit will be cancelledif it rains
is 0.9. If you aresurethatB if A, e.g. thatit has 4 sides if it's square,then
b(A&B) = b(A) and b(A&-B) = 0; your degree of belief in -B if A is 0.
You are nearly sure to the extent that b(A&B)and b(A) are close, and
b(A&-B) is a smallfractionof b(A)andof b(A&B).
Ramseysuggestedthe Thesis:
If two people are arguing"Ifp will q?"andarebothin doubtas
to p, they areaddingp hypotheticallyto theirstockof knowledge
and arguing on that basis about q; ... they are fixing their degrees
of belief in q givenp. (Ramsey1931, p. 247)
5.3. Furtherfeaturesof conditionaldegreesof belief needcomment.
(1) The ratiob(A&B)Ib(A)is not definedwhenb(A)= 0. It is plausible
thatthe indicativeconditionalis usedonly if the antecedentis takenas an

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 265

epistemicpossibility,as not certainlyfalse, by the speaker(thinker)-at

least for the sake of argument,at least temporarily,at least to co-operate
with her audience.So this is, in the firstinstance,an exercise in under-
standing indicative conditionals. Certaintybeing a vague and shifty
notion28,therearefew thingsthatyou cannottakeas an epistemicpossi-
bility,as all the famousscepticalhypothesesshow.Descartessearchedfor
suchthings,andhis findingslendsomesupportto thisrestrictionon indic-
ativeconditionals.A conditionalmaybegin"IfI hadnot existed",or "IfI
don't exist tomorrow"."If I did not exist yesterday"may get off the
groundin the contextof a discussionof scepticism;butthereis no thought
whichbegins"IfI don'texist now":this is a non-starter.
Adams (1975 Ch. 4) and Brian Skyrms (1981, 1994) suggest (in
slightlydifferentways)thattheThesiscanbe extendedto counterfactuals,
alongthe followinglines: confidencein the counterfactual expressesthe
that
judgement it was probable thatB givenA, at a time when A hadnon-
zero probability,even if it no longer does; and even if you do not now
have a high degreeof belief in B givenA.
Probabilitieschange with time, as live possibilities get eliminated.
Thinkof yourfavouritethriller:theherois doomed,escapeswithamazing
luck, victory seems assured when luck switches to the villain .... You bet
on 3 headsin a row.Yourprobabilityof winningis '/8. Afterone toss, your
probabilityof winninghas changed-it is either1/4 or 0, dependingon the
outcomeof the firsttoss. We makejudgementsaboutwhat was proba-
ble-was to be expected, was to be expected if we assume such-and-
such-as well as aboutwhatis probable.
If theThesisappliesto indicativeconditionals,anextensionto counter-
factualsis primafaciedesirable.Thecloselinksbetweenconditionaljudge-
mentsof thedifferentformsstronglysuggestthatwhatmakesyouconfident
that(e.g.) you will be ill if you eat the apple,also makesyou confident,
afteryou havethrownit away,thatyou wouldhavebeenill if you hadeaten
it-the conditionalprobabilityof illnessgiveneatingwas high.Returning
to anotherexample(p. 240 above),the doctorobservescertainsymptoms.
Herdegreeof belief thatthe patienthas thesesymptomsis roughly1; and
the assumptionthatthe patienttook arsenichas no effect on it: she thinks
thatif the patienttook arsenic,he has these symptoms;andif the patient
didn'ttakearsenic,he has these symptoms.But she thinksthatthe condi-
tionalprobabilityof symptomsgiven arsenicwas high, while the condi-
tionalprobabilityof symptomsgivenno arsenicwas low;thatis, shethinks
it was likelythatthe patientwouldget thesesymptoms,giventhathe took
arsenic;andwas unlikelythathe wouldget thesesymptoms,giventhathe
28 I mean,thereis no sharpcontext-freedistinctionbetweencertaintyand its
nearneighbours.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
266 DorothyEdgington

did not take arsenic. She infers from these judgements thatthe patientprob-
ably took arsenic. Returning to yet another example: I can think it was
improbable,beforethe killing,thatanyoneotherthanOswaldwouldkill
Kennedy.I agreewith thejudgementwhich,beforethe killing,wouldbe
expressedby "If Oswalddoesn'tdo it, no one else will".But, in the light
of whatis nowknown,I amsurethatsomeoneelse didit if Oswalddidn't.
But this is looking ahead.We shall restrictattentionto indicativesfor
the time being.
(2) In mathematicalexpositionsof probabilityone reads"p(BgivenA)
=dfp(A&B)Ip(A)(providedp(A) ? 0)", andsome philosophers(forexam-
ple, Lewis, 1976,p. 133)follow suit.A mathematicalexpositionwill start
with a completeprobabilitydistributionover a partition-an assignment
of numbersto themembersor"worlds",29 whichsumto 1.Theprobability
of any (unconditional)propositionis the sum of the probabilitiesof the
worldsin whichit is true.The distributiondeterminesp(A) andp(A&B),
in termsof whichp(BIA) is defined.Thisis finemathematically, butit is
at best misleadingin epistemicapplicationsof the theory;for it suggests
thatyou need to have determinedb(A) andb(A&B)in orderto arriveat
b(B givenA). Thatwouldpreventus fromworkingout b(A&B)fromb(A)
and b(B IA).In the example of the exam (p. 263) -hardly untypical in struc-
ture -we ended up with a partition, constructed from the inputs b(nice),
b(passlnice) and b(passlnot nice). Ramsey's multiplicationrule, CB,
wouldcollapseintotheidentity,b(A&B)= b(A&B),usingthis"definition".
Also, we oftenhavea degreeof beliefin B givenAwhenwe havenotdeter-
minedwhatwe thinkaboutA.Oneimportantcaseis whenI amdeliberating
aboutwhatto do:A hasthe form"Ido x", andB is a possibleconsequence
of doingx. It wouldbe absurdto hold thatI haveto figureout how likely
it is thatI will do x, beforeI can arriveat a judgementb(BIA).
The naturalorderof humanthinkingis not the samethingas the most
elegant orderof mathematicalexposition.30Humansare not endowed
with completebelief-distributions over the finestpartitionsthey need to
consider.They need to work out some degrees of belief (as the need
29 It is convenientto thinkof theelementsof a partition-the finestdistinctions

amongpossibilitieswhichareneededforthe purposeat hand-as "worlds".They

are not ultimate possibilities-not complete ways the worldmightbe, however.
We arenot capableof thinkingof possibilitiesin completedetail.Theyare what
Kripkecalls "mini-worlds"in the Prefaceto Naming and Necessity, (1972) pp.
16-18.
30 Russell,in the introduction to the secondeditionof Principia Mathematica,
takesthe reductionof all truth-functions to the Shefferstroketo be "themostdef-
inite improvement...duringthe past fourteenyears [since the first edition]"
(1962, p. xiii). But "Neither (neither P nor P) nor (neither Q nor Q)"is hardly epis-
temicallymorebasic-than"P andQ".Norcould-onecomebythe truththat2+2=4
via the theoremto thateffect in Volume2 of Principia Mathematica.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals267

arises)in termsof otherswhich are morereadilyaccessible.b(BIA) can

be accessibleen routeto b(A&B),andcan be accessiblewhenb(A)is not
(for humans,fortunately,are capableof supposing).The same must be
said of another"definition": A andB areprobabilistically independentiff
p(A) = p(A given B). This does not preventme fromjudgingthatrainin
Chinatomorrowis independentof my finishing this papernext week
withouthaving first decided what these probabilities,conditionaland
unconditional,are andthen noticingthatthey are equal.Whetherwe see
epistemicprobabilitytheoryas the logic of uncertainty(as I do) or as the
mechanicsof cognition(as some do) it does not work"fromthe bottom
up"31:judgementsof independenceconstrainjudgementsof uncondi-
tional and conditionalprobabilityandjudgementsof conditionalproba-
bility fix ratiosof unconditionalprobabilities.32
It is perfectlypossibleto
operatewiththe constraintthat(unlessor untilsomethingshouldchange
my mind) b(A&B) = 0.9b(A) in advance of settling b(A). The Thesis says
thatdoingso is tantamountto being90%certainthatif A, B. The thought
process behind this judgement:assumeA; underthat assumption,I'm
90%certainthatB.
A ratiocan be determinatewhose numeratoranddenominatorarenot.
I look at my speedometer,whichtells me, in its inscrutableway,thatI am
now doing30 miles perhour."Now"refersnotto a dimensionlessinstant,
but to a short but indeterminatestretch of time. Velocity is distance
divided by time. We do not have to resolve the indeterminacyof the
time-for howeverwe do so, the ratiois fixed (withinlimits) andmean-
ingful.Conditionaldegreeof belief is an interestingconceptto the extent
thatthe ratiosare stablefixturesof a belief system,which can be settled
independentlyof b(A)andb(A&B).33
(3) A featureof the Thesisto whichsome peopleobject34is thatbelief
thatA&Bis sufficientfor belief thatB if A. The Thesissharesthis feature
withthe truth-functional conditional,andwith Stalnaker'spossible-world
analysisof indicativeconditionals:if you believeA&B,you believe a suf-
ficientconditionfor the truthof ADB;and you believe thatthe A-world
which differs minimallyfrom the actualworld-viz., the actualworld
I ' Thinkof a "world"as a state-description.Supposetherearesix logicallyin-
dependentpropositionsto be consideredin a givenproblem.The64 state-descrip-
tions (or lines of a truthtable)of 6 conjunctsforma partition."Fromthe bottom
up"meansthatwe startby assigningprobabilitiesto these.
32 Pearl(1988) does muchto makethe theorycomputationally tractableas the
mechanicsof cognition,by giving beliefs aboutindependenciesa fundamental
role.
33 D. H. Mellor(1993) dismissesthe ratioas an accountof conditionalbelief
on the groundsthatb(B if A) can exist whenb(A) does not.This is an over-reac-
tion. No defenderof theThesisthinksb(A) mustbe fixedbeforeb(B if A).
34 ForinstanceMellor(1993), Pendlebury (1989), andRead(1995).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
268 DorothyEdgington

itself- is a B-world.Still, it is commonlycomplained,conditionalswith

partswhich aremutuallyirrelevant,like "If Napoleonis dead,Oxfordis
in England"arenot acceptable,or even, false.
Ourinterestin conditionalscentreson the case wherewe're not sure
whetherA, not surewhetherB, butthe suppositionthatA has some bear-
ing on whetherB. I'm not surewhetherJimwill pass,butprettysurethat
if a Nice Topic comes up he will pass. But let us considerthe conse-
quencesof the Thesisfor the less interestingcases. First,supposeyou are
alreadysurethatA. Then supposingthatA changesnothing:yourbelief
distributionb alreadyrulesout -A, andso is the sameas bA. Yourdegree
of belief in any proposition,B, on the assumptionthatA, is simplyyour
degreeof belief in B. So, if you are alreadysurethatB as well as A, you
aresurethatB if A. Most instancesof this kindwill be of no interest.But
it is too muchto ask thatall acceptableconditionalsbe interesting.It is
enoughthatyou do not doubtthatB is true,on the assumptionthatA is.
Supposeyou thinkthatB is true.It does not follow thatfor any suppo-
sition,A, you will believe thatB if A. ForA mightbe the sortof supposi-
tion whichwouldundermineyourbelief thatB. But if you considerA to
be irrelevantto B, the suppositionthatA leaves yourbelief thatB undis-
turbed.Forinstance,I believe thatthe matchwill be cancelled;for all the
playershaveflu. I believe thatthe matchwill be cancelledwhetheror not
it rains.I thinkit will be cancelledif it rains,and I thinkit will be can-
celled if it doesn'train.(Saying"Thematchwill be cancelledif it rains"
is likely to be misleadingin this situation.To rejectGrice'sdefenceof the
truth-functional conditionalis notto rejectwholesalethe Griceanthought
thatyou can misleadyouraudienceby expressinga belief, whenthereis
somethingmore appropriateyou could have said.) On the other hand,
althoughI believe thatthe matchwill be cancelled,I don'tbelieve thatif
the playersmakea very speedyrecoverythe matchwill be cancelled.For
thatsuppositiondoes unsettlemy belief.
Mellor (1993) defendsa close cousin of the Thesis, butjettisonsthis
feature:he does not, he says, accept "If France is big, Egypt is hot"
(althoughhe is certainof bothconjuncts)."Iamnotat all disposedto infer
Egypt'sheat fromFrance'ssize" (pp. 247-8). Being disposedto inferB
fromA is one way he characterisesacceptinga conditional,whichhe elab-
orates:"Inotherwords,fully to accept... 'If P, Q' is to be disposedfully
to believe Q if I fully believe P". He also endorsesRamsey'sexplanation
of conditionaldegreesof belief in termsof conditionalbets:"Mychoice
of oddsfor ... a conditionalbet [on Q, conditionaluponP] I taketo meas-
urethe degreeof belief I now believe I am disposedto havein Q if I fully
believe P" (p. 234, fn5). But if bettingtests work at all, they will show
Mellorbelieving"IfFranceis big, Egyptis hot".A conditionalbet is a bet

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 269

on the consequent,whichis calledoff if the antecedentis false. He is sure,

in this case, thatthe antecedentis true,and hence thatthe bet won't be
called off. It is, in his eyes, equivalentto a bet on the consequent.His
choice of odds will reflecthis belief thatthe consequentis true,and his
belief thatit is trueif the antecedentis. So Mellordoes nothave a consist-
ent position on this issue. Elsewhere in his argument,Mellor says he
accepts,speakingof a visiblybluebird,"Ifthat'sa canary,it's notyellow"
(p. 245). So, I think,he should.But he is no more (or less) disposedto
infer "It'snot yellow"from "It'sa canary"thanhe is to infer "Egyptis
hot"from"Franceis big".On a liberalinterpretation, "to accepta condi-
tionalis to be disposedto inferconsequentfrom antecedent"does apply
here. Yourotherbeliefs must, in general,be appealedto in these infer-
ences. But then, if Q is one of your other beliefs, Q follows from the
assumptionthatP togetherwith yourotherbeliefs.
Relevanceis a context-dependent matter.Any two contingentproposi-
tions are mutuallyrelevantin some contexts,and mutuallyirrelevantin
others.Thomson(1990) has the exampleof someonecominghome say-
ing "Ifthere'sa bookon my coffee table,two GreatDanesarrivedat Pad-
dingtonStationthis morning"- and tells a storyin which the remarkis
apposite.The relevanceof the symptomsto the questionwhetherI have
the diseaseceases aftermoredirecttests havebeen carriedout. The rele-
vance of antecedentto consequentin a contingentconditionalceases
when the truthvalue of the consequentis establishedby perception.A
defenderof the Thesis (or the truth-functional conditional,or Stalnaker's
conditional)will claimthatquestionsof relevancebelongto the pragmat-
ics of communication.This is, I think,whatLewis calls a "spoilsto the
victor"issue: if the best overalltheoryallows thatthere are boringbut
acceptableconditionalswithmutuallyirrelevantparts,so be it. If not,not.
(4) It wouldbe wrongto read"Itis probablethatB given A" as "IfA,
then(it is probablethatB)". This wouldbe like the so-called"modalfal-
lacy"-of reading"Ifhe's sittingdown,thennecessarily,he's notstanding
up"with "necessarily"qualifyingthe consequentratherthanthe whole
thought.The modalfallacyhas the consequencethatall truthsareneces-
sarytruths:if A, then necessarilyA. And in this contextit has the conse-
quencethatall probabilitiesare 1 or 0. Forthe probabilityof A givenA is
1, andthe probabilityof A given -A is 0. If we readthis:If A, thenp(A)
= 1; if -A, then p(A) = 0; then, granted Av-A, we could validly derive
thatp(A) = 1 orp(A) = 0. "I'msurethatA if A"does not have the conse-
quencethatif A (is true),thenI'm surethatA (is true).
It is less of a howlerto thinkof yourdegreeof belief in B givenA as the
degreeof belief you would have in B if you were certainthatA. This is
typicallycorrect,but not invariablyso. Forthereareall sortof ways you

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
270 Dorothy Edgington

mightlearnthatA. Youthinkthe matchwill light if struck.Youlearnthat

it is being struck.Typically,you then thinkit will light. But not if you
learnthatit is being struckat the bottomof a swimmingpool.
Admittedly,in this last example,you did not expect it to be struckin
this way, or else you wouldn'thave thoughtit would light if struck.So
perhapsthinkingthatB if A is beingpresentlydisposedto believeB if you
learn that A. Mellor (1993) suggests this.35 But this won't quite do
either- thereis one particularkindof counterexample:
If Reaganwas in the pay of the KGB,we'll neverfind out.
SupposeReaganwas in the pay of the KGB;then,I judge,it's likely that
we'll neverfind out. But if I were to learnthathe was in the pay of the
KGB,I wouldnotthinkit likely thatwe'll neverfindout!Nor,pace Mel-
lor, am I presentlydisposed to believe the consequenton learningthe
antecedent.Whatthe exampleshows is thatsupposingthatsomethingis
trueis not alwaysequivalentto supposingyou knowit's true,or pretend-
ing you're certainthat it's true.36"Theprobabilitythat (yourdegreeof
belief in) B on the suppositionthatA"is intelligibleas it stands:thesefur-
therattemptsto gloss it areunsuccessfulandunnecessary.
Anotheridiomcomes in handy.Takethe specialcase whereyou havea
partition,fine enoughfor the problemat hand,of equallylikely alterna-
tives, or "worlds".The probabilityof an unconditionalproposition(A,
-A, A&B,etc.) is the proportionof worldsin whichit is true.The proba-
bilityof B givenA is the proportionof A-worldswhichareB-worlds.(The
proportionof A&B-worldsis the proportionof A-worlds,multipliedby
the proportionof A-worldswhichareB-worlds.)If we dropthe simplify-
ing assumptionthateach alternativeis equallylikely,we haveto replace
"proportion" by "weightedproportion", wherethe weightsarethe proba-
bilitiesof theworlds.Justfocusingon a singleprobabilitydistribution, we
can stick to the simpleridiom by artificiallysubdividingthe weightier
worldsintoslimmerones,indistinguishable forthe purposesathand,until
each propositionis true at some numberof equallylikely worlds.37We
mirrorthe structureof conditionalandunconditionalprobabilitiesby the
phrases"proportion of A-worldswhichareB-worlds"and"proportion of
worldswhichareA-worlds".
35Mellordoes not call this a judgementof conditionalprobability,reserving
thatnamefor somethinghe distinguishesfromit andrejects.See fn. 33 above.
36 W. V. 0. Quine(1966, pp.22-3), makesthisdistinctionin his solutionto the
surpriseexaminationparadox.
Exampleslike this aredueto Thomason(see VanFraassen,1980,p. 503). This
exampleis Lewis's(1986, p. 155).
37 To the diagramon p. 264, superimpose as manyequally-spacedhorizontal
lines as you needto get each memberof the partitiontruein some numberof el-
ementsof the resultingpartitionof equally-likely"worlds".

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 271

6. The bombshell38

6.1. At the beginningof ?5 we considereda formof objectionto various

proposedtruthconditionsfor conditionals:that a clear-headedperson
could have less confidencein the conditionalthanin the proposedtruth
condition,or vice versa.At the end of ?5.1 we asked:can we finda truth
conditionwhichis immunefromthis objection?Weturnedto a theoryof
uncertainty,and arrivedat the Thesis: a logical relationwhich governs
yourdegreesof belief in B if A, A&B,andA. Canwe findtruthconditions
for conditionals39whichfit the Thesis?Takeany two logicallyindepend-
ent propositions,e.g. "Annis in Paris"and"Bill is in Paris";call themA
and B; suppose you have a conditionaldegree of belief in B given A.
Whichtruthconditions(if any)aresuchthatyourdegreeof belief in their
obtainingmustmatchyourconditionaldegreeof belief in B givenA? Can
we finda propositionX suchthat,in anyconsistentbelief distributionover
the relevantpropositionsin whichb(A)? 0,
b(X) = b(B given A) [henceforth, the Equation],
so thatwe may say, consistentlywith the Thesis, "IfA, B" is trueif and
only if X?
Therecan be differentepistemicattitudesto the sameproposition.We
seek an X such that b(X) and b(BIA) = b(A&B)lb(A) cannot coherently
come apart:an interpretation such thatfor all consistentdistributionsof
belief over the relevantdomain40,b(X)= b(A&B)lb(A).In a given belief
distribution,theremayof coursebe a propositionC (orseveralsuchprop-
ositions)whichyou believe to the samedegreeas you believe B givenA.
But if someoneelse, or you in a differentinformationstate,may consist-
entlythinkC more,or less, likely thanB givenA, C is not the X we seek.
The bombshellis thatno propositionat all satisfiesthe Equation.If we
stickby the Thesis,we mustnot thinkof conditionalsas propositions,as
truthbearers.If belief thatif A, B fitsthe Thesis,it is nonsenseeven to say
thingsof the form"'IfA, B' is trueif andonly if, if A, B".Yourdegreeof
belief thatB is true,on the suppositionthatA is true,cannotbe consist-
ently and systematicallyequatedto yourdegreeof belief thatsomething
is true,simpliciter.
38 So described by Bas Van Fraassen (1976, p. 273); and by Stalnakerin a letter
to Van Fraassen published there, p. 302.
39 We are concerned here with indicative conditionals (see pp. 265-6 above).
The possibility of extending the Thesis to counterfactuals is discussed in ? 10.
40 I
shall take the qualification "in which b(A) ? 0" for granted except when it
is specially important.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
272 DorothyEdgington

Some pre-bombshellwritingsforetoldthis result.GilbertRyle (1950)

recommendedthinkingof conditionalsas "inferencetickets"ratherthan
statements.John Mackie (1973, p. 93) construedsaying "If A, B" as
assertingthatB withinthe scope of the suppositionthatA, and said this
view "abandonstheclaimthatconditionalsarein a strictsensestatements,
... thatthey arein generalsimplytrueor simplyfalse".Adams(1965 pp.
169-70, 1966 pp. 265-6) expresseddoubtsaboutthe applicationof truth
to conditionals, and developed a logic for conditionals construedin
accordancewith the Thesis.
But therewas no strongreasonfor holdingthattheremustbe an oppo-
sition between (e.g.) assertingthatB underthe suppositionthatA, and
sayingsomethingtrue.Indeed,it was (andis) hardto see how therecould
be an opposition:musttherenot be a distinctionbetweenwhenit is right,
andwhenit is wrong,to assertthatB underthe suppositionthatA, which
will yield a notionof truthandfalsity?
Stalnaker(1968) fully endorsed Ramsey's account of conditional
belief:addthe antecedenthypotheticallyto yourstockof beliefs,andcon-
sider whetheryou believe the consequent under that hypothesis. He
soughtappropriate truthconditionsto match:
Now we have foundan answerto the question,"Howdo we de-
cide whetheror not we believe a conditional statement?"the
problemis to makethe transitionfrombelief conditionsto truth
conditions;thatis, to find a set of truthconditionsfor statements
of conditionalformwhichexplainswhy we use themethodwe do
use to evaluatethem.Theconceptof a possibleworldis just what
we needto makethis transition,since a possibleworldis the on-
tologicalanalogueof a stockof hypotheticalbeliefs. Thefollow-
ing set of truth conditions, using this notion, is a first
approximation to the accountI shallpropose:
Considera possibleworldin whichA is true,andwhichotherwise
differs minimallyfrom the actualworld. "IfA, then B" is true
(false) just in case B is true (false) in that possible world. (1968,
pp. 33-4)
This was the first appearancein printof the "nearestpossible world"
approachto conditionals,andit was designedto providethe "ontological
analogue"of the Thesis. Stalnakerleft probabilisticconsiderationsaside
in this introductorypaper,announcingin a footnote (n. 17, p. 43) that
these would be elaboratedsubsequently(1970). But they were in the
background.The samefootnoterefersto Adams(1966), who hadshown,
for instance, that there are plausible counterexamplesto transitivity,
strengthening andcontraposition,andthatthiswas to be expectedfor con-
ditionalswhich satisfiedthe Thesis. (The logic Adamsdevelopedon the
basis of the Thesis is discussedin ?7.2.) Stalnakergave his own coun-
terexamplesto these inferencepatterns(1968, pp. 38-9), andformulated

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 273

a logic (p. 37) which is identicalto Adams'over theircommondomain.

(Adams'logic is restrictedto sentencesin which "if", if it occursat all,
occursas the main connective.Stalnaker'sis not so restricted:once we
have truthconditions, we have somethingwhich embeds naturallyin
largersentences.)
Coincidencein logic does not guaranteecoincidencein interpretation.
But Stalnaker'sattitudeto interpretationwas minimalist. He did not
expecta reductiveanalysisof "if"in termsof a substantivenotion,"min-
imallydifferentA-world".Rather,we shouldthinkof the "minimallydif-
ferentA-world"as the worldthatwill be actualif A is. He laterdescribed
his conditionalpropositionsas "aprojectipnof epistemicstrategyontothe
world"(1984, p. 119).
Allan Gibbard(1981, p. 211) describesit as "littlemore thana coin-
cidence"thatAdams'and Stalnaker'slogics agree. In one way it is no
coincidence:each is motivatedby the same notionof conditionalbelief,
which Stalnaker'struthconditionswere intendedto fit. Nor is it so sur-
prising that the rich frameworkof possible-world semantics, in the
hands of someone as expert as Stalnakerin its manipulation,should
yield the right structure.With agreementin logic, and no leverage on
the semanticsindependentlyof the notion of conditionalbelief, Stal-
naker'sclaim,to have identifiedthe propositionwhose belief conditions
fit the Thesis (1970, p. 107, p. 120), had an irrefutableair.So it came as
a bombshellwhen Lewis, at the 1972 meeting of the CanadianPhilo-
sophicalAssociation,refutedit, provingthat there is no propositionat
all such that your degree of belief in its truthsystematicallymatches
your degree of belief in B given A (Lewis, 1976, 1986a).A conditional
degreeof belief is not equivalentto a degreeof belief that[somethingor
other]is true.
6.2. Therearefourbombshellscomingup. Hereis the base result.There
is no propositionX suchthatp(X) =p(B IA) in all probabilitydistributions
in whichthese aredefined.(A probabilitydistributionis an assignmentof
non-negativenumbersto the membersof a partitionwhichsum to 1.)
Supposethereis suchanX. Wefirstshow somethingof the logicalrela-
tionshipsbetweenX andA:X is (a) compatiblewithA, and(b) compatible
with -A, but (c) not entailedby -A, i.e., X may or may not be trueif -A
is true.
Proofs,in reverseorder:(c) Thereareprobabilitydistributionsin which
p(-A ) is high andp(BIA) low. (e.g., let p(-A ) = 0.9; p(A&B)= 0.01;
p(A &-B) = 0.09. p(BI A) = 0. 1.) So thereareprobabilitydistributions in
whichp(-A ) is highandp(X) [=p(BIA)] low. So -A cannotentailX: if
it did, X wouldbe truethroughoutthe -A-worlds, and could not be less
probablethanp(--A). So, in some -A -worlds,X is not true.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
274 DorothyEdgington

Similarly:(b) there are probabilitydistributionsin which p(BIA) is

high andp(A) is low; hence in whichp(X) is high andp(A) is low. So X
cannotentailA: X mustbe truein some -A-worlds. And a parallelargu-
ment will show that (a) X cannotentail -A: X must be truein some A-
worlds.Thereis nothingsurprisingin thesefacts.Now to the mainpartof
the proof:
(i) p(B IA) depends only on how probabilities are distributed in the A-
worlds(thepartof the partitionin whichA is true).Fixp(A) andp(A&B),
andp(BIA)is fixed.
(ii) Any propositionX whichsatisfiestheEquationmustbe truein some
butnot all -A-worlds, andtruein someA-worlds,as was shown.So p(X)
dependsnot onlyon how probabilitiesaredistributedin theA-worlds,but
also on how they aredistributedin the -A-worlds.
(iii) There are distinct probabilitydistributionswhich agree in all
assignmentsin the A-worlds, but disagree in assignmentsin the -A -
worlds.They will agreeon p(A&B) andp(A),andhenceonp(BIA). And
they will agreeon p(A&X).But they will disagreeon p(-A&X). As p(X)
= p(A&X) + p(-A&X), they will disagree on p(X). So there are distribu-
tions in which p(BIA) ? p(X). Endof proof.41
We can illustratethe proofby takingStalnaker'sconditional("A>B")
for X. We get a partitionof logical possibilitiesof the followingshape.

A B A>B Pi P2

1. T T T 0.4 0.4
2. T F F 0.1 0.1
3. F T 0.4 0.1
4. F F 0.1 0.4

(Note: (a) If A is true, the minimally differentA-world is the actual

world;this explainsthe top two lines; (b) if A is false, then,whateverthe
truthvalue of B, it may or may not be the case thatthe minimallydiffer-
ent A-worldis a B-world.This explainslines 3 and4. (c) Appearancesto
the contrary,we need not assume that the propositionX (hereA>B) is
necessarily either true or false. What is at issue is whetherthere is a
propositiontheprobabilityof whose truthis, in all distributions,p(BIA).
We can leave open whether,wheneverit is not true,it is false. Hencewe
4' A sketchof a proofalong these lines is given by I. Carlstromand C. Hill
(1978), in theirreviewof Adams(1975).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals275

could construe, here and below "F" as "not true", and "-_" as "It is not
truethat".)
On the rightwe havetwo probabilityfunctionsoverthe partition.They
agreein theA-worlds.Ineachp(BIA) = 0.4/0.5 =0.8. Inthe first,p,(A>B)
= 0.4 + 0.4 = 0.8 = pl(BIA). In the second, p2(A>B) = 0.4 + 0.1 = 0.5 ?
p2(BI A).
Does the baseresultrefuteStalnaker'sclaim?An argumentthatit does
wouldbe this.Wehaveabovefourclearlyspecifiedpossibilities,one and
onlyone of whichwill obtain,like a 4-horserace.Allow background infor-
mationto vary.Any probabilitydistributionover four such possibilities
mightrepresenta not-irrational belief distributionin some stateof back-
groundinformation.(A crudeillustration:as I ponderthefourpossibilities
above,anOracletells me "I'llgive you a hint:either2 or 3 is thetrueone".
Acceptingthehint,I dividemy belief equallybetween2 and3. If I amcer-
tainthatthe Oraclespoketruly,my b(BIA) andb(A>B)arerespectively0
and0.5. If I amnearlycertainthattheOracleis right,theyareclose to these
numbers.)Thisargumentwill be blockedif it canbe shown,in a non-ques-
tion-beggingway, thatfor some probabilitydistributions(the conflicting
ones) thereis no stateof informationin whichtheywouldrepresenta rea-
sonablebelief-distribution. I do not know any such argument.
But thereis an interpretation of Stalnakerwhich is immunefrom the
base result.He couldbe interpretedas stipulating that,as well as satisfy-
ing the partitionprinciple,belief distributionsinvolvingconditionalprop-
ositionsandtheirpartsmustsatisfythe Thesis:P2 is to be ruledout from
the class of consistent belief-distributions.This fits with Stalnaker's
image of a conditionalpropositionas a "projectionof epistemicstrategy
ontothe world".Thefallacyin the argumentabovewas, I suppose,to take
"conditionalproposition"too realistically:thereareno facts about"near-
est A-worlds"independentlyof ourepistemicstrategies.
The stipulationis consistentfor a single conditionalin a single belief
distribution.But stipulationshave consequences.This one has untenable
ones when we considerthe originalconditionalin differentbelief distri-
butions(as Lewis showed);it also has untenableconsequencesfor other
conditionalsin the samebeliefdistribution(as Stalnakerhimselfshowed).
6.3. Hereis a simplifiedandrelativelyinformalversionof Lewis'sproof
(1976).Theproofinvolvesaninitialbeliefdistribution,whichis ex hypoth-
esi reasonable,in whichb(A&B)andb(A&-1B) arebothnon-zero.Wefind
out how it mustchange(given the Equation)if the believerwere to learn
certainthings,anddeducewhatit musthavebeenlike in the firstplacefor
suchchangesto be rationallypermissible.Forthislastpartwe needa prin-
cipleaboutbeliefrevision.I shallappealto thefollowingprinciple,which
is weakerthanLewis's,a consequenceof his, butall thathe needs:

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
276 DorothyEdgington

(RP) If a person rationallyhas a non-zero degree of belief in

C&D;andthenlearnsC for certain,andnothingelse of relevance,
it is alwaysrationalfor himto continueto havea non-zerodegree
of belief in D.
Contraposing theprinciple:if learningC forcertainrendersimpermissible
any degreeof belief in D otherthan0, thenb(C&D)cannotrationallybe
otherthan0 beforeyou learntC.
First,supposethe personwere to learnfor certainthatB, andnothing
else of relevance.In the new distributionb' thatwould result,b'(B) = 1.
As his original b(A&B) ? 0, b'(A) ? 0 (by RP). So b'(BI A) = b'(A&B)I
b'(A) = 1. So b'(X) = 1, and b'(-X) =0. So, by RP, in the former distribu-
tion, it is not rationallypermissibleto have a non-zerodegreeof belief in
B&-X. (If it were,learningB wouldnot force-X to zero.)
Second,supposethe personwereto learnfor certainthat-B, andnoth-
ing else of relevance.In thenew distributionb"whichwouldresult,b"(B)
= 0. As b(A&-B) was non-zero, b" (A) ? 0. So b" (B IA) = b" (A&B)Ib"(A)
= 0. So b"(X)= 0. By the samereasoningas above,this meansthatin his
originaldistributionb, b(-,B&X) must have been 0. For if it were not,
learningjust -B wouldnot forcehim to assign0 to X.
We have proved that in any reasonablebelief distributionin which
b(A&B) and b(A&-B) are non-zero, b(B&-X) = 0 and b(-1B&X)= 0. It is
an elementary consequence of these two facts that b(B) = b(Xi). [b(B) =
b(B&X); and b(X) = b(B&X).] So, in any such distribution, b(B) = b(X) =
b(BIA). But this is absurd!Take any three-waypartitione.g. {C&D,
C&-D, -_C} and a distributionof belief which is positive (e.g. 1/3) for
each member b(C&D)Ib(C) = '/2. b(C&D) = 1/3. b((C&D)IC) ? b(C&D).
Yet this example satisfiesLewis's initial conditions;b(C&(C&D))and
b(C&-1(C&D)) are both positive. The absurditymay be statedroughly
thus:takeany two propositionssuchthatthe firstdoes not entailthe sec-
ond;learningthatone is truehas no relevanceto how muchyou should
believe the other.42
6.4. VanFraassen(1976) objectedthatit is unreasonableto assumethat
the sentence"IfA, B" be interpretedas the same propositionin different
belief distributions.He labelledthis assumptionLewis's "metaphysical
realism"(p. 252). The label is hardlyfair in this context:the assumption
amountsonly to the claimthatwe cantakedifferentepistemicattitudesto
a propositionwithoutchangingthe subject(see Lewis, 1976, p. 138).
Replyingby letterto VanFraassenin 1974 (VanFraassen,1976,pp. 303-
4), Stalnakerproducedhis own versionof the bombshell.It is described
42 Anthony'Appiah (1986) objectedto Lewis'sproof,claimingthatit is never
reasonableto have a degree of belief of strictly 1 in a contingentproposition.
Lewis (1986a)gave a new proofwhichdoes not requirethatassumption.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 277

by Gibbard(1981, pp. 219-20). Wedo not needto invokedifferentbelief

distributionsto derivean absurdity,at leastfor Stalnaker'sproposedtruth
conditions.He showsthatif, in somedistribution,b(A>B)= b(BI A), then
thereexist, in the same distribution,two furtherpropositions,C and D,
suchthat,demonstrably,b(C>D)? b(DIC).
Hereagain,on the left, is a partitionfor Stalnaker'sconditional

A B A>B b C D E C>D

1. T T T 0.25 T F F F

2. T F F 0.25 T T F T
3. F T 0.25 F F T F
4. F F 0.25 T F F F

Considera belief distributionwhichassigns0.25 to each line, andso sat-

isfies the Equation:b(A>B) = b(BIA) = 0.5. (Any numbersotherthan
zeroswhichsatisfythe Equationwill do. I just pick the easiest.)Let E be
-A&(A>B): E is the propositiontruejust at line 3. Let C be -E. C is the
propositiontrueat lines 1, 2 and4. Let D be A&-B. D is truejust at line
2. Now b(DIC) = 1/3 [line2 (0.25) dividedby the sumof lines 1, 2 and4
(0.75]. WhataboutC>D? It is trueat line 2 (becauseits antecedentand
consequentare);false at lines 1 and4 (becauseits antecedentis trueand
its consequentfalse). Whataboutline 3? StalnakershowsthatC>D andE
are incompatible.So C>D is false at line 3. So it is truejust at line 2. So
b(C>D) = 0.25 ? b(DI C). Why areC>D andE incompatible?C has the
form(A v (-A &G)) [I abstractfromthe structureof G]. So C>D has the
form (A v (-A&G)) > (A&-B). This implies A>-B43, while E implies
A>B. For consistentA, A>B andA>-B are incompatibleon Stalnaker's
logic andsemantics.
Admittedly,"C>D"is a somewhatcontrivedproposition.But its exist-
ence is forceduponus by the assumptionthatthereare truthconditions
43 A proofthatC>D entailsA>-B, in termsof Stalnaker'ssemantics:C>D
says thatthe nearest(A v (-A &G))-worldis an (A&-B)-world. Supposethat's
true.Now the nearest(Av (-A &G))-worldis eitheranA-worldor a -A -world.
Supposeit's a -A-world. Thenthe conditionalsays it's an (A&-B) world.Con-
tradiction.So the nearest(A v (-A&G))-worldmustbe an A-world.Go to the
nearestA-world.It will be the nearest(A v (-A &G)) world.So it's an A&-B-
world.So it's a -B world.So A>-B. (Theproof,of course,canbe doneformally
in Stalnaker'slogic. The only not entirelytrivialstepsinvolve(1) the incompati-
bility of A>B andA>-B, for consistentA; and(2) that((A>C)v (B>C))follows
from((AvB) > C).)

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
278 DorothyEdgington

compatiblewith the Thesis. (It's not easy to get your mind roundthe
G6delsentence,either.)44
6.5. The following argumentyields, I believe, a diagnosisof the trouble.
Let us examinethe relationshipbetweenb(B givenA) andb(ADB). There
aretwo specialcases in whichtheymustbe equal:(1) you arecertainthat
A&-B is false (but not certainthatA is false); then b(B given A) and
b(ADB) are both 1; (2) you are certain thatA; then b(B given A) = b(ADB)
= b(B). These cases apart, in all belief distributionsb(B given A) <
b(ADB).
Theeasiestway to see thisis to comparehowmuch b(BIA) andb(ADB)
differ from certainty.Here is a partition.Adjustingthe positionsof the
innerhorizontallines will representdifferentbelief distributionsover it.

A A&BR
A 1

-At

Theamountby whichADBdiffersfromcertaintyis simplythe proportion

of the whole assigned to A&-B: writing"u"for "the uncertaintyof',
u(ADB) = (1 - b(ADB)) = b(A&-B). The amount by which b(BIA) differs
fromcertaintyis theproportionof A whichis assignedto A&-nB:u(BIA)
= (1 - b(B IA)) = b(A&-_B)/b(A). Now, b(A&-B) is a greaterproportionof
b(A) than it is of the whole-except when b(A&-B) = 0, or b(A) = 1.
Henceb(BIA) is moreuncertainthanb(ADB),exceptin thesetwo special
cases,wheretheyareequal.If -A is large,A&-J.Bmustbe small;butA&B
maybe smallerstill, in whichcase b(BIA) is low but b(ADB)is high. An
exampleof the difference:how likely is it thatif this (fair)die landsan
even number,it will land6? b(six given even) = 1/3. b(evenv six) = b(not
even, or six) = 2/3: if it lands 1, 3, 5 or 6, the truth-functional
conditional
is true.
Whenwe tryto equateb(BIA) with a degreeof belief in a proposition,
b(X),we findwe haveincompatiblerequirements uponit:
44 A moregeneralproofthat,withina single belief distribution, not all condi-
tionalprobabilitiescan be probabilitiesof the truthof a proposition,is given by
Alan Hajek(1989).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals279

(1) Clearly,ADB doesn't entailX. (If it did, one could not coherently
havea higherdegreeof belief in ADBthanin X; butin general,b(ADB) >
b(BI A) = b(X).) So therearepossiblesituationsin which ADB is trueand
X is not true.Hence,someonewithjust enoughinformationto be certain
thatADB does nothaveenoughinformationto be certainthatX:rulingout
just those situationsin which ADB is false, i.e. rulingout just A&-B,
leaves openthe possibilitythatADBis trueandX is not.
(2) But, on the contrary,by the firstspecialcase, rulingoutjustA&-B
is enoughfor b(BIA) = 1 = b(X). Contradiction.
(Theprincipleappealedto in (1) is:
If C does not entailD (if therearepossiblesituationsin whichC
is trueandD is not true),thencertaintythatC is consistentwith
less-than-certaintythatD.
Here is a putativeobjection:let D be "I am certainthat C". C does not
entail"IamcertainthatC".But,it mightbe held,certaintythatC is incon-
sistentwithless-than-certainty thatI am certainthatC. Now, eitherwe do
not have infallible access to our own epistemic states, or we do. If we
don't, we have no counterexample: being less-than-certain
thatI'm cer-
tainis not incompatiblewith being certain.If we do, we restrictthe prin-
ciple to beliefs about whose truthuncertaintyis possible. Uncertainty
aboutconditionalsis possible,so my use of the principlesurvives.
Moregenerally:the only possiblesourceof troublefor the principle,as
far as I can see, will come frombeliefs aboutone's own epistemicstate
(troubleakinto Moore'sparadox:p andI don'tbelieve thatp). Provided
thatconditionalsaboutmatches,kangaroos,Ann'sandBill's whereabouts,
etc. are not propositionsaboutthe believer'smentalstate,the use of the
principlestands.)
If we acceptthis principle,the above argumentthrowssome light on
the puzzlewhicharoseat the end of ?2, p. 247. Twoprimafacie desirable
propertiesof indicativeconditionaljudgements:
(i) Minimal certainty that AvB (ruling out just -A &-,B) is
enoughfor certaintythat if -A, B; changingthe negationsign,
minimalcertaintythat-AvB (rulingout just A&-_B)is enough
for certaintythatif A, B.
(ii) It is not necessarilyirrationalto disbelieveA yet disbelieve
thatif A, B.
Thetruth-functional accountsatisfies(i) butnot (ii). Strongertruthcondi-
tionsmay satisfy(ii), if they allowthatthe conditionalmaybe false when
A is false. Butthey cannotsatisfy(i): for any strongertruthcondition,rul-
ing outjust A&-B leaves open the possibilitythat"IfA, B" is not true.
The Thesis satisfies both (i) and (ii): (i) ruling out just A&-B makes
b(B IA) = 1. Yet(ii) it is possibleto have b(-A ) highyet b(B IA) low. So

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
280 DorothyEdgington

it is incompatible with both truth-functional truth conditions, and

stronger-than-truth-functional truthconditions.
We may generalize. Take any proposition. Either it is entailed by
-(A&-B), or it is not. If it is, it will satisfy(i) but not (ii) (when substi-
tutedfor "if A, B"). If it is not, it may satisfy (ii), but cannotsatisfy (i).
Conditionaljudgementsinterpretedaccordingto the Thesis satisfyboth
(i) and(ii). So they cannotbe interpretedas belief in any proposition.
How does the Thesis achievewhatbelief in no propositioncan?Well,
supposeA&-B has been ruledout. This is enoughfor certaintythatB
givenA, not becausesomepropositionor otheris truewheneverA&-B is
false; but because B is true in all the worlds that concernthe question
whetherB if A-the A-worlds.Whatgoes on in the -A-worlds has noth-
ing whateverto do with thoughtsabouthow likely it is thatB givenA. A
high degreeof belief in -A is consistentwith a low degreeof belief in B
given A, not becausesome propositionis false in some -A -worlds;but
becausethe fact thatb(-A) is highhas no bearingat all on whethermost,
or the most probable,A-worldsareB-worlds.
6.6. Anyone interestedin the concept of truthshould take note of this
result.It is an empiricalquestionhow well the Thesisfits ourpracticeof
assessingconditionals,andit is a deeperquestionwhether,andif so why,
it is a good practiceif it does. But-to say the least-there could be
peoplewho use "if"this way.The resulttells us thatthey do not use "if"
to express propositions, evaluable in terms of truth. A previously
unnoticedtest for the applicabilityof the conceptof truthhas presented
itself: if judgements of a given type are subject to uncertainty, do
uncertainjudgementsof this type fit the structureappropriateto uncer-
taintyabouttruthbearers?Someonemay objectthatthis whole theoryof
uncertainty,based on the structureof probability,is wrong. Then it is
incumbentupon him or her to give an alternativetheoryof the logic of
uncertainty.It is too importanta phenomenon-as it appliesto condition-
als, andas it appliesto otherjudgements-to ignore.

7. Is truth necessary?

.7.]. Compounds. Couldit be thatthe mistakephilosophershavemade,in

tryingto understandconditionals,is to treatthem as partof fact-stating
discourse- as representingthe world as being a certainway- and that
this is not their function?If so, pressingquestionsarise. What are we
doingwhenwe say or thinkthatif A, B, if not sayingor thinkingthatthis
is how thingsare?Whatdo we aim at, if not to stateor thinkthe truth?

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 281

Whatis it to be rightor wrongin so sayingor thinking?How are we to

understandthe role of conditionalsin arguments,if not in termsof pre-
servingtruth?And thereis the questionLewis raises:
I have no conclusiveobjectionto the hypothesisthatindicative
conditionals are non-truth-valuedsentences .... I have an incon-
clusive objection,however:the hypothesisrequirestoo muchof
a fresh start.It burdensus with too muchwork still to be done,
andwastestoo muchthathasbeendonealready.... [W]hatabout
compoundsentencesthathave conditionalsas constituents?We
thinkwe knowhow the truthconditionsfor compoundsentences
of variouskindsaredeterminedby the truthconditionsof constit-
uentsubsentences,butthis knowledgewouldbe useless if anyof
these subsentenceslackedtruthconditions.Eitherwe need new
semanticrulesfor manyfamiliarconnectivesandoperatorswhen
applied to indicative conditionals ... or else we need to explain
away all seemingexamplesof compoundsentenceswith condi-
tionalconstituents.(1976, pp. 141-2)
Too muchof the ship would need rebuilding,says Lewis. However,the
particularplankon whichhe restshis case is farfromsound.We do think
we know how the truthconditions of compoundsentences of various
kindsaredeterminedby the truthconditionsof constituentsubsentences.
But this knowledgeis useless whenit comes to conditionalsubsentences.
We do not have a satisfactorygeneralaccountof sentenceswith condi-
tionalconstituents.This may be becausewe have not yet figuredout the
truthconditionsof conditionals.Orit maybe becausetheydon'thaveany.
First, the truth-functionalaccount45gives bizarreresults for com-
poundsof conditionals.Forexample,
Either,if the Queenis athomeshe is worryingaboutme, or,if the
Queenis not at home she is worryingaboutme
is a tautology;so, if I reject the first disjunct,I must (on this account)
acceptthe second.And the following argumentis valid if we treat"if"
truth-functionally:
If God does not exist, then it's not the case that if I pray my
prayerswill be answered.I do not pray.ThereforeGod exists.
Second,the attemptsby GriceandJacksonto explainawaythe seemingly
paradoxicalfeaturesof truth-functional conditionals,focusexclusivelyon
whatmoreis neededto justifythe assertion of "IfA, B", beyondthebelief
thatthe truthconditionis satisfied.Theyaresilentaboutthe occurrenceof
conditionals,unasserted,as constituentsof longer sentences.Jacksonis
expliciton thispoint."Itsimplydoesn'tfollow fromthefactthatI give (A
45 Lewis holds thatindicativeconditionalshave truth-functionaltruthcondi-
tions,andacceptsJackson'saccountof theirassertabilityconditions.See thePost-
of ConditionalsandConditionalProbabilities"
scriptto "Probabilities (1986, pp.
152-6).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
282 Dorothy Edgington

-> B) truthconditions that I must find, say, [(A-> B) - C)] ... a meaning-
ful sentence" (1987, p. 129). That is, it is compatible with his account of
conditionals as (a) having truth-functionaltruth conditions and (b) being
subject to a special rule of assertability, that unasserted conditionals are
meaningless. If we want to give them meaning, we have more work to do.
Third, non-truth-functional truth conditions also have controversial
consequences for compounds of conditionals. Stalnaker adopts as a logi-
cal truththe Law of Conditional Excluded Middle: (if A, B) or (if A, -B).
Lewis admits that there is much to be said for this -he calls it "[t]he prin-
cipal virtue and the principal vice of Stalnaker's theory" (1973, p. 79)
but thinks there is more to be said against it.46 Another controversy is
whether "If A, then if B then C" is equivalent to "If A and B, then C". We
do treat these forms as interchangeable, it seems. But on Stalnaker's
semantics (or Lewis's for counterfactuals) neither entails the other. Con-
sider (1) "If it rains or snows tomorrow, and it doesn't rain tomorrow, it
will snow tomorrow".That, it is agreed, is unassailable. Now consider (2)
"If it rains or snows tomorrow, then if it doesn't rain tomorrow, it will
snow". We read that in the same way-just as trivial. But on Stalnaker's
semantics (2) may well be false. If snow is a far-out possibility, and rain
a close-in possibility, then in all the closest worlds in which it rains or
snows, it rains but doesn't snow. Then, the closest world in which it rains
or snows (viz. rains) may be such that the closest world to it in which it
doesn't rain, it doesn't snow either. So for Stalnaker "If it rains or snows
tomorrow, then if it doesn't rain, it won't snow" may be true.
This is somewhat counterintuitive. However, maintaining the equiva-
lence of (1) and (2) also exacts a price: modus ponens for conditionals
with conditional consequents. I accept, as trivial, "If it rains or snows,
then if it doesn't rain, it will snow". I accept that it will rain or snow
(because I am nearly certain that it will rain). But I deny that if it doesn't
rain it will snow (because I'm virtually certain that if it doesn't rain, it
won't snow either).47
Turning from particular theories to the phenomena themselves, let's
first consider disjunctions of conditionals. "Or" is a very useful word,
especially when it connects things we can be uncertainabout, for often we
can be confident that A or B, while not knowing which. We can be uncer-
46
Lewis's remark is about the tenability of this law for counterfactuals.
Stalnaker does not think that there always must be a closest A-world. When B
is true in some but not all of the closest, he holds that each disjunct is indetermi-
nate but the disjunction determinately true.The analogue is the treatmentof vague
terms such that an object may be not determinately red, nor determinately orange,
but determinately either red or orange (1981).
47 Examples like this are the topic of Vann McGee's "A Counterexample to
Modus Ponens" (1985). The phenomenon is mentioned by Adams (1975, p. 33).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 283

tainaboutconditionals.Yet"Either(if A, B) or (if C, D) -but I don'tknow

which"is a formof thoughtthatis virtuallyuninstantiated. An agile mind
will leap to the challenge and instantiateit-but once you've seen the
results,you see why we have no use for suchthoughts."Eitherif I go out
I'll get wet, or if I turnthe television on I'll see cricket(I don't know
which)."That'snottoo hardto interpret:if it's rainingandI go outI'll get
wet; if it's not rainingandI turnthe televisionon, I'll see cricket;andof
course,it's eitherrainingor it isn't. (The disjunctionof conditionalshas
disappeared.)Othersneed morebackground:"Either,if you open box A,
you'll get ten pounds,or, if you open box B, you'll get a button,I don't
know which."If Fredis in a good mood he has put ten poundsin box A
andtwentypoundsin box B. If Fredis not in a good mood he has put a
paper clip in box A and a button in box B ... . Again, the disjunction of
conditionalsis an exceedinglybad way to convey the informationyou
have, and once the necessary backgroundis filled in the disjunction
belongselsewhere.On the otherhand,our genuineneed for disjunctions
shows up naturallyinside a conditional:"IfA, theneitherB or C (I don't
know which)".Some apparentdisjunctionsof conditionalsare no such
thing:"Eitherwe'll have fish, if Johnarrives,or we'll have left-overs,if
he doesn't".
Turnto negations.If someonemakesa remark,e.g. "Itwill rain",you
may disagreein two ways, one strongerthanthe other.Youmay say "No
it won't";or you may say "I wouldn'tbe so sure".In the firstcase, you
assertthe negationof the firststatement;in the second,you areprepared
to assertneitherit norits negation.Similarly,if someonesays "Ifit rains,
they will be delayed",you may disagreein two ways. If you disagree
strongly,you will say "No, if it rains,they won't be delayed".Or again,
you may go less far, andexpressuncertaintyaboutwhetherthey will be
delayedif it rains.If the analogyholds, thenA is to -A as "IfA, B" is to
"IfA, -B". And "It'snot the case thatif A, B" has no clear established
sense distinguishablefromthis.
Conditionalsin antecedentsof otherconditionalsarealso problematic.
Gibbardsuggests (1981, pp. 234-8) that we have no general way of
decodingthem, and some cannotbe deciphered,for example"If Kripke
was thereif Strawsonwas, thenAnscombewas there".If someoneutters
a sentenceof this form,we do ourbest to interpretit by adhoc strategies.
Forinstance,we can sometimesidentify,in context,the obviousbasis,D,
for an assertionof "IfA, B", andinterpret"If (B if A), then C" as "If D
thenC":"Ifthe lightwill go on if you pressthe switch,the electricianhas
called"(If the poweris on, the electricianhas called).MichaelDummett
(1973, pp. 351-4; see also 1992, pp. 171-2) suggeststhat some may be
understoodas saying "If you acceptthatB if A, you must surelyaccept

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
284 Dorothy Edgington

this":"IfJohnshouldbe punishedif he tookthe money,thenMaryshould

be punishedif she tookthe money".
Considerthe schema,
"IfA, B" is trueif andonly if, if A, B.
It makestwo claims:
If (if A, B), then"IfA, B" is true.
If not (if A, B), then"IfA, B" is not true.
If negationof conditionals,and conditionalantecedents,are ill-under-
stood,so is the schema48,as Dummettcomments(1992, p. 171):"we do
not know how to interpretthis, becauseit is not our normalpracticeto
applynegationto an entireconditionalstatement";and "we have hardly
any use, in naturallanguage,for conditionalsentences ... in which the
antecedentis itself a conditional,andhence we cannotgraspthe content
of the principle".
Conditionalsdo not go into truth-functionalcontexts, or into each
other,easily, then. (Appiah(1985, pp. 205-10) argueslikewise.) Those
we do understand,e.g. conditionalsin consequents,we understandas
equivalentto sentenceswithoutembeddedconditionals.The facts square
at leastas well withthehypothesisthatconditionalsdo nothavetruthval-
ues as with the hypothesisthatthey do. (In ?9.4 I examinesome creative
attemptsto develop a languagewith compoundsof conditionalswhich
satisfythe Thesis.)
7.2. Validity. Turnto the question of the validity of argumentswhich
involveconditionals.Anotherreasonfor disinclinationto rebuildthe ship
mightbe put:"Validityis thenecessarypreservation of truth.Conditionals
occurin valid arguments.So conditionalsmusthave truthvalues".This
conceptionof validitymay be too narrow,independentlyof conditionals.
Thereare valid argumentsinvolvingmoraljudgements,but it is contro-
versialwhethermoraljudgementshave truthvalues.Legalexpertsspend
theirlives derivingconsequencesfromlaws, yet it's notobviousthatlaws
have truthvalues.
Adams(1966; 1975, Ch. 2) gives an accountof validityfor arguments
involving conditionalswhich conformto the Thesis. His methodis far
from ad hoc: it teaches us something about classical validity too. He
shows that classical, truth-preserving valid argumentsare, in a special
senseto be madeprecise,probability-preserving. Andthispropertycanbe
generalizedto applyto argumentswith conditionals.The valid ones are
those which, in the requiredsense, preserveprobabilityor conditional
probability.
48 Wright(1992, pp. 12-20) arguesin this way (not in connectionwithcondi-

tionals)thatthe schemahas moresubstancethanmightappearat firstblush.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 285

Begin with valid argumentswhich don't containconditionals.We use

themin arguingfromcontingentpremisseswhichareoftenbelievedwith
less thancertainty.The questionarises:how certaincan we be of the con-
clusion of the argument,given thatwe think,but are not sure, that the
premissesare true?Call the uncertaintyof a propositionone minus its
probability.Adamsshowsthis:if (andonly if) the argumentis valid,then
in no probabilitydistributiondoes the uncertaintyof the conclusion
exceed the sum of the uncertaintiesof the premisses.Thus, if I have a
valid argumentwith two premisseseach at least99%probable,this guar-
antees thatthe conclusionis at least 98% probable.In this sense, valid
argumentsare probability-preserving. (They are not probabilitypreserv-
ing in a different,strongersense:the probabilityof the conclusioncan be
less thanthe probabilityof each individualpremiss.The LotteryParadox
showsthisvividly.Wecan'texpectthatmuch.Theconclusioncan inherit
a risk of falsehoodfrom each premiss,and hence be less probablethan
each. Still, Adams'resultvindicatesdeductivereasoningfrom uncertain
premisses,providedthey arenot too uncertainandtherearenot too many
of them.)
Thisis an independentlyusefulandimportantconsequenceof classical
validity,then.Now Adamsextendsthisideato argumentscontainingcon-
ditionals.Takea languagewith "and","or","not",and "if"-but with
"if" occurringonly as the mainconnectivein a sentence.(Thuswe put
aside compoundsof conditionals.)Takeany argumentformulatedin this
language.Considerany probabilitydistributionover the sentencesin the
argumentwhich assigns non-zeroprobabilityto the antecedentsof all
conditionals,thatis, any assignmentof numbersto the non-conditional
sentenceswhichconformsto the PartitionPrinciple,andanassignmentof
numbersto the conditionalsentenceswhich conformsto the Thesis:p(B
if A) = p(A&B)lp(A).Extendthe term"uncertainty" to cover conditional
uncertainty:the uncertaintyof "If A, B" is one minus p(A&B)lp(A).
Definea valid argumentas one such thatthereis no probabilityfunction
in whichthe uncertaintyof the conclusionexceeds the sum of the uncer-
tainties of the premisses.And a nice logic emerges-the same as that
given by Stalnaker(1968),restrictedto simpleconditionals.Forexample,
if p(A) = 0.9 andp(B if A) = 0.9, we can show that the lower limit for p(B)
is 0.81 (modusponenshas a slightlyhigherlowerlimitfor the conclusion
thancan be guaranteedin general).
Wesaw above(p. 278) thatin all distributions,p(ADB) ? p(BIA). Take
an argumentwithconditionalsamongthe premissesbuta non-conditional
conclusion.Supposeit is validif we interpret"if"truthfunctionally.Then
it is also validin this probabilisticsense -if the conclusionfollows from
the weakerADB premiss,it follows fromthe stronger"IfA, B". But not

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
286 Dorothy Edgington

all truth-functionallyvalid argumentswith conditional conclusions

remainvalid:the premissesmay entailthe weakerADByet not entail"If
A,B".
ConditionalProoffails. Forexample,(1) "-(A&B);A; so -B" is valid,
but(2) "-(A&B);so, if A, -B" is not.
Probabilitiescan be modelledby proportions,and I shall use them to
illustratethe structurebehindthesefacts.(1) If almosteverythingis A, and
almostnothingis A&B,it follows thatalmostnothingis B. Indeed,if 99%
of the things in questionareA, and only 1%areA&B, so that99% are
-(A &B),it follows thatat most2%areB, at least98%are-B. That'sthe
structurebehindthe validityof the firstcase. (2) Supposethat99%of the
thingsareneitherA norB, andthe remaining1%areA&B.Thus,99%are
-,(A&B)but every A is B-0% of the As are -B. That's the structure
behindthe invalidityof the secondcase.49(Ann andBill areinseparable.
I can believe that it's not the case thatAnn-and-Billare there,without
believingthatif Ann is there,Bill isn't.)
All the departuresfrom truth-functional validitycan be tracedto the
failureof ConditionalProof.In the followinglist,the inferenceon the left
is valid, its partneron the right,derivableby a step of conditionalproof,
is not.
Valid Invalid
(1)A;BI-A A HIfB,A
(2)AvB; -A [-B AvBI-If -A,B
(3)-(A&B);A H-B -(A&B) HIf A, -B
(4) If A, B; if B, C; A [ C IfA,B;ifB,C HIfA,C
(5) IfA, B; -B -A If A, B If -B, -A.
Modelslike the abovewill show this.
Hereis one reasonwhy the argumentson the rightseem valid:if you
are 100%certainof the premisses-if you give themprobabilityor con-
ditionalprobability1-you mustgive the conclusionprobabilityor con-
ditionalprobability1.Thecounterexamples dependcruciallyuponatleast
one premissbeing, howeverslightly,less thancertain.Whereuncertain
premissesare not at issue-in mathematics,say-these inferenceforms
won't let us down. (Analogy:if all As areB andall Bs are C, thenall As
areC;butwe canhave,allAs areB, almostall Bs areC, yet allAs are -,C.)
We could call an argument"valid"if it satisfiedthis Certaintycriterion,
andthusreinstatethe argumentson the right.But, in arguingaboutcon-
tingentmatters,100%certaintyfor our premissesis rare;moreover,it is
491Proportions isomorphism.If you imagine
providejust a model- a structural
the space of possibilitiesdividedup into enoughequallyprobablelittle bits, or
"worlds",you can translatethe modelto "almostall worldsareA-worlds",etc.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals287

hardto distinguishfromits nearneighbours.Knowingthatan argumentis

"valid"in this sense would be of little use: it would guaranteenothing
aboutwhatwe shouldthinkaboutthe conclusionwhenourpremissesare
only a hair'sbreadthawayfromcertainty.
7.3. Speech Acts. "Whatam I doing, when I say, or think,thatif A, B, if
not saying(thinking)thatsomethingis the case?"As faras thinkinggoes,
the idea of believingthatB underthe suppositionthatA, of havinga con-
ditionalbelief, of believing somethinggiven a hypothesis-Ramsey's
idea-is, I hope, clearenough.HereI focus-on saying.
Someoneasksme who will win theBoatRace.I say "Oxfordwill win".
I expressa belief. But I speakaboutthe world.If I say "If the wateris
calm,Oxfordwill win",I expressa conditionalbelief;butit is implausible
thatthatis all I do: I also speakaboutthe world-about the BoatRace-
albeitconditionally.The answerwhichfits the Thesisbest is this:I make
a conditionalassertion.My high degreeof belief thatOxfordwill win if
the wateris calm,amountsto thinking(thewateris calmandOxfordwill
win) is muchmorelikely than(the wateris calm andOxfordwon't win).
I takemyself to be in a position(ceterisparibus)to assertthatOxfordwill
win, not categorically,butconditionallyuponthe waterbeing calm.50
Any kindof speechact canbe performedunconditionallyor condition-
ally. Thereare conditionalquestions,commands,promises,agreements,
offers, etc., as well as conditionalassertions.Any kind of propositional
attitudecanoccurwithinthe scopeof a supposition.Thereareconditional
beliefs, desires,hopes,fears,etc. "Ifhe phones,whatshallI say?";"Ifhe
phones,hangup immediately"; "Iwantto speakto him if he phones";"If
50Ceteris paribus. Some people interpretthe Thesis as an account of when a
conditional sentence is assertable: a conditional is assertable to the extent that b(B
given A) is high. Adams did in his early writings, but not in his book (1975) or
subsequently. Appiah (1985) does, as do Lewis and Jackson (see ?9.1). I do not.
I interpretit as an account of belief that B if A, to various degrees. Firstly, whether
a sentence is assertable depends on all sorts of Gricean contextual factors, which
have to be put aside for an account in terms of assertability. Secondly, both for an
unconditional claim (Oxford will win) and a conditional one (they'll win if the
water is calm), how high b(B) or b(B IA)has to be for an assertion unqualified by
"probably", or "I think", is a context-dependent matter. Thirdly, this may be a
question not just of how close to certain one is, but of the nature and prominence
of this uncertainty.Dudman (1992) says we don't assert "I won't win the lottery"
or "if I buy a ticket I won't win", even if our chance of winning is one in fifty mil-
lion. Lowe (1995) says likewise. I'm not sure whether they are right, but if they
are this has no bearing on the Thesis as I understand it, the claim that someone
who knows the chance has a high degree of belief that she won't win/won't win
if she buys a ticket.
If no uncertaintyis compatible with unqualified assertion we should assert very
little. We do assert, unqualified,-many things, conditionally or otherwise, where
there is more than a chance of one in fifty million that our expectations will be
thwarted.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
288 DorothyEdgington

he phones,I hopeyou won'tbe rude".It is overwhelminglyplausiblethat

the clause, "if he phones",does the samejob in conditionalstatements,
commands,questions,promises,expressionsof wish, etc.; andhencethat
a theoryof conditionalsshouldbe applicableto more thanconditional
statements.
This is quite a severe test. Try applyingStalnaker'stheoryto condi-
tional commands.Interpret"If it rains, take your umbrella"as "In the
closest possible worldin which it rains,take your umbrella".Now sup-
pose I have forgottenyourcommandor alternativelyam inclinedto dis-
regardit. However,it doesn'train.In the closest worldsin which it does
rain,though,I don'ttakemy umbrella.So, on Stalnaker'sanalysis,I have
disobeyed you. Similarlyfor conditionalpromises:on this analysis, I
couldbreakmy promiseto go to the doctorif the paingets worse,even if
the pain gets better.This is wrong:conditionalcommandsand promises
arenot requirementson my behaviourin otherpossibleworlds.
We have alreadyseen thatconditionalbelief is not belief in the truth-
functionalconditional.Nor are conditionalcommandsor expressionsof
desire,commandsetc. thatthetruth-functional conditionalbe true."Ifyou
writethe article,submitit to Mind."Now "Eitheryou won'twritethe arti-
cle, or submitit to Mind"is a non-starter,not even grammatical.5'1 Con-
struedas a commandto make the truth-functionalconditionaltrue, it
amountsto the command"Eitherdon't write the article,or submitit to
Mind".But I am not urgingthat:you couldeasily makethattruein ways
whichwould pleaseme least of all. Turnthe exampleinto an expression
of desire:"If you writethe article(W), I want you to submitit to Mind
(S)". The conditional desire amounts to a preference for W&Sover
W&-S. It does not amountto a preferencefor -,WvSover W&-S.52 For
althoughW&-S is less desirablethanW&S, it may be very muchmore
desirablethan-_W,which,alas,is a farfromimplausibleway that -,WvS
couldbe true.
The claim that we use if-sentencesto make conditionalassertionsis
madeby von Wright(1957, p. 131), and is mentionedby Quine(1952).
Quinesays:
Now underwhat circumstancesis a conditionaltrue?Even to
raisethis questionis to departfromeverydayattitudes.An affir-
mationof the form"ifp thenq"is commonlyfelt less as an affir-
mationof a conditionalthanas a conditionalaffirmationof the
consequent.If, afterwe havemadesuchan affirmation,the ante-
cedentturnsout true,then we considerourselvescommittedto
S lThe grammatical factthatconditionalshavea mainanda subordinate clause
fits the view thatthey are used to do whateverthe mainclause does, but condi-
tionally.
52 Thatis, it does not entailthis preference;
althoughit is compatiblewith it.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 289

the consequent,and are readyto acknowledgeerrorif it proves

false. If on the otherhandthe antecedentturnsout to have been
false, ourconditionalaffirmationis as if it hadneverbeen made
(p. 19).
As it stands,this last sentenceis absurd.It is not absurdif we delete the
word"conditional" fromit. It is not absurdto hold thatI do not countas
havingmadean assertionunlessthe antecedentis true.But it is absurdto
say it is as if I hadnotmadea conditionalassertion-as if I hadsaidnoth-
ing at all. I say to you "If you pressthatswitch, therewill be an explo-
sion".As a consequence,you don't pressit. HadI said nothingat all, let
us supposeyou wouldhavepressedit. A disasteris avoided,as a resultof
this piece of linguisticcommunication.It is not as if nothinghad been
said. This is no objectionto the idea thatI did not (categorically)assert
anything.Forlet us supposethatI am understoodas havingmadea con-
ditionalassertionof the consequent.My hearerunderstandsthat if she
pressesit, my assertignof theconsequenthascategoricalforce;and,given
thatshe takesme to be trustworthyandreliable,if it does acquirecategor-
ical force,it is muchmorelikely to be truethanfalse. So she too acquires
reasonto thinkthattherewill be an explosionif she pressesit, andhence
a reasonnot to pressit.
Dummett,likeQuine,misrepresents thenotionof a conditionalassertion
whenhe saysit is "asif [someone]hadhandedhis hearersa sealedenvelope
marked 'Open only in the event that ... ."' (1992, p. 115). If it were like
that,modustollenswouldbe impossible,as Dummettpointsout.Whereas,
on thelinesof theexampleabove,we canexplainwhy someoneinfersthat
-A whenhe knowsthatB is false anda trustworthy personhasjustasserted
B conditionallyuponA. Elsewhere(1973, p. 341ff), Dummettis sensitive
to the differencebetween"no(categorical)assertionhas been made"and
"nothinghasbeensaid".It'snotjustthatthesealed-envelopeinterpretation
cannotbe trueof ouruse of conditionals.Thereis somethingintrinsically
absurdin theideathatunderstanding a sentenceshouldrequireyou, in cer-
taincircumstances,to behaveas thoughit hadnot been said."Wecannot
lay down a conventionthatno one is to be influenced"(op. cit. p. 342). I
undertaketo carefor yourchildrenif you die. Even if you don't die, my
conditionalundertaking has consequencesfor you andfor me.
It is the analogueof the mistaken,sealed-envelopeinterpretation for
conditional commandsthat leads Dummettto say that a conditional
imperativewherethe antecedentis in the agent'spower"must... be inter-
pretedas a commandto makethematerialconditionaltrue"(1973, p. 340;
see also 1959,pp. 8-9.). A childis told "Ifyou go out, wearyourcoat".If
he cannotfindhis coat,he staysin, in orderto complywiththe command.
On my interpretation,if the child can't find his coat, he has a choice
betweendisobeyingthe command,and behavingin such a way thatno

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
290 DorothyEdgington

categoricalcommandhas been made (not: behavingas thoughnothing

had been said). If he wishes not to disobey,he must stay in. Dummett
claims thatthereis no distinctionbetweennot disobeyinga conditional
command,andobeyingit. But otherexamplesmakethis implausible.If,
in theemergencyward,you'retold"Ifthepatientis still alivein themorn-
ing, changethe drip",andyou smotherthe patient,you can hardlyclaim
to have merelycarriedout an order.
A conditionalassertion"IfA, B" is anassertionof B whenA is true,and
an assertionof nothingwhenA is false. It is naturalthen,to say my con-
ditionalassertionis trueif A andB arebothtrue,andfalse if A is trueand
B is not, and has no truthvalue whenA is false. This is compatiblewith
the Thesis, providedwe interpretthis assignmentof truthvalues with
care.Belief thatif A, B is notbelief thatit is true.Forit is trueonly if A&B,
and we may believe that if A, B withoutbelieving thatA&B. Nor is it
belief thatit is not false. For it is not false providedthat -1(A&-B),i.e.
ADB;andwe can believe thatit is not false withoutbelievingthatif A, B.
Belief thatif A, B is a conditionalbelief thatit is truegiven thatit has a
truthvalue-belief thatit is truegiven thatit is eithertrueor false. Now
my degreeof belief that"IfA, B" is true,given thatit has a truthvalue,is
just b(A&B)Ib(A), as it should be. (The bombshellis avoided because
belief thatif A, B is not belief thatsomethingis true.)For a proposition
assumedto be eithertrueor false, yourdegreeof beliefthatit is truegiven
thatit has a truthvalue,is yourdegreeof belief thatit is true.So ourpro-
posalis not ad hoc. It has as a specialcase thatfor a bivalentproposition,
to believe it is to believe thatit is true.
In makingconditionalassertions,we do not aim at truth(for we don't
assertthem only if we believe A&B);nor do we aim at avoidingfalsity
(forwe don'tassertthemwheneverwe believe-{A&-B)); ouraimis that
they be truegiven thatthey have a truthvalue-that if it turnsout thatA,
we get B as well, ratherthan-B.
Dummett(1959, pp. 10-4) rightlysays that giving a truthtable for a
statementwith threevalues,T, F, and X, gives you little guidanceas to
how the statementis to be used.Does the speakerintendto ruleout X (as
in the case of emptynames),or not? On the conditional,Dummettsays
"[thespeaker]is not takenas havingmisusedthe statementor misledhis
hearersif he envisagesit as a possibilitythatthatcase will arisein which
he is said not to have madea statementtrueor false"(p. 11).Thatis cor-
rect, but it does not follow thatfor conditionalsX is really a species of
truth,as Dummettclaims(p. 12). Thereis a differencebetweenthe claim
thatI may use this statementcorrectlywhen the antecedentis false, and
the claimthatwheneverthe antecedentis false, I have usedthe statement
correctly.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals291

The "true,false, neither"classificationdoes not yield an interesting3-

valuedlogic or a promisingtreatmentof compoundsof conditionals(for
it is not a case of there being some "designated"value or values). It
helps only in minor ways. It allows us to say that a conditional is
straightforwardly false if its antecedentis trueandthe consequentfalse;
andthatit is straightforwardly trueif the antecedentandconsequentare
both true.Thereis nothingcomparablystraightforward to say when the
antecedentis false.
An unconditionalassertion,e.g. thatJohnis in London,canbe rightby
luck, or wrongby bad luck:my reasonscan be good yet I'm wrong,my
reasonscan be bad yet I'm right.Or I can have no reasons,yet guess, or
havea hunchthatJohnis in London.Likewisefor a conditionalassertion,
thatJohnis in Londonif Maryis. Mackie(1973,p. 107)hasa fathersaying
to a child "If you putyourfingerthroughthe bar [of the monkey'scage],
it will be bittenoff'. The child does so nevertheless.In one scenario,the
monkeypaysno attention,buta birdsweepsdownandbitesoff the finger.
The conditionalassertionwas true,for an unexpectedreason.In another
scenario,the monkeyis aboutto bite whena rockfalls, squashesthe cage
andkills the monkey.The assertionwas false, for an unexpectedreason.
If it is plausiblethata conditionalassertion,like anunconditionalone, can
be rightor wrongby luck,thisis an argumentagainstthosewho insistthat
the antecedentmustbe "relevant"to the consequent(see pp.267-9). And
it addsplausibilityto thefeatureof theThesisto whichsuchpeopleobject:
belief thatA&B is sufficientfor belief thatB if A. If, as you believe,A&B
is true,so is yourassertionof B on the conditionthatA.53

8. Objectivityand its limits

8.1. The truthvaluespermittedby the notionof conditionalassertionare

little morethanepiphenomenal-theydon't significantlychangethe pic-
turewe hadwithoutthem.Weneedmoreobjectivityfor conditionalsthan
they provide,it will be complained:we need an accountof how a condi-
tionalcan be right,or wrong,even if its antecedentis false. "Ifthatlump
of sugaris placedin water,it will dissolve"is true, it will be said;"Ifthat
lumpof graniteis placedin water,it will dissolve"is false, even if neither
is placedin water.Yetmoredecisive,"Ifit's square,it has4 sides"and"If
I3I became more aware of the relevance of other conditional speech acts on
reading Michael Firestone's thesis, "The Meaning of 'If"' (Australian National
University). Michael Woods' manuscriptalso treats "simple conditionals" as con-
ditional assertions; he too has a careful examination of the different conditional
speech acts.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
292 DorothyEdgington

it's square,it has 5 sides"are(surely!)trueandfalse respectively,even if

the objectin questionis not square.Orconsidermy conditionalassertion
of "Ifyou pressthe button,therewill be anexplosion".Youdon'tpressit.
Weholda postmortem,tryingto establishwhethertherewouldhavebeen
an explosionif you hadpressedit (NB). It couldendwith"Yousee, I was
right",or with "Youwerewrong-there wouldhavebeen no explosionif
I hadpressedit".
Fromwhatwe haveseen so far,theThesisneednotruleouttherebeing
an objectively correct thing to think about whetherB if A. The right
degreesof belief thatit has4 sides/5 sides given thatit's squareare 1 and
0 respectively.In the case of the sugarlump and the granite,the right
degreesof belief (in normalcases-putting asidesuper-saturation andthe
like) areat leastveryclose to 1 and0. 90%of theredballsin thisbaghave
blackspots.Youareto shakeit, putyourhandin, andpicka ball.Thereis
a rightdegreeof belief thatthe ball you pick will have a blackspot if it's
red. For the best opinionaboutwhetheryou'll be curedif you have this
operation,ask the best doctoryou can find.The chancethatthe chemical
substancewill emit dangerousradiationif storedunderground,will be
best estimatedby a chemistor physicist.Youreadin the newspaperthat
if you eat garlic,you are less likely to get heartdisease.You watchthe
weatherforecast.And so on. We all have the idea of a right,or at least a
better,opinion.Thatis, we have the idea of objectiveprobability-or at
leastthe idea thatsome degreesof belief areworthmorethanothers.An
expertis someonewho hasacquiredgoodjudgementin a given area-and
moreover,has access to morerelevantinformationthanthe restof us, in
thatarea.An expertis someonewhoseadvicewe do well to heed,in form-
ing ourown beliefs andplansfor action.54
Primafacie, thereis roomfor an accountof objectivelycorrectcondi-
tionalthoughts.It doesn'tfollow thatthey have truthconditions.The fol-
lowing has been suggested55
"IfA, B" is trueiff the objectiveprobabilityof B givenA is suffi-
cientlyhigh.
Thisis notcompatiblewiththeThesis,andis independently objectionable.
(I do not objectto the fact thatthe truthconditionis vague.)Presumably,
in a context,eitherthereis some numberless than 1 whichis sufficiently
high;or thereis somenumbergreaterthan0 whichis notsufficientlyhigh;
or (mostlikely) both.Takean examplewhereobjectiveprobabilitiesare
relativelyeasy to estimate-balls in bags, say. Call the proposedtruth
54 Adamsis no subjectivistaboutprobability.
A sectionof his book (1975) is
entitled"Amotivefor wantingto arriveat correctprobabilityestimates".
55 See SimonBlackburn (1986, pp.213-5); MichaelWoods(manuscript); and
the suggestioncropsup orallyfromtime to time.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 293

conditionS. Firstsuppose0.9, say, is sufficientlyhigh, and I am certain

thatthe objectiveprobabilityof B givenA is 0.9. My degreeof belief in B
given A is 0.9. Accordingto the Thesis, I am 90%confidentthatif A, B.
But I am certainthatS, hence, certainthatthe conditionalis true.By the
truthcondition,I am 100%confidentthatif A, B. (Thetruthconditionhas
the additionalembarrassing consequencethatthetruthof "IfA, B"is com-
patiblewith the truthof A&-B.) Second,supposesome numbergreater
than0 is not sufficientlyhigh-0.5 say. SupposethatI am certainthatthe
objectiveprobabilityof B givenA is 0.5, andso have degreeof belief 0.5
thatB givenA. By the Thesis,I am 50%confidentthatif A, B. Now I am
certainthatS is false,hencecertainthattheconditionalis false.By thetruth
condition,I am0%confidentthatif A, B. (Thetruthconditionalso hasthe
consequencethatthe truthof A&B is compatiblewith the certainfalsity
of "IfA, B".Not everyonemindsthat.I thinkit's wrongfor me to say "It
is certainlyfalse that,if you approach,thedog will bite",whenI knowthat
theobjectiveconditionalprobabilityof its biting,giventhatyou approach,
is 0.5; andfurther,to admitno errorwhen you approachandarebitten-
to stickto my judgementthatthe conditionalwas certainlyfalse. But not
everyoneagreeswith me.56)
If we are to have objective values, we need values intermediate
betweentruthandfalsity.But thereis an obstacleto objectivevalues-to
therebeing a rightthingto think-to whichI now turn.
8.2. Gibbard(1981, pp. 231-2) presentedan argumentfor the Thesis and
against truthfor indicative conditionals, which threatensto wipe out
objectivityalongwithtruth.Thisis its structureas I see it. (1) If two state-
mentsarecompatible,so thattheycanbothbe true,a personmayconsist-
ently believe bothof themsimultaneously.(2) For consistentA, and any
B, people do not simultaneouslybelieve both "If A, B" and "If A, -B"
(unlessby oversight),norconsiderit permissibleto do so; rather,to accept
"IfA, B" is to reject"IfA, -B". (This accordswith the Thesis:if b(BIA)
is high, b(-B IA)is low. It also accordswith Stalnaker'struthconditions
but not the truth-functional account.)So, by (1) "IfA, B" and"IfA, -B"
can't both be true:if they could, why shouldn'tsomeonereadilyaccept
both?But (3) one personX can haveimpeccablereasonsfor believing"If
A, B",while anotherpersonYhasimpeccablereasonsfor believing"IfA,
-B"; (a) the situationis symmetric:thereis no reasonto preferX's belief
to Y's,or vice versa;no case canbe madefor sayingjust one of thebeliefs
is false;(b) neitherof themis makinganysortof mistake;eachis rational,
andbaseshisjudgementon knowntruths;no case canbe madefor saying
bothbeliefs arefalse. So: they can'tbothbe true,theycan'tbothbe false,
56 Pendlebury(1989), Read(1995) andothersarguethatthe truthof A&Bis
compatiblewiththe falsityof "IfA, B".

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
294 DorothyEdgington

andit can'tbe thatjust one of themis true.Truthandfalsity arenot suit-

abletermsof assessmentfor conditionals.
Gibbard'smuch-discussedexampleis the Sly Pete Story.It concernsa
pokergame.Jacksaw thatPete hadthe losing hand,andbelieves"IfPete
called,he lost".ZackknowsthatPete,the cheat,knewthe contentsof his
opponent'shand,andthatPete alwaysplays to win. He believes "IfPete
called,he won".In fact, Pete didn'tcall. Oncethey learnthis, neitherhas
any use for a thoughtbeginning"If Pete called".
Gibbard'sexampleis perhapsnot perfectlysymmetric,andsome have
arguedthatJack'sbelief is betterthanZack's.Pendlebury(1989, p. 182)
claimsthatfromthe God'seye pointof view,Jack'sconditionalis the true
one. Thispoints,at best,to an imperfectionof the example.Hereis a bor-
ing, perfectlysymmetricone. In a game,(1) all redsquarecardsareworth
10 points, and (2) all large squarecards are worthnothing.X caughta
glimpse as Z picked a card and saw that it was red. Knowing (1), he
believes"IfZ pickeda squarecard,it's worth10 points".YXseeingit bulg-
ing underZ's jacket, whereZ is keepingit out of view, knows it's large.
Knowing(2), he believes "IfZ pickeda squarecard,it's worthnothing".
(Someonewho knowsall the relevantfactsknowsit isn't square,andhas
no use for a conditionalbeginning"Ifit's square".)
Thereis little hope for objectivelycorrectopinion,if one personcan
have a completelyadequatereasonto accept"If A, B", andreject"IfA,
-B"; while anotherhas a completelyadequatereasonto do preciselythe
opposite.It is not as though,if eitherhad more information,he would
knowwhatto think.If he hadmorerelevantinformation,he wouldknow
that-A, andhave no use for eitherconditional,"eachof whichis a ticket
for anintellectualjourneystartingat a placewherehe knowshe will never
be" (Bennett,1988, p. 520).
How widespreadis theGibbardphenomenon?Gibbard(1981, pp.226-
9) thinkstherearetwo kindsof conditionals:"epistemic"ones, whichsat-
isfy the Thesis,andare subjectto this phenomenon,paradigmsof which
are past-tenseindicatives;and"nearnessconditionals",to be treated'ala
Stalnaker,whicharenot subjectto the phenomenon,paradigmsof which
are subjunctiveconditionals.Future-tenseindicatives can function as
either(p. 228). I don'tthinkconditionalsdividein this way.Forany con-
tingentconditional,the worldmay be suchthatthe Gibbardphenomenon
can arise.Hereis anotherexample.
Supposetherearetwovaccinesagainstacertaindisease,AandB. Neither
is completelyeffectiveagainstthe disease.Everyonewho hasA andgets
the disease,gets a side effect S. Everyonewho has B andgets the disease,
doesn't get S. Having both vaccines is, however,completely effective
againstthe disease(thoughnot manypeoplehave both).These scientific

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 295

factsareknown.X knowsthatJoneshas hadA, andsays "IfJonesgets the

disease,he'll get S ". Y knows thatJoneshas hadB, and says "If he gets
the disease,he won't get S".(If Jonesis meanwhilerunoverby a bus and
killed, these can go counterfactual:
"If he had got the disease,he would
havegot S","Ifhe hadgot the disease,he wouldn'thavegot S".Butcoun-
terfactualsareto be putasideuntil?10.) In all thesecases, if the full story
is known,the conditionalsbecomeuseless.Forinstance,the doctorgiving
JonesB says "I can't guaranteethatyou won't get the disease,but if you
do, you won'tget S". "ButI've alreadyhadA", saysJones."Ohwell, then,
youwon'tgetthedisease(thequestionof whatwill happenif youdodoesn't
arise)".57
Now let us changethe initial story slightly.Everyonewho has just A
andgets the disease,gets S. Everyonewho hasjust B andgets the disease,
doesn't get S. Few people have both. Of those who do, very few-say
0.01%- get the disease(whereasabout1%of those who havejust A, or
just B, get the disea4e).Anyonewho has both vaccinesandgets the dis-
ease has a 50%chanceof gettingS. These areknownfacts.As before,X
knows that Jones has had A. He has groundsfor near-certaintythat if
Jonesgets the disease,he will get S. Y,who knowsthatJoneshas hadB,
has groundsfor near-certainty thatif he gets the disease,he won't get S.
This time, neitherhas the rightopinion.Thereis furtherobtainableinfor-
mationwhichwouldleadeachto a betteropinion:anyonewho knowsthe
relevantscientificfacts, andthatJoneshas hadbothvaccines,thinksthat
the chancethatJoneswill get S if he gets the diseaseis 0.5. Doctor,having
givenJonesB: "Ican'tguaranteethatyou won'tget the disease,butif you
do, you won't get S". Jones:"ButI've alreadyhadA". Doctor:"Oh,then
I mustcorrectwhatI said:it's very unlikelythatyou will get the disease;
andit's 50%likely thatyou will get S if you do get the disease".
TheGibbardphenomenonarisesif andonly if therearecurrentlyascer-
tainablefacts which rule out A.58In one direction:let F be a set of cur-
rentlyascertainable factsfromwhichwe canderivethat-A. Thatis, A&F
is inconsistent.Thatis, A&F entailsa contradiction.Furthermore, it will
generallybe the case thatif we knowF, we canlearnthat-A by assuming
A andderiving,fromA&F,a contradiction,B&-B. Now, F is true,there-
fore consistent.It is the additionof A thatenablesus to deriveB&-B. So
I7I have given elsewhere a more careful example, which relies just on the gas
laws (Edgington 1991, pp. 206-7). The examples show that conditionals which
are based on "objective connections in the world" are not immune from the Gib-
bard phenomenon.
58 We must include general truths among the ascertainable facts; in the exam-
ples we have considered, some were founded on the rules of a game; and there
was "Pete always plays to win"; and the scientific general truths about the vac-
cine.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
296 Dorothy Edgington

theremustbe two subsetsof F (not necessarilydisjoint),F, andF2,such

thatfromA&F, we can deriveB, and fromA&F2we can derive-B. So
someonewho knowsjust F, has adequatereasonto believe "IfA, B"; and
someonewho knowsjust F2has adequatereasonto believe "IfA, -B".
In the otherdirection:supposethereis no set of currentlyascertainable
factswhichruleoutA: all currentlyascertainablefactsareconsistentwith
A. Therefore,therearenottwo subsetsF, andF2suchthatA&F1entailsB
andA&F2entails -B. The Gibbardphenomenondoes not arise. There
may be two subsets which renderit probable,and improbable,respec-
tively,thatB given A. But this is no threatto objectivity.Peoplein these
statesof informationcan improvetheiropinionsby learningmore.There
is no obstacleto therebeing an objectivelyrightthingto think,basedon
all the relevantcurrentlyavailableinformation.
If currentlyascertainablefacts are sufficientto ruleoutA, thenA has a
currentobjectiveprobabilityof 0. Hencethe presentobjectiveprobability
of B givenA is undefined.Hencethereis no ideal,objectivethingto think.
(Theremayof coursebe moreor less rationalways of assessingthe infor-
mationyou have. But the world does not make one judgementbest.) A
necessaryconditionfor therebeing an optimumjudgement,then, is that
the antecedenthas a presentnon-zerochanceof being true.59
8.3. One way of approachingthe idea of correctjudgementis via the
device of an ideal epistemicperspective.Thereare differentdegreesto
which we can idealise.Takethe extreme:G, knows everything-all the
facts,past,presentandfuture.Considerthe conditional"Ifit rainstonight
(R), the riverwill overflowits bankstomorrow(Q)".60 GIknowsR&O, or
knows R&-1O, or knows --R. He knows too much to have any use for
indicativeconditionals.He can pronouncethemtriviallytrueor false in
the first two cases. If the thirdcase obtains,the questiondoesn't arise.
(How G, fares with the questionwhetherit would have overflowedif it
hadrained,I leave until?10.)
G, is a littletoo idealfor us to relateto. Beforeturningto G2,let us look
at some featuresof the conceptof objectivechance.
The concept of objective chance gets some purchasewhen, at least
apparently,like causes do not have like effects;andmoreover,in a class
of apparentlyrelevantlyidenticalcases, the proportionsof the various
sortsof outcomearerelativelystable,althoughtheseproportionsaregen-
eratedin an apparentlyrandomway.Thesefacts aboutproportionscould
59For presentpurposes,nothingwhich is now causallypossible has a zero
chanceof beingtrue.Wecan give infinitesimalprobabilitiesto hittinga particular
pointon a dart-board,etc. See Lewis (1980, p. 89), McGee(1994).
60 This exampleis Bennett's
(1988, p. 521), wherehe also uses the device of
an idealepistemicperspective.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 297

be brute(as those who interpretobjectivechance as relativefrequency

think);or they could be explainedas just the sortof patternof outcomes
one wouldexpectif (e.g.) eachF has ap% chanceof resultingin a G. The
strongestnotionof objectivechance appliesonly if we remove "appar-
ently"from the above: relevantlyqualitativelyidenticalstates do have
differentoutcomes.Shortof that,whichdifferentoutcomeoccursmaybe
arbitrarilysensitiveto the exact valuesof the variableswhichspecify the
cause:theremay be no degreeof similarityshortof qualitativeidentity
suchthat,if two situationsarethatsimilar,the outcomewill be similar.6'
As we cannotmeasureto an infinitedegreeof accuracy,we needto apply
theconceptof chance.Andtherewill be manymorecases wheredetection
of relevantdifferenceis not practicable,andapplicationof the conceptof
chance is useful. We could call these three cases "absoluteindetermin-
ism","quasi-indeterminism" and"practicalindeterminism"; therewill be
correlativenotionsof determinism,dependingon which they ruleout. It
will not matterwhichway you interpret"indeterminism" and"determin-
ism"in whatfollows. They representdifferentdegreesof idealisationof
epistemicperspectives.
The pointof the conceptof objectivechanceis thatknowledgeof the
chancesof futureeventsgives us reasonsfor ourexpectations.Thereare
laws aboutchances,whichwe tryto ascertain.
The idea of backward-looking objectivechanceis as rechercheas the
ideaof backwardcausation.62 A pasteventmayhavehad a realchanceof
notcomingabout.Butthetimefortherealisationof thatchancehaspassed.
The event happened,andhas no presentchanceof not havinghappened.
(Theremaybe anexception:the unobservedwave packetof quantumthe-
ory whichdidn'tcollapse.Ourgreatdifficultyin understanding this phe-
nomenonunderlinesmy point abouthow we normallythink.)Chances
changewithtime,accordingto theoutcomesof interveningchanceevents.
The probabilityof my winningmy bet on 3 headsis '/8. Afterone toss, it
haschangedto either1/4 or 0, forthe probabilityof '/2thatthe firsttoss had
of landingheadshasnow"collapsed"to 1 or0. A mangoes througha maze
at constantspeed,decidinghis pathby a randomdevice.At anypoint,we
cancalculatethechancethathe will be atthecentreby noon.Thiscanvary
as he makesunluckyor luckyturns,untilnoon,whenit becomes 1 or 0.63
Ourreasoningaboutthe past respectsthis asymmetryin the direction
of chance.Wetryto findthe best explanationof whatwe currentlyknow,
andexaminehypothesesaboutpastforward-looking chances.The chance
6! This gives rise to Chaos Theory. See Gleick (1987); Smith (1991).
62
I leave open the question whether it is incoherent, or merely very weird.
63 This is Lewis's delightful example of the "garden of forking paths" (1980,
p. 91). My discussion of chance owes much to this article.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
298 Dorothy Edgington

was highthathe wouldget thesesymptoms,if he tookarsenic;the chance

was low thathe wouldget these symptomsif he didn't.
Now let us returnto the ideal epistemicperspective.G2does not have
magicalknowledgeof the future(or anythingelse). For a given condi-
tional,e.g. "Ifit rainstonight,theriverwill overflowits bankstomorrow",
he knows all he needs to know about the past and the laws of nature.
(Theseare thingswe aspireto.) Now distinguishtwo cases: if determin-
ism is true,G2is in as good a positionas G,. Fromhis knowledgeof the
pastandthelaws, he caninferthefuture.Again,he knowsR&O, or knows
R&-1O, or knows --R. He has no non-trivialuse for indicativecondition-
als. If determinismis false, however,G2canknowthatR has somechance
of occurring,andmoreover,can knowthe presentchanceof 0 given R-
it may be 1 or 0 or somethingin-between.
But supposeG2knowsthatthe chanceof rainis now 0. Thenp(OlR) is
undefined.Thereis no objectivefact abouthow likely it is, now, thatthe
riverwill overflowgiven thatit rains,whenit is now causallyimpossible
that it rain. Comparethe doctor and the vaccination, above (p. 295).
(Theremay be an objectivefact abouthow likely it is thatit wouldhave
overflowedif it hadrained.Moreon thatlater.)Objectivitybreaksdown
whenthe presentchanceof the antecedent'sbeing trueis 0.
Whenthereis a rightthing to think,it is only temporarilyright.The
chanceof a futurepossibilitychangeswith time;andin due course,flips
to 1 or 0, accordingto what actuallyhappens.Consider"If it rainedon
Mondaynight, the river overflowedits banks on Tuesdayafternoon".
Evenif I knowthatthe chanceof 0 givenR was 1, someoneelse canhave
just as good reasonto say "If it rainedon Mondaynight,the riverdidn't
overflowon Tuesdayafternoon" -either because she saw thatthe river
didn't overflow,or perhapssaw from the state of the river on Tuesday
morning,or on Wednesday,thatit couldn'toverflow(orhaveoverflowed)
on Tuesdayafternoon.Forbackward-looking indicatives,thereis the bet-
teror worsemanagementof uncertainty, but no idealview.
Objectivelycorrectpresentvaluesremain,then,forfuture-lookingcon-
ditionalswhose antecedenthas a presentnon-zerochanceof being true
(thereare no currentlyascertainablefacts to ruleit out;it is still causally
possible).Buta difficultyarisesfortheprojectof providingstronger-than-
truth-functionaltruthconditions for such conditionals.People hanker
after somethingalong Goodman'sor Lewis's lines, some sort of strict
conditional:thetruthof "IfA, B" requiresthatthetruthof B be guaranteed
by the truthof A, togetherwith otherfacts,given the laws -at least if the
antecedentis false. (Thisqualificationis unnecessaryfor those who deny
thatA&B is, sufficientfor "If A, B" -Pendlebury (1989), Lowe (1995)
andRead(1995), for instance.Others,like Lewis, acceptthatthe condi-

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 299

tionalis trueif it turnsout thatA&B, even if it is causallypossiblethat

A&-B. (Lewis is, of course,only giving a theoryof counterfactuals; but
he appliesthe theorywidely, for exampleto the forward-lookingcondi-
tionalsneededfor decisiontheory(1981, pp. 325-35).)
A fairdose of determinism,then,is requiredfor truth-to ensurea con-
nectionbetweenA andB, andto ensurethatall the initialconditionsare
in place. But total determinismwill meanthat,if the antecedentis false,
it is now causallyimpossible,andthereis no idealthingto think.I do not
say that happy combinations of determinismand indeterminismare
impossible.But it is hardto see what would groundour confidencethat
many conditionalsare true. (No generalpredilectiontowardsdetermin-
ism, or indeterminism,would.)
If so much is requiredfor truth,so little is requiredfor falsity,on the
above views. Consider"If you toss the coin ten times, it will landheads
at least once".Add thatit is improbablethatyou will toss it. The condi-
tional is either certainlytfalse (on the strongeralternative)or probably
false (on the weakeralternative-it is trueonly if you do toss it, and it
landsheads at least once, but it is improbablethatyou will toss it). Our
everydayconditionalsrungreatrisksof beingcertainlyfalse, or muchtoo
probablyfalse. It may be repliedthat charityrequiresthat we interpret
them with a silent "it will be very probablethat"insertedin the conse-
quent:"Ifyou toss the coin ten times,therewill be a veryhighchancethat
it will landheadsat least once".64Wellthen,truth,as opposedto falsity,
for the plainconditionalswe utter,isn't whatmattersaboutthem.We are
happyenoughwith falsity,providedthatthe chanceof consequentgiven
antecedentis sufficientlyhigh. It is rathermore straightforward to con-
struepeopleas doingtheirbest, on thebasisof as muchrelevantinforma-
tion as it is worth their while to acquire,to estimate as accuratelyas
possiblehow likely it is thatB givenA.
The conceptof objectivechanceis philosophicallypuzzling.But ordi-
nary people do understandit, when they read that (e.g.) eating garlic
reducesone's chanceof heartdisease.Those who wouldexplainit away
need a surrogatefor it. I hope thatI have not reliedon anythingtoo theo-
reticallycontentiousin the discussionabove.

64
The conditionals Lewis uses in the analysis of causation and decision, when
the assumption of determinism is dropped, have chances in the consequents
(Lewis 1986, pp. 175-84; 1981, pp. 329-35).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
300 DorothyEdgington

9. Is truthpossible after all?

9.1. Accordingto Jackson(1979, 1981, 1987) we can explain why the

Thesis gives the rightassertabilityconditionfor indicativeconditionals,
while maintainingthat their truthconditionis the truth-functionalone.
The relationbetween "ADB"and "If A, B" is analogousto the relation
between"AandB"and"AbutB",he claims.The latterpairhavethe same
truthconditions;butit is partof the meaningof "but"thatit is usedto sig-
nala contrastbetweenthepropositionsitjoins. Inthecase of "if",it is part
of its meaningthatit signalsnot only thatthe speakerbelieves thatADB,
but thatthis belief is robustwith respectto the antecedent:the speaker
wouldnot abandonthe belief if she wereto learnthatA. So someonewho
asserts"IfA, B" must not only have a high degreeof belief in ADB,but
mustalso have a high degreeof belief in ADB given A. But b((ADB)lA)
= b((-,AvB)IA) = b(BIA). So we assertconditionalswhenwe havea high
degreeof belief in the consequentgiven the antecedent.
Jacksonclaimsan explanatoryadvantageoverthosewho takethe The-
sis as primitive-the "no-explanation" theorists,he calls them (1987, p.
55). "Tohave assertabilityconditionsbest explainedby certaintruthcon-
ditionsis to have those truthconditions"(1987, p. 58). It is a plausible
methodologicalmaxim that the value of an explanationdependson its
explainingmore thanthe datait accommodates.So let us firstask what
else, if anything,the truthconditionsexplain.
Armedwith truthconditions,can we explainthe occurrenceof condi-
tionalsasconstituentsof longersentences?Wecannot. WeknowhowADB
behavesin compoundsentences.ButJackson'stheoryis notthat"IfA, B"
meansthesameasADB:theirtruthconditionsarethesame,buttheirassert-
ability conditionsdiffer.The theoryhas no implicationsfor how condi-
tionalsbehavein contextsin whichthey areunasserted.In fact he adopts
thestrategyof explainingawaysuchoccurrences(1987,pp. 127-37) along
the lines of ?7.1 above.
Do thetruthconditionsexplainthevalidityof argumentsinvolvingcon-
ditionals?A mainsourceof dissatisfactionwith the truth-functional con-
ditionalis a clashbetweenourintuitionsaboutvalidityandthe arguments
it licensesas valid(see the exampleon p. 281 above).Jacksonclaimsthat
ourintuitionsareat faulthere:we confusepreservationof truthwithpres-
ervationof assertability(1987, pp.50-1). Thesegenerallystayin line,but,
becauseof thespecialruleforassertingconditionals,heretheycomeapart.
Adamsgave an accountof validityin termsof preservationof probability
orconditionalprobability(assertabilityforJackson),whichcoincideswith
an accountin termsof preservationof truthfor argumentswithoutcondi-
tionals. Lewis, agreeingthatour intuitionsaboutvaliditygo betterwith

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 301

Adams,says "As to whether'validity'shouldbe the word for truth-or

assertability-preservation, that seems a non-issueif ever therewas one"
(1986, p. 153). Lewis, presumably,meansnot merelythatwe can use the
word"valid"anyway we like, butthateitherchoicewouldbe reasonable.
Adams'guidingthoughtwas a pragmaticone. Takean argumentwithtwo
or threepremisses.Wouldit be useful to classify it as "valid"when one
canbe arbitrarily close to certainof thepremissesyet rejecttheconclusion
utterly(or in Jackson'sterms,if the premissesare very highly assertable
andthe conclusioncompletelyunassertable)? Adamsthoughtnot.
Jackson'saccountdoes have the advantagethat if A is true and B is
false, "If A, B" is straightforwardly false. A defenderof the Thesis can
pointoutthata high b(BIA) commitsyou to a high b(ADB), andso com-
mitsyou to somethingfalse if A is trueandB is false. Butthisis somewhat
indirect.However,in ?7.3 I arguedthatone could, compatiblywith the
Thesis,interpretasserting"IfA, B"as makinga conditionalassertion,true
if A&B,false if A&-B, truth-value-less otherwise(see abovepp. 390-1).
We have yet to see any explanatoryadvantageof Jackson'stheory.I
turnnow to somedisadvantages.In ?7.3 I claimedthatthenotionof a con-
ditional assertionwas part of a uniformtheory of conditional speech
acts-conditional commands,etc. A theoryof conditionalstatements,I
claimed,should allow that an if-clause, "If he phones",plays the same
role in "If he phones,Marywill be pleased"and "If he phones,hangup
immediately"(p. 287-8). This was a difficultyfor Stalnaker'stheory,and
forthetruth-functional theory,I argued.Letus tryto extendJackson'sthe-
ory to conditional commands: to commandthatif A, B, is notonly to com-
mandthat(ADB),but also to signalthatyou wouldstill command(ADB)
if you believedthatA were true.Returnto the example"If the patientis
alive in the morning,changethe drip"(p. 290). On this analysisI com-
mand"Makeit the case thateitherthe patientis not alive in the morning,
or you changethe drip".You obey my commandif you kill the patient.
Thereis no obviousreasonwhy you shouldconcernyourselfwith whatI
would have commandedhad I believed that the patientwould be still
alive. (Thisis quitedistinctfromthe notionof a conditionalcommand-
an utterancewhich has the force of a commandto makethe consequent
true,on the conditionthatthe antecedentis true.)
I turnto a difficultyfor Jackson'sexplanationof why we asserta con-
ditionalwhenb(B IA) is high, noticedby Lewis. Havinga high degreeof
belief in B given A (call this, following Lewis, robustness,)does not
alwaysmeanthatyou would(or even thinkyou would)be confidentin B
if you learnedA (call thisrobustness2). Consideragain"IfReaganis in the
pay of the KGB,I'll neverfindout"(see above,p. 270). I do have a high
conditionaldegreeof belief in consequentgiven antecedent.I do assert

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
302 DorothyEdgington

theconditional.ButI knowin advancethatI won'tbelievethe consequent

if I learnthe antecedent.Lewis says
Whatreallymattersis robustness2,so it wouldbe moreusefulto
signal that.On the otherhandit would be much easierto signal
robustness1....It may be no easy thingto judge whatwould be
learnedif [A] were learned,in view of the variety of ways in
whichsomethingmightbe learned.Forthemostpart,robustness1
is a reasonable guide to the robustness2that really matters .... So
it is unsurprisingthat what we have the means to signal is the
formerratherthanthe latter.And if this gets conventionalized,it
should be unsurprisingto find that we signal robustness1even
whenthatclearlydivergesfromrobustness2. Thatis exactlywhat
happens ... . I say "If Reagan works for the KGB, I'll never be-
lieve it". (Lewis, 1986 p. 155-6)
So: when we uttera conditional"If A, B", we convey that we believe
ADB;andwe wouldlike to convey,in addition,thatif we wereto learnA,
we wouldstil acceptADB.But thatis a hardthingto be confidentabout,
as Lewis admits.65We settleinsteadfor somethingeasier:ourconditional
degreeof belief in B givenA -not becauseit is intrinsicallyinterestingin
itself, butbecauseit is a good butfallibleguideto whatwe wouldbelieve
if we learnedA. As this conventionis established,even in cases where
b(BIA) is highbut we wouldnot believe B if we learnedA, we assert"If
A, B".
My point aboutconditionalcommands,combinedwith this difficulty,
mightmakeus wonderwhetherit is robustness2thatreallymatters.With
conditionalcommands,I need not concernmyself with whatyou would
commandif you learnedsomethingelse; I needonly concernmyself with
what you do command,albeit conditionally.Similarlywith conditional
assertions,one might wonderwhy I should be interestedin what you
wouldassertif you learnedthatA -after all, as has been pointedout, this
is a difficultthingto be confidentabout.I aminterestedin thefactthatyou
areconfidentthatB on the suppositionthatA-confident enoughto assert
thatB, conditionallyuponthe truthof A. Is Jackson'stheoryof the mean-
ing of "if" plausible?His favouriteanalogyis with the meaningof "but"
(1987, p. 26). Consultingthe dictionaryon "but",I see "in contrast".On
"if",I see "onconditionthat;providedthat;supposingthat";I do not find
anythingto suggestthat"if"conventionallymeansanythingaboutwhatI
wouldassertif I learnedsomething.
65 This,Jacksoninsists(1987, p. 33), andLewis no doubtagrees,is a subjunc-
tive (counterfactual)
conditional.Jackson'stheoryof these is similarto Lewis's.
Andit certainlyis hardto be surethatin all close worldsin whichI learnA, I will
acceptB. Fortheremightbe, for all I know,unexpectedways of learningA-like
learningthatthe matchwas struck,butat the bottomof the swimmingpool.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 303

A thesis like Jackson's-belief thatthe truth-conditions of a statement

S arefulfilledis not sufficientfor it to be asserted-might be testedby try-
ing to find out the conditionsunderwhich people believe that S, even
whenthey wouldnot be preparedto assertit. Takea co-operativesubject
in a contextwhere it is clear that,for purposesof research,we want to
elicit herbeliefs. Firstconsider"but".She believes thatAnn is poor,and
thatAnn is honest.She sees no reasonto contrastthese statesof affairs.
She wouldn'tsay "She'spoorbut she's honest"(exceptin a specialcon-
text where she's huntingfor a poor crook). She is asked whethershe
assentsto that sentence.She might hesitate.She might say "Yes-but I
wouldn'tput it that way";she might say "She's poor, she's honest,but
why the 'but'?".Now she's asked whethershe assentsto "If the Tories
win, they will nationalisethe motorindustry".She thinksthe Torieshave
a very smallchanceof winning.Still, she unhesitatinglysays "No".(She
mightadd"That'scertainlyfalse".)
Jacksoit-is well aware that it is impossible to supporthis theoryby
elicitingevidenceaboutwhenpeoplebelieve conditionals.He advocates
an errortheoryhere-people arewrongabout"if".We speakandthinkas
though there were a conditional connective, "*", such that b(A*B) =
b(BIA). Not manypeopleknowLewis's 1976resultthatthereisn't (1987,
p. 39). (It is not surprising,then, that dictionariesdo not give the true
meaningof "if".)
Now in the sense (if any) that ordinarypeople can be creditedwith
believing thatthereis a conditional"connective",thereis a conditional
connective:take two suitablesentences;make,if necessary,some gram-
maticalchanges,add an "if" in an appropriateplace and you have one
usable sentence.It is philosophers,not ordinarypeople, who have mis-
construedit.
What about the response "That'scertainlyfalse" to the conditional
aboutthe Tories?Doesn't a defenderof the Thesishave to attributeerror
to the speaker?I don'tthinkso. If you say to her"Youmeanit's certainly
false thattheTorieswill nationalisethemotorindustry,on the assumption
thattheywin thenextelection?",I thinkshe wouldacceptthatparaphrase.
In Jackson'sview, if we were free from error-if we stoppedbeing flat-
earthers(1987, p. 40)-we would see thatthatconditionalis very likely
to be truewhenit's verylikely thattheTorieswon'twin. In fact,however,
we arebetteroff in "error".As I said earlier(p. 244), we wouldbe intel-
lectuallydisabledwithoutthe abilityto discriminatebetweenbelievable
andunbelievableconditionalswhose antecedentwe thinkis false.
9.2. Mellordefendsa positionin some ways like Jackson's.It concerns
not assertability(1993 p. 234 fn. 6), but acceptability:to accepta condi-
tional is to be disposedto infer its consequentfrom its antecedent;my

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
304 DorothyEdgington

degreeof acceptanceof "IfA, B" is the degreeof belief I am disposedto

have in B if I fully believe A. This is close enough to the Thesis: any
peripheraldifferencesarenot my presentconcern(see above,pp. 268-9).
To accepta conditionalis not to have a belief, but to have an inferential
disposition.Buthe argues,thisdoes not depriveconditionalsof truthcon-
ditions. WhatLewis showed (1976) is that there is no proposition,no
objectof belief, suchthatyourdegreeof belief in its truthsystematically
matchesyour conditionaldegree of belief in B given A. Once we have
decidedthatto accepta conditionalis not to believe something,Mellor
claims,Lewis's resultbecomesirrelevantto the questionwhethercondi-
tionalshave truthconditions.
Mellorremindsus thatbeliefs are not the only mentalstatesthathave
truthconditions.Desires, fears, and other propositionalattitudeshave
them too. So why shouldn't the dispositional states which constitute
acceptanceof conditionalsalso have them?"All Lewis shows is thatan
Adams 'If P, Q' cannot express a belief in ... [its truth-conditional] con-
tent.So nothingstopsthe [dispositional]theorycreditingall ... condition-
als with ... truthconditions" (Mellor 1993, p. 238).
Mellor thinksthereare two kinds of conditionals,and has been con-
vincedby Dudmanthatthetraditionalline was misplaced(Dudman1988;
and see above, p. 238). Backward-lookingconditionalslike "If Oswald
didn't do it, someone else did", have truth-functionaltruthconditions.
Forward-looking conditionals,like "IfOswalddoesn'tdo it, someoneelse
will", behavelike those traditionallycalled "subjunctive".These,Mellor
suggests,have Stalnaker-Lewis-style truthconditions.
Takethetruth-functional case.Belief thatthetruthconditionis satisfied
is not enoughto acceptthe conditional,for well-knownreasons.But if
you do accept"IfA, B", the truthof ADB is whatensuresthatyou won't
end up with a false belief in B, shouldyou learnA.
Thatis so, but inferringthe consequentfromthe antecedentis not the
only thing we do with conditionalswe accept.They have otherroles in
practicalandtheoreticalreasoning.And acceptingtruth-functionally true
ones could get you into all sorts of trouble.Being in a fragile state of
mind,I accept thatif the Queenwas at home this last hour,she has been
worryingaboutwhereI am. So I hadbettertry andphone.I am liable to
be arrestedfor makingnuisancecalls, thoughthe conditionalI acceptis
true,for she is not at home. Had the WarrenCommissionaccepted"If
Oswalddidn'tkill Kennedy,M15did",Anglo-Americanrelationswould
have sunkto an all-timelow, even if theirconditionalwas true(Oswald
did it). We shouldtrynot to acceptconditionalswhicharetruth-function-
allyfalse, everyoneagrees;butacceptinga conditionalcanbe a prettybad
inferentialdispositionto have,even if its materialcounterpart is true.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 305

The converse problem arises for the conditionals to which Mellor

ascribesstrongtruthconditions:I may be disposedto have a high degree
of belief in B on learningA, yet be fairlysurethat"IfA, B" is not true.To
repeatthe boringold example,I am disposedto a highdegreeof belief in
"thecoin will landheadsat leastonce"shouldI acquirethebelief thatyou
aregoing to toss it ten times. So I acceptthe correspondingconditional.I
alsohappento thinkit's unlikelythatyou will toss it tentimes.OnLewis's
andStalnaker'struthconditions,it is trueonly if you do toss it ten times
andget at least one head- andthis is unlikely,for it is unlikelythatyou
will toss it. On Lewis's truthconditions,it is otherwisefalse-either you
toss it and get no heads, or you don't toss it, but there are some close
worldsin whichyou toss it andget no heads.On Stalnaker's,it is false if
you toss andget no heads,indeterminate if you don'ttoss it. Forboth,the
probabilityof its truthis low. (I use the coin examplebecauseits structure
is transparent;but many everydayexamplesproducethe same result.)
Mellormisses this discrepancybecause,althoughhe allows thatthereare
degreesof acceptanceandof belief, he does not considertheirapplication
in his discussion of possible-worldtruthconditions.The discrepancy
betweenhis acceptanceconditionandthe truthconditiondoes not show
up in the case of full belief. If I'm certainthatall relevantA-worldsareB-
worlds,I will fully believe B if I learnA, andvice versa.The discrepancy
showsup if we replace"all"with "almostall".
It is not clear what role the truthconditions play when they fit the
acceptance condition badly (as, of course, they must, given Lewis's
result).Mellorremindedus thatdesires,hopes,fearsalso have truthcon-
ditions.So they do: to desirethatp is to desirethatit is truethatp, i.e.,
thatp's truthconditionobtains.Thatis to say, these stateshave proposi-
tional content.But a conditionaldoes not have a propositionalcontent,
accordingto Mellor,rather,"ithas notone content,buttwo, namelyP and
Q" (1993, p. 238). To accepta conditionalis not to acceptthatit is true,
that its truthcondition obtains. And yet it has a truthcondition-one
whichI may believe while not acceptingthe conditional,or one which I
maydisbelievewhile acceptingthe conditional.Now "hasa propositional
content"and"hasa truthcondition"aresomewhattechnicalterms,butwe
use theminterchangeably. It hasbeenhardenoughto get ourmindsround
the idea thatconditionalshave neither.It is harderstill, I think,to accept
thatthey have one butnot the other.
9.3. Conditionalsmay have truthconditionswhich are radicallycontext
dependent.Van Fraassen's complaint against Lewis's proof was the
assumptionthata conditionalwill expressthe samepropositionin differ-
ent belief states:"thelogical disasterwas precipitatednot by Stalnaker's
Thesis [theequationof the probabilityof a propositionwith a conditional

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
306 DorothyEdgington

probability],butby [Stalnaker's]ThesiscoupledwithLewis's metaphys-

ical realism"(VanFraassen1976, p. 275). Lewis had said "presumably
ourindicativeconditionalhas a fixedinterpretation, the samefor speakers
withdifferentbeliefs, andfor one speakerbeforeandaftera changein his
beliefs. Else how are disagreementsabout a conditional possible, or
changesof mind?"(Lewis 1976,p. 138). Stalnakershowedthateven in a
single belief distribution,the Equationcannothold for all conditionals
(see ?6.4 above);but,for a simpleconditionalwith no embeddedcondi-
tionals,we can,perhaps,alwaysfindsome propositionfor it to expressin
a given belief distribution.Stalnaker(1975) arguedfor a context-depend-
ent interpretation of his truthconditionsfor indicativeconditionals.Later
he said that"to play theirmethodologicalrole, [indicative]conditionals
mustbe too closely tied to the agentswho utterthemfor thosecondition-
als to expresspropositionswhichcouldbe separatedfromthe contextsin
whichthey areaccepted"(Stalnaker1984,p. 111).
Stalnaker(1975) addressesa problemwhichis the mirrorimageof that
addressedby Onrice andJackson.In ?6.5, p. 278, I mentionedtwo prima
facie desirablepropertiesof indicativeconditionals:(i) minimalcertainty
thatAvB (rulingoutjust -A &-B) is enoughfor certaintythatif -A, B;
and(ii) it is not necessarilyirrationalto disbelieveA yet disbelievethatif
A, B. I showedthatno propositioncan satisfyboth.The truth-functional
conditionalsatisfiesthe firstand not the second.GriceandJacksonpro-
vide a surrogatefor the second:if you disbelieveA, then, although-A
entails"IfA, B", the lattermay still be unassertable.Stalnaker'scondi-
tionalsatisfiesthe secondbutnotthe first.He arguesfor a surrogateof the
first.Althoughthe inference
Eitherthe butleror the gardenerdid it. Therefore,if the butler
didn'tdo it, the gardenerdid
is invalidon his semantics,nevertheless,wheneverthe firstis assertable,
so is the second.
Like Grice, Stalnakerappeals to the pragmaticsof communication.
Communicationtakes place in a context-a set of presuppositions,
knowledge or assumption. Call the "context-set"the set of possible
worldsat whichthe presuppositions hold.Now (1) indicativeas supposed
to subjunctiveconditionalsshouldhave antecedentscompatiblewith the
context-set.And (2) worldsin the context-setcountas closerto the actual
worldthanworldsoutsideit. So, if we areconsideringwhetherif A, B, by
(1) thereshouldbe some A-worldsin the context-set,and (2) the closest
A-world,whichis in the context-set,mustbe a B-worldif the conditional
is true. (3) A disjunction,AvB, is assertableonly if both A&-1B, and
-A &B, are takenas open possibilities,but -A &--B has been ruledout.
Hence,if it is assertable,the context-setcontainssome -A &B-worlds,no

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 307

-A&---B-worlds.So the nearest-A-world must be a B-world-indeed,

all the -A -worldsin the context-setareB-worlds.So, in this context,"If
-A, B" is true.
This is neat, but it does requireradicalcontext-dependence of condi-
tionalpropositions.Supposewe startoff in a contextwherewe have no
relevantinformationaboutA and B. Stalnaker'sconditionalbeing non-
truth-functional when the antecedentis false, the following six possibili-
ties arecompatiblewith ourcontext:

A -A B AvB -A>B

1. T F T T T
2. T F T T F
3. T F F T T
4. T F F T F
5. F T T T T
6. F T F F F

Now we learnthatAvB, andnothingelse of relevance.We eliminateline

6, nothingmore.The propositionexpressedin the old contextby "-A >B"
has not been shownto be true;for lines 2 and4, at which-A>B is false,
remainpossibilities in that context. Eliminatingone possibilitycannot
ipso facto rule out other,incompatiblepossibilities.Rather,our context
has changed.The new propositionwhich "-A>B" expressesin our new
contextis true.For,in this new context,all the -A-worlds areB-worlds.
The new conditionalwhichwe know to be trueis not the old conditional
whichwe werewonderingabout,buta differentconditional,governedby
a differentnearnessrelation.
Stalnaker'sremarkthat conditionals cannot "express propositions
which could be separatedfromthe contextsin which they are accepted"
was in the contextof the Sly Pete Story(see p. 294 above).Jack'scondi-
tionalis true-for-Jack-truein his context;the samesentenceis false-for-
Zack-false in Zack'scontext.All disagreement,all changeof mind, is
equivocation.I don't wish to deny thatthereis such a thing as context
dependence.But to appealto it to suchan extentmightbe thoughtto viti-
ate the pointof the appellation"true".
Stalnakermakes mattersratherworse by saying that the presupposi-
tions which definethe context-set"neednot be beliefs which are really
commonto the speakerand audience;in fact, they need not be beliefs at

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
308 DorothyEdgington

all.Thepresuppositions will includewhateverthe speakerfindsit conven-

ientto takefor granted,or to pretendto takefor granted"(Stalnaker,1975
pp. 141-2). But now, anyold contingentconditionalcan come out truein
some context-set.We take for grantedthat eitherwe won't do our rain
dance,or it will rain.In this context,"Ifwe do ourraindance,it will rain"
is true.We dance, and the droughtcontinues.Still, our conditionalwas
true in its originalcontext. This providesfurtherreasonfor doubtthat
such"truthconditions"deservetheirname.
9.4. VanFraassen(1976, pp. 279-82) hadan ideafor adaptingStalnaker's
semanticsso that your degree of belief in a StalnakerconditionalA>B
equals your b(B given A). Closely related ideas are found in McGee
(1989), Jeffrey(1991) and StalnakerandJeffrey(1994). The proposition
A>Bwill notbe independentof yourbelief state,butit will yield a theory
of what you shouldbelieve aboutcompoundsof conditionals.This has
been the focus of recentwork.
Stalnaker'sformalsemanticsis equippedwith a selectionfunction,f
whichselects,for anyworldw andanypropositionA, the "nearest"world
w' to w in whichA is true.Let w be the actualworld.If A is actuallytrue
f selects the actualworld.If A is actuallyfalse,f selects the "nearest"A-
world.A>B is trueiff B is trueat the worldf selects for A. Now suppose
you thinkit's 80%likely thatB givenA. Forexpositorypurposesonly,let
me expressthis:you think80%of theA-worldsareB-worlds.In theA&B-
worlds,A>Bis true.In theA&-B -worlds,A>Bis false. If the actualworld
is a -A -world,is it one for whichf selectsanA&Bworld,or one forwhich
f selects an A&-B1-world?Well, you don't know; and theremay be no
determinateanswerto the question,for theremay be nothingto choose
betweendifferentA-worlds.Stalnakernever did believe thattherewere
hardfacts aboutwhich worlds were "nearest",or how actualselection
functionswork:this was his way of "projectingepistemicstrategiesonto
the world"(see p. 273 above). The best projection,Van Fraassensug-
gested, would be this. If A is false, let the selection function select an A-
worldat random.Thenhow likely is it (foryou) thatit selectsa B-world?
80%,becauseyou think80%of the A-worldsareB-worlds.So A>B has
an 80%probabilityof beingtrueif anA-worldobtains,andan 80%prob-
abilityof beingtrueif a -A-world obtains;so an 80%probabilityof being
true.b(A>B) = b(BIA).66
Suppose90%of the redballs have blackspots.How likely is it that,if
you pick a redball (R), it will have a blackspot ( B)?Yourb(BIR) = 0.9.
R>Bis trueif R&B,false if R&-B; if -J?, thereis a 90%chancethatan
R&B-worldis "selected"and (R>B)is true.So b(R>B)= 0.9. You think
66 VanFraassenprovedfarfromtrivialresultsshowinghow to applythis idea

to an infiniteset of worlds.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 309

it's very likely thatif they areat home (H), the lightswill be on (L). Sup-
pose they're not at home. Then the selection functionis very likely to
select an H&L-worldratherthanan H&-L world.So b(H>L)is high.
Thisis to give up genuinetruthvaluesforthe conditionalwhenits ante-
cedentis false.The-A -worldsdon'treallydivideintothosein whichA>B
is trueandthosein whichA>Bis false.A>Bis indeterminate in all the-A-
worlds(whenyourb(BIA) is neither1 nor0). This would,I think,block
Stalnaker'sversionof thebombshell(see ?6.4). His proofdid assumethat
the -A-worlds divide into the -A &(A>B)-worlds, and the -A &-(A>B)-
worlds.
Jeffrey(1991) got the same effect by giving "IfA, B" an intermediate
"semanticvalue"equalto yourb(BIA) whenA is false. The conditional
is, as it were, 80%true,if -A, whenyour b(BIA) is 0.8. If we write"1"
for "T"and "0" for "F",we get a "truthtable"for the conditionalthat
looks like this

b A B IfA,B

0.4 1 1 1
0.1 1 0 0

0.5 0 0.8

Extendingthe notionof degreeof belief to the case wherethe objectof

belief is a three-valuedentity,he takesthe weightedaverageof its seman-
tic value. For degreesof belief in A&B,A&-B and-A as above,we get
b(ifA, B)=(0.4x 1)+(0.1 xO)+(0.5 x0.8)=0.8= b(BIA).
Jeffrey (1991) devised ways of assigning degrees of belief to com-
poundsof conditionalsfromthisbasis.StalnakerandJeffrey(1994) show
thatthis constructionis equivalentto VanFraassen's.Withina givenbelief
distribution,a conditionaldegreeof belief can be equatedwith a degree
of belief (in this extendedsense) in a three-valuedentity,anddegreesof
belief in compoundsof conditionalscan be assigned.
It is, admittedly,a ratherweirdthree-valuedentity.The 1 and0 aretruth
values.The0.8 is a degreeof belief. Thisis an odd mixtureof ingredients
in a weightedaverage.Replyingto Jeffrey(Edgington1991, pp. 203-5),
I thoughtwe could get the sameeffect by takingthe thirdvalueto be the
objectiveprobabilityof B given A (wherethis exists); andthatour epis-
temic estimationof this objectivethree-valuedentity would still be our
b(BIA). ButI was wrong.Thebombshellextends:thereis no three-valued
entitysuch that,in all belief distributions,yourepistemicestimateof its
value is your b(BIA).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
310 DorothyEdgington

McGee (1989) tackleddirectlythe questionof assigningprobabilities

to e.g. conjunctionsandnegationsof conditionalswhose values arecon-
ditionalprobabilities.Partof his methodologywas to investigatewhat
wouldbe fairbettingodds on conjunctionsof conditionals.
If we can finda generalway of assigningprobabilitiesto conjunctions
andnegationsof conditionals,we have a meansof assigningtruthvalues
to them(and,moreobviously,vice versa).Startwitha set of conditionals.
Formfrom it a set of "state-descriptions"-conjunctions which contain,
foreveryconditional,eitherit or its negation.Theseforma partition.They
are surrogatepossible worlds. Suppose we can assign probabilitiesto
these,whichsumto 1. The probabilityof a conditionalshouldbe the sum
of the probabilitiesof the state-descriptionsin whichit is "true",viz. un-
negated(see Adams1975,pp. 32-3). McGeeendsby showingthaton his
constructionconditionalswhich satisfy the Thesis can be construedas
Stalnaker-likeconditionalswith a randomselection function,like Van
Fraassen's.
I say "Stalnaker-like"because McGee makes one modification.An
importantpartof McGee's constructionis the equivalenceof "If A&B,
thenC" and"IfA, thenif B thenC".This is invalidon Stalnaker'sseman-
tics (see above,p. 282). But only a smallchangeis neededto modifythe
semanticsin this respect.This is one differencebetween Jeffrey'sand
McGee's constructions.Another,connected difference is that Jeffrey
applies his methodsto conditionalsin antecedentsof conditionals,and
McGee does not: McGee's antecedentsare always "factual"sentences.
Apartfromthesedifferences,they get the sameresults,by differentmeth-
ods.
Unfortunately,their resultsaboutcompoundsof conditionalsare not
altogetherpleasing.Lance(1991) has a plausiblecounterexample to their
commonaccountof conjunctionsof conditionals.I raisedsome further
difficulties(Edgington1991,pp.200-2). StalnakerandJeffrey(1994) and
McGee(privatecommunication) concedethattheirtheorieshavecounter-
intuitiveconsequences.Intuition,though,is not very robuston this sub-
ject. We seem to knowenoughaboutcompoundsof conditionalsto reject
certainclaims (see ?7.1 above). Othersremaincontroversial,and some
we don'tknowhow to understand. This is an areain whichthereis more
workgoing on, andtheremaybe moreto be learned.
It is hardto decide whetherthereis moreto be learned.Conditionals,
these theoristsconcede, are not ordinarypropositions,so there is no a
priori reason why there should be general routines for decoding
compoundsof them.On the otherhand,they are not thatdifferentfrom
ordinarypropositions;and they stand in definite relationsto ordinary
propositions (A&B entails "if A, B", which entails ADB, on Adams'

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 311

account).In ?7.1 I agreedwith Gibbardthat,with the help of a context,

we adoptadhoc strategiesfor findingsuitableinterpretations.
Sometimes,
without a suitable context, we fail to understandthem. These general
theoriesgive the impressionof going beyondthe data,andgoing beyond
our practicalneeds. But a generaltheorywhich did not clash with those
intuitionswe do have wouldbe an achievement,one for whichit is likely
we wouldfinda use.

10. Counterfactuals (by any other name)

10.1. Wereturnto the issues of ?1. Consider:

(ODD)If Oswalddidn'tkill Kennedy,no one else did;
(ODW)If Oswalddoesn'tkill Kennedy,no one else will;
(OHW) If Oswald hadn't killed Kennedy, no one else would
have.
"O"standsfor Oswald."DD","DW"and "HW"standfor the types of
conditional, "did/didn't-did","does/doesn't-will" and "had/hadn't-
would".The terminologyis borrowedfromBennett(1995). On the tradi-
tionalview, HWis segregatedfromDD andDW.Logic textbooks,if they
botherat all, tell us to set HW aside,anddrawfreelyon examplesof DD
andDW.HWis the subjectof specialistarticlesandmonographs-in the
early days, often relatedto the philosophyof science, less so since the
flourishingof possible-worldsemantics.
Dudmanhas convincedmanythatthe differencebetweenDW andHW
is merelyone of tense.67HW expressesat a latertime what DW would
expressat an earliertime.TherelationbetweenDW andHWis muchlike
the relationbetween"Johnis in Paristoday"said now, and"Johnwas in
Parislast week"said in a week's time.
Thereis considerableevidencethatthis is correct.I say "If you touch
thatsocket you will get an electricshock".You don't touchit. I use my
circuit-testinginstrument,to show you:"Yousee, I was right- if you had
touchedit, you would have got a shock".Or, if the result is different,
"Funny,the power must be off. I was wrong.You wouldn'thave got a
shockif you hadtouchedit".Traditionaltheoriesmightbe ableto explain
why establishingHW is sufficient to show thatmy earlierDW was right.
Mostsuchtheorieshavethe counterfactual as strongerthanthe indicative.
67
See Dudman (1983, 1984, 1984a, 1988, 1989, 1994). Smiley (1984), Ben-
nett (1988) and Mellor (1993) were among the converts. Bennett (1995) has re-
vertedto the traditionalview. Independentlyof Dudman,Ayers(1965), Adams
(1975, Ch. 4) and Ellis (1984) put forwardsimilarviews. See also the example
fromStrawson(1986) quotedon p. 245 above.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
312 DorothyEdgington

Buttheycannotexplainwhy it is necessaryandsufficient.In otherwords,

they cannotexplainhow it can show I was wrong-for a strongerstate-
mentmay be wrongwhile the weakerwas right.Yet,it appears,the one
andonly one thing we can and mustestablish,to decide whetherwhatI
saidwas rightor wrong,is whetheryou wouldhavegot a shockif you had
touchedit.
In similar vein, Adams (1975, pp. 114-5) has someone arguing:
"They'llbe hereby eight.If they'rehereby eight,we will eat at nine. So
we'll eat at nine."The argumentis rephrased,hungrily,at ten: "They
shouldhave beenhereby eight.If they hadbeenhereby eight,we would
have eatenat nine. So we shouldhave eatenat nine."68
Dudmanevinces linguisticevidencefor this view. "Wouldhave"is the
past tense of "will" (compare should havelshould, might havelmay, could
havelcan).Further,thereis the primafacie puzzlingfeatureof the formof
the verb in the antecedentsof DW and HW: "If it rainsldoesn'train
tomorrow, ..." "If it had rainedlhadn't rained yesterday ...". "It rains
tomorrow"is not a self-standingsentence-or at least,the tense is wrong
for it to convey,as a self-standingsentence,whatit seemsto conveyin the
if-clause.(Perhapswe can imagineGod,havingdecidedhow thingsshall
be, saying"Itrainstomorrow"-compare"I am seeing Janetomorrow".
But this is irrelevantto the interpretation
in the conditional.)These odd
tensesdo servea function,in clarifyingthe temporalrelationbetweenthe
subordinateand main clauses. "If she receives the lettertomorrow,she
will be in a badmood."Construethis as expressinga judgementthatshe
will be in a bad mood,qualifiedby the if-clause(Dudman1984, p. 151).
The if-clausemustbe aboutwhatcomes earlier,or at least not later,than
whatthe mainclauseis about.69Weneeda differentformwhenthis is not
so: "If she will receive the lettertomorrow(anyway),I shall warnher
aboutit today";not "If she receivesthe lettertomorrow,I shall warnher
aboutit today".(This exampleis MichaelWoods'.) So we can explain
why we get the "backwardtense shift"in the antecedentof DW -to clar-
ify the temporalrelations;HW,being the pasttense of DW,has its ante-
cedentin the past-pasttense, for the same reason(see Bennett1988, pp.
513-5).
Whatdoes it meanto say thatHWis the pasttenseof DW,or,adapting
a sloganfromBennett,"everyHadn't-wouldwas once a Doesn't-will"?It
68 "Should have" here indicates that something was to be expected-the epis-
temic "should". Note that replacing "will" by "should" in the first premiss and the
conclusion of the first argument would leave the point intact. Adams thinks of
counterfactuals as "epistemic past tense" like "should have".
69 This is not exceptionless. See Dudman
(1984, pp. 150-2). He argues that
where it fails, the conditional is, in a more subtle way, still to be interpretedas for-
ward-looking (p. 152).

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 313

doesn'tmeanthis:if you now accepta hadn't-would,you earlieraccepted

a doesn't-will.Thatwas so in some of the examplesabove-the dinner
guests,andthe firstoutcomeof the electricitytest, butit is not so in gen-
eral.Consider(ODW) and (OHW).You accept(OHW).Before the kill-
ing, you didn'taccept(ODW):you had neverheardof Oswald.Perhaps
you weren'talive then.Perhapsno one, beforethe killing,hadthe infor-
mationneededto accept(ODW).We gatherinformationretrospectively,
abouthow thingsstood at the time in question(cf. the circuit-testing).If
you accept(OHW),then,by yourpresentepistemiclights,giventhe infor-
mationyou now have, you thinkthatthe world,before the killing, was
such as to make(ODW)the rightthingto think.The conditionalthought
now expressedby (OHW)is the sameconditionalthoughtas wouldhave
been expressedby the doesn't-will,beforethe killing.70The situationis
analogouswith "Johnwas in Parislast week".
I thinkthis is essentiallyrightfor a largenumberof centralcases-the
most commonlymet, andthe most important,counterfactuals. For some
counterfactualsit will be hardto find a temporaldimension.To cover
them,I thinkthereis a naturalgeneralisationof the claimas we have it. I
turnto the otherhalf of the Dudmanandpost-Dudmanstory.
10.2. WhataboutDD? Dudmanthinksthis is a differentspeciesfromDW
andHW.His linguisticevidenceis thatherewe do havetwo self-standing
sentencesin theircustomarysense.Fromthat,he concludesthattheyhave
differentroles.He thinksDD is to be interpretedas a condensedargument
(1984, p. 147), DW and HW as judgementsor "verdicts"on the main
clause,qualifiedby the subordinateclause(1984, pp. 151-2).
It is not clearwhy thesetwo accountsof whatyou aredoing,in putting
forwardan "if"-sentence,are incompatible.If I give a condensedargu-
ment,I am in a positionto reachthe verdict,qualifiedby the "if";con-
versely,if my qualifiedverdictis challenged,then,like anyverdict,if it is
to stand,I am obligedto give reasonsfor it-to spell out whatcould be
construedas the argumentwhichwas condensed.I mayof coursefail, and
say "Well,I still have a hunchthatMarywill come if Johndoesn't".And
of course,my reasonsneed not have been vivid, or even operative,at the
time. But this distinction-to whatextentyou are actuallydoing a bit of
inferring,as opposedto "imaginativelyprojecting"(whichDudmansees
as the processbehindthe verdictin DW and HW)-doesn't necessarily
coincidewiththe grammaticaldistinction."If Maryleaves, the Headship
will be offeredto John(I inferthisfromthe generalprocedureslaid down
in the minutes;every other possibility can be ruled out)"-condensed
argument."If Fred was there, there was a fight (I just feel it in my
70 Nothing
hangs on the existence of precise general criteria for sameness of
thought here: an everyday sense will do.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
314 DorothyEdgington

bones)"-imaginative projection.JudgingthatB on the suppositionthat

A coversboth these cases. Dudman'sdistinctionis too fine, and doesn't
coincide with the grammaticaldifferenceit seeks to explain (see A. J.
Dale, 1985).
How thento explainthe grammaticaldifference?I tentativelysuggest
this.Sometimesit is usefulto constrainthetemporalrelationsbetweenthe
two clauses,as we do in DW andHW.But it is usefulalso to have a way
of leaving them unconstrained,as they are in DD-so we don't tamper
withthe verbin the if-clause."Ifhe atethe apple,he was ill"couldbe said
becauseyou thinkthe applewas poisoned;or it couldbe saidbecauseyou
know he only eats fruitwhen he is ill. Or it could be said to rebutsome-
one's claim thathe couldn'thave been ill, for he took an applewith him
for his lunch,andan applea day keepsthe doctoraway.
Those philosophersworkingin termsof traditionaltheoreticalframe-
works(which Dudmanrejects)who have been influencedby Dudman's
workhaveagreedwithhimthatDD is to be segregatedfromDW andHW.
This hts had the effect of shrinking the class of "indicatives" and
expandingthe class of "subjunctives"(as we used to call them). There
was some comfortin this, for Gibbard'sargument(1981), in particular,
had carriedconviction that DDs were "purelyepistemic", devoid of
factualcontent.Dudman'sworkprovideda reasonfor savingDWs from
this fate-they could be treatedas possible-world conditionals, with
objective truthconditions. Further,it was taken for granted,the OK
cases-the fact that you can accept (OHW) and reject (ODD)-does
show thattheremustbe two kindsof conditionals.Thatleaves us free to
relocateDW (see Bennett,1988).
I thinkthis is wrong.I will show thatthereis a naturalexplanationof
the OK cases if we thinkof all conditionalsin termsof conditionalprob-
ability. Indeed, doing so predicts the OK cases. This way of thinking
aboutcounterfactuals explainswhy andhow counterfactuals matterto us.
AndI shallarguethatwe haveamplereasonto treatDD andDW as a uni-
fied class. I am a traditionalistin that I think the bigger distinctionis
betweenDD andDW on the one hand,andHWon the other.But the dis-
tinctionisn't thatbig-Dudman was rightto thinkit is (mainly)one of
tense.
10.3 Howto explainthe OKcases?It is a commonplacethatprobabilities,
andconditionalprobabilities,changewith time. Yourdegreeof belief in
B, or in B given A, changeswith new information;and-in as informal
sense as possible-the forward-lookingobjective chances of future
events-of winningthe election,of recoveryor death,of gettinga partic-
ularjob, of solving the murder,of catchingthe train-go up and down,
untilthe eventscome aboutandthey settledownto 1 or 0. Ourepistemic

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 315

degreesof belief arethebest opinionswe havebeenableto arriveat about

these objectivechances.So we have plentyof use for thejudgementthat
it was probablethat such-and-such-it was to be expected-but it no
longeris (or vice versa).I get a messageby telephonethatI have won a
lotteryprize.It could be a hoax, but it doesn't seem like one. I thinkit's
probablethatI have won. Soon after,I am certainthatI have won. But it
was (I believe) far fromprobable,a little earlier,thatI was going to win
(thatI wouldwin); my winningwas not to be expected.I endorsean ear-
lier very low degree of belief in my winning.Similarlywith Oswald:I
thinkhe did it;I amcertainthatif he didn'tsomeoneelse did.Let 0, S and
N stand for "Oswald ...", "Someone else ..." and "No one else ... ". My
present b(S given 0) and b(N given 0) are in the vicinity of 1 and 0
respectively.Nevertheless, I think that before the killing, it was very
improbablethat S on the assumptionthat 0; by my presentepistemic
lights,p(S given 0) was low andp(N given 0) was high, beforethe kill-
ing; in accepting(OHW),I endorsea hypotheticalb'(N given 0) which
couldbe expressed,beforethe killing,by "IfOswalddoesn'tdo it, no one
else will".
If thisis right,it defeatsthe argument(pp.237-8 above)thattheremust
be two conditionalconnectives.The argumentwas: the sametwo propo-
sitions, N and 0, conditionallyconnected,express (OHW)and (ODD).
We accept (OHW) and reject (ODD). So the conditionalconnectivein
(OHW)doesn'tmeanthe same as thatin (ODD). On the above analysis,
we do have the sametwo propositions,N and0, conditionallyconnected
in the same way, but concerningdifferenttimes.Acceptingsomethingas
probablenow,andacceptingthe samethingas havingbeenprobablethen,
aremutuallyindependentjudgements.
The independenceis worthstressing.Compare(ODD) and (OHW)to
the pairwiththe sameantecedentsbut"someoneelse did/wouldhave"in
the consequents(whichwe consideredin ?1, p. 237). Call the latterpair
(ODD') and (OHW').You accept(OHW)and reject(ODD);you accept
(ODD')andreject(OHW').Wegive a uniform,mirrorimage,explanation
of the two cases. Anyone who holds thatthe counterfactualis logically
strongerthanthe indicative(LewisandJackson,for instance)cannotgive
a uniformexplanationof the two pairs. They can explain (ODD') and
(OHW')-you acceptthe weakerbut rejectthe stronger.But they cannot
offer the same explanationof (ODD) and (OHW);for on their view,
(OHW)entails(ODD). Yet (it seems) you acceptthe strongerand reject
the weaker.Lewis (1973 p. 72) arguesthatalthough(ODD) follows from
(OHW),it is not assertablewheneverone has groundsto believe thatit is
true.Thisis less attractivethananentirelyuniformexplanationof thetwo
pairs.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
316 DorothyEdgington

Could we make the case for a merely temporaldistinctionbetween

(ODD) and (OHW)using not the notionof probabilitybut the notionof
truth?Not so readily.Wehavereliedjust on the platitudethatwhatis now
probablemayearlierhavebeenimprobable(andvice versa).Theideathat
somethingwas truebut is now false (or vice versa)is less palatable.Dif-
ferencein truthvalue is usuallytakenas sufficientfor differencein con-
tent. "It was truethat it is raining,but it is now false that it is raining"
showsthatwe havetwo, temporallyindexed,propositions,not one. Well,
we could providetemporalindicesfor the conditionalitself, as opposed
to its constituents,so thatwe couldhave (If A, B),trueand(If A, B)t,false.
Those who say conditionals express context-dependentpropositions
would not object.But it is revisionary,comparedwith the independently
familiarfact thatprobabilitieschangewith time.
10.4. Therehas always been a vague worryaboutwhy counterfactuals
matter.Arethey notjust idle speculation,"Mondaymorningquarterback-
ing", in,Adams'phrase(1975, p. 133; see also Adams, 1993)? It is an
advantageof the probabilisticpast-tenseapproachto counterfactuals
that
such past-tenseprobabilityjudgementsare alreadygenerallyrecognised
as indispensableto empiricalreasoning.We observesomething,which,
we think, was to be expected if A; and was not to be expected if -A. This
gives us a reasonto believe A. The doctorobservesthe symptoms.They
were to be expected if the patient took arsenic; they were not to be
expectedif the patientdidn'ttakearsenic.
Consideragain"They'renot at home;for the lightsareoff; andif they
hadbeen at home,the lightswouldhavebeenon".I can see thatthe lights
are off. Whetheror not they'reat home, the lights are off. If they are at
home,the lightsareoff, andif they arenot at home,the lightsareoff. It is
not my presentconditionalprobabilitiesthat matter,but my judgement
aboutpastones: whatwas to be expected(beforeI saw thatthe lights are
off) aboutwhetherthe lightswouldbe on or off, if they were at home.
This "counterfactualmodustollens"is useful, but defeasible,in two
ways. First,my companionreplies"Iagreethatit was to be expectedthat
the lights would be on if they were at home;but, they always leave the
lights on when they go out. So it was also to be expectedthatthe lights
would be on, had they gone out. So somethingstrangehas happened-
perhapstherewas a powercut, andthey may be in afterall".
Second,"I agreethatit was to be expectedthatthe lights wouldbe on
if they wereat home.But theypromisedme they wouldbe in. Of course,
maybeanemergencycalledthemaway.Ormaybetherehas beena power
cut, andthey areat home afterall".

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 317

These cases are partand parcel of a well-understoodframeworkof

probabilisticreasoning,which deals with how yourpast conditionaland
unconditionaljudgementsinteractwith yournew information.
If HWis the pasttenseof DW,is therea past-tenseof DD? I thinkthere
is, butit is of littleinterest.Marysays "Johndid the shopping".I say later
"Marysaid thatJohnhad done the shopping".Marysays: "Johndid the
shoppingif Fred didn't".I say later"Marysaid thatJohnhad done the
shoppingif Fredhadn't".Supposewe discover,to our amazement,that
Kennedyis alive and well and living in Bolivia: it was a doppelganger
who was killed. Hardlya soul knew.Whatdid we believe beforewe dis-
coveredthis?Webelievedthatif Oswaldhadn'tkilledKennedy,someone
else had done it. Judgementslike this are of little value, comparedwith
ourjudgementsaboutpastforward-lookingconditionals.
Does Doesn't-willhave a futuretense, as well as a pasttense?No, and
for a goodreason.Youcanthink"It'sprobablenow,butit wasn'tprobable
yesterday".You can't reasonablythink"It'sprobablenow, but it will be
improbabletomorrow".If you believe it will be improbabletomorrow,
thatis reasonto believe thatit is improbablenow.
10.5. I turnto positive reasonsfor a unifiedtreatmentof DD and DW.
First,withboth,Ramsey'sThesisis plausible:you believe "IfA, B" to the
extentthatyou havea high degreeof belief in B on the suppositionthatA;
you takeyourself,ceterisparibus,to be justifiedin assertingB, condition-
ally uponA; ceterisparibus,you will inferB if you learnthatA. On the
otherhand,to accept(OHW)is not to have a high degreeof belief in N
given 0. Youthinkit was likelythatno one was goingto if Oswaldhadn't,
but you don'tthinkit is likely thatno one did if Oswalddidn't;you cer-
tainly don't assert "No one did" conditionallyupon "Oswalddidn't".
(Youdo assert"No one wouldhave"(no one was going to) on the suppo-
sitionthatOswaldhadn't.)Youcertainlywouldn'tinfer"No one did"on
learning"Oswalddidn't".
Second,we saw (? 1.1, pp. 238-9), futureOK cases can be constructed
to distinguishDW from HW. Returnto my two prisoners,Smith and
Jones,one of whom,I'm now sure,is planningto escape.I'm prettysure
it is Jones.But whatif I'm wrong?Thenit's Smith.Thatcanbe expressed
as DW:if Jonesdoesn't,Smithwill. Whatif Joneshadnotbeen goingto?
HereI do not abandonmy belief thathe will. I consideran earliertime,
beforehe haddecidedto escape,andconsiderwhether,at thatearliertime
it wouldbe rightto say "IfJonesdoesn't,Smithwill".I have no reasonto
thinkso. If Joneshadnot been going to, then,most likely,the evidenceI
did come acrosswouldnot havebeenthere.I reject"IfJoneshadn'tbeen
going to escape,Smithwouldescape".I rejectHW.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
318 Dorothy Edgington

Here let us digressto considera reasonBennett(1995, ?8) gives for

rejectingthe claim which he once held, and I have arguedfor: thatHW,
now, standsor falls with a DW,backthen.He has someonewho is com-
pletely certainthat some futureevent will happen.Moreover,this cer-
taintyis centralto his belief system:the whole systemwouldbe in a state
of shock if he is wrongaboutthis. For example,a scientistis sure thata
comet will hit Jupiternext week. He accepts"If the comet hadn'tbeen
going to hit Jupiternext week, Jupiter'ssurfacewouldbe F in two weeks
from now".But he doesn't accept"If the comet doesn'thit Jupiternext
week,Jupiter'ssurfacewill be F two weeksfromnow";insteadhe accepts
"Ifthe cometdoesn'thit Jupiternext week, somethingmustbe so wrong
in my understanding of the situationthatI don'tknow whatto think".
Now thisis like a forward-looking OKcase -like the case of Smithand
Jones.The scientistbelieves thatnothing,now,can preventthe collision.
"Whatif it doesn'thappen?ThenI don'tknowwhatto think"."Whatif it
hadn'tbeengoingto happen?"-the scientistadoptsa hypotheticalearlier
standpoitnt, relativeto which it can't be ruled out that it won't happen.
(Compare"Whatif Joneshadn'tbeengoingto escape?")Fromthatstand-
point,his amplegeneralknowledgetells himthatJupiter'ssurfacewill be
F two weeksfromnow.His acceptanceof "Ifthe comethadn'tbeengoing
to collidewithJupiter,Jupiter'ssurfacewouldbe F two weeksfromnow"
(HW) standsor falls with his acceptingthatthe correspondingDW was
the rightthingto thinkat anearliertime.DW is nottherightthingto think
now, given the evidencehe now has. But if the comet had not been on a
collision coursewith Jupiter,he wouldn'thave the evidencethathe does
have, thatcollision is now inevitable.
So this is not a counterexampleto the claim thatHW now, standsor
falls withDWbackthen.It is a counterexample to the claimthatHWnow,
standsor falls with DW now.But, as probabilitieschangewith time, DW
can be the wrong thing to think now in the light of presentevidence,
althoughit was the rightthingto thinkfroma hypotheticalearlierstand-
point,when its antecedentwas an open possibility.Comparethe familiar
OK case.
Let us consideranotherpuzzlecase, raisedby Adams(1975, pp. 129-
33)- "a counterexample", he calledit, to his own theoryof counterfactu-
als as past-tenseindicatives.This examplehas the same structureas his.
Johnis ill. Thereare two drugs,D and E, which would help him. If he
takesbothor neither,he'll get worse.If he takesjust one, he'll get better.
(The harmfuleffects of takingboth are not well known, however,and
Johnwon't realisethis.) Both drugsare in exceedinglyshortsupply,and
it's very unlikelythatJohnwill get eitherdrug.But it's around100 times
less likely that he'll get E than it is that he'll get D, and immensely

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 319

unlikelythathe'll get hold of both. (Thereis no causallink betweenthe

suppliersof the two drugs.) I think "if John takes E, he'll get better"
(DW).Now,Johndoes get better.Muchthe mostlikelyexplanationis that
he got hold of D. So now I think,"if he had takenE, he would not have
got better".My presentHW disagreeswith my previousDW.
Well, my firstjudgementwas rightabouthow the chanceslay before
Johngot eitherdrug.My new information,thatJohnrecovered,makesme
thinkthatthis stateof affairsis mostlikely to havebeenbroughtaboutby
John's getting D. On this hypothesis, taking E would preventJohn's
recovery.My presentHWcorrespondsto a DW acceptedin a hypothetical
stateof informationin whichI knowhe obtainedD butnotE: the situation
I have reasonto thinkobtains.This is no counterexampleto the general
claimthatevery HW correspondsto an earlierDW.It bringsout some of
the subtletiesinvolvedin tryingto assessthe objectivechances.Thereare
conceptuallycomplexissues involvedin makingthe slogan "EveryHW
was once a DW"moreprecise(see Skyrms,1981, 1994,forfurtherreflec-
tions on thisAkind of case).
Continuingwith reasonsfor a unifiedtreatmentof DD andDW:third,
we saw that the Gibbardphenomenonapplies to DW as well as DD.
Recall the example of the vaccine (?8, p. 295). The phenomenoncan
applyto HWas well. WehadJonesrunoverby a bus, andX andYsaying
respectively"If Joneshad got the disease,he wouldhave got S" and "If
Joneshad got the disease,he wouldnot have got S". Hereit is naturalto
interpretX and Yas talkingabouthow thingsstoodjust beforeJoneswas
run over, and at thattime, the Gibbardphenomenonapplies-each has
adequatereasonfor his opinion,and the world rules out therebeing an
objectivelycorrectopinion,for it rulesoutJones'gettingthe disease.Had
they been talkingaboutan earliertime, beforeJoneshad/hadone or both
vaccines,therecouldbe an objectivecorrectopinionabouthow likely it
is thathe get S if he gets the disease.I discussin the nextsectionthe ques-
tion of objectivityfor counterfactuals.
Fourth,a pointof Jackson's(1990) in favourof the traditionaldistinc-
tion: we can believe (HW) "If Oswald hadn't killed Kennedy,things
wouldhavebeendifferentfromthe way they actuallyare(orwere,or will
be)";but neither(DW) "If it rains,thingswill be differentfromthe way
they actuallywill be";nor (DD) "If Oswalddidn'tkill Kennedy,things
weredifferentfromthe way they actuallywere".This is explicableon the
view I have been defending: (DD) and (DW) representyour present
beliefs abouthow the world(actually)will be or was, conditionallyupon
a supposition.(HW)representsyourbelief aboutwhatwas going to hap-
pen, conditionallyupon somethingelse's havinghappenedwhich, often,
you thinkactuallydidn't.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
320 DorothyEdgington

Jackson'spointhasa converse:we canalso say,non-trivially,"Ifhe had

takenarsenic,thingswouldbe just as they actuallyare";"Iftherewere a
bus coming,peoplewouldbe behavingin just this way";while it is trivial
to say "Ifhe took arsenic,thingsarejust as they actuallyare";"Ifit rains,
thingswill be just as they actuallywill be".The non-trivialthoughtplays
an importantrole in empiricalinvestigation.
Finally,the if-clause in DD and DW can occur in conditionalcom-
mandsandpromises.ForDW,this is obvious.ForDD: "Ifhe didn'tgive
his lecture,tell the Principal";"If Jonesdid it, I promisenot to tell".On
the otherhand,"If he hadn'tgiven his lecture,tell the Principal"and"If
Joneshadn'tdoneit, I promisenot to tell"arenonsense.(I owe this point
to MichaelFirestone.)Thereare no past-tensecommandsandpromises:
"Iforbidyou to havegone out lastnight";"Ipromisenot to havegone out
last night".
10.6. In ?8 I raisedmetaphysicaldoubtsabouttruthfor indicativecondi-
tionals.Wherethereis truth,thereshouldbe an ideal thingto think.Gib-
bard'sargumentshows that, with DD, thereis never a non-trivialideal
thingto thinkabout"IfA, B": the fully-informedpersonknowsA&B, or
knows A&-B, or the questionwhetherif A, B doesn't arise;partially-
informedpeople can come to good but conflictingopinions.Thereis no
way of segregatingyour ignoranceof whetherA from otherknowledge
you would need in orderto have the best opinionpossible (see Bennett,
1988, pp. 520-1). DW is no differentif our "idealperspective"allows
magicalknowledgeof thefuture(G,, p. 296), or even if it doesn't,if deter-
minismis true(G2,p. 298). If determinismis not true,therecan be pres-
ently optimalopinionson DW, providedthe antecedentcannotnow be
ruledout (see p. 298). If thesedo not yield a probabilityof 1 or 0 for B if
A, theyareonly temporarilyideal,for probabilitieschange.Andwhenthe
futurehas become past, then,even if they were I or 0, they cease to be
optimal.I know thatthe objectivechancewas 1 thatif it rainedon Mon-
day, the riveroverflowedits bankson Tuesday.Someoneelse, who saw
thatthe riverdidn'toverflow,is just as well-placedto think"If it rained
on Monday,the riverdidn'toverflowits bankson Tuesday"(see p. 298).
How do G, andG2farewithHWs?If theylive in a deterministicworld,
andknow everything,it is dubiouswhetherthey have any use for them:
for any proposition,they knowits truthvalue,andknow thatit neverhad
any chanceof being different.If they live in indeterministicworlds,they
canconsidera time whenA hadsome chanceof happening,andknowthe
truechanceof B given A at thattime. It may be 1, it may be 0, it may be
somethingin-between.G, can know whatthe chancewas thatthe river
wouldhave overflowed,hadit rained(see p. 296). (I don't see how even
G, couldknowmorethanthat.If indeterminismgovernsthe case, can he

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 321

tell you how the coin wouldhave landedif you hadtossedit, as opposed
to whatthe chanceswere?)G2may know thatthe presentchanceof rain
is 0, butcanalso pronounceon how likelyit was thattheriverwouldover-
flow if it hadrained,by consideringan earliertime whenthe chancewas
not 0 (see p. 298).
In one sense, HWs are the most objectivekind of conditional:if they
are counterfactual,then the conditionalchance,just beforethey became
counterfactual, staysfixed.If it does rain,the actualoutcomesextinguish
the chances.If it doesn'train,the chancethatthe riverwouldhave over-
flowedif it had,remains.It need not, however,be 1 or 0.
I am not, of course, denying that there is plenty of "objectivity"in
anothersense of thatterm,in our assessmentof conditionalsof all sorts.
We can and shouldweigh the evidencewe have with skill and care.We
should acquiremore evidence if necessary.I am denying that there is
somethingfixedto aim at -a truthvalue.
10.7. I ramin apparentconflictwith Lewis when I say, threeparagraphs
back,thatGIandG2haveno use forcounterfactuals if determinismis true.
Lewis's "littlemiracles"allow counterfactuals to be truein a determinis-
tic world(see Lewis 1973p. 75,,1979;andsee above,pp.254-6). Lewis's
"littlemiracles"providea surrogatefor indeterminism,to cover the pos-
sibility thatour world, which is not transparently deterministic,should
really be deterministicat bottom.They are appealedto, if necessary,in
accountingfor ourthoughtaboutwhatwouldhavehappenedhadthe tree
not blown over, or had Nixon pressedthe button,when, at least appar-
ently, these things had some chance of happening.If determinismwere
transparent (as it wouldbe to GIandG2)it is notclearthatwe wouldhave
use for thesethought-experiments involvingsmallmiracles.But if appar-
ent indeterminismis necessaryfor counterfactuals to get off the groundin
termsof smallmiracles,the structureexhibitedby the coin problem(see
pp. 249-50 andpp. 258-9 above)becomesprevalent.Supposedetermin-
ism is truebutfine-grained.Therewill be manywaysof realisingthe ante-
cedentof "Ifyou hadtossedthecoin tentimes,it wouldhavelandedheads
atleastonce".Mostbutnotquiteall smallmiracleswhichdo so will guar-
anteethe truthof the consequent.So the counterfactual is certainlyfalse.
Likewisefor a largeproportionof the counterfactuals we use. If determin-
ism is true,Lewis cannotoffer us the surrogatethoughtthatif you had
tossed it ten times, therewouldbe a very high chancethatit wouldland
headsat least once. For he holds that,if determinismis true,all chances
are 1 or0(see Lewis 1980,pp. 117-21).
10.8. Conditionalprobabilities,I have argued,play an importantrole in
empiricalreasoning.Manyphilosophershave arguedthatwhen it comes
to practicalreasoning,we need a differentkind of conditional(see e.g

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
322 DorothyEdgington

Appiah(1985) ch. 3); conditionalprobabilitiesmay represent"evidential

conditionals"but not the sortof "causalconditionals"we need for deci-
sion theory.Lewis (1981) surveysa varietyof proposalson thistopic,and
makeshis own.
Therearealmostalwaysbeliefs aboutcausalstructurebehindthe con-
ditionalswe accept.But only somekindsof causalstructuregive us a rea-
son for actingin a particularway.A fantasyattributedto the statistician
RonaldFisheris thatsmokingdoes not cause cancer,but more smokers
thannon-smokersget cancerbecausea tendencyto smokeanda tendency
to cancerareproducedby a commongene:-if I smoke,thatmakesit more
likely that I'll get cancer;but, on this fantasy,that isn't a reasonnot to
smoke.(It may be a reasonto get life insurance.)
Conditionalbeliefs with antecedents"IfI do x" bearon decision-mak-
ing in differentways dependingon the agent'sbeliefs aboutthe causal
structureof the case.Thisis a problemfor a unitarytheoryof conditionals
if, and only if, you aim to analyzecausationin termsof conditionals,as
LewisK(1973,and Postscripts1986) does. This projectfaces difficulties,
andLewis allows thatthereis "unfinishedbusiness"(1994, p. 473, refer-
ringto PeterMenzies(1989)). Supposeyou don'tthinkcausationcan be
analyzedin termsof conditionals,or don'tthinkthatcausationcanbe ana-
lyzed at all (see e.g. Stalnaker1984 pp. 157-60).Thenconditionalproba-
bilitiestogetherwithbeliefsaboutcausalstructureareneededfordecision
theory(see e.g. Pearl, 1988). (In my discussionof objectivityin ?8 and
? 10.5,1 haveappealedto causation,in motivatingtheconceptof objective
chance,andin speakingof whatis now causallypossible.)
10.9. Consider"IfI hadbeen 8 feet tall, I wouldhave been a freak".The
following is ratherstrained:it's not now causally possible that you
become 8 feet tall. Go backin time to whenyou were a child, andcould
have been fed lots of steroidsandthe like; thenit wouldhave been right
to say "If he becomes 8 feet tall, he'll be a freak".And that's what I
expressnow by saying"Ifyou hadbeen 8 feet tall, you wouldhavebeen
a freak". If that is strained, what about "If I were a blackbird ...", "If kan-
garooshadno tails","Ifgravitywentby the inversecubeof distance"?A
textbookon elementarynumbertheorycontainsthe sentence"Were6 and
8 relativelyprime,of course,this situationwouldnotarise"(Burton1980,
p. 28).
Some of the above are doubtfulcandidatesfor literalsense. (I found
myself, concerningJacksonandLewis, aboutto write"If I were they".)
Thosewhichdo makeliteralsense canbe understoodby analogywiththe
standardtemporalcase. In the temporalcase, you use HW,or something
like it, to adopt,hypothetically,an epistemicperspective,differentfrom
youractualone, which you take to be suitedto some time in the past, a

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 323

time when the antecedentwas (you think)a live possibility.These other

casescanalsobe seenas adoptinghypotheticallyanepistemicperspective,
differentfromyouractualone,in whichtheantecedentis anepistemicpos-
sibility,but not necessarilyone whichyou thinkit wouldever be rightto
takeas a live possibility.Therewill be no generalpreciserulesfordeciding
whateffect the suppositionhas on yourotherbeliefs, thoughtherewill be
easy cases. The first,"freak"case may be one. (On the otherhand,if you
hadbeen 8 feet tall, maybeotherpeoplewouldhavebeen too).
The term"counterfactual" fits well enough some uses of HW:those
whereyou takeyourselfto know the antecedentis false (I didn'tpressit,
but if I had ...); those in which you take yourself to know the consequent
is false, (butit wouldhavebeen trueif ... had).It fits ill withthosewhere
you takeyourselfto know thatthe consequentis true(andmoreover,this
is what would have happenedif ...). What about "subjunctive"?I am
inclined to thinkthat the doubly hypotheticalaspect of HW-not only
conditional,but assessedfroma hypotheticalperspectiveotherthanyour
actualcfpe-might be registeredby this mood, whose functionseems to
be to emphasisethata sentenceis not being used assertorically.But who
am Ito say?
The studyof conditionalsraises hardquestions-about truth,object-
ivity, propositions.Who would have thoughtthatone little word could
disturbso much?I am remindedof Russell's remark,on anotherlittle
word,to whichhe was aboutto devotetwo chapters:"itis a wordof very
greatimportance:... I wouldgive the doctrineof thiswordif I were 'dead
fromthe waistdown'andnotmerelyin a prison"(Russell1919,p. 167).1

Department of Philosophy DOROTHY EDGINGTON

Birkbeck College
University of London
Malet Street
London WCIE 7HX
UK
[email protected]

7 ' A finalwordof thanksis due to MarkSainsburyandthe Mindteam,espe-

cially GinnyWatkins,for theirskill, good-natureandforbearance.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
324 Dorothy Edgington

REFERENCES
Adams,E. W. 1965:"A Logic of Conditionals".Inquiry,8, pp. 166-97.
1966: "Probabilityand the Logic of Conditionals",in Hintikka,J.
andSuppes,P. (eds.) 1966, pp. 256-316.
1970: "Subjunctiveand IndicativeConditionals".Foundationsof
Language, 6, pp. 89-94.
1975: The Logic of Conditionals. Dordrecht:Reidel.
PacificPhilo-
1993:"Onthe Rightnessof CertainCounterfactuals".
sophical Quarterly, 74, pp. 1-10.
Anderson,Alan Ross 1951:"A Note on SubjunctiveandCounterfactual
Conditionals".Analysis,12, pp. 35-8.
Appiah, A. 1985: Assertion and Conditionals. Cambridge: Cambridge
UniversityPress.
1986:"TheImportanceof Triviality".PhilosophicalReview,95, pp.
209-31.
andSubjunctiveConditionals".
Ayers,M. R. 1965:"Counterfactual Mind
74, pp. 347-64.
Bayes, Thomas1940:"AnEssayTowardsSolvinga Problemin the Doc-
trineof Chances",in Deming,W. E. (ed.) 1940.Originallypublished
in Transactions of the Royal Society of London, 53, 1763, pp. 370-
418.
Cana-
Bennett,Jonathan1974:Reviewof DavidLewis, Counterfactuals.
dian Journal of Philosophy, 4, pp. 381-402.
1984: "Counterfactualsand TemporalDirection".Philosophical
Review,93, pp. 57-91.
1988: "Farewellto the PhlogistonTheoryof Conditionals".Mind,
97, pp. 509-27.
1995: "ClassifyingConditionals:The TraditionalWay is Right".
Mind, 104,414, pp. 331-4.
Bernoulli,Jacques1713:Ars Conjectandi.Basle.
Black,Maxed. 1950:PhilosophicalAnalysis.EnglewoodCliffs:Prentice
Hall.
Blackburn,Simon, 1986: "How Can We Tell Whethera Commitment
Has a TruthCondition?",in Travis,C. (ed.) 1986, pp. 201-32.
Burton,David M. 1980:ElementaryNumberTheory.Boston:Allyn and
Bacon
Carlstrom,I. andHill, C. 1978:Review of Adams1975,in Philosophyof
Science, 45, pp. 155-8.
Carnap,R. 1936: "Testabilityand Meaning".Philosophyof Science, 3,
pp. 419-71; 4, pp. 1-40.
Conditional".
Chisholm,R. M. 1946:"TheContrary-to-fact Mind,55, pp.
289-307.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 325

Davidson,Donald 1980: "MentalEvents"in his Essays on Actionsand

Events.Oxford:ClarendonPress,pp. 207-25.
Dale, A.J. 1985:"IstheFutureUnreasonable?".
Analysis,45, pp. 179-83.
Deming, W. E. 1940: Facsimiles of Two Papers by Bayes. Washington
D.C: US Departmentof Agriculture.
Dudman,V. H. 1983"TenseandTimein EnglishVerbClustersof thePri-
mary Pattern".Australian Journal of Linguistics, 3, pp. 25- 44.
1984:"Parsing'If'-sentences".Analysis,44, pp. 145-53.
1984a:"ConditionalInterpretationsof 'If'-sentences".Australian
Journal of Linguistics, 4, pp. 143-204.
1986:"AntecedentsandConsequents".Theoria,52, pp. 168-99.
1987:"Appiahon 'If"'.Analysis,47, pp. 74-9.
1988:"Indicativeand Subjunctive".
Analysis,48, pp. 113-22.
1989:"Vivela Revolution!".Mind,98, pp. 591-603.
1992:"ProbabilityandAssertion".Analysis,52, 4, pp. 204-11.
1994:"OnConditionals". Journalof Philosophy,91, pp. 113-28.
Dummett,Michael1959:"Truth",in Dummett,M. 1978, pp. 1-24.
1973: Frege: The Philosophy of Language. London: Duckworth.
1978: Truthand Other Enigmas. London: Duckworth.
1992: The Logical Basis of Metaphysics. London: Duckworth.
Edgington,Dorothy1986:"Do ConditionalsHaveTruthConditions?"in
Jackson,F. (ed.) 1991,pp. 176-201. Firstpublishedin Critica18,52,
1986, pp. 3-30.
1991:"TheMysteryof the MissingMatterof Fact".Proceedingsof
the Aristotelian Society Supplementary Volume 65, pp. 185-209.
Eells, E. and Skyrms, B. (eds.) 1994: Probability and Conditionals. Cam-
bridge:CambridgeUniversityPress.
Ellis, Brian,1973:"TheLogic of SubjectiveProbability".
BritishJournal
for the Philosophy of Science, 24, pp. 125-52.
1979: Rational Belief Systems. Oxford: Basil Blackwell.
1984: "Two Theories of Indicative Conditionals".Australasian
Journal of Philosophy, 62, pp. 50-66.
Fine, K. 1975: Criticalnotice of David Lewis's Counterfactuals.Mind,
84, pp. 451-58.
Firestone,Michael, 1995: "TheMeaningof 'If' A Studyof the Condi-
tional".MA Thesis,AustralianNationalUniversity.
Fowler,H. W. 1965:A Dictionaryof ModernEnglishUsage, secondedi-
tion, revisedby Sir ErnestGowers.Oxford:ClarendonPress.
Frege,G. 1960:Begriffsschriftin GeachandBlack 1960, pp. 1-20. First
publishedin 1879.
1979: Posthumous Writings. Oxford: Basil Blackwell.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
326 Dorothy Edgington

1980: Philosophical and Mathematical Correspondence. Oxford:

Basil Blackwell.
Geach, Peter and Black, Max 1960: Translationsfrom the Philosophical
Writingsof Gottlob Frege. Oxford: Basil Blackwell.
Gibbard,A. 1981: "Two Recent Theoriesof Conditionals"in Harper,
StalnakerandPearce(eds.) 1981, pp. 211-47.
Gleick,James1987:Chaos.PenguinBooks.
Goodman,N. 1947:"TheProblemof Counterfactual Conditionals".
Jour-
nal of Philosophy, 44, pp. 113-28.
1955: Fact, Fiction and Forecast. Indianapolis: Bobbs-Merrill.
Grandy,R. E. and Warner,R. eds. 1986: Philosophical Groundsof
Rationality.Oxford:ClarendonPress.
Grice, H. P. 1989: Studies in the Way of Words. Cambridge MA: Harvard
UniversityPress.
of Conditionals-Revisited". Journalof
Hajek,Alan 1989:"Probabilities
Philosophical Logic, 18, pp. 423-8.
Harper,W. L. andHooker,C. A. (eds.) 1976:Foundationsof Probability
Theory, Statistical Inference, and Statistical Theories of Science,
VolumeI. Dordrecht:Reidel.
Harper,W. L., Stalnaker,R., andPearce,C. T. (eds.) 1981:Ifs. Dordrecht:
Reidel.
Hintikka,J. and Suppes, P. (eds.) 1966: Aspects of InductiveLogic.
Amsterdam:NorthHolland.
Jackson,Frank1977: "A CausalTheoryof Counterfactuals". Australa-
sian Journal of Philosophy, 55, pp. 3-21.
1979: "On Assertionand IndicativeConditionals".Philosophical
Review, 88, pp. 565-89.
1981: "Conditionals and Possibilia". Proceedings of the Aristotelian
Society,81, pp. 125-37.
1987:Conditionals.Oxford:Basil Blackwell.
1990:"ClassifyingConditionals".Analysis50, pp. 134 47.
(ed.) 1991:Conditionals.Oxford:ClarendonPress.
Jeffrey,Richard1964:"If'. Journalof Philosophy,61, pp. 702-3.
1991: "Matterof Fact Conditionals". Proceedings of the Aristotelian
Society Supplementary Volume 65, pp. 161-83.
Kripke,Saul 1963: "SemanticalConsiderationson ModalLogic".Acta
Philosophica Fennica 16, pp. 83-94.
1972: Naming and Necessity. Oxford: Basil Blackwell.
Lance, Mark 1991: "ProbabilisticDependenceamong Conditionals"'.
Philosophical Review, 100, pp. 269-76.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 327

Laplace, Pierre Simon 1951: A Philosophical Essay on Probabilities.

Translatedby Truscott,F. W. and Emory,F. L. New York:Dover
Publications.
Oxford:Basil Blackwell.
Lewis, David 1973:Counterfactuals.
1973a: "Causation". Journal of Philosophy, 70, pp. 556-67. Page
referenccesto Lewis 1986.
1976:"Probabilitiesof ConditionalsandConditionalProbabilities".
Philosophical Review, 85, pp. 297-315. Page references to Lewis
1986.
1979: "Counterfactualsand Time's Arrow".Nous, 13, pp. 455-76.
Page referencesto Lewis 1986.
1980: "A Subjectivist's Guide to Objective Chance",in Harper,
W.L., Stalnaker,R., and Pearce,C.T. (eds) 1981, pp. 267-97. Page
referencesto Lewis 1986. Firstpublishedin Jeffrey,R. (ed.) 1980:
Studies in Inductive Logic and Probability, vol. 2, Berkley and Los
Angeles:Universityof CaliforniaPress,pp. 263-93.
1981: "Causal Decision Theory". Australasian Journal of Philoso-
phy, 59, pp. 5-30. Pagereferencesto Lewis 1986.
1986: PhilosophicalPapers Volume 2. Oxford:OxfordUniversity
Press.
1986a:"Probabilitiesof Conditionalsand ConditionalProbabilities
II". Philosophical Review, 5, pp. 581-9.
1994:"HumeanSupervenienceDebugged".Mind,103, pp. 473-90.
Lowe, E. J. 1990:"Conditionals,ContextandTransitivity".
Analysis,50,
2, pp. 80-7.
1995:"TheTruthaboutCounterfactuals". PhilosophicalQuarterly,
43,pp.41-59.
Mackie, J. 1973: Truth, Probability and Paradox. Oxford: Clarendon
Press.
McGee, Vann 1985:"A Counterexample
to ModusPonens".Journalof
Philosophy, 82, pp. 462-7 1.
1989:"ConditionalProbabilitiesandCompoundsof Conditionals".
Philosophical Review, 98, pp. 485-542.
1994:"Learningthe Impossible",in Eells, E. and Skyrms,B. (eds.)
1994, pp. 179-99.
Mellor,D. H. 1993:"Howto Believe a Conditional". Journalof Philoso-
phy, 90,5, pp. 233-48.
Menzies,Peter 1989: "ProbabilisticCausationand CausalProcesses:A
Critique of Lewis". Philosophy of Science, 56, pp. 642-63.
Pearl, Judea 1988: Probabilistic Reasoning in Intelligent Systems. San
Mateo,California:MorganKaufmann.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
328 Dorothy Edgington

Pendlebury,Michael 1989:"TheProjectionStrategyandthe TruthCon-

ditionsof ConditionalStatements".Mind,390, pp. 179-205.
Quine,W. V. 0. 1952:Methodsof Logic. London:RoutledgeandKegan
Paul.Pagereferencesto thirdedition, 1974.
1966:"Ona SupposedAntinomy",in Quine,W. V. O., TheWaysof
Paradox. New York: RandomHouse, pp. 21-3. First publishedin
Mind,62, 1953, as "Ona So-calledParadox".
Ramsey, Frank 1931: The Foundations of Mathematics. London:
RoutledgeandKeganPaul.
Read, Stephen1995: "Conditionalsand the RamseyTest".Proceedings
of the Aristotelian Society Supplementary Volume 69.
Russell, B. 1919: Introduction to Mathematical Philosophy. London:
GeorgeAllen andUnwin.
Russell, B. and Whitehead, A. N. 1962. Principia Mathematica to *56.
Cambridge:CambridgeUniversityPress.Firstpublished1910.
Ryle, Gilbert1950:"'If', 'so' and 'because"',in Black(ed.) 1950.
Sanford,"DavidH. 1989: If P, then Q: Conditionals and the Foundations
of Reasoning. London: Routledge.
Skyrms,B. 1981:"ThePriorPropensityAccountof SubjunctiveCondi-
tionals",in Harper,W. L., Stalnaker,R. and Pearce,G. (eds.) 1981,
pp. 259-65.
1994: "AdamsConditionals",in Eells, E. and Skyrms, B. (eds.)
1994, pp. 13-26.
Smiley, Timothy 1984: "Hunteron Conditionals".Proceedings of the
Aristotelian Society, 84, pp. 113-22.
Smith, Peter 1991: "The Butterfly Effect". Proceedings of the Aristotelian
Society,91, pp. 247-67.
Stalnaker,R. 1968:"ATheoryof Conditionals"in Studiesin LogicalThe-
ory, American Philosophical Quarterly Monograph Series, 2.
Oxford:Blackwell,pp. 98-112. Reprintedin Jackson,F. (ed.) 1991.
Pagereferencesto 1991.
1970:"ProbabilityandConditionals". Philosophyof Science,37, pp.
64-80. Reprintedin Harper,W. L., Stalnaker,R. andPearce,G. eds.
1981. Page referencesto 1981.
1975: "Indicative Conditionals". Philosophia, 5, pp. 269-86,
reprintedin Jackson,F. (ed.) 1991, pp. 136-54. Page referencesto
1991.
1981: "A Defense of ConditionalExcludedMiddle"in Harper,W.
L., Stalnaker,R. andPearce,G. (eds.) 1981, pp. 87-104.
1984:Inquiry.CambridgeMA: MITPress.
Stalnaker,R. andJeffrey,R. 1994:"Conditionalsas RandomVariables",
in Eells, E. andSkyrms,B. (eds.) 1994, pp. 31-46.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions
 On Conditionals 329

Strawson,P. F. 1986: "'If' and 'D"', in Grandy,R. E. and Warner,R.

1986,pp. 229-42.
Thomson, James 1990: "In Defense of D". Journal of Philosophy, 87, pp.
56-70.
Tichy, Pavel 1976:"A Counterexample to the Stalnaker-LewisAnalysis
of Counterfactuals".
PhilosophicalStudies,29, pp. 271-3.
Travis, Charles, ed. 1986: Meaning and Interpretation. Oxford: Basil
Blackwell.
Van Fraassen,Bas 1976: "Probabilitiesof Conditionals",in Harper,W.
andHooker,C. (eds.), 1976, pp. 261-308.
1980: Review of Ellis 1979. Canadian Journal of Philosophy, 10,
pp. 497-511.
Von Wright,G. H. 1957:Logical Studies.London:Routledge& Kegan
Paul.
Woods,Michael,"Conditionals".Unpublishedmanuscript.
Wright,Crispin1983: "KeepingTrackof Nozick".Analysis,43, 3, pp.
134-40.
1992: Truthand Objectivity.CambridgeMA: HarvardUniversity
Press.

This content downloaded from 128.235.251.160 on Tue, 27 Jan 2015 18:12:47 PM

All use subject to JSTOR Terms and Conditions