Professional Documents
Culture Documents
Fearon, James D. 1991. "Counterfactuals and Hypothesis Testing in Political Science." World Politics 43 (2) - 169-95
Fearon, James D. 1991. "Counterfactuals and Hypothesis Testing in Political Science." World Politics 43 (2) - 169-95
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Cambridge University Press and Trustees of Princeton University are collaborating with JSTOR to digitize,
preserve and extend access to World Politics.
http://www.jstor.org
ACTUALCASE COMPARISONS,
COUNTERFACTUALS, ANDTHE LOGIC OF
INFERENCE
If Pi is false,then E(X'e) does not equal zero. If E(X'e) does not equal
zero, Pi cannot be true(Q.E.D.).
This argumentsays thatassumingthatE(X'e) -=o in a quasi experi-
ment is equivalentto assuming the truthof a counterfactualproposition
about what would have happened if we could have altered a variable's
value forany case in the sample. Althoughone may not thinkabout the
ceteris paribus assumption in termsof a counterfactualproposition,a
counterfactualpropositionis necessarilyinvolved nonetheless.In actual
experimentsrandom assignmentguaranteesthe truthof Pi (within the
limitsof sampling variance). In quasi experiments,a causal interpreta-
tionof estimatedcoefficients requiresbeliefin thecredibilityof thecoun-
terfactualPi for justification.If we believe the resultsof a regression
analysis,we must be willing to believe that,say, if Joe Respondenthad
been a Republican as opposed to a Democrat, he would have been
roughly"so much" more likelyto have voted forReagan in i984; or that
if Israel had not moved troopsquicklyto itsnortheasternborderin Sep-
tember I970, then Syria would have been much less likelyto have been
deterredfromadvancing on the Jordaniancapital.'3
If both strategiesof confirmation
are means of solvingthe same statis-
tical problem,and if both depend in some measure on counterfactuals,
then how do theydiffer?The answer is thateach strategyprovides its
"empirical" confirmationfora causal hypothesisin a differentway.
In the actual case strategysupportfora hypothesizedcausal connec-
tion comes principallyin the formof a frequencyor magnitudeof asso-
ciation across actual cases. Of course, theory-drivenassumptions-
among them a counterfactualone are needed to supportor justifyany
regressionresult.But the resultone looksfor in regressionanalysis is an
treatmentand controlgroups.See Leland Neuberg,ConceptualAnomaliesin Economicsand
Statistics:
Lessonsfrom theSocial Experiment(Cambridge:CambridgeUniversityPress, i988).
Among otherthings,Neuberg shows thata counterfactual assumptionis needed to justify
estimatesof samplingvariancein actual experiments.I suggestbelow, however,thatcoun-
terfactualsplay a key role in quasi-experimentalhypothesistestingthattheydo not play in
actual experiments.
i The laterexamplecomes fromHuth (fn.8), 97, who oftenuses counterfactual argument
about particularcases to make more plausiblethe resultsof his regressionanalysis.
other causes are and about how theymight be related to the variables
being testedexplicitly.
Two othercontrastsbetween the counterfactualand actual case strat-
egies should be noted. The firstconcernsthe appraisal of relativecausal
weight.In the actual case strategy,such appraisals can be carried out in
several ways, essentiallyby contrastingour estimatesof the effectsof
differentindependentvariables.'9Ultimately,we can do thisbecause we
have a sample fromwhich relevantfrequenciesand magnitudescan be
extracted.In the counterfactualstrategy,by contrast,we have no con-
crete frequenciesor magnitudes,and the degrees of freedomproblem
will bite every time we introducea new variable that may have influ-
enced the particulareventto be explained.Explicitjustificationof claims
about relativeeffectswill requirea proliferationofcounterfactualcases.
Suppose, forexample, a historianor politicalscientistwishes to argue
thatbothA and B were causes ofeventE, but thatA was more important
than B. The above analysiswould suggestthatwe now need not one but
at least two counterfactualscenarios to support this claim. We would
need to contrasta counterfactualcase where A is presentbut B absent
with one where B is presentbut A absent,and theninvoke general prin-
ciples and relevantfactsto argue thatE would have been more likelyto
have occurredin the firstinstance.20
One might well object that such argumentsabout what would have
happened in multiplecounterfactualscenarioswill be veryimpreciseand
uncertain.The second contrastbetweenthe two strategiesrelatesto this
issue of "precision of estimates."In the actual case strategywhen N is
large, frequenciesand magnitudesallow the researcherto get an idea of
how much risk attachesto the beliefthatthe truecausal effectof a vari-
able is as distinctfrom the null hypothesisas the resultsshow. In the
counterfactualstrategythereis no such formalcriterionforgauging the
risk of error associated with some independentvariable. All depends
instead on the plausibilityof argumentsabout what would have hap-
pened. As will be seen in the example of the debate on the origins of
World War I, argumentsabout the relativeimportanceof possiblecauses
becomeargumentsabout the relativeplausibilityof different counterfactual
scenarios.
19There is, however,more than one meaningfulsense to the idea of causal importancein
a regressionmodel. See J.MerrillShanks,"The Importanceof Importance"(Workingpaper,
SurveyResearchCenter,Universityof California,Berkeley,i982); ChristopherAchen,In-
and UsingRegression
terpreting (BeverlyHills, Calif.: Sage Publications,i982).
20 Some philosophersof historyworkingon theproblemof how historianscan and should
attributecausal weightingshave proposed similarcriteria.See Raymond Martin,"Causes,
Conditions, and Causal Importance,"Historyand Theory21 (i982), 53-74, and citations
therein.
tatorshipin Brazil," WorldPolitics40 (October i987), 30-54; Hyug Baeg Im, "The Rise of
BureaucraticAuthoritarianismin South Korea," WorldPolitics39 (Januaryi987), 231-57;
JoanneGowa, "Hegemons, 1Os, and Markets:The Case of theSubstitutionAccount,"Inter-
nationalOrganization38 (Autumn I984), 66I-83.
24 For examples,see articlesin FredericDeyo, ed., The PoliticalEconomyoftheNew Asian
Industrialism(Ithaca, N.Y.: Cornell UniversityPress,i987).
25 As the precedingdiscussionshould suggest,an N = I case studyin which causal infer-
N = I EXAMPLES
37 Ibid., 68-69.
38 The fortunateabsence of actual cases of nuclearconflicthas led a numberof historians
and politicalscientiststo reflecton the role of counterfactuals in nuclear history.See John
Lewis Gaddis, "Nuclear Weapons and InternationalSystemicStability,"AmericanAcademy
of Arts and Sciences Occasional Paper No. 2 (Cambridge: AAAS, i990). This paper was
preparedforan AAAS workshopentitled"Nuclear Historyand theUse ofCounterfactuals."
In a different vein,RichardNed Lebow and JaniceGross Stein("Beyond Deterrence,"Jour-
nal of Social Issues43 [Winter i987], 3-7i) have brieflydiscussedthe role of counterfactuals
in defininga sample of cases of successfuldeterrence.
39 Juan Linz and AlfredStepan, eds., The Breakdownof DemocraticRegimes(Baltimore,
Md.: JohnsHopkins UniversityPress,1978).
4 Stepan, "Political Leadership and Regime Breakdown: Brazil," ibid. For otherexam-
ples,see RichardSmoke, War:Controlling Escalation(Cambridge:Harvard UniversityPress,
1977), and citationsin fn.23.
4' Stepan (fn.40), 134, and see also 120.
7"Goodman (fn.57, i983), 15-17. See also references in fn. 14. Goodman pointsout thatit
is quite problematicto use a counterfactual to definegeneraltruthconditionsforcounterfac-
tuals. See Mackie (fn. 14) fora possibleway around thisproblem(which at any rate may be
of greaterinterestto philosophersthan to politicalscientists).
I should note thatElster(fn. Io) is well aware of the issue of cotenability,
which he refers
to as "compossibility"(p. 177) and also "compatibility"(p. i83). Indeed, his "branching
worlds" theoryfor assessing the truthof counterfactuals can be seen as a suggestionfor
assessingcotenability.
7' Note the similarity of the cotenabilityconditionto PI, the key assumptionjustifyinga
causal interpretation of regressioncoefficients derived fromquasi-experimentaldata. The
likenessunderscoresthe point thatquasi experimentsand the counterfactual strategyshare
relianceon counterfactual suppositions.
CONCLUSION