Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques OxfordHandbooksOnline CaseSelectionforCase‐ StudyAnalysis:QualitativeandQuantitative Techniques JohnGerring TheOxfordHandbookofPoliticalMethodology EditedbyJanetM.Box-Steffensmeier,HenryE.Brady,andDavidCollier PrintPublicationDate: Aug2008 OnlinePublicationDate: Sep 2009 Subject: PoliticalScience,ComparativePolitics,Political Methodology DOI: 10.1093/oxfordhb/9780199286546.003.0028 AbstractandKeywords Thisarticlepresentssomeguidancebycatalogingninedifferenttechniquesforcaseselection:typical,diverse, extreme,deviant,influential,crucial,pathway,mostsimilar,andmostdifferent.Italsoindicatesthatifthe researcherisstartingfromaquantitativedatabase,thenmethodsforfindinginfluentialoutlierscanbeused.In particular,thearticleclarifiesthegeneralprinciplesthatmightguidetheprocessofcaseselectionincase-study research.Casesaremoreorlessrepresentativeofsomebroaderphenomenonand,onthatscore,maybe consideredbetterorworsesubjectsforintensiveanalysis.Thearticlethendrawsattentiontotwoambiguitiesin case-selectionstrategiesincase-studyresearch.Thefirstconcernstheadmixtureofseveralcase-selection strategies.Thesecondconcernsthechangingstatusofacaseasastudyproceeds.Somecasestudiesfollow onlyonestrategyofcaseselection. Keywords:typicalcase,diversecase,extremecase,deviantcase,influentialcase,crucialcase,pathwaycase,mostsimilarcase,mostdifferent case,case-studyanalysis CASE ‐studyanalysisfocusesononeorseveralcasesthatareexpectedtoprovideinsightintoalargerpopulation. Thispresentstheresearcherwithaformidableproblemofcaseselection:Whichcasesshouldsheorhechoose? Inlarge‐sampleresearch,thetaskofcaseselectionisusuallyhandledbysomeversionofrandomization. However,incase‐studyresearchthesampleissmall(bydefinition)andthismakesrandomsamplingproblematic, foranygivensamplemaybewildlyunrepresentative.Moreover,thereisnoguaranteethatafewcases,chosen randomly,willprovideleverageintotheresearchquestionofinterest. Inordertoisolateasampleofcasesthatbothreproducestherelevantcausalfeaturesofalargeruniverse (representativeness)andprovidesvariationalongthedimensionsoftheoreticalinterest(causalleverage),case selectionforverysmallsamplesmustemploypurposive(nonrandom)selectionprocedures.Ninesuchmethods arediscussedinthischapter,eachofwhichmaybeidentifiedwithadistinctcase‐study(p.646) “type:”typical, diverse,extreme,deviant,influential,crucial,pathway,most‐similar,andmost‐different.Table28.1summarizes eachtype,includingitsgeneraldefinition,atechniqueforlocatingitwithinapopulationofpotentialcases,itsuses, anditsprobablerepresentativeness. Whileeachofthesetechniquesisnormallypracticedononeorseveralcases(thediverse,most‐similar,andmost‐ differentmethodsrequireatleasttwo),allmayemployadditionalcases—withtheprovisothat,atsomepoint,they willnolongerofferanopportunityforin‐depthanalysisandwillthusnolongerbe“casestudies”intheusualsense (Gerring2007,ch.2).Itwillalsobeseenthatsmall‐Ncase‐selectionproceduresrest,atleastimplicitly,uponan analysisofalargerpopulationofpotentialcases(asdoesrandomization).Thecase(s)identifiedforintensive studyischosenfromapopulationandthereasonsforthischoicehingeuponthewayinwhichitissituatedwithin Page 1 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques thatpopulation.Thisistheoriginoftheterminology—typical,diverse,extreme,etal.Itfollowsthatcase‐selection proceduresincase‐studyresearchmaybuilduponpriorcross‐caseanalysisandthattheydepend,atthevery least,uponcertainassumptionsaboutthebroaderpopulation. Incertaincircumstances,thecase‐selectionproceduremaybestructuredbyaquantitativeanalysisofthelarger population.Here,severalcaveatsmustbesatisfied.First,theinferencemustpertaintomorethanafewdozen cases;otherwise,statisticalanalysisisproblematic.Second,relevantdatamustbeavailableforthatpopulation,or asignificantsampleofthatpopulation,onkeyvariables,andtheresearchermustfeelreasonablyconfidentinthe accuracyandconceptualvalidityofthesevariables.Third,allthestandardassumptionsofstatisticalresearch (e.g.identification,specification,robustness)mustbecarefullyconsidered,andwhereverpossible,tested.Ishall notdilatefurtheronthesefamiliarissuesexcepttowarntheresearcheragainsttheunreflectiveuseofstatistical techniques.1Whentheserequirementsarenotmet,theresearchermustemployaqualitativeapproachtocase selection. Thepointofthischapteristoelucidategeneralprinciplesthatmightguidetheprocessofcaseselectionincase‐ studyresearch,buildinguponearlierworkbyHarryEckstein,ArendLijphart,andothers.Sometimes,these principlescanbeappliedinaquantitativeframeworkandsometimestheyarelimitedtoaqualitativeframework.In eithercase,thelogicofcaseselectionremainsquitesimilar,whetherpracticedinsmall‐Norlarge‐Ncontexts. Beforewebegin,abitofnotationisnecessary.Inthischapter“N”referstocases,notobservations.Here,Iam concernedprimarilywithcausalinference,ratherthaninferencesthataredescriptiveorpredictiveinnature.Thus, allhypothesesinvolveatleastoneindependentvariable(X)andonedependentvariable(Y).Forconvenience,I shalllabelthecausalfactorofspecialtheoreticalinterestX1,andthecontrolvariable,orvectorofcontrols(if thereareany),X2 .Ifthewriterisconcernedtoexplainapuzzlingoutcome,buthasnopreconceptionsaboutits causes,thentheresearchwillbedescribedasY‐centered.Ifaresearcherisconcernedto(p.647) (p.648) investigatetheeffectsofaparticularcause,withnopreconceptionsaboutwhattheseeffectsmightbe,the researchwillbedescribedasX‐centered.Ifaresearcherisconcernedtoinvestigateaparticularcausal relationship,theresearchwillbedescribedasX1/Y‐centered,foritconnectsaparticularcausewithaparticular outcome.2 X‐orY‐centeredresearchisexploratory;itspurposeistogeneratenewhypotheses.X1/Y‐centered research,bycontrast,isconfirmatory/disconfirmatory;itspurposeistotestanexistinghypothesis. Table28.1.Techniquesofcaseselection 1.Typical ºDefinition:Cases(1ormore)aretypicalexamplesofsomecross‐caserelationship. ºCross‐casetechnique:Alow‐residualcase(on‐lier). ºUses:Hypothesis‐testing. ºRepresentativeness:Bydefinition,thetypicalcaseisrepresentative. 2.Diverse ºDefinition:Cases(2ormore)illuminatethefullrangeofvariationonX1,,Y,orX1/Y. ºCross‐casetechnique:Diversitymaybecalculatedby(a)categoricalvaluesofX1orY(e.g.Jewish, Catholic,Protestant),(b)standarddeviationsofX1orY(ifcontinuous),(c)combinationsofvalues(e.g.based oncross‐tabulations,factoranalysis,ordiscriminantanalysis). ºUses:Hypothesisgeneratingorhypothesistesting. ºRepresentativeness:Diversecasesarelikelytoberepresentativeintheminimalsenseofrepresentingthe fullvariationofthepopulation(thoughtheymightnotmirrorthedistributionofthatvariationinthepopulation). Page 2 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques fullvariationofthepopulation(thoughtheymightnotmirrorthedistributionofthatvariationinthepopulation). 3.Extreme ºDefinition:Cases(1ormore)exemplifyextremeorunusualvaluesofX1orYrelativetosomeunivariate distribution. ºCross‐casetechnique:AcaselyingmanystandarddeviationsawayfromthemeanofX1orY. ºUses:Hypothesis‐generating(open‐endedprobeofX1orY). ºRepresentativeness:Achievableonlyincomparisonwithalargersampleofcases. 4.Deviant ºDefinition:Cases(1ormore)deviatefromsomecross‐caserelationship. ºCross‐casetechnique:Ahigh‐residualcase(outlier). ºUses:Hypothesis‐generating(todevelopnewexplanationsforY). ºRepresentativeness:Afterthecasestudyisconducteditmaybecorroboratedbyacross‐casetest,which includesageneralhypothesis(anewvariable)basedonthecase‐studyresearch.Ifthecaseisnowanon‐ lier,itmaybeconsideredrepresentativeofthenewrelationship. 5.Influential ºDefinition:Cases(1ormore)withinfluentialconfigurationsoftheindependentvariables. ºCross‐casetechnique:HatmatrixorCook'sDistance. ºUses:Hypothesis‐testing(toverifythestatusofcasesthatmayinfluencetheresultsofacross‐case analysis). ºRepresentativeness:Notpertinent,giventhegoalsoftheinfluential‐casestudy. 6.Crucial ºDefinition:Cases(1ormore)aremostorleastlikelytoexhibitagivenoutcome. ºCross‐casetechnique:Qualitativeassessmentofrelativecrucialness. ºUses:Hypothesis‐testing(confirmatoryordisconfirmatory). ºRepresentativeness:Oftendifficulttoassess. 7.Pathway ºDefinition:Cases(1ormore)thatembodyadistinctcausalpathfromX1toY. ºCross‐casetechnique:Cross‐tab(forcategoricalvariables)orresidualanalysis(forcontinuousvariables). ºUses:Hypothesis‐testing(toprobecausalmechanisms). Page 3 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques ºRepresentativeness:Maybetestedbyexaminingresidualsforthechosencases. 8.Most‐similar ºDefinition/:Cases(2ormore)aresimilaronspecifiedvariablesotherthanX/1/and/orY. ºCross‐casetechnique:Matching. ºUses:Hypothesis‐generatingorhypothesis‐testing. ºRepresentativeness:Maybetestedbyexaminingresidualsforthechosencases. 9.Most‐different ºDefinition:Cases(2ormore)aredifferentonspecifiedvariablesotherthanX1andY. ºCross‐casetechnique:Theinverseofthemost‐similarmethodoflarge‐Ncaseselection(seeabove). ºUses:Hypothesis‐generatingorhypothesis‐testing(eliminatingdeterministiccauses). ºRepresentativeness:Maybetestedbyexaminingresidualsforthechosencases. 1TypicalCase Inorderforafocusedcasestudytoprovideinsightintoabroaderphenomenonitmustberepresentativeofa broadersetofcases.Itisinthiscontextthatonemayspeakofatypical‐caseapproachtocaseselection.The typicalcaseexemplifieswhatisconsideredtobeatypicalsetofvalues,givensomegeneralunderstandingofa phenomenon.Byconstruction,thetypicalcaseisalsoarepresentativecase. Sometypicalcasesserveanexploratoryrole.Here,theauthorchoosesacasebaseduponasetofdescriptive characteristicsandthenprobesforcausalrelationships.RobertandHelenLynd(1929/1956)selectedasinglecity “tobeasrepresentativeaspossibleofcontemporaryAmericanlife.”Specifically,theywerelookingforacitywith 1)atemperateclimate;2)asufficientlyrapidrateofgrowthtoensurethepresenceofaplentiful assortmentofthegrowingpainsaccompanyingcontemporarysocialchange;3)anindustrialculturewith modern,high‐speedmachineproduction;4)theabsenceofdominanceofthecity'sindustrybyasingle plant(i.e.,notaone‐industrytown);5)asubstantiallocalartistic(p.649) lifetobalanceitsindustrial activity…;and6)theabsenceofanyoutstandingpeculiaritiesoracutelocalproblemswhichwouldmark thecityofffromthemidchannelsortofAmericancommunity.(LyndandLynd1929/1956,quotedinYin 2004,29–30) AfterexamininganumberofoptionstheLyndsdecidedthatMuncie,Indiana,wasmorerepresentativethan,orat leastasrepresentativeas,othermidsizedcitiesinAmerica,thusqualifyingasatypicalcase. Thisisaninductiveapproachtocaseselection.Notethattypicalitymaybeunderstoodaccordingtothemean, median,ormodeonaparticulardimension;theremaybemultipledimensions(asintheforegoingexample);and eachmaybedifferentlyweighted(somedimensionsmaybemoreimportantthanothers).Wheretheselection criteriaaremultidimensionalandalargesampleofpotentialcasesisinplay,someformoffactoranalysismaybe usefulinidentifyingthemost‐typicalcase(s). However,themorecommonemploymentofthetypical‐casemethodinvolvesacausalmodelofsomephenomenon oftheoreticalinterest.Here,theresearcherhasidentifiedaparticularoutcome(Y),andperhapsaspecificX1/Y Page 4 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques hypothesis,whichshewishestoinvestigate.Inordertodoso,shelooksforatypicalexampleofthatcausal relationship.Intuitively,oneimaginesthatacaseselectedaccordingtothemeanvaluesofallparametersmustbe atypicalcaserelativetosomecausalrelationship.However,thisisbynomeansassured. SupposethattheLyndswereprimarilyinterestedinexplainingfeelingsoftrust/distrustamongmembersofdifferent socialclasses(oneoftheimplicitresearchgoalsoftheMiddletownstudy).Thisoutcomeislikelytobeaffectedby manyfactors,onlysomeofwhichareincludedintheirsixselectioncriteria.Sochoosingcaseswithrespecttoa causalhypothesisinvolves,firstofall,identifyingtherelevantparameters.Itinvolves,secondly,theselectionofa casethathasa“typical”valuerelativetotheoverallcausalmodel;itiswellexplained.Caseswithuntypical scoresonaparticulardimension(e.g.veryhighorverylow)maystillbetypicalexamplesofacausalrelationship. Indeed,theymaybemoretypicalthancaseswhosevalueslieclosetothemean.Thus,adescriptive understandingoftypicalityisquitedifferentfromacausalunderstandingoftypicality.Sinceitisthelatterversion thatismorecommon,Ishalladoptthisunderstandingoftypicalityintheremainderofthediscussion. Fromaqualitativeperspective,causaltypicalityinvolvestheselectionofacasethatconformstoexpectations aboutsomegeneralcausalrelationship.Itperformsasexpected.Inaquantitativesetting,thisnotionismeasured bythesizeofacase'sresidualinalarge‐Ncross‐casemodel.Typicalcaseslieonorneartheregressionline; theirresidualsaresmall.Insofarasthemodeliscorrectlyspecified,thesizeofacase'sresidual(i.e.thenumberof standarddeviationsthatseparatetheactualvaluefromthefittedvalue)providesahelpfulcluetohow representativethatcaseislikelytobe.“Outliers”areunlikelytoberepresentativeofthetargetpopulation. Ofcourse,justbecauseacasehasalowresidualdoesnotnecessarilymeanthatitisarepresentativecase(with respecttothecausalrelationshipofinterest).Indeed,(p.650) theissueofcaserepresentativenessisanissue thatcanneverbedefinitivelysettled.Whenonereferstoa“typicalcase”oneissaying,ineffect,thatthe probabilityofacase'srepresentativenessishigh,relativetoothercases.Thistestoftypicalityismisleadingifthe statisticalmodelismis‐specified.Anditprovideslittleinsuranceagainsterrorsthatarepurelystochastic.Acase mayliedirectlyontheregressionlinebutstillbe,insomeimportantrespect,atypical.Forexample,itmighthavean oddcombinationofvalues;theinteractionofvariablesmightbedifferentfromothercases;oradditionalcausal mechanismsmightbeatwork.Forthisreason,itisimportanttosupplementastatisticalanalysisofcaseswith evidencedrawnfromthecaseinquestion(thecasestudyitself)andwithourdeductiveknowledgeoftheworld. Oneshouldneverjudgeacasesolelybyitsresidual.Yet,allotherthingsbeingequal,acasewithalowresidualis lesslikelytobeunusualthanacasewithahighresidual,andtothisextentthemethodofcaseselectionoutlined heremaybeahelpfulguidetocase‐studyresearchersfacedwithalargenumberofpotentialcases. Bywayofconclusion,itshouldbenotedthatbecausethetypicalcaseembodiesatypicalvalueonsomesetof causallyrelevantdimensions,thevarianceofinteresttotheresearchermustliewithinthatcase.Specifically,the typicalcaseofsomephenomenonmaybehelpfulinexploringcausalmechanismsandinsolvingidentification problems(e.g.endogeneitybetweenX1andY,anomittedvariablethatmayaccountforX1andY,orsomeother spuriouscausalassociation).Dependingupontheresultsofthecasestudy,theauthormayconfirmanexisting hypothesis,disconfirmthathypothesis,orreframeitinawaythatisconsistentwiththefindingsofthecasestudy. Thesearetheusesofthetypical‐casestudy. 2DiverseCases Asecondcase‐selectionstrategyhasasitsprimaryobjectivetheachievementofmaximumvariancealong relevantdimensions.Irefertothisasadiverse‐casemethod.Forobviousreasons,thismethodrequiresthe selectionofasetofcases—atminimum,two—whichareintendedtorepresentthefullrangeofvalues characterizingX1,Y,orsomeparticularX1/Yrelationship.3 Wheretheindividualvariableofinterestiscategorical(on/off,red/black/blue,Jewish/Protestant/Catholic),the identificationofdiversityisreadilyapparent.Theinvestigatorsimplychoosesonecasefromeachcategory.Fora continuousvariable,(p.651) thechoicesarenotsoobvious.However,theresearcherusuallychoosesboth extremevalues(highandlow),andperhapsthemeanormedianaswell.Theresearchermayalsolookforbreak‐ pointsinthedistributionthatseemtocorrespondtocategoricaldifferencesamongcases.Orshemayfollowa theoreticalhunchaboutwhichthresholdvaluescount,i.e.whicharelikelytoproducedifferentvaluesonY. Page 5 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Anothersortofdiversecasetakesaccountofthevaluesofmultiplevariables(i.e.avector),ratherthanasingle variable.Ifthesevariablesarecategorical,theidentificationofcausaltypesrestsupontheintersectionofeach category.Twodichotomousvariablesproduceamatrixwithfourcells.Threetrichotomousvariablesproducea matrixofeightcells.Andsoforth.Ifallvariablesaredeemedrelevanttotheanalysis,theselectionofdiverse casesmandatestheselectionofonecasedrawnfromwithineachcell.Letussaythatanoutcomeisthoughttobe affectedbysex,race(black/white),andmaritalstatus.Here,adiverse‐casestrategyofcaseselectionwould identifyonecasewithineachoftheseintersectingcells—atotalofeightcases.Thingsbecomeslightlymore complicatedwhenoneormoreofthefactorsiscontinuous,ratherthancategorical.Here,thediversityofcase valuesdonotfallneatlyintocells.Rather,thesecellsmustbecreatedbyfiat—e.g.high,medium,low. Itwillbeseenthatwheremultiplevariablesareunderconsideration,thelogicofdiverse‐caseanalysisrestsupon thelogicoftypologicaltheorizing—wheredifferentcombinationsofvariablesareassumedtohaveeffectsonan outcomethatvaryacrosstypes(Elman2005;GeorgeandBennett2005,235;LazarsfeldandBarton1951).George andSmoke,forexample,wishtoexploredifferenttypesofdeterrencefailure—by“faitaccompli,”by“limited probe,”andby“controlledpressure.”Consequently,theywishtofindcasesthatexemplifyeachtypeofcausal mechanism.4 DiversitymaythusrefertoarangeofvariationonXorY,ortoaparticularcombinationofcausalfactors(withor withoutaconsiderationoftheoutcome).Ineachinstance,thegoalofcaseselectionistocapturethefullrangeof variationalongthedimension(s)ofinterest. Sincediversitycanmeanmanythings,itsemploymentinalarge‐Nsettingisnecessarilydependentuponhowthis keytermisdefined.Ifitisunderstoodtopertainonlytoasinglevariable(X1orY),thenthetaskisfairlysimple.A categoricalvariablemandatesthechoiceofatleastonecasefromeachcategory—twoifdichotomous,threeif trichotomous,andsoforth.Acontinuousvariablesuggeststhechoiceofatleastone“high”and“low”value,and perhapsonedrawnfromthemeanormedian.Butotherchoicesmightalsobejustified,accordingtoone'shunch abouttheunderlyingcausalrelationshiporaccordingtonaturalthresholdsfoundinthedata,whichmaybe groupedintodiscretecategories.Single‐variabletraitsareusuallyeasytodiscoverinalarge‐Nsettingthrough descriptivestatisticsorthroughvisualinspectionofthedata. (p.652) Wherediversityreferstoparticularcombinationsofvariables,therelevantcross‐casetechniqueis someversionofstratifiedrandomsampling(inaprobabilisticsetting)orQualitativeComparativeAnalysis(ina deterministicsetting)(Ragin2000).Iftheresearchersuspectsthatacausalrelationshipisaffectednotonlyby combinationsoffactorsbutalsobytheirsequencing,thenthetechniqueofanalysismustincorporatetemporal elements(Abbott2001;AbbottandForrest1986;AbbottandTsay2000).Thus,themethodofidentifyingcausal typesrestsuponwhatevermethodofidentifyingcausalrelationshipsisemployedinthelarge‐Nsample. Notethattheidentificationofdistinctcasetypesisintendedtoidentifygroupsofcasesthatareinternally homogeneous(inallrespectsthatmightaffectthecausalrelationshipofinterest).Thus,thechoiceofcaseswithin eachgroupshouldnotbeproblematic,andmaybeaccomplishedthroughrandomsamplingorpurposivecase selection.However,ifthereissuspecteddiversitywithineachcategory,thenmeasuresshouldbetakentoassure thatthechosencasesaretypicalofeachcategory.Acasestudyshouldnotfocusonanatypicalmemberofa subgroup. Indeed,considerationsofdiversityandtypicalityoftengotogether.Thus,inastudyofglobalizationandsocial welfaresystems,DuaneSwank(2002)firstidentifiesthreedistinctivegroupsofwelfarestates:“universalistic” (socialdemocratic),“corporatistconservative,”and“liberal.”Next,helookswithineachgrouptofindthemost‐ typicalcases.HedecidesthattheNordiccountriesaremoretypicaloftheuniversalisticmodelthanthe Netherlandssincethelatterhas“somecharacteristicsoftheoccupationallybasedprogramstructureanda politicalcontextofChristianDemocratic‐ledgovernmentstypicalofthecorporatistconservativenations”(Swank 2002,11;seealsoEsping‐Andersen1990).Thus,theNordiccountriesarechosenasrepresentativecaseswithin theuniversalisticcasetype,andareaccompaniedinthecase‐studyportionofhisanalysisbyothercaseschosen torepresenttheotherwelfarestatetypes(corporatistconservativeandliberal). Evidently,whenasampleencompassesafullrangeofvariationonrelevantparametersoneislikelytoenhance therepresentativenessofthatsample(relativetosomepopulation).Thisisadistinctadvantage.Ofcourse,the inclusionofafullrangeofvariationmaydistorttheactualdistributionofcasesacrossthisspectrum.Ifthereare Page 6 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques more“high”casesthan“low”casesinapopulationandtheresearcherchoosesonlyonehighcaseandonelow case,theresultingsampleoftwoisnotperfectlyrepresentative.Evenso,thediverse‐casemethodprobablyhas strongerclaimstorepresentativenessthananyothersmall‐Nsample(includingthestandalonetypicalcase).The selectionofdiversecaseshastheadditionaladvantageofintroducingvariationonthekeyvariablesofinterest.A setofdiversecasesis,bydefinition,asetofcasesthatencompassesarangeofhighandlowvaluesonrelevant dimensions.Thereis,therefore,muchtorecommendthismethodofcaseselection.Isuspectthatthese advantagesarecommonlyunderstoodandareappliedonanintuitivelevelbycase‐studyresearchers.However, thelackofarecognizablename—andanexplicitmethodologicaldefense—hasmadeitdifficultforcase‐study researcherstoutilizethismethodofcaseselection,andtodosoinanexplicitandself‐consciousfashion. Neologismhasitsuses. (p.653) 3ExtremeCase Theextreme‐casemethodselectsacasebecauseofitsextremevalueonanindependent(X1)ordependent(Y) variableofinterest.Thus,studiesofdomesticviolencemaychoosetofocusonextremeinstancesofabuse (Browne1987).Studiesofaltruismmayfocusonthoserareindividualswhoriskedtheirlivestohelpothers(e.g. Holocaustresisters)(Monroe1996).Studiesofethnicpoliticsmayfocusonthemostheterogeneoussocieties(e.g. PapuaNewGuinea)inordertobetterunderstandtheroleofethnicityinademocraticsetting(Reilly2000–1). Studiesofindustrialpolicyoftenfocusonthemostsuccessfulcountries(i.e.theNICS)(Deyo1987).Andsoforth.5 Oftenanextremecasecorrespondstoacasethatisconsideredtobeprototypicalorparadigmaticofsome phenomenaofinterest.Thisisbecauseconceptsareoftendefinedbytheirextremes,i.e.theiridealtypes.Italian FascismdefinestheconceptofFascism,inpart,becauseitofferedthemostextremeexampleofthatphenomenon. However,themethodologicalvalueofthiscase,andotherslikeit,derivesfromitsextremity(alongsome dimensionofinterest),notitstheoreticalstatusoritsstatusintheliteratureonasubject. Thenotionof“extreme”maynowbedefinedmoreprecisely.Anextremevalueisanobservationthatliesfaraway fromthemeanofagivendistribution.Thismaybemeasured(iftherearesufficientobservations)byacase's“Z score”—thenumberofstandarddeviationsbetweenacaseandthemeanvalueforthatsample.Extremecases havehighZscores,andforthisreasonmayserveasusefulsubjectsforintensiveanalysis. Foracontinuousvariable,thedistancefromthemeanmaybeineitherdirection(positiveornegative).Fora dichotomousvariable(present/absent),extremenessmaybeinterpretedasunusual.Ifmostcasesarepositive alongagivendimension,thenanegativecaseconstitutesanextremecase.Ifmostcasesarenegative,thena positivecaseconstitutesanextremecase.Itshouldbeclearthatresearchersarenotsimplyconcernedwith caseswheresomething“happened,”butalsowithcaseswheresomethingdidnot.Itistherarenessofthevalue thatmakesacasevaluable,inthiscontext,notitspositiveornegativevalue.6 Thus,ifoneisstudyingstate capacity,acaseofstatefailureisprobablymoreinformativethanacaseofstateendurancesimplybecausethe formerismoreunusual.Similarly,ifoneisinterestedinincesttaboosaculturewheretheincesttabooisabsentor weakisprobablymoreusefulthanaculturewhereitispresentorstrong.Fascismismoreimportantthan nonfascism.Andsoforth.Thereisagoodreason,therefore,whycasestudiesofrevolutiontendtofocuson “revolutionary”cases.ThedaSkocpol(1979)hadmuchmoretolearnfromFrancethanfromAustro‐Hungarysince FrancewasmoreunusualthanAustro‐HungarywithinthepopulationofnationstatesthatSkocpolwas(p.654) concernedtoexplain.Thereasonisquitesimple:Therearefewerrevolutionarycasesthannonrevolutionary cases;thus,thevariationthatweexploreasacluetocausalrelationshipsisencapsulatedinthesecases,against abackgroundofnonrevolutionarycases. Notethattheextreme‐casemethodofcaseselectionappearstoviolatethesocialsciencefolkwisdomwarningus notto“selectonthedependentvariable.”7 Selectingcasesonthedependentvariableisindeedproblematicifa numberofcasesarechosen,allofwhichlieononeendofavariable'sspectrum(theyareallpositiveornegative), andiftheresearcherthensubjectsthissampletocross‐caseanalysisasifitwererepresentativeofapopulation.8 Resultsforthissortofanalysiswouldalmostassuredlybebiased.Moreover,therewillbelittlevariationtoexplain sincethevaluesofeachcaseareexplicitlyconstrained. However,thisisnottheproperemploymentoftheextreme‐casemethod.(Itismoreappropriatelylabeledan extreme‐samplemethod.)Theextreme‐casemethodactuallyrefersbacktoalargersampleofcasesthatlieinthe Page 7 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques backgroundoftheanalysisandprovideafullrangeofvariationaswellasamorerepresentativepictureofthe population.Itisaself‐consciousattempttomaximizevarianceonthedimensionofinterest,nottominimizeit.If thispopulationofcasesiswellunderstood—eitherthroughtheauthor'sowncross‐caseanalysis,throughthework ofothers,orthroughcommonsense—thenaresearchermayjustifytheselectionofasinglecaseexemplifyingan extremevalueforwithin‐caseanalysis.Ifnot,theresearchermaybewelladvisedtofollowadiverse‐casemethod, asdiscussedabove. Bywayofconclusion,letusreturntotheproblemofrepresentativeness.Itwillbeseenthatanextremecasemay betypicalordeviant.ThereissimplynowaytotellbecausetheresearcherhasnotyetspecifiedanX1/Ycausal proposition.Oncesuchacausalpropositionhasbeenspecifiedonemaythenaskwhetherthecaseinquestionis similartosomepopulationofcasesinallrespectsthatmightaffecttheX1/Yrelationshipofinterest(i.e.unit homogeneous).Itisatthispointthatitbecomespossibletosay,withinthecontextofacross‐casestatistical model,whetheracaseliesnearto,orfarfrom,theregressionline.However,thissortofanalysismeansthatthe researcherisnolongerpursuinganextreme‐casemethod.Theextreme‐casemethodispurelyexploratory—a wayofprobingpossiblecausesofY,orpossibleeffectsofX,inanopen‐endedfashion.Iftheresearcherhassome notionofwhatadditionalfactorsmightaffecttheoutcomeofinterest,orofwhatrelationshipthecausalfactorof interestmighthavewithY,thensheoughttopursueoneoftheothermethodsexploredinthischapter.Thisalso impliesthatanextreme‐casemethodmaytransformintoadifferentkindofapproachasastudyevolves;thatis,as amorespecifichypothesiscomestolight.Usefulextremecasesattheoutsetofastudymayprovelessusefulata laterstageofanalysis. (p.655) 4DeviantCase Thedeviant‐casemethodselectsthatcase(s)which,byreferencetosomegeneralunderstandingofatopic (eitheraspecifictheoryorcommonsense),demonstratesasurprisingvalue.Itisthusthecontraryofthetypical case.BarbaraGeddes(2003)notestheimportanceofdeviantcasesinmedicalscience,whereresearchersare habituallyfocusedonthatwhichis“pathological”(accordingtostandardtheoryandpractice).TheNewEngland JournalofMedicine,oneofthepremierjournalsofthefield,carriesaregularfeatureentitledCaseRecordsofthe MassachusettsGeneralHospital.Thesearticlesbeartitleslikethefollowing:“An80‐Year‐OldWomanwithSudden UnilateralBlindness”or“A76‐Year‐OldManwithFever,Dyspnea,PulmonaryInfiltrates,PleuralEffusions,and Confusion.”9 Anotherinterestingexampledrawnfromthefieldofmedicineconcernstheextensivestudynow devotedtoasmallnumberofpersonswhoseemresistanttotheAIDSvirus(BuchbinderandVittinghoff1999; Haynes,Pantaleo,andFauci1996).Whyaretheyresistant?Whatisdifferentaboutthesepeople?Whatcanwe learnaboutAIDSinotherpatientsbyobservingpeoplewhohavebuilt‐inresistancetothisdisease? Likewise,inpsychologyandsociologycasestudiesmaybecomprisedofdeviant(inthesocialsense)personsor groups.Ineconomics,casestudiesmayconsistofcountriesorbusinessesthatoverperform(e.g.Botswana; Microsoft)orunderperform(e.g.Britainthroughmostofthetwentiethcentury;Searsinrecentdecades)relativeto somesetofexpectations.Inpoliticalscience,casestudiesmayfocusoncountrieswherethewelfarestateismore developed(e.g.Sweden)orlessdeveloped(e.g.theUnitedStates)thanonewouldexpect,givenasetofgeneral expectationsaboutwelfarestatedevelopment.Thedeviantcaseiscloselylinkedtotheinvestigationoftheoretical anomalies.Indeed,tosaydeviantistoimply“anomalous.”10 Notethatwhileextremecasesarejudgedrelativetothemeanofasingledistribution(thedistributionofvalues alongasinglevariable),deviantcasesarejudgedrelativetosomegeneralmodelofcausalrelations.Thedeviant‐ casemethodselectscaseswhich,byreferencetosome(presumably)generalrelationship,demonstratea surprisingvalue.Theyare“deviant”inthattheyarepoorlyexplainedbythemultivariatemodel.Theimportant pointisthatdeviant‐nesscanonlybeassessedrelativetothegeneral(quantitativeorqualitative)model.This meansthattherelativedeviant‐nessofacaseislikelytochangewheneverthegeneralmodelisaltered.For example,theUnitedStatesisadeviantwelfarestatewhenthisoutcomeisgaugedrelativetosocietalwealth.Butit islessdeviant—andperhapsnotdeviantatall—whencertainadditional(politicalandsocietal)factorsareincluded inthemodel,asdiscussedin(p.656) theepilogue.Devianceismodeldependent.Thus,whendiscussingthe conceptofthedeviantcaseitishelpfultoaskthefollowingquestion:Relativetowhatgeneralmodel(orsetof backgroundfactors)isCaseAdeviant? Page 8 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Conceptually,wehavesaidthatthedeviantcaseisthelogicalcontraryofthetypicalcase.Thistranslatesintoa directlycontrastingstatisticalmeasurement.Whilethetypicalcaseisonewithalowresidual(insomegeneral modelofcausalrelations),adeviantcaseisonewithahighresidual.Thismeans,followingourprevious discussion,thatthedeviantcaseislikelytobeanunrepresentativecase,andinthisrespectappearstoviolatethe suppositionthatcase‐studysamplesshouldseektoreproducefeaturesofalargerpopulation. However,itmustbeborneinmindthattheprimarypurposeofadeviant‐caseanalysisistoprobefornew—butas yetunspecified—explanations.(IfthepurposeistodisproveanextanttheoryIshallrefertothestudyascrucial‐ case,asdiscussedbelow.)Theresearcherhopesthatcausalprocessesidentifiedwithinthedeviantcasewill illustratesomecausalfactorthatisapplicabletoother(moreorlessdeviant)cases.Thismeansthatadeviant‐ casestudyusuallyculminatesinageneralproposition,onethatmaybeappliedtoothercasesinthepopulation. Oncethisgeneralpropositionhasbeenintroducedintotheoverallmodel,theexpectationisthatthechosencase willnolongerbeanoutlier.Indeed,thehopeisthatitwillnowbetypical,asjudgedbyitssmallresidualinthe adjustedmodel.(Theexceptionwouldbeacircumstanceinwhichacase'soutcomeisdeemedtobe“accidental,” andthereforeinexplicablebyanygeneralmodel.) Thisfeatureofthedeviant‐casestudyshouldhelptoresolvequestionsaboutitsrepresentativeness.Evenifitis notpossibletomeasurethenewcausalfactor(andthustointroduceitintoalarge‐Ncross‐casemodel),itmaystill beplausibletoassert(basedongeneralknowledgeofthephenomenon)thatthechosencaseisrepresentativeof abroaderpopulation. 5InfluentialCase Sometimes,thechoiceofacaseismotivatedsolelybytheneedtoverifytheassumptionsbehindageneralmodel ofcausalrelations.Here,theanalystattemptstoprovidearationalefordisregardingaproblematiccaseorasetof problematiccases.Thatistosay,sheattemptstoshowwhyapparentdeviationsfromthenormarenotreally deviant,ordonotchallengethecoreofthetheory,oncethecircumstancesofthespecialcaseorcasesarefully understood.Across‐caseanalysismay,afterall,bemarredbyseveralclassesofproblemsincluding measurementerror,specificationerror,errorsinestablishingproperboundariesfortheinference(thescopeofthe argument),andstochasticerror(fluctuationsinthephenomenonunderstudythataretreatedasrandom,given availabletheoreticalresources).Ifpoorlyfittingcases(p.657) canbeexplainedawaybyreferencetothese kindsofproblems,thenthetheoryofinterestisthatmuchstronger.Thissortofdeviant‐caseanalysisanswersthe question,“WhataboutCaseA(orcasesoftypeA)?Howdoesthat,seeminglydisconfirming,casefitthemodel?” Becauseitsunderlyingpurposeisdifferentfromtheusualdeviant‐casestudy,Iofferanewtermforthismethod. Theinfluentialcaseisacasethatcastsdoubtuponatheory,andforthatreasonwarrantscloseinspection.This investigationmayreveal,afterall,thatthetheoryisvalidated—perhapsinsomeslightlyalteredform.Inthisguise, theinfluentialcaseisthe“casethatprovestherule.”Inotherinstances,theinfluential‐caseanalysismay contributetodisconfirming,orreconceptualizing,atheory.Thekeypointisthatthevalueofthecaseisjudged relativetosomeextantcross‐casemodel. Asimpleversionofinfluential‐caseanalysisinvolvestheconfirmationofakeycase'sscoreonsomecritical dimension.Thisisessentiallyaquestionofmeasurement.Sometimescasesarepoorlyexplainedsimplybecause theyarepoorlyunderstood.Acloseexaminationofaparticularcontextmayrevealthatanapparentlyfalsifying casehasbeenmiscoded.Ifso,theinitialchallengepresentedbythatcasetosomegeneraltheoryhasbeen obviated. However,themoreusualemploymentoftheinfluential‐casemethodculminatesinasubstantivereinterpretationof thecase—perhapsevenofthegeneralmodel.Itisnotjustaquestionofmeasurement.ConsiderThomasErtman's (1997)studyofstatebuildinginWesternEurope,assummarizedbyGerardoMunck.Thisstudyargues thattheinteractionofa)thetypeoflocalgovernmentduringthefirstperiodofstatebuilding,withb)the timingofincreasesingeopoliticalcompetition,stronglyinfluencesthekindofregimeandstatethat emerge.[Ertman]teststhishypothesisagainstthehistoricalexperienceofEuropeandfindsthatmost countriesfithispredictions.Denmark,however,isamajorexception.InDenmark,sustainedgeopolitical competitionbeganrelativelylateandlocalgovernmentatthebeginningofthestatebuildingperiodwas Page 9 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques generallyparticipatory,whichshouldhaveledthecountrytodevelop“patrimonialconstitutionalism.”But infact,itdeveloped“bureaucraticabsolutism.”Ertmancarefullyexplorestheprocessthroughwhich DenmarkcametohaveabureaucraticabsolutiststateandfindsthatDenmarkhadtheearlymarksofa patrimonialconstitutionaliststate.However,thecountrywaspushedoffthisdevelopmentalpathbythe influenceofGermanknights,whoenteredDenmarkandbroughtwiththemGermaninstitutionsoflocal government.Ertmanthentracesthecausalprocessthroughwhichtheseimportedinstitutionspushed Denmarktodevelopbureaucraticabsolutism,concludingthatthisdevelopmentwascausedbyafactor welloutsidehisexplanatoryframework. (Munck2004,118) Ertman'soverallframeworkisconfirmedinsofarashehasbeenabletoshow,byanin‐depthdiscussionof Denmark,thatthecausalprocessesstipulatedbythegeneraltheoryholdeveninthisapparentlydisconfirming case.Denmarkisstilldeviant,butitissobecauseof“contingenthistoricalcircumstances”thatareexogenousto thetheory(Ertman1997,316). Evidently,theinfluential‐caseanalysisissimilartothedeviant‐caseanalysis.Bothfocusonoutliers.However,as weshallsee,theyfocusondifferentkindsofoutliers.(p.658) Moreover,theanimatinggoalsofthesetwo researchdesignsarequitedifferent.Theinfluential‐casestudybeginswiththeaimofconfirmingageneralmodel, whilethedeviant‐casestudyhastheaimofgeneratinganewhypothesisthatmodifiesanexistinggeneralmodel. Theconfusionstemsfromthefactthatthesamecasestudymayfulfillbothobjectives—qualifyingageneralmodel and,atthesametime,confirmingitscorehypothesis. Thus,intheirstudyofRobertoMichels's“ironlawofoligarchy,”Lipset,Trow,andColeman(1956)choosetofocus onanorganization—theInternationalTypographicalUnion—thatappearstoviolatethecentralpresupposition.The ITU,asnotedbyoneoftheauthors,has“along‐termtwo‐partysystemwithfreeelectionsandfrequentturnoverin office”andisthusanythingbutoligarchic(Lipset1959,70).Assuch,itcallsintoquestionMichels'sgrand generalizationaboutorganizationalbehavior.Theauthorsexplainthiscuriousresultbytheextraordinarilyhigh levelofeducationamongthemembersofthisunion.Michels'slawisshowntobetrueformostorganizations,but notall.Itistrue,withqualifications.Notethattherespecificationoftheoriginalmodel(ineffect,Lipset,Trow,and Colemanintroduceanewcontrolvariableorboundarycondition)involvestheexplorationofanewhypothesis.In thisinstance,therefore,theuseofaninfluentialcasetoconfirmanexistingtheoryisquitesimilartotheuseofa deviantcasetoexploreanewtheory. Inaquantitativeidiom,influentialcasesarethosethat,ifcounterfactuallyassignedadifferentvalueonthe dependentvariable,wouldmostsubstantiallychangetheresultingestimates.Theymayormaynotbeoutliers (high‐residualcases).Twoquantitativemeasuresofinfluencearecommonlyappliedinregressiondiagnostics (Belsey,Kuh,andWelsch2004).Thefirst,oftenreferredtoastheleverageofacase,derivesfromwhatiscalled thehatmatrix.Basedsolelyoneachcase'sscoresontheindependentvariables,thehatmatrixtellsushowmuch achangein(orameasurementerroron)thedependentvariableforthatcasewouldaffecttheoverallregression line.ThesecondisCook'sdistance,ameasureoftheextenttowhichtheestimatesofalltheparameterswould changeifagivencasewereomittedfromtheanalysis.CaseswithalargeleverageorCook'sdistancecontribute quitealottotheinferencesdrawnfromacross‐caseanalysis.Inthissense,suchcasesarevitalformaintaining analyticconclusions.Discoveringasignificantmeasurementerroronthedependentvariableoranimportant omittedvariableforsuchacasemaydramaticallyreviseestimatesoftheoverallrelationships.Hence,itmaybe quitesensibletoselectinfluentialcasesforin‐depthstudy. Notethattheuseofaninfluential‐casestrategyofcaseselectionislimitedtoinstancesinwhicharesearcherhas reasontobeconcernedthatherresultsarebeingdrivenbyoneorafewcases.Thisismostlikelytobetruein smalltomoderate‐sizedsamples.WhereNisverylarge—greaterthan1,000,letussay—itisextremelyunlikely thatasmallsetofcases(muchlessanindividualcase)willplayan“influential”role.Ofcourse,theremaybe influentialsetsofcases,e.g.countrieswithinaparticularcontinentorculturalregion,orpersonsofIrishextraction. Setsofinfluentialobservationsareoftenproblematicinatime‐seriescross‐sectiondata‐setwhereeach(p.659) unit(e.g.country)containsmultipleobservations(throughtime),andhencemayhaveastronginfluenceon aggregateresults.Still,thegeneralruleis:thelargerthesample,thelessimportantindividualcasesarelikelytobe and,hence,thelesslikelyaresearcheristouseaninfluential‐caseapproachtocaseselection. Page 10 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques 6CrucialCase Ofalltheextantmethodsofcaseselectionperhapsthemoststoried—andcertainlythemostcontroversial—isthe crucial‐casemethod,introducedtothesocialscienceworldseveraldecadesagobyHarryEckstein.Inhisseminal essay,Eckstein(1975,118)describesthecrucialcaseasone“thatmustcloselyfitatheoryifoneistohave confidenceinthetheory'svalidity,or,conversely,mustnotfitequallywellanyrulecontrarytothatproposed.”A caseiscrucialinasomewhatweaker—butmuchmorecommon—sensewhenitismost,orleast,likelytofulfilla theoreticalprediction.A“most‐likely”caseisonethat,onalldimensionsexceptthedimensionoftheoretical interest,ispredictedtoachieveacertainoutcome,andyetdoesnot.Itisthereforeusedtodisconfirmatheory.A “least‐likely”caseisonethat,onalldimensionsexceptthedimensionoftheoreticalinterest,ispredictednotto achieveacertainoutcome,andyetdoesso.Itisthereforeusedtoconfirmatheory.Inallformulations,thecrucial‐ caseoffersamost‐difficulttestforanargument,andhenceprovideswhatisperhapsthestrongestsortof evidencepossibleinanonexperimental,single‐casesetting. SincethepublicationofEckstein'sinfluentialessay,thecrucial‐caseapproachhasbeenclaimedinamultitudeof studiesacrossseveralsocialsciencedisciplinesandhascometoberecognizedasastapleofthecase‐study method.11Yettheideaofanysinglecaseplayingacrucial(or“critical”)roleisnotwidelyacceptedamongmost methodologists(e.g.Sekhon2004).(Evenitsprogenitorseemstohavehaddoubts.) Letusbeginwiththeconfirmatory(a.k.a.least‐likely)crucialcase.Theimplicitlogicofthisresearchdesignmaybe summarizedasfollows.Givenasetoffacts,weareaskedtocontemplatetheprobabilitythatagiventheoryistrue. Whilethefactsmatter,tobesure,theeffectivenessofthissortofresearchalsorestsupontheformalpropertiesof thetheoryinquestion.Specifically,thedegreetowhichatheoryisamenabletoconfirmationiscontingentupon howmanypredictionscanbederivedfromthetheoryandonhow“risky”eachindividualpredictionis.InPopper's (1963,36)words,“Confirmationsshouldcountonlyiftheyaretheresultofriskypredictions;thatistosay,if, unenlightenedbythetheoryinquestion,weshouldhaveexpectedan(p.660) eventwhichwasincompatible withthetheory—andeventwhichwouldhaverefutedthetheory.Every‘good’scientifictheoryisaprohibition;it forbidscertainthingstohappen.Themoreatheoryforbids,thebetteritis”(seealsoPopper1934/1968).Arisky predictionisthereforeonethatishighlypreciseanddeterminate,andthereforeunlikelytobeachievedbythe productofothercausalfactors(externaltothetheoryofinterest)orthroughstochasticprocesses.Atheory producesmanysuchpredictionsifitisfullyelaborated,issuingpredictionsnotonlyonthecentraloutcomeof interestbutalsoonspecificcausalmechanisms,andifitisbroadinpurview.(Thenotionofriskinessmayalsobe conceptualizedwithinthePopperianlexiconasdegreesoffalsifiability.) ThesepointscanalsobearticulatedinBayesianterms.ColinHowsonandPeterUrbachexplain:“Thedegreeto whichh[ahypothesis]isconfirmedbye[asetofevidence]depends…ontheextenttowhichP(eČh)exceedsP (e),thatis,onhowmuchmoreprobableeisrelativetothehypothesisandbackgroundassumptionsthanitis relativejusttobackgroundassumptions.”Again,“confirmationiscorrelatedwithhowmuchmoreprobablethe evidenceisifthehypothesisistruethanifitisfalse”(HowsonandUrlbach1989,86).Thus,thestrangerthe predictionofferedbyatheory—relativetowhatwewouldnormallyexpect—thegreaterthedegreeofconfirmation thatwillbeaffordedbytheevidence.Asanintuitiveexample,HowsonandUrbach(1989,86)offerthefollowing: Ifasoothsayerpredictsthatyouwillmeetadarkstrangersometimeandyoudoinfact,yourfaithinhis powersofprecognitionwouldnotbemuchenhanced:youwouldprobablycontinuetothinkhispredictions werejusttheresultofguesswork.However,ifthepredictionalsogavethecorrectnumberofhairsonthe headofthatstranger,yourpreviousscepticismwouldnodoubtbeseverelyshaken. WhilethesePopperian/Bayesiannotions12 arerelevanttoallempiricalresearchdesigns,theyareespecially relevanttocase‐studyresearchdesigns,forinthesesettingsasinglecase(or,atmost,asmallnumberofcases) isrequiredtobearaheavyburdenofproof.Itshouldbenosurprise,therefore,thatPopper'sideaof“riskiness” wastobeappropriatedbycase‐studyresearcherslikeHarryEcksteintovalidatetheenterpriseofsingle‐case analysis.(AlthoughEcksteindoesnotcitePoppertheintellectuallineageisclear.)Riskiness,here,isanalogousto whatisusuallyreferredtoasa“most‐difficult”researchdesign,whichinacase‐studyresearchdesignwouldbe understoodasa“least‐likely”case.Notealsothatthedistinctionbetweena“must‐fit”caseandaleast‐likelycase —that,intheevent,actuallydoesfitthetermsofatheory—isamatterofdegree.Casesaremoreorlesscrucialfor confirmingtheories.Thepointisthat,insomecircumstances,apaucityofempiricalevidencemaybe Page 11 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques compensatedbytheriskinessofthetheory. Thecrucial‐caseresearchdesignis,perforce,ahighlydeductiveenterprise;muchdependsonthequalityofthe theoryunderinvestigation.Itfollowsthatthetheoriesmostamenabletocrucial‐caseanalysisarethosewhichare lawlikeintheirprecision,(p.661) degreeofelaboration,consistency,andscope.Themoreatheoryattainsthe statusofacausallaw,theeasieritwillbetoconfirm,ortodisconfirm,withasinglecase.Indeed,riskypredictions arecommoninnaturalsciencefieldssuchasphysics,whichinturnservedasthetemplateforthedeductive‐ nomological(“covering‐law”)modelofsciencethatinfluencedEcksteinandothersinthepostwardecades(e.g. Hempel1942). Afrequentlycitedexampleisthefirstimportantempiricaldemonstrationofthetheoryofrelativity,whichtookthe formofasingle‐eventpredictionontheoccasionoftheMay29,1919,solareclipse(Eckstein1975;Popper1963). StephenVanEvera(1997,66–7)describestheimpactofthispredictiononthevalidationofEinstein'stheory. Einstein'stheorypredictedthatgravitywouldbendthepathoflighttowardagravitysourcebyaspecific amount.Henceitpredictedthatduringasolareclipsestarsnearthesunwouldappeardisplaced—stars actuallybehindthesunwouldappearnexttoit,andstarslyingnexttothesunwouldappearfartherfromit —anditpredictedtheamountofapparentdisplacement.Noothertheorymadethesepredictions.The passageofthisonesingle‐case‐studytestbroughtthetheorywideacceptancebecausethetested predictionswereunique—therewasnoplausiblecompetingexplanationforthepredictedresult—hence thepassedtestwasverystrong. Thestrengthofthistestistheextraordinaryfitbetweenthetheoryandasetoffactsfoundinasinglecase,and thecorrespondinglackoffitbetweenallothertheoriesandthissetoffacts.Einsteinofferedanexplanationofa particularsetofanomalousfindingsthatnootherexistingtheorycouldmakesenseof.Ofcourse,onemust assumethattherewasno—orlimited—measurementerror.Andonemustassumethatthephenomenonofinterest islargelyinvariant;lightdoesnotbenddifferentlyatdifferenttimesandplaces(exceptinwaysthatcanbe understoodthroughthetheoryofrelativity).Andonemustassume,finally,thatthetheoryitselfmakessenseon othergrounds(otherthanthecaseofspecialinterest);itisaplausiblegeneraltheory.Ifoneiswillingtoaccept theseaprioriassumptions,thenthe1919“casestudy”providesaverystrongconfirmationofthetheory.Itis difficulttoimagineastrongerproofofthetheoryfromwithinanobservational(nonexperimental)setting. Insocialsciencesettings,bycontrast,onedoesnotcommonlyfindsingle‐casestudiesofferingknockoutevidence foratheory.Thisis,inmyview,largelyaproductofthelooseness(theunderspecification)ofmostsocialscience theories.GeorgeandBennettpointoutthatwhilethethesisofthedemocraticpeaceisasclosetoa“law”as socialsciencehasyetseen,itcannotbeconfirmed(orrefuted)bylookingatspecificcausalmechanismsbecause thecausalpathwaysmandatedbythetheoryaremultipleanddiverse.Underthecircumstances,nosingle‐case testcanofferstrongconfirmationofthetheory(GeorgeandBennett2005,209). However,ifoneadoptsasofterversionofthecrucial‐casemethod—theleast‐likely(mostdifficult)case—then possibilitiesabound.Indeed,Isuspectthat,implicitly,mostcase‐studyworkthatmakesapositiveargument focusingonasinglecase(withoutacorrespondingcross‐caseanalysis)relieslargelyonthelogicoftheleast‐ likelycase.Rarelyisthislogicmadeexplicit,exceptperhapsinapassingphraseortwo.Yetthedeductivelogicof the“risky”predictioniscentraltothecase‐studyenterprise.(p.662) Whetheracasestudyisconvincingornot oftenrestsonthereader'sevaluationofhowstrongtheevidenceforanargumentmightbe,andthisinturn— wherevercross‐caseevidenceislimitedandnomanipulatedtreatmentcanbedevised—restsuponanestimation ofthedegreeof“fit”betweenatheoryandtheevidenceathand,asdiscussed. LilyTsai's(2007)investigationofgovernanceatthevillagelevelinChinaemploysseveralin‐depthcasestudiesof villageswhicharechosen(inpart)becauseoftheirleast‐likelystatusrelativetothetheoryofinterest.Tsai's hypothesisisthatvillageswithgreatersocialsolidarity(basedonpreexistingreligiousorfamilialnetworks)will developahigherlevelofsocialtrustandmutualobligationand,asaresult,willexperiencebettergovernance. Crucialcases,therefore,arevillagesthatevidenceahighlevelofsocialsolidaritybutwhich,alongother dimensions,wouldbejudgedleastlikelytodevelopgoodgovernance,e.g.theyarepoor,isolated,andlack democraticinstitutionsoraccountabilitymechanismsfromabove.“LiSettlement,”inFujianprovince,issucha case.Thefactthatthisimpoverishedvillagenonethelessboastsanimpressivesetofinfrastructural accomplishmentssuchaspavedroadswithdrainageditches(ararityinruralChina)suggeststhatsomething Page 12 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques ratherunusualisgoingonhere.Becausehercaseiscarefullychosentoeliminaterivalexplanations,Tsai's conclusionsaboutthespecialroleofsocialsolidarityaredifficulttogainsay.Howelseisonetoexplainthis otherwiseanomalousresult?Thisisthestrengthoftheleast‐likelycase,whereallotherplausiblecausalfactorsfor anoutcomehavebeenminimized.13 JackLevy(2002,144)referstothis,evocatively,asa“Sinatrainference:”ifitcanmakeithere,itcanmakeit anywhere(seealsoKhong1992,49;Sagan1995,49;Shafer1988,14–6).Thus,ifsocialsolidarityhasthe hypothesizedeffectinLiSettlementitshouldhavethesameeffectinmorepropitioussettings(e.g.wherethereis greatereconomicsurplus).Thesameimplicitlogicinformsmanycase‐studyanalyseswheretheintentofthestudy istoconfirmahypothesisonthebasisofasinglecase. Anothersortofcrucialcaseisemployedforthepurposeofdisconfirmingacausalhypothesis.AcentralPopperian insightisthatitiseasiertodisconfirmaninferencethantoconfirmthatsameinference.(Indeed,Popperdoubted thatanyinferencecouldbefullyconfirmed,andforthisreasonpreferredtheterm“corroborate.”)Thisis particularlytrueofcase‐studyresearchdesigns,whereevidenceislimitedtooneorseveralcases.Thekey provisoisthatthetheoryunderinvestigationmusttakeaconsistent(a.k.a.invariant,deterministic)form,evenifits predictionsarenotterrificallyprecise,wellelaborated,orbroad. Asithappens,thereareafairnumberofinvariantpropositionsfloatingaroundthesocialsciencedisciplines (GoertzandLevyforthcoming;GoertzandStarr2003).Itusedtobeargued,forexample,thatpoliticalstability wouldoccuronlyincountriesthatarerelativelyhomogeneous,orwhereexistingheterogeneitiesaremitigatedby (p.663) cross‐cuttingcleavages(Almond1956;Bentley1908/1967;Lipset1960/1963;Truman1951).Arend Lijphart's(1968)studyoftheNetherlands,apeacefulcountrywithreinforcingsocialcleavages,iscommonly viewedasrefutingthistheoryonthebasisofasinglein‐depthcaseanalysis.14 Granted,itmaybequestionedwhetherpresumedinvarianttheoriesarereallyinvariant;perhapstheyarebetter understoodasprobabilistic.Perhaps,thatis,thetheoryofcross‐cuttingcleavagesisstilltrue,probabilistically, despitetheapparentDutchexception.Orperhapsthetheoryisstilltrue,deterministically,withinasubsetofcases thatdoesnotincludetheNetherlands.(Thissortofclaimseemsunlikelyinthisparticularinstance,butitisquite plausibleinmanyothers.)Orperhapsthetheoryisinneedofreframing;itistrue,deterministically,butappliesonly tocross‐cuttingethnic/racialcleavages,nottocleavagesthatareprimarilyreligious.Onecanquibbleoverwhatit meansto“disconfirm”atheory.Thepointisthatthecrucialcasehas,inallthesecircumstances,provided importantupdatingofatheoreticalprior. Heretofore,Ihavetreatedcausalfactorsasdichotomous.Countrieshaveeitherreinforcingorcross‐cutting cleavagesandtheyhaveregimesthatareeitherpeacefulorconflictual.Evidently,thesesortsofparametersare oftenmattersofdegree.Inthisreadingofthetheory,casesaremoreorlesscrucial.Accordingly,themostuseful —i.e.mostcrucial—caseforLijphart'spurposeisonethathasthemostsegregatedsocialgroupsandthemost peacefulanddemocratictrackrecord.Intheserespects,theNetherlandswasaverygoodchoice.Indeed,the degreeofdisconfirmationofferedbythiscasestudyisprobablygreaterthanthedegreeofdisconfirmationthat mighthavebeenprovidedbyothercasessuchasIndiaorPapuaNewGuinea—countrieswheresocialpeacehas notalwaysbeensecure.Thepointisthatwherevariablesarecontinuousratherthandichotomousitispossibleto evaluatepotentialcasesintermsoftheirdegreeofcrucialness. Notethatthecrucial‐casemethodofcase‐selection,whetheremployedinaconfirmatoryordisconfirmatorymode, cannotbeemployedinalarge‐Ncontext.Thisisbecauseanexplicitcross‐casemodelwouldrenderthecrucial‐ casestudyredundant.Onceoneidentifiestherelevantparametersandthescoresofallcasesonthose parameters,onehasineffectconstructedacross‐casemodelthatconfirmsordisconfirmsthetheoryinquestion. Thecasestudyisthenceforthirrelevant,atleastasameansofdecisiveconfirmationordisconfirmation.15It remainshighlyrelevantasameansofexploringcausalmechanisms,ofcourse.Yet,becausethisobjectiveisquite differentfromthatwhichisusuallyassociatedwiththeterm,Ienlistanewtermforthistechnique. (p.664) 7PathwayCase Oneofthemostimportantfunctionsofcase‐studyresearchistheelucidationofcausalmechanisms.Butwhichsort ofcaseismostusefulforthispurpose?Althoughallcasestudiespresumablyshedlightoncausalmechanisms,not Page 13 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques allcasesareequallytransparent.Insituationswhereacausalhypothesisisclearandhasalreadybeenconfirmed bycross‐caseanalysis,researchersarewelladvisedtofocusonacasewherethecausaleffectofX1onYcan beisolatedfromotherpotentiallyconfoundingfactors(X2 ).Ishallcallthisapathwaycasetoindicateitsuniquely penetratinginsightintocausalmechanisms.Incontrasttothecrucialcase,thissortofmethodispracticableonlyin circumstanceswherecross‐casecovariationalpatternsarewellstudiedandwherethemechanismlinkingX1and Yremainsdim.Becausethepathwaycasebuildsonpriorcross‐caseanalysis,theproblemofcaseselectionmust besituatedwithinthatsample.Thereisnostandalonepathwaycase. Thelogicofthepathwaycaseisclearestinsituationsofcausalsufficiency—whereacausalfactorofinterest,X1, issufficientbyitself(thoughperhapsnotnecessary)toaccountforY'svalue(0or1).TheothercausesofY,about whichweneedmakenoassumptions,aredesignatedasavector,X2 . Notethatwherevervariouscausalfactorsaresubstitutableforoneanother,eachfactorisconceptualized (individually)assufficient(Braumoeller2003).Thus,situationsofcausalequifinalitypresumecausalsufficiencyon thepartofeachfactororsetofconjointfactors.Anexampleisprovidedbytheliteratureondemocratization, whichstipulatesthreemainavenuesofregimechange:leadership‐initiatedreform,acontrolledopeningto opposition,orthecollapseofanauthoritarianregime(Colomer1991).Thecase‐studyformatconstrainsusto analyzeoneatatime,soletuslimitourscopetothefirstone—leadership‐initiatedreform.Soconsidered,a causal‐pathwaycasewouldbeonewiththefollowingfeatures:(a)democratization,(b)leadership‐initiatedreform, (c)nocontrolledopeningtotheopposition,(d)nocollapseofthepreviousauthoritarianregime,and(e)noother extraneousfactorsthatmightaffecttheprocessofdemocratization.Inacaseofthistype,thecausalmechanisms bywhichleadership‐initiatedreformmayleadtodemocratizationwillbeeasiesttostudy.Notethatitisnot necessarytoassumethatleadership‐initiatedreformalwaysleadstodemocratization;itmayormaynotbea deterministiccause.Butitisnecessarytoassumethatleadership‐initiatedreformcansometimesleadto democratizationonitsown(givencertainbackgroundfeatures). Nowletusmovefromtheseexamplestoageneral‐purposemodel.Forheuristicpurposes,letuspresumethatall variablesinthatmodelaredichotomous(codedas0or1)andthatthemodeliscomplete(allcausesofYare included).Allcausalrelationshipswillbecodedsoastobepositive:X1andYcovaryasdoX2 andY.Thisallows ustovisualizearangeofpossiblecombinationsataglance. Recallthatthepathwaycaseisalwaysfocused,bydefinition,onasinglecausalfactor,denotedX1.(The researcher'sfocusmayshifttoothercausalfactors,butmayonlyfocusononecausalfactoratatime.)Inthis scenario,andregardlessofhow(p.665) manyadditionalcausesofYtheremightbe(denotedX2 ,avectorof controls),thereareonlyeightrelevantcasetypes,asillustratedinTable28.2.Identifyingthesecasetypesisa relativelysimplematter,andcanbeaccomplishedinasmall‐Nsamplebytheconstructionofatruth‐table (modeledafterTable28.2)orinalarge‐Nsamplebytheuseofcross‐tabs. Page 14 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Table28.2.Pathwaycasewithdichotomouscausalfactors Casetypes X1 X2 Y A 1 1 1 B 0 0 0 C 0 1 1 D 0 0 1 E 1 0 0 F 1 1 0 G 1 0 1 H 0 1 0 Notes:X1=thevariableoftheoreticalinterest.X2 =avectorofcontrols(ascoreof0indicatesthatallcontrol variableshaveascoreof0,whileascoreof1indicatesthatallcontrolvariableshaveascoreof1).Y=the outcomeofinterest.A–H=casetypes(theNforeachcasetypeisindeterminate).G,H=possiblepathway cases.Samplesize=indeterminate. Assumptions:(a)allvariablescanbecodeddichotomously(abinarycodingoftheconceptisvalid);(b)all independentvariablesarepositivelycorrelatedwithYinthegeneralcase;(c)X1is(atleastsometimes)a sufficientcauseofY. Notethatthetotalnumberofcombinationsofvaluesdependsonthenumberofcontrolvariables,whichwehave representedwithasinglevector,X2 .Ifthisvectorconsistsofasinglevariablethenthereareonlyeightcase types.Ifthisvectorconsistsoftwovariables(X2a ,X2b )thenthetotalnumberofpossiblecombinationsincreases fromeight(23 )tosixteen(24 ).Andsoforth.However,noneofthesecombinationsisrelevantforpresentpurposes exceptthosewhereX2a andX2b havethesamevalue(0or1).“Mixed”casesarenotcausalpathwaycases,for reasonsthatshouldbecomeclear. Thepathwaycase,followingthelogicofthecrucialcase,isonewherethecausalfactorofinterest,X1,correctly predictsYwhileallotherpossiblecausesofY(representedbythevector,X2 )make“wrong”predictions.IfX1is— atleastinsomecircumstances—asufficientcauseofY,thenitisthesesortsofcasesthatshouldbemostuseful fortracingcausalmechanisms.ThereareonlytwosuchcasesinTable28.2—GandH.Inallothercases,the mechanismrunningfromX1toYwouldbedifficulttodiscerneitherbecauseX1andYarenotcorrelatedinthe usualway(constitutinganunusualcase,inthetermsofourhypothesis)orbecauseotherconfoundingfactors(X 2 )intrude.IncaseA,forexample,thepositivevalueonY(p.666) couldbeaproductofX1orX2 .Anin‐depth examinationofthiscaseisnotlikelytobeveryrevealing. Keepinmindthatbecausetheresearcheralreadyknowsfromhercross‐caseexaminationwhatthegeneral causalrelationshipsare,sheknows(priortothecase‐studyinvestigation)whatconstitutesacorrectorincorrect prediction.Inthecrucial‐casemethod,bycontrast,theseexpectationsaredeductiveratherthanempirical.Thisis whatdifferentiatesthetwomethods.Andthisiswhythecausalpathwaycaseisusefulprincipallyforelucidating causalmechanismsratherthanverifyingorfalsifyinggeneralpropositions(whicharealreadymoreorless apparentfromthecross‐caseevidence).Ofcourse,wemustleaveopenthepossibilitythattheinvestigationof causalmechanismswouldinvalidateageneralclaim,ifthatclaimisutterlycontingentuponaspecificsetofcausal mechanismsandthecasestudyshowsthatnosuchmechanismsarepresent.However,thisisratherunlikelyin Page 15 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques mostsocialsciencesettings.Usually,theresultofsuchafindingwillbeareformulationofthecausalprocessesby whichX1causesY—or,alternatively,arealizationthatthecaseunderinvestigationisaberrant(atypicalofthe generalpopulationofcases). Sometimes,theresearchquestionisframedasaunidirectionalcause:oneisinterestedinwhy0becomes1(or viceversa)butnotinwhy1becomes0.Inourpreviousexample,weaskedwhydemocraciesfail,notwhy countriesbecomedemocraticorauthoritarian.Soframed,therecanbeonlyonetypeofcausal‐pathwaycase. (Whetherregimefailureiscodedas0or1isamatteroftaste.)Whereresearchersareinterestedinbidirectional causality—amovementfrom0to1aswellasfrom1to0—therearetwopossiblecausal‐pathwaycases,GandH. Inpractice,however,oneofthesecasetypesisalmostalwaysmoreusefulthantheother.Thus,itseems reasonabletoemploytheterm“pathwaycase”inthesingular.Inordertodeterminewhichofthesetwocasetypes willbemoreusefulforintensiveanalysistheresearchershouldlooktoseewhethereachcasetypeexhibits desirablefeaturessuchas:(a)arare(unusual)valueonX1orY(designated“extreme”inourprevious discussion),(b)observabletemporalvariationinX1,(c)anX1/Yrelationshipthatiseasiertostudy(ithasmore visiblefeatures;itismoretransparent),or(d)alowerresidual(thusindicatingamoretypicalcase,withintheterms ofthegeneralmodel).Usually,thechoicebetweenGandHisintuitivelyobvious. Now,letusconsiderascenarioinwhichall(ormost)variablesofconcerntothemodelarecontinuous,ratherthan dichotomous.Here,thejobofcaseselectionisconsiderablymorecomplex,forcausal“sufficiency”(intheusual sense)cannotbeinvoked.Itisnolongerplausibletoassumethatagivencausecanbeentirelypartitioned,i.e. rivalfactorseliminated.However,thesearchforapathwaycasemaystillbeviable.Whatwearelookingforinthis scenarioisacasethatsatisfiestwocriteria:(1)itisnotanoutlier(oratleastnotanextremeoutlier)inthegeneral modeland(2)itsscoreontheoutcome(Y)isstronglyinfluencedbythetheoreticalvariableofinterest(X1),taking allotherfactorsintoaccount(X2 ).Inthissortofcaseitshouldbeeasiestto“see”thecausalmechanismsthatlie betweenX1andY. Achievingtheseconddesideratarequiresabitofmanipulation.Inordertodeterminewhich(nonoutlier)casesare moststronglyaffectedbyX1,givenalltheother(p.667) parametersinthemodel,onemustcomparethesizeof theresidualsforeachcaseinareducedformmodel,Y=Constant+X2 +Resreduced ,withthesizeofthe residualsforeachcaseinafullmodel,Y=Constant+X2 +X1+Resfull.Thepathwaycaseisthatcase,orsetof cases,whichshowsthegreatestdifferencebetweentheresidualforthereduced‐formmodelandthefullmodel (ΔResidual).Thus, Notethattheresidualforacasemustbesmallerinthefullmodelthaninthereduced‐formmodel;otherwise,the additionofthevariableofinterest(X1)pullsthecaseawayfromtheregressionline.Wewanttofindacasewhere theadditionofX1pushesthecasetowardstheregressionline,i.e.ithelpsto“explain”thatcase. Asanexample,letussupposethatweareinterestedinexploringtheeffectofmineralwealthontheprospectsfor democracyinasociety.Accordingtoagooddealofworkonthissubject,countrieswithabountyofnatural resources—particularlyoil—arelesslikelytodemocratize(oroncehavingundergoneademocratictransition,are morelikelytoreverttoauthoritarianrule)(Barro1999;Humphreys2005;Ross2001).Thecross‐countryevidence isrobust.Yetasisoftenthecase,thecausalmechanismsremainratherobscure.Inordertobetterunderstandthis phenomenonitmaybeworthwhiletoexploitthefindingsofcross‐countryregressionmodelsinordertoidentifya countrywhoseregimetype(i.e.itsdemocracy“score”onsomegeneralindex)isstronglyaffectedbyitsnatural‐ researchwealth,allotherthingsheldconstant.Ananalysisofthissortidentifiestwocountries—theUnitedArab EmiratesandKuwait—withhighΔResidualvaluesandmodestresidualsinthefullmodel(signifyingthatthese casesarenotoutliers).Researchersseekingtoexploretheeffectofoilwealthonregimetypemightdowellto focusonthesetwocasessincetheirpatternsofdemocracycannotbewellexplainedbyotherfactors—e.g. economicdevelopment,religion,Europeaninfluence,orethnicfractionalization.Thepresenceofoilwealthin thesecountrieswouldappeartohaveastrongindependenteffectontheprospectsfordemocratizationinthese cases,aneffectthatiswellmodeledbygeneraltheoryandbytheavailablecross‐caseevidence. Toreiterate,thelogicofcausal“elimination”ismuchmorecompellingwherevariablesaredichotomousandwhere causalsufficiencycanbeassumed(X1issufficientbyitself,atleastinsomecircumstances,tocauseY).Where variablesarecontinuous,thestrategyofthepathwaycaseismoredubious,forpotentiallyconfoundingcausal factors(X2 )cannotbeneatlypartitioned.Evenso,wehaveindicatedwhytheselectionofapathwaycasemaybe Page 16 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques alogicalapproachtocase‐studyanalysisinmanycircumstances. Theexceptionsmaybebrieflynoted.Sometimes,whereallvariablesinamodelaredichotomous,thereareno pathwaycases,i.e.nocasesoftypeGorH(inTable28.2).Thisisknownasthe“emptycell”problem,ora problemofseverecausalmulticollinearity.Theuniverseofobservationaldatadoesnotalwaysobligeuswith casesthatallowustoindependentlytestagivenhypothesis.Wherevariablesarecontinuous,theanalogous problemisthatofacausalvariableofinterest(X1)that(p.668) hasonlyminimaleffectsontheoutcomeof interest.Thatis,itsroleinthegeneralmodelisquiteminor.Inthesesituations,theonlycasesthatarestrongly affectedbyX1—ifthereareanyatall—maybeextremeoutliers,andthesesortsofcasesarenotproperly regardedasprovidingconfirmatoryevidenceforaproposition,forreasonsthatareabundantlyclearbynow. Finally,itshouldbeclarifiedthattheidentificationofacausalpathwaycasedoesnotobviatetheutilityofexploring othercases.Onemight,forexample,wanttocomparebothsortsofpotentialpathwaycases—GandH—witheach other.Manyothercombinationssuggestthemselves.However,thissortofmulti‐caseinvestigationmovesbeyond thelogicofthecausal‐pathwaycase. 8Most‐similarCases Themost‐similarmethodemploysaminimumoftwocases.16 Initspurestform,thechosenpairofcasesissimilarin allrespectsexceptthevariable(s)ofinterest.Ifthestudyisexploratory(i.e.hypothesisgenerating),the researcherlooksforcasesthatdifferontheoutcomeoftheoreticalinterestbutaresimilaronvariousfactorsthat mighthavecontributedtothatoutcome,asillustratedinTable28.3(A).Thisisacommonformofcaseselectionat theinitialstageofresearch.Often,fruitfulanalysisbeginswithanapparentanomaly:twocasesareapparently quitesimilar,andyetdemonstratesurprisinglydifferentoutcomes.Thehopeisthatintensivestudyofthesecases willrevealone—oratmostseveral—factorsthatdifferacrossthesecases.Thesedifferingfactors(X1)arelooked uponasputativecauses.Atthisstage,theresearchmaybedescribedbytheseconddiagraminTable28.3(B). Sometimes,aresearcherbeginswithastronghypothesis,inwhichcaseherresearchdesignisconfirmatory (hypothesistesting)fromtheget‐go.Thatis,shestrivestoidentifycasesthatexhibitdifferentoutcomes,different scoresonthefactorofinterest,andsimilarscoresonallotherpossiblecausalfactors,asillustratedinthesecond (hypothesis‐testing)diagraminTable28.3(B). Thepointisthatthepurposeofamost‐similarresearchdesign,andhenceitsbasicsetup,oftenchangesasa researchermovesfromanexploratorytoaconfirmatorymodeofanalysis.However,regardlessofwhereone begins,theresults,whenpublished,looklikeahypothesis‐testingresearchdesign.Questionmarkshavebeen removed:(A)becomes(B)inTable28.3. Asanexample,letusconsiderLeonEpstein'sclassicstudyofpartycohesion,whichfocusesontwo“most‐similar” countries,theUnitedStatesandCanada.Canadahashighlydisciplinedpartieswhosemembersvotetogetheron theflooroftheHouseof(p.669) CommonswhiletheUnitedStateshasweak,undisciplinedparties,whose membersoftendefectonfloorvotesinCongress.Inexplainingthesedivergentoutcomes,persistentovermany years,Epsteinfirstdiscussespossiblecausalfactorsthatareheldmoreorlessconstantacrossthetwocases. BoththeUnitedStatesandCanadainheritedEnglishpoliticalcultures,bothhavelargeterritoriesand heterogeneouspopulations,botharefederal,andbothhavefairlyloosepartystructureswithstrongregionalbases andaweakcenter.Thesearethe“control”variables.Wheretheydifferisinoneconstitutionalfeature:Canadais parliamentarywhiletheUnitedStatesispresidential.AnditisthisinstitutionaldifferencethatEpsteinidentifiesas thecrucial(differentiating)cause.(Forfurtherexamplesofthemost‐similarmethodseeBrenner1976;Hamilton 1977;Lipset1968;Miguel2004;Moulder1977;Posner2004.) Page 17 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Table28.3.Most‐similaranalysiswithtwocasetypes Casetypes X1 X2 Y (A)Hypothesis‐generating(Y‐centered): A ? 0 1 B ? 0 0 A 1 0 1 B 0 0 0 (B)Hypothesis‐testing(X1/Y‐centered): X1=thevariableoftheoreticalinterest.X2 =avectorofcontrols.Y=theoutcomeofinterest. Severalcaveatsapplytoanymost‐similaranalysis(inadditiontotheusualsetofassumptionsapplyingtoallcase‐ studyanalysis).First,eachcausalfactorisunderstoodashavinganindependentandadditiveeffectonthe outcome;thereareno“interaction”effects.Second,onemustcodecasesdichotomously(high/low, present/absent).Thisisstraightforwardiftheunderlyingvariablesarealsodichotomous(e.g.federal/unitary). However,itisoftenthecasethatvariablesofconcerninthemodelarecontinuous(e.g.partycohesion).Inthis setting,theresearchermust“dichotomize”thescoringofcasessoastosimplifythetwo‐caseanalysis.(Some flexibilityisadmissibleonthevectorofcontrols(X2 )thatare“heldconstant”acrossthecases.Nonidentityis tolerableifthedeviationrunscountertothepredictedhypothesis.Forexample,EpsteindescribesboththeUnited StatesandCanadaashavingstrongregionalbasesofpower,afactorthatisprobablymoresignificantinrecent CanadianhistorythaninrecentAmericanhistory.However,becauseregionalbasesofpowershouldleadto weakerparties,ratherthanstrongerparties,thiselementofnonidentitydoesnotchallengeEpstein'sconclusions. Indeed,itsetsupamost‐difficultresearchscenario,asdiscussedabove.) Inonerespecttherequirementsforcasecontrolarenotsostringent.Specifically,itisnotusuallynecessaryto measurecontrolvariables(atleastnotwithahighdegreeofprecision)inordertocontrolforthem.Iftwocountries canbeassumed(p.670) tohavesimilarculturalheritagesoneneedn'tworryaboutconstructingvariablesto measurethatheritage.Onecansimplyassertthat,whatevertheyare,theyaremoreorlessconstantacrossthe twocases.Thisissimilartothetechniqueemployedinarandomizedexperiment,wheretheresearchertypically doesnotattempttomeasureallthefactorsthatmightaffectthecausalrelationshipofinterest.Sheassumes, rather,thattheseunknownfactorshavebeenneutralizedacrossthetreatmentandcontrolgroupsby randomizationorbythechoiceofasamplethatisinternallyhomogeneous. Themostusefulstatisticaltoolforidentifyingcasesforin‐depthanalysisinamost‐similarsettingisprobablysome varietyofmatchingstrategy—e.g.exactmatching,approximatematching,orpropensity‐scorematching.17 The productofthisprocedureisasetofmatchedcasesthatcanbecomparedinwhateverwaytheresearcherdeems appropriate.Thesearethe“most‐similar”cases.RosenbaumandSilber(2001,223)summarize: Unlikemodel‐basedadjustments,where[individuals]vanishandarereplacedbythecoefficientsofa model,inmatching,ostensiblycomparablepatternsarecompareddirectly,onebyone.Modernmatching methodsinvolvestatisticalmodelingandcombinatorialalgorithms,buttheendresultisacollectionofpairs orsetsofpeoplewholookcomparable,atleastonaverage.Inmatching,peopleretaintheirintegrityas people,sotheycanbeexaminedandtheirstoriescanbetoldindividually. Matching,concludetheauthors,“facilitates,ratherthaninhibits,thickdescription”(RosenbaumandSilber2001, 223). Page 18 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Inprinciple,thesamematchingtechniquesthathavebeenusedsuccessfullyinobservationalstudiesofmedical treatmentsmightalsobeadaptedtothestudyofnationstates,politicalparties,cities,orindeedanytraditional pairedcasesinthesocialsciences.Indeed,thecurrentpopularityofmatchingamongstatisticians—relative,that is,togarden‐varietyregressionmodels—restsuponwhatqualitativeresearcherswouldrecognizeasa“case‐ based”approachtocausalanalysis.IfRosenbaumandSilberarecorrect,itmaybeperfectlyreasonableto appropriatethislarge‐Nmethodofanalysisforcase‐studypurposes. Aswithothermethodsofcaseselection,themost‐similarmethodispronetoproblemsofnonrepresentativeness.If employedinaqualitativefashion(withoutasystematiccross‐caseselectionstrategy),potentialbiasesinthe chosencasemustbeaddressedinaspeculativeway.Iftheresearcheremploysamatchingtechniqueofcase selectionwithinalarge‐Nsample,theproblemofpotentialbiascanbeaddressedbyassuringthechoiceofcases thatarenotextremeoutliers,asjudgedbytheirresidualsinthefullmodel.Most‐similarcasesshouldalsobe “typical”cases,thoughsomescopefordeviancearoundtheregressionlinemaybeacceptableforpurposesof findingagoodfitamongcases.(p.671) Table28.4.Most‐differentanalysiswithtwocases Casetypes X1 X2a X2b X2c X2d Y A 1 1 0 1 0 1 B 1 0 1 0 1 1 X1=thevariableoftheoreticalinterest.X2a–d =avectorofcontrols.Y=theoutcomeofinterest. 9Most‐differentCases Afinalcase‐selectionmethodisthereverseimageofthepreviousmethod.Here,variationonindependent variablesisprized,whilevariationontheoutcomeiseschewed.Ratherthanlookingforcasesthataremost‐ similar,onelooksforcasesthataremost‐different.Specifically,theresearchertriestoidentifycaseswherejust oneindependentvariable(X1),aswellasthedependentvariable(Y),covary,whileallotherplausiblefactors(X 18 2a–d )showdifferentvalues. Thesimplestformofthistwo‐casecomparisonisillustratedinTable28.4.CasesAandBaredeemed“most different,”thoughtheyaresimilarintwoessentialrespects—thecausalvariableofinterestandtheoutcome. Asanexample,IfollowMarcHoward's(2003)recentwork,whichexplorestheenduringimpactofCommunismon civilsociety.19 Cross‐nationalsurveysshowastrongcorrelationbetweenformerCommunistregimesandlow socialcapital,controllingforavarietyofpossibleconfounders.Itisastrongresult.Howardwonderswhythis relationshipissostrongandwhyitpersists,andperhapsevenstrengthens,incountriesthatarenolongersocialist orauthoritarian.Inordertoanswerthisquestion,hefocusesontwomost‐differentcases,RussiaandEast Germany.Thesetwocountrieswerequitedifferent—inallwaysotherthantheirCommunistexperience—priorto theSovietera,duringtheSovietera(sinceEastGermanyreceivedsubstantialsubsidiesfromWestGermany),and inthepost‐Sovietera,asEastGermanywasabsorbedintoWestGermany.Yet,theybothscorenearthebottomof variouscross‐nationalindicesintendedtomeasuretheprevalenceofcivicengagementinthecurrentera.Thus, Howard's(2003,6–9)caseselectionproceduremeetstherequirementsofthemost‐differentresearchdesign: Varianceisfoundonall(ormost)dimensions(p.672) asidefromthekeyfactorofinterest(Communism)andthe outcome(civicengagement). Whatleverageisbroughttotheanalysisfromthisapproach?Howard'scasestudiescombineevidencedrawnfrom masssurveysandfromin‐depthinterviewsofsmall,stratifiedsamplesofRussiansandEastGermans.(Thisisa goodillustration,incidentally,ofhowquantitativeandqualitativeevidencecanbefruitfullycombinedinthe intensivestudyofseveralcases.)Theproductofthisanalysisistheidentificationofthreecausalpathwaysthat, Howard(2003,122)claims,helptoexplainthelaggardstatusofcivilsocietyinpost‐Communistpolities:“the Page 19 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques mistrustofcommunistorganizations,thepersistenceoffriendshipnetworks,andthedisappointmentwithpost‐ communism.”Simplyput,Howard(2003,145)concludes,“agreatnumberofcitizensinRussiaandEastern Germanyfeelastrongandlingeringsenseofdistrustofanykindofpublicorganization,ageneralsatisfactionwith theirownpersonalnetworks(accompaniedbyasenseofdeterioratingrelationswithinsocietyoverall),and disappointmentinthedevelopmentsofpost‐communism.” Thestrengthofthismost‐differentcaseanalysisisthattheresultsobtainedinEastGermanyandRussiashould alsoapplyinotherpost‐Communistpolities(e.g.Lithuania,Poland,Bulgaria,Albania).Bychoosinga heterogeneoussample,Howardsolvestheproblemofrepresentativenessinhisrestrictedsample.However,this sampleisdemonstrablynotrepresentativeacrossthepopulationoftheinference,whichisintendedtocoverall countriesoftheworld. Moreproblematicisthelackofvariationonkeycausalfactorsofinterest—Communismanditsputativecausal pathways.Forthisreason,itisdifficulttoreachconclusionsaboutthecausalstatusofthesefactorsonthebasisof themost‐differentanalysisalone.Itispossible,thatis,thatthethreecausalpathwaysidentifiedbyHowardalso operatewithinpolitiesthatneverexperiencedCommunistrule. Nordoesitseempossibletoconclusivelyeliminaterivalhypothesesonthebasisofthismost‐differentanalysis. Indeed,thisisnotHoward'sintention.Hewishesmerelytoshowthatwhateverinfluenceoncivilsocietymightbe attributedtoeconomic,cultural,andotherfactorsdoesnotexhaustthissubject. Myconsideredjudgmentisthatthemost‐differentresearchdesignprovidesminimalleverageintotheproblemof whyCommunistsystemsappeartosuppresscivicengagement,yearsaftertheirdisappearance.Fortunately,thisis nottheonlyresearchdesignemployedbyHowardinhisadmirablestudy.Indeed,theauthoremploystwoother small‐Ncross‐casemethods,aswellasalarge‐Ncross‐countrystatisticalanalysis.Thesemethodsdomostofthe analyticwork.EastGermanymayberegardedasacausalpathwaycase(seeabove).Ithasalltheattributes normallyassumedtofostercivicengagement(e.g.agrowingeconomy,multipartycompetition,civilliberties,afree press,closeassociationwithWesternEuropeancultureandpolitics),butnonethelessshowslittleorno improvementonthisdimensionduringthepost‐transitionera(Howard2003,8).Itisplausibletoattributethislack ofchangetoitsCommunistpast,asHowarddoes,inwhichcaseEastGermanyshouldbeafruitfulcaseforthe investigationofcausalmechanisms.ThecontrastbetweenEastandWestGermanyprovidesamost‐similar analysissincethetwopolitiesshare(p.673) virtuallyeverythingexceptaCommunistpast.Thisvariationisalso deftlyexploitedbyHoward. Idonotwishtodismissthemost‐differentresearchmethodentirely.Surely,Howard'sfindingsarestrongerwiththe intensiveanalysisofRussiathantheywouldbewithout.Yethisbookwouldnotstandsecurelyontheempirical foundationprovidedbymost‐differentanalysisalone.Ifonestripsawaythepathway‐case(EastGermany)andthe most‐similaranalysis(East/WestGermany)thereislittleleftuponwhichtobaseananalysisofcausalrelations (asidefromthelarge‐Ncross‐nationalanalysis).Indeed,mostscholarswhoemploythemost‐differentmethoddo soinconjunctionwithothermethods.20 Itisrarely,ifever,astandalonemethod.21 GeneralizingfromthisdiscussionofMarcHoward'swork,Iofferthefollowingsummaryremarksonthemost‐ differentmethodofcaseanalysis.(Ileaveasideissuesfacedbyallcase‐studyanalyses,issuesthatareexplored inGerring2007.) LetusbeginwithamethodologicalobstaclethatisfacedbybothMilleanstylesofanalysis—thenecessityof dichotomizingeveryvariableintheanalysis.Recallthat,aswithmost‐similaranalysis,differencesacrosscases mustgenerallybesizeableenoughtobeinterpretableinanessentiallydichotomousfashion(e.g.high/low, present/absent)andsimilaritiesmustbecloseenoughtobeunderstoodasessentiallyidentical(e.g.high/high, present/present).OtherwisetheresultsofaMilleanstyleanalysisarenotinterpretable.Theproblemof“degrees” isdeadlyifthevariablesunderconsiderationare,bynature,continuous(e.g.GDP).Thisisaparticularconcernin Howard'sanalysis,whereEastGermanyscoressomewhathigherthanRussiaincivicengagement;theyareboth low,butRussiaisquiteabitlower.Howardassumesthatthisdivergenceisminimalenoughtobeunderstoodasa differenceofdegreesratherthanofkinds,ajudgmentthatmightbequestioned.Intheserespects,most‐different analysisisnomoresecure—butalsonoless—thanmost‐similaranalysis. Inonerespect,most‐differentanalysisissuperiortomost‐similaranalysis.Ifthecodingassumptionsaresound,the Page 20 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques most‐differentresearchdesignmaybequiteusefulforeliminatingnecessarycauses.Causalfactorsthatdonot appearacrossthechosencases—e.g.X2a–d inTable28.4—areevidentlyunnecessaryfortheproductionofY. However,itdoesnotfollowthatthemost‐differentmethodisthebestmethodforeliminatingnecessarycauses. Notethatthedefiningfeatureofthismethodisthe(p.674) sharedelementacrosscases—X1inTable28.4.This featuredoesnothelponetoeliminatenecessarycauses.Indeed,ifonewerefocusedsolelyoneliminating necessarycausesonewouldpresumablyseekoutcasesthatregisterthesameoutcomesandhavemaximum diversityonotherattributes.InTable28.4,thiswouldbeasetofcasesthatsatisfyconditionsX2a–d ,butnotX1. Thus,eventhepresumedstrengthofthemost‐differentanalysisisnotsostrong. Usually,case‐studyanalysisisfocusedontheidentification(orclarification)ofcausalrelations,nottheelimination ofpossiblecauses.Inthissetting,themost‐differenttechniqueisuseful,butonlyifassumptionsofcausal uniquenesshold.By“causaluniqueness,”Imeanasituationinwhichagivenoutcomeistheproductofonlyone cause:YcannotoccurexceptinthepresenceofX.Xisnecessary,andinsomesituations(givencertain backgroundconditions)sufficient,tocauseY.22 Considerthefollowinghypotheticalexample.Supposethatanewdisease,aboutwhichlittleisknown,has appearedinCountryA.Therearehundredsofinfectedpersonsacrossdozensofaffectedcommunitiesinthat country.InCountryB,locatedattheotherendoftheworld,severalnewcasesofthediseasesurfaceinasingle community.Inthissetting,wecanimaginetwosortsofMilleananalyses.Thefirstexaminestwosimilarcommunities withinCountryA,oneofwhichhasdevelopedthediseaseandtheotherofwhichhasnot.Thisisthemost‐similar styleofcasecomparison,andfocusesaccordinglyontheidentificationofadifferencebetweenthetwocasesthat mightaccountforvariationacrossthesample.Asecondapproachfocusesoncommunitieswherethediseasehas appearedacrossthetwocountriesandsearchesforanysimilaritiesthatmightaccountforthesesimilaroutcomes. Thisisthemost‐differentresearchdesign. Bothareplausibleapproachestothisparticularproblem,andwecanimagineepidemiologistsemployingthem simultaneously.However,themost‐differentdesigndemandsstrongerassumptionsabouttheunderlyingfactorsat work.Itsupposesthatthediseasearisesfromthesamecauseinanysetting.Thisisoftenareasonableoperating assumptionwhenoneisdealingwithnaturalphenomena,thoughtherearecertainlymanyexceptions.Death,for example,hasmanycauses.Forthisreason,itwouldnotoccurtoustolookformost‐differentcasesofhigh mortalityaroundtheworld.Inorderforthemost‐differentresearchdesigntoeffectivelyidentifyacausalfactorat workinagivenoutcome,theresearchermustassumethatX1—thefactorheldconstantacrossthediversecases —istheonlypossiblecauseofY(seeTable28.4).Thisassumptionrarelyholdsinsocial‐scientificsettings.Most outcomesofinteresttoanthropologists,economists,politicalscientists,andsociologistshavemultiplecauses. Therearemanywaystowinanelection,tobuildawelfarestate,togetintoawar,tooverthrowagovernment,or— returningtoMarcHoward'swork—tobuildastrongcivilsociety.Anditisforthisreasonthatmost‐differentanalysis israrelyappliedinsocialscienceworkand,whereapplied,israrelyconvincing. Ifthisseemsatadsevere,thereisamorecharitablewayofapproachingthemost‐differentmethod.Arguably,this isnotapure“method”atallbutmerelya(p.675) supplement,awayofincorporatingdiversityinthesub‐sample ofcasesthatprovidetheunusualoutcomeofinterest.Iftheunusualoutcomeisrevolutions,onemightwishto encompassawidevarietyofrevolutionsinone'sanalysis.Iftheunusualoutcomeispost‐Communistcivilsociety,it seemsappropriatetoincludeadiversesetofpost‐Communistpolitiesinone'ssampleofcasestudies,asMarc Howarddoes.Fromthisperspective,themost‐differentmethod(so‐called)mightbebetterlabeledadiverse‐case method,asexploredabove. 10Conclusions Inordertobeacaseofsomethingbroaderthanitself,thechosencasemustberepresentative(insomerespects) ofalargerpopulation.Otherwise—ifitispurelyidiosyncratic(“unique”)—itisuninformativeaboutanythinglying outsidethebordersofthecaseitself.Astudybasedonanonrepresentativesamplehasno(orverylittle)external validity.Tobesure,nophenomenonispurelyidiosyncratic;thenotionofauniquecaseisamatterthatwouldbe difficulttodefine.Oneisconcerned,asalways,withmattersofdegree.Casesaremoreorlessrepresentativeof somebroaderphenomenonand,onthatscore,maybeconsideredbetterorworsesubjectsforintensiveanalysis. (Theoneexception,asnoted,istheinfluentialcase.) Page 21 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Ofalltheproblemsbesettingcase‐studyanalysis,perhapsthemostpersistent—andthemostpersistently bemoaned—istheproblemofsamplebias(AchenandSnidal1989;CollierandMahoney1996;Geddes1990;King, Keohane,andVerba1994;Rohlfing2004;Sekhon2004).LisaMartin(1992,5)findsthattheoveremphasisof internationalrelationsscholarsonafewwell‐knowncasesofeconomicsanctions—mostofwhichfailedtoelicit anychangeinthesanctionedcountry—“hasdistortedanalystsviewofthedynamicsandcharacteristicsof economicsanctions.”BarbaraGeddes(1990)chargesthatmanyanalysesofindustrialpolicyhavefocused exclusivelyonthemostsuccessfulcases—primarilytheEastAsianNICs—leadingtobiasedinferences.Anna BremanandCarolynShelton(2001)showthatcase‐studyworkonthequestionofstructuraladjustmentis systematicallybiasedinsofarasresearcherstendtofocusondisastercases—thosewherestructuraladjustmentis associatedwithverypoorhealthandhumandevelopmentoutcomes.Thesecases,oftenlocatedinsub‐Saharan Africa,arebynomeansrepresentativeoftheentirepopulation.Consequently,scholarshiponthequestionof structuraladjustmentishighlyskewedinaparticularideologicaldirection(againstneoliberalism)(seealsoGerring, Thacker,andMoreno2005). Theseexamplesmightbemultipliedmanytimes.Indeed,formanytopicsthemost‐studiedcasesareacknowledged tobelessthanrepresentative.Itisworthreflectinguponthefactthatourknowledgeoftheworldisheavilycolored byafew“big”(populous,rich,powerful)countries,andthatagoodportionofthedisciplinesofeconomics, politicalscience,andsociologyarebuiltuponscholars'familiaritywith(p.676) theeconomics,politicalscience, andsociologyofonecountry,theUnitedStates.23 Case‐studyworkisparticularlypronetoproblemsof investigatorbiassincesomuchridesontheresearcher'sselectionofone(orafew)cases.Eveniftheinvestigator isunbiased,hersamplemaystillbebiasedsimplybyvirtueof“random”error(whichmaybeunderstoodas measurementerror,errorinthedata‐generationprocess,orasanunderlyingcausalfeatureoftheuniverse). Thereareonlytwosituationsinwhichacase‐studyresearcherneednotbeconcernedwiththe representativenessofherchosencase.Thefirstistheinfluentialcaseresearchdesign,whereacaseischosen becauseofitspossibleinfluenceonacross‐casemodel,andhenceisnotexpectedtoberepresentativeofa largersample.Thesecondisthedeviant‐casemethod,wherethechosencaseisemployedtoconfirmabroader cross‐caseargumenttowhichthecasestandsasanapparentexception.Yetevenherethechosencaseis expectedtoberepresentativeofabroadersetofcases—those,inparticular,thatarepoorlyexplainedbythe extantmodel. Inallothercircumstances,casesmustberepresentativeofthepopulationofinterestinwhateverwaysmightbe relevanttothepropositioninquestion.Notethatwherearesearcherisattemptingtodisconfirmadeterministic propositionthequestionofrepresentativenessisperhapsmoreappropriatelyunderstoodasaquestionof classification:Isthechosencaseappropriatelyclassifiedasamemberofthedesignatedpopulation?Ifso,thenitis fodderforadisconfirmingcasestudy. Iftheresearcherisattemptingtoconfirmadeterministicproposition,ortomakeprobabilisticargumentsabouta causalrelationship,thentheproblemofrepresentativenessisofthemoreusualsort:IscaseAunit‐homogeneous relativetoothercasesinthepopulation?Thisisnotaneasymattertotest.However,inalarge‐Ncontextthe residualforthatcase(inwhatevermodeltheresearcherhasgreatestconfidencein)isareasonableplacetostart. Ofcourse,thistestisonlyasgoodasthemodelathand.Anyincorrectspecificationsorincorrectmodeling procedureswilllikelybiastheresultsandgiveanincorrectassessmentofeachcase's“typicality.”Inaddition, thereisthepossibilityofstochasticerror,errorsthatcannotbemodeledinageneralframework.Giventhe explanatoryweightthatindividualcasesareaskedtobearinacase‐studyanalysis,itiswisetoconsidermore thanjusttheresidualtestofrepresentativeness.Deductivelogicandanin‐depthknowledgeofthecasein questionareoftenmorereliabletoolsthantheresultsofacross‐casemodel. Inanycase,thereisnodispensingwiththequestion.Casestudies(withthetwoexceptionsalreadynoted)rest uponanassumedsynecdoche:Thecaseshouldstandforapopulation.Ifthisisnottrue,orifthereisreasonto doubtthisassumption,thentheutilityofthecasestudyisbroughtseverelyintoquestion. Fortunately,thereissomesafetyinnumbers.Insofarascase‐studyevidenceiscombinedwithcross‐case evidencetheissueofsamplebiasismitigated.Indeed,thesuspicionofcase‐studyworkthatonefindsinthesocial sciencestodayis,inmyview,aproductofatoo‐literalinterpretationofthecase‐studymethod.Acasestudytout (p.677) courtisthoughttomeanacasestudytoutseul.Insofarascasestudiesandcross‐casestudiescanbe Page 22 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques enlistedwithinthesameinvestigation(eitherinthesamestudyorbyreferencetootherstudiesinthesame subfield),problemsofrepresentativenessarelessworrisome.Thisisthevirtueofcross‐levelwork,a.k.a. “triangulation.” 11Ambiguities Beforeconcluding,Iwishtodrawattentiontotwoambiguitiesincase‐selectionstrategiesincase‐studyresearch. Thefirstconcernstheadmixtureofseveralcase‐selectionstrategies.Thesecondconcernsthechangingstatus ofacaseasastudyproceeds. Somecasestudiesfollowonlyonestrategyofcaseselection.Theyaretypical,diverse,extreme,deviant, influential,crucial,pathway,most‐similar,ormost‐differentresearchdesigns,asdiscussed.However,many casestudiesmixandmatchamongthesecase‐selectionstrategies.Indeed,insofarasallcasestudiesseek representativesamples,theyarealwaysinsearchof“typical”cases.Thus,itiscommonforwriterstodeclarethat theircaseis,forexample,bothextremeandtypical;ithasanextremevalueonX1orYbutisnot,inother respects,idiosyncratic.Thereisnotmuchthatonecansayaboutthesecombinationsofstrategiesexceptthat, wherethecasesallowforavarietyofempiricalstrategies,thereisnoreasonnottopursuethem.Andwherethe samecasescanserveseveralfunctionsatonce(withoutfurthereffortontheresearcher'spart),thereislittlecost toamulti‐prongedapproachtocaseanalysis. Thesecondissuethatdeservesemphasisisthechangingstatusofacaseduringthecourseofaresearcher's investigation—whichmaylastforyears,ifnotdecades.Theproblemisacutewhereveraresearcherbeginsinan exploratorymodeandproceedstohypothesis‐testing(thatis,shedevelopsaspecificX1/Yproposition)orwhere theoperativehypothesisorkeycontrolvariablechanges(anewcausalfactorisdiscoveredoranotheroutcome becomesthefocusofanalysis).Thingschange.Anditisthemarkofagoodresearchertokeephermindopento newevidenceandnewinsights.Toooften,methodologicaldiscussionsgivethemisleadingimpressionthat hypothesesareclearandremainfixedoverthecourseofastudy'sdevelopment.Nothingcouldbefurtherfromthe truth.Theunofficialtranscriptsofacademia—accessibleininformalsettings,whereresearcherslettheirguards down(particularlyifinebriated)—arefilledwithstoriesaboutdead‐ends,unexpectedfindings,anddrastically revisedtheorychapters.Itwouldbeinteresting,inthisvein,tocomparepublishedworkwithdissertation prospectusesandfellowshipapplications.Idoubtifthecorrelationbetweenthesetwostagesofresearchis particularlystrong. Research,afterall,isaboutdiscovery,notsimplytheverificationorfalsificationofstatichypotheses.Thatsaid,itis alsotruethatresearchonaparticulartopicshouldmovefromhypothesisgeneratingtohypothesis‐testing.This markstheprogressofa(p.678) field,andofascholar'sownwork.Asarule,researchthatbeginswithanopen‐ ended(X‐orY‐centered)analysisshouldconcludewithadeterminateX1/Yhypothesis. Theproblemisthatresearchstrategiesthatareidealforexplorationarenotalwaysidealforconfirmation.The extreme‐casemethodisinherentlyexploratorysincethereisnoclearcausalhypothesis;theresearcheris concernedmerelytoexplorevariationonasingledimension(XorY).Othermethodscanbeemployedineitheran open‐ended(exploratory)orahypothesis‐testing(confirmatory/disconfirmatory)mode.Thedifficultyisthatonce theresearcherhasarrivedatadeterminatehypothesistheoriginallychosenresearchdesignmaynolonger appeartobesowelldesigned. Thisisunfortunate,butinevitable.Onecannotconstructtheperfectresearchdesignuntil(a)onehasaspecific hypothesisand(b)oneisreasonablycertainaboutwhatoneisgoingtofind“outthere”intheempiricalworld.This isparticularlytrueofobservationalresearchdesigns,butitalsoappliestomanyexperimentalresearchdesigns: Usually,thereisa“good”(informative)finding,andafindingthatislessinsightful.Inshort,theperfectcase‐study researchdesignisusuallyapparentonlyexpostfacto. Therearethreewaystohandlethis.Onecanexplain,straightforwardly,thattheinitialresearchwasundertakenin anexploratoryfashion,andthereforenotconstructedtotestthespecifichypothesisthatis—now—theprimary argument.Alternatively,onecantrytoredesignthestudyafterthenew(orrevised)hypothesishasbeen formulated.Thismayrequireadditionalfieldresearchorperhapstheintegrationofadditionalcasesorvariables thatcanbeobtainedthroughsecondarysourcesorthroughconsultationofexperts.Afinalapproachistosimply Page 23 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques jettison,orde‐emphasize,theportionofresearchthatnolongeraddressesthe(revised)keyhypothesis.Athree‐ casestudymaybecomeatwo‐casestudy,andsoforth.Losttimeandeffortarethecostsofthisdownsizing. Intheevent,practicalconsiderationswillprobablydeterminewhichofthesethreestrategies,orcombinationsof strategies,istobefollowed.(Theyarenotmutuallyexclusive.)Thepointtorememberisthatrevisionofone's cross‐caseresearchdesignisnormalandperhapstobeexpected.Notalltwistsandturnsonthemeanderingtrail oftruthcanbeanticipated. 12AreThereOtherMethodsofCaseSelection? AttheoutsetofthischapterIsummarizedthetaskofcaseselectionasamatterofachievingtwoobjectives: representativeness(typicality)andvariation(causalleverage).Evidently,thereareotherobjectivesaswell.For example,onewishestoidentifycasesthatareindependentofeachother.Ifchosencasesareaffectedby(p. 679) eachother(sometimesknownasGalton'sproblemoraproblemofdiffusion),thisproblemmustbecorrected beforeanalysiscantakeplace.Ihaveneglectedthisissuebecauseitisusuallyapparenttotheresearcherand,in anycase,therearenosimpletechniquesthatmightbeutilizedtocorrectforsuchbiases.(Forfurtherdiscussion ofthisandotherfactorsimpinginguponcaseselectionseeGerring2001,178–81.) Ihavealsodisregardedpragmatic/logisticalissuesthatmightaffectcaseselection.Evidently,caseselectionis ofteninfluencedbyaresearcher'sfamiliaritywiththelanguageofacountry,apersonalentréeintothatlocale, specialaccesstoimportantdata,orfundingthatcoversonearchiveratherthananother.Pragmaticconsiderations areoften—andquiterightly—decisiveinthecase‐selectionprocess. Afinalconsiderationconcernsthetheoreticalprominenceofaparticularcasewithintheliteratureonasubject. Researchersaresometimesobligedtostudycasesthathavereceivedextensiveattentioninpreviousstudies. Thesearesometimesreferredtoas“paradigmatic”casesor“exemplars”(Flyvbjerg2004,427). However,neitherpragmatic/logisticalutilitynortheoreticalprominencequalifiesasamethodologicalfactorincase selection.Thatis,thesefeaturesofacasehavenobearingonthevalidityofthefindingsstemmingfromastudy. Assuch,itisappropriatetogranttheseissuesaperipheralstatusinthischapter. Onefinalcaveatmustbeissued.Whileitistraditionaltodistinguishamongthetasksofcaseselectionandcase analysis,acloselookattheseprocessesshowsthemtobeindistinctandoverlapping.Onecannotchooseacase withoutconsideringthesortofanalysisthatitmightbesubjectedto,andviceversa.Thus,thereadershould considerchoosingcasesbyemployingtheninetechniqueslaidoutinthischapteralongwithanyconsiderations thatmightbeintroducedbyvirtueofacase'squasi‐experimentalqualities,atopictakenupelsewhere(Gerring 2007,ch.6). References ABADIE,A.,DRUKKER,D.,HERR,J.L.,andIMBENS ,G.W.2001.Implementingmatchingestimatorsforaveragetreatment effectsinStata.StataJournal,1:1–18. ABBOTT,A.2001.TimeMatters:OnTheoryandMethod.Chicago:UniversityofChicagoPress. —— andTSAY,A.2000.Sequenceanalysisandoptimalmatchingmethodsinsociology.SociologicalMethodsand Research,29:3–33. —— andFORREST,J.1986.Optimalmatchingmethodsforhistoricalsequences.JournalofInterdisciplinaryHistory, 16:471–94. ACHEN,C.H.,andSNIDAL ,D.1989.Rationaldeterrencetheoryandcomparativecasestudies.WorldPolitics,41: 143–69. ALLEN,W.S.1965.TheNaziSeizureofPower:TheExperienceofaSingleGermanTown,1930–1935.NewYork: Watts. Page 24 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques ALMOND,G.A.1956.Comparativepoliticalsystems.JournalofPolitics,18:391–409. (p.680) AMENTA,E.1991.Makingthemostofacasestudy:theoriesofthewelfarestateandtheAmerican experience.Pp.172–94inIssuesandAlternativesinComparativeSocialResearched.C.C.Ragin.Leiden:E.J. Brill. BARRO,R.J.1999.Determinantsofdemocracy.JournalofPoliticalEconomy,107:158–83. BELSEY,D.A.,KUH,E.,andWELSCH,R.E.2004.RegressionDiagnostics:IdentifyingInfluentialDataandSourcesof Collinearity.NewYork:Wiley. BENNETT,A.,LEPGOLD,J.,andUNGER,D.1994.Burden‐sharinginthePersianGulfWar.InternationalOrganization,48: 39–75. BENTLEY,A.1908/1967.TheProcessofGovernment.Cambridge,Mass.:HarvardUniversityPress. BRADY,H.E.,andCOLLIER,D.(eds.)2004.RethinkingSocialInquiry:DiverseTools,SharedStandards.Lanham,Md.: RowmanandLittlefield. BRAUMOELLER,B.F.2003.Causalcomplexityandthestudyofpolitics.PoliticalAnalysis,11:209–33. BREMAN,A.,andSHELTON,C.2001.Structuraladjustmentandhealth:aliteraturereviewofthedebate,itsrole‐players andpresentedempiricalevidence.CMHWorkingPaperSeries,PaperNo.WG6:6.WHO,Commissionon MacroeconomicsandHealth. BRENNER,R.1976.Agrarianclassstructureandeconomicdevelopmentinpre‐industrialEurope.PastandPresent, 70:30–75. BROWNE,A.1987.WhenBatteredWomenKill.NewYork:FreePress. BUCHBINDER,S.,andVITTINGHOFF,E.1999.HIV‐infectedlong‐termnonprogressors:epidemiology,mechanismsof delayedprogression,andclinicalandresearchimplications.MicrobesInfect,1:1113–20. COHEN,M.R.,andNAGEL ,E.1934.AnIntroductiontoLogicandScientificMethod.NewYork:Harcourt,Braceand Company. COLLIER,D.,andMAHONEY,J.1996.Insightsandpitfalls:selectionbiasinqualitativeresearch.WorldPolitics,49:56– 91. COLLIER,R.B.,andCOLLIER,D.1991/2002.ShapingthePoliticalArena:CriticalJunctures,theLaborMovement,and RegimeDynamicsinLatinAmerica.NotreDame,Ind.:UniversityofNotreDamePress. COLOMER,J.M.1991.Transitionsbyagreement:modelingtheSpanishway.AmericanPoliticalScienceReview,85: 1283–302. CONVERSE,P.E.,andDUPEUX,G.1962.PoliticizationoftheelectorateinFranceandtheUnitedStates.PublicOpinion Quarterly,16:1–23. COPPEDGE,M.J.2004.TheconditionalimpactoftheeconomyondemocracyinLatinAmerica.Presentedatthe conference“DemocraticAdvancementsandSetbacks:WhatHaveWeLearnt?”,UppsalaUniversity,June11–13. DEFELICE,E.G.1986.Causalinferenceandcomparativemethods.ComparativePoliticalStudies,19:415–37. DESCH,M.C.2002.Democracyandvictory:whyregimetypehardlymatters.InternationalSecurity,27:5–47. DEYO,F.(ed.)1987.ThePoliticalEconomyoftheNewAsianIndustrialism.Ithaca,NY:CornellUniversityPress. DION,D.1998.Evidenceandinferenceinthecomparativecasestudy.ComparativePolitics,30:127–45. ECKSTEIN,H.1975.Casestudiesandtheoryinpoliticalscience.InHandbookofPoliticalScience,vii:Political Science:ScopeandTheory,ed.F.I.GreensteinandN.W.Polsby.Reading,Mass.:Addison‐Wesley. Page 25 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques (p.681) EGGAN,F.1954.Socialanthropologyandthemethodofcontrolledcomparison.AmericanAnthropologist, 56:743–63. ELMAN,C.2003.LessonsfromLakatos.InProgressinInternationalRelationsTheory:AppraisingtheField,ed.C. ElmanandM.F.Elman.Cambridge,Mass.:MITPress. —— 2005.Explanatorytypologiesinqualitativestudiesofinternationalpolitics.InternationalOrganization,59: 293–326. EMIGH,R.1997.Thepowerofnegativethinking:theuseofnegativecasemethodologyinthedevelopmentof sociologicaltheory.TheoryandSociety,26:649–84. EPSTEIN,L.D.1964.AcomparativestudyofCanadianparties.AmericanPoliticalScienceReview,58:46–59. ERTMAN,T.1997.BirthoftheLeviathan:BuildingStatesandRegimesinMedievalandEarlyModernEurope. Cambridge:CambridgeUniversityPress. ESPING‐ ANDERSEN,G.1990.TheThreeWorldsofWelfareCapitalism.Princeton,NJ:PrincetonUniversityPress. FLYVBJERG,B.2004.Fivemisunderstandingsaboutcase‐studyresearch.Pp.420–34inQualitativeResearch Practice,ed.C.Seale,G.Gobo,J.F.Gubrium,andD.Silverman.London:Sage. GEDDES ,B.1990.Howthecasesyouchooseaffecttheanswersyouget:selectionbiasincomparativepolitics.In PoliticalAnalysis,vol.ii,ed.J.A.Stimson.AnnArbor:UniversityofMichiganPress. —— 2003.ParadigmsandSandCastles:TheoryBuildingandResearchDesigninComparativePolitics.Ann Arbor:UniversityofMichiganPress. GEORGE,A.L.,andBENNETT,A.2005.CaseStudiesandTheoryDevelopment.Cambridge,Mass.:MITPress. —— andSMOKE,R.1974.DeterrenceinAmericanForeignPolicy:TheoryandPractice.NewYork:Columbia UniversityPress. GERRING,J.2001.SocialScienceMethodology:ACriterialFramework.Cambridge:CambridgeUniversityPress. —— 2007.CaseStudyResearch:PrinciplesandPractices.Cambridge:CambridgeUniversityPress. —— THACKER,S.andMORENO,C.2005.Doneoliberalpoliciessavelives?Unpublishedmanuscript. GOERTZ,G.andSTARR,H.(eds.)2003.NecessaryConditions:Theory,MethodologyandApplications.NewYork: RowmanandLittlefield. —— andLEVY,J.(eds.)forthcoming.Causalexplanations,necessaryconditions,andcasestudies:WorldWarIand theendoftheColdWar.Manuscript. GOODIN,R.E.andSMITSMAN,A.2000.Placingwelfarestates:theNetherlandsasacrucialtestcase.Journalof ComparativePolicyAnalysis,2:39–64. GUJARATI,D.N.2003.BasicEconometrics,4thedn.NewYork:McGraw‐Hill. HAMILTON,G.G.1977.Chineseconsumptionofforeigncommodities:acomparativeperspective.American SociologicalReview,42:877–91. HAYNES ,B.F.PANTALEO,G.andFAUCI,A.S.1996.Towardanunderstandingofthecorrelatesofprotectiveimmunityto HIVinfection.Science,271:324–8. HEMPEL ,C.G.1942.Thefunctionofgenerallawsinhistory.JournalofPhilosophy,39:35–48. HO,D.E.IMAI,K.KING,G.andSTUART,E.A.2004.Matchingasnonparametricpreprocessingforreducingmodel dependenceinparametriccausalinference.Manuscript. HOWARD,M.M.2003.TheWeaknessofCivilSocietyinPost‐CommunistEurope.Cambridge:CambridgeUniversity Page 26 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques Press. (p.682) HOWSON,C.andURBACH,P.1989.ScientificReasoning:TheBayesianApproach.LaSalle,Ill.:OpenCourt. HUMPHREYS ,M.2005.Naturalresources,conflict,andconflictresolution:uncoveringthemechanisms.Journalof ConflictResolution,49:508–37. JENICEK,M.2001.ClinicalCaseReportinginEvidence‐BasedMedicine,2ndedn.Oxford:OxfordUniversityPress. KARL ,T.L.1997.TheParadoxofPlenty:OilBoomsandPetro‐states.Berkeley:UniversityofCaliforniaPress. KAZANCIGIL ,A.1994.Thedeviantcaseincomparativeanalysis:highstatenessincomparativeanalysis.Pp.213–38 inComparingNations:Concepts,Strategies,Substance,ed.M.DoganandA.Kazancigil.Cambridge:Blackwell. KEMP,K.A.1986.Race,ethnicity,classandurbanspatialconflict:ChicagoasacrucialcaseUrbanStudies,23: 197–208. KENDALL ,P.L.andWOLF,K.M.1949/1955.Theanalysisofdeviantcasesincommunicationsresearch.In CommunicationsResearch,1948–1949,ed.P.F.LazarsfeldandF.N.Stanton.NewYork:HarperandBrothers. Reprintedaspp.167–70inTheLanguageofSocialResearch,ed.P.F.LazarsfeldandM.Rosenberg.NewYork: FreePress. KENNEDY,C.H.2005.Single‐caseDesignsforEducationalResearch.Boston:AllynandBacon. KENNEDY,P.2003.AGuidetoEconometrics,5thedn.Cambridge,Mass.:MITPress. KHONG,Y.F.1992.AnalogiesatWar:Korea,Munich,DienBienPhu,andtheVietnamDecisionsof1965.Princeton, NJ:PrincetonUniversityPress. KING,G.KEOHANE,R.O.andVERBA,S.1994.DesigningSocialInquiry:ScientificInferenceinQualitativeResearch. Princeton,NJ:PrincetonUniversityPress. LAKATOS ,I.1978.TheMethodologyofScientificResearchProgrammes.Cambridge:CambridgeUniversityPress. LAZARSFELD,P.F.andBARTON,A.H.1951.Qualitativemeasurementinthesocialsciences:classification,typologies, andindices.InThePolicySciences,ed.D.LernerandH.D.Lass‐well.Stanford,Calif.:StanfordUniversityPress. LEVY,J.S.2002.Qualitativemethodsininternationalrelations.InEvaluatingMethodologyinInternationalStudies, ed.F.P.HarveyandM.Brecher.AnnArbor:UniversityofMichiganPress. LIJPHART,A.1968.ThePoliticsofAccommodation:PluralismandDemocracyintheNetherlands.Berkeley: UniversityofCaliforniaPress. —— 1969.Consociationaldemocracy.WorldPolitics,21:207–25. —— 1971.Comparativepoliticsandthecomparativemethod.AmericanPoliticalScienceReview,65:682–93. —— 1975.Thecomparablecasesstrategyincomparativeresearch.ComparativePoliticalStudies,8:158–77. LIPSET,S.M.1959.Somesocialrequisitesofdemocracy:economicdevelopmentandpoliticaldevelopment. AmericanPoliticalScienceReview,53:69–105. —— 1960/1963.PoliticalMan:TheSocialBasesofPolitics.GardenCity,NY:Anchor. —— 1968.AgrarianSocialism:TheCooperativeCommonwealthFederationinSaskatchewan.AStudyinPolitical Sociology.GardenCity,NY:Doubleday. —— TROW,M.A.andCOLEMAN,J.S.1956.UnionDemocracy:TheInternalPoliticsoftheInternational TypographicalUnion.NewYork:FreePress. LYND,R.S.andLYND,H.M.1929/1956.Middletown:AStudyinAmericanCulture.NewYork:Harcourt,Brace. Page 27 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques MAHONEY,J.andGOERTZ,G.2004.Thepossibilityprinciple:choosingnegativecasesincomparativeresearch. AmericanPoliticalScienceReview,98:653–69. (p.683) MARTIN,L.L.1992.CoerciveCooperation:ExplainingMultilateralEconomicSanctions.Princeton,NJ: PrincetonUniversityPress. MAYO,D.G.1996.ErrorandtheGrowthofExperimentalKnowledge.Chicago:UniversityofChicagoPress. MECKSTROTH,T.1975.“Mostdifferentsystems”and“mostsimilarsystems:”astudyinthelogicofcomparative inquiry.ComparativePoliticalStudies,8:133–77. MIGUEL ,E.2004.Tribeornation:nation‐buildingandpublicgoodsinKenyaversusTanzania.WorldPolitics,56: 327–62. MILL ,J.S.1843/1872.TheSystemofLogic,8thedn.London:Longmans,Green. MONROE,K.R.1996.TheHeartofAltruism:PerceptionsofaCommonHumanity.Princeton,NJ:PrincetonUniversity Press. MOORE,B.,JR.1966.SocialOriginsofDictatorshipandDemocracy:LordandPeasantintheMakingoftheModern World.Boston:BeaconPress. MORGAN,S.L.andHARDING,D.J.2005.Matchingestimatorsofcausaleffects:fromstratificationandweightingto practicaldataanalysisroutines.Manuscript. MOULDER,F.V.1977.Japan,ChinaandtheModernWorldEconomy:TowardaReinterpretationofEastAsian Developmentca.1600toca.1918.Cambridge:CambridgeUniversityPress. MUNCK,G.L.2004.Toolsforqualitativeresearch.Pp.105–21inRethinkingSocialInquiry:DiverseTools,Shared Standards,ed.H.E.BradyandD.Collier.Lanham,Md.:RowmanandLittlefield. NJOLSTAD,O.1990.Learningfromhistory?Casestudiesandthelimitstotheory‐building.Pp.220–46inArmsRaces: TechnologicalandPoliticalDynamics,ed.O.Njolstad.ThousandOaks,Calif.:Sage. PATTON,M.Q.2002.QualitativeEvaluationandResearchMethods.NewburyPark,Calif.:Sage. POPPER,K.1934/1968.TheLogicofScientificDiscovery.NewYork:HarperandRow. —— 1963.ConjecturesandRefutations.London:RoutledgeandKeganPaul. POSNER,D.2004.Thepoliticalsalienceofculturaldifference:whyChewasandTumbukasarealliesinZambiaand adversariesinMalawi.AmericanPoliticalScienceReview,98:529–46. PRZEWORSKI,A.andTEUNE,H.1970.TheLogicofComparativeSocialInquiry.NewYork:JohnWiley. QUEEN,S.1928.Roundtableonthecasestudyinsociologicalresearch.PublicationsoftheAmericanSociological Society,PapersandProceedings,22:225–7. RAGIN,C.C.2000.Fuzzy‐setSocialScience.Chicago:UniversityofChicagoPress. —— 2004.Turningthetables.Pp.123–38inRethinkingSocialInquiry:DiverseTools,SharedStandards,ed.H.E. BradyandD.Collier.Lanham,Md.:RowmanandLittlefield. REILLY,B.2000–1.Democracy,ethnicfragmentation,andinternalconflict:confusedtheories,faultydata,andthe “crucialcase”ofPapuaNewGuinea.InternationalSecurity,25:162–85. —— andPHILLPOT,R.2003.“Makingdemocracywork”inPapuaNewGuinea:socialcapitalandprovincial developmentinanethnicallyfragmentedsociety.AsianSurvey,42:906–27. ROGOWSKI,R.1995.Theroleoftheoryandanomalyinsocial‐scientificinference.AmericanPoliticalScience Review,89:467–70. Page 28 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques ROHLFING,I.2004.Haveyouchosentherightcase?Uncertaintyincaseselectionforsinglecasestudies.Working Paper,InternationalUniversity,Bremen. ROSENBAUM,P.R.2004.Matchinginobservationalstudies.InAppliedBayesianModelingandCausalInferencefrom anIncomplete‐dataPerspective,ed.A.GelmanandX.‐L.Meng.NewYork:JohnWiley. —— andSILBER,J.H.2001.Matchingandthickdescriptioninanobservationalstudyofmortalityaftersurgery. Biostatistics,2:217–32. ROSS ,M.2001.Doesoilhinderdemocracy?WorldPolitics,53:325–61. (p.684) SAGAN,S.D.1995.LimitsofSafety:Organizations,Accidents,andNuclearWeapons.Princeton,NJ: PrincetonUniversityPress. SEKHON,J.S.2004.Qualitymeetsquantity:casestudies,conditionalprobabilityandcounter‐factuals.Perspectives inPolitics,2:281–93. SHAFER,M.D.1988.DeadlyParadigms:TheFailureofU.S.CounterinsurgencyPolicy.Princeton,NJ:Princeton UniversityPress. SKOCPOL ,T.1979.StatesandSocialRevolutions:AComparativeAnalysisofFrance,Russia,andChina. Cambridge:CambridgeUniversityPress. —— andSOMERS ,M.1980.Theusesofcomparativehistoryinmacrosocialinquiry.ComparativeStudiesinSociety andHistory,22:147–97. STINCHCOMBE,A.L.1968.ConstructingSocialTheories.NewYork:Harcourt,Brace. SWANK,D.H.2002.GlobalCapital,PoliticalInstitutions,andPolicyChangeinDevelopedWelfareStates. Cambridge:CambridgeUniversityPress. TENDLER,J.1997.GoodGovernmentintheTropics.Baltimore:JohnsHopkinsUniversityPress. TRUMAN,D.B.1951.TheGovernmentalProcess.NewYork:AlfredA.Knopf. TSAI,L.2007.AccountabilitywithoutDemocracy:HowSolidaryGroupsProvidePublicGoodsinRuralChina. Cambridge:CambridgeUniversityPress. VANEVERA,S.1997.GuidetoMethodsforStudentsofPoliticalScience.Ithaca,NY:CornellUniversityPress. WAHLKE,J.C.1979.Pre‐behavioralisminpoliticalscience.AmericanPoliticalScienceReview,73:9–31. YASHAR,D.J.2005.ContestingCitizenshipinLatinAmerica:TheRiseofIndigenousMovementsandthe PostliberalChallenge.Cambridge:CambridgeUniversityPress. YIN,R.K.2004.CaseStudyAnthology.ThousandOaks,Calif.:Sage. Notes: (1)Gujarati(2003);Kennedy(2003).Interestingly,thepotentialofcross‐casestatisticsinhelpingtochoosecases forin‐depthanalysisisrecognizedinsomeoftheearliestdiscussionsofthecase‐studymethod(e.g.Queen1928, 226). (2)ThisexpandsonMill(1843/1872,253),whowroteofscientificenquiryastwofold:“eitherinquiriesintothe causeofagiveneffectorintotheeffectsorpropertiesofagivencause.” (3)Thismethodhasnotreceivedmuchattentiononthepartofqualitativemethodologists;hence,theabsenceofa generallyrecognizedname.ItbearssomeresemblancetoJ.S.Mill'sJointMethodofAgreementandDifference(Mill 1843/1872),whichistosayamixtureofmost‐similarandmost‐differentanalysis,asdiscussedbelow.Patton (2002,234)employstheconceptof“maximumvariation(heterogeneity)sampling.” Page 29 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques (4)Moreprecisely,GeorgeandSmoke(1974,534,522–36,ch.18;seealsodiscussioninCollierandMahoney 1996,78)setouttoinvestigatecausalpathwaysanddiscovered,throughthecourseoftheirinvestigationofmany cases,thesethreecausaltypes.Yet,forourpurposeswhatisimportantisthatthefinalsampleincludesatleast onerepresentativeofeach“type.” (5)ForfurtherexamplesseeCollierandMahoney(1996);Geddes(1990);Tendler(1997). (6)Traditionally,methodologistshaveconceptualizedcasesashaving“positive”or“negative”values(e.g.Emigh 1997;MahoneyandGoertz2004;Ragin2000,60;2004,126). (7)Geddes(1990);King,Keohane,andVerba(1994).SeealsodiscussioninBradyandCollier(2004);Collierand Mahoney(1996);Rogowski(1995). (8)Theexceptionwouldbeacircumstanceinwhichtheresearcherintendstodisproveadeterministicargument (Dion1998). (9)Geddes(2003,131).Forotherexamplesofcaseworkfromtheannalsofmedicinesee“Clinicalreports”inthe Lancet,“Casestudies”inCanadianMedicalAssociationJournal,andvariousissuesoftheJournalofObstetrics andGynecology,oftendevotedtoclinicalcases(discussedinJenicek2001,7).Forexamplesfromthesubfieldof comparativepoliticsseeKazancigil(1994). (10)ForadiscussionoftheimportantroleofanomaliesinthedevelopmentofscientifictheorizingseeElman (2003);Lakatos(1978).Forexamplesofdeviant‐caseresearchdesignsinthesocialsciencesseeAmenta(1991); Coppedge(2004);Eckstein(1975);Emigh(1997);KendallandWolf(1949/1955). (11)Forexamplesofthecrucial‐casemethodseeBennett,Lepgold,andUnger(1994);Desch(2002);Goodinand Smitsman(2000);Kemp(1986);ReillyandPhillpot(2003).ForgeneraldiscussionseeGeorgeandBennett(2005); Levy(2002);Stinchcombe(1968,24–8). (12)Athirdposition,whichpurportstobeneitherPopperianorBayesian,hasbeenarticulatedbyMayo(1996,ch. 6).Fromthisperspective,thesameideaisarticulatedasamatterof“severetests.” (13)ItshouldbenotedthatTsai'sconclusionsdonotrestsolelyonthiscrucialcase.Indeed,sheemploysabroad rangeofmethodologicaltools,encompassingcase‐studyandcross‐casemethods. (14)SeealsothediscussioninEckstein(1975)andLijphart(1969).Foradditionalexamplesofcasestudies disconfirminggeneralpropositionsofadeterministicnatureseeAllen(1965);Lipset,Trow,andColeman(1956); Njolstad(1990);Reilly(2000–1);anddiscussioninDion(1998);Rogowski(1995). (15)Granted,insofarascase‐studyanalysisprovidesawindowintocausalmechanisms,andcausalmechanisms areintegraltoagiventheory,asinglecasemaybeenlistedtoconfirmordisconfirmaproposition.However,ifthe casestudyupholdsapositedpatternofX/Ycovariation,andfindsfaultonlywiththestipulatedcausalmechanism, itwouldbemoreaccuratetosaythatthestudyforcesthereformulationofagiventheory,ratherthanits confirmationordisconfirmation.Seefurtherdiscussioninthefollowingsection. (16)Sometimes,themost‐similarmethodisknownasthe“methodofdifference,”afteritsinventor(Mill1843/1872). ForlatertreatmentsseeCohenandNagel(1934);Eggan(1954);Gerring(2001,ch.9);Lijphart(1971;1975); Meckstroth(1975);PrzeworskiandTeune(1970);SkocpolandSomers(1980). (17)ForgoodintroductionsseeHoetal.(2004);MorganandHarding(2005);Rosenbaum(2004);Rosenbaumand Silber(2001).ForadiscussionofmatchingproceduresinStataseeAbadieetal.(2001). (18)Themost‐differentmethodisalsosometimesreferredtoasthe“methodofagreement,”followingitsinventor, J.S.Mill(1843/1872).SeealsoDeFelice(1986);Gerring(2001,212–14);Lijphart(1971;1975);Meckstroth(1975); PrzeworskiandTeune(1970);SkocpolandSomers(1980).ForexamplesofthismethodseeCollierandCollier (1991/2002);ConverseandDupeux(1962);Karl(1997);Moore(1966);Skocpol(1979);Yashar(2005,23). However,mostofthesestudiesaredescribedascombiningmost‐similarandmost‐differentmethods. (19)InthefollowingdiscussionItreatthetermssocialcapital,civilsociety,andcivicengagementinterchangeably. Page 30 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015 Case Selection for Case‐Study Analysis: Qualitative and Quantitative Techniques (20)E.g.CollierandCollier(1991/2002);Karl(1997);Moore(1966);Skocpol(1979);Yashar(2005,23).Karl (1997),whichaffectstobeamost‐differentsystemanalysis(20),isaparticularlyclearexampleofthis.Herstudy, focusedostensiblyonpetro‐states(stateswithlargeoilreserves),makestwosortsofinferences.Thefirst concernsthe(usually)obstructiveroleofoilinpoliticalandeconomicdevelopment.Thesecondsortofinference concernsvariationwithinthepopulationofpetro‐states,showingthatsomecountries(e.g.Norway,Indonesia) managetoavoidthepathologiesbroughtonelsewherebyoilresources.Whenattemptingtoexplainthe constrainingroleofoilonpetro‐states,Karlusuallyreliesoncontrastsbetweenpetro‐statesandnonpetro‐states (e.g.ch.10).Onlywhenattemptingtoexplaindifferencesamongpetro‐statesdoessherestricthersampleto petro‐states.Inmyopinion,verylittleuseismadeofthemost‐differentresearchdesign. (21)Thiswasrecognized,atleastimplicitly,byMill(1843/1872,258–9).Skepticismhasbeenechoedby methodologistsintheinterveningyears(e.g.CohenandNagel1934,251–6;Gerring2001;SkocpolandSomers 1980).Indeed,explicitdefensesofthemost‐differentmethodarerare(butseeDeFelice1986). (22)AnotherwayofstatingthisistosaythatXisa“nontrivialnecessarycondition”ofY. (23)Wahlke(1979,13)writesofthefailingsofthe“behavioralist”modeofpoliticalscienceanalysis:“Itrarelyaims atgeneralization;researcheffortshavebeenconfinedessentiallytocasestudiesofsinglepoliticalsystems,most ofthemdealing…withtheAmericansystem.” JohnGerring JohnGerringisProfessorofPoliticalScience,BostonUniversity. Page 31 of 31 PRINTED FROM OXFORD HANDBOOKS ONLINE (www.oxfordhandbooks.com). (c) Oxford University Press, 2015. All Rights Reserved. Under the terms of the licence agreement, an individual user may print out a PDF of a single chapter of a title in Oxford Handbooks Online for personal use (for details see Privacy Policy). Subscriber: Minnesota State University, Mankato; date: 31 July 2015