George and McKeown2

7/27/2019 George and McKeown2

1/38

F. h",ff' ('y'"t ";ntl*

CASE STUDIESAND THEORIESOF ORGANIZATIONALDECISIONMAKING

Alexander L. Georgeand Timothy J. McKeown

I. INTRODUCTIONln spite of fluctuations n their fashionability,caspstudies oforganizational ecisionmakinghave ong beenone of the mostimportant methods by which researchershave investigatedorganizationalbehavior and improved their theoreticalunder-standingof that behavior.For example,our understanding fAmerican oreignpolicydecisionmakinghasbeenconsiderablyenrichedn thepast 5 years ycase tudieshathave lucidatedheAdvsncesn InformationProcessingn Organizetions, ol. 2, peges l-5tCopyrightCI 1985by JAt Pressnc.All rights of reproductionn enyform reserved.ISBN: U89232-42U2


2/38

r22 A. L. GEORGEand T. J. McKEOM/Nimportance of organizational information-processing ystems '(Allison, l97l); limitations on human information-processingcapabilitiesSteinbruner, 974);and bargainingNeustadt,1970;Paige,1968). The role of case tudiesn the development f our 'understanding f decisionmaking in businessirms is equallyimportant.)Theseaccomplishmentseem ll the more emarkablewhen we consider hat casestudieshave ong been egardedbymany socialscientists ith suspicion. o someone ho views helogic of explanationsolely n statisticalor quasi-experimentalterms, he epistemologicaltrategyembodiedn case tudies eemsfundamentallywrong-headed.ven hosewho arewillingto grantthat theremight besomething f value n case tudiesmayhesitateto usecase tudiesaspart of a research rogram,becauseherearefew methodolocialguidelinescomparable o those that exist inexperimentaland quasi-experimentalesearchThis situation suggestswo needs.First, it is important toexplicate he underlying epistemological trategyof casestudies.This "reconstructed ogic" (Kaplan, 1964) s useful for makingsense f case tudies o a generation f scholarsrained o think ofexplanationsolely n statistical erms. t also helpsus to be moreself-criticalabout the casestudies we produce,by making ourexpectationsabout what they can and cannot accomplishmoreexplicit. Second,we attempt to develop a useful and explicitmethodology or case tudies. his methodologys consistentwithour "reconstructedogic" of the case-studyrocess, nd s alsoanattempt o codify thewisdomof practitioners.We present flexiblebut disciplined and rather self-consciousmethodology-themethod of "Qt -uctured;ocused comparison" (George, 1979a;1979b;1982)hat hasbeenused.in.gese"prchrogramshat rely oncasestudies. n what follows ltggbt discuss mportant contrastsbetween ase tudies ndquaslfrp"timentdqF hq4s.we(fiil )consider severaldifferentapproacheso the use of casestudies,with each successive pproach representinga more dramaticdeparture from the logic of quasi-experimentalesearch.This isfollowed by a preserrl,ationf the methodof structured, ocusedcomparison.Finally, we speculate n the evolutionof researchprograms elyingon case tudies.


3/38

- ,_1

Cnse tudies ndTheoriesfOrganizationalecision akingII. SOMECONTRASTS ETWEEN ASESruDIES

AND QUASI-EXPERIMENTALETHODSOur startingpoint is the claim hat observationsre theory aden,thathumanbeingsmpose ome onceptual cheme n theirsensoryinputsso as o bring order out of chaos. his position s by now afamiliar one in the philosophyof science.lWe will not disputeDavid Hume'swell-knownclaim that causation annotbe nferredfrom "mere" observation.However, f observation s somehowinformedby a theoryor pre-theory, hen t is logicallypossibleorthe observer to integrateobservationswith presuppositionsnorder o arriveat causalnferences.ndeed,he actualbehaviorofhumans n everyday ituations mounts ojust this(Nisbett& Ross,1980). he difficulty s that theproblemof drawingcorrectcausalinferences s eSsentially he same or researchers s for everyoneelse;without any methodological afeguards, ne would expectcase-study esearcherso make the same sorts of errors thatpeoplemake n their everyday nferences.The standardquasi-experimentalpproach o this problem s toreducehe ikelihood of erroneousnferencesy makingsystematic.comparisons f varianceon the relevantvariablesacrossa largenumber of cases.By using statistical echniques ne can analyzethecovariance f variables nd therebydeterminewhethera causalpatternattributedto a givenclassof situations s ndeedconsistentwith a setof observations. owever, n some ituations his methodis either uneconomicalor simply not feasible.When performingcasestudies, ne s typically working with sucha small number ofcaseshat conventional tatisticalanalysiss not possibleor useful.Although onecan o someextentmitigate he difficultiesof too fewcasesimplybygenerating orecasessee ectionII in thispaper),it is typically true that the number of variables nvolved greatlyexceedshe number of cases. hus, additionalcases oupled osystematic nalysisof variations n variablesacrosscases rovideonly a modicumof protectionagainsterroneous ausal nferences.The case-studymethod differs from quasi-experimentalap-proachesn its heavy relianceon within-cese nalysisas a way ofevaluatingclaims about causalprocesses. uch a method relies

23


4/38

r24 A. L. GEORGE ndT. J.McKEOWNonly to a limitedextenton comparing ariations n variables the"congruence"method, iscussedn SectionV); morecentrals thelogicof "processracing" Section ),whichdoes ot relysolely ncomparing variations in order to make claims about causalprocessest work.A second istinction etween uasi-experimentalndcase-studyapproachesies n their treatmentof thedistinctionbetweenheoryformationand heory esting.Theconventional uasi-experimentalposition is founded on a clear-cutdistinctionbetween hese woactivities.2Theoryformationis subjecthat normally s ruled "outof bounds"-statistical analysis, or example, s about testinghypotheses, ot about generatinghem.3By contrast, he case-study techniqueoften entailsviewinghypothesisormation as anobjectiveof the study. n case tudies, cleardemarcation etweenhypothesisformation and hypothesis esting often is absent,because he researchprocessoften involves terative cycling inwhich hypothesesare successively fitted" to observations.Problemswith the fit then leadto revisionof the hypotheses ndmay also ncitesearchesor additionaldata.Theway n whichcasestudiesgenerallyare reported n journals and monographsdoesnot reveal hat such a process ccurs.Both of theseconceptualdistinctionsbetween asestudiesandmore conventionalquasi-experimental pproachessuggest hatdiscussing asestudiesas if they wereconventionalquasi-experi-mental designshandicappedby absurdly ow Ns is neither veryaccuratenor very helpful n evaluatingheir usefulness. owever,becausehe imited amount of reflectionabout the relationof casestudies to theory developmedtoften proceeds rom just thisposition,we beginour consideration f variousstrategiesor usingcase tudieswith thosesuggestionshat raise he fewest hallengesto the conventionalquasi-experimentaliew of research.

III. STRATEGIESFORPERFORMINGCASESTLIDIES:CONTROLLEDCOMPARISONone strategyntended o strengthenhe amountof confidence ecan attach o causal nferencesn case tudiess to use he com-


5/38

** l

CaseStudies nd TheoriesfOrganizational ecision aking 25parativemethod. Within the domain of interest, he researcherselects numberof casesor analysis, aking care hat the casesdiffer rom one anotheron somedimensionsudgedsignificant ythe researcher. he researcheran then compareoutcomesn thevariouscasesand attempt to relate he occurrence f differingoutcomeso the presencef differing nitial conditions,or relatecommonoutcomes o the presence f common nitial conditions.As Lijphart (1971)argues, The comparativemethod resemblesthe statisticalmethod n all respectsut one.Thecrucialdifferenceis that the number of casest dealswith is too small to permitsystematic ontrol by meansof partial correlations" p. 684).From such a startingpoint Lijphart (1971)considers everalstrategiesor minimizing he weaknessesf th is approach:

l. Increasinghe numberof cases2. Reducing he numberof variables onsidered3. Focusingon cases hat are highly "comparable"-that is,cases hat possess ighly similar initial conditions,so thatwhateverdifferencesn outcomes reobserved anpotentiallybe explained n termsof those ewdimensions n which nitialconditionsare not identical4. Working with and attempting o construct elativelysimpletheories f few variables 685-691).

Attempts to improve comparisonssimply by increasingthenumber of cases tudiedcan hardly be criticizedon logical ormethodological rounds,but case tudies s hat term traditionallyis understoodare rathef lengthy, costly affairs. It is thereforeunlikely hat thenumberof cases anbe enlargedo thepoint whereconventional tatistical nalysiswould be convincing.Lijphart also suggestsmaking comparisonsmore tractable byreducing he number of variablesunder consideration, y usingsimple rather than complex theoretical argumentsand by usingcoarse- ather than fine-grained artitioningsof variables.Thesesuggestionsremotivatedby thedesireo create ross-classificationschemes ufficientlysimple so that evena small number of casesleavesew empty cells.Again, one can hardly argue hat simpletheoriesare bad theories,or that eliminatingempty cells s not a


6/38

-r26 A. L. GEORGE nd T. I. McKEOWNuseful idea. However, we must also note that the case-studyresearcher ften has a clinician'snterest n developing heoriesthat are useful or makingdiscriminatingdiagnoses;uch nterestsare not well servedby analyzingcasesn terms of such simpletheories.This strategy, ike the previousone, probably will beinadequatef the objective s to create ituationswhereorthodoxstatistical nalysis anproceed.Smelser1968)presents n analysis f the comparativemethodthat in manyrespectsoreshadows ijpharts's.He argues hat thecomparativeresearcher an treat some sourcesof variance inoutcomesas parametersconstants), y choosingcases n whichthe values or thosevariablesare the sameacross ases.Whatevervariablesare not constantacrosscases re then the "operative"(independent)variables. Smelser also offers a suggestionnotmentionedby Lijphart. If one begins esearchwith some heory,often one can identify behavioral mplicationsof the theory in avarietyof domainsother han the one hat originallywas he primeconcernof the researcher.Replicationat differentanalytic evels"(investigationof the empiricalaccuracyof a theory in differentsettings) an"increase r decreasehe nvestigator'sonfidencen asuspected ssociation etween onditionsand the phenomenonobe explained"(Smelser, 968;p. 73).SmelserconsidersDurk-heim's study of suicideamong hemilitary as an excellent xampleof this approach.Diesing(1971)presentswo practicalsuggestionsor ensuringthat casesare as comparableas possibleevenwhen there is notheory hat tellsuson whatdiminensions imilarity s to besought:(l) select aseswithin similargeographicalontexts,and (2) selectcases hat have common histories pp. 187-188). A third sug-gestionof Diesingwill be discussedn a subsequentectionof thispaper).Diesingalsonotes he drawbacksof his own suggestions.The geographicmethod s vulnerable o the fact that contiguouscases re not necessarilyimilarat all; historicalcontrols are rarelypossible-partly because he imputation of a common historyoften requiresa level of knowledgeof the past hat the researchercannotpossesspp. 188-189).The essentialogicsof the comparativemethod,as numerouswriters have noted,a are derived from John Stuart Mill's dis-


7/38

r-1

Cttse tudiesand Theories fOrganizationalDecisionMaking 27' quisition n A System f Logic (1843), n which he discussedhe

"method of agreement"and the "method of difference."Themethodof agreement ttempts o identifysimilaritiesn indepen-' dentvariables ssociated itha common utcomen two or morecases. he methodof difference ttempts o identify ndependentvariables ssociated ith differingoutcomes. oth methodsutilizethe logic of eliminationused n solvinga system f simultaneousequations:by "subtracting" actors ound across he setof casesone attempts o "solve" the cases or the critical independentvariables.Mill himselfemphasizedhe serious bstacleso makingeffectiveuse of these wo logicsof eliminationbecause f thedifficulty of finding nstances f a phenomenonhat are sufficientlysimilar.SinceMill's time,other ogicians nd methodologistsaveexpressedvenstronger eservations. ummarizing nd elaboratingMill's own concerns nd caveats,Cohen and Nagel (1934)con-cluded hat the methodsof agreement nd difference id not satisfythe requirementsof either the logic of discoveryor the logic ofverification.Becauseocialsystems an seldombe reatedasclosedsystems, he multiplicity of environmental nfluencesgenerallyacting on any givensocialsystemgreatlycomplicateshe task offinding two or more instancesof a phenomenon n which thedifferences n environmentalcontext are so slight that Mill'smethods have a chance of working. The attempt to use thesemethods n small-Ncontrolledcomparisonss further complicatedwhen,as s oftenthecase n socialscience,he phenomenonunderstudy has complexmultipledeterminants r alternatiuedeterminantsrather than single independent ariablesof presumedcausalsignificance. n investigatorworking with a small number ofvariablesand a smallnumberof casess liable to misseither ofthese ossibilities. inally, heres theobviousdifficulty hat,giventhe small number of observations hat form the basis for thethese methods, the possibilityof highly idiosyncratic factors orobservation/measurementrrorscontaminating he procedure squite strong.sIn spite of these ormidible logical difficulties,Mill's methodsstill constitute he basis or the strategyof controlledcomparison.Therefore,n evaluating he resultsof suchstudies t is necessaryojudge, as best as onecan, how well the researcher ascopedwith


8/38

A. L. GEORGE ndT.I. McKEOWNdifficulties n achieving control" among the cases,and.erany conclusionshat may havebeendrawn are fatally:d by inadequatesafeguards gainstplausibleoccurrencesdifficultiesust noted.)caseor theeffective tilizationof controlled omparisonorI developmentan be strengthenedf the comparison an be'med in a highly systematic ay. This insight ies behindrg's (1971) hird strategy or achievingsome measureof>l n comparative nalysis. e suggestshat theselection nd

rrisonof cases eguidedby an explicit ypology p. 189). f argy s used o guidecase election o that at leastonecase stt in each ellof the ypology, hen wo casesn adjacent ellsiffer on only one endogenousimen$ion.f exogenousi.e.,rtual) differences etween ases an be ruled out as a sourceiations n outcome admittedly ot a simplematter, n viewofecedingdiscussion),hen differencesn outcomes n the two;ould be attributed o differencesn the onedimensionof thergy.However, iven hedifficultyof doing his, t is ikely hatpology will servea more modestbut still important role.typologyspecifies setof mutuallyexclusive nd exhaustiveconditions, hen, f cases reselected o hat at leastonecaseen from eachcell, the researcher as a strategy or system-y searching he universeof possible nitial conditions andng that all logicalpossibilities avebeen overed. he setofchosen n this fashiondoesnot comprisea representativee in a statistical ense, ecause o attention s paid to the:ncydistributions f cases cross ells.However, neat leaste sure that the set of cases electedepresents ll logicallyrle nitial conditions.: typolory constitutes claim about what similaritiesacrossare basic, and, as Diesingnotes, "the questionof whatrities are basic ather han superficials theoretical."Typesypologiesare only implicit theoriesand "typologiesmustrally by controlled by [explicit] theory of some sort to bele" (1971, . 189).Still,asDiesing otes,ypologies rovideations for developinga theory. In this respect,he achieve-of control via typeshas advantages ver relianceon geo-


9/38

.- rrj

CaseStudies nd Theories fOrganizational ecisiortMaking 29graphical ontinuity or common history. nasmuchas typologiesare subject o refinementand revision,whereasgeographyorhistory are inherentlyself-limitingbases or the classification fcasesp. 190).Given he difficultiesencounteredn attempts o makecontrolledcomparisons,t is hardly surprising hat researchershould beskepticalof thegeneral tility of this approach.Not all researchers,however,believe hat the problemsare so intractable under anyand all circumstancess to warrant abandoning his approachaltogether.Lijphart (1971), or example,believes hat usefulresearch an be accomplished y studiesusing the method ofcontrolled comparison, n spite of the formidible difficultiespreviously oted.Of course, he many researchersho haveusedthecomparativemethodalsoaremakingan mplicit statementhatthey consider he methodworthwhile.

ry. STRATEGIESFORPERFORMINGCASESTUDIES:TI{E CONGRUENCEPROCEDUREOf major significanceor dealingwith theproblemof control is theutilization of "within-case" observations o assess he causalrelationshipbetweenvariables.This possibilityhas been over-looked n discussionshat emphasizehe imitationsof the methodsof agreement nd difference.ndeed, nsofar as controlledcom-parison eliessolelyupon.the ogicsassociated ith the method ofeliminationto infer or to assessausal elationships,ts contribu-tions o theorydevelopment ill be"severelyandicapped. ut, t isarguedhere,controlledcomparisonneednot and shouldnot relysolelyon this method.What we have n mind can be conveyedbycontrasting within-case"explanationwith explanationhat reliesupon "cross-case" omparisonof relationships etween ariables.Statistical rocedures,he modern-day ounterpart o Mill's logicsof elimination,compareassociationsf variablesacross asesnorder to infer possible ausal elationships.A simpleexampleof


10/38

IIT T

A. L. GEORGE nd T. J.McKEOWNapproach is the analysisof alternative causal models byrining correlationcoefficients.)n the absence f a sufficiently: number of casesdrawn with an appropriatesamplingpro-re, it is hazardous o utilizethe logic of suchcross-caseom- I;ons to infer causal relationships.n contrast,within-caselnation doesnot require he nvestigator o apply themethodimination o a comparison f cases.re congruence rocedures the first within-casemethod weider. This procedurecan be employedeither n a single-case

y or for eachof the casesn a comparativestudy.It requiresething that is not requiredby the comparativemethod-ary that predicts outcomeson the basis of specified nitiallitions. Sucha theory may be providedby existing ormal orrmal theories,or may have been formulated n the light ofIts obtained in previouscase studiesor quasi-experimentalk. Dependingon the level of developmentof the theory, itslictionsmay be abundantand precise,or they may be scarcehighly generalOncea theory s identified, he researcherirstblishes hevalueof the ndependent nd dependent ariablesnrrticularcase.Theresearcherhen compareshe observed aluehe dependent ariableswith thosepredictedby the theory,6:n the observed ndependentvariables. If the outcome issistent with the prediction, then the possibility of a causaltionship s strengthened.rlthough consistency ften is takenas providingsupport for asal nterpretation, hispracticeobviously s open o misuseandject to challenge.Ways must be found to safeguardagainstustified rnputationof a causal elationshipon the basisof merersistency,ust as safeguards avebeen developedn statistical.lysiso dealwith thepossibilityof spurious orrelation.Helpfulhis respectwill be the useof inferentialcriteria that attempt to,vide he functionalequivalent f orthodoxcontrol. The claimr causal elationshipbased n the indingof consistency etween;ervedand theoreticallypostulatedvaluesshouldbe subjectedtwo screening uestionsbefore t is takenseriously.First, howusible s the claim of "consistency ot spurious"? Second, ow


11/38

Cttse tudies nd Theories fOrganintionalDecisiort akingsensitive s the judgement of "consistencynot spurious"idiosyncratic spects f the ndividualcase?

How Plausible s the Claimof"ConsistencyNot Spurious?"The possibility that consistency etween he values of theindependent nd dependentariabless causal, ot spurious, ainssupport f the relationship etween ariabless consistentwith apreviously ormulated general heory; the strongerand moreaccurate he general heory, the more confidencewe ought to

attach o claims hat consistencys not spurious.The ustificationfor this viewpoint is to be found in Bayesiandecision heory(Gregg & Simon, 1967). f one adoptsa Bayesian iew of theprobability hat the observed onsistencysgenuinely ausal, hen,usingBayes'Theorem,one'sposteriorassessmentf the relationbetween ependent nd ndependentariabless partly a functionof one'sprior estimate f the probability hat the relationship sindeedcausal.The higher the probability assignedn the priorassessment,he higher the a posteriori probability assigned othe claim that consistencys not spurious.To the extent hat acceptance f a model dependson the priorprobability hat we assign o the model, t seems ppropriate oinquire nto theproblemof howprior probabilities reset.Bayesiananalysis reatsprior probabilities sgiven,and inquiresonly intoappropriateproceduresor revisingprior probabilities n the faceof additionalevidence. ut asLeamer 1978) otes, t is oftentimesvery difficult for a resoarchero decide ust what constitutesareasonablerior probability or a givenmodel.One heuristic orsettingprior probabilitiess to evaluate he modelunderconsid-eration n terms of whether t is consistentwith or analogous o

other models hat alreadyare accepted s true.This evaluation slikely nformal,and consists f suchasquestions s he following:l. Does he new model ake a novel view of humancognitiveprocesses?

31to


12/38

32 A. L. GEORGE nd T. J.McKEOWNDoes the new modelposit a novel set of motivations,or anovel relationbetweenmotivationsand action?Is the nature of the evidence equired to support the newmodel of a kind that hitherto has not been equiredor used?Do the predictionsof the new model differ from the pre-dictions of previouslyacceptedmodels or the same nitialconditions? f no previousmodelof thephenomenon xists,do thepredictions f the newmodelseemsonsistent ith thepredictionsof a model of an analogousprocess?The effectof using hese euristics ill be to assign igherpriorprobabilitiesto models hat are generallysimilar to models hathavealreadybeendeveloped. his seems ensiblenasmuchas theacceptance f a novel model that is generally nconsistentwithothermodelswould raisedoubtsabout he accuracy f all the otherpreviouslyacceptedmodels.Becauseevisingone'sentire stock oftheories s verycostly,onewants o be sure hat this is undertakenonly when it is clearlyrequired.

How Robust Is the judgement of"ConsistencyNot Spurious?"A problem confronting case-studyresearcherss that theinfluenceof contextual actorscannot be eliminatedby laboratorymethodsor controlledby statisticalmeans, ut generallymustbegaugedby direct empirical nvestigationof each actor. Given thecommon situation n which the list of contextual actorspossiblyoperatingwithin a"cases much onger han the ist of contextual

factors hat are systematicallynvestigated,nemustconsiderhepossibility hat theseunobserved ontextualvariables perhaps nconjunction with some observedvariables)constitute a causalprocess hat affectsoutcomes n a given case.Therefore, n asituation in which one beginswith some relatively simple andapparentlysoundtheory claiming hat X causesY, if one alsosuspectshat contextualactorsshapehe outcome,n addition othoseenumeratedn the theory,onemust confront he possibilitythat in a sl ightlydifferentcontextY might haveoccurred n the

2.3.4.


13/38

-" ]

Crtse tudiesud Theories fOrganizational ecisionMaking 33absence f X, or X might haveoccurredwithout leading o Y. Ifeitherresult occurs, he existing heory s obviously ncomplete,becauset doesnot take nto account he effectof these ontextualchanges.Supposehatone'sheorypredicts everal aysn whichoutcomeY could occur. f the theoryclaims hat eachpossible ausalpathleading o outcomeY requires hepresence f variableX, thenonecan with relativelymoreconfidence redict hat shifts n contextwould not affect a judgementabout the causalsignificance fvariableX. Becausehe importance f X asa necessarybut notsufficient) onditionhas hus far been ound to be insensitiveocontext,a reasonable orkingassumptions that it will continueto be found insensitiveo context. f, however,variableX is acomponentonly of sontepostulated ausalprocesseseading ooutcomeY, then it is reasonableo operateunder the workingassumption hat the imputation of causal significance o theassociation f X and Y is relativelymorevulnerableo contextualchanges. s the research ffort proceeds nd as more cases rom avariety of contextsare examined heseworking assumptions resubject o revision.How muchconcern s devoted o the possiblesignificance'ofshifts in unobserved ontextual actors at anygiven point in the investigation s thereforea function of theresearcher'sprior theory. Any theory that identifies differentcausalprocesses t work in differentcontextswill impart moreplausibility o the claim that "contextmatters" han will a theorythat doesnot claim that causalprocessest work varyasa functionof context.If theory s not sufficientlywell developed o permit udgementsabout the likely courseof events nder a varietyof contexts, ndif a sufficient arietyof cases ppropriateor controlled omparisonis not available,he resdarcher ay rely on disciplined nalyticalimagination o assesshe plausibilityof an argumentabout causalprocesses.n considering omeplausible ounterfactual ituationchosenor its general imilarity o the situationactuallyanalyzed,the researcher ay attempt he following thoughtexperiment: fmy argumentabout causalprocessest work in the observed aseis correct, what would occur in the hypotheticalcase f it were


14/38

A. L. GEORGE nd T. J.McKEOWNred by the samecausalprocesses?We are not bashful inng that the researcher's reconceptual feel" for the cir-rnces f thehistoricalcases mportantandhelpful n makingr thought experimentworthwhile. If the researcher as ar of preconceptual, ontextualknowledgeof the case,andrmaginedesultsseembizarre, his suggestshe researchers:d in downgrading he strengthof claims about the causalcanceof a givenvariable.procedureswe haveoutlined to discipline he useof theJenceprocedureshould serve o guide across-case ausalretationsso that they are essarbitrary. Because econceiveresearch rojectasa series f iterationsof theory ormation,ration,and theory revision,we take for granted he fact thatne times (particularly early in the researchprocess), e-rerswill attach considerablymore (or less)plausibility toes han theywould if they hadperfect nformation about there system n question ctuallybehaves. owever, he nitialItions of the processest work are not critical f the terativeJure succeeds n stimulating constructive revisions ofruslyheld beliefs.

V. STRATEGTESORPERFORMINGCASEUDIES:THE PROCESS-TTTACINGROCEDUREfar, the analysisof within-caseexplanationhasconsideredrrocedures hat supplement he assessmentf the degreeof;tencybetween bserved nd predictedoutcomes s he basisrdgmentsabout causal processes t work. However, thisuenceprocedurehardly exhaustshe possibilitiesor within-rnalysis. As even a cursory examinationof such fields asnatic history would indicate, case-study esearchers ftencausal nferences ithoutany relianceon quasi-experimentaliques or techniques(such as the congruenceprocedure)

ed by a quasi-experimentalogic. There is a well-knowne about the scientificstatusof arguments bout causation n'ical works,T but a more pragmatic and logically prior


15/38

'- Yi l

CaseStudies ndTheoriesfOrganizational ecisiort akingconcernseemso havebeen ost in the shuffie:how exactlydohistorians o about makingcausalnferencesn the first place?We argue hat historicalarguments bout causalprocessesnstudies f humanand organizational ecisionmakingoften nvolvea "process-tracing" rocedure George, 1979a,1979b).8As itsname mplies, hisprocedures ntended o investigate ndexplainthe decision rocess y which various nitial conditionsare trans-lated nto outcomes. process-tracingpproach ntailsabandon-ment of the strategyof "black-boxing" the decisionprocess;instead,hisdecision-makingrocesss hecenter f investigation.eThe process-tracing pproachattempts o uncover what stimulithe actorsattend o; the decision rocesshat makesuseof thesestimuli o arriveat decisions;he actualbehavior hat then occurs;the effect of various institutional arrangements n attention,processing, nd behavior; and the effec t of other variablesofintereston attention,processing,nd behavior.In sociology nd anthropologyherehas ong beencontroversyover the extent to which casestudiesor ethnographies houldworldttempt to recreate he world as actors experiencet, or whetherthesestudiesshould simply imposea set of researcher-designedexperiencet, whethercategories nd concepts n the experience f actorsand translatethat experiencento the researcher'sanguageSmelser, 968).Aswe understandprocess-tracing,t involvesboth an attempt toreconstructactors' definitionsof the situationand an attempt todevelopa theory of action.The frameworkwithin which actor'sperceptions nd actions are described s given by the researcher,not by the actors hemselves. owever,becausehat frameworkoften s revisedafter observations f behavior in fact, the originalframework itself may-be based on previous observationsofbehavior), he externally mposed heoretica l ramework is cer-tainly shapedby actors' self-definitions, ven f it is not designedspecificallyo capturesuchself-definitions.Process-tracings not a statistical r quasi-experimentalethodof analyzing ausal rocesses.n termsof an underlying hilosophyof explanation, t owesmuchto the patternmodel of explanation(Diesing, l97l: Kaplan, 1964:Raser,Campbell,& Chadwick,1970).From this perspective,he processof constructingan

35


16/38

36 A. L. GEORGE nd T. f . McKEOWNexplanationsmuch ike heconstruction f a webor network.Theresearcher ssembles its and piecesof evidencento a pattern;whethera pieces to be changed r addeddepends n whether hechange its with what alreadyhasbeenconstructed, nd whetherit strengthenshe web's structure.Does the modificationof theexplanation reate nternal nconsistenciesn the heory?Does hemodificationof the explanation reatemore new puzzles han itsolves?f yes s the answer o these uestions,he modification srejected.Modificationsthat are consistentand producesmaller,more ocalized,and ess requent esearch uzzles re o bevalued.The growth of the web orients the search or new pieces,ust asthegrowthof a ig-sawpuzzleguideshe searchor pieceshat willfit togetherwith what is alreadyassembled.The process-tracing pproach also differs from conventionalquasi-experimentalesearchn that each case" in quasi-experi-mentalresearchs represented sa singledatapoint, whereas, ycontrast, n case tudies ne ypically s interested ot in a singlebit of behaviorbut, rather, n a streamof behavior hrough ime(Brown, 1974;Campbell,1975;Lundberg, 94l). The behaviorof the system s not summarizedby a singledata point, but by aseriesof points or curvesplotted through time. Any explanationof theprocessest work in the case husnot only must explain hefinal outcome,but alsomust account or this streamof behavior.As Chassan1961)observes,n a sensehe case erves s ts own"control," becausehangesn the streamof behavior an be inkedmore confidently o changes n theoretically elevantvariables.This is possiblewhen the history and attributesof the decisionmakers emainunchanged, r change ery slowly relative o therateat whichobserved ehavior hanges.The process-tracingechnique pplied o the analysis f humanand organizational ecisionmaking also reduceshe difficultiesassociatedwith unobservedcontextual variables. Because helimitations on the perceptualand information-processingapa-bilities of humansare well known and pronounced,he process-tracing techniquehas a chance of constructinga reasonablycompleteaccountof the stimuli o whichan actorattends.More-


17/38

CaseStudies ndTheoriesfOrganizational ecision aking 37over, f the study in question s one of organizational ecisionmaking, the researchers further assistedby the fact that thisactivity s a socialenterprise. he socialnatureof decisionmakingimplies hat actorsmustcommunicate ith oneanother n makingdecisions;hat the contentof thiscommunication ill revealmuchabout the attention ocus, he decision ules,and the behaviorofactors(even f it cannot be taken at face value); and that thiscommunicationoften is accessibleo researchers. hese eaturesof the casestudy do not guarantee hat researchers sing theprocess-tracingechniquewill be able o attain the samedegreeofconfidencen their conclusions bout the working of socialpro-cessesswould be possiblef theycould readilyemploystandardexperimentalor quasi-experimental ethods; they do imply,however, hat the conclusions hat emerge rom casestudiesarelessa productof the subjective tateof the researcherhan is oftensupposed.Strategiesnalogouso theprocess-tracingechnique avebeenadvanced y researchers ith interestsand backgroundsdifferentthanours.Scriven 1976) asproposed "modus operandi" MO)method of performing evaluation research.The MO procedureassumes hat the researcher lready possessesn inventory ofcausalprocesseshat may beat work in a particularsituation,andthat the researcher lso knows what sortsof empirical symptomsor tracesare manifestedwhena givencausalprocesss at work.The research ask is then one of pattern recognition. Such anactivity s feasiblef the numberof patterns or which o searchsrelatively malland if the patternsare well defined.loScrivendoesnot explicitly addresshe critical questionsof howinventoriesof causalprocessesre constructed n the first place,and how empirical nvestigations an lead researcherso revisetheir understanding f causal rocessesnd he related ymptoms.Given the well-knownhuman propensity o imposepatternsondata even when the data are generatedby random processes(Steinbruner, 974:pp. I l0- I l2), perhapswecan ake or grantedthat researchers ill "see" somesort of causalprocessn the data,evenf theybeginwith no explicit heory.As sgenerallyecognized,


18/38

38 A. L. GEORGE nd T. J.McKEOWNwhether he perceived atternsare to be accepted s "real" (i.e.,whetherone has identified awsthat generatehe patterns f theappropriatenitialconditions represent) epends n the esults fadditionalobservations.f the data are usedas the basis or ageneralization,hey cannot simultaneously erveas a convincingtestof that generalization.Leamer's (1978)discussionof "data-instigated"models ineconometrics offers an important qualification to this strictseparationof model formulation and model testing. Orthodoxstatisticaltheory, whether of the classicalor Bayesianvariety,condemnsdouble-counting"he data: f oneusesnformation ogenerate ypotheses,ne cannot hen use hat same nformationto "test" thosesamehypotheses. n the otherhand, t we are tobelieve eamer's haracterizationsf how econometriciansctuallypracticeeconometrics1978, . vi), post-observationlterationofhypothesesr thewriting of entirelynew hypothesesrecommonprocedures,even though the way that research s reported injournals suggestsotherwise. Leamer attempts to resolve thecontradictionbetween heory and practiceby suggesting pro-cedurewherebypost-observationmodel revision is legitimized,but the valueof the data asa subsequentestof a revisedmodel sdiscounted o reflect the double usage. Leamer's nontechnicaldescriptionof this procedure s noteworthy,particularly in viewof the similaritiesbetween is accountof the research rocess ndotherviewswe havepreviously ited:

I like to describe his asSherlockHolmes nference.Sherlocksolves he caseby weaving ogetherall the bits of evidencento a plausiblestory. He wouldthink it preposterousf anyonesuggestedhat he shouldconstructa functionindicating the probability of the particularevideirce t hand for all possiblehypotheses nd then assignprior probabilities o the hypotheses. e advisesinstead, No data yet . . . It is a capital mistake o theorizebeforeyou haveall the evidence. t biaseshe udgments."

There s, ncidentally,a tendency mongsocialscientists, articularly hosemost trained n statistical nference, o disparageSherlockHolmes nference."Boy, he really went on a fishing expedition hat time, didn't he?" The factthat Sherlock Holmes procedures nvalidate statistical nferencer is evensometimes aken to mean that SherlockHolmes inference s "unscientific."Nothing could be further from the truth.r2


19/38

CaseStudies nd Theories fOrganizational ecision akingLeamer'sargument hat SherlockHolmesprocedures re an

essentialpart of science s illustrated by Kuhn's argumentconcerninghe distinctionbetween strologyand astronomy:Compare he situationsof the astronomerand the astrologer. f an astrono-mer's prediction failed and his calculationschecked,he could hope to setthe situation ight. Perhapshe datawereat fault: old observationsould bere-examined nd new measurements ade, askswhich poseda host o[ cal-culationsand instrumentalpuzzles.Or perhaps heory neededadjustment,either by the manipulationof epicycles, ccentrics, quants,etc., or by morefundamental eformsof astronomical echnique.For more thana millenniumthesewere he theoreticaland mathematicalpuzzlesaround which, togetherwith their instrumentalcounterparts,he astronomical esearchradition wasconstituted.The astrologer,by contrast,had no suchpuzzles. he occurrenceof failurescould be explained,but particular failuresdid not give rise to re-searchpuzzles, or no man , howeverskilled,could make useof them in aconstructive ttempt o revise heastrologicalradition.Therewere oo manypossible ources f difficulty,mostof thembeyond heastrologer's nowledge,control, or responsibility. ndividual failures were correspondingly unin-formative, and they did not reflecton the competence f the prognosticatorin the eyes of his professionalcompeers . . ln short, though astrologersmade testablepredictions and recognized hat the predictions sometimesfailed, heydid not and couldnot engagen the sortsof activities hat normallycharacterize ll recognized ciences.Kuhn, 1970, p. 9-10, quoted n Leamer,1978,pp. 286-287)

The ssue s thuswhether discipline an makeuseof the researchpuzzlest faces o make constructive evisionsn its models.Suchan activity must entail the formulation of hypothesesafter thepuzzleshave beengenerated,with the intent of disposingof thepuzzles. The test of suchpost-observationalormulationsonthe originaldata setcan.then ereadasan indicationof how wellthe new ormulationdoes ndeed ispose f thepuzzlesn a partic-ular situation).t already asbeen ecognizedhatsucha procedurecarries he danger hat the resulting eformulationswill result n areductionof the overall coherence nd power of a model ratherthan in an improvement Lakatos,1970),but this dangerseemsacceptablef the alternatives to behave ike an astrologer atherthan an astronomer.Using case studies o attack anomaliesexposed y quasi-experimentalnalysiss one obviousexample fthis strategyhat hasalreadybeen ried r3 one alsocould usecase

39


20/38

40 A. L. GEORGE nd T. J. McKEOWNstudies o investigate nomalies ncoveredby earliercasestudies.In either case,we think it more appropriate o conceiveof casestudiesas a complement ather than as a substitute or quasi-experimentalor experimentalwork.

Thecase or discounting ather han disallowinga test hat usesthe samedata from which a theory s constructed estspartly onthe observationhat the systematic pplication f somestatisticalmodel (e.g.,ordinary eastsquares)mposes differentstructureon the data than preconceptual umanpatternhunting. Becausethe testingmethoddiffers rom the methodby which the model sextracted rom the data, somegreater-than-zero egreeof con-fidence houldbeattachedo test esults, venwhen hedata s thesame.Thecase or discounting ather han completelydisallowingalsohasa morepractical ide:any model hat fits a givendatasetwellmayalso it otherdatasetswell. n the social ciences, odelsthat fit data setswell are scarce.They should not be discardedlightly, even if they are not immaculatelyconceived.Testing amodel against he samedata that inspired t at leastsuppliesareasonably recisedea of how well the model really fits that dataset. If the fit is good, one can then proceed o test the model inother settings. f the fit is not good, he process f model revision,basedon the observation f the model's ailures, an lead to animprovedfit.As the abovediscussionmplies,we certainlyhaveno objectionto orthodox claims hat one can not double-count ata, and thattestsbasedon the samedata that inspired he model are hardlyconvincing.The problemwith the orthodoxview s not that it iswrongbut that it is critically ncomplete, ecauset hasnothing osay about the process y which researcherso about developingmodels.The discussions f so-calledprocess-tracingistoricalstudies,modusoperandievaluation esearch, r SherlockHolmeseconometric echniquesall constituteattempts to fill this voidby making methodologicallyself-conscious tatementsabout .procedures f theorydevelopment.The primary difficultywith the process-tracing ethod s notthat it fails to conform to quasi-experimental rthodoxy, but,rather, hat in many situations, o perform t is not practicable.Although the process-tracing ethod offers the possibility of


21/38

CttseStudies ndTheoriesfOrganizational ecisiotr aking 4lexplanationshat are more powerful than thoseyieldedby thecongruence rocedure,ts data requirements re substantiallygreater.Documentinghe process y which nputsare convertedto outputs equiresmuchmore evidencehan doessimplynotingthe nputsand outputsand treatingwhat iesbetweenhem asablackbox. Available istoricalmaterials ftenare nadequateorpurposes f processracing: he researcher ay need o expendconsiderable ime and resourceso locate historical data viaarchivalsearchor interviews.Use of the congruence roceduredoesnot requireas much nformationabout the case n questionand is. therefore,more practicable lthough essambitious andconvincing.

VI. THE METI{OD OF STRUCTURED,FOCUSEDCOMPARISONSo far we have discussed n a fairly abstract evel the variousproblems hat arise rom attempts o usecasestudies n theorydevelopment.We now turn to our second oncern-the statementof a usefuland explicitmethodologyor casestudies.Our treat-ment of this topic builds on earlierdiscussionsf the method ofstructured, focusedcomparison (George, 1979b; George &Smoke,1974).A comparison f two or more casess focusedinsofaras heresearchereals electively ith only thoseaspects feachcasehat arebelievedo be relevanto theresearch bjectivesand data requirements f the study. Similarly, controlled com-parison is structured when the researcher,n designing heitudy, defines nd standardizeshe datarequirements f the casestudies. his s accomplishedy formulatingheoreticallyelevantgeneralquestionso guide he examinatr+n f eachcase.Themethod s presentedsa wayof performing smallnumberof casestudiesn suchawayas o make heassessmentf thecomparabilityof thecasesmuch more systematic nd defensible, ut thegeneralapproachwe outline also s well suited o the performance f asinglecasestudy.The importanceof introducinga selectiveheoreticalocus ntocasestudieswill be evident f we recall hat one reasonso many


22/38

42 A. L. GEORGE nd T.l. McKEOWNcasestudies n the past contributedunevenlyand meagerly otheory development s that they lacked a clearly definedandcommon focus (George, 1979a, 1982).Different investigatorsdoing research n a phenomenonended o bring diverse he-oretical andnontheoretical)nterestso bear n theircase tudies.However nteresting ndwelldone n andof itself,each ase tudytended o pursue ather diosyncratic esearch roblemsand toinvestige set of dependent nd ndependentariableshat oftenwerecorrespondinglydiosyncratic.Moreover,manyof these asestudiesacked clear ocus, ecausehe nvestigator asnot guidedby a well-definedheoretical bjective nd wasdrawn nstead ndirections ictated y themost eadilyavailable istoricalmaterialsor by aspects f the case hat wereudged nteresting n intuitivegrounds. t is not surprising,herefore,hat later researchershodid have a well-defined heoreticalnterest n certain historicalcases ound that earliercase tudies ftenwereof little value ortheir own purposes nd had to be redone.In this connection t is important to recognizehat a singlehistoricalevent anbe relevantor researchmbracing varietyoftheoretical nterests. or example,he Cubanmissile risisoffersuseful material or developingmanydifferent heories; his casemay be (indeed,has been) egarded nd usedas an instanceofdeterrence,oercivediplomacy,crisis management,egotiation,domestic nfluenceon foreignpolicy,personalitynvolvement ndecisionmaking,etc.Eachof these iverseheoreticalnterestsnthe Cubanmissile risis equiresheresearchero adopta differentfocus, o developor usea different heoreticalramework,and toidentify a different set of data requirements.Treatment of theCuban missile risis n a case tudymustbe selectitelyocusednaccordwith the type of theory hat the nvestigators attemptingto develop.As alreadynoted, t is alsonecessaryo structure hecomparison f caseso ensure hat theyare done n a comparablefashion.This s essentialf the results f the ndividiral ases re ocumulate.Structured omparison an beachieved y askinga setof standardi:ed, eneral uestionsf eachcase Georgeand Smoke,1974, p.97-103; Smoke,1977, p. 37-39).The interrogationof the case s akin to the process hereby ognitive sychologistsinterrogate esearch ubjectsn order to learn somethingof the


23/38

CaseStudies ndTheories fOrganizational ecisiott aking 43thoughtprocesseshe subjects mploy n solvingproblems Gregg& Simon, 1967).When interviewing s possible, he analogy scomplete, ut evenwhen he researchers restrictedo analysis fhistoricalmaterials, he interrogation of the written record ssimilar to this procedureusedby cognitivepsychologists,n thesensehat the path of one'ssearchs dependent n the questionsthat alreadyhave beenaskedand the answershat alreadyhavebeengiven.This is only to say-and to insist-that controlledcomparisonof a small N should ollow a procedure f systematic ata com-pilation.The questions skedof eachcasemust be of a generalnature; heyshouldnot becouchedn overlyspecificerms elevantto onecaseonly but shouldbe applicableo all caseswithin theclass f eventswith which hestudy s concerned.Asking he samequestions f each asen a controlled omparison oesnot preventthe investigatorrom eitherdealingwith more specific spects fthe case r bringingout idiosycraticeatures f each ase hat maybe of interest n and of themselves,f not also for the theory-development xercise.)The set of standardized eneralquestionsmust be carefullydeveloped o adequately eflect the researchobjectivesand the-oretical ocus of the study.Certainly,comparabledata will beproducedn thecase tudies y questionshat do not link up withtheexisting r postulatedheoryaddressedn the nvestigation ndare not well formulated to further the researchobjectivesof thestudy,but such data will be lackingor limited in its theoreticalrelevance.As theseremarks ndicate,a purely mechanicalorritualisticuseof the method of structured, ocusedcomparisonwill not guaranteeood esults. he ocusing nd structuringof thecontrolled comparison s only part of the task of designingsuchstudies.

VII. DESIGNAND IMPLEMENTATIONOF CASESTUDIESThereare our phasesn the design nd mplementationf theory-orientedcasestudies.n Phase the desienand structureof the


24/38

M A. L. GEORGE nd T. f . McKEOWNresearchs formulated. n Phase each ase tudy s carriedout inaccordwith the design. n Phase the researcherrawsupon theresults f thecase tudy n order o assess,eformulate, r elaboratethe initial theory stated n Phase . In Phase4 the researcherpresentshe resultsof the research.

Phase .:DesignA well-designedingle-caser controlled-comparisontudymustdeal satisfactorily ith five asks.These asksare not novel; they

reflect familiar requirements or all types of systematic, heory-oriented esearch.What sperhaps ovel,however,s the assertionthat heseive asks houldbeappliedn case-studyesearch swell.Task 1

A clear specificationof the research roblem and the researchobjectives f the study.What kind of phenomenon r behavior s beingsingledoutfor examination-that s, what s theclass f events f whichthesingle ase r casesn the controlled-caseomparisonwillbe an instance?Is thephenomenono beexplainedhought o bean empiricaluniversal (i.e., no empirical variation in the. dependentvariable), o that the research roblem s to account or lackof variation n theoutcemes f thecasesn theclass f events?Or is what is to be explainedan observable ariation in thedependentariable?What theoretical rameworkwill be employed-is thereanexisting heory or rival candidateheorieshat bearon thoseaspects f the phenomenon r behavior hat are to be ex-plained? f not, what provisional heoryor theorieswill theresearcherormulatefor the purposeof the study?Whichaspects f theexistingheoryor theorieswill besingledout for refinement nd elaboration?


25/38

CaseStudies ndTheoriesfOrganizational ecisiott akingTask2

45

Specificationf theelementsconditions, arameters,ariables)thatwill beemployedn analyzinghe historical ase r controlledcompanson:what is hedependentor outcome) ariableo beexprained?what independentand intervening) ariables omprise hetheoreticalrameworkof the study?which of thesewill beheldconstant serve sparameters)ndwhichwill be allowed o varyacross aseso be ncluded n thecontrolled omparison

Task3Selection f a historicalcaseor cases uidedby a typologytn

aa

developedrom the work in Tasks and 2:The universe r class f eventsrom whichcases re selectedmustbewelldefined.n a controlled omparison, ll the casesmustbe nstances f the same lass.Thecasesmustbe selectedo provide he kind of control andvariation requiredby the researchproblem-that is, caseselectionmust be appropriate o whether he methodbeingemployed s the method of agreement r the method ofdifference.

Task 4Consideration f the way n whichvariance f variables anbestbe describedo further the.assessmentr refinement f existingtheory.A judgementon this matteroften s subject o revisionasthehe empiricalwork proceeds. he researcher'secisionabout howto describe ariance an be critical for achieving esearch bjec-empiricaltives,becausehe discoveryof causalrelationshipmay indeeddependon how the variance n the these ariabless postulated.Basingdesigndecisions n a priori judgementsmay be risky andunproductive; he investigators more likely to developsensitive


26/38

46 A. L. GEORGE ndT. J.McKEOWNwaysof describing ariancen the variables fter he or she hasbecomeamiliarwith thewaysn which heyvary n historical ases.An iterative procedure or determininghow best to describevariance s therefore ecommended.Task 5

Formulation of the data requirementso be satisfied n theanalysisof the historicalcase(s)-that is, thegeneralquestionsobe asked of each case.When the researcher ttempts o use aprocess-tracing pproach, he data requirements enerallyaremuch more extensivehan they are in situationswhereonly thecongruence rocedures employed.Althoughthe exactnatureofthese equirements ill varydepending n the heoreticalocusofthe studyand the natureof the situationbeing nvestigated,herequirement o achievea thorough understanding f decisionprocessesenerally ntails longer,moredetailed etof questionsto be asked.Several bservations ustbemadeabout heseivedesignasks.First, in any givenstudy these asksshouldbe viewedas com-prisingan integratedwhole.The researcherhouldkeep n mindthat these asksare nterrelated nd nterdependent.or example,theway n whichTask 2 is performedshouldbe consistentwith thespecification f Task l. Similarly,both the selection f casesnTask3 and the theoreticalrameworkdevelopedn Task 4 mustbeappropriateand serviceablerom the standpointof the determi-nationsmade or Tasks and 2. And finally, he dentification fdata requirements ustbe guidedby theperformance f rasks I,2. and3.second, he researcherhouldkeep n mind that a satisfactoryintegrationof the ive asksusually annotbeaccomplishedn th;first try. A good designdoesnot comeeasily. onsiderableter-ation and respecificationf the various asksmay be necessarybeforea satisfactoryesearch esign s achieved. he researcherwill need o gain amiliaritywith the phenomenonn questionbyundertaking preliminary xamination f avarietyof cases, eforeselecting ne or severalor intensivenvestigation.


27/38

CaseStudies ndTheoriesf Organizational ecision aking 47Third, despite he researcher's estefforts, the formulation ofthe designs likely to remain mperfectn oneor another espect.These imperfections may emergeand become evident to theinvestigator fterhe or she s well nto Phase or evenPhase ofthe study. If thesedefectsare sufficientlyserious, he researchershouldconsiderhalting further work and redesigninghe study,even f this means hat someof the casestudieswill have to beredone. n drawingconclusionsrom the study, he researcherorotherswho evaluate t) may be able to state someuseful essonsfor a betterdesignof a new studyof the problem.

Phase2: The C-ase tudiesEach of the.case tudies n the controlledcomparison or thesingle ase tudy, f that swhat he esearchonsists f ) is analyzedfrom the standpointof the data requirementsdentified n Task 5of the previousphase.The valueof the dependent ariable(s)sestablished hrough standardproceduresof historical inquiry.(By that we do not mean hat quantitativeanalysiss not used. f itfits the purposesof the research,here s no reasonwhy variablescannot be quantifiedand scaledn some ashion.)tsWhetheror

not quantitativeanalysis s used, he researcher hould articulatecriteria for "scoring" the variablesso as to provide a basis forintercoder eliability.The next step is the developmentof the explanations or theoutcome(s)n eachcase.As alreadyargued, n case tudies his is amatterof detectivework andhistoricalanalysis ather hana matterof applying an orthodox quasi-experimentaldesign. Socialscientistsperforming such casestudieswill need to familiarizethemselveswith the "craft" aspectsof the historian's trade-learning, for the context in which they are working, the specialdifficulties presentedby various kinds of evidence hat may beavailable; singmultipleweak nferencesather hansingle tronginferences o buttress conclusions; developing procedures forsearching hrough largemasses f data whenthe objectivesof thesearchare not easily summarizedby a few simple searchrules.


28/38

48 A. L. GEORGE nd T. J.McKEOWNHistoricalexplanations re alwaysprovisionaland so will be theexplanations enerated y case'studyesearchers.he researchermay find that his or her explanations resuccessfullyhallenged yothers on one or another ground-for example, the originalresearchmay have overlooked elevantdata or misunderstoodtssignificance; r it may have ailed o consider n importantrivalhypothesis, nd soforth. If case xplanations re atersuccessfullychallenged,he researcher ill have o reassesshe mplications orany theorythat has beendevelopedrom suchcase tudies.Suchareassessmentlso would be necessaryn the event that new his-torical databearingon the case(s)ecome vailable t a laterdateand ead o a successfulhallenge f earlierexplanations.hus, heprovisional atureof explanationsf casesmplies hat the theorydevelopedrom thosecases lso s provisional.Severalother observationsmust be made about Phase2. lnaddition to developing specific xplanation or eachcase, heresearcher lso transforms he specificexplanation nto the con-ceptsand variablesof the generalheoretical ramework specified.in Task 2. (ln Eckstein's975] terminology,he researchs "dis-ciplined-configurative"ather than "configurative-idiographic.")For such a transformation of specificexplanations nto generaltheoretical terms to be possible, the researcher's heoreticalframework must be broad enough o capture he major elementsof the historicalexplanation. hat is, the set of independent ndinterveningvariablesmust be adequateo captureand record theessentials f a causalaccountof the outcome n the case.Thedividing inebetween hat s "essential" ndwhat s not s whetheror not aspects f a causalprocessn a givencaseare found or aresuspectedo operate across he entire classof casesunder con-sideration. For example, f some instanceof organizationaldecisionmaking weredecisively ffected y the fact that oneof thekey participants n the decisionprocesscaught a cold and wasunable o attendan mportantmeeting,hiswouldnotconstitutebasis or revisingour theoryof organizational ecisionmaking oendogenizehe susceptibilityof actors o disease.t would,how-ever,constitutea basis or a general rgumentabouthowoutcomes


29/38

CnseStudies ndTheories fOrganizational ecisionMakingare affectedsimply by the presence r absence f importantpotential articipants.Somehistorians bject o this procedure f transforming richand detailedhistorical explanation nto a more abstract andselective ne couched n theoretical oncepts, rguing hat someuniquequalitiesof the explanationnevitablywill be lost in theprocess. his is undoubtedly rue; some ossof informationandsome implifications inherentn any effortat theory ormulation.The criticalquestion, owever,s whether he ossof informationand the simplificationeopardize he validity of the conclusionsdrawn from thecasesor the theoryandthe utility of that theory.This question annotbe answered bstractly.t is indeedpossiblethat the transition rom specifico general nalysismay result nthe dismissalof causalprocesses t work in the case, simplybecause hey are not already captured by the general theory,or becausehe researcherails to recognizehat variablesdeemedto play only a situation-specificole actuallyhavea much moregeneral ignificance.o say hat theavoidance f these rrors s amatter of the sensitivity'andudgemento[ the researcher, hiletrue, snot veryhelpful n avoiding hem.Oneslightlymore specificguidelines that the susceptibilityo this error seems o depend,first, on whether he research asas a primary interest he searchfor new causal atterns, r whether he effort s to evaluate laimsabout somecausalpatternsalreadyhypothesizedo be operatingin a particularcase;and second, n how the researcher lects odescribehe variance n eachof the variables. s the descriptionof the variance becomesmore concreteand fine-grained, heanalysismore readily will accommodatea more differentiateddescription f the causalprocessest work.

Phase3:Drawing the Implicationsor TheoryRecall that in Task I of the study the researcher tated theexisting heoryor candidate heories hat were o be assessedymeansof the cases. avingperformedhe case tudiesn Phase ,the researcher roceedso draw upon the resultsof thesecases n


30/38

50 A. L. GEORGEandT. ]. McKEOWNorder to assess,efine, or elaborate hese heories.There arelimits o what case tudies anaccomplishn this respect, iven hevariousmethodologicalroblemshat wehavediscussed. ecisivetheory estingn controlled-comparisontudies ften snot possiblebecause any social-scientificheories renot yet ormulatedwithsufficient larity or detail o permit rigorous esting.Preciselyorthese easons he use of casestudies n building, refining,andelaborating heory often is more important than their valueas adecisive est of existing heories.It has been ecognizedEckstein,1975) hat in principle t ispossibleo conducta "critical test" using he case-study ethod.Even a single case, f strategicallyselected nd properly carriedout,mayserve o invalidate theory.How cana candidate ase orsuch a critical test be identified?As Ecksteinnotes,sucha casemust"fit" at leastoneexisting heorywell; that s,given he nitialconditionspresentn the case,he given heory eads s o attachahigh degree of confidence o a particular prediction about be-havior in that case.r6 deally, any rival theories hat exist willgenerate onflictingpredictions boutbehavior n thecase, o hat,regardless f the observedoutcome, he outcome will be incon-sistentwith at leastone heory.However,even f only one heory sunder consideration, he performanceof a single"critical" casecould serve o reducesubstantially he amount of confidenceweattach o that theory.Although we argue that "testing" of theories s sometimespossiblen case tudies, he morecommonproblem s drawing heright "lessons" from a casestudy. How do we decidewhetheragiven outcome s consistent r inconsistentwith a theoreticallybased rediction?Assuminghe nitial theories f thephenomenonin question urn out to be not completely ccurate,what can thecase tudy ell us about theways n which he heory might usefullybe modified?Judgements f prediction "success"may not be easy in asituationwherestatistical pproacheso this problemcannotbeapplied.Essentially, rediction success"s a relative oncept-wecan say hat one heory smore successfulhan another f outcomesareobservedo be"closer" o itspredictionshan o thepredictions


31/38

CaseStudies ndTheories fOrganizational ecisiottMaking 51of some ther heory.Sucha udgements facilitatedwhena theorymakesclaims about the decision-making rocess, ather thanrestrictingtself o statements bout the relationsbetweennputsand outputs. When claims about processare available or com-parisonwith observations boutprocess, qualitativecomparisonof observedand predictedbehaviors s much more convincingthanwhenone simply s comparingonesingleobserved utcometo onesinglepredicted utcome.Wherepredictions nd outcomesarequalitative, t is also desirableo havemultiple theories hatpredict mutually exclusive utcomes. n such a situation,pre-diction "success"has a muchclearermeaning han it doeswhenonly one theory s beingconsidered, r whenmultiple theoriesarebeingconsidered, ut the theories recomplementaryather hancompeting explanations. n the latter situations one simplycomparesqualitatively one pattern of observedoutcomeswithanother pattern of predictedoutcomes,and makessome udge-ment about whether the degreeof similarity betweenoutcomesis sufficiently lose.The task of learning the right lessons rom case studies s acritical one. One may achieveprediction success, et concludethat the model must be revisedor refined so that it can makepredictions hat aremoreprecise nd more specific. his canoccurwhen he case tudyuncovers ausalprocessesithertoundetected,or leads o a reformulation of previously postulatedcausalpro-cesses.Of course, the search or alternative causalprocessesslikely to be more intensive n situationswhereexistingmodelsseem o be performingpoorly.One can, of course, est the hypotheses nd modelsdevelopedby case tudiesanywayonepleases. o the extent hat case tudiesassist n clarifying the implicationsof a set of theoreticalproposi-tions, they will directly aid subsequentestingby making it easierto decidewhat to testand how to interpret he results.Researchersinterested n developing more refined knowledge of a givenphenomenonan movebackand forth betweenntensive nalysisof appropriately chosencases nd quasi-experimental r experi-mentaldesignsRussett, 970).We are nclined o believehat casestudieswill be particularly useful at an early stage of theory


32/38

52 A. L. GEORGEand T. J.McKEOWNdevelopment,when considerations f theorydevelopment reatlyoutweigh hoseof testing, ut wealsoagreewith Eckstein's1975)idea that casestudies ould prove useful at virtually any point inthe life cycleof a research rogram.As theory developmentproceedsand as empirical findingscumulate,heoristswill trade ich, ambiguous oncepts ndclaimsfor conceptsand claims that are considerablemore preciseandloadedwith fewer connotations. he domainto which a theoryapplieswill becomemore clearly demarcated, nd the focus ofempirical nvestigations ill become essdiffuse.t7Theorieswilltend to becomemore formalized,and the relationshipbetweentheoreticalpropositionswill tend to be recast n termsof logicalnecessity. neapproacho highlydevelopedormaltheory s thatof microeconomics. notherexample,which webelieve aspartic-ular relevance o casestudies, s the computer simulation. ncomputer simulationsa theory of the behaviorof somesystem sstatedn a specialanguage. he resulting utput behaviorof thatsystem ollows logically from the propositionsprogrammed odescribe ts behavior. The computersimulation possesseshestandardadvantagesenerally ttributed o formal microeconomictheory-lack of ambiguity n the meaningof propositions, nd aclear ogical elationshipbetween ropositionsdescribing systemand claims about the behavioraloutputsof that system. Anyonewho has everperformedany computerprogrammingknows thatwhencomputer-languageropositionsare stated n ambiguousorincorrectgrammaror syntax, he system implydoesnot run )r8In addition, he computersimulationseems ettersuited o capturedecisionprocesseshan doeg ormal theory of the microeconomicvariety. Microeconomic heory is geneiallyconstruedas makingclaimsonly about nput-output relations; he mainstreamreat-ment of the decisionprocesss to deem t irrelevant(Friedman,1953).One could drop this view of microeconomicheory andmodel decisionprocessesy treatingorganizatiohal ecisions sconsistingof a series f input-output relations,with the output ofeachstage erving s he nput to thenextstage. he ntermediateinput-output relationswould then serveas statements bout theprocess. owever, s ong asparsimonysvaluedmorehighly han


33/38

CaseStudies nd TheoriesfOrganizational ecision aking 53depictionof process,t is unlikely that this coursewill be taken,becausehe system s much more briefly described y equationsthat skip the ntermediate teps nd simplydescribeinal outputsin terms of original inputs. Given our substantiventerest naccurate epictionof processesnd a lackof interestn parsimonyfor its own sake,we are nclined o favor a way of stating formaltheorieshat bearsmore esemblenceo computer imulationshanto existing microeconomic heory. Even if invoked only as ametaphor, he computersimulationseems o be a considerablymore useful organizingconcept han microeconomicheory. Inparticular, he presence f situation-specificausalprocessesnconjunctionwith highly generalprocessesan be viewed as asituationakin to a program n which varioussubroutines an becalled,depending pon which branch s takenearlier n the pro-gram. In some situations certain causal processes and thesubroutines epresenting hosecausalprocesses) ould appear;in others heywould not.

Phase : Presentinghe Results f theStudyTo the extent that the case-studymethod has arisen from thepractice of historians, t has tended to follow certain of theirprocedures hat are not really appropriate practices or socialscientists. ne eatureof mosthistorians'work s a relative ack ofconcernwith or discussion f methodologicalssues ncounteredin theperformance f research. nother feature s a heavy elianceon simple chronologicalnarrativeas an organizingdevice forpresenting he case-studymaterials.We believe hat the firsttendencyshould be replacedby a more methodologically elf-conscious nd explicitdiscussion f the research ilemmas hatthe case-study esearcheracesn the analysisof the case,and thejustificationsor solving hosedilemmasn a particularway.Someindicationof how initial expectationsbout behaviorand initialdata-collectionuleswere evisedn the course f the studywouldpermit the readerof thecaseo makea more nformedanalysis ftheprocess y whicha case nd theconclusions asedon the case


34/38

54 A. L. GEORGE ndT. J.McKEOWNwerereached. he basis or the selection f cases nd particularproblemsor opportunities n gainingevidence bout behavior nthe case houldalsobe discussed.We seeno reasonwhy presentationof the casemust alwaysfollow a chronological narrative. As theory becomesbetterdeveloped nd as researchocuses n more tightly defined argets,there will be lesspoint in presentingong narratives.Ordinarylanguageaccountsof the casecan also be supplemented r evenreplacedentirely by such devices s computerprograms,decisiontrees, or other constructsmeant to display the logic of actors'decisions.

VI[. SUMMARYAlthough case tudiesand controlledcomparisonof cases ffer anessentially ualitativeapproach o theorydevelopment, e haveargued n this paper that qualitatiueand unscientifcarc hardlysynonymous. lthough he difficultiesn makingcausalnferencesin case tudies houldnot be dismissed,f sensible ethodologicalcriteria are applied to the performanceof the cases, he risksinherent n such nferences an be reduced o a manageableevel.Although we have discussedprocedures hat facilitate theanalysisof phenomenaby cross-caseomparisons, nd althoughwebelieve hat suchproceduresansignificantlymprovecompara-tive research,we believe hat the most distinctive eatureof casestudies s what we have termed a process-tracing pproach oanalyzing behavior. Such an. approachrepresents break fromorthodox quasi-experimentalogic, both in its emphasis n thehypgthesis-formation rocessand in i1s approach o suppgrtingclaims or the existencef some ausal rocessn agivensituation.The orthodox ogic s of l ittle usewhenone acks he ngredientsupon which its success epends-a reasonably xplicitand well-developedheory,andenoughdata o test he heoryusingstandardstatisticalmethods.For all situationswhere hese ngredientsarelacking, we believeour approachhassomething o offer.


35/38

CnscStudies ndTheories fOrganizational ecisiotr akingNOTES

I. The reader nterestedn a philosophical iscussions referred o Hanson( e58).2. A clearstatementof thisdistinction can be found in Popper 1959).3. A very interestingattempt to app ly econometric echniqueso the problemof hypothesisormation s Leamer's rgument or "data-instigated"models 1978).4. SeeLipjhart 1971, . 68a).5. A more detaileddiscussion f Mill's methods s found in George 1982).6. In practice he theory s sometimesmplicit rather than explicit.7. See. or example, ray (1957);Gardiner 1952);Weber 1949).8. Because f our lack of familiari ty with other ieldsof history.wedo not arguethat all historicalcasestudies n which argumentsabout causalprocesses re made

(e.g.,casestudiesof economicdevelopment) mploy an implicit process-tracingprocedure.9. In studies f socialsystemsather han single ndividuals he natureof rhestrategic nteractionamong actorscan be considered art of this decisionprocess.10. Further discussion f the MO methodcan be found in LawrenceMohr'spaper n this volume.

I l. Hill (1980)attempts o accomodate eamer's rgumentwithin a Bayesianframework by arguing that post-observation ypotheses re always "lurking" inthe mind of the researcher.He proposes reating the prior probability of these"lurking" hypothesess some nfinitesimally mallepsilon.12. Judging by the references e cites,Leamer was not aware of most of theliteraturewe cite.I 3. See, or example.Russett's 963)case tudy of the Japanese ttack on PearlHarbor, which was motivated in part by the unexplained indings of his earlierquasi-experimentalesearch.14. See he discussion n Section3.15. A goodexampleof theuseof quantitativeestimatesn a case tudy s Maoz's( 98l ) case tudyof the sraelidecisiono conduct he raidat Entebbe. n this studyMaoz deployedan elaborateproceduredesigned o generate robability and utilityestimates or various outcomes so that Israeli decision-makingcould then beanalyzedwithin an expected-utility ramework.16. It also seems ossible o select he single critical" case o that, under theinitial conditionsobtaining n thecase, he theoryat handgenerates dererminareprediction about behavior,but also givesus reason o attach relatively ittle con-fidence o that prediction.Allison (1971), or example,seems o have had thisstrategy n mind whenhe selectedhe Cuban missile risis or hiscase tudy.SinceAllison apparently startedwith the belief hat the circumstances f an acute nter-national crisiswould be the most propitious ones or the suppression f bureau-cratic politics by strong presidential eadership, he observation of significant

f,J


36/38

56 A. L. GEORGEndT. f. McKEOWNbureaucratic olitics n that casewouldconstituteevidencehat bureaucratic-politicsmodels of governmentdecisionmaking are more robust than previously suspected.17. As Riker (1957, pp.67-69) argues, t is desirable o possessheories hatfocus on "small," welldefinedevents. he problem s to specifya mode of analysisthat hasa reasonable hanceof reaching his felicitous state of affairs from a startingpoint of partial or complete gDorance.Somemore recentwork in the history andphilosophyof science uggestshat the strategywe haveoutlinedhasplayeda centralhistoricalrole in the developmentof such ieldsas chemistryand physics;by con-trast, it has only been n this century (if at all) that orthodox quasi-experimentalanalysis has become an importaDt part of actual researchpractice (Harre andSecord,1973:69).

18. As mathematicalmodelsbecomemorecomplex, hey oo may fail to "1's11"-not because f any flaws n their statement, ut simply becausehey are not readilysolvable. n suchsituations, he developers f suchmodelsoften turn to computersimulations.

REFERENCESAllison, G. T. (1971). Essence f decision:Explaining the Cuban ntissile crisis.

Boston: Little, Brown.Brown, S. R. (1974). ntensiveanalysis n political research. olitical Methodology,1. l-25.Campbefl, D. T. (1975). 'Degrees of freedom' and the case study. Comparatiue

Political Studies, , 178- 193.Chassan, . B. (1961).Stochasticmodelsof the singlecaseas the basisof clinicalresearchdesign.Behaaoral Science,6, 42- 50.Cohen, M. R., & Nagel, E. (1934).An introduction o logic and scientificmethod.New York: Harcourt, Brace& Co.Diesing, P. (1971).Patterns of dbcouery n the social sciences.Chicago: Aldine-Atherton.Dray, W. (1957). awsandexplanationn hbtory. London: Oxford University Press.Eckstein,H. (1975).Casestudy and theory in political science. n F. L Greensteinand N. W. Polsby (Eds.), Handbookof political science Vol. VII). Reading.MA. Addison-Wesley.Friedman, M. (1953).Essaysn positiueeconomics. hicago: University of ChicagoPress.Gardiner, P. (1952). The nature of historicalexplanation.London: Oxford Uni-versityPress.George, A. L. (1979a).The causalnexus betweencognitive beliefsand decision-making behavior:The "operationalcode" beliefsystem. n L, Falkowski (Ed.),Psychoog cal modelsand internationalpolitics.Boulder : Westview.


37/38

CaseStudiesand Theoriesof Organizational DecisionMaking 57' George,A. L. (1979b).Casestudiesand theory development:The method of struc-''

tured focused comparison. In P. G. Lauren (Ed.), Diplomatic history: NewApproaches.New York: The Free Press.. George,A. L. (1982).Casestudiesandtheory detselopment. npublished manuscript.George,A. L., & Smoke,R. (1974).Deterrmce n Americanforetgn olicy: Theory

andpractice.New York: Columbia Univers ity Press.Gregg, L. W., & Simon, H. A. (1967).Processmodelsand stochastic heoriesof

simple conceptformation. J ournal of M athematicalPsychoog , 4, 26-27 6.Hanson, R. N. (1958). Patterns of discouery.Cambridge: Cambridge University

Prcss.Harre, R., & Secord,P. F.(1973). Theexplanation f socialbehauiour. otowa, NJ:

Littlefield,Adams.Hill, B. (1980).Review of Edward E. Leamer, Specifcationsearches:Ad hoc in-

ference with nonexperimental ata. Journal of the American StatisticalAssocia-tion, 75,252-253.Kaplan, A. (1964).Theconduct f inquiry.SanFrancisco:Chandler.Kuhn, T. S. (1970).Logic of discoveryor psychologyof research? n I. Lakatos

& A. Musgrave (Eds.), Criticism and the growth of knowledge.Cambridge:CambridgeUniversityPress.

Lakatos, I. (1970).Falsification and the methodologf of scientificresearchpro-grammes. In I. Lakatos and A. Musgrave (Eds.), Criticism and the gron'th ofknowledge pp. 9l-196). Cambridge: CambridgeUniversity Press.Leamer, E. E. ( 1978) Specificationsearches:Ad hoc inferencewith nonexperimental

dara. New York: Wiley.Lijphart, A. (1971).Comparative oliticsand the comparativemethod.Americ'unPo itica I ScienceReuieu,, 5, 682-693.Lundberg, G. A. (1941).Case-studiesersus tatistics-an issuebasedon misunder-standing. Sociometry,4, 379-383.Maoz, Z. (1981).The decision o raid Entebbe.Journal of Conflict Resolution,25,

677-707.Mill, John Stuart. (1843).A systemof ogic.Neustadt,R. E. (1970).Alliancepolitics.New York: Columbia University Press.Nisbett, R., & Ross, L. (1980I Human inference: Strategies and shortcomingsof

social udgment. Englewood Cliffs, NJ: Prentice-Hall.Paige,G. (1968).The Koreandecision.New York: The Free Press.Popper, K. (1959).The ogic of scientific iscotery.New York: Basic Books.Raser, . R., Campbell,D. T., & Chadwick,R. (1970).Gamingand simulation ordeveloping theory revelant to internationa l relations.General Systems,15,

I 83- 97.Riker, W. H. (1957).Eventsand situations. ournal of Philosophy,54,57-70.Russett,B. (1963).Pearl Harbor. Deterrence heory and decision heory. Journal

of PeaceResearch,2, 9-106.


38/38

58 A. L. GEORGEndT. |. McKEOWNRussett,B. (1970). nternationalbehavior research.Casestudiesand cumulation. '

In M. Haas & H. S. Kariel (Eds.),Approacheso thestudyof potitical science.'SanFrancisco:Chandler.Scriven,M. (1976).Maximizing the power of causal nvestigations.The modusoperandimethod. In GeneV. Glass(H.), EualuationStudiesReuiew nnual(pp. l0l- l l8) . Bever lyHil ls: Sage.Smelser,N. J. (1968).The methodologyof comparativeanalysisof economicactivity. In N.J. Smelser Ed.), Essat'sn sociological xplanarionpp.62-75).EnglewoodCliffs Prentice-Hall.Smoke, R. (1977). War: Controllingescalalron.Cambridge: Harvard University

Press.Steinbruner, J. (1974). The cybernetic heory of decision.Princeton: Princeton

UniversityPress.Weber, M. (1949).Themethodology f thesocial sciencesE. A. Shils& Henry A.Finch,Trans.and Eds.).New York: The FreePress.

George and McKeown2

Documents