POPULARPARTICIPATIONANDPOLITICALVIOLENCEByGregoryWallsworthADISSERTATIONSubmittedtoMichiganStateUniversityinpartialfulÞllmentoftherequirementsforthedegreeofEconomicsÐDoctorofPhilosophy2016ABSTRACTPOPULARPARTICIPATIONANDPOLITICALVIOLENCEByGregoryWallsworthTheÞrstessayÒProtest:Onsetand(De)EscalationÓarguesthattheliteratureonCivilConßicthasreachedapointofmaturityinidentifyingthecorrelatesofconßict;however,theriskfactorsforconßictarefarmorecommonthanconßictitself.Evenwhenunderlyingconditionsappearsimilar,asforthecountriesimpactedbytheArabSpring,diverseoutcomescanarise.Bymodelingtheescalationprocessandincorporatingprotestasasignaltothegovernmentandpotentialdissidentsinsociety,thispapershowshowsimilarstartingconditionscanleadtoprotest,governmentcon-cessions,orevencivilwar.Thispaperalsocontributestounderstandingtherelationshipbetweenrepressionanddissent.Wearguethatrepressionmayreduceoveralldissent,butcausedissentthatoccurstobecomemoreviolent.Finallyweexaminesomepredictionsofthemodel;thisisdonebycomplementingtraditionalconßictdatafromUppsalaConßictDataProgram(UCDP)withdataonprotestfromtheSocialConßictinAfricaDatabase(SCAD).TheanalysisÞndssupportfortwokeypredictionsinthemodel:thelikelihoodofaconcessionincreaseswithprotestsize,andanon-monotonicrelationshipbetweenprotestsizeandtheprobabilityofescalationtoconßict.ThesecondessayÒElectoralViolence:AnEmpiricalExaminationofExistingTheoriesÓar-guesthatrecentstudiesofelectionviolencehavefoundthatviolencemarsasmanyas80percentofAfricanelections.However,thewaysinwhichviolenceisusedtoinßuenceelectionsarestillunclear.Twotheoreticalframeworkshavebeensuggested.TheÞrstarguesthatviolenceistargeteddirectlyatcoreoppositionsupportersinanattempttopreventthemfromvoting.Thesecondismorenuancedandarguesthatitismorefeasibletodeterunalignedvoterswithuntargetedvio-lencebecausetheyarelesscommittedtovoteforanyparticularparty.Apartycouldincreasetheirvotesharebyexcludingunalignedvotersiftheyhaveastrongeradvantageincoresupportersthanunalignedvoters,becauseexcludingunalignedvotersplacesmoreweightoneachpartyÕscoresup-porters.BycombiningsurveydatafromtheRound4AfrobarometersurveywitheventdatafromtheSocialConßictinAfricaDatabase,wecomparethevalidityofthesetheories.First,weconÞrmthatviolenceisassociatedwithareducedlikelihoodofvoting.Moreimportantly,weÞndimpor-tantheterogeneityinthisassociation.SupportingtheÞrstframework,voterswithastrongpoliticalafÞliationdoceasevotingiftheypersonallyfearviolence.Insupportofthesecondframework,weÞndthatunalignedvotersaretheonlygroupsigniÞcantlylesslikelytovoteinthepresenceofviolence,evenwithoutreportingagreaterfearofviolence.Weconcludethatbothtargetedanduntargetedviolencearepotentiallyeffectivestrategies,butuntargetedviolenceappearstobemorecommon.Finally,thethirdessayÒProÞlinginViolentElectionsÓarguesthatrecenttheoreticalandem-piricalresearchonelectionviolencehaspresentedseveralpotentialwaysinwhichviolencemaybeusedtoinßuencetheelectoralprocess.Akeydifferentiationbetweenemergingtheories,ashighlightedinWallsworth(2016),iswhetherviolenceistargeteddirectlyatoppositionsupport-ersorindirectlyatunalignedvotersmorelikelytovotefortheopposition.Wallsworth(2016)demonstratedthatreactionstoviolenceareconsistentwithbothstrategies.Targetedviolenceisassociatedwithalowerlikelihoodofvoting,andunalignedvotersweretheonlygrouptoreacttoindirectlytargetedviolence.OnewaytodistinguishwhichtheoryismoreviableinagivencountryistounravelhowsuccessfullyapotentialperpetratorofviolencecouldproÞletheopposition.ThispaperexaminestheviabilityofproÞling,whichcharacteristicsmaybeusedtoproÞle,whoappearstobetargetedbyviolence,andhowcharacteristicswhichcorrelatewithanindividualÕspoliticalafÞliationalsocorrelatewiththeirfearofviolence.TABLEOFCONTENTSLISTOFTABLES.......................................viLISTOFFIGURES......................................viiiCHAPTER1Protest:Onsetand(De)Escalation....................11.1Introduction......................................11.2LiteratureReview...................................41.3TheModel.......................................81.3.1BaselineModel................................81.3.2Solution....................................101.3.3Equilibria...................................131.3.4StrategicRepression.............................151.4ModelAnalysisandImplications...........................201.4.1Theuseandtypeofdissent..........................201.4.2Protest:EscalationorDe-Escalation.....................221.4.3ModelAssumptions..............................231.5DataSources.....................................251.6Results.........................................271.6.1TheImpactofProtest.............................271.7Conclusion......................................31APPENDICES.......................................33APPENDIXAFIGURESFORCHAPTER1....................34APPENDIXBTABLESFORCHAPTER1.....................42REFERENCES.......................................46CHAPTER2ElectoralViolence:AnEmpiricalExaminationofExistingTheories.492.1Introduction......................................492.2LiteratureReview...................................532.2.1CanElectionsReplaceViolence?.......................532.2.2TheImpactofViolence............................542.2.3CausesandTypesofElectoralViolence...................552.2.4TheoriesPoliticalCompetitionandViolence.................572.3ModelsandHypotheses................................592.4ComparingSimpliÞedModels............................612.4.1CollierandVicenteModel(CV).......................612.4.2ChaturvediModel(CH)............................632.4.3HypothesesfromBaselineModels......................642.5DataSources.....................................662.5.1PoliticalAfÞliation..............................662.5.2PoliticalParticipation.............................682.5.3ViolenceMeasure...............................682.6EmpiricalInvestigation................................72iv2.6.1ReactionstoViolence:TargetedandUntargeted...............722.6.2WhoFearsViolenceinElections?......................792.6.3WhereDoesViolenceOccur?........................812.7Conclusion......................................82APPENDIX.........................................84REFERENCES.......................................93CHAPTER3ProÞlinginViolentElections.......................963.1Introduction......................................963.2HowcantheviabilityofproÞlinghelpdistinguishbetweenmodelsofelectionviolence?.......................................983.3Data..........................................993.4Results.........................................1013.4.1Whofearsviolenceandwhere?........................1013.4.2WhatpredictsanindividualÕspoliticalafÞliation?..............1033.4.3Howareindividualstargetedwithviolence?.................1043.5Conclusion......................................108APPENDIX.........................................110REFERENCES.......................................120vLISTOFTABLESTable1.1PoliticalDissentData...............................43Table1.2FreedomHouseRatings..............................43Table1.3ProtestSizeandConcessions...........................44Table1.4ProtestSizeandEscalationtoConßict......................45Table1.5TransitionRates..................................45Table2.1PoliticalAfÞliationSummary...........................85Table2.2AgreementBetweenMeasures..........................86Table2.3IsPoliticalAfÞliationAccuratelyReported?...................86Table2.4ViolenceSummary,RegionalLevel........................86Table2.5SimpleCorrelationsSurveyandEventViolenceMeasures............87Table2.6AdditionalComparisonofViolenceMeasures..................87Table2.7BaselineImpactofViolenceonVoting......................88Table2.8BaselineResults,HeterogenousRelationshipBetweenVotingandViolence...89Table2.9ExpandedSpecifcationsoftheimpactofViolenceonVoting...........90Table2.10PerceptionsofViolence..............................91Table2.11DivisionoftheElectorateandViolence:UnalignedVoters............92Table3.1MeanViolenceandPoliticalAfÞliationBreakdownbyCountry.........112Table3.2MeansofDemograhicData:OverallandbyCountry...............113Table3.3LinearProbabilityModelsPredictingProbabilityofSupportingMajorityParty.114Table3.4LinearProbabilityModelsPredictingProbabilityofSupportingLargestOppo-sitionParty.....................................115viTable3.5DemographicTraitsandFearofViolence.....................116Table3.6ComparingPredictorsofViolenceandPoliticalAfÞliation............117Table3.7VarianceinFearofViolenceExplained......................118Table3.8MeanviolencebyAfÞliation,DemeanedbyDistrict...............119viiLISTOFFIGURESFigure1.1FormalGameTree.................................35Figure1.2ImpactofCgonEquilibriumStrategies......................36Figure1.3EquilibriumPathsfor!and"0..........................36Figure1.4ImpactofCronEquilibriumStrategies......................37Figure1.5Impactof#0onEquilibriumStrategies......................37Figure1.6ProtestSizeandEscalation............................38Figure1.7EscalationversusProtestsize...........................38Figure1.8EscalationversusProtestsize...........................39Figure1.9EscalationversusProtestsize...........................39Figure1.10EscalationversusProtestsize...........................40Figure1.11EscalationversusProtestsize...........................40Figure1.12EscalationversusProtestsize...........................41Figure1.13ProtestSizeandConßictEscalation........................41Figure3.1ComparingPoliticalafÞliationandFearofViolence...............111viiiCHAPTER1Protest:Onsetand(De)Escalation1.1IntroductionCivilwarisarelativelyrarephenomenon,thoughtheriskfactorsforwararenot.AsstatedinWalter(2009),Òexistingstudiescannotexplainvariationintheoutbreakofviolenceacrosscountriesthatareatsimilarriskofcivilwar.ÓShecontinuestoarguethat,ÒByviewingthedecisiontoÞghtaspartofalargerbargainingprocessandnotasanisolatedevent...scholarscanbetterexplainwhyviolenceismorelikelyinsomecountriesÓ(p.244).TheeventsintheArabSpringdemonstratetheaccuracyofWalterÕsstatementsandservetomotivatethispaper.Protestsoccurred,tosomedegree,inmorethanadozencountriesacrossNorthernAfricaandtheMiddleEast.Existingmacroeconomicconditionsweresimilarinmanyofthesecountries,aswasthewayinwhichthemassmovementsagainsttheirrulingregimesstarted.However,thesequenceofeventsfollowingtheprotestsvaried,asdidtheexpectationstheseregimeshadestablishedforhowtheywouldrespondtoprotest.Intheend,someregimesstayedinplacewithlittlechange,sometoppledpeacefully,othersviolently,andsomestillremainembroiledincivilwar.Thispaperprovidesapotentialexplanationforhowdramaticallydivergentresultscanarisefromsimilarinitialconditions,likethosefollowingtheinitialwaveofprotestintheArabSpring.ByincorporatingprotestintoamodelofconßictwecanaddressWalterÕscommentsdirectly.Withprotestasanoption,usedtosignalthestrengthofpotentialrebelgroupstothegovernment,conßictoccursonlywhenthesignalconveysinaccurateinformation.Thisset-upleadstonovelconclusions.Aswithmanymodels,dissentisdrivenbyeconomicconditions.However,weÞndtheformittakesisdrivenbyexpectedstrategicinteractions,speciÞcallyexpectationsofrepres-1sion,andnon-economicdeterminantsoftheviabilityofconßict.Thisimpliesthatsimilarmacroe-conomicconditionscouldleadtodifferentoutcomes.Additionally,modelingrepressionasacosttoprotestallowsustoexaminewhenitisaneffectivetooltoquelldissent.However,wearguethatgovernmentsmaynotbecapableofquicklyloweringtheexpectedlevelofrepression,whichcanleadtoconßictevenwhenbothpartieswouldprefertoavoidit;because,toohighanexpectedlevelofrepressioncancauseviolentdissenttooccur.WealsoÞndacounterintuitive,non-monotonicrelationshipbetweenprotestandescalationtoconßict,wherethelikelihoodofconßictinitiallyincreasesinprotestsize,thendropstozeroonceprotestsizesurpassesathreshold.Theworlddepictedinthefollowingmodelisonewheretwogroupsinsociety,thegovernmentandtherebels,mustagreetosplitavailableresources.Atthebeginningoftheinteraction,policydeterminestheallocationofresourcestoeachgroup.Afterseeingtheallocationtherebelsdecidetousedissent,andwhetherornotthatdissentwillbeviolent.Protestisusedbydissidentstolearnabouttheirownstrength,whilesimultaneouslysendinganimperfectsignalofthatstrengthtothegovernment.Itischosenwhenthecost,ortheexpectedlevelofrepression,isrelativelylow.Ontheotherhand,iftherebelsarefairlyconÞdentoftheirownstrengthandtheexpectedlevelofrepressionisrelativelyhigh,theymaychoosetouseviolentdissentimmediately.Afterprotest,therebelslearntheirtypeprecisely,whereasthegovernmentonlyobservesanoisysignal,e.g.protest,whichislikelytobelargeriftherebelsarestrong.Afterobservingthesignal,thegovernmentisthenabletoadjustthepolicyinanattempttoappeasetherebelgroup,orquelltherebellion.Variousconditionsinthemodelcanpredictpeace,immediateconßict,orprotestthateitherescalatesintoconßictorresultsinapeacefulbargain.Immediateconßictoccursinresponsetogovernmentsthatareexpectedtovigorouslysuppressprotest,makingthecostofprotesttoohighforrebelgroupstouse.Protestoccurswhenevertheexogenouslysetpolicyisunfavorabletotherebelsandthethreatofrepressionislow.Followingprotesttheinteractionendsorescalationoccurs.Escalationtoconßictoccursafterprotestonlyifthegovernmentmakestoosmallacon-cessiontoarebelgroupthatwasinfactstrong.Peacemayfollowprotestinoneoftwoways:a2successfulprotest,wheretheproteststhemselvesgarnermajorconcessions,orafailedprotestbyaweakrebelgroup,wheretherebelsrationallyacceptminorconcessions.Previousresearchhasdocumentedacommonpatternbetweenviolentandnon-violentdissent,periodsofviolentdissentareoftenprecededbyperiodsofnon-violentdissent(Gurr,2000).ToexamineourmodelÕspredictionsdescribingtherelationshipbetweenviolentandnon-violentdis-sent,wecombinetwodatasetswithcomplementarymeasuresofpoliticaldissent.TheÞrstisthePeaceResearchInstituteatOsloÕs(PRIO)ArmedConßictDataset(ACD),whichweusetomea-surecivilconßict;thesecondistheSocialConßictinAfricaDatabase(SCAD),fromwhichwederivemeasuresofprotest.Thesedatacover42Africancountriesfrom1990-2012.Thesedataprovideauniqueopportunitytoexaminethefullspectrumofpoliticaldissent,andthedeterminantsofescalation.AtÞrstglancethedatashowthatnon-violentdissentisapredictorofviolentdissent.Themaincomponentoftheanalysiswhichfollowsisexploringtheconditionsthatmakeprotestmorelikelytode-escalate,escalate,orgarnerconcessions.WeÞndsupportforourmainhypotheses.First,weobserveanon-monotonicrelationshipbetweenprotestsizeandconßictescalation,wherethelargestprotestsobservedinthedatadonotescalate.Second,weÞndthepredictedlinearrelationshipbetweenprotestsize,andtheprobabilityofaconcession,measuredasanincreasesincivillibertiesorpoliticalrightsusingFreedomHousedata.Finally,weexpandthemodelandallowthegovernmenttosetthelevelofrepressionstrategi-cally.Wearguethatthebaselinemodel,withrepressiontakenasexogenous,maybeappropriateifthegovernmenthasalonghistoryofbeinghighlyrepressive,orcannotcrediblycommittolowerlevelsofrepression.Endogenizingrepression,undertheassumptionthegovernmentcannotcom-mit,leadstoasimilarsetofresults,andweÞndthatthecostassociatedwithconßictbecomesthemaindeterminantofthetypeofdissent(protestorconßict)therebelgroupwilluseintheÞrstperiod.31.2LiteratureReviewThereisavastliterature,boththeoreticalandempirical,examiningthecausesofcivilconßict.Un-derlyingmuchofthisliteratureistheassumptionthatconßictisrational,aconceptformalizedbyFearon(1995).Thebindingprincipleamongrationalconßictmodelsisthatsomeagreementex-iststhatallsideswouldprefertoavoidconßict;however,someinefÞciencypreventsanagreementfrombeingreached.Inhisseminalpaper,Fearon(1995)laysoutthreepossiblecausesofrationalconßict:informationasymmetrieswithanincentivetomisrepresent,commitmentproblems,andissueindivisibilities.Severalrecentstudies,including(BlattmanandMiguel,2010;Walter,2009),askwhymoreimplicationsofrationalconßictmodelshavenotbeentested.ThebasicpremiseoftheseargumentsissummarizedwellbyWalter(2009):Moststudiesofcivilwarhavefocusedontheunderlyingstructuralconditionsthatencouragegroupstogotowarratherthanonthebargainingproblemsthatmaystandinthewayofsettlement...ByviewingthedecisiontoÞghtaspartofalargerbargainingprocessandnotasasingleisolatedevent,scholarscanbetterexplainwhybargainsaresorareincivilwarsandwhyviolenceismorelikelyinsomecountriesthanothers.(p.244)Usingprotestasasignal,ofrebelstrength,inarationalmodelofconßictallowsustoderiveandtesthypothesesrelatedtotheaforementionedbargainingproblems,asopposedtoexaminingonlythestructuralconditionswhichleadtoconßict.Wearguethatthebasicconditionsthathavebeenfoundtoprecipitateconßict,suchaspooreconomicconditionsorexistingpoliticalinstabilitydoinducedissentMigueletal.(2004),HegreandSambanis(2006).However,theformofdissentdependsontheoutcomeofactionstakenbypotentialdissidentsandthereputationestablishedbythegovernmentforreactingtonon-violentdissent,speciÞcallytherepressionofprotest.Thetheoreticalliteratureonconßictandprotestbothhaveadiverseselectionofmodelsusingasymmetricinformationastheprimaryfrictiondrivingconßict(Powell,2002).ModelslikeChas-4sangandPadro-iMiquel(2009)useinaccurateinformationonthestateoftheworldtogenerateconßict,otherssuchasWittman(1979)ormorerecently,BaligaandSjıstrım(2004),useprivateinformationonthemilitarycapabilitiesofeachside;ourmodelmakesasimilarassumption.Thefocusinthispaper,asopposedtomuchoftheliterature,isnotinexaminingtheconditionsun-derwhichasymmetricinformationdoesordoesnotleadtoconßict,butratheronexaminingifprotestispotentiallyonewaytoconveythisprivateinformation.AspointedoutbyFearon(1995),asymmetricinformationaloneisnotenoughtogenerateconßict;itrequiresanincentivetomis-srepresentthatinformationoraninabilitytoconveyitaccurately.Weexpandonthisliteraturebyinvestigatingifprotestwaybeonewaytoconveythatassymetricinformationinanattempttoavoidconßict.Manymodelsalsoexistwhichuseprotestasasignalinaglobalgamessetting(CarlssonandVanDamme,1993;MorrisandShin,2001).Thispaperdeviatesfromthetraditionalset-upofmodelslikeMelosh(2012)tofocusontheescalationprocessandthegovernmentÕsresponseratherthantryingtoexplainhowmovementsovercomeproblemsofcollectiveaction.Thecanonicalframeworkforsuchamodelusesathresholdvalueforthesizeoftheprotest,Granovetter(1978),andgenerallyarguesthatifsuchavalueisexceeded,theprotestmovementwillgrowmassiveandsucceedinobtainingitsdemands.Weagreewiththebasicintuitionbehindthresholdmodels:largeenoughprotestsdosucceedinconvincinggovernmentstograntmajorconcessions.However,ourmodelmakesamajordeparturewhenexaminingwhathappensastheprotestsizeincreasestowardsthisthreshold.Wepredictthatitisinthesecircumstances,whenitismostlikelythatthegovernmenthasmisidentiÞedastrongrebelgroup,thatconßictismostlikelytoensue.Empirically,thispapercontributestoasmallbutgrowingliteratureattemptingtoviewconßictonagreatercontinuum.EarlyworksuchasMigueletal.(2004)acknowledgedmeasuringonlytheoccurenceasaweaknessrelatedtotheavailabledata.OtherrecentworksuchasChaudoinetal.(2013)andBesleyandPersson(2009)haveleveragedimprovementsinavailabledata.OurcontributionrunsclosesttoBesleyandPersson(2009)inthisregard:ratherthantryingtomoreaccuratelypredictthebreadthorintensityofconßict,wearemostinterestedinexaminingthe5determinantsofdifferenttypesofdissent.SpeciÞcally,weexaminewhennon-violentdissentismostlikelytoescalatetoviolence.Anadditionalcontributionofthemodelinthispaperisanattempttobringtogetherliteraturesonprotest,repressionanddissent,andconßict.BydoingthisweareabletoprovideapossibleexplanationforDavenportÕsÔpunishmentpuzzleÕ(Davenport,2007):theideathatdissentisalmostalwayspositivelycorrelatedwithrepression;however,theimpactofrepressionondissentishighlyinconsistent.Inparticularweprovideananswertoonequestionhehighlights:ÒUnderwhatcircumstancescanauthoritiesreducedissent?ÓOurmodelarguesthatdependingonthestateoftheworld,repressioncanhavenoimpactondissent,eliminatedissentcompletely,orinduceachangeinthetypeofdissentrebelgroupswilluse.Weprovidepotentiallytestableimplicationsthatmayhelptoexplainotherwiseinconsistentresultsintheempiricalliteratureexaminingtheimpactofrepressionondissent.Thispapermakessomekeydeparturesfromthetheoreticalliteratureonrepressionanddissent.Byreturningtoanearlierviewpoint,whererepressionwasmodeledasincreasingthecostofactsviewedasthreateningtothepowerofthestate,(Goldstein,2001).However,wereÞnethisbyarguingthatrepressioniseffectiveonlyagainstnon-violentdissent.Ifagroupdecidestouseviolenceagainstthegovernment,theyhavealreadyacknowledgedthatitbecomesacceptableforthegovernmenttorespondinkind.SimilartoPierskalla(2010),wearguerepressionisabletoquellprotest.However,weareabletogenerateanyofthemodelÕspaths:conßict,peace,orprotest,asequilibriumoutcomeswithoutneedingtointroduceathirdparty.ThekeydeparturefrompreviousworkforthispaperandinPierskalla(2010),isincorporatingthiscostintoastrategicinteractionwherebothactors!,therebelsandthegovernment,decisionprocessesareexplicitlymodeled.PreviousmodelssuchasLichbach(1987)focusedonthetypeofdissentusedbyrebels,whereasothermodelssuchasMoore(2000)focusedonthegovernment!sdecisiononhowtouserepression.ThedecisionprocessinbothofthosemodelswasdrivenbyacostbeneÞtanalysisofthechoicesfacedbythemodelÕsagents.TakeforexampleMoore(2000),themodelÕsdissidentsusewhichevertypeofdissentis6mostcost-effectivetogetthegovernmenttoaltertheirpolicies;however,thegovernmentÕsdecisionisnotmodeled.Byincorporatingbothdecisions,wecanseewhenandwhyprotestispreferredtoconßict.Arecenttrendintheempiricalliteratureondissenthasbeencomparingviolentandnon-violentdissent(ChenowethandCunningham,2013),thispaperalsocontributestothisliteraturebydistin-guishingbetweenthetypesofdissentusedbypotentialrebelgroupsCunningham(2013).Muchoftherecentworkintheliteraturecomparingviolentandnon-violentdissenthasfocusedonthecomplementarityofdifferenttypesofpoliticaldissent.Wedepartfromthisandexamineanempir-icalobservationmadeinthispaper,andinearlierworkbyGurr(2000):whydoesprotestsooftenprecedeconßict,andwhatdeterminesthepathittakes.Byjoiningtogetherintuitionfromrelatedliteraturesonrepression,protest,andconßict,thispaperprovidessomepotentialexplanationsforunresolvedquestionsfoundineachliterature.WeprovideonepotentialexplanationforinconsistentÞndingsontheimpactofrepressionondis-sent,highlighttheroleofbargaininginexplainingwhysomecountriesareabletoavoidconßict,andmoveinthedirectionoftestingimplicationsofrationalconßictmodelsratherthananalyz-ingwhatstructuralconditionsencouragedissent.ThoughwedonotprovidedeÞnitiveanswerstoanyoftheseopenquestions,weprovidedirectionandhighlighttheadvantagesofdrawingfromthediverseliteratureencompassingpoliticaldissentasawholeratherthanfocusingsolelyononeparticulartypeofdissent.71.3TheModel1.3.1BaselineModelTheRebels(R)andGovernment(G)shareaprizenormalizedto1,andtheirrespectivesharesare"and1!".ThisleadstothefollowingpayofffunctionsUR=",and(1.1)UG=1!".(1.2)Tostart,"="0,and"0istakenasexogenous.ForG,strengthissetandispublicknowledge.However,RÕsstrengthisunknowntobothRandG.Wearguegovernmentresourcesarerelativelywellknown,whereastheactuallevelofsupportforthosewishingtoopposethegovernmentisnotlikelytobewidelyshared.SpeciÞcally,Risoneoftwotypes:strong(Rs)orweak(Rw).GandRshareacommonpriorof#0onRÕstypebeingRs1.Afterobserving"0,Rmayeitherpayacost!toprotest(P),allowingGtoshift"from"0to"1,orinitiateconßict(C).Gmayonlyadjust"followingaprotest2.IfPwaschosen,Rpaysacost!,learnstheirtype,andanoisysignal,P,isobservedbyG.Pcanbethoughtofasthesizeoftheprotest.ThesignalP"Fi(p)issuchthatforP1>P0F(P1|s)F(P1|w)>F(P0|s)F(P0|w)3.Thisimpliesthatasprotestsizeincreases,theprobabiltityitwasgeneratedbyastrongrebelgroupincreases.Wearguethatprotestcouldberandomforanumberofreasons:weatherandothernaturalphenomonamaypreventsomeindividualsfromattending,issuesofcollectiveactionmaytakelongertoovercomethantheinitialwaveofprotest,ortimingmay1ThisparticularinformationstructureisasimpliÞcation.WhatisnecessaryisRstartingwithsomeuncertaintyovertheirowntypeandreceivingamoreprecisesignalthanG.2ThisassumptionisjustiÞedbyarguingthatseizingmoresurplusorsettinglessfavorablepolicies,ifnotdoneinresponsetocivilunrest,wouldimposetoohighacostintheformofpotentialinternationalsanctionsagainstthecountry.Forexamplesomepreferentialtradeagreementstiehumanrightscompliancetotheirexistence,Davenport(2007)Hafner-Burton(2005).Thisisoneexampleofacosttorepressionforthegovernment.3Fi(p)isanyCDFthathasthemonotonelikelihoodratiopropertyandanunboundedlikelihoodratio.8bepoorlycommunicatedasamovementbegins.Regardlessoftheexactreason,wearguethatthoseinvolvedintheprotestwillgainmoreaccurateinformationthanthegovernment,whoonlyobservestheprotestÕsvisibleoutcome.Thecostofprotest,!,couldbethoughtofasrepresentingmanythingsbutismeanttocapturetheexpectedlevelofrepressioninthegovernmentÕsresponsetoprotest4.Followingprotest,GmakesanoffertoRbasedonGÕsupdatedbeliefofRÕstype.Shouldconßictoccur,byeitherRrejectingGÕsofferorRchoosingconßictintheÞrstperiod,eachsideincursacostthatispaidregardlessoftheconßictÕsoutcome.CristhecostofconßictforR,andCgisthecostofconßictforG.Additionally,TiistheprobabilityRiwins,withthisprobabilitybeinghigherforstrongrebelgroups.Thevictorcanset"atanypointoftheirchoosing.Setting"attheirpreferredpoint,of1or0,impliesthefollowingexpectedpayoffsfromconßict5E[UR(C)]=Ti!Cr,and(1.3)E[UG(C)]=1!Ti!Cg.(1.4)Formally,thetimingofthegameisasfollows.1.Thepolicy"isexogenouslysetat"0,andnaturedrawsrebeltype,eitherRsorRw,withprior#0onthetypebeingRs.Atthisstage,allplayershavethesameinformation,withrebeltypebeingunknown.2.Rthenhasthechoiceto,initiateconßict(C),protest(P),orstayhome(!).FollowingCor!,thegameends,witheachreceivingtheconßictpayofforpayoffdeterminedby"0,respectively.3.IfPwaschosen,Rpaysacost!,learnstheirtypeandanoisysignalisobservedbyG.4ThegovernmentÕsstrategiclevelofrepressionisexaminedinthefollowingsection.5ApayoffstructurelikethiscouldbederivedfromatraditionalconßictsuccessfunctionSkaperdas(1996).94.Followingthesignal,GupdatestheirbeliefsofRÕstype,andisabletochange"0to"1anywhereontheunitinterval.5.Finally,Rcaneitheraccept"1orchoosetoinitiateconßict(C).ThelayoutofthegamecanbeseeninFigure1.1.1.3.2SolutionAperfectBayesianequilibrium(PBE)oftheextensiveformgame!isastrategyproÞleforR,G,andbeliefs.ThesebeliefsareconsistentandupdatedusingBayesÕrule.Actorsmaximizeindividualpayoffs,andweassumeintheeventofatie,conßictisavoided.ForR,astrategyconsistsofaninitialdecisionof{P,C,!},andadecisionto{Accept,C}asafunctionofGÕsofferfollowingprotest,"1,andtheirowntype.ForG,astrategyconsistsofanoffer"1#(0,1)asafunctionofthesignal,wherethechoiceof"1isconsistentandrational.Inordertolookatpossibleequilibriainthegame,itisusefultostartbyderivingseveralconditions.Usingbackwardsinduction,theofferseachrebeltypewillacceptattheÞnalnodeareRs:"1$Ts!Cr%"&&,and(1.5)Rw:"1$Tw!Cr%"&,(1.6)wherethesereservationpayofflevelscomefromtheconßictoption.Backinguptothepreviousnode,itisclearthatGwilloffereither"&or"&&for"1.Anyofferintherange("&&,1]isdominatedbyoffering"&&.AhigherofferwouldstrictlylowerGÕspayoffbecauseeitherrebeltypewouldaccept"&&.Foroffers#("&,"&&),Gwouldbebetteroffering"&.OffersinthisrangeareonlyacceptedbyRw,implyingGcouldraisetheirpayoffbyofferingthelowestofferacceptedbyRw:"&.Finally,offersin[0,"&)canbeeliminatedbycomparingGÕspayofffromguaranteedconßicttooffering"&,becauseofferingbelow"&guaranteesconßict.UsingEquation(1.4)and"&weseethatoffering"&andavoidingconßictwithRwimprovesGÕs10payoffbyCr+Cg.Thisisbecausethegovernmentisabletosavethecostofconßict,byofferingtherebelstheirconßict-basedreservationpayoff.DeÞning#1asGÕsupdatedbeliefoffacingthestrongtypefollowingtherealizationofthesignalP,wecanseeGwilloffer"&&ifandonlyif1!"&&$(1!#1)!1!"&"+#1(1!Ts!Cg).(1.7)TheLHSoftheinequalityistheguaranteedpayoffofmakingalargeconcession:offeringthestrongrebeltypeÕsreservationpayoff.ThisisbalancedagainsttheRHS,whichcorrespondstooffering"&.Iftherebeltypeturnsouttobeweak(probability1!#1),thentheofferisaccepted.Otherwise,conßictwiththestrongtypeensues.RearrangingEquation(1.7)andusingEquations(1.5)and(1.6)towriteitintermsofthelikelihoodratioyieldsthefollowingconditionforGtomakethehigheroffer:#11!#1$Ts!TwCr+Cg.(1.8)UsingBayesÕrule,theLHSofequation(1.8)canberewrittenasafunctionoftheprior,#0,andprotestsize,P.Thisleadsto#11!#1=#01!#0'f(P|Rs)f(P|Rw)$Ts!TwCr+Cg.(1.9)DeÞnedimplicitlyinEquation(1.9)isathresholdvalueforP,callitP',whichmakesEquation(1.9)holdwithequality.ForvaluesofP$P',Gwilloffer"&&,whileforP<P',Goffers"&.P'isthesmallestprotestforwhichthegovernmentÞndsitoptimaltomakealargeconcession,ratherthanriskingconßictwithstrongrebeltypes.UsingthegovernmentÕsdecisionrule,thepayoffstotherebelsÕstrategychoices,C,P,and!canbederived.TherebelsÕÞrst-perioddecisioncanbecharacterizedbycomparingthepayoffofthethreestrategychoices.First,thepayofffordoingnothing,!,isÞxedat"0.Second,theexpectedpayoffforimmediateconßict,C,comesfromaweightedaverageofthe11conßictpayoffforthetwopossiblerebeltypesbecausetherebelgroupdoesnotknowitsowntypebeforechoosinganinitialstrategy.Usingtheprior#0andeachtypeÕsexpectedconßictpayoffleadsto#0'Ts+(1!#0)'Tw!Cr%"c.(1.10)Finally,therebelsÕpayofftoprotestcanbederived.Afterprotestoccurs,thepayoffisdepen-dentonGÕsdecisionruleonlyiftherebelgroupendsupbeingweak.Thisisbecausethestrongrebelgroupwilleitheracceptanofferof"&&orinitiateconßict;bothpayTs!Cr.However,fortheweakrebelstheirpayoffcouldbeTs!CrorTw!Cr,dependingontheprotestÕsoutcome.UsingthisresultandP'asdeÞnedabove,wecanwritetheexpectedpayoffforprotestas#0(Ts!Cr)+(1!#0)[Fw(P')'Tw+(1!Fw(P'))Ts!Cr]!!="c+(1!#0)(1!Fw(P'))(Ts!Tw)!!.(1.11)TheÞrsttermistheprobabilitytherebelgroupisstrong,timestheirguaranteedpayoffofTs!Cr,thesecondtermistheprobabilityofaweakrebelgrouptimestheweightedaverageoftheirpotentialpayoffs,whichdependontheprotestÕsoutcome.TheequationcanthenberewrittenasEquation(1.11),whereweseetheexpectedpayofftoprotestistheexpectedconßictpayoffplusapremiumrelatedtotheprobabilitytherebelsareweak,lessthecost!.ComparingEquations(1.10)and(1.11)leadsustothefollowingconditionforrebelstopreferprotesttoconßict:!((1!#0)(1!Fw(P'))(Ts!Tw)%$(1.12)ExaminingEquation(1.12)yieldsintuitionastowhenRprefersprotesttoconßict.TheLHS!isthecostonepaysforchoosingprotest,andtheRHSisthebeneÞt,whichcanbebrokenintothreepieces.First,(1!#0)istheprobabilityRisaweaktypewhocanpotentiallybeneÞtfromprotest.Second,(1!Fw(P'))istheprobabilityaweakrebelwillproduceasignallargeenough12toconvinceGtooffer"&&.Finally,(Ts!Tw)istheactualdifferencebetweenthestrongandweakrebelsÕoffers,"&&!"&.1.3.3EquilibriaWedeÞnethepossibleequilibriawithrespecttotherebelsÕinitialstrategychoice.Thisisdonebecausethedecisionruleforthegovernmentisnotdependentonthestrategychoiceoftherebelsbutratherontheoutcomeofthesignal.Thisleadsustothefollowingpotentialequilibria:1.ProtestEquilbriumRebels:{P;(Rs:acceptif"1$"&&;Rw:acceptif"1$"&)}2.ImmediateConßictEquilibriumRebels:{C;(Rs:acceptif"1$"&&;Rw:acceptif"1$"&)}3.PeaceEquilibriumRebels:{!;(Rs:acceptif"1$"&&;Rw:acceptif"1$"&)}¥ForallequilibriaGovernment:offer"&&ifP$P',"&o.w.Proposition(1)deÞneswheneachofthepotentialequilibriaoccur.Proposition1.Equilibriumwithexogenousrepression:ThereisauniqueperfectBayesianequi-libriumdeterminedasfollows:1.For"0<"cand!<$,Protest2.For"0<"cand!>$,ImmediateConßict3.For"0>"cand!<"c+$!"0,Protest4.For"0>"cand!>"c+$!"0,Peace13Figure1dividesthe("0,!)parameterspaceaccordingtotherebelsÕequilibriumstrategychoiceandcanbeusedtoproveProposition1graphically.ThethreelinesinthegraphrepresenttheindifferenceconditionsbetweeneachpairofstrategiesandareequivalenttotheconditionspresentedinProposition1.Theverticallineat"cistheexpectedvalueofconßict,orthepeace-conßictindifferencecurve.Thehorizontallineat$correspondstotheconßict-protestindifferencecurve.Finally,thedownwardslopinglineistheprotest-peaceindifferencecurve.TheÞguredepictsclearlywheneachofthepotentialequilibriaoccur.For"0<"c,dissental-waysoccursbecauseconßictdominatespeaceinthisregion.Findingthatpoorconditionsgenerateconßictisnotsurprising;however,thetypeofdissentdependsonthelevelofrepressionexpectedbytherebels,!,comparedto$.Fortoohighalevelofrepression,protestbecomescostlytotherebelsrelativetoconßict.Thisisakeyresult,thatalowstateoftheworldencouragesdissentbutdoesnotnecessarilycauseconßict.Furthermore,wegenerateacounterintuitiveresultwithrespecttorepression;ifconditionsareverypoor,repressionisunlikelytobeaneffectivetoolatquellingdissent,deÞnedasprotestorconßict.Thisisbecausetherebelsalwayshaveconßictasanoption,andalthoughrepressionmaypreventprotest,byraisingitscost,itleavesconßictastheonlyviablealternative.Finally,weseeforvaluesof"0greaterthan"c,therebelschooseeitherpeace(!)orprotest.Protestischosenforrelativelylowcombinationsof"0and!.Whenpeaceweaklydominatesconßict,protestcanstillachieveshigherpayoffsthanconßict,grossofrepression.Thisresultsfromthepremiumprotestearnsoverconßict.Hereweseeanothercounterintuitveresultwithrepression:onlywheneconomicconditionsarebetterthansomeminimumthreshold,"c,canrepressionbeusedtoquelldissent.Perhapsthemostinterestingresultfromthisset-upisseeingexplicitlyhowhavingconßictasareservationpayoffdriveswhenpoliticaldissentoccurs.Ifthestateoftheworldiseverbelowthisvalue,someformofdissentisgoingtohappen.However,theformpoliticaldissenttakesdependsonthegovernmentÕsactions.Thisisinterestingbecausewegenerateafewcounterintuitiveresultsrelatedtotheimpactofrepression,andprovideapossiblewaytoexplainthevariationinconßict14outbreakacrossostensiblysimilarcountries.Lowlevelsofrepressioncanleadtomorepoliticaldissent,intheformofprotest,buthighlevelsofrepressionswitchthetypeofdissenttoconßict.Additionally,repressionisonlyeffectivewhenthestateoftheworldisbetterthantheconßictoption.Thisimpliesthatcountrieswithespeciallypoorconditionsmayactuallywanttoencour-ageprotestandtakeactionstopreventconßictfromoccurring.Ironically,thisalsoimpliesthatcountrieswhereconditionsareslightlybettermaywanttouserepressiontopreventanychanceofconßict,whichcouldresultfromfailedprotests.Theseresultsraisethequestionastowhattheoptimallevelrepressionshouldbe,whichisanalyzedinthefollowingsection.However,modeling!asÞxedisappropriateifweareinterestedinshortrunchanges.Consideringthattherebelsmustmaketheirdecisiontoprotestbeforerepres-sionactuallyoccurs,theyarelikelytoplaceagreatdealofweightonpastexperienceinteractingwiththegovernment.Itwouldbeveryeasyforagovernmenttomaketheresponsetoaverysmallincidentdramaticandpublic,raisingtheexpectedlevelofrepression,butitcouldtakemuchlongerforagovernmenttoconvincinglycommittoalowerlevelofrepression.1.3.4StrategicRepressionIntheliteratureonrepression,twoimportantquestionsare:whatpurposeitservesandwhatisitsrelationshipwithdissent?Sometheoreticalmodelshavearguedthatveryhighlevelsofrepressionshouldquellalldissent,forexamplePierskalla(2010).Forthepurposeofthissection,wewillexaminewhatthelevelofrepressionwouldbeifGcouldcostlesslyset!atthebeginningofthegame.Whenarangeofvaluesispossible,weassumeGchoosestheminimumpossiblelevelofrepression6.Examiningtheuseofrepressionhighlightstheissuecausedbytheopposingnatureofthepref-erencesforthegovernmentandrebelsinthegame.Thegovernmentcanuserepressiontoimpactthetypeofdissentusedbytherebel.Figure1highlightedthistradeoff;repressioncaninßuencethechoicebetweenprotestandconßictforvaluesof"0<"c,andthechoicebetweenpeaceand6Thiscouldbedonebyimposingasmallincreasingcostforrepression,orthroughtheuseoflexicographicpreferences.15protestforvaluesof"0#["c,"c+$).Usingourunderstandingofhowthelevelofrepressionim-pactstherebelsÕequilibriumstrategychoice,andcontinuingwiththeuseofbackwardsinduction,weonlyneedtoÞgureoutwhenthegovernmentpreferswhichrebelstrategychoice.Westartbyderivingtheexpectedpayoffstothegovernmentforeachrebelstrategy.Thetotalpayoffavailableinthegameex-anteis1!(Cr+Cg)'P(ConßictOccurs).Withpeace,conßictdoesnotoccur,andtherebelsreceiveapayoffof"0.SinceRandGshareatotalpayoffof1,thisimpliestheGovernmentÕspayoffisPeace:1!"0(1.13)Withprotest,conßictoccursif:(1)therebelgroupisinfactstrong,probability#0;and(2)thestrongrebelgroupfailstoaccuratelyconveytheyarestrongtothegovernment.Thatoccurswhentheprotestisrelativelysmall,speciÞcallylessP',probabilityFs(P').Thisleavesanaggregatepayoff,grossofrepression,of1!#0Fs(P')(Cr+Cg).Thisisthetotalavailablesurplus,lessthedeadweightlossofconßicttimestheprobabilityconßictwilloccur;aftersubtractingofftherebelsÕpayoffwegetProtest:1!"c!$!#0Fs(P')(Cr+Cg).(1.14)Finally,forconßict,theaggregatepayoffis1!(Cr+Cg).SubtractingofftherebelsÕexpectedpayofffromconßictleavesthegovernmentwithImmediateConßict:1!"c!(Cr+Cg).(1.15)Fromhere,wecancomparethepayoffsfromeachtodeterminewhenthegovernmentpreferseachstrategy.WefocusÞrstonexaminingthecasewhenProtestdominatesConßict.Substitutingtherespec-tivevaluesintothepayoffsforeach,weseethatprotestdominatesconßictif16Cr+CgTs!Tw$(1!#0)(1!Fw(P'))(1!#0Fs(P'))%A.(1.16)ItisstraightforwardtoshowtheRHSoftheinequalityisalwayslessthanone.ThisgivesLemma1.Lemma1.ForCr+CgTs!Twlargeenough($1issufÞcient),thegovernmentprefersprotesttoimmediateconßict.Thisconditionsaysthatwhenthecostofconßict,Cr+Cg,islargerelativetothepotentialgainfromriskingconßict,Ts!Tw,thenthegovernmentwouldprefertoallowsprotestinordertogainmoreinformationontherebelÕsstrength.Nextweexaminewhenpeacedominatesprotest,thisisdonebyadirectcomparisonofthegovernmentÕspayoffsforeach,andresultsinthefollowing"0("c+$+#0Fs(P')(Cr+Cg).(1.17)ToÞndthegovernmentschosenlevelofrepression,itisusefultounderstandwhenandhowrepressionimpactstherebelsÕstrategychoice.ThisismadeclearinProposition1andFigure1.2.For"0<"c,therebelswillonlychoosebetweenprotestandimmediateconßict,andrepressiondirectlyimpactsthatchoicebyactingasacosttoprotest.Forvaluesof"0#("c,"c+$),therebelswillchooseeitherprotestorpeace,andagainthischoicedependsonthelevelofrepression.Finally,for"0>"c+$,therebelsalwayschoosepeace,andthelevelofrepressionhasnoimpactonthatchoice.Thisimpliestherearetwocasestoexamine:"0<"c,and"0#("c,"c+$).If"0<"c,therebelswillonlyuseprotestorimmediateconßict.So,Equation(1.16)determinesthegovernmentÕspreferredstrategychoice.ExaminingEquation(1.16)showsthisisindependentof"0,andthatforhighenoughcostofconßictitisoptimaltoallowprotest,whileforlowenoughconßictcostsGwillsetrepressionhighenoughtoinduceconßict7.Because$isthelevelofrepressionwhich7Thisismostclearlyseenusinglemma1andseeingthatthisisalwaystrueforCr+Cg>(Ts!Tw),as(Ts!Tw)isactuallytheupperboundbecauseA(1.17equatesthepayoffsofprotestandimmediateconßictfortherebels,itcanbeusedtocalculatetheoptimallevelofrepressionfor"0<"c.WithCr+Cg$(Ts!Tw)'A,theoptimallevelisanyvalueof!($,orbyassumption!=0.ForCr+Cg<(Ts!Tw)'A,Gwishestoinduceconßictandneedstoset!>$,implying!=$+%,forsome%closetozero.Thesecondcaseisforvaluesof"0#("c,"c+$).Hereifrepressionislowenough,therebelswilluseprotest,otherwisetheywillchoosepeace.ExaminingEquation(1.17),weseethatinthisrange,thegovernmentalwayspreferspeace.Thisimpliessetting!$"c+$!"0,or!="c+$!"0byassumption.Because!mustbepositive,beyond("c+$)Gsets!=0.Thisrangeof"0isexactlywhenGwouldlikedissenttooccur,sohecouldlowertherebelsÕshareofresources.However,therebelschoose!regardlessofthelevelof!.CombiningtheseresultsleadsustothefollowingpropositiondeterminingthegovernmentÕsoptimallevelofrepression.Proposition2.StrategicRepression:TheminimallevelofrepressionrequiredtoinducetherebelstotakethegovernmentÕspreferredstrategyisdeterminedasfollows:1.For"0<"c(a)IfCr+Cg$(Ts!Tw)'A,!=0(b)IfCr+Cg<(Ts!Tw)'A,!=$+%2.For"0$"c,!=max{"c+$!"0,0}Thissectionassumedthatitwascostlessforthegovernmenttorepress;whatifitwasnot?Repressiongeneratesadiscretechangetotheexpectedpayoffforthegovernmentbypotentiallyinducingtherebelstochangetheirstrategy.Forlowvaluesof"0thiscanchangeprotesttoconßict,andforhighervaluesitcanpreventprotest.Includingacostwouldnotchangethelevelsofrepression,unlessthelevelneededtoeitherinduceconßict,orpreventprotestcostmorethanthechangeinexpectedpayoffforthegovernment.Forexample,itcouldinducethegovernmenttoswitchtosetting!=0inCase1or2ofProposition2.18Evenwithrepressionendogenous,protestanditsassociateddynamicsarestillpotentialout-comes.Theinterestingconsiderationfromthisdiscussionwouldbeunderstandingtheimpactofpoliciesintendedtoraisethecostofrepressionforagovernment.Althoughthiscouldpotentiallyreducethelikelihoodofimmediateconßictoccurring,suchpoliciesraisethelevelofnon-violentdissent,whichcanescalate.Noonewantstoencouragerepression,soperhapstherealsolutiontopreventingprotestescalationislookingagainatthemodelÕsargumentforwhyitescalates.Es-calationofprotestistheresultofprotestbeinganoisysignal;howeverreducingthenoiseinthesignalcouldleadtoimmediateconßict.Whatthisimpliesisthatinordertoreducethechanceofconßict,policieswhichboth,reducethelevelofrepressionandtheamountofnoiseinprotestmustbeenactedtogether.Forexamplepolicieswhichincreasecivillibertiesandfreedomofthepressmayallowinformationtotravelmorefreely,reducingthenoiseofprotest,whilesimultaneouslyreducingtheexpectedlevelofrepression.Inshort,wemustconsiderthecostsofencouragingpoliticaldissent,andensurethatdoingsodoesnotsimplyleadtowardsviolence.191.4ModelAnalysisandImplicationsFocusingonthebaselinemodel,whichtakesrepressionasexogenous,wecanderivepotentiallytestableimplicationswhichhighlightthemaincontributionofthemodel:distinguishingnotonlywhendissentoccurs,butalsowhatformittakes.Foreachparameterinthemodel,wediscusshoweachpotentialstrategyÕspayoffisimpactedandthenhowitmoveseachoftheindifferencecurves.Thenthemodelisexaminedstartinginastateofprotest,andthedeterminantsofthepotentialpathsfollowingprotestareanalyzed.Finally,thedeterminantsoftheoptimallevelofrepression,fromthemodelÕsextension,areexplored.1.4.1TheuseandtypeofdissentConsistentwithmanymodelsofpoliticaldissent,asthestateoftheworldimproves,overalldissentdecreases.Inthemodel,thiscorrespondswithincreasesin"0.Unlikesomeothers,forexampleDalBŠandDalBŠ(2011),"0doesnotinßuencethepayoffofpoliticaldissentdirectly.However,as"0increases,thepayoffofeitherformofdissent,conßictorprotest,willfallrelativetopeace.HowtherebelsÕequilibriumstrategieschangewith"0canbeseeninFigure(1.2)bymovingalongthehorizontalaxis.Similarily,theimpactofvarying!,theexpectedlevelofrepression,canbeseeninFigure(1.2)bymovingalongtheverticalaxis.As!increases,protestbecomeslesslikely;whetherthatleadstoconßictorpeacedependson"0.Forotherparametersinthemodel,itiseasiesttoexaminehowtheyimpact$and"candthenmovetherespectivecurvesinFigure(1.2).Startingwiththeimpactofthecostofconßict,thegovernmentÕsportionCgonlyimpactstherebelsÕpayoffforprotestanddoessothroughitsimpactonP'.AsCgincreasesthelikelihoodofprotestincreases,becauseGlowersthethresholdprotestsize,P',atwhichtheywillmakealargeconcession.ThisimpactcanbeseeninFigure(1.3)byraising$andshiftingtheprotest-peacein-differencecurveoutward.TheÞgurehighlightsaninterestingimplication:increasingCgraisestheoveralloccurenceofdissent.AlthoughtheuseofconßictintheÞrstperioddecreases,theoverall20chanceofconßictoccurringmaygoupordown.Thisisduetothechanceofconßictfollowingprotest,andtheincreaseduseofprotest.Figure(1.3)showsthesechanges;thehighlightedareatotherightof"cwaspreviouslypeace,whereastheareatotheleftwaspreviouslyconßict.TherebelsÕcostofconßict,Cr,impactsthepayoffsforbothconßictandprotest.Forconßict,thepredictionissimple,asitdirectlyreducesthepayoff.However,forprotesttheimpacthastwoparts.TheÞrstpartisadirectreductionequaltoCr.ThesecondpartoffsetsthisreductionbecausethegovernmentlowersP'.Inmostcases,thedirectnegativeeffectdominates.TheendresultisamuchmoreintuitiveonethanwithCg:asCrincreases,thereisareductionintheoccurenceofdissentandaswitchfromimmediateconßicttoprotest.AllofthiscanbeseeninFigure(1.4).Asthepriorprobabilityofbeingstrong,#0,increases,therebelsÕpayoffforanytypeofdissentalsoincreases.Theintuitionforthisisstraightforward,whetherusingconßictorprotest,themorelikelytherebelsareastrongtype,themorelikelyitistheyendupwiththestrongtypeÕsreservationpayoff.Thisresultsinaclearriseintheoveralloccurenceofdissent,ascanbeseeninFigure(1.5).However,theimpactof#0on$isambiguous,implyingwecannotmakeanypredictionsonhowthecompositionofdissentchangeswith#08.ThelastpiecetoconsideristheimpactoftherebelsÕstrengthparameters,TsandTw.Weconsidertwopossiblechangeshere:holdingTs!TwÞxedwhileraisingTs+Tw,orholdingTs+TwÞxedwhileraisingTs!Tw.TheÞrstcasecouldbethoughtofasincreasingthestrengthoftherebelsrelativetothegovernment.Thisraisesthemeanvalueof"c,whichclearlyleadstomoredissentoverall.However,becausethegovernment!sdecisiononlytakesintoaccountwhichgroupheislikelytoface,nottheleveloftheaveragepayoff,themixofdissentwillremainunchanged.Inotherwords,thevalueforP'isnotimpactedifwechangeonlythemeanvalueoftherebels!strength.Figure(1.5)istheexactpictureofthisscenario.Forthesecondcase,increasingthespreadofTs!Twratherthanthelevel,theimpactondissentdependson#0,becausethedirectioninwhich"cmovesisalsodependenton#0.ItismoreinterestingtoinvestigatethecasewhereweÞx"c,whileraisingthevarianceinconßictsoutcome,8Figure(1.5)Þxes$toshowonlytheincreaseinoveralldissent.21thespreadofTs!Tw.AchangelikethisraisesP',switchingsomedissentfromnon-violenttoviolent.Thisoccursbecausethegovernmentbecomesmorewillingtoriskconßict,orlesswillingtogiveamajorconcession,asthedifferencebetweenthereservationpayoffsincreases.1.4.2Protest:EscalationorDe-EscalationItisofvaluetoexaminewheneachofthethreepotentialpathsfollowingaprotestarepredictedtooccur:peacewithoutconcession,escalation,orpeacewithconcession.Hereconcessionreferstothegovernmentoffering"&&withoutconßictoccuring.Thekeydeterminantsofwhichpathoccursarethesizeoftheprotestandthethresholdprotestvalue,P',whichistheprotestsizeabovewhichGmakesaconcession.Thisleadstooneofthemainrelationshipswetest.AboveP',Gshouldmakeanoffereithertypewillacceptandconßictshouldnotoccur.HoweverasprotestsizeincreasestowardsP'theprobabilityofconßictincreases.Thisisbecausethelargertheprotestthehigherthelikelihooditcamefromastrongrebeltype,butitisnotuntilGseesaprotestlargerthanP'thattheyarewillingtomakealargeconcession.ThisimpliestheprobabilityofescalationÞrstincreases,thendecreasesinprotestsize.TobepreciseFigure1.6,trackstheprobabilityofconßictoccuringasafunctionofprotestsize,theincreasingportionofthecurveisexactlyequalto#1,afterP'theprobabilityofescalationdropstozeroasthegovernmentwillthenmakealargeconcession.UnderstandingthedeterminantsofP'isinstrumentalinunderstandingthepathfollowingprotest.P'decreasesinthecostofconßictand#0,increasesinTs!Tw,andisunimpactedby"09.Thisimplies,conditionalonprotesthavingoccurred,higher#0,Cr,orCgallincreasethelikelihoodofamajorconcessionbyloweringP'.Additionally,asTs!Twincreases,theprobabil-ityofaconcessiondecreases.TheintuitionforthisresultisthatGstandstogainmorebyriskingconßictasTs!Twincreases.9ThesecanbeseenclearlybyexaminingEquation(1.12).221.4.3ModelAssumptionsInthismodel,protestoccursbecausetherebelsgainaninformationaladvantagewhenprotesting,andweakrebelscandonoworsethantheirreservationpayoffbyprotesting.However,thisgainisinpartoffsetbyprotestÕscost:repression.First,itisworthnotingthataslongastherebelsexpecttogainmoreinformation,eveniftheinformationisnotperfectasthemodelset-upimplies,therewillstillbeapremiumfromprotesting.Second,thekeyconditionfortheprotesttobeinformativerequiresthatthesignalfollowstheMLRP:largerprotestsmustbemorelikelywhenrebelgroupsarestrong,andthegovernmentandrebelsmustreceivedifferentsignals.ThisraisesafewquestionsaboutwhenthemodelÕsassumptionswouldholdandwhytherebelsmaygainmoreinformationthanthegovernmentwhenprotesting.SpeciÞcally,thisraisesconcernsaboutwhatwouldhappenif:1)notallrebelswouldbewillingtoengageinviolence;2)whetherornotprotestsizeiscorrelatedwithrebelstrength,and3)whytherebelsmaygainmoreinformationintheprotestprocessthanthegovernment.TheÞrsttwoconcernsarecloselyrelated.Onewaytoexplainwhyprotestsizeiscorrelatedwithrebelstrengthwouldbetocreateamodelwhereindividualparticipationinprotestwasex-plainedexplicitly.Ifindividualpreferencesforgovernmentpoliciesdeterminedparticipationinprotestandafractionofindividualswouldbewillingtouseviolence,thenanytimethegovern-mentsawaprotesttheywouldknowsomefractionofthoseprotestingwouldberebels.Arelatedissueiswhetherornotprotestmovementsareactuallylinkedtogroupswillingtocommitviolence.However,thesamebasicargumentholds,aslongassomefractionoftheotheranti-governmentmovementsupportersmayjoinwiththerebellionÕscause,thesignalwouldstillbecorrelatedwiththestrengthoftherebelmovement.Furthermore,ifthegovernmenthaslessinformationthantherebelsdoabouttherelationshipbetweentherebelsandtheprotestors,thatmayprovideoneexplanationforthedifferenceinthesignalsthegovernmentandrebelsreceive.Thereareanumberofpotentialexplanationsforwhytherebelsmayreceiveamoreaccuratesignalthanthegovernment.Forone,iftheywereinpartresponsiblefororganizingtheprotest,23theymayhaveamuchmoreaccuratesenseofthedegreeofconvictionamongthosewhoshowedup,andthosewhodidnotshowup,thanthegovernment.Inthecasewheretheprotestswerenotactuallyorganizedbypotentialrebels,theymaystillgainmoreinformationthanthegovernmentbyparticipatingandrecruitingothergroupswhoopposethegovernment.Additionally,theprotestmayprovideanopportunityforlearningbydoing.TheactoforganizingprotestorsmaytestsomeoftherebelsÕcapabilitiesthatwouldalsobeusefulwhencommittingmoreviolentactsofdeÞance,suchastheabilitytoorganizeandcommunicate.However,acaveatofthisworkisthatthemodelisamuchbetterÞtinsituationswherethereisadirectlinkbetweenprotestandrebelmovements.Asthedisconnectgrowsstronger,themodelÕsassumptionsbecomelessrealisticandthemodelÕspredictionsarelikelytobelessaccurate.241.5DataSourcesToinvestigatetheimplicationsofthismodel,weturntodatafromtheSocialConßictinAfricaDatabase(SCAD)andPeaceResearchInstituteatOsloÕs(PRIO)ArmedConßictDatabase(ACD).CombiningthispoliticaldissentdatawithFreedomHouseratings,WorldBankWorldDevelop-mentIndicators(WDI),anddatafromtheCingranelli-Richards(CIRI)humanrightsdataprojectleadsustoasampleof48countriescoveringthetimeperiod1990-2012.10Combiningthenon-violenteventsdatafromSCADwithmoretraditionalconßictdataintheACDprovidesauniquewaytoexaminethefullcontinuumofpoliticaldissent,andallowsseveralofthemodelÕskeypredictionstobeexamined.Tomeasureconßict,PRIOÕsACDdatabasewaschosenfortwomainreasons.First,theSCADdataweremeanttobenon-overlappingwiththePRIOdata,sothisshouldpreventthedoublecountingofeventsbetweendatasets.Second,thePRIOdatacoverstheentiresampleintheSCADdataandhasavailablealleventsatahighlydissagregatedlevel.Forprotest,therearefewersourcesavailable;SCADwaschosenduetoitsexhaustivecoverageandtheeaseofidentifyingnon-violenteventswhichinvolvedthegovernmentasatarget.ThispaperusestheminorconßictthresholdintheACDtoconsideracountryinconßict;however,yearscodedasinactiveintheACDarenotcountedasinconßict.11Allperiodsnotconsideredasinconßictareincludedinthefollowinganalysis,regardlessofthelevelofprotestthatoccured.Nodistinctionismadebetweenperiodsofpeaceandprotest,weonlymakeadistinctionforcodingperiodsasinconßict.Thispaperusesmultiplecontinuousmeasuresofprotestintensityandmeasuresprotestonacontinuumratherthananindicatorforoccurrence.Table(1.1)belowdisplayssummarystatisticsforthepoliticaldissentdata.Thedatacontains51onsetsofconßictwithinthe1101countryyearsofdata,with10conßictsthatwereongoingin1990.Overall,22.1percentofallperiodswerecodedasbeinginconßict.Forprotestseveraldif-10From1990-1992thereare47countries,Eritreabecameindependentin1993.11ThedeÞnitionofconßictusedintheACDis:Òacontestedincompatibilitythatconcernsgovernmentand/orterritorywheretheuseofarmedforcebetweentwoparties,ofwhichatleastoneisthegovernmentofastate,resultsinatleast25battle-relateddeaths.ÓStrandetal.(2003)(pg.1)25ferentmeasureswerederivedfromtheeventdata.TheÞrst,Protest1,isthenumberofeventsandonlycapturestheoccurenceofprotest;regardlessofthesizeordurationofanevent,itaddsonlyonetothecount.Thesecond,Protest2,measurecountsthenumberofprotest-daysinacountry-year.Thethird,Protest3,countsthenumberofdistinctlocation-daysinacountry-year.SCADcodesdistinctlyseperatelocations,forthesameevent,asadditionaldatapoints.Forexampleaprotestmovementwithprotestsintwomajorcitieswouldlikelyendupaddingtwotothecountinthismeasure.Thisisthepreferredmeasure,asitcapturesboththeoccurenceandextentofprotest.Longerprotestmovementscoveringmoreareaarecapturedbetterwiththismeasure.Foranalysiswefocusonresultsusingthenaturallogoftheprotestmeasures;thisisdonetosmoothoutsomeoftheover-dispersionpresentinthedata.Anidealmeasurewouldcontainastrongindicationofthenumberofparticipants,butareliableestimateofthisisunavailable.TomeasureConcessions,changesintheFreedomHouse(FH)ratingsforacountrywereused.Anidealset-upwouldallowustotraceexactlywhodemandedwhatandwhatconcessionsweremade;however,nosuchdataexist.TheFHratingsdocapturemanyfundamentalrights,forwhichgroupsmaybewillingtoÞght.Eachyeartheypublishtwoindicesforeverycountry,oneforPo-liticalRightsandanotherforCivilLiberties.TheirindexiswellsuitedtomeasuringconcessionswithinagivencountrybecausetheyusetheprioryearÕsscoreasabenchmark.Thiscouldcom-plicatecross-countrycomparisons,butweareinterestedinwithin-countrychanges.FromFHÕsdescriptionoftheratingprocess,ÒAscoreistypicallychangedonlyiftherehasbeenareal-worlddevelopmentduringtheyearthatwarrantsadeclineorimprovementÓHouse(2014).Fromtheperspectiveofcapturngmajorconcessions,thisisideal.Table(1.2),whichfollows,summarizesthelevelsandfrequencyofchangestotheFHratings.Eachofthetwoindicesisratedona1to7scale,with1beingthemostfree.Aconcessionismeasuredasanincreasetowardsgreaterfreedom,oneitherindex,betweenthecurrentyearandthenext.1212TheFHdataincludes2013,soweareabletocreatethislagfortheentiretimeperiod.261.6ResultsTheempiricsinthispaperfocusonprotest;speciÞcally,whetherornotitoccursandsubsequently,ifitescalates.Examiningthisleadstotwomainconclusions.First,althoughwedonotrejectothertheoriesofprotest,suchasthoseintheglobalgamesliteraturehighlightingitsabilitytoovercomeissuesofcollectiveaction,weÞndevidencesuggestingitisapotentialsignalingmechanismusedtoavoidmoreviolenttypesofdissent.Thisimpliesthatpolicieswhichallowprotesttoaccuratelyconveypublicsupportmaybeeffectiveinreducingtheoccurrenceofviolentdissent.Second,thedataonprotesthavenotreachedthesamestateofmaturityasthatonconßict.SpeciÞcallytheabilitytomatchprotestmovementstospeciÞcgroupsanddemandsisnotavailable.Ifdatalikethisbecomesavailable,additionalimplicationsofthemodelcouldbeexamined.However,thepatternsfoundareconsistentwiththepredictionsofthemodel:largerprotestsgenerateconcessionsmoreoften,andthelargestprotestsaretheleastlikelytoescalatetoconßict.1.6.1TheImpactofProtestStartingfromaperiodofprotest,themodelpredictsthatoneofthreethingscouldoccur:esca-lationtoconßict,aconcession,orpeacewithoutconcession.Equation(1.9),whichdeÞnesthegovernmentÕsofferdecision,determineswhenthegovernmentwouldmakeaconcession,deÞnedasoffering"&&inthemodel.Theprobabilityofescalationinthemodeliscloselyrelatedtothesameequation,howeveritisalsodependentontherebelsÕtruetype.Westartbyanalyzingcon-cessionandthenmovetowardspredictingescalationtoconßict.ThemodelÕsset-upcanbetranslatedintoalatentutilityframeworktoanalyzeconcession.Thegovernmentiscomparingtheexpectedutilityfromtwopotentialchoices:offering"&andriskingconßictwithstrongtypes,oroffering"&&andÒpayingforpeace.ÓEquation(1.9)comparesthisde-cision.Embeddingthisconditionintoalatentutilitymodelrequirescomparingtheexpectedutilityfromeachpossiblechoice,andtheadditionofanerrortermmeanttocaptureunobservedfactorsinßuencingthisdecision.Doingthis,andmovingallitemstotheLHSleadstothefollowing,27whichdescribestheprobabilityaconcessionoccurs.P(Concession)=(#01!#0'f(P|Rs)f(P|Rw)!Ts!TwCr+Cg+%$0).(1.18)Equation(1.18)demonstrateswhentheutilityfrommakingaconcession,offering"&&,islargeenoughrelativetoriskingconßict,offering"&.Usingcomparativestaticsderivedinsection4,weknowthatthisprobabilitywillincreasewithincreasesin#0,protestsize,Cr,orCg,andwilldecreasewithincreasesinTs!Tw.Thetheorydoesnotpredictarelationshipbetweenconcessionand"0,conditionalonprotestoccurring.ThisleadstoourÞrstformhypotheses:Hypothesis1.Theprobabilityofconcessionincreasesinprotestsize.ToexamineHypothesis1,weuseÞxedeffectsregressionsofthefollowingform:Concessit=&1Protestit+&2Xit+$i+'t+%it(1.19)Whereiindicatescountry,tindicatestheyear,$and'arecountryandtimeÞxedeffects,Xitisavectorofcontrolvariables,and%itistheerrorterm.ForthedependentvariablewegenerallylookatthenetchangeoftheFHratingsfromoneyeartothenext,andweuseloggedvaluesofallthreemeasuresofprotestsizediscussedinthedatasection.ExaminingthesespeciÞcationsinTable(1.3)weseeaconsistentpositiverelationshipbetweenprotestsizeandtheprobabilityofconcession.Toquantifythisimpact,itiseasiesttoexaminetheÞnalcolumn,whereweuseanindicatorthatisequaltooneifeitherscaleimprovedfromthecurrentyeartothenext.WeseeacoefÞcientof.08,whichmeansthateachdoublingofprotestsizecorrespondstoanincreaseofabouteightpercentagepointsinthelikelihoodofaconcession.Toputthisintermsrelativetothedata,aprotestonestandarddeviationabovethenormraisesthelikelihoodofobservinganincreaseintheFreedomHouseratingsbynearlyÞfty-percent.Wealsoaddressanumberofpossibleconcernsabouttherobustnessofthisimpact.First,mostspeciÞcationsincludebothcountryandyearÞxed-effects,thisremovesheterogeneityattherespectivelevels.Alternatively,weexaminetherelationshipwithouteachoftheseandseeasimilar28pattern.Wealsoexaminealternativefunctionalforms,includingthesquaretermofprotestsizeanduseaÞxed-effectsprobitestimationinsteadofalinearprobabilitymodelandÞndsimilarresults.Finally,weinvestigatethepossibilitythatperiodsofprotestarethemselvestumultuous,andresultinatemporaryreductionintheFHratingsforacountry,sowhatweobservewouldactuallybearecoverytonormallevels.ToaddressthiswechangethedependentvariabletobealaggedversionofthenetchangeintheFHratings,andÞndaninsigniÞcantnegativeimpactwithamuchsmallermagnitude.Thesecondrelationshipweanalyzeiswhichprotestsescalate.Therelationshipbetweenprotestsizeandtheprobabilityofescalationismorenuanced.Escalationoccursonlywiththestrongrebelgroupandonlywhentheyfailtoproducealargeenoughsignalforthegovernmenttooffer"&&.BelowP',thisprobabilitymatchesexactly#1,andshouldbeincreasinginprotestsize.HoweveroncetheprotestsizepassesP',thegovernmentoffers"&&andconßictshouldbeavoided.Takenliterally,thisimpliesaprobabilitywhichincreasestoapointandthensharplydropsofftozero.Thisleadstooursecondhypotheses:Hypothesis2.Therelationshipbetweenprotestsizeandconßictescalationisnon-monotonic,Þrstincreasingthendecreasing.WeexamineHypothesis2usingnon-parametricspeciÞcationsandlookforapatternmatchingFigure(1.6);wealsolookforaninverted-ushapedrelationshipwithregressionspeciÞcations.Thissuddendropintheprobabilityofescalatingtoconßictpredictedbythetheoryisnotfarfromwhatweseeinthedata.Table(1.4)showstheresultsfromaÞxedeffectsregression,whenexaminingprotestdaysandlocation-days,weseesigniÞcantresultsontheprotestsizeanditssquare,aftercontrollingforcountryandyearÞxedeffects.Toputtheimpactinperspective,wecalculateandgraphthemarginsfortherangeofvaluesprotesttakesoninthedata.InbothspeciÞcations,weseethelargestprotestsinthedatapredictanegativelikelihoodofescalation.Overall,theobservedpatternisexactlywhatispredictedbytheory,aninitiallyincreasingthendecreasingprobabilityofescalation.Giventhenon-monotonicnatureofourprediction,wehavenoapriorireasontobelievea29quadratictermisthecorrectwaytocapturethetruerelationshipbetweenprotestsizeandthelikelihoodofescalatingtoconßict.Onewaytoaddressthisistouselocallylinearpolynomialestimation,whichessentiallyestimatesandsmoothsaseriesoflinearregressions.TodothisweemploySTATAÕslpolycommandandtracetherelationshipbetweenprotestsizeandtheprobabilityofconßictoccurringinthenextperiod.Wedothisintwoways,Þrstasimpleunivariaterelation-ship,regressingprotestsizeontheindicatorforescalatingtoconßict.Thenbecausewewouldliketoeliminatetimeandcountryinvariantunobservables,weÞrstpartialouttheÞxedeffects,andthenconductthesameregressionusingtheresidualsleftafterpartiallingouttheÞxedeffects.Weagainfocusonprotest2andprotest3,becauseprotest1lackssufÞcientvariationtoidentifytherelationship.ExaminingFigures(1.9)Ð(1.12)weseethatbothmeasuresofprotest,withandwithouttheinclusionofÞxedeffects,cometothesameconclusionastheregressionspeciÞcations.Whenprotestsarerelativelysmalltherelationshipbetweensizeandescalationtoconßictisstableorslightlyincreasing;however,thelargestprotestsinthedatadonotescalate.Theresultsheremaynotbecausal;however,therelationshipweobserveandpredictisaratherspeciÞconeandfewifanyalternativeexplanationscometomindquicklyfortheresultinginverted-ushapedrelationshipbetweenprotestsizeandthelikelihoodofescalationtoconßict.Furthermore,theuseoftimeandcountryÞxedeffectsgoalongwaytoeliminatesourcesofheterogeneity.Inshort,thecorrelationsweÞndmaynotbecausal,buttheyarestrikinglysimilartothemodelÕspredictions.Furthermore,commonintuitionwouldoftensuggestthatthelargestprotestsaretheonesthatescalateintofullscaleconßict,anintuitionrejectedhereintheoryandwithdata.Perhapsmostimportantly,wealsoÞndthepredictednon-monotonicrelationshipbetweenprotestsizeandescalation,arelationshipspeciÞctothismodel.301.7ConclusionExaminingtheescalationtoconßictinarationalsettingmayhelpustobetterunderstandwhycon-ßictoccursinlieuofnegotiatedsettlements.ThemodelpresentedhereisclearlyasimpliÞcation,andfocusesononepossibleexplanationfortherolesofprotestandrepression.Nonetheless,theinsightprovidedbythemodel,allowsustoexamineifasymmetricinformationisadrivingforcecausingconßict.Theresultsinthispaper,boththeoreticalandempiricalpointtowardssucharelationship.Ifasymmetricinformationisadrivingforceforconßict,thisshouldinßuencethetypesofpoli-ciestoencourageincountriesmovingalongthepoliticalspectrumtowardsdemocracy.Transitionssuchasthosewhichoccurredthroughoutthesampleperiodcoveredbythedatausedinthispaperareknowntobeconßict-prone,(HegreandSambanis,2006;Hegreetal.,2001).Ifthisisthecase,helpingtocreatetransparentmechanismsforexpressingpoliticalviewpointsmaybeakeyfactorinavoidingconßict.Reducingthecostofusingnon-violentmechanismsofdissentdoesleadtomoreuseasshowninCunningham(2013).However,simplyencouragingprotestcouldleadtoconßictiftheinformationconveyedthroughprotestisnotaccurate.Furtherresearchwillbeneededifwearetotrulyunderstandtherelationshipbetweenviolentandnon-violentpoliticaldissent.Theanalysisheresuffersfromseveralshortfalls,mostpromi-nentlyaninabilitytobreakoutwhotheparticipantsinagiveneventarebeyondclassifyingthemasciviliansandgovernmentswithinacountry.AsbetterdatabecomesavailablewemaybeabletoreÞnethehypotheseswewishtotest,andbegintocomparethemtoalternatives.Forexample,itwouldbeinterestingtocompareamodelofasymmetricinformationtoamodelexaminingcom-mitmentproblems.Wouldthisresultinsimilarpredictionsastotheuseofprotestandrepression?Ifnot,canwecomparethetwopossibilities?Insuchacasesimplyadvocatingtransparentmech-anismsforexpressingpoliticalviewpointswouldnotlikelybeenough;andcreatinginternationalmechanismstoenforceagreementsmaybeapreferablepathtoavoidfutureconßict.Nonetheless,whilecommonintuitionsuggeststhelargestprotestsarethemostlikelytoescalatetoconßict,this31intuitionisrejectedbyboththetheoryandempiricalresultsofthispaper.Theliteratureoncivilwarhasastrongtraditionofuncoveringrelationships,thoughrarelyaretheactualcausalmechanismsintheserelationshipspinneddown.Thispaperprovidesastepinthatdirection,bydevelopingandtestingthepredictionsofamodelofasymmetricinformation.Additionally,highlightingtheideathatbothviolentandnon-violentpoliticaldissenthavethesameunderlyingcauses,thisresearchsuggeststhatalotcanbelearnedbyexaminingwhatdifferencesprovokeoneovertheother.32APPENDICES33APPENDIXAFIGURESFORCHAPTER134Figure1.1FormalGameTreeNatureRs(Ts!Cr,1!Ts!Cg)CNatureP("0,1!"0)!#0Rw(Tw!Cr,1!Tw!Cg)CNatureP("0,1!"0)!1!#0GGRs("1!!,1!"1)Accept(Ts!Cr!!,1!Ts!Cg)CRw("1!!,1!"1)Accept(Tw!Cr!!,1!Tw!Cg)CGsets"1Gsets"1Gupdates#0to#1ProtestObservedProtestObserved35Figure1.2ImpactofCgonEquilibriumStrategies"0!$"cProtest=PeaceConßict=PeaceProtest=Conßict"c+$Protestfollowedby?ImmediateConßictImmediatePeaceFigure1.3EquilibriumPathsfor!and"0"0!$$&"cProtest=PeaceConßict=PeaceProtest=ConßictPeacetoProtestConßicttoProtest36Figure1.4ImpactofCronEquilibriumStrategies"0!$$&"c"&cProtest=PeaceConßict=PeaceProtest=ConßictConßicttoPeaceConßicttoProtestProtesttoPeaceFigure1.5Impactof#0onEquilibriumStrategies"0!$"c"&cProtest=PeaceConßict=PeaceProtest=ConßictPeacetoProtestPeacetoConßict37Figure1.6ProtestSizeandEscalationProbabilityofEscalationProtestSizeP'#1Goffers"&Goffers"&&Figure1.7EscalationversusProtestsize-.1-.050.05.1Escalation Probability01.422.93.564.455.56Log Protest 2Fixed Effects Regression: Margins with 95% CIs38Figure1.8EscalationversusProtestsize-.15-.1-.050.05.1Escalation Probability01.422.93.564.4567Log Protest 3Fixed Effects Regression: Margins with 95% CIsFigure1.9EscalationversusProtestsize0.02.04.06.08.1Escalation Probability0246Log Protest 295% CIlpoly smoothkernel = epanechnikov, degree = 0, bandwidth = .88, pwidth = 1.32Local polynomial smooth39Figure1.10EscalationversusProtestsize-.050.05.1Escalation Probability02468Log Protest 395% CIlpoly smoothkernel = epanechnikov, degree = 0, bandwidth = .86, pwidth = 1.28Local polynomial smoothFigure1.11EscalationversusProtestsize-.1-.050.05.1Escalation Probability0246Log Protest 295% CIlpoly smoothkernel = epanechnikov, degree = 0, bandwidth = .85, pwidth = 1.27Local polynomial smooth: Partialled out Fixed Effects40Figure1.12EscalationversusProtestsize-.1-.050.05Escalation Probability02468Log Protest 395% CIlpoly smoothkernel = epanechnikov, degree = 0, bandwidth = 1, pwidth = 1.5Local polynomial smooth: Partialled out Fixed EffectsFigure1.13ProtestSizeandConßictEscalation-.050.05.1.15Probability of Escalation to Conflict-505Z score of Protest Size95% CIlpoly smoothkernel = epanechnikov, degree = 0, bandwidth = .82, pwidth = 1.23Local polynomial smooth41APPENDIXBTABLESFORCHAPTER142Table1.1PoliticalDissentDataVariableObsMeanStd.Dev.MinMaxProtest111014.7411.500282Protest2110122.7246.590366Protest3110130.7077.260884P3Percapita11010.371.17017.33logofP311011.771.8006.78ProtestIndicator11010.530.5001ConßictIndicator11010.220.4101MeansExcludingConßictPeriodsVariableObsMeanStd.Dev.MinMaxProtest18584.6212.100282Protest285822.5847.510365Protest385831.7783.190884ProtestIndicator8580.680.4701Table1.2FreedomHouseRatingsVariableObsMeanStd.Dev.MinMaxFHPoliticalRights11014.671.8117FHCivilLiberties11014.401.3717ConcessionPoliticalRights11010.100.3001ConcessionCivilLiberties11010.110.3101ConcessionFH11010.170.3801NetChangeBothRatings1101-.081.02-69FHPoliticalRights,andFHCivilLibertiesarethemeanvaluesoftheactualPoliticalrightsandcivillibertiesratings,1representsthegreat-estdegreeoffreedom.Concession,isanindicatorforconcession,animprovementintherespectiverating,withConcessionFHimplyinganimprovementineitherrating.43Table1.3ProtestSizeandConcessionsConcessionMeasureNetChangeinFHRatingLaggedIncreaseNetChangeIndicatorLogProtest10.13***0.16**0.09*0.128*0.0976**0.0728-0.04120.0813***(0.05)(0.07)(0.05)(0.07)(0.04)(0.10)(0.07)(0.03)LogPopulation-1.09-0.97-0.97-2.74**0.01-0.0296-2.772**-0.825-1.0891.027-0.605*(0.76)(0.75)(0.75)(1.11)(0.03)(0.03)(1.08)(0.79)(0.76)(0.65)(0.31)LogProtest20.07**(0.03)LogProtest30.07**(0.03)Protest3PerCapita0.07*(0.04)LogProtest1-0.19***CountryDemeaned(0.07)TimeTrend0.04(0.03)EconomicShock-0.04(0.23)LogProtest1-0.02Squareterm(0.04)YearFixedEffectsYesYesYesYesYesYesYesNoYesYesYesYesCountryFixedEffectsYesYesYesYesYesNoNoYesYesYesYesYesInclusion111122221111Observations818818818818550550550550768818820583R-squared0.0440.0430.0440.0390.0720.0580.0510.0360.0340.0440.040.06NumberofCountries47474747444446474645Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1444444Table1.4ProtestSizeandEscalationtoConßictOutcome:ConßictOnsetConßictOnsetConßictOnsetProtestMeasure1:Events2:EventDays3:EventLocation-DaysLogProtest0.020.043**0.037**(0.03)(0.02)(0.02)LogProtestSquared0.00-0.009**-0.007**(0.01)(0.00)(0.00)LogPopulation-0.24-0.26-0.26(0.24)(0.24)(0.24)Observations818818818R-squared0.0250.0320.031NumberofCountries474747Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1Table1.5TransitionRatesPeacet+1Protestt+1Conßictt+1Peacet0.570.380.05Protestt0.150.780.08Conßictt0.060.160.78AperiodisclassiÞedbythemostextremepolit-icaldissenttohaveoccured,anyprotesteventinSCADqualiÞesasprotest,andPRIOÕsminorcon-ßictthresholdisusedforconßict.45REFERENCES46REFERENCESBaliga,S.,&Sjırstrım,T.(2004).Armsracesandnegotiations.TheReviewofEconomicsStudies,71,351Ð369.Besley,T.,&Persson,T.(2009).Repressionorcivilcar?AmericanEconomicReview,99,292Ð297.Blattman,C.,&Miguel,E.(2010).Civilwar.JournalofEconomicLiterature,48,3Ð57.Carlson,J.,&VanDamme,E.(1993).Globalgamesandequilibriumselection.Econometrica,61,989Ð1018.Chassang,S.,&Padro-iMiguel,G.(2009).Economicsshocksandcivilwar.QuarterlyJournalofPoliticalScience,4,211Ð228.Chaudoin,S.,Peskowitz,Z.,&Stanton,C.(2013).Beyondzeroesandones:Theseverityandevolutionofcivilconßict.AvailableatSSRN2135780.Chenoweth,E.,&Cunningham,K.G.(2013).Understandingnonviolentresistanceanintroduction.JournalofPeaceResearch,50,271Ð376.Cunningham,K.G.(2013).UnderstandingstrategicchoiceThedeterminantsofcivilwarandnonviolentcampaigninself-determinationdisputes.JournalofPeaceResearch,50,291Ð304.DalBŠ,E.&DalBŠ,P.(2011).Workers,warriors,andcriminals:Socialconßictingeneralequilibrium.JournaloftheEuropeanEconomicAssociation,9,646Ð677.Davenport,C.(2007).Staterepressionandpoliticalorder.AnnualReviewofPoliticalScience,10,1Ð23.Fearon,J.(1995).Rationalistexplanationsforwar.InternationalOrganization,49,379Ð414.Gleditsch,N.P.,Wallensteen,P.,Eriksson,M.,Sollenberg,M.,&Strand,H.(2002).Armedconßict1946-2001:Anewdataset.JournalofPeaceResearch,39,615Ð637.Goldstein,R.J.(2001).PoliticalrepressioninmodernAmericafrom1870to1976.UniversityofIllisnoisPress.Granovetter,M.(1978).Thresholdmodelsofcollectivebehavior.AmericanJournalofSociology,1420Ð1443.Gurr,T.R.(2000).Peoplesversusstates:Minoritiesatriskinthenewcentury.U.S.InsituteofPeacePress.Hafner-Burton,E.M.(2005).Tradinghumanrights:Howpreferentialtradeagreementsinßuencegovernmentrepression.InternationalOrganization,59,593Ð629.Hegre,H.&Sambanis,N.(2006).Sensitivityanalysisofempiricalresultsoncivilwaronset.JournalofConßictResolution,50,508Ð535.47Hegre,H.,Ellingsen,T.,Gates,S.,&Gleditsch,N.P.(2001).Towardsademocraticcivilpeace?Democracy,politicalchange,andcivilwar,1816-1992.AmericanPoliticalScienceReview,95,33Ð48.FreedomHouse.(2014).Freedomintheworld2014methodology[Codebook].Retrievedfromhttps://freedomhouse.org/sites/default/Þles/Methodology%20FIW%202014.pdfLichbach,M.I.(1987).Deterrenceorescalation?Thepuzzleofaggregatestudiesofrepressionanddissent.JournalofConßictResolution,31,266Ð297.Kricheli,R.,Livne,Y.,&Magaloni,B.(2011).Takingtothestreets:Theoryandevidenceonprotestsunderauthoritarianism.InAPSA2010AnnualMeetingPaper.Miguel,E.,Satyanath,S.,&Sergenti,E.(2004).Economicshocksandcivilconßict:Aninstru-mentalvariablesapproach.JournalofPoliticalEconomy,112,725Ð753.Moore,W.H.(2000).Therepressionofdissentasubstitutionmodelofgovernmentcoercion.JournalofConßictResolution,44,107Ð127.Morris,S.,&Shin,H.S.(2001).Globalgames:Theoryandapplications.AvailableatSSRN284813.Pettersson,T.,&Wallensteen,P.(2015).Armedconßict1946-2001:Anewdataset.JournalofPeaceResearch,52,536Ð550.Pierskalla,J.H.(2010).Protest,deterrence,andescalation:Thestrategiccalculusofgovernmentrepression.JournalofConßictResolution,54,117Ð145.Powell,R.(2002).Bargainingtheoryandinternationalconßict.AnnualReviewofPoliticalSci-ence,5,1Ð30.Skaperdas,S.(1996).Contestsuccessfunctions.EconomicTheory,7,283Ð290.Strand,H.,Wilhelmsen,L.,&Gleditsch,N.P.(2003).Armedconßictdatasetcodebook[Codebook].Retrievedfromhttp://www.pcr.uu.se/digitalAssets/124/124920_1codebook_ucdp_prio-armed-conßict-dataset-v4_2014a.pdfWalter,B.(2009).Bargainingfailuresandcivilwar.AnnualReviewofPoliticalScience,12,243Ð261.Wittman,D.(1979).Howawarendsarationalmodelapproach.JournalofConßictResolution,4,743Ð763.48CHAPTER2ElectoralViolence:AnEmpiricalExaminationofExistingTheories2.1IntroductionElectionsareintendedtoprovideapeacefulmethodforgroupstocompeteforpoliticalpower,sometimesdirectlyreplacingviolentalternatives.However,signiÞcantdebateexistsoverhowelectionscanbeusedinpost-conßict,orconßict-prone,societiesasasolutiontoviolence(Chan-dra,2001:Reilly,2002).Insituationswherepoliticalpartieshavenotyetacceptedtherulesofthegamesetforthbyelections,violenceistheorizedtobeeffectiveatmanipulatingelectionresults(CollierandVicente,2012;Machadoetal.,2011;RobinsonandTorvik,2009).Althoughmanytheoriesofelectionviolenceexist,theyandtheeffectivenessofelectoralviolenceremainlargelyuntested(Hickman,2009).Thispapercriticallyexaminestwoexistingtheoreticalframeworksforhowviolenceisusedtoinßuenceelectionoutcomes.Thispaper,likeothers,suchasGoldsmith(2015)andHıglund(2009),focusesonelectionviolenceasseparatefromothertypesofpoliticalviolenceandstartswiththefollowingbroaddeÞnitionofelectionviolencefromFischer(2002):Anyrandomororganizedactorthreattointimidate,physicallyharm,blackmailorabuseapoliticalstakeholderinseekingtodetermine,delayortootherwiseinßuenceanelectoralprocess.WithsuchabroaddeÞnitionofelectoralviolenceandsomanypossiblereasonsforitsoccur-rence,itisnecessarytoreÞnethefocusofelectoralviolenceforthispaper.Wechoosetofocus49onviolenceasatacticofvotermanipulationandexamineinwhatwaypatternsofelectoralvio-lenceareconsistentwiththestrategicuseofviolencetoinßuencetheoutcomeofanelectionbythepurposefultargetingandattemptedexclusionofparticulargroupsofvoters.Thistypeofelec-toralviolencehasbeenthesubjectofsigniÞcantandvariedacademicwork,includingcasestudies,empiricalworklikeinHickman(2009),andthetwotheoreticalpaperswefocuson:Chaturvedi(2005),hereafterCH,andCollierandVicente(2014),hereafterCV.Thesepapersoutlinetwodifferentframeworksregardinghowviolencemaybeusedtoinßuenceelectionoutcomes.InCH,violenceistargetedatoppositionsupporterswiththeintentionofpreventingthemfromvoting.InCV,amorenuancedpossibilityisexamined;theyargueviolenceisnottargeteddirectlyattheoppositionbutinsteadthattheuseofuntargetedviolencedisenfranchisesunalignedvoters,whohavelessatstakeintheelection.Whenunalignedvotersceasevotingduetoviolence,moreweightisplacedoneachpartyÕscoresupporters.Thus,ifonepartystartswithanadvantageincoresupporters,relativetotheirsupportamongunalignedvoters,theywillincreasetheirvotesharebyusinguntargetedviolence.Thereislimitedempiricalevidenceonhowviolenceimpactsindividualvotingbehavior,withtheexceptionofafewstudieslike(Bratton,2008;CollierandVicente,2014;andHickman,2009),focusingonsingleelections.Forthisreason,wefocusonexamininghowindividualsrespondtoelectionviolence.WedothisusingsurveydatafromAfrobarometerandtheconßicteventdatafromtheSocialConßictinAfricaDatabase(SCAD).Thisallowsustoexamineabroadersampleofcountriesandheterogeneityinresponseto,andfearsof,violence.UnderstandinghowelectoralviolenceinßuencesbehaviorinelectionsisafundamentalpartofunravelingthemechanismswhichdriveelectoralviolenceandwillallowustoreÞneexistingtheories.TheAfrobarometerround4surveyof20countrieswasintendedtomeasurepoliticalattitudesacrossAfrica(AfrobarometerData,2008).Weuseittomeasureperceptionsofviolence,politi-calparticipation,andpoliticalafÞliation.ThisisamongtheÞrstpaperstousesurveymeasuresofviolence,andwearguethisprovidesseveralpotentialadvantagesovertraditionaleventcountmeasures,especiallywhenweareinterestedinunderstandinghowindividualcharacteristicsare50associatedwithdifferentperceptionsofviolence.Wealsocomparemeasuresofviolencefromthesurveytomoretraditionaleventcountdata.Althoughthereisastrongcorrelationbetweenthesemeasures,wedonotbelievetheymeasuretheexactsamephenomena.Wearguethataskingindividualsabouttheirfearofbeingvictimizediswellsuitedforanalyzingwhoisimpacted,ortargeted,byviolence.ThismeasureshowssigniÞcantvariationbetweenrespondents,evenwithinsmallgeographicalareaswhereallindividualsarelikelytoobservethesameoccurrencesofvi-olence.Inotherwords,theoccurrenceofeventsdoesnotnecessarilyresultinallindividualsintheareafearingviolence.Therefore,thesurveyresponsesmaybecapturingindividualdifferenceswhicharemostlikelytheresultofsomeindividualshavingexperiencedtargetedviolence.Leveragingtheseuniquemeasuresofviolence,weareabletoexaminehowwellreactionstoviolencealignwiththeframeworkspresentedinCHandCV.Consistentwiththeframeworkpro-videdinCH,whicharguesthatindividualstargetedwithviolencewillceasevoting,anyindividualwhofearsviolenceislesslikelytovote.Additionally,consistentwiththestoryofuntargetedvio-lenceinCV,weÞndthatswingorunalignedvotersaretheonlygroupmorelikelytoabstainfromvotinginthepresenceofviolence,evenwhentheydonotreportagreaterfearofviolence.Thisleadsustoconclude,consistentwithbothframeworks,thattargetedviolencecouldcauseabsten-tionamongvotersofanypoliticalafÞliation,whereasuntargetedviolenceisonlylikelytoreduceturnoutamongunalignedvoters.Whenexaminingwhofearsviolence,ratherthanhowindividualsreacttoviolence,weÞndanorderingofthosefearsbasedonpoliticalafÞliation.Supportersofthemajorityorrulingpartyfearviolencetheleast,followedbyswingorunalignedvoters,whilesupportersofotherpoliticalpartiesreportthehighestfearofviolenceinelections.AnumberoftheoriesofelectionviolencecouldsupporttheseÞndings;howeveroneexplanationconsistentwiththisÞndingwouldbetheuseofthestatesecurityapparatustorepresscompetition.Wearguethatmoredetaileddataontheactualeventswouldbethebestwaytodeterminewhythisorderingoccursinthedata.Finally,weturntoexaminingwhereviolenceoccursandhowthatrelatestothedivisionoftheelectorate.WeÞndthatdistrictswithahigherfractionofunalignedvotersalsohavemore51occurrencesofviolence.Thisimpliesthatalthoughbothframeworksarepotentiallyviable,theonelaidoutbyCVappearstobetterexplainthepatternsobservedinthedata.Furtherworkisneededtoinvestigatewhatconditionsdeterminewheneachframeworkismorelikelyandtoruleoutotherpotentialexplanations.Thispapermakesconsiderableprogressinexplainingwhoreactstoelectionviolenceandhow.GiventheprevalenceofviolenceinelectionsandthedebateoverhowÒelectoralengineeringÓcanbeusedtohelpmitigateviolenceinelections,havingathoroughunderstandingofhowviolenceinßuenceselectoralbehaviorisimportantifwearetotrytomitigateitsimpact.1Furthermore,suggestiveevidencesupportstheideathateveninsocietieswheretheremaybeelections,theuseofelectionviolencemaybepreventingseriousoppositionfromforming.Thisimpliesinternationalresponsesshouldbestrongwhenobviousinfractionsareobserved.Otherwise,thispotentiallyeffectivestrategyofviolenceislikelytopersist.1ForathoroughcomparisonofthetwoleadingtheoriesofÒelectoralengineering,ÓandaccompanyingexamplesoftheirsuccessandfailureseeReilly(2002).522.2LiteratureReviewElectionsplayafundamentalroleindemocraticgovernance,ÒThey[elections]providelegitimacythroughdirectpopularparticipation,and,inturn,legitimacycreatescapacityforeffectivegover-nanceÓ(Brown,2003).Violenceinthisprocess,then,hasthecapacitytounderminenotonlytheelections,butalsothelegitimacyofthegovernmentitself.Alargebodyofworkhasexaminedtheuseofelectionsinpost-conßictsocieties;however,astheAfrobarometerdata,andpreviousresearchshow,someviolenceisquitecommoninAfricanelections,Goldsmith(2015),Lindberg(2006),andStrausandTaylor(2009).Giventheimportanceofelectionsandthehighdegreeofsufferingviolencecancause,itisnotdifÞculttomotivatetheneedtounderstandelectoralviolence.2.2.1CanElectionsReplaceViolence?Inpost-conßictsocieties,electionsareoftenintendedtotradeÔbulletsforballotsÕ;however,aspointedoutinRapoportandWeinberg(2000),successionisthemostturbulenttimeforanytypeofgovernment.AgreatdealofresearchhasconsideredtheconsequencesofÒelectoralengineer-ingÓ,orhowanelectoralsystemcanmitigateorencouragetheuseofviolencebasedonelectionresults.Although,violenceisnottheonlydimensionaffectedbythechoiceofelectoralsystem,aspointedoutinNorris(2007):ÒElectoralrulesarenotneutral.ÓThecombinedimpactofviolenceandtheelectoralsystemneedstobeevaluatedforhowitinßuencesrepresentation,especiallyforcontentiousgroupsinconßict-pronesocieties.Insuchsituations,theviolencemaynotbelimitedtoelections;infact,akeyconditionforrecurrentcivilconßictdiscussedbyWalter(2004)isthelackofanon-violentmethodtoinßuencegovernance.Therearetwomajortheoriesonhowelectionscanbeusedtomitigateconßict;bothofthesetheoriesarediscussedatlengthinReilly(2002).TheÞrst,consociationalism,arguesallgroupsshouldbegivenafairvoiceingovernance.Thetheorysuggestsaproportionalrepresentationsystemmayalleviateconßictbygivinggroupsadequaterepresentation.Thealternativetheoryarguesthatdoingsoencouragesdivisionalongexistingethno-religiousdividesanddoesnothing53toaddressunderlyingissues.Insteadtheysuggestareciprocal-voteapproach,whichisintendedtofosterinter-groupcooperation2.2.2.2TheImpactofViolenceTheexistingresearchexaminingthedirectlinkbetweenelectionoutcomes,electoralparticipation,andviolenceislimitedtoafewstudiesanalyzingparticularcountriesindetail.TheÞrst,Hickman(2009),examinestheimpactofviolenceonvoterturnoutandelectionresultsinSriLanka.HeÞndsviolenceperpetratedbyindividualsassociatedwithonepoliticalparty,leadstoareductioninturnoutfortheoppositioninthatdistrict.Hickmanarguesthattheimpactissmall,didnotchangetheelectionresults,andthattheuseofviolencebybothsidescancelsoutintheaggregate.AnotherpaperusingtheAfrobarometerdatafromNigeria,showsthatvotebuyingisfarmoreeffectivethanviolence(Bratton2009).BrattonalsodemonstratesthatthemostcommonresponsetoanyformofÒillegitimatecampaigningÓ,e.g.violenceorvote-buying,isabstention.Weextendthisliteraturebyexaminingmultiplecountriesandlookingatheterogeneityintheimpactofviolence.Empirically,Blattman(2009)andCollierandVicente(2014)providethemostconvincingcausalevidenceoftheimpactofviolenceonpoliticalparticipation.TheÞrst,Blattman(2009),usesvariationinexposuretoviolenceinUganda,whichhearguestobeexogenous,toestimatetheimpactofthatexposureonpoliticalparticipation.HeÞndsthathavingbeenexposedtoviolence,nearlytwodecadesearlier,makesanindividualmorelikelytobeinvolvedinthecommunityandmorelikelytovote.CollierandVicente(2014),userandomplacementofanti-violencecampaignsduringthe2007Nigerianelectiontocreateexogenousvariation.Thecampaignsuccessfullyre-ducedelectionviolenceintheregionsofimplementation,andtheyfound,likewedo,thatviolencereducesvoterturnout.Thetimingofviolenceineachstudylikelyexplainstheircontradictoryconclusions.InBlattman(2009),theexposuretoviolenceisalmosttwodecadesearlier,andincreasespoliticalparticipation.InCollierandVicente(2014),theviolenceiscurrent,andtheyÞnditdecreasesvoterturnout.The2ForathoroughexaminationoftheseissuesastheyrelatespeciÞcallytoAfricancountries,seeLindberg(2005).54differencecouldbeexplainedbydifferentiallongandshortruneffects.Inthelongrunobservingviolencemaymakeanindividualwishtobemoreinvolvedinpoliticsbecauseofpastexperiences.However,intheshortrun,itmaymaketheactofvotingtoodangerousorcostly.Forsuchaprevalenttopic,relativelylittleempiricalliteratureexists.Datalimitationsarelikelytheprimaryreasonforthis.Furthermoretheirregularnatureofelectionsinmanycountries,es-peciallywhereelectionviolenceisacommonproblem,yieldsseriousestimationissues.SomeoftheseissuesarediscussedinGoldsmith(2014)andCheibubet.all(2012).Weaddresssomeoftheissuesrelatedtodatalimitations,byexploringtheuseofmoregeneralpoliticalviolencedatatomeasureelectionviolenceanddemonstratingthatsurveydatacanbeusedtomeasureelectionviolence.2.2.3CausesandTypesofElectoralViolenceUnderstandingthecausesofelectoralviolencerequiresasolidunderstandingofwhatwemeanbyelectoralviolence.Broadlyspeaking,thispaperusesthedeÞnitionfromFischer(2002).AnotherpaperKehailia(2014),publishedaspartofaseriesofcasestudiesonelectionviolencefrompractionerexperienceatInternationalFoundationforElectoralSystems(IFES),(Cyllahet.all2014),laysoutausefultypeologyofelectionviolence.ThistypologyclassiÞeselectionviolenceintoeighttypes,basedonwhoisinvolvedintheincident.Theexamplesheprovidesforeachtype,demonstratehowdifferenttypesofelectoralviolencecanhavefundamentallydifferentcauses,andbeusedfordifferentpurposes.Suchatypologyisusefulwhenstudyingelectionviolence,especiallywhenconsideringtheun-derlyingcausesorintentofelectionviolence.However,evenwithinasingletypethecauses,goals,andpurposemaydiffer.Thispaperfocusesonelectionviolencewheretheintentisvotedeterrence,speciÞcallystrategicallymanipulatingwhovoteswiththeintentofinßuencingtheelection.Eveninsuchanarrowrangeasvotedeterrence,thepurposeofdoingsomaynotalwaysbemanipu-latingtheoutcomeonpollingday;somegroupsmaysimplyhavetheintentofunderminingtheelectoralprocessasawhole.Takeoneoftheexamplesfrom(Kehailia2014),whereunidentiÞed55gunmenopenedÞreonapollingstationintheDemocraticRepublicoftheCongo(DRC),withsupportersofbothpartiespresent.Aswewillshow,thiscouldbeconsistentwiththemodelsthatfollowassumingtheycouldinferthatthevotersinattendancewerefarmorelikelytosupportonepartyovertheother,butactionslikethiscouldalsobeintendedtounderminetheelectoralprocess.Nonetheless,focusingonwhoperpetratesactsofelectoralviolenceandtheirintendedtargetisanilluminatingthoughtprocess,andacrucialmissingpieceincurrentlyavailabledata.Weareunawareofacompletesurveyofthecausesofelectionviolenceexistsatthistime,thoughanexcellentsummarycanbefoundinGoldsmith(2015).Wealsodiscussseveralpotentialcausesofelectoralviolencethatarebeyondthescopeofthispaper.Onecommonlyrecurringmes-sagethroughouttheliteraturewrittenbyelectionpractionersisthisso-calledcultureofimpunity,mentionedinBekoe(2011)andCyllahet,all(2014).Basically,timeandtimeagaincountriesinAfricaexperiencesomelevelofelectionviolence,formcommitteestoinvestigatetheoccur-rences,andmakefewprosecutionsorfailtocreateindependentbodieswiththepowertoenforcepunishments.Theseactionssetthestageforviolencetooccuragain.GroupafÞliationsandethnicityareoftenusedbypartiestoencourageviolence.Forexample,politiciansinthe2011electioncampaignintheDRCrefferredtomembersofapoliticalgroupalignedwithanethnicminorityasÒmosquitoesÓandadvocatedtoÒspraysomeinsecticide.ÓInNigeriaformerpresidentOlusegunObasanjosaidtheelectionwasaÒDo-or-dieÓaffair.Soof-tenviolenceisencouragedpubliclyfromleadingpoliticalÞgureheads,whouseexistingethnictensionstodivideandconquer.Asizeableacademicliteraturehasalsosuggestedthatelectionvi-olenceoftenmirrorsexistingethnic,religious,andpoliticaltensionsinsociety.AsstatedinBates(1983),Òelectoralcompetitionarousesethnicconßict.ÓAÞnaluseofelectoralviolencewewilldiscussbrießy,onewhichthispaperÕsresultssuggestisplausible,isgovernmentrepressionofpotentialcompetition.Thisrelatestoaclassicalexampleofrationalconßict,theÞrst-strikeadvantage.Thiscanbesummarizedbyarguingthateveniftheuseofviolenceiscostlynow,ifitpermanentlyeliminatescompetitionitmaybestrategicallyvi-able.Therearecountlessexamplesofcandidateassassinationsandunjustprosecution.Onerecent56exampleofsuchbehaviorincludesthearrestoftwooppositioncandidatesbeforetheirpoliticalrallysinUgandaÕs2015electionsBariyo(2015),thoughthisparticularabuseofstateresourceswasnon-violentitisagoodexampleofhowstatebackedpoliticalpartiescanabusethesystemintheirfavor.NigeriaÕs2007electionsrepresentanother,moreviolent,exampleoftheabuseofstatesresourcesduringanelectioncampaign.Theelectionsresultedinaboutseventydeaths,whichaccordingtoSuberu(2007),includedassassinations.Thissortofmanipulationisthesubjectofseveraltheoreticalpapers,whichaddressabuseofthestatesecurityapparatustorepresscompeti-tioninelections.Itiseasytoseehowpreventingcompetitionfromformingmaybeonewayinwhichsemi-autocraticgovernmentscanstayinpowerviatheuseofelections.Eventhoughelectoralviolenceisdistinctfrompoliticalviolenceingeneral,itisclearthatitcantakemanyforms,andoccurformanyreasons.Thispaperfocusesontwotheoreticalframeworksthatclarifyhowviolencemaybeusedtomanipulatetheoutcomeofanelection.Beforeturningtoadetaileddiscussionofthesepapers,wediscussrelatedtheoreticalpapersonpoliticalcompetition,whichincorporateviolence.2.2.4TheoriesPoliticalCompetitionandViolenceThetheorieswefocusonarenottheonlypossibleexplanationsforelectoralviolence,andourÞndingspotentiallylendsupporttoavarietyoftheories.ThisinvestigationsuggeststhatfurtherresearchwillbeneededtodeterminetheconditionsthatmakeaparticulartheoryÞtacountryorsituationbetterthananother.OnepaperofparticularinterestEllmanandWantchekon(2000)introducesthethreatofviolencebyathirdpartyintoatraditionalHotelingmodelofpoliticalcompetition.Inthismodel,thethreatofviolencebyathirdpartycaninßuencethepartyÕspoliciesandhowindividualvotersvote,withoutactuallyoccurring.WethinkthispapercomplementsourownÞnding,thatthefearofviolencealoneissufÞcienttoinßuencevotingbehavior.AnotherpaperRobinsonandTorvik(2009),focusesontherelatedquestionofwhichgroupofvotersapartybeneÞtsbyintimidating,undecidedvotersoroppositionsupporters.Theyderiveaseriesofpropositionsthatanalyzehowcompetitivetheelectionsarewithandwithoutparticulargroupsof57votersparticipatingintheelectionandconcludethatswingvotersareusuallytheoptimaltargetforviolence.Incorporatingviolenceintotraditionalmodelsofpoliticalcompetition,asisdonebymanyofthemodelsdiscussedinthispaper,canbeusefultounderstandthemechanismsdrivingitsoccurrence.Thepapershighlightedinthissection,andexaminedmorethoroughlyinthefollowingsectionareimportantstepsinthatdirection.Wefeelthatmuchoftheexistingtheoreticalliteraturetakestoonarrowaviewofelectionviolence,byexaminingonlyasinglecountry.Weperhapssteptoofarintheoppositedirection,byexaminingmanycountries;however,wefeelthisisanimportantsteptowardsunderstandinghowgeneralizableexistingmodelsareforstudyingelectionviolence.582.3ModelsandHypothesesThissectioncontrastshowviolenceinßuenceselectoralbehaviorintwoleadingmodelsofelectionviolence;Chaturvedi(2005),hereafterCHandCollierandVicente(2012),hereafterCV.WithfewpapersasidefromBratton(2008)andGoldsmith(2015)whichexaminehowviolenceinßuencesvotingbehaviorandframeworkswhichproducedifferingpredictionsabouttheuseofviolenceinelectionswearguethisisthenaturalstartingpointforinvestigatingelectionviolence.TheÞrstmodel,CH,setsupatwo-partywinner-takes-allelectionwhereeachsideattemptstomaximizethedifferencebetweentheirexpectedvoteshares.Inthesecondmodel,CV,eachsideseekstomaximizetheirvoteshare.Inbothmodels,violenceisusedtoinßuencetheelectionresultsbychangingthecompositionoftheelectoratewhichvotesonElectionDay.Themodelsdifferinwhoabstainsfromvotingfollowingtheuseofviolence.InCVtheassumptionmadeisthatifviolenceoccurs,allswingvoterswouldabstainfromvoting.Incontrasttothisassumption,CHassumesthatviolenceisdirectlytargeted;eachpartytargetstheopponentÕscoresupportersandswingvotersarenotdirectlyimpactedbyviolence.Eachmodeldescribesandsetsupswingvotersinaslightlydifferentfashion.Wetakelibertyinalteringandsimplifyingthesemodelstobetterfacilitatecomparison.Thecoreassumptioninbothmodelsisthatviolencecanreducevoterturnout,weturnthisintoourÞrsthypothesis.Hypothesis3.Violence,perceived,threatened,oractuallyoccurring,reducesthelikelihoodofanindividualvoting.Asdiscussed,thesetwoframeworksprovidetwodifferentwaysinwhichviolencemaybeusedtoinßuencevotingbehavior.TheÞrstframeworkassumesthatanyonewhoistargetedbyviolencewillbelesslikelytovoteasaresultoffearandintimidation.ThisframeworkisusedinChaturvedi(2005),wheretheyassumethatviolenceistargeteddirectlyatoppositionsupporters.ThesecondframeworkfromCollierandVicente(2012)assumesthatinsteadoftargetingindividualswithvio-lenceitmaybeusedtocreatehavocinanareaandcertaingroupsofpeople,speciÞcallyunaligned59voters,wouldbemorelikelytoabstainfromvotinginthepresenceofviolencethanothergroups,suchasvoterswithastrongpoliticalafÞliation.TheAfrobarometerdata,complementedbytheSocialConßictinAfrica(SCAD)data,allowsustoexaminebothpossibilities,ifthepresenceofviolencegenerallyorfearsoftargetedviolenceinßuencevotingbehavior.Ifviolenceisusedasascaretactic,intendedtodisenfranchiselargegroupsofswingvoters3,asintheCVmodel,wewouldexpecttoseedifferentialreactionstoviolence.CVassumesthatswingvoters,sincetheycareleastabouttheoutcome,willhavestrongerreactionstoviolence.Thisleadstooursecondhypothesis.Hypothesis4.IndividualswithoutastrongpoliticalafÞliationshouldbelesslikelytovote,iftheyexpectorfearviolence,thanthosewith.TheframeworkinCHarguesthatcorepartysupportersaretargetedbyviolence,andthereforeceasevoting.Ifthisisthecase,wewouldexpecttoseecorepartysupportersreportingagreaterfearofviolenceinelections,thisleadstoournexthypothesis.Hypothesis5.IndividualswithastrongpoliticalafÞliationshouldbemorelikelytofearviolenceduringanelection.Startingwithasolidunderstandingofhowpeoplereacttothethreatofviolenceduringelectionsisessentialifwewishtodesignempiricalmodelscapableofpredictingitsoccurrence.3Therationalforthisisexplainedinthefollowingsection,theintuitionforwhowishestodothisrelatestowhohasmorecoresupportersinanelectionwhowillcontinuetovoteevenwhenviolenceoccurs602.4ComparingSimpliÞedModelsTheprecedingsectiondiscussedtheassumptionsmadewithregardstohowindividualsreacttoviolenceinthetwomodels.ThissectionexaminessimpliÞedversionsofthesemodelsfocusingonhowviolenceimpactseachpartyÕsvoteshare,whileremovingthealternativespresentintheoriginalmodels.ThissimpliÞcationfacilitatesamoredirectcomparisonofthemodels.2.4.1CollierandVicenteModel(CV)Eachpoliticalparty,C,thechallengerandI,theincumbent,wishtomaximizeitsshareofthevotes.Aunitmassofvotersaredividedintothreegroups,C,I,&S.C,isthefractionofvoterssupportingthechallenger;I,isthefractionofvoterssupportingtheincumbent;andS,isthefractionofunalignedorswingvoters.Ifneithercandidateweretotakeanyactions,afractionofswingvoters,#,wouldvotefortheincumbent,and(1!#)forthechallenger.VotersingroupCandIalwaysvotefortheirrespectivecandidate.Eachpartyhastheabilitytouseviolenceinanelectiontointimidateswingvoters.Coresupporters,CandI,arenotaffected4.Weassumethatviolencehasalinearimpactonvoterparticipation,afterviolence,oflevelv,thefractionofvoterswhocontinuetovoteis(1!v).Thisimpliesthatifswingvotersaretargeted,withviolenceoflevelv,then(1!v)Softhemvote.Withoutlossofgenerality(WLOG),wefocusonresultsfortheincumbent,wecandothisbe-causeeachpartyfacesanoppositebutsymmetricsetofchoices.Absentviolence,theincumbentÕsvoteshareis,VoteShare=I+#SC+I+S.(1.1)Includingviolence,IÕsvoteshareis4Onecouldinterpretthisascomingfromacost-beneÞtcalculation:withtheamountofviolenceneededtodetercoresupportersbeingtoocostly.61VoteShare=I+(1!v)#SC+I+(1!v)S(1.2)Notethatviolenceisnotindexedbyparty,ratherthetotalamountofviolencebetweenbothpartiesiswhatmatters.However,onlyonesidewilluseviolenceinequilibriumundertheseassumptions.Thederivative,orgaininvotesharefromviolence,aftersomesimpliÞcationis:(VoteShare(v=((1!#)I!#C)S(C+I+(1!v)S)2=MB(1.3)Rearrangingshowsthattheincumbentwillraisetheirvoteshareusingviolenceiff,1!##>CI.(1.4)Assumingaconstantmarginalcostofviolence,$,thefollowingpropositiondeterminestheequilibriumlevelofviolence,Proposition3.PartyIwillsetv=1iff:theirvoteshareincreasesbyexcludingtheaverageswingvoter,andthemarginalbeneÞttoexcludingallswingvotersexceedsthemarginalcost,((1!#)I!#C)S(C+I+S)(C+I)>$(1.5)Theproofisasfollows,if1!##>CI,thenIgainsvotesharebydisenfranchisingswingvoters.Thisimpliestheoppositionpartywouldnotuseviolencebecauseofthemodel!ssymmetry.Thesecondcondition,((1!#)I!#C)S(C+I+S)(C+I)>$determinesifthetotalbeneÞtexceedsthetotalcost.Ifbothconditionshold,Igainsenoughvoteshareforviolencetoappeartobeaviablestrategy.Theintuitionforthisisthatmoreweightisplacedoncoresupportersasswingvotersstopvoting.Soapartywouldonlywishtouseviolence,anddisenfranchiseunalignedvoters,iftheyhaveagreateradvantageamongcoresupportersthentheyhaveamongswingvoters.Anumberofinterestingpredictionscanbederivedfromthismodel.Theimpactofviolence62onapartyÕsvoteshareincreasesinS,thefractionofswingvoters.Thisistheresultofmorevotersbeingdiscouragedwhenviolenceoccurs.Theimpactofviolencealsoincreasesinownpartysupport.Thisisbecauseasswingvotersleavetheelection,duetoviolence,moreweightisplacedontheratioofcoresupporters.Finally,theimpactofviolencedecreasesinoppositionsupport,Casset-upabove,andinthefractionofswingorunalignedvoterssupportingtheperpetratingparty,#asset-up.2.4.2ChaturvediModel(CH)TwomaindifferencesoccurwhenexaminingtheCHmodel,Þrstthegoalofeachpartyistomaximizetheirpluralityratherthantheirshareofthevotes;CHdeÞnespluralityasthedifferenceinexpectedvotes.Second,violenceistargetedattheoppositionssupportbase.Inordertopreservethepredictionsfromtheoriginalmodel,wecontinuetomaximizetheplurality.Thecompositionofthevotersremainsthesame,aunitmassofvotersaredividedintothreegroups,C,I,&S.C,isthefractionofvoterssupportingthechallenger;I,isthefractionofvoterssupportingtheincumbent;andS,isthefractionofundecidedorswingvoters.Ifneithercandidateweretotakeanyactions,afractionofswingvoters,#,wouldvotefortheincumbent5.Inthismodelviolenceisindexedbythepartyusingit,becauseitistargeted,againWLOGwefocusontheincumbentÕsdecisiontouseviolence.Theincumbentseekstomaximizetheirplurality,differenceinvoteshare.Thiscanbederived,fortheincumbent,asfollows:IncumbentVotes=(1!vC)I+#S(1.6)ChallengerVotes=(1!vI)C+(1!#)S(1.7)Plurality=(1!vC)I!(1!vI)C+(2#!1)S.(1.8)5Thislastassumptionisthebiggestdivergencefromtheoriginalmodel,howeveritleavesthedirectionofthepredictionsunchanged.63Weseeimmediatelythatthereturnonviolenceisproportionaltotheopponentssupportbase,(Plurality(vI=C.(1.9)ThissimpliÞedmodeldoeslosesomeofthenuancedpredictionsoftheoriginalmodel;how-ever,thecorepredictionsremain.Theimpactofviolenceincreaseswiththeoppositionsvoteshare,andtheuseofviolenceincreasesasthefractionofswingvotersdecreases.BothofthesepredictionsareoppositethoseintheCVinspiredset-up.ContrastingthiswiththeCVmodel,weseethatinsteadofonlyonesideusingviolence,bothsidesalwaysusesomeviolence6.Thisisbecauseaslongastheopponenthassomesupportersintheelection,eliminatingsomeofthosesupporterswouldraisetheperpetratingpartyÕsplurality.Inotherwords,inthismodelbothsidesalwayshaveapositivemarginalbeneÞtfromviolence,whereasonlyonesidehasapositivemarginalbeneÞtinCV.EvenwiththesesimpliÞcations,weareabletoseehowdifferentassumptionsaboutwhoistargetedby,andreactsto,violencecanleadtodrasticallydifferentpredictionsforitsuse.Weturntodiscussingthesedifferencesinthenextsection.2.4.3HypothesesfromBaselineModelsSwingvotersreceiveagreatdealofattentioninmodelsofelectoralcompetition,andareaprimaryfocusofbothmodelsofelectionviolencehighlightedinthispaper.Theyarealsooneoftheprimarydistinctionsbetweenthetwomodels.Theintuitionforthisdifferenceisalsofairlystraightforward.IntheCVmodelasthefractionofswingvotersincreases,eliminatingthemfromtheelectionplacesgreaterweightoneachpartyÕscoresupporters,sowhicheversidehasastrongeradvantageincoresupporterstheninunalignedsupportersstandstogainmorebyexcludingundecidedvoters7.TheCHmodelmakestheoppositeprediction,asthefractionofswingvotersincreasesthereturnto6AddingacostofviolencetothissimpliÞedmodelcouldchangethat,forexamplewithaconstantmarginalcostwecouldseenoviolencebeingusedifthatcostwastoohigh.7Thispredictionassumesthatthedominantpartiespredictedfractionofvotersamongswingvotersislessthantheirfractionofcoresupporters,ifthisdoesnotholdtheweakerpartywoulduseviolence.Theresultingpredictionthatviolenceincreaseswiththefractionofswingvoterswouldremainunchanged.64violencedecreases.InthesimpliÞedmodel,theintuitionforthisisthatareductioninswingvotersimpliesanincreaseincoresupporters.Thisleadstothefollowinghypothesescomparingthepredictionsofthetwomodels:Hypothesis6.Observedviolenceshouldincrease(CV),ordecrease(CH)asthefractionofswingvotersintheelectorateincreases.Onekeydifferencebetweenthesemodels,asidefromwhothemodelsassumeshouldbetar-getedwithviolence,isthedifferenceinobjectivefunctions:maximizingthedifferenceinvotesversusthemaximizingvoteshare.Inpart,testingbetweenthesemodelswouldalsobetestingthefunctionalformdifferencesinthem.However,themainresearchquestioninthispaperfocusesonexaminingwhetherreactionstoviolenceareconsistentwiththeassumptionsineitherframework,notspeciÞcallyonwhichmodelorobjectivefunctionÕspredictionsbetterÞtthedata.So,althoughthedifferingobjectivefunctionsdodrivesomeofthepredictionsineachmodel,wefocusonex-aminingwhetherreactionstoviolenceareconsistentwitheitherofthesetheories.SpeciÞcally,dounalignedvotersshowstrongerreactionstoviolence,anddoesdirectlytargetedviolencedissuadealignedvotersfromvoting?652.5DataSourcesTheprimarydatasourcesusedinthispaperaretheround4Afrobarometer(AB)surveysandtheSocialConßictinAfricaDatabase(SCAD).TheABsurveysprovideauniquelookatpoliticalsentimentacross20Africancountries.Theyincludedavarietyofquestionsrelatedtoelections,politicalparticipation,andelectionviolence.TheSCADdataallowssurveymeasuresofviolencetobecomparedtomoretraditionaleventdata.Combined,thesedatasourcesprovideauniqueperspectivetoexamineelectionviolence.2.5.1PoliticalAfÞliationInthetheoriesdiscussedinthispaper,politicalafÞliationisakeydeterminantofhowanindividualreactstoelectoralviolence.Furthermore,thetheoriesusedivisionoftheelectorateasanimportantpredictorofelectoralviolence.TheABsurveyprovidestwopotentialmeasurementsofpoliticalafÞliation,whichcanalsobeusedtoestimatedivisionoftheelectorate.Thepreferredmeasureiscreatedfromthefollowingquestions.ThesurveyÞrstasks,ÒDoyoufeelclosetoanyparticularpoliticalparty?Ó.Iftherespondentansweredyes,theywerethenasked,ÒWhichpartyisthat?ÓRespondentswhosaidÒNoÓ,totheÞrstquestionareclassiÞedasunalignedvoters.RespondentswhoprovidedaspeciÞcpartyidentiÞcationarethenplacedinoneofthreegroups:1)thosewhosupportthepartywiththehighestnumberofresponsesinthecountry,2)thosewhosupportthepartywiththesecondhighestnumberofresponses,and3)thosewhosupportanyotherparty.Thesurveyasksanotherrelatedquestion,ÒIfapresidentialelectionwereheldtomorrow,whichpartyÕscandidatewouldyouvotefor?ÓRespondentsaredividedintoanequivalentsetofgroupsusingthismeasure.WefocusontheÞrstsetofquestionsfortworeasons.First,thetheoriescareaboutpartysupporters,andtheÞrstsetofquestionsasksdirectlyaboutpartyafÞliation.Second,weshowthatthereisahighdegreeofoverlapbetweenthemeasures.Finally,usinginformationgatheredfromtheInter-ParliamentaryUnion,wecodedthequestionÒWhichpartyisthat?Óas1iftherulingpartyorcoalitionwasindicatedoras0ifitwasnot.66Table(2.1)displayssummarystatisticsonpoliticalafÞliation.Slightlyoverone-thirdofre-spondentssupportthestrongestpoliticalparty,MajorityParty,almostathreetooneratiowhencomparedtothenextlargestparty,SecondParty.Whenexaminingwhichpartyindividualsclaimtheywillvotefor,thisratiorises.Table2.2,demonstratesconsiderableoverlapbetweenthetwomeasuresofpoliticalafÞliation(i.e.,supportandvoting).Theexceptiontothisisthatmanyin-dividualsmaynothaveastrongpoliticalafÞliationwithapartybutstillknowwhichpartytheywouldvotefor.Bothmeasuresshareoneparticularlyproblematicweakness:politicalintimidationmayhaveswayedanindividualÕsreportingbehaviorinoneoftwopossibleways.ThosewhoarefearfulofintimidationmaybemorelikelytoreportanafÞliationwiththeintimidatingpartyorrefusetoreportoneatall.ToexaminewhetherthereportedpoliticalafÞliationseemstobebiasedbyintimidation,weleveragedquestionsavailableintheABsurvey.TheABdatacontainsawidevarietyofquestionsrelatedtothepoliticalatmosphereineachcountry.Weareabletousethistoinvestigatewhetherindividualsmaybemisreportingtheirpolit-icalafÞliationbecausetheyfearrepercussionfromadmittinganafÞliationwithminoritypoliticalparties.Inparticular,wefocusononequestionfromthesurvey,Òhowoften:dopeoplehavetobecarefulofwhattheysayaboutpolitics?Ó,andexamineitscorrelationwithanindividualÕspoliticalafÞliation.IfindividualsareafraidtoaccuratelyreporttheirpoliticalafÞliation,wemayÞndthatpeoplewhofeeltheymustbecarefulwhattheysayaboutpoliticsaremorelikelytoreportanafÞliationwiththerulingpartyortoreportnoafÞliationatall.TheOLSregressionsinTable(2.3)examinethispossibility.WeÞndthatthemorecarefulanindividualfeelstheymustbeaboutwhattheysay,thelesslikelytheyaretoreportbeingamemberoftherulingparty.Thisprovidesevidencethatdespitetheirneedtomonitorhowtheytalkaboutpolitics,peoplearenotmisreportingtheirpoliticalafÞlia-tioninfavoroftherulingparty.Whenexaminingthesecondpossibility,wedoseethatindividualswhofeeltheymustÒalwaysÓbecarefulwhattheysayareslightlymorelikelytobecategorizedasanunalignedvoterbyourprimarymeasure;however,weÞndnosuchrelationshipwiththealter-67nativemeasureofpoliticalafÞliationwhichaskswhichpartytheywouldvotefor.Thisevidencesuggeststhattheeffortsbythesurveyenumeratorstoinstilltrustwhendiscussingsuchcontentioustopicsmusthavealleviatedatleastsomeconcernsamongamajorityofrespondents.2.5.2PoliticalParticipationTomeasurepoliticalparticipationwederivetwovariablesfromtheABdata.TheÞrstisretro-spectiveandistheprimaryvariableofinterest.ThesecondishypotheticalandisderivedfromthesamequestionasoursecondsetofpoliticalafÞliationquestions.Thetwovariablesareasfollows:¥Voted:ReportedthattheyvotedinthepreviousNationalElectioninthecountry.72%ofrespondentsreporttheyvotedinthelastelection.¥WontVote:Whenasked,ÒIfapresidentialelectionwereheldtomorrow,whichpartyÕscan-didatewouldyouvotefor?Ó;respondentreportedtheywouldnotvote.8%ofrespondentsreporttheywouldnotvoteinahypotheticalelection.2.5.3ViolenceMeasureOfparticularimportancetothispaperishowtheABdatawereusedtomeasureviolence,es-peciallyhowwecapturethedifferencebetweentargetedanduntargetedviolence.Mostpaperstryingtopredictviolenceorunderstanditsimpactuseeventdata,whichcountsthenumberofeventsreportedinnewspapersmeetingsomecriteriainagiventimeandspace.Weuseeventdatabutcomplementitwithtwosurveymeasuresofviolence.Oneofourmeasuresofviolenceisfundamentallydifferentfromevendata.Itisasurveyresponsetothequestion,ÒDuringelectioncampaignsinthiscountry,howmuchdoyoupersonallyfearbecomingavictimofpoliticalintim-idationorviolence?ÓRespondentscouldchooseoneoffourcategories;ÒAlotÓ,ÒSomewhatÓ,ÒAlittlebitÓ,orÒNotatallÓ.Thisisthenturnedintoanorderedcategoricalvariableonazerotothreescale,withthreecorrespondingtofearingviolence,ÒAlotÓ.Wecallthisvariable,ÒfearviolenceÓ.Wechoosethisasourmeasureoftargetedviolencebecauseofthewaythequestionwasworded,68referringspeciÞcallytotheindividual,butretaineditasameasureoftargetedviolencebecauseitdifferssigniÞcantlyfromothermeasuresofviolenceevenwithinsmallgeographicareas,implyingthattheremainingvariationmustbeattributableindividualdifferencesamongtherespondents.Itispossibletheseindividualdifferencesarenottheresultoftargeting,butpatternsweseesuchaspoliticalafÞliationbeingsigniÞcantevenaftercontrollingfordistrictÞxedeffectssuggestthisisareasonablemeasurefortargetedviolence.Inordertomeasureuntargetedviolence,weusetheeventdatafromtheSCADandthesurveyquestionwhichasks,ÒInyouropinion,howoften,inthiscountry:Doescompetitionbetweenpoliticalpartiesleadtoviolentconßict?ÓThis,liketheprevioussurveyquestion,isalsoplacedonazerotothreescale.ThissurveyquestionhasslightlymorevarianceexplainedbygeographicheterogeneitywithanR-squaredofaboutfourpercentagepointshigher,thantheotherquestion.Also,asdemonstratedinTable2.5Column(2),ithasastrongercorrelationwiththeeventdata.ForthesereasonswearguethattheÞrstquestionmeasurestargetedviolence,likeintheCHframework,andthesecondquestion,alongwiththeeventcountdata,bettermeasuresuntargetedviolence.Thetermuntargetedviolencewaschosentohighlightthekeydifferencebetweenthemodelswearefocusingon,butneitherthedatanorthetheoryrequiresthisviolencebetrulyuntargeted.TakethisexamplefromaHumanRightsWatchreport,WitnessesreportednumerousincidentstoHumanRightsWatchinwhicharmedthugs,usuallythoughnotexclusivelyfromthePDP,shotintotheairorotherwisethreatenedvoterswithviolence,createdchaos,andthenranawaywiththeballotboxes.Insomeinstances,thesegroupsshotdirectlyatindividualsfromopposingparties.Inothercases,theirthreateningbehaviorandpublicdisplayofweaponsrangingfromknivestoÞrearmswassufÞcienttoscareofftheiropponents,aswellasordinaryvoters.(pg.6)Theuntargetedviolenceinthisexampleisthepublicdisplay,whichimpactedallvoters.An-otherwaytothinkofuntargetedviolenceisgeneralorpervasiveviolence.TomeasurethisweusemeasuresofviolencethatwecanÕtidentifywhowastargeted,evenwhentheeventswhichcaused69themeasuretoincreasemayhavebeentargeted.However,forthehypothesiswewishtofocusonwiththesemeasuresofuntargetedviolencethatisÞne,becauseweareinterestedinseeingifunalignedvotersreactmorestronglytoviolence,evenifitisnotdirectedspeciÞcallyatthemandthesemeasuresofviolencedonotsuggestthisviolencewasdirectedprimarilyatunalignedvoters.Eachapproachtomeasuringviolencehasitsownstrengthsandweaknesses.Themainadvan-tageofusingasurveyresponseisthatthesurveyprovidesmeasuresofeachpersonÕsperceivedlevelofthethreat,individual-leveldemographiccharacteristics,aswellaseachrespondentÕspo-liticalafÞliationandbehavior.Furthermore,notallviolencecanbecapturedineventdata.Forexample,threatsofharmareoneparticularcategoryincludedinthedeÞnitionofelectoralviolencethateventcountdatawouldbeincapableofcapturing.Theabilitytocontrolforheterogeneityinsmallgeographicregionsandnotwipeoutourmeasuresofviolencewiththosecontrolsiskeytoourabilitytoexaminetargetedanduntargetedviolence.Theprimarydisadvantagesofthesurveymeasuresrelatetotheinabilitytoascertainwhatexperienceledarespondenttoreportahigherlevelofperceivedviolenceanditspossiblecorrela-tionwithindividualattitudes.Toovercomesomeoftheissuesthisposes,weusecomplementaryquestionsontheABsurveyandcontrolforavarietyofalternativeexplanationsfortheobservedcorrelationswhenwemovetoourresultssection.Thefollowingtablessummarizeandcomparetheviolencemeasuresusedinthispaper.Ofparticularinterestishowthesurveymeasuresofviolenceandeventcountdatacorrelate.FollowingGoldsmith(2015),weusealleventsinawindowaroundelectionsfromtheSCADdatatocaptureelectionviolence.Weusethedatesofallelections,whichoverlapwithbothSCADÕs1990-2012timeframeandTheNationalElectionsacrossDemocracyandAutocracy(NELDA)dataÕstimeframe,whichendsin2010.Table(2.4)showsthemeansofthesemeasures.BecauseoftheoverÐdispersionpresentintheeventdata,theloggedvaluesandindicatorsareusedinanalyses8.Table(2.5)examineshowwellthesurveymeasuresofviolencepredictviolenteventsusingregionallevelaveragesofviolenceinOLSregressionswithcountry-levelÞxedeffects9.Themea-8Theactualconversiontakesthenaturallogofthenumberofeventsplusone.9Theregionisusuallyacountry!sÞrst-leveladministrativedivision70sure,FearViolence,showsamuchweakercorrelationthanViolentCompetition.Thisagainsup-portsourideathatthesemeasurescapturedifferentphenomenaandtheideathattheÒfearviolenceÓmeasuremaybewellsuitedtocapturingtargetedviolence.712.6EmpiricalInvestigation2.6.1ReactionstoViolence:TargetedandUntargetedFundamentaltoanytheoryofelectionviolenceisanunderstandingofhowviolenceinßuencesbe-haviorinelections.Bothofthemodelsassumethatviolencereducesvoterturnout,anassumptionwhichhasfoundvalidityinotherempiricalworksuchasCollierandVicente(2014)andBratton(2008).Toexaminethisandthetwopotentialframeworksforhowviolenceinßuencesvotingbe-havior,weexplorehowself-reportedmeasuresofvotingintheABdataarerelatedtoviolenceusingavarietyofviolencemeasures.Giventheuniquestructureoftheavailabledata,carefulconsiderationneedstobegiventothemethodofanalysisused.TheABsurveywasconductedbetween2008and2009across20differentAfricancountries,alltakenattimesthatdonotdirectlycorrespondwiththecountryÕselections,whichposesuniquedataanalysisissues.However,cross-countrysurveydataisbecomingmorecommonlyusedinpoliticalscienceresearch.TwomajorexamplesaretheABsurveyswhichweuseandtheLatinAmericanPublicOpinionProject,Seligsonet.all,whichexaminessimilarquestionsinLatin-Americancountries,e.g.,Machadoet.all.Withtheintroductionofthesedatasets,somedebatehasarisensurroundingtheiruse.Severalissuesraisedinclude:thevalidityofpoolingresultsacrosscountries,whetherornotÞxedeffectsatthesurveyorcountrylevelaresufÞcienttodealwithheterogeneity,andtheproperlevelatwhichtoclusterstandarderrors.Toaddressissuesrelatedtocountrylevelheterogeneityandthevalidityofpoolingresultsacrosscountries,weincludevaryinglevelsofÞxedeffects:country,region,anddistrict.CountrylevelÞxedeffectsareusedprimarilytocontrolformajordifferencesacrosscountrieswhichwillaffectthesurveyresponses,suchastheoverallatmosphereofviolence,thetypeofelectoralsystem,andamountoftimesincethelastelectionoccurred.Regionanddistricteffectsareusedtoexaminevariationwithinsmallergeographicareas.Re-gionscorrespondtothecountriesÕÞrstorsecondleveladministrativedivision.RegionalÞxed72effectsareusedprimarilybecausesomevariablesshowlittlevariationatthedistrictlevelandtocontrasttheresultsacrosslevelsofÞxedeffects.ThedistrictisrelatedtothePrimarySamplingUnit(PSU)andisaclusterofabouteightrespondentslivingwithinrelativelycloseproximitytoeachother,usuallythesametownorvillage.PrimarilywefocusonspeciÞcationswithdistrictlevelÞxedeffectsbecauseallindividualswithintheserelativelysmallareasaresubjecttoasimilarsetofexperiencesandwecanthenattributedifferencesinsurveyresponsestovariationinindividuallevelcharacteristicsaswellastargetingasaresultofthesecharacteristics.Lastly,weaddresstheissueofstandarderrorsbyclusteringattheregionallevel;thisavoidsthecommonissueofmanyobservationswithtoofewclusters.Thislevelofclusteringisalsoappropriategiventhatmanyindependentvariablesaremeasuredatthislevel.Finally,toexplorethesourcesofheterogeneityandhowtheyimpacttheresults,weincludeinteractiontermsofcountrylevelvariableswithindividualcharacteristics.Doingthisallowsustoexploresomesourcesofcross-countryheterogeneitywhilecontrollingforotherunobservedsourcesofcountry-levelheterogeneity.ThistechniqueisusedinsteadofcomparingtherangeofcoefÞcientscountry-by-country,largelybecausewehavetoofewobservationsinsomerelevantsub-populationswithinspeciÞccountries.ThisleadsustobaselinespeciÞcationstoexaminetheimpactofviolenceonvotingasfollows:Yij=&Vij+$j+%ij.(1.10)Whereidenotestheindividualandjdenotesthelocation(country,region,ordistrict).Yijisanindicatorthatequalsoneiftherespondentireportedtheyvoted,Vijisthemeasureofviolence,forindividuali,$jisaÞxedeffect,thelevelofwhichvariesacrossspeciÞcations.Finally,%ijisanindividualerrorterm.Additionally,weexaminespeciÞcationslikethefollowingtoinvestigatethesourcesofcross-countryheterogeneity.Yij=&Vij+'Vij'Vj+$j+%i.(1.11)73WherenotationremainsthesameasthepreviousspeciÞcation,withtheadditionofVj,thecountrylevelmeanofviolence.Thisappearsonlyasaninteractionbecausethebasetermissubsumedby$j.Table(2.7)presentstheprimaryresultsforH1:Violence,perceived,threatened,oractuallyoccurring,reducesthelikelihoodofanindividualvoting.Theseresultsincludenocontrolvari-ablesanddifferinglevelsofÞxedeffectsacrossthreemeasuresofviolence.Controlvariables,suchasdemographicfactors,areexcludedfromthesebaselinespeciÞcationsbecausewebelieve,andwilldemonstrate,thatgroupswithalowerpropensitytovoteareoftentargetedwithviolence,i.e.unalignedvoters.Normally,onewouldattempttoremovethatbias;however,ifthosedemo-graphicfactors,likeethnicity,areusedtoidentifyanindividualÕspoliticalafÞliation,removingtheobservedcorrelationwouldresultinadownwardbiasonourcoefÞcientofinterest.Forexample,ifethnicityisusedtoidentifypoliticalafÞliationandtargetanindividualwithviolence,control-lingforethnicitymayreducetheimpactofviolenceonvoting,byattributingareductioninthelikelihoodofvotingtoethnicity,whenitisduetothoseindividualsbeingtargetedwithviolencebecauseoftheirethnicity.Differentmeasuresofviolenceusedaremeanttocaptureexposureindifferentways.Asdis-cussedintheprevioussection,FearViolenceismeanttomeasuretargetedviolence,whereastheothertwocaptureuntargetedviolence.Finally,thevaryinglevelsofÞxedeffectsareincludedinordertoeliminateidiosyncraticdifferencesacrossareasandcountriesthatmayaffectvotingbe-havior.Theyareespeciallyusefulatthesmallestlevel,thePSU,tounderstandhowdifferingfearsofviolencewithinevenasmalltownorvillagecanhavedifferentialimpactsonvotingbehavior.Itisreasonabletoassumethatindividualswithinoneofthesedistrictsareexposedtoanearlyidenticalsetofviolenteventsandthatobserveddifferencesinresponsesandbehaviorarelargelyattributabletotheidiosyncrasiesinperceivedriskofviolence.Interpretingtheseresults,weÞndthatindividualswhoreportfearingviolenceÒAlotÓorÒSomewhatÓarebetweenthreeandfourpercentagepointslesslikelytovotethanthosewhofearviolenceÒAlittlebitÓorÒNotatallÓ.ThisisseeninTable(2.7).Whenlookingattheeventdata,74thebottomrowofTable(2.7),weseethatindividualsinaregionwithatleastoneeventreportedarealittlemorethanfourpercentagepointslesslikelytovote.Column(2)inTable(2.7)interactstheindividualviolencemeasurewiththecountrylevelaverageoftherespectivemeasureofviolence.Forourmeasureoftargetedviolence,themeanisinsigniÞcantandtheresultsareunchanged.Forthemeasuresofuntargetedviolence,weseeanincreaseintheprimarycoefÞcient,resultinginagreaterreductionofvotinginareaswhereviolenceoccurs.However,thisincreaseismitigatedbythesigniÞcantandpositivecoefÞcientontheinteractionterm,incountrieswherethemeanlevelofviolenceisgreatertheeffectofanindividualeventislessened.Wearguethisisactuallytheresultofsingleeventscausingagreaterreductioninvoterturnoutwhenviolenceislesscommon.Beforeaddressingconcernsaboutthevalidityoftheseestimates,weturntooursecondhypoth-esis:individualswithoutastrongpoliticalafÞliationshouldbemorelikelytobedeterredfromthepollsbyviolence.ThishypothesiscomesfromtheassumptionintheCVmodelthatunalignedvotersarethemostresponsivetoviolence.Toinvestigatethiswecreateanindicatordividingvot-ersintotwogroups.Thesurveyaskedrespondentsiftheyfeltclosetoaparticularpoliticalparty.RespondentswhoansweredÒNoÓwereclassiÞedasunalignedvotersinourvariableUnalignedVoter.Thefollow-upquestioniftheyansweredÒYesÓ,ÒWhichpartyisthat?Ó,wasusedtode-terminetheactualpoliticalafÞliationamongdecidedvoters.ThefollowingequationshowsthebaselinespeciÞcationtoinvestigateheterogeneousreactionstoviolence,Yij=&Vij+!PAij+'PAij'Vij+$j+%ij.(1.12)TheadditioninthisspeciÞcationisPAi,anindividualpoliticalafÞliationindicator,whichequals1iftherespondentdoesnothaveastrongafÞliationwithanypoliticalparty.ThespeciÞcationsinTable(2.8)showaninterestingpatternofhowviolenceinßuencesvotingbehavior.ThecoefÞcientonFearViolenceisnegativeandsigniÞcantinallspeciÞcations,withnoheterogeneousrelationshipbetweenpoliticalafÞliationandfearofviolence.Inotherwords,anyonewhofearsviolenceislesslikelytovote.Withtheothermeasuresofuntargetedviolence,ViolentCompetitionandtheeventcountmeasure,weÞndtheoppositepattern:thereisnoprimary75effectbutthereisaheterogeneouseffect,bypoliticalafÞliation.Onlyrespondentswhoareclassi-Þedasunalignedvotersarelesslikelytovoteinthepresenceofuntargetedviolence.AlthoughthelevelofsigniÞcanceontheinteractionvaries,thepointestimatesareconsistent.Furthermore,aswediscussedpreviously,theimpactontheeventmeasuresarelikelytobeattenuatedtowardzero.Weseeafairlyconsistentestimateofaonecategoryincreaseinreportedviolencecorrespondingtoabouta1.5percentagepointreductioninthelikelihoodofvoting,regardlessofpoliticalafÞliation,forpersonalfearofviolenceFearViolence.Fortheeventmeasureofviolence,theresultsindicatethatinregionswhereaneventoccurred,unalignedvotersarebetweentwoandfourpercentlesslikelytovote,Columns(4)-(5)Table(2.8).Finally,thesurveymeasureofoccurrence,ViolentCompetition,showsresultssimilartotheeventmeasure.Thisdichotomyofresultsisimportantandsupportsthekeyassumptionsinbothmodelsofhowviolenceinßuencesvotingbehavior.Itimpliesthattargetedviolencemaybeeffectiveatpreventingallvoters,includingcorepartysupporters,fromvoting,asintheCHmodel.Italsoimpliesthatuntargetedviolencemaybeenoughtodisenfranchiseunalignedvoters,asintheCVmodel.Itisreasonabletoassumethatthefearofviolenceinelectionsmayinfactcauseanindividualtoabstainfromvoting;however,thenatureofthisdatadoesnotallowdirectclaimsofcausality.Furthermore,thereareseveralalternative,non-causal,explanationsfortheobservednegativecor-relationbetweenviolenceandvoting.Weinvestigateandruleoutseveraloftheseexplanations,including:omittedvariables,responseorrecallbias,non-linearimpacts,andcorrelationswithalternativeformsofmanipulation.First,toaddressomittedvariables,were-runsomeoftheabovespeciÞcationswithasetofdemographiccontrolvariablesandsubsequentlyaddressthepossibilitythatsomeindividualsmaysystematicallyreporthigherlevelsofperceivedviolenceforreasonsthatarecorrelatedwithvotingbehavior.ForthesewefocusonFearViolenceasthemeasureofviolence.Aftertheinclusionofdemographiccontrolsforage,gender,education,andsocio-economicstatus,thereisamodestreductiononthecoefÞcientofabouthalfapercentagepoint,butnochangeinsignorlevelofsigniÞcance,Table(2.9)Columns(1)and(2).ThedifferencebetweenthesecolumnsisthatCol-76umn(2)includesindicatorsforeachofthepossibleresponses.Demographicfactorsarecertainlyimportantdeterminantsofvotingbehavior,thoughtheydonotaltertheobservedrelationship.Anotherconcernisthatindividualsmaybereportingafearofviolenceforreasonsnotcon-nectedspeciÞcallytoelectionviolence.Toaddressthisweincludefearcrimewiththeassumptionthatthismaycapturethecorrelationbetweenothertypesofviolenceandahigherpropensitytoreportgreaterfearofelectionviolence.IncludingfearcrimeinTable(8)Column(5)resultsinacoefÞcientnearzeroonfearcrimeandnochangetoourcoefÞcientofinterest.Thisprovidessomeevidencethatourresultsarenotdrivenbyother,moregeneralfearsofviolence.Additionally,wemayworrythatindividualswhoaredisillusionedwiththedemocraticprocessandthuslesslikelytovotemayalsobemorelikelytoreportthatviolenceoccursthroughouttheelectionprocess.Thiscouldleadtoabiasedestimateoftheimpactofviolenceonvoting.Toexaminethisalternativeexplanation,anABitemwhichasksindividualsabouttheirsatisfactionwithdemocracywasincludedinanadditionalspeciÞcation.Thisshowednochangeonthevio-lencecoefÞcient,butasonewouldexpect,itdidhaveastrongpositivecorrelationwithvoting,Table(2.9)Column(6).TheseadditionalspeciÞcationscannoteliminateallpossiblealternativeexplanations,butdoeliminatesomepotentialnon-causalexplanationsfortheobservednegativecorrelationbetweenviolenceandvoting.Asecondpotentialconcernisthatstructuralconditionswhichencourageviolencemayalsodis-couragevoting.Forexample,remoteareasmaybemoreviolence-proneandentailhighervotingcosts.OurbaselinespeciÞcationsmakesigniÞcantheadwayineliminatingthisasanalternativeexplanationthroughtheuseofÞxedeffects.ThesmallestlevelofÞxedeffectsweused,thedis-trict,istheintendedprimarysamplingunit(PSU)intheABsurveys.Insuchasmallgeographicareainwhichanenumeratorisabletowalkfromhouse-to-housecollectingsurveyresponses,allindividualswouldlikelybeexposedtoidenticalstructuralconditions.Giventhatweseethatfear-ingviolenceretainsasigniÞcantcorrelationwithvotingwhencontrollingfordistrictÞxedeffects(i.e.,wipingoutdistrict-levelvarianceinstructuralconditions),thisseemstobeanunlikelyexpla-nation.ThishighlydisaggregatedlevelofÞxedeffectsalsoaddressesanumberofotherconcerns77withregardstotheendogeneityofviolence.Forexample,onemaybeconcernedwiththeen-dogeneityofthetimingofviolence.However,beingthatweseereactionstoviolencevaryevenwithinasmalltownorvillageandacrossdifferentpoliticalafÞliationssuchconcernswithfactorscommonacrossnational,orevensub-nationallevels,aremitigated.Additionally,weexaminedanothermeasureviolencewhichshouldnotbecorrelatedwithvotingbutwouldbecorrelatedwithstructuralconditions:post-electionviolence.Ahighnumberofpost-electionviolenteventswouldindicateagenerallyviolentdistrictandthusmorefearofviolence.However,whenincludingthenumberofpost-electionviolenteventsinanadditionalspeciÞcation,wedonotseeasigniÞcantreductioninvotingbehavior,againdemonstratingthatstructuralconditionsarenotdrivingtheobservedcorrelation,thesespeciÞcationswereommitted.Anotherpossibleestimationissueisthattheimpactofviolenceisnon-linear.Furthermore,giventhatourdependentvariableisdichotomous,linearprobabilitymodelshaveanobviouslim-itation:theydonotrestricttheestimatetothezero-onerange.Giventhatthedependentvariablevotinghasameanofabout.7,thisisanunlikelyconcenr.Toexamineifoureffectsarenon-linear,weranavarietyofalternativespeciÞcationsandÞndresultstobesimilartoourbaselinespeci-Þcations.Figure(??)graphstheestimatedprobabilitiesofvotingfortheaveragerespondentinthesampleusingaÞxedeffectsprobitmodel.ThisÞgurealsodisplaysalineartrend.However,Column(2)inTable(9)doesshowthatthosereportingfearingviolenceÒSomewhatÓorÒAlotÓ,thetwohighestcategories,driveourresults.However,theestimatedmagnitudesdonotchangeverymuch.ThelastpossibilityweexamineisthatpastvotingbehaviorstronglyinßuencesanindividualÕsviewofviolence.Toaddressthisconcern,weexaminetheimpactofviolenceonthehypotheti-calWontVotemeasurewhilecontrollingforanindividualÕspastvotingbehavior.Onceagainthepredictedinßuenceofviolenceonvoting,thistimemeasuredbyWontVote,remainsunchanged.Giventhatthenegativecorrelationofviolencewithvotingremainsafterconsideringanumberofnon-causalexplanations,itisreasonabletoconcludethatthereisacausalimpact.Evenaddressingallofthesepotentialalternativeexplanations,itispossiblethatotherfactors78maybedrivingtheobservedcorrelation.Takeforexampletheendogeneityofelectionsthemselves,electionsmaybepushedquicklyduringviolenttimesasahopetoresolveaconßict,howevereveninsuchcasestherewouldbesigniÞcantvariationwithregardstowhereviolenceoccurswithinacountry,andweseethatvotingratesarelowerinmoreviolentareas.ThefactthattheseresultsholdacrosssuchadiversesetofcountriesinAfrica,suggeststhatviolencedoesreducethelikelihoodanindividualvotes.Theabilitytocontrolforothersourcesofviolence,andtheheterogeneityinreactionstoviolenceareallresultsthatwouldbedifÞculttoattributetoareversedrelationship.Insummary,fearingviolenceseemstohaveaconsistentlynegativeimpactonvotingbehavior.WealsoÞndthatgeneraloruntargetedviolenceappearstoreducevotingamongunalignedvoters.Forthisreason,wenowturntoanalyzinghowreportedfearsofelectionviolencearerelatedtopoliticalafÞliation.2.6.2WhoFearsViolenceinElections?Boththeoriesmakeimplicitassumptionswithregardstowhoshouldfearviolenceinanelection.Usingtheseassumptions,wederivedH3:corepartysupportersaremorelikelytofearviolenceduringanelection.ThiswasbasedontheCVmodelÕsassumptionthatviolenceisdirectedatcorepartysupporters.Totestthis,weranaseriesofOLSregressions,thistimefocusingonpoliticalafÞliation.BecausedemographiccharacteristicsmaypredictpoliticalafÞliation,weagainstartedbyexcludingcontrolsandreliedheavilyontheuseofdifferinglevelsofÞxedeffectstoseeifdifferentgroupsofpeoplereportedsystematicallydifferentfearsofviolencewithinthesameelection.FocusingontheresultsinTable(2.10),wedonotseedirectsupportforthesimplehypotheseslaidoutinH3.However,politicalafÞliationisastrongpredictorofelectoralviolence.Focus-ingÞrstontheresultsinColumn(1),whichincludesnoÞxedeffects,thebasicpatternweseethroughoutthisanalysisemerges:supportersoftherulingpartyfearviolencetheleast,followedbyunalignedvotersandsupportersofthesecondlargestparty.However,asidefromthosesup-portingthemajorityorrulingparty,differencesaresmallandinsigniÞcant.Theabilitytocontrol79forheterogeneityatsuchasmallgeographicunitisimportant;thereisconsiderablegeographicvariationintheconcentrationofpoliticalpartiesinthedataandwewouldliketoensurewearecontrollingforthatwhenwecomparefearsofviolenceacrossdifferentpoliticalgroups.Exam-iningColumns(2)Ð(5),whichincludedistrictÞxedeffects,weseethepatternremainslargelyunchanged.Column(2)replicatesColumn(1),butintroducesÞxedeffectstocontrolforthepos-sibilityofindividualsofthesameafÞliationlivinginparticularlyviolentornon-violentareas.Column(3)usesadichotomousmeasureofviolence,dividingrespondentsintohighandlowcate-gories.Column(5)doesthesame,butusesthealternativesurveymeasureofviolence.Column(4)usesthealternativeclassiÞcationofpoliticalafÞliationwhichcategorizesindividualsassupportersoftherulingparty,oranotherparty.RegardlessofthespeciÞcation,theresultsaresimilar.ThisÞnding,thatsmallpoliticalpartiesreportthehighestandmajoritypartymembersthelowestfearofviolence,isconsistentwithanalternativeexplanationoftheuseofelectionviolence:repressionandcontrolofthestatesecurityapparatus.ThisismostevidentinColumn(4)inTable(2.10),whereweseethatthosewhosupporttherulingpartyfearviolencesigniÞcantlylessthanallothergroups.Overall,theresultsfromexaminingthecoreassumptionsmadeinthesemodelsaremixed.WeÞndstrongevidencethatviolencereducesthelikelihoodofvoting,althoughthereisimportantheterogeneityinthoseresults.Theobservedbehaviorisconsistentwithbothmodels.AsintheCVmodel,weseethatunalignedvotersaremorereactivetoviolence:simplytheoccurrenceofviolenceisenoughtodeterthem.WecanalsosupporttheassumptionoftheCHmodelthatvio-lencetargetedattheoppositionÕssupporterswillreduceturnout:regardlessofpoliticalafÞliation,anyonewhopersonallyfearsbecomingvictimizedbyviolenceislesslikelytovote.WedonotseeapatternconsistentwitheithermodelwhenexaminingwhoistargetedbasedsolelyonpoliticalafÞliation.Takenasawhole,thisevidencesuggeststhatpeoplereacttoelectoralviolenceinamethodconsistentwiththeoriesofvotermanipulation,althoughthisdoesnottellusanythingdi-rectlyabouthowitisactuallyusedinelections.Toexaminethat,weturntoaggregateresultsandanalyzehowdivisionoftheelectoratepredictsthelevelofviolenceinanareawithinacountry.802.6.3WhereDoesViolenceOccur?WetestonehypothesisderivedfromthemodelsÕpredictionsrelatedtodivisionoftheelectorate,H4:Observedviolenceshouldincrease(CV)ordecrease(CH)asthefractionofunalignedvotersintheelectorateincreases.Thisallowsforadirecttestbetweenthetwomodels,whichhasusefulimplicationsfortryingtopredicttheoccurrenceofelectionviolence.Totestthishypothesis,weexaminehowthefractionofunalignedvoters,measuredusingUnalignedVoter,isrelatedtothreemeasuresofviolence.LookingattheresultsinTable(2.11),weseenorelationshipbetweenthefractionofunalignedvotersandrespondentsÕreportedfearofviolence.However,theassociationbetweenviolenceandthefractionofunalignedvotersdiffersdependingonthemeasureofviolence.Usingtheeventcountmeasure,weseeasigniÞcantpositivecorrelation.However,usingthesurveymeasureofuntargetedviolence,ViolentCompetition,weseethattheassociation,althoughpositive,isinsigniÞcant10.Fortheeventcountmeasure,aonepercentagepointincreaseinthefractionofunalignedvotersleadstoanincreaseofapproximatelyone-halfapercentagepointinthelikelihoodaneventoccurs.Resultsaresimilarusingthelognumberofevents,buttheinterpretationisfarclearerwiththeoriginalindicator.TheinsigniÞcantrelationshipbetweenfearingviolence,whichisourmeasureoftargetedviolence,wasexpected.IfviolenceisuntargetedasintheCVmodel,wewouldnotexpectanincreaseinfearoftargetedviolence.Amorethoroughtestbetweenthesemodelswouldrequireamoreaccuratepictureofwhomtheviolenceintheseareaswasaimedatandwhoperpetratedit.Takingtheevidenceforthepredictionsandassumptionstogether,theCVmodelismoresupported,suchthatviolenceisusedprimarilytodeterunalignedvotersfromvoting.Furthermore,theevidencealsopointstowardthemajoritypartybeingtheperpetratorsaswellasanalternativeexplanation:violencemaybeintendedtokeepeveryoneexceptmajoritypartiesfromparticipatinginelections.ThisisnotsurprisinggiventhatthevastmajorityofelectionsinAfricaarewonbyincumbentpoliticalparties.10RemovingtheweightingraisesthevalueofthecoefÞcientbyaboutone-thirdandresultsinasigniÞcanctrela-tionship,thisislikelyduetotheCVmodelbeingabetterÞtincountrieswithmoreobservations,whichrecievelessweightintheregressionsafterincludingsurveyweights.812.7ConclusionDemocraciesare,unfortunately,notimmunetoviolence.Thehopeisthatelectionscanbeusedasameansofpeacefulcompetition;however,thisisnotalwaysthecase.Themodelsexaminedinthispaperprovideonepotentialexplanationfortheuseofviolenceinelections:strategicma-nipulationofwhovotesonpollingday.ThereisvalidityforeachframeworkÕsassumptionofhowviolenceinßuencesbehaviorinelections:bothtargetedanduntargetedviolenceseemtobeeffec-tiveatreducingturnout.Wefoundimportantsourcesofheterogeneityinreactionsto,andinfearsof,electionviolence.Allvoterswhofearbeingvictimizedbyviolencereactinthesameway:theyaremorelikelytoabstainfromvoting.However,notallvotersceasevotingsimplybecauseviolencehasoccurredingeneral.WeÞndthatunalignedvotersaretheonlygroupwhoreacttotheoccurrenceofviolenteventsevenwhentheirleveloffearisunchanged.Wearguecorepartysupportersmustbetargetedbyviolencetoceasevoting,butunalignedvotersneednotbetargeted.Giventheimportanceoffearingviolenceinpredictingvotingbehavior,wealsoexplorethecorrelatesofthatfear.WeÞndpoliticalafÞliationpredictsanindividualÕsfearofelectoralvio-lence.Weshowanorderingexistsamongvotergroups:supportersofoppositiongroupsandsmallpoliticalpartiesreportthehighestlevelsoffear,followedbyunalignedvotersandvotersalignedwiththemajorityorrulingparty.Examiningtheimplicationsofthemodelsdiscussedinthispaperleadstoweakerconclusions.WeÞndsomeevidenceforthepredictionsmadebytheCVmodel:greaterviolenceisobservedinregionswithmoreunalignedvoters,whenusingeventcountmeasures.Theirmodel,though,isnottheonlystoryconsistentwithsuchapattern;itispossiblethatmoreunalignedvoterscouldcorrespondtogreatercompetitioninanelection,whichmayitselfleadtomoreviolence.Nonethe-less,thesemodelshavehighlightedanimportantdirectionforfutureresearch:uncoveringtherelationshipbetweenelectionviolenceandthedivisionoftheelectorate.Withbetterdataontheperpetratorsofelectionviolence,wecouldmorethoroughlytestbetweenthesetheoriesratherthanonlyexaminethevalidityoftheirassumptions.82However,thispaperhasmadestridesincircumventingdatalimitations.Wedemonstratethevalidityofusingsurveydatatomeasureelectionviolence.Theuseofsurveydataprovidessomeuniqueadvantages;wecanseehowfearofviolencevariesevenwhenindividualsareexposedtothesamesetofevents.Doingso,weareabletoconcludethatsomeelectoralviolenceislikelybeingtargetedatparticulargroups.Wefound,evenwhencontrollingforheterogeneityatrelativelysmallgeographicscales,townsorvillages,thatdifferentialfearsofviolencearerelatedtopoliticalafÞliation.Second,wefoundthatevenwhenthereisrelativelylittleelectoralviolence,asmeasuredbyeventdata,someindividualsreportafearofelectoralviolenceandreacttothatfear.TheseÞndingsareimportantbecauserecentpaperssuchasGoldsmith(2015)andLindberg(2005)havearguedviolencehashadaminimalimpactonmostelections.WearguethatcautionmustbetakenwiththeseconclusionsbecauseweÞndpeoplereacttolowlevelsofviolenceandthattheestimatedimpacts,whenusingmoregeneralmeasuresofpoliticalviolence,areunderestimated.Furthermore,ifthreatsofviolencearesufÞcienttogenerateafearofviolence,thentheimplicationsofmodelslikethatinEllmanandWantchekon(2000),whereoffequilibriumviolencecanaltertheelectionresults,needtobeseriouslyconsidered.Understandingwhyapoliticalpartyusesviolenceinanelectionisessentialtodesigningpoli-ciestoreduceitsoccurrence.If,asthetheorieswefocusedonsuggest,violenceisusedtostrategi-callymanipulatevoterturnoutinanelection,thekeytopreventingviolencemaylieintheelectoralsystemitself.Byincreasingtheneedforpoliticalpartiestoappealtovotersotherthantheircoresupporters,theviabilityofviolenceasanelectoralstrategycouldbesigniÞcantlyreduced.Regard-lessoftheunderlyingmechanisms,ourresultssuggestthatpreventionstrategiesshouldfocusonareaswithahighfractionofunalignedvotersandonmembersofrelativelysmallpoliticalparties.However,withoutidentifyingtheunderlyingcausesofelectionviolence,apermanentsolutionisunlikelytoemerge.Moreresearchisneededtotestbetweentheoriesofelectionviolence.Al-thoughthispapertakesimportantstepsinthatdirection,itscontributionidentiÞesreactionsto,andnotthesourcesof,electionviolence.83APPENDIX84Table2.1PoliticalAfÞliationSummaryIndividualCountryVariableObsMeanStd.Dev.MeanMinMaxTrustRuling261731.601.13TrustOpposition254531.221.06CarefulWhatSay264451.751.12SupportsRulingParty277130.340.470.340.000.71MajoirtyParty277130.360.480.370.190.71SecondParty277130.110.310.100.020.27OtherParty277130.090.290.090.010.32NoParty\Unaligned277130.410.490.410.180.63VotesMajorityParty277130.440.500.450.220.78SecondParty277130.130.340.130.020.31OtherParty277130.110.310.110.010.30NoParty\Unaligned277130.320.470.320.090.6485Table2.2AgreementBetweenMeasuresWouldVoteForSupportUnalignedMajoritySecondOtherPartyTotalUnaligned0.620.240.080.0711,407Majority0.070.850.070.029,985Second0.070.220.690.032,944OtherParty0.140.070.040.752,549Total8,29011,9953,6702,93026,885Eachcellisthefractionofindividualswhoreportedthatcombina-tion,dividedbythenumberofindividualswhoreportedsupportingthatgroup.ThedenominatorforeachcellistheÞnalcolumn.Table2.3IsPoliticalAfÞliationAccuratelyReported?VARIABLESSupportsRulingSupportsRulingUnalignedUndecidedCarefulwhatyousayRarely0.00-0.010.000.00(0.01)(0.01)(0.01)(0.01)Often-0.02-0.04***0.000.00(0.01)(0.01)(0.01)(0.01)Always-0.04***-0.05***0.03**0.02(0.01)(0.01)(0.01)(0.01)FixedEfffectsCountryDistrictDistrictDistrictObservations26,44526,44526,44526,445R-squared0.150.300.180.18Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1UnalignedcorrespondstoourprimarymeasuresofpoliticalafÞliation,askingifrespon-dentsfeelÒclosetoÓapoliticalparty.Thesecondmeasure,undecided,askswhatpartytheywouldvoteforinahypotheticalelection.Table2.4ViolenceSummary,RegionalLevelMeanS.D.MinMaxFearViolence3721.110.550.032.75Voted3720.720.130.181.00ViolentCompetition3721.430.520.042.89LogEventsBefore3500.230.530.003.18LogEventsAfter3500.370.690.003.6486Table2.5SimpleCorrelationsSurveyandEventViolenceMeasuresVARIABLESFearViolenceViolentCompetitionViolentCompetitionLogEvents0.02510.0838**0.0641Before(0.04)(0.03)(0.06)LogEvents0.0219After(0.05)FixedEffectsCountryCountryCountryObservations350350350R-squared0.5660.6810.682Standarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1Table2.6AdditionalComparisonofViolenceMeasuresVARIABLESLogEventsLogEventsViolentCompetition0.198**(0.09)FearViolence0.021(0.07)FixedEffectsCountryCountryObservations350350R-squared0.310.30Standarderrorsinparentheses***p<0.01,**p<0.05,*p<0.187Table2.7BaselineImpactofViolenceonVotingVotedVotedVotedFearViolence-0.038***-0.066**-0.0366***(0.01)(0.02)(0.01)FearViolence*0.0254CountryMean(0.02)N=26984R-squared0.0430.0430.14ViolentCompetition-0.013-0.0875***-0.003(0.01)(0.03)(0.01)ViolentComp*0.0515**CountryMean(0.02)N=26120R-squared0.0410.0420.14EventIndicator-0.0435***-0.0635**Not(0.01)(0.02)applicableLogEvents*0.047CountryMean(0.05)N=25249R-squared0.0420.0430.14FixedEffectsCountryCountryDistrictStandardErrorsClusteredbyRegion***p<0.01,**p<0.05,*p<0.1Column(1)and(2)differbyaninteractionwiththecountrylevelmeanoftheviolencemeasure.Column(3)includesÞxedeffectsforthe1820districts,Eventsaremeasuredatahigherlevel,andthereforedropsout.€Violenceherereferstothesamemeasureofviolencelisteddirectlyabove.Allmeasuresareindicators,eventsindicateatleastoneeventintheregion,thesurveymeasuresrefertothetoptwocategories.88Table2.8BaselineResults,HeterogenousRelationshipBetweenVotingandViolence(1)(2)(3)(4)(5)DependentVariable:VotedVotedVotedVotedVotedViolenceMeasure:FearViolenceHighFearViolentCompetitionHighViolenceEventIndicatorViolence-0.016***-0.029***-0.0050.001-0.022(0.00)(0.01)(0.00)(0.01)(0.01)UnalignedVoter-0.134***-0.130***-0.123***-0.127***-0.122***(0.01)(0.01)(0.02)(0.02)(0.01)Violence*UnalignedVoter-0.002-0.006-0.012*-0.025*-0.034*(0.01)(0.01)(0.01)(0.01)(0.02)SupportsMajority0.030***0.030***0.030***0.030***0.0287***(0.01)(0.01)(0.01)(0.01)(0.01)FixedEffectsCountryCountryCountryCountryCountryObservations2698426120265132576326120DependentVariable:VotedVotedVotedVotedVotedViolenceMeasure:FearViolenceHighFearViolentCompetitionHighViolenceEventIndicatorViolence-0.016***-0.033***-0.0010.007Not(0.00)(0.01)(0.00)(0.01)ApplicableUnalignedVoter-0.134***-0.127***-0.118***-0.114***-0.121***(0.01)(0.01)(0.02)(0.01)(0.01)Violence*UnalignedVoter0.001-0.000-0.011€-0.026*-0.027(0.01)(0.02)(0.01)(0.01)(0.02)SupportsMajority0.020*0.022**0.023**0.023**0.0230**(0.01)(0.01)(0.01)(0.01)(0.01)FixedEffectsDistrictDistrictDistrictDistrictDistrictObservations2698426120265132576326120StandardErrorsClusteredbyRegion***p<0.01,**p<0.05,*p<0.1€P-valueof.112.HighFearandHighViolencevariablesareindicatorvariablesforthetwohighestcategories898989Table2.9ExpandedSpecifcationsoftheimpactofViolenceonVoting(1)(2)(3)(4)(5)(6)(7)(8)(9)(10)VARIABLESVotedVotedVotedVotedVotedVotedVotedWontVoteVotedVotedFearViolence-0.012***-0.013***-0.012***-0.012***-0.011***-0.012***0.005*-0.015**-0.009(0.003)(0.003)(0.003)(0.003)(0.003)(0.003)(0.003)(0.007)(0.006)Alittle-0.009(0.009)Somewhat-0.034***(0.011)Alot-0.033***(0.010)age0.009***0.009***0.009***0.009***0.009***0.009***0.009***0.0004**0.009***(0.000)(0.000)(0.000)(0.000)(0.000)(0.000)(0.000)(0.000)(0.000)Female-0.021***-0.021***-0.017**-0.026***-0.020***-0.018**-0.018**0.010**-0.023***(0.007)(0.007)(0.007)(0.007)(0.007)(0.007)(0.007)(0.004)(0.007)Wealth-0.001-0.001-0.001-0.002-0.001-0.003-0.000-0.010***-0.001(0.004)(0.004)(0.004)(0.004)(0.004)(0.004)(0.004)(0.003)(0.004)Education0.005**0.005**0.004**0.0010.005**0.004*0.0030.005***0.005**(0.002)(0.002)(0.002)(0.002)(0.002)(0.002)(0.002)(0.002)(0.002)ViolentComp-0.001(0.004)LogEvents-0.028**(0.011)FearsCrime-0.003(0.003)DemocraticSatisfaction0.011***(0.004)VotenotSecret-0.010**(0.004)Voted-0.073***(0.008)Violence*Timesince-0.000-0.000election(0.000)(0.000)FixedEffectsRegionRegionRegionCountryRegionRegionRegionRegionRegionRegionObservations25,71125,71124,52424,45725,64323,82324,15025,71125,74924,629R-squared0.1610.1620.1630.1280.1610.1680.1620.10.0440.161Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1909090Table2.10PerceptionsofViolenceVARIABLESFearViolenceFearViolenceHighFearHighFearHighViolence(1)(2)(3)(4)(5)SupportsMajority-0.071***-0.119***-0.042***-0.049***(0.03)(0.03)(0.01)(0.01)SupportsSecond0.037-0.004-0.012-0.019(0.03)(0.04)(0.02)(0.02)UnalignedVoter0.0201-0.040-0.009-0.032**(0.03)(0.03)(0.01)(0.01)SupportsRuling-0.039***(0.01)SupportsOther0.006(0.01)FixedEffectsNoneDistrictDistrictDistrictDistrictObservations26,98426,98426,98426,98426,120R-squared0.0010.2710.2320.2320.259Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1TheommitedcategoryinallspeciÞcations,exceptColumn4,issupportsotherminorityparty.Incolumn4itisunalignedvoters,asallotherpartiesareputintotheSupportsOthercategory.91Table2.11DivisionoftheElectorateandViolence:UnalignedVotersVARIABLESLogEventsLogEventsViolentCompFearViolenceFraction0.531**0.528**0.145-0.063Unaligned€(0.23)(0.23)(0.13)(0.14)SupportsMajorityà-0.001-0.09***-0.10***(0.02)(0.02)(0.02)Unalignedà-0.00-0.06**-0.02(0.01)(0.02)(0.03)FixedEffectsCountryCountryCountryCountryObservations26,27326,27326,12026,984R-squared0.240.240.1830.165Robuststandarderrorsinparentheses***p<0.01,**p<0.05,*p<0.1€Thisistheregionalfractionofunalignedvoters.àTheseareindividualpoliticalafÞliationindicators.92REFERENCES93REFERENCESAfrobarometerData,[Countries:Benin,Botswana,BurkinaFaso,CapeVerde,Ghan,Kenya,Lesotho,Liberia,Madagascar,Malawi,Mali,Mozambique,Namibia,Nigeria,Senegal,SouthAfrica,Tanzania,Uganda,Zambia,Zimbabwe.](2007,2008,2009)availableathttp://www.afrobarometer.org.Bekoe,B.(2011)NiegeriaÕs2011elections:Bestrun,butmostviolent.UnitedStatesInsituteforPeace,.Blattman,C.(2009)Fromviolencetovoting:WarandpoliticalparticipationinUganda.AmericanPoliticalScienceReview,103,231Ð247.Brown,M.M.(2003)Democraticgovernance:Towardaframeworkforsustainablepeace.Univer-sityofIllisnois.UnpublishedManuscript.,Chandra,K.(2001).CumulativeÞndingsinthestudyofethnicpolitics.APSA-CPNewsletter,12,7Ð25.Chaturvedi,A.(2005).Riggingelectionswithviolence.PublicChoice,125,189Ð202.Cheibub,J.A.,Hays,J.,&Savun,B.(2012).ElectionsandcivilwarinAfrica.PublicChoice,125,189Ð202.Collier,P.&Vicente,P.(2012).Violence,bribery,andfraud:Thepoliticaleconomyofelectionsinsub-saharanAfrica.PublicChoice,153,117Ð147.Collier,P.&Vicente,P.(2014).Votesandviolence:EvidencefromaÞeldexperimentinNigeria.TheEconomicJournal,124,F327ÐF355.Ellman,M.&Wantchekon,L.(2000).Electoralcompetitionunderthethreatofpoliticalunrest.QuarterlyJournalofEconomics,115,499-531.Fischer(2002).Electoralconßictandviolence.IFESWhitePaper,1Goldsmith,A.(2015).ElectoralviolenceinAfricarevisited.TerrorismandPoliticalViolence,27.5,818Ð837.Hickman,J.(2009).Iselectoralviolenceeffective?EvidencefromSriLankaÕs2005presidentialelection.ContemporarySouthAsia,17,412Ð427.HumanRightsWatch.(2004).NigeriaÕs2003elections:Theunacknowledgedviolence.Retrievedfromhttps://www.hrw.org/sites/default/Þles/reports/nigeria0604.pdfHıglund,K.(2009).Electoralviolenceinconßict-riddensocieties:Concepts,causes,andconse-quences.TerrorismandPoliticalViolence,21,412Ð427.Lindberg,S.(2005).ConsequencesofelectoralsystemsinAfrica:Apreliminaryinquiry.ElectoralStudies,24,41Ð64.94Lindberg,S.(2005).ConsequencesofelectoralsystemsinAfrica:Apreliminaryinquiry.ElectoralStudies,24,41Ð64.Lindberg,S.(2006).DemocracyandelectionsinAfrica.JHUPress.Machado,F.,Scartascini,C.,&Tommasi,M.(2011).PoliticalinstitutionsandstreetprotestsinLatinAmerica.JournalofConßictResolution,55,340Ð365.Norris,P.(1997).Choosingelectoralsystems:Proportional,majoritarianandmixedsystems.InternationalPoliticalScienceReview,18,297Ð312.LatinAmericanPublicOpinionProject,Sleigson,M.,Kutzinski,V.,&Zephyr,P.(1997).Choos-ingelectoralsystems:Proportional,majoritarianandmixedsystems.VanderbuiltUniversity.Rapoport,D.&Weinberg,L.(2000).ElectionsandViolence.TerrorismandPoliticalViolence,12,15Ð50.Reilly,B.(2002).Electoralsystemsfordividedsocieties.JournalofDemocracy,13,156Ð170.Robinson,J.&Torvik,R.(2009).TherealswingvoterÕscurse.TheAmericanEconomicReview,99,310Ð315.Strauss,S.,&Taylor,C.(2009).Democratizationandelectoralviolenceinsub-saharanAfrica,1990Ð2007.InAPSA2009TorontoMeetingPaper.Suberu,R.(2007).NigerieaÕsmuddledelections.JournalofDemocracy,18,95Ð110.Walter,B.(2004).Doesconßictbegetconßict?Explainingrecurringcivilwar.JournalofPeaceResearch,41,371Ð388.95CHAPTER3ProÞlinginViolentElections3.1IntroductionRecenttheoreticalandempiricalresearchonelectionviolencehaspresentedseveralpotentialwaysinwhichviolencemaybeusedtoinßuencetheelectoralprocess.Akeydifferentiationbetweenemergingtheories,ashighlightedinWallsworth(2016),iswhetherviolenceistargeteddirectlyatoppositionsupportersorindirectlyatunalignedvoterswhoaremorelikelytovotefortheop-position.Wallsworth(2016)foundthatreactionstoviolenceareconsistentwithbothstrategies.Targetedviolenceisassociatedwithalowerlikelihoodofvoting,andunalignedvotersweretheonlygrouptoreacttoindirectlytargetedviolence.UsingdatafromabroadsampleoftwentyAfricancountries,neitherframeworkemergedasaclearlybetterÞtinthedata.Onewaytodis-tinguishwhichtheoryismoreviableinagivencountryistounravelhowsuccessfullyapotentialperpetratorofviolencecouldproÞlesupportersoftheopposition.ThispaperexaminestheviabilityofproÞlingbyidentifyingthecharacteristicsthatareassociatedwithfearingviolence.BeforewedelveintounderstandingifandhowproÞlingcouldbeoccurringintheseelectionsitisusefulandinterestingtoexaminewhofearsviolencethemostineachcountry.Infact,theoptimaltargetofviolenceisthetopicofrelatedtheoreticalworkonelectionviolenceinRobinsonandTorvik(2009).TodothisweusedatafromAfrobarometerÕsRound4surveys,whichaskrespondentshowmuchtheyfearbecomingavictimofelectionviolence.Mostoften,weÞndthatoppositionsupportersreportthehighestfearofviolence.Then,usingthisdataweexamineeasilyidentiÞablecharacteristicswhichcanbeusedtoinfersomeoneÕspoliticalafÞliationandinvestigatewhetherornotthesecharacteristicsarealsocorrelatedwitharespondentÕsfearofviolence.Un-surprisingly,weÞndethnicitytobethestrongestcharacteristicwhichpredictsbothanindividualÕs96politicalafÞliationandtheirfearofviolence.WhenweturntoexamininghowandifproÞlingoccurs,basedonotheridentiÞablecharacteris-tics,theresultsvaried.Socio-economiccharacteristicsandwhereanindividuallivespredictfearsofviolence,anindividualÕspoliticalafÞliation,orbothinsomebutnotallcountries.WearguecharacteristicswhicharecorrelatedwithbothanindividualÕsfearofviolenceandtheirpoliticalafÞliationmaybeparticularlyusefulinmatchingmodelsofelectionviolencetovariouscountries.Finally,weseektoidentifywhichcharacteristicsexplaintheobservedvariationinfearsofvi-olence,focusingonwhetherself-reportedorpredictedmeasuresofpoliticalafÞliationcanexplainmoreofthevariation.Theresultsfromthisanalysisshowthatbothpredictedandreportedmea-suresofpoliticalafÞliationaresigniÞcantpredictorsofanindividualÕsfearofviolence,althoughneitherexplainmuchofthevariance.WeÞndthatethnicityconsistentlyexplainsmoreofthevariationinfearacrossindividualsthanlocationorpoliticalafÞliation.UnderstandinghowproÞlingoccursinelectionsisusefulbeyondsimplyallowingustobuildmoreaccuratemodelsofelectionviolence.Havingabetterunderstandingofwhothemostlikelyvictimsofelectionviolenceareinagivencountrycanhelpauthoritiestaskedwithpreventingelectionviolenceidentifythemostat-riskgroupsandbetterallocatewhatmaybelimitedresourcestopreventviolence.973.2HowcantheviabilityofproÞlinghelpdistinguishbetweenmodelsofelectionviolence?Existingtheoreticalmodelsofelectionviolencevaryonafewkeydimensions,includingthere-searchquestion,whoperpetratesviolence,andhoworatwhomviolenceistargeted.UnderstandingtheviabilityofproÞlinghelpsthemostwhendistinguishingonthelastdimension:howoratwhomelectionviolenceistargeted.SomemodelslikethoseinEllmanandWantchekon(2000)orCollierandVicente(2012)donotassumetheneedtoidentifythevictimsofviolenceforittobeeffective.InCollierandVicente(2012),theyassumethatdifferenttypesofvoters,speciÞcallystrongandweakpartysupporters,reacttoviolencedifferently.Theyarguethatweakpartysupporters,whichonecanthinkofasunalignedvoterswholeantowardssupportingagivenparty,willstopvotingifanyviolenceoccurs.Ifthisisthecase,apartyneednottargetoppositionsupportersdirectly.Instead,theycangainvotesharebyusinguntargetedviolenceinareaswheretheyhaveaninitialadvantageamongstrongpartysupportersrelativetoweaksupporters.ModelsliketheonepresentedinEllmanandWantchekon(2000)alsodonotneedtheabilitytotargetviolenceforittobeaneffectivestrategytomanipulateelections.Intheirmodel,allindividualsareimpactedbyviolence.Theydoassumethatoppositionsupportersbearagreatercostwhenviolenceoccursthanthosewhocontrolthethreatofviolence.Althoughthismodelismoreapplicabletoviolencethatmaybeusedtoforciblyoverturnelectionsresultsratherthanmanipulateanelection,understandingwhetherornotproÞlingisaviableoptionmaystillhelpdetermineifthismodelofviolencecouldbeappliedtoaparticularcountry.Ontheotherhand,somemodelsexplicitlyexaminewhoshouldbetargetedbyviolenceinelectionsandunderwhatconditions.ThatistheprimaryquestioninvestigatedinRobinsonandTorvik(2009).Theydividetheelectorateintothreegroups.Twoofthegroupsleanideologicallytowardsthetwoparties,andthethirdgroupconsistsoftheswingvoters.Theiranalysisfocusesonwhenitismostviabletouseviolenceandwhichgroupthatviolenceshouldbedirectedat.98TheirinterestingconclusionisthatitcanbeÒmoreattractiveforanincumbenttodisenfranchisetheswingvotersthanthecoresupportersoftheopposition.ÓHowever,forsuchamodeltobeviable,itmustbepossibletoidentifythegroupofvotersyouplantotarget.ThelastmodelwediscussisChaturvedi(2005).Inhismodel,resourcescanbesplitamongtraditionalideologicalcampaigningandviolentcampaigning.Violenceistargeteddirectlyatop-positionsupporterstopreventthemfromvoting,whereasideologicalcampaigningismeanttopersuadeunalignedvoters.Thismodel,likethelastone,assumesnocollateraldamageoccurswhenviolenceisusedasacampaigntactic.Thepaperspresentingthesemodelssuccesfullymotivatethemwithexamples.However,wemaybeabletoimproveonthatbyempiricallyexaminingwhetherproÞlingseemstobepossibleinaparticularcountry.Ifthereisnowaytoidentifyoppositionsupporters,perhapsviolencelikethatproposedinEllmanandWantchekon(2000)maybethebestÞtbecausetheviolenceintheirmodelisuntargetedbutresultsinabeneÞtfortheperpetratingparty.Iflocationisthestrongestpredictor,amodelwhichbasestheeffectivenessofviolenceontherelativedistributionofvotertypesinanelectionlikeCollierandVicente(2012)becomesastrongcandidate.Finally,ifviolencecanbetargeteddirectlyatoppositionsupportersonthebasisofsomeobservablecharacteristicslikeethnicity,thenmodelslikeRobinsonandTorvik(2009)andChaturvedi(2005)becomemoreviable.3.3DataThedatausedinthispapercomesfromtheRound4Afrobarometersurveys.Thisdataisuniquelysuitedtoanalyzingwhoistargetedbyviolenceinelectionsasitasksquestionsaboutanindivid-ualÕsdemographiccharacteristics,politicalafÞliation,andfearofviolence.Inordertomeasureviolence,weuseaquestionwhichasks,ÒDuringelectioncampaignsinthiscountry,howmuchdoyoupersonallyfearbecomingavictimofpoliticalintimidationorviolence?ÓRespondentscouldchooseoneoffourresponseoptions:ÒAlotÓ,ÒSomewhatÓ,ÒAlittlebitÓ,orÒNotatallÓ.Thisisthentreatedasanorderedcategoricalvariableonazerotothreescale,with99threecorrespondingtofearingviolenceÒAlotÓ.Throughout,weassumethatanindividualÕsfearofviolenceisdirectlycorrelatedwiththeprobabilitythatanyonewithsimilarcharacteristicswouldbetargetedorvictimizedbyviolenceduringanelectionintheircountry.InordertomeasurepoliticalafÞliation,weuseaseriesofquestionswhichaskÞrst,ÒDoyoufeelclosetoanyparticularpoliticalparty?Ó.Then,iftherespondentansweredyes,theywereaskedÒWhichpartyisthat?ÓUsingthisquestionwedividevotersintofourgroups:thosewhosupportthelargestandsecondlargestpartiesinacountry,thosewhosupportanyotherparty,andthosewhodonotsupportanypoliticalparty.ThefractionofrespondentsfallingintoeachofthesefourcategoriesandtheirrespectivemeanfearsofviolencearepresentedinTable(3.1).Finally,wealsousesomequestionsonsocio-economicanddemographiccharacteristicsoftherespondents.Thisincludesquestionsaboutgender,age,livingconditions,wealth,andeducation.Forgenderweusedadichotomousvariable,calledfemale,whichequalsoneiftherespondentisfemale.AgeistherespondentÕsagetruncatedat80yearsoldbyAfrobarometertoprotecttherespondentÕsidentity.Livingconditions,wealth,andeducationareallcategoricalvariableswithhighercategoriescorrespondingtomorewealth,betterlivingconditions,orhavingcompletedahigherlevelofeducation.Themeansforthesevariables,overallandbycountry,arepresentedinTable(3.2).1003.4Results3.4.1Whofearsviolenceandwhere?InordertounderstandifproÞlingisoccurringorhowsomeonemaybetargetedwithviolenceduringelections,weÞrstneedtoknowwhoisbeingtargeted.Tomeasurethisweusethequestion,ÒHowmuchdoyoupersonallyfearbecomingavictimofviolenceduringelectionsinyourcountryÓWeexaminehowresponsesvaryacrossself-reportedpoliticalafÞliationbrokendownintofourcategories:supportsmajorityparty,supportsthesecondlargestparty,supportsanyotherparty,andsupportsnoparty.Table(3.1)presents,bycountry,thefractionofrespondentsfallingintoeachcategory,themeanresponsefortheirfearofviolence,andwhetherornottheirmeanresponseisstatisticallydifferentfromsupportersofthemajorityparty.Table(3.8)intheappendixexaminesthesamevariables,demeanedatthedistrictlevel.BothtablesreportthemeansasdifferencesfromthemajoritypartyÕsmeanfearofviolence,tohighlightwhichgroupsfearviolencethemostineachcountry.Inmostcountries,oppositionsupporters,eitherofthesecondlargestpartyorthosesupportingothersmallparties,reportthehighestfearsofviolence.Themostcommonexceptiontothisisafewcountriesinwhichindividualswhodonotreportsupportingapoliticalpartyhavethehighestreportedleveloffear.Onlyinonecountry,Zimbabwe,dothemajoritypartysupportersreportthegreatestfearofviolence.Thissuggeststhatthemostlikelyperpetratorofelectionviolenceisthemajorityparty.WeconcludethisbecauseinmostcountriesthemajoritypartyÕsmostrelevantoppositionreportsthehighestfearofviolence.Oneexplanationforthisisthattheymaybeabletoperpetrateviolencewithoutfearofreprisalandcouldbeabusingstatecontrolledresourcestodoso.Thereareseveralexceptionstothispattern.TheÞrstexceptionofnoteisinthecountrieswhereviolencedoesnotchangebasedonpoliticalafÞliation.Somecountriesinthiscategoryinclude:Malawi,Mali,Lesotho,andKenya.ForKenya,thismaybebecausebythesurveywas101takennotlongaftertheextremeviolencefollowingtheirturbulent2008electioncycle,whichmayhaveraisedfearsofviolencedramaticalllyforeveryone.Thenextthreeexceptions,SouthAfrica,Liberia,andBotswana,shareanotherpattern.Allthreeofthesecountrieshaveunalignedindividualsreportingthegreatestfearofviolence.Takingacloserlookatthecountrieswhereunalignedvotersreportthegreatestfearofviolenceilluminatessomeotherinterestingpatterns.Inthreeofthesecountries,themajoritypartyhasamuchhigherlevelofsupportthananyotherparty.InSouthAfricaandBotswanathemajoritypartyhasmorethanfourtimesthelevelofsupportthanthenextlargestparty,inLiberiaitismuchcloserbutthemajoritypartystillhasÞftypercentmoresupportersinthedatathanthesecondlargestparty.Thissuggestsanalternativewayinwhichthemajoritypartymaybeusingviolencetokeeppowerthroughtheelectioncycle:byoppressingpotentialcompetitiontopreventthemfromformingaviableoppositionparty.Withsuchadominantpluralityofthevote,noonepartyisstrongenoughtotakedownthemajorityparty,socreatinganatmosphereoffearthroughouttheelectionprocessanddeterringparticipationmaybethemosteffectivestrategytomanipulatetheelection.Inmostcountries,itappearsthatthemajoritypartyisthemostlikelyperpetratorofviolence,andthatthisviolenceisdirectedattheiropposition,whetherthathappenstobeanotherlargeparty,anumberofsmallparties,oroppressingthepopulationingeneral.However,thisdoesnottellusanythingabouthowtheyidentifyandtargettheoppositionwithviolence.ComparingTable(3.1)toTable(3.8)intheappendixbeginstoilluminateonepotentialwayinwhichthismaybeoccurring:usingviolenceinareasthataremorelikelytocontainoppositionsupporters.Innearlyhalfofthecountries,statisticalsigniÞcanceonthepredictivepowerofpoliticalafÞliationreducesordissapearsafteradjustingformeandifferencesofviolencebylocation.ToinvestigatethisfurtherwenowturntoexaminingwhatpredictspoliticalafÞliation,speciÞcallylookingatdemographiccharacteristicsthatmaybeeasytoinferbasedonanindividualÕsappearance.1023.4.2WhatpredictsanindividualÕspoliticalafÞliation?AspointedoutinCollierandVicente(2012),theeasieritistoinferoneÕspoliticalafÞliation,theeasieritistouseviolencetomanipulateanelection.However,whichcharacteristicspredictthatafÞliationandtheaccuracyofthosepredictionsvariesdramaticallybycountry.Wenowturntoan-alyzingwhichvisiblecharacteristicspredictanindividualÕspoliticalafÞliation.TodothisweranaseriesoflinearprobabilitymodelsbycountryforthetwoprimarycategoriesofpoliticalafÞliation:supportsmajority,andsupportsthelargestoppositionparty.1Thisanalysishastwopurposes:Þrst,itilluminateswhichfactorsarestrongpredictorsofpoliticalafÞliationineachcountry;second,wecanusetheknowledgeofwhatidentiÞablecharacteristicspredictoneÕspoliticalafÞliationinordertoexaminehowpartiesmaybeproÞlingtheiroppositionÕssupporterswithviolence.TheAfrobarometersurveycontainsalargenumberofvariableswhichcouldpotentiallybeusedtoinfersomeoneÕspoliticalafÞliation,includinghowwelltheytrustparticularpoliticalpar-ties,whatissuestheyfeelaremostsalient,andquestionsrelatedtosocio-economicstatus.Wefocussolelyonthosequestionswhichcouldreasonablybeassociatedwithoutwardlyvisiblechar-acteristics:gender,age,education,wealth,livingconditions,ethnicity,andwheretherespondentlives.Weomittedtwoothervariableswhichfallintothiscategory,religionandlanguage,becausetheywereoftenalmostperfectlycollinearwithethnicityorlocation.Thisledustorunaseriesoflinearprobabilitymodelswiththefollowingform:Pr[PAi=J]=&1Femalei+&2Agei+&3LivingConditions+&4Wealth+&5EDucation+&6Ethnicity+&7Location,(2.1)whereJiseitherbeingamemberofthemajority,orsecondlargestpoliticalparty.SpeciÞcationsarerunseperatelybyafÞliationandcountry.Femaleisadummyvariablewhichequalsoneif1Otherpartieswereomittedfortwomainreasons,ÞrstasidefromthetoptwocategoriesthereweretwofewobservationsinanygivenpartytoaccuratelypredicttheprobabilityofonebeingafÞliatedwiththatparty.Second,itseemedunlikelythatunalignedindividualswouldappeartobeacoherentgroup.103therespondentwasafemale,andethnicityandlocationareaseriesofdummyvariablesforallpotentialethnicitiesanddistrictsinthecountry.Anon-linearmodel,likeprobitorlogit,wouldordinarilybepreferredtoalinearprobabilitymodelincaseswhenyouwishtousethepredictedvalues.However,thelargenumberofobserva-tions,insomecountriesashighasseventy-Þvepercent,whichhavetheoutcomevariableperfectlypredictedbyeitherlocationorethnicityeliminatesthisastheoptimalestimationtechnique.Tables(3.3)and(3.4)reporttherawcoefÞcientsforthedemographiccharacteristics,alongwiththeirsigniÞcancelevelsincolumns(4)-(8).Columns(9)and(10)presentthep-valuesforajointF-testofthesigniÞcanceoflocationandethnicitydummies.WhatpredictsanindividualÕspoliticalafÞliationvariesdramaticallybycountry.Female,ageandeducationaresigniÞcantpredictorsinaboutÞftypercentofcountriesinthesample.EthnicityisastatisticallysigniÞcantpredictorofafÞliationinalltwentycountries,andlocationmatters,independentlyofethnicity,inabouthalfofthesample.Althoughnotsurprising,thislendssupporttotheideathatpoliticalafÞliationcanbeinferredfromvisiblecharacteristics.Ethnicity,location,gender,andageareallcharacteristicsonecangleanwithoutsigniÞcanteffortortheuseofinformants.Socio-economiccharacteristicsarealsolikelytoberelativelyeasytoidentify.Thenextquestionthenistwo-fold:doesanindividualÕspredictedpoliticalafÞliationcorrelatewithviolencebetterthantheirreportedafÞliation,andaresigniÞcantpredictorsofanindividualÕspoliticalafÞliationmorelikelytobecorrelatedwiththeirfearofviolence?Thesequestionsaddressacommonunderlyingidea:ifviolenceisdirectedatparticulargroupsofpeople,lookinglikeyoubelongtothatgroupmaybemoredetrimentalthanactuallybelongingtothatgroup.Ifthishypothesisissupportedbythedata,wemaygainabetterunderstandingofwhatcharacteristicsarebeingusedtotargetpeoplewithviolence.3.4.3Howareindividualstargetedwithviolence?TogetabetterideaofhowproÞlingmayoccur,weexaminehowthepredictedvaluesofpoliticalafÞliation,speciÞcallytheprobabilityonesupportsthemajoritypartyortheprimaryopposition104party,correlatewithfearsofviolence.Figure(3.1)presentsthedifferenceinmeanreportedfearsofviolencebycountrybetweenmajorityandoppositionsupportersforbothself-reportedandpre-dictedpoliticalafÞliation.Foreachcountry,theleftmostbarusestheself-reportedmeasureofpoliticalafÞliationandthebarontherightusespredictedafÞliationfromthemodelsdescribedintheprevioussection.Thenumbersaboveorbeloweachbararethep-valuesforthedifference,whilethedifferenceitselfisrepresentedbytheheightofthebar.Wereportthedifference,ratherthanaseriesofabsolutemeans,becausethedifferencehelpsilluminatewhichgroupfearsvio-lencethemostineachcountry.Apositivevalueindicatesthatmajoritymembersreportagreaterfearofviolence,whereasanegativevalueindicatesoppositionsupportersreportagreaterleveloffear.ThisÞguredemonstratesthatthepredictedvaluesofpoliticalafÞliationproduceresultswhichmirrortheself-reportedvalues,withtheexceptionofthelargermagnitudesinthedifference.Theselargermagnitudesareatleastinpartexplainedbythefactthatthesevaluescomefromaregressionframeworkwhereavalueof1actuallyindicatessomeonereportedthatafÞliationforthereportedmeasure,whereastheaveragepredictedvaluewouldrangefrom.3to.5dependingontheirafÞliation,resultinginalargermagnitudewhencomparingthecoefÞcients.PredictedafÞliationisessentiallyaweightedaverageofdemographiccharacteristics.WewouldliketobeabletoidentifywhichcharacteristicsareusedinwhichcountrytoproÞlein-dividuals.Todothis,werunanadditionalsetofregressionsandcomparetheresultstothemodelswhichwereusedtopredictpoliticalafÞliation.VariableswhicharesigniÞcantpredictorsofbothanindividualÕsfearofviolenceandtheirpoliticalafÞliationareofparticularinterestbecausetheyarethemostlikelycharacteristicstobeusedtotargetindividualswithviolence.Table(3.5)presentstheresultsoftheregressionsanalyzinghowthesecharacteristicsareafÞliatedwithfearsofvio-lenceandTable(3.6)compareswhichpredictorsaresigniÞcantforpredictingpoliticalafÞliation,fearingviolence,orboth.Perhapsmostinteresting,andleastsurprising,isthesigniÞcantdegreeofoverlappingpredict-ingpowerforethnicity.EthnicityissigniÞcantlyassociatedwithviolenceinallbutthreecountriesandisasigniÞcantpredictorofpoliticalafÞliationinallcountries.Thisimpliesthatethnicityis105oneparticularlystrongcharacteristicwhichcouldbeusedtoinferpoliticalafÞliationandtargetanindividualwithviolence.AnotherÞndingworthnotingisthatlocationsigniÞcantlypredictsanindividualÕspoliticalafÞliationfarmoreoftenthanitdoestheirfearofviolence.Takentogether,thissuggeststhatdirecttargetingofviolencemaybethemoreoftusedstrategy,evenifindirecttargetingisapotentiallyviabletactic.Thesocio-economiccharacteristicsofgender,age,livingconditions,wealth,andeducation,havemuchmorevariedpredictivepower.GenderandlivingconditionssigniÞcantlypredictbothinÞvecountries,withtheothersoftenpredictingeitherpoliticalafÞliationorfearofviolencebutnotboth.Insomerespectsthismakessensetherearelikelytobepeopleofallgenders,ages,andvariedsocio-economicstatusinmostpoliticalpartiessothesecharacteristicsarelesslikelytobeusefulindifferentiatingbetweensupportersofthevariouspoliticalparties.Ontheotherhand,whenthesearesigniÞcantpredictorsofboth,itmayprovideusefulinformationastowhatexactlydividestheseparties.TakeforexampleNigeria,wherewealthisastrongpredictorofwhichpartyanindividualsupportsandtheirfearofviolence.Thismaybebecausecontrolofoilrevenuesisamajorpoliticalissue.Furthermore,itisunsurprisingtoseelocationisasigniÞcantpredictorofpoliticalafÞliationinNigeriagiventheircountryÕstraditionofalternatingrulebetweensomeonefromthenorthandsomeonefromthesouth.Finally,inTable(3.7)weanalyzewhatbestexplainsthevariationinindividualfearsofvio-lence.ThistablereportstheR-squaredvaluesfromcountrylevelregressionsincludingafullsetofdummiesforlocationinColumn(1),ethnicityinColumn(2),andthereportedorpredictedvaluesofsupportformajorityorlargestoppositionpartyinColumns(3)and(4).AllColumnsusetheexactsamesetofobservationsineachcountry,andtheR-squaredvaluesaredirectlycomparableinColumns(3)and(4)becauseeachisaregressionwithtwovariablesandanintercept.However,Columns(1)and(2)arenotaseasilycompared,thenumberofethnicitiesanddistrictsineachofthecountrieshassubstantialvariation.Again,weÞndthatethnicitydoesthebestjobinexplainingthevariationinfearingviolence.Weunfortunatelythencannotdetermineifthishastodowithexistingethnictensions,orthefactthatethnicityisastrongpredictorofpoliticalafÞliation.106Overall,theseresultssupporttheideathattheeasieroneÕspoliticalafÞliationistoidentify,themoreviableastrategyviolencebecomes.Furthermore,itisclearthatnosinglemodelisgoingtosuccessfullypredictthewhen,where,andwhyofelectionviolenceinallcountries.However,theintuitionofexistingmodelsdoesprovidealotofusefulcorrelatestoinvestigateandraisestheideathatproÞlingmaybeonewayinwhichindividualsaretargetedwithviolencethroughouttheelectoralprocess.1073.5ConclusionAnalysissuggeststhatinthemajorityofcountriesexamined,whenviolenceisaconcernintheelectorate,itseemstobeprimarilyperpetratedbythemajorityparty.Furthermore,theviolenceseemstobemostoftendirectlytargetedatoppositionsupporters.ThisisaninterestingadditiontotheÞndinginWallsworth(2016)thatbothtypesofviolenceappeartobeeffectiveatmanipulatingelections.TheÞndingsinthispapersuggestthatthetypeofviolenceusedcoulddependonwhatcharacteristicscanbeusedtoinferanindividualÕspoliticalafÞliation.WhenweturntoanalyzinganindividualÕspoliticalafÞliation,weÞndthatethnicityisoneofitsstrongestandmostconsistentpredictors.EthnicityalsopredictsanindividualÕsfearofviolence.However,wedonothavetheabilitytodisentangleifthereasonethnicitypredictsanindividualÕsfearofviolenceisactuallytheresultofproÞling.Finally,weusethesepredictorsofpoliticalafÞliation,alongwiththepredictedvalueitself,andexaminehowtheycorrelatewithanindividualÕsfearofviolence.ThepredictedvaluesofpoliticalafÞliationtendtomimictheresultsusingtheactualvalues.WearguethatthissuggeststhatvisiblepredictorsmaybeusefulforidentifyinganindividualÕspoliticalafÞliation.Thisevidenceleadsustoconcludethreethings.First,therulingpartygenerallyappearstobetheperpetratorofviolence.Thereareplentyofinterestingandimportantexceptionstothis,butthisleadsustosuggestthatinternationalauthorities,andindependentagencieswithingovernments,arewherereformtopreventviolencemustoriginatefrom.Itistooeasyforthosewhocontrolthestatesecurityapparatusandjudicialsystemstocommitviolencewithoutfearofrepercussionandthistrendneedstostopifwehopetopreventviolenceinelections.Second,violencetendstobetargetedatthemajoritypartyÕsmostrelevantopposition.Atleastobservationally,thecountriesinwhichunalignedvotersfearviolencethemostareallcountrieswherethereisnotanoppositionpartystrongenoughtochallengethemajorityparty.Furtherinvestigationwouldbeneededtovalidatethisclaim,butthepatterniscertainlyinteresting.TheÞnalconclusionwedrawisthatthewayinwhichviolenceistargetedisstronglycorrelatedwiththeeasiestwaystoinferpolitical108afÞliation,bethatlocation,ethnicity,orotherdemographiccharacteristics.Richerdataonsocio-demographictraitsofindividualsinthesecountriesalongwiththemake-upofthepoliticalpartieswithinthemcouldstrengthenthisclaimsigniÞcantly.TheseÞndingsarealsousefultovalidateassumptionsinexistingtheory,andhelptodeterminewhichtheoriesaremostapplicabletowhichsituations.Allfourofthetheoriesdiscussedinthispapercouldbeappliedtosome,butnotall,countriesinthissample.InÞfteenofthetwentycountries,ethnicitymayenableviolencetobedirectlytargeted.Inasmallsampleofcountries,indirecttargetingofviolencemaybeoccurringonthebasisoflocation.Understandinghowitispossibletouseviolenceinagivencountrycanhelpresearcherstodeveloptheoriescateredtoaparticularsituation,andtochoosewhichtheorytobasetheirdecisionsupon.Overall,itisclearviolencehasbecomeanunfortunatelyintegralpartofelectionsinAfrica.Withthosewhoperpetrateviolencemostoftenbeingimmunetopunishment,signiÞcantreformsneedtobemadefromtheoutsideandindependentenforcementagenciesshouldbecreatedwithincountrieswhereelectionviolenceisproblematic.Furthermore,encouragingpartiestoformalonglesscontentiousdivides,likeethnicity,couldalsomaketheuseofviolencelesseffectivebymak-ingithardertoinferanindividualÕspoliticalafÞliation.Lookingatviolencethroughthelensofeconomictheoryprovidesusefultoolstoattempttopredictandpreventitsoccurence.Theanal-ysisinthispapersuggeststhatthetheoriespresentedinChaturvedi(2005),CollierandVicente(2012),andRobinsonandTorvik(2009),andotherpapersprovideasolidstructuretobuildonwhenattemptingtounderstandelectoralviolence.109APPENDIX110Figure3.1ComparingPoliticalafÞliationandFearofViolence.03.77.04.010.00.190.00.41.080.00.87.85.50.00.28.170.00.2.02.02.01.2.16.160.00.1.24.1.04.13.23.15.49.010.00.690.000.00.010.00-3-2-101Fear of Violence (Majority - Opposition)BeninBotswanaBurkina FasoCape VerdeGhanaKenyaLesothoLiberiaMadagascarMalawiMaliMozambiqueNamibiaNigeriaSenegalSouth AfricaTanzaniaUgandaZambiaZimbabwe42CountryPredicted AffiliationActual AffiliationNumber above\below bar is p-value for differenceFear of Violence: Predicted versus Actual Affiliation111Table3.1MeanViolenceandPoliticalAfÞliationBreakdownbyCountryPartyAfÞliation:MajoritySecondOtherNoneNMeanFractionMeanFractionMeanFractionMeanFractionMeanCountryViolenceSupportingViolenceSupportingViolenceOtherViolenceSupportingViolenceBenin12000.580.190.750.04-0.35**0.140.25***0.64-0.25***Botswana12000.270.550.230.130.090.09-0.060.230.1**Burkina_Faso12001.010.381.030.020.370.080.38***0.52-0.10Cape_Verde12640.500.280.610.27-0.110.02-0.070.43-0.2***Ghana12000.910.360.780.190.33***0.03-0.170.430.15**Kenya11041.820.411.830.14-0.150.11-0.080.330.05Lesotho12000.980.330.930.110.150.06-0.040.500.08Liberia12000.950.200.680.130.18*0.12-0.010.560.37***Madagascar13500.850.240.810.030.36**0.060.180.670.03Malawi12001.010.480.980.090.21*0.08-0.050.350.04Mali12320.980.240.930.100.160.32-0.020.340.05Mozambique12000.880.630.870.030.240.01-0.150.330.00Namibia12001.130.441.050.060.080.120.23**0.380.09Nigeria23241.430.221.320.13-0.2**0.12-0.14*0.520.25***Senegal12000.830.310.740.070.39***0.150.21**0.480.06South_Africa24000.970.390.930.06-0.060.08-0.040.480.08*Tanzania12080.930.710.850.050.81***0.04-0.270.200.25***Uganda24311.590.381.470.170.43***0.08-0.17*0.370.11**Zambia12001.090.220.950.190.33***0.14-0.32***0.460.21**Zimbabwe12002.410.372.510.08-0.52***0.01-0.540.55-0.12****p<0.01,**p<0.05,*p<0.1Thep-valuesindicatewhetherthemeanisstatisticallydifferentfromthemeanfearofviolencereportedbyrespondentsafÞliatedwiththemajorityparty.MeanviolenceforSecond,Other,andNonereportedasdifferencefrommeanofMajorityParty.112112112Table3.2MeansofDemograhicData:OverallandbyCountryFemaleAgeLivingCondWealthEducationOverallMean0.5036.302.632.783.15OverallMin018110OverallMax180559Benin0.5035.382.372.742.08Botswana0.5040.152.502.663.38BurkinaFaso0.5036.522.752.971.49CapeVerde0.5037.782.832.893.20Ghana0.5038.892.753.022.86Kenya0.5035.192.132.503.77Lesotho0.5041.302.062.412.85Liberia0.5035.782.703.093.02Madagascar0.5039.502.702.752.88Malawi0.5035.452.692.842.44Mali0.5039.102.462.741.20Mozambique0.5030.642.892.833.07Namibia0.5034.732.932.914.04Nigeria0.5031.303.173.104.40Senegal0.5138.872.062.572.00SouthAfrica0.5037.772.772.954.31Tanzania0.5037.512.332.393.06Uganda0.5033.712.562.513.37Zambia0.5035.082.512.763.40Zimbabwe0.5036.532.742.853.83113Table3.3LinearProbabilityModelsPredictingProbabilityofSupportingMajorityParty(1)(2)(3)(4)(5)(6)(7)(8)(9)(10)CountryObsFractionFemaleAgeLivingCondWealthEducationEthnicity€Location€Benin11260.19-0.06**-0.010.010.010.010***0.33Botswana11700.550.06**0.02**0.04**0.00-0.02**0***0.48Burkina_Faso10620.38-0.07**0.000.020.04**-0.03***0.00***0.52Cape_Verde10920.280.000.03**0.05***0.000.02**0.00***0.14Ghana11000.36-0.03-0.010.04***0.00-0.010.00***0.06*Kenya10600.41-0.05*-0.010.02-0.020.000.00***0.00***Lesotho11490.330.05*0.010.03**0.00-0.03***0.00***0.31Liberia11780.200.030.04***0.000.02*0.01**0.230.11Madagascar12950.24-0.06**0.000.010.010.010***0***Malawi11130.480.03-0.010.05***0.010.000.00***0.08*Mali11970.24-0.09***0.000.010.04**0.000.00***0.25Mozambique10560.630.010.03**0.020.020.03**0.01***0.42Namibia11850.44-0.010.02**0.000.02-0.03***0.00***0.00***Nigeria22190.22-0.07***0.01-0.010.02*0.010.00***0.00***Senegal11060.310.000.04***0.00-0.04*-0.01*0.01***0.5South_Africa22900.39-0.020.000.000.01-0.02**0.00***0.00***Tanzania11800.710.1***0.05***0.010.02-0.02*0.00***0.31Uganda23640.380.010.03***0.03***0.01-0.010.00***0.00***Zambia11330.22-0.040.02*0.02*0.00-0.010.00***0.13Zimbabwe11600.37-0.040.010.00-0.020.03***0.00***0.54***p<0.01,**p<0.05,*p<0.1€Thesecolumnsreportthep-valuesfromaF-testofjointsigniÞcanceforalltherelevantdummies,foreitherethnicityorlocation.Observationsisthenumberofobservationsforthecountry,andusedobservationsisthenumberofobser-vationswhichwerekeptforestimation.Ageisindecades.114114114Table3.4LinearProbabilityModelsPredictingProbabilityofSupportingLargestOppositionParty(1)(2)(3)(4)(5)(6)(7)(8)(9)(10)CountryObsFractionFemaleAgeLivingCondWealthEducationEthnicity€Location€Benin11260.040.000.00-0.02**0.010.01*0.00***1Botswana11700.130.000.00-0.010.010.000.00***0.83Burkina_Faso10620.02-0.010.000.010.000.01**0.00***0.28Cape_Verde10920.270.000.00-0.04**0.00-0.02**0.00***0.9Ghana11000.19-0.07***0.00-0.03***-0.010.000.16*0.02**Kenya10600.14-0.010.01-0.010.020.000.02**0***Lesotho11490.11-0.04**-0.01*-0.010.010.01*0.05**0.77Liberia11780.13-0.05**-0.04***0.010.000.010.04**0***Madagascar12950.03-0.010.000.010.000.000.00***0.98Malawi11130.09-0.02-0.02**-0.02**-0.010.000.00***0***Mali11970.100.01-0.01-0.02**-0.02*0.000.00***0.03**Mozambique10560.03-0.010.00-0.010.000.000.270.97Namibia11850.06-0.010.000.01-0.010.02***0.761Nigeria22190.13-0.03**0.0100.02***0.000.00***0***Senegal11060.070.020.00-0.01-0.010.010.260.08*South_Africa22900.06-0.010.000.000.000.000.00***0.00***Tanzania11800.05-0.04***-0.01***-0.01**-0.01**0.000.00***0.31Uganda23640.17-0.09***-0.03***-0.02**0.000.02***0.00***0.00***Zambia11330.19-0.08***-0.03***-0.01-0.02**0.010.00***0.00***Zimbabwe11600.080.000.000.00-0.01-0.01*0.00***0.06****p<0.01,**p<0.05,*p<0.1€Thesecolumnsreportthep-valuesfromaF-testofjointsigniÞcanceforalltherelevantdummies,foreitherethnicityorlocation.Observationsisthenumberofobservationsforthecountry,andusedobservationsisthenumberofobservationswhichwerekeptforestimation.Ageisindecades.115115115Table3.5DemographicTraitsandFearofViolenceCountryFemaleAgeLivingCondWealthEducationEthnicity€Location€Benin0.040-0.020.010.05***0.02**0.09*Botswana0.08*0-0.04*-0.0300.03**0.15BurkinaFaso0.18**0.01*-0.04-0.030.040.01**0.48CapeVerde0.040-0.020.010.010***0.03**Ghana0.0900.02-0.1***-0.030***0.07*Kenya0.05-0.01**-0.04-0.01-0.04*0***0.2Lesotho0.13*0-0.050.08**-0.04*0***0.27Liberia0.020-0.03-0.030.020***0.86Madagascar0.080-0.08**-0.030.010***0.58Malawi0.02-0.01***-0.1***-0.03-0.030***0.74Mali0.0800.06*-0.11**-0.04*0.110.37Mozambique0.05-0.01**-0.03-0.020.020.240.52Namibia0.0100-0.03-0.020***0.02**Nigeria0.16***00.03-0.11***-0.03**0***0.16Senegal0.02-0.01**-0.03-0.040.030***0.68SouthAfrica0.08*00.04*-0.06***-0.020***0.41Tanzania0.050-0.12***0.02-0.010***0***Uganda0.060-0.06**0.010.04***0***0.07*Zambia0.18***0-0.06**0.040.020.160.32Zimbabwe0.06-0.01**-0.01-0.050.010.06*0.18***p<0.01,**p<0.05,*p<0.1€Thesecolumnsreportthep-valuesfromaF-testofjointsigniÞcance.116Table3.6ComparingPredictorsofViolenceandPoliticalAfÞliationCountryFemaleAgeLivingCondWealthEDucationEthnicityLocationBeninAfÞliationxAfÞliationxBothBothViolenceBotswanaBothAfÞliationBothxAfÞliationBothxBurkinaFasoBothViolencexAfÞliationAfÞliationBothxCapeVerdexAfÞliationAfÞliationxAfÞliationBothViolenceGhanaAfÞliationxAfÞliationViolencexBothBothKenyaAfÞliationViolencexxViolenceBothAfÞliationLesothoBothAfÞliationAfÞliationViolenceBothBothxLiberiaAfÞliationAfÞliationxAfÞliationAfÞliationBothAfÞliationMadagascarAfÞliationxViolencexxBothAfÞliationMalawixBothBothxxBothAfÞliationMaliAfÞliationxBothBothViolenceAfÞliationAfÞliationMozambiquexBothxxAfÞliationAfÞliationxNamibiaxAfÞliationxxAfÞliationBothBothNigeriaBothxxBothViolenceBothAfÞliationSenegalxBothxAfÞliationAfÞliationBothAfÞliationSouthAfricaViolencexViolenceViolenceAfÞliationBothAfÞliationTanzaniaAfÞliationAfÞliationBothAfÞliationAfÞliationBothViolenceUgandaAfÞliationAfÞliationBothxBothBothBothZambiaBothAfÞliationBothAfÞliationxAfÞliationAfÞliationZimbabwexViolencexxAfÞliationBothAfÞliationTotalsBoth53523154AfÞliation8853738Violence1323322Neither66812706BothindicatesthevariablewasasigniÞcantpredictorofbeingamemberofeitherpoliticalparty,andofviolence.ViolenceandAfÞliationindicateitwasasigniÞcantpredictorintherespectivecategory,xfornone.AtenpercentsigniÞcancelevelwasassumedthroughoutthetable.117Table3.7VarianceinFearofViolenceExplainedPoliticalAfÞliationCountryDistanceEthnicityPredictedReportedBenin0.030.090.000.01Botswana0.040.140.010.00Burkina_Faso0.030.100.010.00Cape_Verde0.020.230.030.01Ghana0.050.160.010.01Kenya0.080.140.010.00Lesotho0.040.050.010.00Liberia0.030.120.000.02Madagascar0.020.130.000.00Malawi0.010.070.010.00Mali0.020.130.000.00Mozambique0.060.110.000.00Namibia0.030.140.010.01Nigeria0.030.250.020.02Senegal0.030.080.000.01South_Africa0.010.180.000.00Tanzania0.160.280.060.03Uganda0.070.130.000.02Zambia0.050.080.010.01Zimbabwe0.050.100.010.02EachcellreportstheR-squaredvalueofaregressionwiththerespectivesetofdummyvariables.PredictedandReportedaf-ÞliationreporttheR-squaredvaluefromalinearregressionwitheitherthereportedorpredictedlikelihoodofbeingafÞl-iatedwiththemajorityorlargestoppositionparty.118Table3.8MeanviolencebyAfÞliation,DemeanedbyDistrictCountryNOverallMeanDifferencesFromMajoritySecondOtherNoneBenin12000.58-0.170.15*-0.2***Botswana12000.270.08-0.110.06Burkina_Faso12001.010.260.27**-0.03Cape_Verde12640.50-0.030.13-0.05Ghana12000.910.18**-0.37**0.18***Kenya11041.82-0.040.18*0.10Lesotho12000.980.020.030.01Liberia12000.950.18*0.010.35***Madagascar13500.850.240.160.02Malawi12001.010.140.05-0.02Mali12320.980.160.03-0.03Mozambique12000.880.25-0.150.01Namibia12001.130.080.15*0.04Nigeria23241.43-0.03-0.100.18***Senegal12000.830.41***0.19**0.05South_Africa24000.970.02-0.010.07*Tanzania12080.930.080.010.17**Uganda24311.590.45***0.100.13**Zambia12001.090.18*-0.23**0.16*Zimbabwe12002.41-0.4***-0.30-0.08Thep-valuesindicatewhetherthemeanisstatisticallydifferentfromthemeanfearofviolencereportedbyrespondentsafÞliatedwiththemajor-ityparty.Allvaluesarereportedasaveragedeviationsfromthedistrictmeanfortherelevantparty.119REFERENCES120REFERENCESChaturvedi,A.(2005).Riggingelectionswithviolence.PublicChoice,125,189Ð202.Collier,P.&Vicente,P.(2012).Violence,bribery,andfraud:Thepoliticaleconomyofelectionsinsub-saharanAfrica.PublicChoice,153,117Ð147.Ellman,M.&Wantchekon,L.(2000).Electoralcompetitionunderthethreatofpoliticalunrest.QuarterlyJournalofEconomics,115,499-531.Robinson,J.&Torvik,R.(2009).TherealswingvoterÕscurse.TheAmericanEconomicReview,99,310Ð315.Wallsworth,G.(2016).Electoralviolence:Anempiricalexaminationofexistingtheories.Unpub-lishedManuscript,121