Spatial Competition in the Network Television Indu

SpatialCompetitionintheNetworkTelevisionIndustryRonaldGoettlerGraduateSchoolo[e-mail][e-mail]17,2001WewouldliketothankMikeEisenberg,GregKasparian,andDavidPoltrackofCBSfortheirhelpinobtainingthedataforthisstudy.WealsoacknowledgeandaregratefulforhelpfulcommentsfromSteveBerry,MosheBuchinsky,ZviEckstein,PhillipLeslie,ArielPakes,JohnRust,threeanonymousreferees,andparticipantsatvariousseminars.1AbstractWepresentanempiricalstudyofspatialcompetitionandamethodologytoestimatede-mandforproductswithunobservablecharacteristics.Usingpaneldata,weestimateadiscretechoicemodelwithlatentproductattributesandunobservedheterogenousconsumerpreferences.Ourapplicationofthemethodologytothenetworktelevisionindustryyieldsestimatesthatareconsistentwithexperts'views.Givenourestimates,wecomputeNashequilibriaofaproductlocationgame,andndthatrms'observedstrategies(suchasthedegreeofproductdieren-tiation)aregenerallyoptimal.Discrepanciesbetweenactualandoptimalstrategiesreectthenetworks'adherencetorulesofthumb,andpossibly,boundedrationalitybehavior.Keywords:Spatialcompetition,discretechoice,paneldata,latentvariables,consumerheterogeneity,maximumsimulatedlikelihood,networktelevision,statedependence.JEL:C51,D12,D40,L10,L82.21IntroductionMostempiricalindustrystudiesfocusonpricecompetition,conditionalonagivensetofproductcharacteristics.Competitioninproductspaceisalsoimportant.Forexample,ininformationindus-tries,suchasmediaandentertainment,thestrategicchoicesareprimarilynonmonetaryproductcharacteristics.Analysisofcompetitionintheseindustriesisoftencomplicatedbythepresenceofunobservableordiculttomeasureproductcharacteristics.Forexample,therelevantattributesoftelevisionshowsarenotobvious.Wepresentanempiricalstudyofspatialcompetitionandamethodologicalapproachtoes-timatingproductcharacteristicsandconsumerpreferencesforproductswhosecharacteristicsareunobservableordiculttomeasure.Weusepaneldataonconsumers'choicestoidentify(a)theattributespaceoverwhichrmscompete,(b)productlocationsontheseattributes,and(c)thedistributionofconsumerpreferences.Theeconometricmethodisappliedtoanalyzecompetitionforviewersinthetelevisionindustry.Theestimatedattributespaceandproductlocationsareconsistentwithexperts'viewsofthisindustry.Forexample,oneoftheattributesrepresentsthedegreeofrealismineachshow.Theestimatedproductlocationsrevealthatrmsusecounter-programming(i.e.,dierentiatedproductsineachtimeslot)andhomogeneousprogramming(i.e.,similarproductsthrougheachnight).Thesestrategiesareconrmedtobeoptimalinanequilib-riumanalysisgivenourestimateddemand.WecomputeaNashequilibriumthatsuggestsrmscanimprovetheirweeklyratingsbyabout10percentbyincreasingbothcounterprogrammingandhomogeneity.Therstpartofthepaper(Sections2{4)presentstheeconometricmethodology,whilethelatterpart(Sections5and6)appliesittothetelevisionindustryandanalyzesspatialcompetition.InSection2wedescribethediscretechoicemodelofviewerbehavior.Consumerutilityisspeciedtohaveanidealpointstructure,withutilitydecreasinginthedistancebetweentheconsumer'smostpreferredleveloftheattributesandaproduct'sattributes.1Theidenticationofshowcharacteristicsandconsumerpreferencesisnotobvious.Wedonotobservetheattributespacerelevanttoviewers'choices,northeattributelevelsforeachshow,northeidealpointofeachconsumer.2Wedoobservepaneldataofeachviewer'schoiceineachperiod,asdescribedinSection3.Thesechoicehistories,evenwithasingleairingofeachshow,providethecovarianceofchoicesthatidentiesthecovariancematrixofutilityfortheproducts.Forexample,twoproductsconsumedbymanyofthesameindividualswillhaveapositivecovarianceofutility.AsdiscussedinSection4.4,weusethelatentattributespacetoparameterizethecovariancematrixofutilitysuchthatproductswithpositivecovariancetermsarelocatedneareachother.Notethatnomeaning1Anderson,dePalma,andThisse(1992)refertotheidealpointmodelastheaddressmodel.2UsingterminologyfromHeckman(1981a),ourmodelisadiscretechoicemodelwithbothstructuralstatedependenceandacomponents-of-variancestructurefortheunobservedcomponentofutility.3isassignedaprioritothedimensionsoftheattributespace.Assuch,interpretingeachestimateddimensionisimportanttounderstandingviewerbehaviorandproductdierentiation.Themodelisestimatedusingmaximumsimulatedlikelihood.Furthermore,inordertoreducesimulationerrorweusebothimportancesamplingandlow-discrepancy,deterministicsequencesasdescribedbyNiederreiter(1978)andtheliteratureonquasimontecarlointegration.TheeectivenessofthesesimulationmethodsisdescribedinSection4.3.InSection5theeconometricapproachisappliedtothetelevisionindustry.Thetelevisionindustryisaneconomicallyimportantindustrywhoseproductsarediculttocharacterize.In2000advertisersspentabout52billiondollarsontelevisionads.3Revenuesdependonaudiencesizeandcomposition.4Despitethesehighstakes,techniquesforusingindustrydatatoanalyzethismarketarenotwelldeveloped.OurstudyisfacilitatedbyNielsenMediaResearch'spaneldatasetofindividuals'vi