4-Recent Developments in the Econometrics of Progr

JournalofEconomicLiterature2009,47:1,5–86http:=10.1257/jel.47.1.551.IntroductionManyempiricalquestionsineconomicsandothersocialsciencesdependoncausaleffectsofprogramsorpolicies.Inthelasttwodecades,muchresearchhasbeendoneontheeconometricandstatisticalanal-ysisofsuchcausaleffects.Thisrecenttheo-reticalliteraturehasbuilton,andcombinedfeaturesof,earlierworkinboththestatisticsandeconometricsliteratures.Ithasbynowreachedalevelofmaturitythatmakesitanimportanttoolinmanyareasofempiricalresearchineconomicsandsuitableforareview.Inthisarticle,weattempttopres-entsuchareview.Wewillfocusonpracti-calissuesforempiricalresearchers,aswellasprovideanhistoricaloverviewoftheareaandgivereferencestomoretechnicalresearch.Thisreviewcomplementsandextendsotherreviewsanddiscussions,includingthosebyRichardBlundellandMonicaCostaDias(2002),GuidoW.Imbens(2004),andJoshuaD.AngristandAlanB.Krueger(1999)andthebooksbyPaulR.Rosenbaum(1995),JudeaPearl(2000),Myoung-JaeLee(2005a),DonaldB.Rubin(2006),MarcoCaliendo(2006),AngristandJörn-SteffenPischke(2009),HowardS.Bloom(2005),StephenL.MorganandChristopherWinship(2007),JeffreyM.Wooldridge(2002)andImbensandRubin(forthcoming).Inaddition,thereviewsinJamesJ.Heckman,RobertJ.LaLonde,RecentDevelopmentsintheEconometricsofProgramEvaluationGuidoW.ImbensandJeffreyM.Wooldridge*Manyempiricalquestionsineconomicsandothersocialsciencesdependoncausaleffectsofprogramsorpolicies.Inthelasttwodecades,muchresearchhasbeendoneontheeconometricandstatisticalanalysisofsuchcausaleffects.Thisrecenttheoreti-calliteraturehasbuilton,andcombinedfeaturesof,earlierworkinboththestatisticsandeconometricsliteratures.Ithasbynowreachedalevelofmaturitythatmakesitanimportanttoolinmanyareasofempiricalresearchineconomics,includinglaboreconomics,publicfinance,developmenteconomics,industrialorganization,andotherareasofempiricalmicroeconomics.Inthisreview,wediscusssomeoftherecentdevelopments.Wefocusprimarilyonpracticalissuesforempiricalresearch-ers,aswellasprovideahistoricaloverviewoftheareaandgivereferencestomoretechnicalresearch.*Imbens:HarvardUniversityandNBER.Wooldridge:MichiganStateUniversity.FinancialsupportforthisresearchwasgenerouslyprovidedthroughNSFgrantsSES0136789,0452590and08.Wearegratefulforcom-mentsbyEstherDuflo,CarolineHoxby,RogerGordon,JonathanBeauchamp,LarryKatz,EduardoMorales,andtwoanonymousreferees.JournalofEconomicLiterature,Vol.XLVII(March2009)6andJeffreyA.Smith(1999),HeckmanandEdwardVytlacil(2007a,2007b),andJaapH.AbbringandHeckman(2007)provideanexcellentoverviewoftheimportanttheoreti-calworkbyHeckmanandhiscoauthorsinthisarea.Thecentralproblemstudiedinthisliter-atureisthatofevaluatingtheeffectoftheexposureofasetofunitstoaprogram,ortreatment,onsomeoutcome.Ineconomicstudies,theunitsaretypicallyeconomicagentssuchasindividuals,households,mar-kets,firms,counties,states,orcountriesbut,inotherdisciplineswhereevaluationmethodsareused,theunitscanbeanimals,plotsofland,orphysicalobjects.Thetreat-mentscanbejobsearchassistanceprograms,educationalprograms,vouchers,lawsorregulations,medicaldrugs,environmentalexposure,ortechnologies.Acriticalfeatureisthat,inprinciple,eachunitcanbeexposedtomultiplelevelsofthetreatment.Moreover,thisliteratureisfocusedonsettingswithobservationsonunitsexposed,andnotexposed,tothetreatment,withtheevalua-tionbasedoncomparisonsofunitsexposedandnotexposed.1Forexample,anindividualmayenrollornotinatrainingprogram,orheorshemayreceiveornotreceiveavoucher,orbesubjecttoaparticularregulationornot.Theobjectofinterestisacomparisonofthetwooutcomesforthesameunitwhenexposed,andwhennotexposed,tothetreat-ment.Theproblemisthatwecanatmostobserveoneoftheseoutcomesbecausetheunitcanbeexposedtoonlyonelevelofthetreatment.PaulW.Holland(1986)referstothisasthefundamentalproblemofcausalinference.Inordertoevaluatetheeffectofthetreatment,wethereforealwaysneedtocomparedistinctunitsreceivingthedifferentlevelsofthetreatment.Suchacomparison1Asoppposedtostudieswherethecausaleffectoffundamentallynewprogramsispredictedthroughdirectidentificationofpreferencesandproductionfunctions.caninvolvedifferentphysicalunitsorthesamephysicalunitatdifferenttimes.Theproblemofevaluatingtheeffectofabinarytreatmentorprogramisawellstudiedproblemwithalonghistoryinbothecono-metricsandstatistics.Thisistruebothinthetheoreticalliteratureaswellasinthemoreappliedliterature.TheeconometricliteraturegoesbacktoearlyworkbyOrleyAshenfelter(1978)andsubsequentworkbyAshenfelterandDavidCard(1985),HeckmanandRichardRobb(1985),LaLonde(1986),ThomasFrakerandRebeccaMaynard(1987),CardandDanielG.Sullivan(1988),andCharlesF.Manski(1990).Motivatedprimarilybyapplicationstotheevaluationoflabormarketprogramsinobservationalset-tings,thefocusintheeconometricliteratureistraditionallyonendogeneity,orself-selec-tion,issues.Individualswhochoosetoenrollinatrainingprogramarebydefinitiondif-ferentfromthosewhochoosenottoenroll.Thesedifferences,iftheyinfluencetheresponse,mayinvalidatecausalcomparisonsofoutcomesbytreatmentstatus,possiblyevenafteradjustingforobservedcovariates.Conse