




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
PublicDisclosureAuthorizedPublicDisclosureAuthorized
PolicyResearchWorkingPaper11059
DesignofPartialPopulationExperimentswithanApplicationtoSpilloversinTaxCompliance
GuillermoCruces
DarioTortarolo
GonzaloVazquez–Bare
WORLDBANKGROUP
DevelopmentEconomics
DevelopmentResearchGroupFebruary2025
ReproducibleResearchRepository
Averifiedreproducibilitypackageforthispaperisavailableat
,click
here
fordirectaccess.
PolicyResearchWorkingPaper11059
Abstract
Thispaperdevelopsaframeworktoanalyzepartialpopula–tionexperiments,ageneralizationoftheclusterexperimentaldesignwhereclustersareassignedtodifferenttreatmentintensities.Theframeworkallowsforheterogeneityinclus–tersizesandoutcomedistributions.Thepaperstudiesthelarge–samplebehaviorofOLSestimatorsandcluster–ro–bustvarianceestimatorsandshowsthat(i)ignoringclusterheterogeneitymayresultinseverelyunderpoweredexper–imentsand(ii)thecluster–robustvarianceestimatormaybeupward–biasedwhenclustersareheterogeneous.Thepaperderivesformulasforpower,minimumdetectable
effects,andoptimalclusterassignmentprobabilities.Alltheresultsapplytoclusterexperiments,aparticularcaseoftheframework.ThepapersetsupapotentialoutcomesframeworktointerprettheOLSestimandsascausaleffects.Itimplementsthemethodsinalarge–scaleexperimenttoestimatethedirectandspillovereffectsofacommunicationcampaignonpropertytaxcompliance.Theanalysisrevealsanincreaseintaxcomplianceamongindividualsdirectlytargetedwiththemailing,aswellascompliancespilloversonuntreatedindividualsinclusterswithahighproportionoftreatedtaxpayers.
ThispaperisaproductoftheDevelopmentResearchGroup,DevelopmentEconomic.ItispartofalargereffortbytheWorldBanktoprovideopenaccesstoitsresearchandmakeacontributiontodevelopmentpolicydiscussionsaroundtheworld.PolicyResearchWorkingPapersarealsopostedontheWebat
/prwp.Theauthors
maybecontactedatdtortarolo@.Averifiedreproducibilitypackageforthispaperisavailableat
http://
,click
here
fordirectaccess.
Y
C
I
A
E
RES
L
O
P
H
C
R
S
TRANSPARENT
P
E
R
W
O
R
K
I
ANALYSIS
A
NGP
ThePolicyResearchWorkingPaperSeriesdisseminatesthefindingsofworkinprogresstoencouragetheexchangeofideasaboutdevelopmentissues.Anobjectiveoftheseriesistogetthefindingsoutquickly,evenifthepresentationsarelessthanfullypolished.Thepaperscarrythenamesoftheauthorsandshouldbecitedaccordingly.Thefindings,interpretations,andconclusionsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelopment/WorldBankanditsaffiliatedorganizations,orthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.
ProducedbytheResearchSupportTeam
DesignofPartialPopulationExperiments
withanApplicationtoSpilloversinTaxComplianc
e*
GuillermoCruces,U.ofNottingham&CONICET-CEDLAS-UNLP
DarioTortarolo,WorldBankDECRG
GonzaloVazquez-Bare,UCSantaBarbara
JELCODES:C01,C93,H71,H26,H21,O23.
KEYWORDS:partialpopulationexperiments,spillovers,randomizedcontrolledtrials,clusterex-periments,two-stagedesigns,propertytax,taxcompliance.
*WethankYuehaoBai,YoussefBenzarti,AugustinBergeron,JavierBirchenall,MatiasCattaneo,MaxFarrell,KelseyJack,HeatherRoyer,DougSteigerwaldandAlisaTazhitdinovaforvaluablediscussionsandsuggestions,andseminarpartici-pantsatthe2021NationalTaxAssociationconference,IFS,CEDLAS-UNLP,andthe2022AdvanceswithFieldExperimentsconference.WethankJulianAmendolaggineandJuanLuisSchiavonifortheirinvaluablesupportthroughouttheproject.WethankBrunoCrponandRolandRathelotfortheirhelpinobtainingtheirdata.Theviewsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelop-ment/WorldBankanditsaffiliatedorganizations,orthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.Correspondingauthor:DarioTortarolo,E-mail:
dtortarolo@.
ThisprojectwasreviewedandapprovedinadvancebytheInstitutionalReviewBoardattheUniversityofNottingham.ThedesignforthisexperimentwaspreregisteredintheAEARCTRegistry(RCTID:AEARCTR-0006569).Allremainingerrorsareourown.
2
1Introduction
Randomizedcontrolledtrials(RCTs)areextensivelyusedineconomics.Alargefractionoftheseexperi-mentsarebasedontheassumptionthatthetreatmentassignmentofoneunitorsubjectdoesnotinfluencetheoutcomesofothers.Theassumptionofnointerference,however,maybeviolatedinmanysettings.Insuchcases,identifyingandmeasuringspilloversbetweenunitsiscrucialforunderstandingthenatureandmagnitudeofinteractionsbetweensubjects,aswellasforaccuratelyassessingthedirectimpactofthetreatment.
Whiletheearlyexperimentalliteratureconsideredtheimpactonuntreatedunitsinanex-postmanner(e.g.
MiguelandKremer,
2004
),fieldexperimentsincorporatingspillovereffectsintotheirdesignhavegainedtractioninappliedresearch.Insettingswhereunitsaregroupedintoindependentclusters,suchasschools,villages,orfirms,acommondesignisthepartialpopulationdesign.Partialpopulationdesignsareageneralizationoftheclustereddesignwhereinclustersassignedtodifferenttreatmentintensitiesorsaturationsarecomparedtopurecontrolclusterswithnotreatedunits(
Moffit,
2001;
DufloandSaez,
2003;
HudgensandHalloran,
2008;
HiranoandHahn,
2010;
Bairdetal.,
2018
).Thevariationintreatmentintensityallowsresearcherstodisentanglethedirectandindirecteffectsofatreatment.Inthispaper,weprovideaframeworktoanalyzethistypeofexperimentwhenclustersareheterogeneous.
Weconsidertwodimensionsofclusterheterogeneitythathaveimportantpracticalimplications:het-erogeneityinclustersizesandheterogeneityinoutcomedistributionsacrossclusters(distributionalhet-erogeneity)
1.
Whenanalyzinganexperimentwithheterogeneousclusters,correctlyaccountingforthisheterogeneityiscrucialforseveralreasons.Ontheonehand,varianceformulashavetobeadjustedac-cordingly,andfailingtodosomayresultinseverelyunderpoweredexperiments.Ontheotherhand,clusterheterogeneitycanaffecttheaccuracyofthelargesamplenormalapproximation,andinferencebasedonthisapproximationcanbemisleadingwhenclustersareveryheterogeneous(
Carter,Schnepel
andSteigerwald,
2017;
Djogbenou,MacKinnonandØrregaardNielsen,
2019;
HansenandLee,
2019;
SasakiandWang,
2022;
Chiang,SasakiandWang,
2023
).
Withthesechallengesinmind,ourpaperprovidesfivecontributions.First,inTheorem
1
,wederiveanasymptoticdistributionalapproximationforOLSregressionestimatorsinasettingwithbetween-clusterheterogeneity.Weconsideradouble-arrayasymptoticsettingwhereclustersizesareallowed,butnotre-quired,togrowwiththesamplesize.WeprovideconditionsunderwhichOLSestimatorsareconsistentforcluster-size-weightedaveragesofwithin-clusterdifferencesinmeans,andareasymptoticallynormal.Wealsoshowthat,inthepresenceofdistributionalheterogeneity,theusualcluster-robustvarianceestima-torisgenerallyupward-biased,andhenceinferencebasedonthisestimatorisconservative(Proposition
1
).Whilesimilarresultshavebeenobtainedindesign-basedsettingswithnon-randompotentialoutcomes(seee.g.
HudgensandHalloran,
2008;
BasseandFeller,
2018;
Abadieetal.,
2022;
Jiang,ImaiandMalani,
1Wenotethatourframeworkallowsforgeneralformsofbetween-clusterheterogeneity,butassumesthatoutcomesareiden-ticallydistributedwithineachcluster.Thegeneralizationofourresultstothecasewhereoutcomedistributionsareheteroge-neouswithinaclusterisleftforfutureresearch.
3
2023
),toourknowledgewearethefirsttoshowthisresultinasuperpopulationsettingunderdistributionalheterogeneity.
Oursecondcontributionistoderiveexplicit,closed-formformulastoconductpowerandminimumdetectableeffect(MDE)calculationsunderthetwoaforementionedsourcesofclusterheterogeneity.Wethenconsideranintermediatesettingwhereclustersdifferinsizebutnotintheiroutcomedistributions,whichsimplifiespowerandminimumdetectableeffectscalculationsandcanbeappliedmoreeasilywhenbaselineoutcomedataisnotavailable.Weshowhowourformulasgeneralizethoseavailableintheexistingmethodologicalliteratureonexperimentaldesign(
Duflo,GlennersterandKremer,
2007;
Hirano
andHahn,
2010;
Bairdetal.,
2018
)byallowingformultipletreatmentintensities,clusterheterogeneity,heteroskedasticityandgeneralformsofintraclustercorrelationinoutcomesandtreatments.
Ourthirdcontributionistoderiveoptimalassignmentprobabilitiesdeterminingtheproportionofclus-terstobeassignedtoeachtreatmentsaturation(Theorem
2
).Weprovideatractable,closed-formsolutiontotheoptimalchoiceproblemofminimizingaweightedaverageofestimators’variances.Wealsodiscusshowalternativeoptimalitycriteriamaybeusedincombinationwithourvarianceformulasusingnumericalmethods.
Ourfourthcontributionistosetupapotentialoutcomesframeworkwithwithin-clusterspillovers,heterogeneoustreatmenteffects,andheterogeneousclusters.WeusethisframeworktoprovidesufficientconditionsforOLSestimandstorecovercausaldirectandspillovereffects.
Fifth,basedonourframework,wedesignedandconductedalarge-scalefieldexperimenttoestimatedirectandspillovereffectsofarandomizedcommunicationcampaignonpropertytaxcomplianceinAr-gentina.Ourexperimentsentpersonalizedletterstorandomlyselecteddwellingswithremindersabouttaxesdue,informationaboutthestatusoftheaccount,duedates,pastduedebt,andpaymentmethods.Whilethereisampleevidenceontheeffectoftaxremindersoncomplianceandcollection(
Antinyanand
Asatryan,
2024
),ourgoalwastofindevidenceonrelativelyelusivespillovereffectsfrominformationcampaignsontaxcollection.Wedesignedtheexperimentbasedonourmethodologicalresultstocapturespillovereffectsofourmailingsonneighborswholiveinthesamestreetblocksoftreatedindividualsbutwhodidnotreceivealetter.Ourresultsrevealhigherpaymentratesfortreatedindividuals,butalsofortheiruntreatedneighborsinthesamestreetblock,comparedtoaccountsinpurecontrolblockswherenoonereceivedtheletter.Spillovereffectsarelowerinmagnitudebutstillsubstantialandpreciselyestimatedinhigh-saturationstreetblocks,especiallywhenaccountingforexpected(pre-registered)heterogeneityinpastcompliance:paymentratesofuntreatedaccountsinhighsaturationblockswithabovemedianpastcomplianceincreasedby2.6percentagepoints,comparedtodirecteffectsofabout5.1percentagepoints.
Comparisonwithcurrentliterature.Ourpapercontributestoagrowingliteratureonexperimentaldesign(
Duflo,GlennersterandKremer,
2007;
BruhnandMcKenzie,
2009;
Bugni,CanayandShaikh,
2018,
2019;
Bai,
2022
)andinparticulartotheliteratureondesignandanalysisofexperimentsunder
4
spilloversorinterference(
HiranoandHahn,
2010;
Athey,EcklesandImbens,
2018;
Bairdetal.,
2018;
Basse,FellerandToulis,
2019;
Jiang,ImaiandMalani,
2023;
Puelzetal.,
2022;
Viviano,
2024;
Leung,
2022;
Liu,
2023
).Morespecifically,ourresultsgeneralizethoseof
HiranoandHahn
(2010
),
Hudgens
andHalloran
(2008
)and
Bairdetal.
(2018
)byallowingforclusterheterogeneity,heteroskedasticity,generaltreatmentassignmentmechanismsandwithin-groupcorrelationstructuresandalternativecriteriaforoptimaltreatmentassignment.
Inrelatedwork,
Athey,EcklesandImbens
(2018
),
Basse,FellerandToulis
(2019
)and
Puelzetal.
(2022
)deriverandomizationinferencetestsforageneralclassofnullhypothesesunderinterference.Acloselyrelatedstudyis
Jiang,ImaiandMalani
(2023
),whoanalyzetwo-stagecompletelyrandomizedexperimentsandproviderandomization-basedvarianceestimatorsandsamplesizeformulas.Ourre-sultscomplementthisliteraturebyconsideringdifferentestimands,differentassignmentmechanismsandbyconductingsuper-population-basedlarge-sample(insteadofdesign-based)inferenceinadoublear-rayasymptoticframework.Ourapproachallowsustodeterminetheroleofclusterheterogeneityintheasymptoticbehaviorofthetreatmenteffectestimators.
Ourpaperisalsorelatedtotheliteratureoninferenceinclusteredexperiments,whichareaparticularcaseofpartialpopulationexperimentswithonlytwosaturationsandnowithin-clustertreatmentvariation.
Bugnietal.
(2023
)studyinferenceinclusteredexperimentswithnon-ignorableclustersizesandderivevarianceestimatorsandvalidinferenceproceduresinasetupwithrandomclustersizes.WefurtherdiscusstherelationshipbetweenourresultsandthatpaperinSection
3.5.
Wealsocontributetoalargeempiricalliteratureonpropertytaxesandasmallbutgrowingempiricalliteratureonspillovereffectsintaxcompliance.Onpropertytaxes,recentcontributionsinclude
Brock-
meyeretal.
(2020
)studyofMexicoCity,
Bergeron,TourekandWeigel
(2024
)and
Weigel
(2020
)fortheDemocraticRepublicofCongo,and
Krause
(2020
)forHaiti,amongothers.Thelattertwoareran-domizedcontrolledtrials,andinbothcases,theauthorsaddressthepresenceofspillovers,butinex-postanalysisratherthanintheexperimentaldesigns.Theeffectofsocialinteractionsintaxcomplianceinter-ventionshasremainedarelativelyelusiveissueinthebroaderexperimentalcomplianceliterature.Somenotableexceptionsare
Pomeranz
(2015
),whodetectsenforcementspilloversuptheVATchaininChileanfirms,
Drago,MengelandTraxler
(2020
)whostudyenforcementspilloversofTVlicensinginspectionsonuntreatedhouseholdsinAustria,and
Boningetal.
(2020
)whoanalyzedirectandnetworkeffectsfromin-personvisitsbyrevenueofficersonvisitedandnon-visitedfirmsintheUnitedStates(seethereviewin
PomeranzandVila-Belda,
2019
,formorestudiescoveringspillovereffects).InArgentina,arecentstudyby
Carrillo,CastroandScartascini
(2021
)findsneighborhoodspillovereffectsfromaprogramthatrandomlyawarded400taxpayerswiththerepairofasidewalk.Whereasthesepapersfindspillovereffectsintaxcompliance,theiroriginalexperimentswerenotdesignedtocapturetheseeffects.Webuildonthesepioneeringworkswithaninterventiondesignedwiththepurposeofcapturingspillovers.
Thepaperisorganizedasfollows.Section
2
illustratesthepracticalimportanceofclusterheterogeneitywhenconductingpowercalculations.InSection
3
,wesetupourframeworkandderivethemainresults.In
5
Section
4
,weimplementourmethodsinalarge-scalerandomizedcommunicationcampaign,wedescribetheadministrativedatausedintheanalysis,theempiricalstrategy,andevidenceofdirectandspillovereffects.Section
5
providessomepracticalrecommendationsfordesigningandanalyzingpartialpopulationexperiments.Section
6
concludes.
2WhyisClusterHeterogeneityImportant?
Weconsiderapopulationwhereunitsaregroupedintomutuallyexclusiveandindependentclusters.Com-monexamplesofthistypeofclusteringarestudentsinschools(
MiguelandKremer,
2004;
Beuermann
etal.,
2015
),familymembersinhouseholds(
Barrera-Osorioetal.,
2011;
FoosanddeRooij,
2017
),jobseekersinlocallabormarkets(
Crponetal.,
2013
),employeesinfirmsororganizations(
DufloandSaez,
2003
),orhouseholdsinneighborhoods,villagesorothergeographicadministrativeunits(
Angelucciand
DeGiorgi,
2009;
IchinoandSch…undeln,
2012;
HaushoferandShapiro,
2016;
GinandMansuri,
2018
).Inourapplication,alocalpropertytaxreminderinformationcampaign,thepopulationofinterestconsistsoftaxpayersinresidentialblocks.Withinthispopulation,westudyanexperimentaldesignwheretreatmentassignmentscanvarybothbetweenandwithinclusters.
Figure
1
showsthedistributionofclustersizesinsixpartialpopulationexperiments,includingouranalysissampleandfivepublishedpapers(
Crponetal.,
2013;
GinandMansuri,
2018;
Haushoferand
Shapiro,
2016;
IchinoandSch…undeln,
2012;
Imai,JiangandMalani,
2021
).Thefigurerevealssubstantialvariationinclustersizes.Whenclustersizesareheterogeneous,itislikelythatthedistributionofoutcomeswillvaryacrossclustersaswell.Forinstance,onemayexpectthemeanandthevarianceoftheoutcometobedifferentinlargeclusterscomparedtosmallclusters.Werefertothevariationinoutcomedistributionsacrossclustersasdistributionalheterogeneity.
Intuitively,withheterogeneousclusters,thevarianceofanestimatorofinterest,suchasadifference
inmeansbetweenunitsintreatedanduntreatedclusters(wedefinetheestimatorsofinterestpreciselyinthenextsection),canbedecomposedintofourparts:
V[]≈varianceunderuncorrelatedobservations(1)
+clusteringwithequally-sizedclusters(2)
+clustersizeheterogeneity(3)
+clusterdistributionalheterogeneity(4)
Thefirsttermisthevariancethatwouldbeobtainedifobservationswereuncorrelatedwithinclusters.Thesecondtermisanadjustmentfactorthataccountsforthewithin-clustercorrelation,oftenknownasthe“designeffect”orthe“Moultonfactor”(after
Moulton,
1986
)thatdependsontheaverageclustersize.Theterminthethirdlinerepresentstheadditionalvariationduetotheheterogeneityinclustersizes,
6
whichintuitivelyaccountsforthevarianceofclustersizes(
Moulton,
1986
,alsoderivesthisadjustmentforarandomeffectsmodel).Finally,thelastcomponentaccountsforthebetween-clusterheterogeneityinoutcomedistributions.Whiletheneedtoaccountforwithin-clustercorrelations(lines(1)and(2))iswell-understoodfordesigningandanalyzingclusteredexperiments,theadjustmenttermsthataccountforclusterheterogeneityaretypicallyassumedawaybytheliteratureonexperimentaldesign(e.g.
Bloom,
2005;
Duflo,GlennersterandKremer,
2007;
HiranoandHahn,
2010;
Bairdetal.,
2018
).
Tonumericallyillustratetheimportanceofappropriatelyaccountingforclusterheterogeneityinthisdesign,weconsiderthesimplesettingofaclusterRCT(whichisaparticularcaseofapartialpopulationexperiment)where“afew”clustersare“large”.Specifically,weconsiderasampleof200clusters,indexedbyg=1,...,200,eachhavingsizeng.Thefirst10clusterscontain100units,ng=100,andtheremaining190clusterscontain25unitseach,ng=25(thesevaluesarechosentomatchthemedianvaluesintheliteratureinFigure
1
).Weassumethetreatmenthasnoeffect,andtheoutcomeofuniti=1,...,nginclustergisgivenbyarandomeffectsmodel:Yig=αg+νg+ωig,νg,1/2),ωigN(0,1/2)withνgindependentofωigandwhereαgisa(non-random)interceptwithαg=0ifng=25andαg=1ifng=100.ThismodelimpliesthattheaverageoutcomeisE[Yig]=1inlargeclustersandE[Yig]=0insmallclusters.Inaddition,V[Yig]=1andthewithin-clustercorrelationbetweenoutcomesiscor(Yig,Yjg)=0.5.
Figure
2
plotsthreepowerfunctionsforthedifferenceinmeansbetweentreatedanduntreatedclustersthataresearchermayconsiderwhendesigningthisexperiment.Theshort-dashedcurverepresentsthepowerfunctionthatisobtainedwhenignoringbothsourcesofheterogeneity,thatis,consideringonlythetermsinlines(1)and(2)ofthevarianceformula.Usingthisformula,theMDEat80%power,giventhissamplesize,is0.29standarddeviations.However,whenaccountingforthevariationinclustersizes,thecorrespondingpowerfunctionisrepresentedbythelong-dashedcurve.Accordingtothiscurve,thepowertodetectaneffectof0.29isnot80%but69%,sotheexperimentisunderpowered.Furthermore,thetruepowerfunctionthataccountsforbothsourcesofheterogeneity(sizesandoutcomedistributions)isrepresentedbythesolidcurve.Thiscurveshowsthatthetruepowertodetectaneffectof0.29inthissettingwithheterogeneousclustersis48%,significantlybelowthedesiredpowerof80%.Thisnumericalexerciseshowshowignoringheterogeneitymayresultinseverelyunderpoweredexperiments.WeprovidefurtherexamplesoftheimportanceofaccountingforheterogeneityinSection
4.
3AnalysisofPartialPopulationExperiments
3.1Setup
Weconsiderasampleofobservations(units)thataredividedintomutuallyindependentclustersg=
1,...,G,whereeachclustergcontainsngobservationsi=1,...,ngandthetotalsamplesizeisn=
7
Σg.Weviewclustersizesasnon-random(see
Bugnietal.,
2023;
SasakiandWang,
2022
,foranalternativesamplingapproachwhereclustersizesarerandom).Inapartialpopulationexperiment,clustersarerandomlydividedintocategoriesorsaturationsdenotedbyTg∈{0,1,2,...,M},wherebyconventionTg=0denotesapurecontrolcluster(i.e.aclusterwherenounitistreated).LetP[Tg=t]=qt∈(0,1)denotetheprobabilitythatclustergisassignedtosaturationt.Withineachcluster,abinarytreatmentDigisassignedtounitswithprobabilityP[Dig=1|Tg=t]whereP[Dig=0|Tg=0]=1
.2
WeletDg=(D1g,D2g,...,Dngg)9bethevectorofunit-leveltreatmentassignmentsinclusterg,D=
(D,...,D)9andT=(T1,...,TG)9.Figure
A.3
providesanexampleofapartialpopulationdesign
withfoursaturations.NoticethatbothstandardRCTswithindependentobservationsandclusterRCTsareparticularcasesofpartialpopulationexperiments,aswefurtherillustrateinSection
3.5.
TheobservedoutcomeofinterestforunitiinclustergisdenotedbyYigandweletYg=(Y1g,...,Yngg)9bethevectorofobservedoutcomesinclusterg.Inpartialpopulationexperiments,theestimandsofin-terestaretypicallycomparisonsofaverageoutcomesbetweentreatedoruntreatedunitsintreatedclusterstopurecontrolunits,E[Yig|Dig=d,Tg=t]-E[Yig|Tg=0],pooledacrossclusters.Inthefirstpartofthepaper,wetaketheseestimandsasgivensincetheyarethemostcommonlyanalyzedestimandsintheempiricalliterature.InSection
3.6
,wesetupapotentialoutcomesframeworktorigorouslyjustifythecausalinterpretationoftheseestimands.Letµg(d,t)=E[Yig|Dig=d,Tg=t]betheconditionalexpectationoftheoutcomeinclusterggivenassignment(d,t).Weconsiderthefollowingsamplemeansestimators:
where1=1(Tg=t),N=Σi1(Dig=d)andY-gd=ΣiYig1(Dig=d)/N,definedwhenever
N>0.TheseestimatorsarecommonlycomputedbyrunninganOLSregressionoftheoutcomeon
afullsetofindicators(1(Dig=d,Tg=t))(d,t),withoutanintercept.Thus,inwhatfollows,werefertotheseestimatorsasOLSestimators.Ourparameterofinterestisthevectorofcluster-size-weightedaverageofcluster-specificdifferencesinmeans:
Wenotethatourframeworkcaneasilyaccommodateotherparameterswithdifferentweightingschemes,
suchasthesimpleaverageacrossclustersΣg(d,t)-µg(0,0))/G.
2Inpractice,somedesiredsaturationsmaynotcoincidewiththeobservedproportionoftreatedunitsforsomeclustersizes.Forinstance,ifP[Dig=1|Tg=t]=0.5butngisodd,theobservedproportionoftreatedcannotbeexactly0.5.Appendix
A.4
proposesanassignmentmechanismthatensuresthattheexpectedproportionoftreatedcoincideswithP[Dig=1|Tg=t].
8
3.2AsymptoticBehaviorofOLSEstimators
WenowstudytheasymptoticdistributionoftheOLSestimatorsdefinedinEquation(
5
)andfunctionsthereof.Weconsideradouble-arrayasymptoticsettingwheretheclustersizesareallowed,butnotre-quired,togrowwiththesamplesize.Thistypeofapproximationismoreappropriatethantheboundedclustersizeapproachwhengroupscanbelargeandheterogeneousinsize,butwenotethatthesettingswithboundedclustersizesand/orequally-sizedclustersarenestedasparticularcasesofouranalysis
.3
Weconsiderthefollowingsamplingscheme.
Assumption1(Sampling)
(i)(Yg,,Dg,,Tgg.
(ii)Foreachgandforalli=1,...,ng,E[Yi|Dig=d,Tg=t]=µ(d,t)forall(d,t)andforalllsuchthatE[|Yig|l|Dig=d,Tg=t]<∞.
(iii)Foreachgandforalli=1,...,ng,P[Dig=d|Tg=t]=pg(d|t)andP[Dig=d,Djg=d,|Tg=
t]=pg(d,d,|t)foralld,d,andt.
Part(i)statesthatclustersaremutuallyindependent,astandardassumptionintheclusteringliterature.
Noticethatwedonotrequireclusterstobeidenticallydistributed,sooutcomedistributionscanbehet-erogeneousacrossclusters.Part(ii)statesthataverageconditionaloutcomesarethesameforallunits
inthesamecluster.Inwhatfollowswedefineµ(d,t)=µg(d,t)forl=1toreducenotation.Part
(iii)statesthattheunit-leveltreatmentprobabilitiesarethesamewithinacluster.Notethatwithin-clusterassignm
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 斜杠人生的茶艺师考试试题及答案
- 2025计算机初级考试核心知识提炼试题及答案
- 2025健康管理师考试应试技巧试题及答案
- 二零二五年度押付签年人工智能教育平台合作协议
- 二零二五年度工伤责任认定及处理协议
- 二零二五年度手车交易风险评估及担保合同
- 2025年度矿山员工劳动合同与矿山应急救援物资储备协议
- 二零二五年度二零二五年度文化娱乐品牌商标许可使用授权协议书
- 二零二五年度模特赛事选手签约合同
- 二零二五年度事业单位员工协商解除劳动合同补偿协议
- 二级营销员考试题及参考答案
- 部编版道德与法治二年下册《我能行》说课稿共2课时(附教学反思)课件
- 中建履带吊安拆安全专项施工方案
- 医学论文格式与写作课件
- 商业秘密保护管理办法
- 市场监监督管理执法讲座
- 2024年天翼云从业者认证考试题库大全(含答案)
- 《国际形势》课件
- 煤矿开采安全管理培训课件
- 2022年高考真题-政治(重庆卷) 含答案
- 校园欺凌教育主题班会课件
评论
0/150
提交评论