世界銀行-利用調(diào)查對(duì)調(diào)查的推斷以低成本填補(bǔ)貧困數(shù)據(jù)缺口：來(lái)自隨機(jī)調(diào)查實(shí)驗(yàn)的證據(jù)

PublicDisclosureAuthorizedPublicDisclosureAuthorized

PolicyResearchWorkingPaper10738

UsingSurvey-to-SurveyImputationtoFillPovertyDataGapsataLowCost

EvidencefromaRandomizedSurveyExperiment

Hai-AnhDang

TalipKilic

VladimirHlasny

KseniyaAbanokova

CalogeroCarletto

WORLDBANKGROUP

DevelopmentEconomics

DevelopmentDataGroup

March2024

PolicyResearchWorkingPaper10738

Abstract

Surveydataonhouseholdconsumptionareoftenunavailableorincomparableovertimeinmanylow-andmiddle-incomecountries.BasedonauniquerandomizedsurveyexperimentimplementedinTanzania,thisstudyoffersnewandrigorousevidencedemonstratingthatsurvey-to-surveyimputationcanfillconsumptiondatagapsandprovidelow-costandreliablepovertyestimates.Basicimputationmodelsfeaturingutilityexpenditures,togetherwithamodestsetofpredictorsondemographics,employment,householdassets,andhousing,yieldaccuratepredictions.Imputationaccuracyisrobusttovaryingthesurveyquestionnairelength,thechoiceofbasesurveysforestimatingtheimputationmodel,differentpovertylines,

andalternative(quarterlyormonthly)ConsumerPriceIndexdeflators.Theproposedapproachtoimputationalsoperformsbetterthanmultipleimputationandarangeofmachinelearningtechniques.Inthecaseofatargetsurveywithmodified(shortenedoraggregated)foodornon-foodconsumptionmodules,imputationmodelsincludingfoodornon-foodconsumptionaspredictorsdowellonlyifthedistributionsofthepredictorsarestandardizedvis-à-visthebasesurvey.Forthebest-performingmodelstoreachacceptablelevelsofaccuracy,theminimumrequiredsamplesizeshouldbe1,000forboththebaseandtargetsurveys.Thediscussionexpandsontheimplicationsofthefindingsforthedesignoffuturesurveys.

ThispaperisaproductoftheDevelopmentDataGroup,DevelopmentEconomics.ItispartofalargereffortbytheWorldBanktoprovideopenaccesstoitsresearchandmakeacontributiontodevelopmentpolicydiscussionsaroundtheworld.PolicyResearchWorkingPapersarealsopostedontheWebat

/prwp.Theauthorsmay

becontactedathdang@andtkilic@.

ThePolicyResearchWorkingPaperSeriesdisseminatesthefindingsofworkinprogresstoencouragetheexchangeofideasaboutdevelopmentissues.Anobjectiveoftheseriesistogetthefindingsoutquickly,evenifthepresentationsarelessthanfullypolished.Thepaperscarrythenamesoftheauthorsandshouldbecitedaccordingly.Thefindings,interpretations,andconclusionsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelopment/WorldBankanditsaffiliatedorganizations,orthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.

ProducedbytheResearchSupportTeam

UsingSurvey-to-SurveyImputationtoFillPovertyDataGapsataLowCost:EvidencefromaRandomizedSurveyExperiment

Hai

-AnhDang,TalipKilic,VladimirHlasny,KseniyaAbanokovaandCalogeroCarletto*

Keywords:consumption,poverty,survey-to-surveyimputation,householdsurveys,Tanzania.

JELCodes:C15,I32,O15.

*TheseniorauthorshipissharedbetweenDangandKilic.Dang

(hdang@;

correspondingauthor)isasenioreconomistintheLivingStandardsandMeasurementStudy(LSMS)UnitattheWorldBankDevelopmentDataGroupinWashington,DCandisalsoaffiliatedwithGLO,IZA,IndianaUniversity,andLondonSchoolofEconomicsandPoliticalScience;Kilic

(tkilic@;

correspondingauthor)istheseniorprogrammanagerfortheLSMSUnitattheWorldBankDevelopmentDataGroupinWashington,DC;Hlasny(vhlasny@)isaneconomicaffairsofficerattheUNESCWAinBeirut,Lebanon;Abanokova

(kabanokova@)

isaneconomistintheLSMSUnitattheWorldBankDevelopmentDataGroupinWashington,DC;andCarletto

(gcarletto@)

istheseniormanagerfortheLSMSandtheStrategyandOperationsUnitsattheWorldBankDevelopmentDataGroupinWashington,DC.WewouldliketothankBenoitDercef,AndrewDillon,AnneSwindale,andparticipantsatthe2023EuropeanSurveyResearchAssociation(ESRA)conference,theIPA/GRPLconference(Northwestern)andvariousseminarsandworkshopsatAustralianNationalUniversity,UniversityofOxford,andtheWorldBankforhelpfuldiscussionandfeedbackontheearlierdrafts.WearegratefulforthefundingfromtheUnitedStatesAgencyforInternationalDevelopment(USAID).

2

1.Introduction

Householdconsumptionsurveydatathatunderliemonetarypovertyestimatesinlow-andmiddle-incomecountriesareoftenunavailable,unreliableorincomparable.Toaddressthesechallenges,imputation-basedmethodshavebecomeincreasinglymorecommonnotonlytofillpovertydatagapsindata-scarceandresource-constrainedcontexts,butalsotoidentifyproject/programbeneficiariesandevaluatedevelopmentproject/programimpactsonpovertyat

lowcost(WorldBank,2021;SmytheandBlumenstock,2022;DangandLanjouw,2023).

1

Buildingontheseminaltechniquethatobtainssmallareaestimatesofmonetarypovertybyimputingfromahouseholdconsumptionsurveyintoacensus(Elbersetal.,2003),survey-to-surveyimputationbuildsanimputationmodelusingappropriatepredictorvariablesfromanexistingolderconsumptionsurvey(basesurvey),whichcanbesubsequentlyappliedtothesamevariablesinanothernon-consumptionsurvey(targetsurvey)toprovidepovertyestimatesforthelattersurvey.Thetargetsurveycanbeeitheranexisting,non-consumptionsurvey,suchasaDemographicandHealthSurvey(DHS)oralaborforcesurvey(StifelandChristiaensen,2007;Douidichetal.,2016),orapurposefullycommissionedsurveythatonlycollectstherequisitepredictors.Recentapplicationsalsoincludesourcingthedatafortherequisitepredictorsfromadministrativerecordstoimputepovertyforhard-to-reachrefugeepopulations(Altindagetal.,2021;DangandVerme,2023),orphonecalldetailrecordstotargettheultra-poor(Aikenetal.,

2023).

Threekeyconceptual,butunderstudied,issuesmotivateourwork.First,theliteratureon

survey-to-surveyimputationhaslongemphasizedtherequirementofhavingidenticalquestions

1Imputationtechniquesarewidelyusedbyinternationalorganizationsandnationalstatisticalagenciestofillinmissingdatagapssuchaseducationstatistics(UOE,2020)andincomedata(USCensusBureau,2017).SeealsoDangandLanjouw(2023)forarecentreviewofpovertyimputationstudies.

3

forpovertypredictorsinbothbaseandtargetsurveys.However,evenifthisrequirementisfulfilled,substantialdifferencesmaystillexistbetweenbaseversustargetsurveysregardinglength,thematicscope,andcomplexityofquestionnaires.Thesedifferencesmayleadtoconsiderabledifferencesininterviewdurationandrespondentburden,whichcanaffectmeasurementindiversewaysthatareultimatelycontext-andsubject-specific(Kreuteretal.,2011;Eckmanetal.,2014).Inourcase,theunderstudiedtopiciswhetherpovertyimputationaccuracycanbeaffectedbythefactthatthetargetsurveyquestionnaire,bydesign,wouldbelighterandlessburdensomethanitsolder,basesurveycounterpart–eveniftherequisitequestionsunderlyingthepovertypredictorsareidenticalacrossbaseandtargetsurveyquestionnaires.TheonlyavailableevidenceregardingthisquestioncomesfromarandomizedexperimentthatwasimplementedinMalawibutnotreplicatedelsewhere,andthatshowsthemeasurementofpovertypredictorscanindeedbeaffectedbythelengthofthetargetsurveyinawaythatcanalsoimpact

predictedpovertyestimates(KilicandSohnesen,2019).

Thesecondandrelatedissueiswhethershorterconsumptionmodulesincludedinatargetsurvey(e.g.,withreducedoraggregateditemlistsvis-à-visthebasesurvey)canprovidecheaper-to-collectbutreliablepredictorsthatcanfurtherboosttheaccuracyofpovertypredictionsundermarginaladditionalcostsofdatacollection.Inthiscase,therequisitequestionsunderlyingthepovertypredictorsmaybenon-identicalacrossbaseandtargetsurveyquestionnaires–relaxingtheaforementionedtraditionalrequirementforsurvey-to-surveyimputation.Inthisrespect,onlytwostudiesexist,andtheyofferinconclusiveevidence.WhileChristiaensenetal.(2022)suggestthatusingconsumptionsub-aggregatesforpovertyimputationonlyworksundercertainstringentconditions,Dangetal.(forthcoming)analyze14surveysfromvariouscountriesanddemonstrate

thataddinghouseholdutilityexpenditurestoabasicimputationmodelwithhousehold

4

demographicandemploymentattributescanproduceaccuratepovertypredictions-consistentlywithinthe95percentconfidenceinternal,andoftenwithinonestandarderror,oftheobserved

“true”povertyrate.

2

Finally,thelastissuemotivatingourworkisthatexistingstudiesthat“validate”imputedpovertyestimateswereimplementedinartificialsettings.Specifically,thesestudiestypicallypursuevalidationbyestimatinganimputationmodelonanolder,baseconsumptionsurveyandapplyingthemodeltoamorerecent,targetconsumptionsurvey,pretendingthattherewerenoconsumptiondatainthelattersurvey.Thesestudiessubsequentlycomparetheresultingimputedestimatetothetruepovertyratebasedontheactualconsumptiondatainthetargetsurvey.Thefactthatthenewersurveyroundservesbothasthetargetsurveyandasthesourceoftruepovertyabstractsawayfromreal-lifedifferencesinbaseversustargetsurveydesignthatmotivateourworkinthefirstplace.Thesetraditionalartificialsettingsalsodifferfrommanypracticalapplicationsforsurvey-to-surveyimputationwhereanewsurveywithadifferentdesignisimplementedasthetargetsurvey(e.g.,asurveythatdoesnotcollectconsumptiondataorthatadministerslighter

consumptionmodules-asinthecaseofmostproxy-meanstests).

Againstthisbackground,wereportonauniquerandomizedandnationallyrepresentativehouseholdsurveyexperimentthatwasimplementedinTanzaniain2022tosystematicallyinvestigatetheunderstudiedtopicsthathaveabearingontheoperational/practicalapplicationsofsurvey-to-surveyimputationtofillpovertydatagaps.Theexperimentfeaturedthreetreatmentarms(TA)thatsampledhouseholdswererandomlyassignedtoandthatdifferedintermsofquestionnairedesign.TreatmentArm1(TA1)householdswereadministeredaquestionnairethat

collectscomprehensivedataonhouseholdconsumptionandallowsforthecomputationof

2Weusetheterm“true”povertyratetorefertothepovertyratethatcanbeestimatedusingtheactualhouseholdconsumptiondata.

5

benchmarkpovertyestimates,whichisidenticaltothequestionnaireforthebasesurveythatpermitstheestimationofawiderangeofcompetingimputationmodels.TA2householdswereadministeredalightquestionnairevariantthatonlyincludedquestionsthatpermittheestimationofadata-modestsubsetofimputationmodels,whichadditionallyincludestheTA1foodconsumptionmodulebutwithareducedlistofkeyfooditems.Finally,TA3householdswereadministeredanalternativelightquestionnairethatsharesthesamecoreastheTA2questionnaireandthatincludesalternate,aggregatedversionsofTA1foodandnon-foodconsumptionmodules.ThesedataareinturncomplementedwiththedatafromthenationallyrepresentativeTanzaniaNationalPanelSurvey(TZNPS),andspecificallytheTZNPS2020/21and2019/20roundsthatareusedasbasesurveysfortheestimationoftheimputationmodelsthatareinturnappliedtoeach

treatmentarmtoobtainacross-yearpredictions.

Throughourresearch,wemakenovelcontributionstotheliteratureby(a)providingexperimentalevidenceregardingtheeffectsoftargetsurveydesignonpovertyimputation,(b)sidesteppingusualconcernsregardingthe“validation”ofimputedestimatesbyofferingareal-lifesettingwithbenchmarkdata,and(c)providingnewevidenceregardingtheminimum-requiredbaseandtargetsurveysamplesizes.Toourknowledge,weofferthefirststudythatleveragesarandomizedandnationallyrepresentativesurveyexperimenttorigorouslystudytheseinter-connected,butlittle-explored,researchquestionsthatareattheheartofsurvey-to-surveyimputation.Inthissense,ourworkisalsobroadlyrelatedtoagrowingliteraturethatreliesonrandomizedsurveyexperimentsinlow-andmiddle-incomecontextstogaugetherelativeaccuracyandcost-effectivenessofcompetingsurveymethodsvis-à-visgold-standardmeasurementapproaches(Beegleetal.,2012;Arthietal.,2018;Gourlayetal.,2019;DeWeerdtetal,2020;

Kilicetal.,2021;Abateetal.,2023).

6

Theanalysisdemonstratesthatifthepredictorsinthetargetsurveyareelicitedthroughquestionsthatareidenticaltotheircounterpartsinthebasesurvey,imputationaccuracyisnotimpactedbytheremainingdifferencesbetweenthebaseandtargetsurveysintermsofscopeandcomplexity.Basicimputationmodels,includingacoresetofpredictorsondemographics,employment,householdassetsandhousing,and/orutilityexpenditures,yieldhighlyaccuratepredictionsvis-à-visthetruepovertyrate.Furthermore,regardingTA2orTA3withmodified(eithershortenedoraggregated)foodandnon-foodconsumptionmodules,imputationmodelsincludingfoodconsumptionornon-foodconsumptionexpendituresaspredictorsdowellonlyifthedistributionsofthepredictorsarestandardizedvis-à-visthebasesurvey(whichcanbeeithertheTZNPSorTA1).Finally,forthebest-performingmodelstoreachacceptablelevelsofaccuracy,theanalysisshowsthattheminimum-requiredsamplesizeshouldbe1,000observationsforboththebasesurveyandthetargetsurvey.Theresultsarerobusttothechoiceofbasesurveysusedforimputationmodelestimation;differentpovertylines;andalternative(quarterlyormonthly)CPIdeflators.Ourproposedapproachtoimputationisalsoshowntoperformbetterthan

multipleimputationandarangeofmachinelearningtechniques.

Thispaperconsistsofsixsections.Section2presentstheexperimentaldesign(Section2.1)anddescriptivestatistics(Section2.2).Section3discussestheanalyticalframework.Section4presentsthemainestimationresults(Section4.1)androbustnesschecks(Section4.2),followedbysection5onvariousextensions.Section6concludes.WeprovideadditionalestimationresultsinAppendixA,furtherdescriptionoftheconsumptionaggregatesinAppendixB,andmore

detaileddiscussionoftheformulasandintuitionbehindthemethodinAppendixC.

7

2.Experimentaldesignanddescriptivestatistics

2.1.Experimentaldesign

ThedatacomefromtheTanzaniaMethodologicalSurveyExperimentonHouseholdConsumptionMeasurement,whichwasconductedfromApriltoJuly2022bytheTanzaniaNationalBureauofStatistics,withtechnicalsupportfromtheWorldBankLivingStandardsMeasurementStudy(LSMS)program.InformedbythepowercalculationsbasedonthepastroundsoftheTanzaniaNationalPanelSurvey(TZNPS)andtheHouseholdBudgetSurvey(HBS),theexperimentspanned143enumerationareas(EAs)acrossMainlandTanzaniaandZanzibar,includingbothurbanandruralareas.IneachsampledEA,25householdswereselectedatrandomfromafreshhouseholdlistingthatwasconducted,outofwhichfivesampledhouseholdswere

assignedatrandomtooneoffivesurveytreatmentarms.

Weanalyzethreesurveytreatmentarmsthataremostrelevantforourstudy.

3

TreatmentArm1(TA1)administeredthestandardTZNPShouseholdquestionnairethatprovidesobservedconsumptionandpovertyestimatesandthatpermitstheestimationofallimputationmodelspresentedinDangetal.(forthcoming),whoseTanzania-specificportionsoftheresearchreliedonthedatafromthepreviousroundsoftheTZNPS.TableA.1inAppendixAshowseachofthe

modelsandtheirpredictors.TheTA1sampleconsistsof711households.

TreatmentArm2(TA2)administeredalightquestionnairethatincludes:

(1)“Coremodules”thatonlyincludethequestionsnecessaryforcomputingthepredictorsfor

adata-modestsubsetofmodelsthatarepresentedinDangetal.(forthcoming)-specifically

3Thetwoadditionaltreatmentarmsthatarenotdiscussed/usedinthispaperwere(a)thesamplethatwassubjecttoa14-daydiaryfordatacollectiononfoodconsumption,followingtheHBS2017/18methodology,andotherwiseidenticalnon-foodconsumptionexpendituremodulesvis-à-visT1;and(b)thesamplethatwassubjecttoamodifiedversionofT1questionnaire,specificallywithafoodconsumptionmodulethatwassetuptobealignedwiththeT1/TZNPSfoodconsumptionmodulebutwiththeHBSfooditemlist.

8

Models1,2,8and9,whichrequirepredictorsrelatedtohouseholddemographics,

employmentattributes,housingcharacteristics,assets,utilityexpenditures,and

(2)AshorterversionoftheTA1foodconsumptionmodule-withanidenticalset-up/setofquestionsbutwithareducedlistoffooditems–alignedwiththeearlierSurveyofHouseholdWelfareandLabourinTanzania(SHWALITA)andspecificallythe“shortlist”

treatmentarminthatstudy.

4

TheTA2foodconsumptionmoduleisslottedimmediatelyaftertheTA2coremodules,covering26itemsoutofthe71itemsincludedinTA1.

5

Theseselecteditemsaccountfor69percentofthemonetaryvalueoffoodconsumptioninTA1,indicatingthatthereducedlistoffoodconsumptionitemsunderTA2missesoutonaconsiderableshareofthefoodexpenditurecomparedtothefullTA1foodconsumptionmodule.Asdiscussedlater,TA2dataonfoodconsumptionareusedtoestimateanadditionalimputationmodel,namelyModel3aspresentedinDangetal.(forthcoming),whichincludeshouseholdfoodconsumptionexpendituresasapredictor.TheTA2sampleconsistsof701households.TableA.2inAppendixApresentsexpendituresonthese

foodcategoriesforTA2andTA3incomparisonwiththosefromTA1.

Finally,TreatmentArm3(TA3)administeredanalternativelightquestionnairevariantthat

includes:

(1)ThesameTA2coremodulesthatallowfortheestimationofModels1,2,8,and9as

presentedinDangetal.(forthcoming),

4FormoreinformationregardingSHWALITA,pleaseseeBeegleetal.(2012)andvisit

https://www.uantwerpen.be/en/staff/joachim-deweerdt/public-data-sets/shwalita/#introduction.

5TA2covers13individualfooditemsand4itemcategoriescorrespondingto13itemsonTA1.The13individualitemsinclude:rice(husked);maize(grain);maize(flour);milletandsorghum(flour);cassavafresh;cassavadry/flour;sweetpotatoes;cookingbananasandplantains;sugar;beefincludingmincedsausage;dried/salted/cannedfishandseafood;freshmilk;cookingoil.The4groupeditemcategories(covering13itemsinTA1)include:peas,beans,lentils,andotherpulses;Onions,tomatoes,carrots,andgreenpeppers;Spinach,cabbage,andothergreenvegetables;andFreshfishandseafood.

9

(2)Anaggregatedfoodconsumptionmodulethatcorrespondstothe“collapsedlist”treatment

armintheSHWALITAstudy,and

(3)Aseriesofshort,aggregatednon-foodconsumptionexpendituremodulesthatwereinformedbythevariantsfromtheSHWALITAstudybutwererefinedinsomeinstancestobetteralignwiththeCOICOPcategories(UnitedNations,2018),related,forinstance,

toeducation,health,andutilitiesexpenditures.

TheTA3collapsedfoodconsumptionmoduleisslottedimmediatelyafterthecoremodules,coveringall12broadfoodcategories(includingalcoholicbeverages),andonlyaskingtherespondenttostatethemonetaryvaluethattheconsumedquantityoftotalfoodinthatcategorywouldhavecost,haditbeenpurchased.

6

TA3non-foodconsumptionexpendituremodulesarethenslottedimmediatelyaftertheTA3collapsedfoodconsumptionmodule,andtogether,thesesetsofmodulespermittheestimationofModels3and4aspresentedinDangetal.(forthcoming).

TheTA3sampleconsistsof698households.

ThesedataareinturncomplementedwiththedatafromthenationallyrepresentativeTZNPS2020/21and2019/20rounds,whichareusedasbasesurveystoestimatetheimputationmodelsthatareinturnappliedtoeachtreatmentarm.Themainresultsarebasedonthe2020/21round,whileAppendixAincludesconsistentfindingsbasedonthe2019/20round,asdiscussedbelow.TheTZNPSisamulti-topic,nationallyrepresentativelongitudinalhouseholdsurveythathasbeenimplementedbytheNBSsince2008,withfinancialandtechnicalsupportfromtheWorldBankLivingStandardsMeasurementStudy–IntegratedSurveysonAgriculture(LSMS-ISA)project.

ThequestionsforthepovertypredictorsrequiredfortheestimationofModels1,2,8and9are

6TA3covers:cerealsandcerealproducts;starches;sugarandsweets;pulses,dry;nutsandseeds;vegetables;fruits;meat,meatproducts,fish;milkandmilkproducts;oilandfats;spicesandotherfoods;alcoholicandnon-alcoholicbeverages.

10

identicalacrosstheconsumptionexperimentaswellastheTZNPS020/21and2019/20rounds.Thesamplesizeswere4,644in2020/21(followingupwithapanelsamplethatwasfirstinterviewedduringthe2014/15round)and1,179householdsin2019/20(followingupwithasubsetofanolderpanelsamplethathadbeeninterviewedaspartoftheTZNPS2008/09,2010/11and2012/13).Asdiscussedabove,therearedifferencesintermsoffoodandnon-foodconsumptionmodulesthatwereintroducedinTA2andTA3tounderstandthepotentialforusing

lighterversionofthesemodulestoobtainaccuratepovertypredictions.

Finally,inTA1andtheTZNPS2020/21and2019/20rounds,thetotalconsumptionistakentobethesumoffood(consumedatandawayfromhome)andnon-foodconsumption(health,education,utilities,furnishingandhouseholdexpenses,transport,communication,retreats,andother).Weprovidemoredetaileddiscussiononthefoodandnon-foodconsumptionexpenditure

aggregatesfortheTZNPSsandthethreeTAsinAppendixB.

2.2.Descriptivestatistics

Wespatiallyandinter-temporallydeflatealltheconsumptionaggregatesinthethreeTAsandtheTZNPSs.ThespatialandtemporalpricedifferencesinnominalhouseholdconsumptionexpenditureswithinallsurveyroundsarecorrectedusingFisherpriceindices.Thesepriceindicesareestimatedwithineachsurveyroundbystratumandquarter(ormonth,inthecaseofthe

experiment),andthebaseperiodineachestimationcomprisestheentireperiodofeachround.

Theacross-surveyintertemporaldeflationisinturnconductedusingtheannualinflationseriesforvariousconsumptiongroups,asobtainedfromtheWorldBankGlobalDatabaseofInflation

(Haetal.2023).

7

Specifically,foodexpenditureisdeflatedusingtheconsumerpriceinflationfor

7Toaccessthedatabase,visit:

/en/research/brief/inflation-database.

11

foodandnon-alcoholicbeverages,whileutilitiesexpenditureisdeflatedusingtheconsumerpriceinflationforenergy(capturinghousing,water,electricity,gasandotherfuels).Remainingnon-foodconsumptionexpenditureisdeflatedusingtheheadlineaverageconsumerpriceinflation.The

year2022isusedasthebaseyear.

Hence,theconsumptionexpendituresasmeasuredinourexperimentin2022aretakenintheirnominalvalues,whiletheexpendituresinpreviousroundsaredeflated.TheexpenditurevalueselicitedduringtheTZNPS2020/21,conductedbetweenDecember2020andJanuary2022,aredeflatedinaccordancewiththe2021-2022inflation.Similarly,theexpenditurevalueselicitedduringtheTZNPS2019/20,conductedbetweenJanuary2019andJanuary2020,aredeflatedinaccordancewiththe2020-2022inflation.Inwhatfollows,allexpendituresarereportedinyear-2022Tanzanianshillings(TSH),andtotalannualconsumptionperadultequivalentiscompared

totheTZNPS2020/21povertylinedeflatedtopricesin2022.

Table1providesdescriptivestatisticsforTZNPS2020/21,2019/20roundsandforeachofthesurveytreatmentarms,coupledwiththeresultsfromthetestsofmeandifferencesamongtheTAs.The“good”newsisthatacrosstreatments,comparisonsoftheprospectivepovertypredictorsthatarerequiredforModels1,2,8and9largelydonotrevealstatisticallysignificantdifferences.Theonlyexceptionsareparticipationinwagework,andbicycleownership,betweenTA1andTA2;radioownership,urban–ruralresidenceandutilityexpenditures(thoughwithmarginaldifferences)betweenTA2andTA3;andaccesstopipedwater,betweenTA1andTA3.Thesefindingsare

instarkcontrastwiththoseofKilicandSohnesen(2019)

8

anddonotraiseflagsregardingthe

8KilicandSohnesen(2019)reportonarandomizedsurveyexperimentthatwasconductedinMalawiin2016andthatshowsthatobservationallyequivalent,aswellasidentical,householdsinfactanswerthesamequestionsdifferentlydependingonwhethertheyareinterviewedwithashortquestionnaireoritslongercounterpart.Theauthorsfindlargeandstatisticallysignificantdifferencesinreportingacrossarangeoftopicsandquestiontypes,whichcanleadtoadifferenceof3to7percentagepointsinpredictedpovertyestimates,dependingontheimputationmodel.Theauthors,however,demonstratethattheimputationmodelusingonlythepredictorsthatareelicitedpriortothe

12

sensitivityofmeasurementtothedifferencesinlengthandcomplexitybetweenthebasesurveyandtargetsurveyquestionnairedesign,providedthattheidenticalquestionsareutilizedacrossthesurveys.Itisthusreasonablethatchangesinthedistributionsofthepredictorvariablesovertimeforthesefourmodelscancapturethechangeinthepovertyratebetweentherounds(i.e.,satisfying

Assumption2inourimputationframeworkdiscussedinthenextsection).

Ontheotherhand,Table1alsoshow