世界银行-使用重新加权和贫困预测模型校正电话调查贫困估计中的抽样和无响应偏差（英）-2023.12..docx

PaUOLnnValnsoosQ.2-qndPolicyResearchWorkingPaper10656CorrectingSamplingandNonresponseBiasinPhoneSurveyPovertyEstimationUsingReweightingandPovertyProjectionModelsP Z'OfnV alnsoosQ.2-qndKexinZhangShinyaTakamatsuNobYoshidaworldBankgroupxt-rPovertyandEquityGlobalPracticeDecember2023PolicyResearchWorkingPaper10656AbstractTo monitor the evolution of household living conditions during the COVID-19 pandemic, the World Bank conducted COVID-19 High-Frequency Phone Su,eys in around 80 countries. Phone sunfeys are cheap and easy to implement, but they have some major limitations, such as the absence of PoVCrty data, sampling bias due to incomplete telephone coverage in many developing countries, and frequent nonresponses to phone interviews. To overcome these limitations, the World Bank conducted pilots in 20 countries where the SUrVey ofWellbeing via Instant and Frequent Tracking, a rapid povem, monitoring tool, was adopted to estimate poverty rates based on 10 to 15 simple questions collected via phone interviews, and where sampling weightswere adjusted to correct the sampling and nonresponse bias.This paper examines whether reweighting procedures andmethodology can eliminate the bias in povcrtr estimation based on the COVID-19 High-Frequency Phone Surveys. Experiments using artificial phone survey samples show that (i) reweighting procedures cannot fully eliminate bias in povert estimates, as previous research has demonstrated, but (ii) when combined with SUrVey of Wellbeing Via Instant and Frequent Tracking PoVerty projections, they effectively eliminate bias in poverty estimates and otherstatistics.'hispaperisaproductofthePovertyandEquityGlobalPractice.ItispartofalargereffortbytheWorldBanktoprovideopenaccesstoitsresearchandmakeacontributiontodevelopmentpolicydiscussionsaroundtheworld.PolicyResearchWorkingPapersarealsopostedontheWebatTheauthorsmaybecontactedatkzhang2worldbank.org.TbePolieyResearchWorkingPaperSeriesdisseminatestfjefindingsofuforkinprogresstoencouragetbeexchanfofideasaboutdeivlopmentissues.Anobecihfeoftheseriesisto般thefindingoutquickfy,evifthepresentationsarelessthanJulbfpolished.ThepaperscanythenamesOftbeauthorsandshouldbecitedaccordingly.Thefinding,interpretations,andconclusionsexpressedinthispaperareentircbfthoseoftheaut!,wrs.TheydonotnecessarilyrfpnsvnttheviewsoftheInternationalBankforKecoftstnictionandDert,o>menf/WorldBa欣anditsafftiate<!oranifalions,orthoseoftheFixecuthvDireetarsoftheWorldRankort!epivrtimeHtstheyrepresent.ProducedbytheResearchSupport,eamCorrectingSamplingandNonresponseBiasinPhoneSurveyPovertyKcxinZhang,ShinyaTakamatsuandNobuoYoshidaKeywords:Phonesurveys,Weighting,Povertyprojections/estimation,CorrectionofsamplingbiasandnonresponsebiasJELcodes:132,C83,C81*TheauthorswouldliketothankKristenHimelein,XueqiLi,AzizAramanov,ChristinaWieser,AnnaLUiSaPaffhausenf（）rtheirsupportf（）rthisstudy.Theauthorsalsogratefultopeerreviewers,KevinMcGeeandDanielGerszonMahler,andallparticipantsattheQualityEnhancementReviewmeetingheldattheWorldBankonNovember2021,UNECEConferenceofEuropeanStatisticians2022,andparticipantsatIARIWConference2022forhelpfulcomments.a）KexinZhang,SchoolofAgriculturalEconomicsandRuralDevelopment,RenminUniversityofChina,zhangkexin629ShinyaTakamatsu,PovertyandEquityGlobalPractice,WorldBank,stakamatsuworldbank.orgNobuoYoshida,PovertjrandEquityGlobalPractice,WorldBank,nyoshidaworldbank.orgI. IntroductionandbackgroundPhoneinterviewingopensnewpossibilitiesforempiricalresearchinsocialsciences.ItgreatlyreducesthecostsofconductingSUrVeySbecauseallcommunicationoccursWithOUttheface-to-faceinteractionCUStOmaryintraditionalsurveys.WiththeoutbreakoftheCOVID-19pandemiclimitingface-to-faceinterviews,phoneSUrVeySaremoreprevalentamongacademicinstitutions,surveycompanies,andindividualresearchersforindividual-levelandhousehold-leveldatacollection.Totrackhouseholds,livingconditionsonatimelybasisduringapandemic,theWorldBanklaunchedtheCOVID-19High-FrequencyPhoneSurveys(HFPS),whichhavebeencarriedoutinaround80countriessinceMarch2020.TheseSUrVeySallowpolicymakerstomonitorawidevarietyofsocioeconomicindicatorsinatimelyandfrequentmanner.Phonesurveyshaveshortcomings.First,itisdifficulttomonitorpovertyusingtheCOVID-19IIFPS.TheCOVID-19HFPSdoesnotcollectconsumptionorincomedata,whicharenecessaformeasuringpovertyrandinequalityunderthetraditionalPOVCrtymonitoringapproach.Itistime-consuming,costly,andcomplextocollectsuchdata.Aninterviewermustaskhouseholdsanumberofcomplexquestionsregardingtheirrecentconsumption,expenditures,anddetailedincomecomponents.Suchanintemewrequiresatleast30minutesanduptotwohours,evenbywell-trainedandqualifiedinterviewers.Itisalsochallengingtoadministertheintenriewindevelopingcountrieswheretelephoneconnectionsarenotalwaysstableenoughtocompletesuchalonginterview.AsolutiontoChallenge1:.S'IF7F7'asarapidpovertymonitoringtoo!InMarch2020,theWorldBanklaunchedapilotfortheuseofarapidpovertymonitoringtool,SurvreyofWellbeingviaInstant,FrequentTracking(SWIFT),toestimatemonetarypovertyusingtheC()V1-19HFPS.SwIFTadds10to15simplequestionstotheCOVID-19HFPSquestionnaire,andtheseadditionalquestionstake3to5minutestoask.Basedonhouseholds,responsestothesequestions,theSWIFTmethodcanbeadoptedtoimputepovertyandinequalityratesamongthepopulationofinterest.SWIFTwaschosenamongmanypovertyprojectionmethodologiesbecauseithasbeenthoroughlytestedandreviewedbyexpertsinsideandoutsidetheWorldBank.TheresultsareencouragingthedifferencebetweenthetrueandSyrlFT-basedpovcrt,projectionsis1.2percentagepointsonaverageandiswithin+/-2percentagepoints,exceptforonecase(seemoredetailsinAppendix5orYoshidaetal.,2022a).Thepilotwasconductedin20countriesaroundtheworld,andtheresultsofsevencountriesaresummarizedinYoshidaetal.(2022b).Second,thesampleoftheC()VI-19HFPSisunlikelytobenationallyrepresentative.Inmanydevelopingcountries,thecoverageofmobilephonesorlandlinesislimited.Poorhouseholdsthatdonotownaphoneareexcludedfromphoneinterviews.Inaddition,phoneinterviewsusuallyfaceamuchhigherrateofnonresponsethanin-personinterviews.Suchnonresponsesareusuallyconcentratedamongurbanresidentswhoaretoobusytoparticipateintheinterviews.Hence,thesamplecollectedinaphonesurreyisoftenfarfromnationallyrepresentativeandcanresultinbiasintheprojectionsofpoverty.AsolutiontoChalkn2:ReWeigbtiHg-samplingadjustmentstoaddresssamplingbiasesManyapproachesforadjustingsamplingweights,orurcwcighting*havebeendevelopedtocorrectforvarioustypesofsamplingbias.Thispaperclassifiesthemintotwomajorcategories:(1)propensityscoreweighting(W)and(2)*non-PSWadjustments.”PropensityscoreweightingisamethodforadjustingsamplingweightsbasedonpropensityscoresandwasdevelopedfromthepropensityscorematchingproposedbyRubinandRosenbaum(1983and1984).Inthecontextofcausalinferences,propensityscorematchingmakesthecontrolandtreatmentgroupscomparable,minimizingbiasinestimatingtreatmenteffects.Unlessthesamplesofthecontrolandtreatmentgroupsareselectedrandomly(whichisusuallynotthecasewithobsen,ationaldata),baselinecharacteristicsmayexhibitdifferencesbetweenthetwogroups.Propensityscoresareoftenestimatedtoaddressthisissueofnoncomparability.Apropensityscoreistheprobabilityofselectingasamplehouseholdorindividualintothetreatmentgroupfromthepopulation,conditionaloncovariates.Ifasetofassumptionsissatisfied(seeRosenbaumandRubin1983and1984fordetails),acomparisonoftheweightedaverageoftheoutcomeindicatorsbetweenthecontrolandtreatmentgroupsprovidesanunbiasedestimateforthetreatmenteffect.Taylor(2000)andLee(2006)adoptedthepropensityscorematchingtechniquetoadjustsamplingweightsandcorrectforsamplingbiasinawebsurvey.First,theychoseareferencesurveyrepresentativeofthepopulationofinterest(e.g.,theentirepopulationofacountn,theurbanpopulation,orallrefugeesinacountry)。TheycombinedthereferenceandWCbsurveysandestimatedpropensityscoresusingthiscombinedsample.TheythendividedthecombinedsampleintoquintilesbasedonthepropensityscoresandadjustedthesamplingweightsofthewebsunreytoequatethesumoftheweightsofthewebsuneywiththoseofthereferenceSUrVeyineachquintile.Anothertypeofpropensit)scoreweightingmethodis*inverseprobabilityapproach/,whichusespropensityscorestoconstructtheoddsofasamplebeingselectedforareferencesurvey,andtheoddsarefurtherusedasasamplingweightofaWeborphoneSUrVey(MorganandTodd2008,SchaferandKang2008,andAustin2011).ThispaperclassifiesthesetypesofreweightingtechniquesbasedonpropensityscoresasPropensityScoreWeighting(PSw)methods.IftheassumptionsinRubinandRosembaum(1983and1984)arcsatisfied,thenSUmmarystatisticsapplyingPSW-basedweightsinaweborphoneSUrVeybecomerepresentativeofthepopulationofinterest.AsecondCategOryofreweightingtechniques,whichwerefertoasnon-PSWapproaches,matchesindicatorsbetweenaphoneorWebsurveyandareferencesurveyatamoreaggregatelevelthanatthehousehold/individuallevel.Thisgroupofapproachesincludesraking,post-stratification,andmaxentropy.Thenon-PSWapproachesselectasetofindicatorsandadjustsamplingweightssothatweightedaveragesoftheselectedindicatorsarcclosely/exactlymatchedbetweenaphonc/wcbsuncyandareferencesurvey.PSWandnon-PSWadjustmentsareoftenconductedtogether.WeconsidertwoadvantagesofthecombinedapplicationsofPSWandnon-PSWapproaches.First,f(>rPSWtoeliminatesamplingbias,asetofassumptions,likestrongignorability,needtobesatisfied,whichcannotbeeasilytested,forwhichnon-PSWservesasacomplementforweightingpurposes.Second,non-PSWapproachesmatchthemeansofkeyindicatorsbetweenaphone/webSUrVeyandareferencesurey,butthereisnoguaranteethatthedistributionsoftheseindicatorsarealsomatched.PSWapproaches,bycontrast,matchthedistributionofpropensityscores.Sincebothapproacheshavetheirownstrengths,conductingbothPSWandnon-PSWadjustmentsisreasonable.ThispaperinvestigateswhetherreweightingcancorrectthebiasofpovertyprojectionsproducedbytheSWIFTmethodology.Theperformanceofreweightingtechniquesdiffersbydataandtargetindicatorsthatwerematched,andthereisagreementintheliteraturethatreweightingtechniquesreducethebiasesintargetstatisticsyetdonoteliminatethem(Lee(2006)andDrezeandSomanchi(2023).DrezeandSomanchi(2023)createdbiasedsamplesbydroppingpoorerhouseholdsfromahouseholdSUrVeyandtestedwhetheranon-PSWreweightingtechnique(maximumentropyreweighting,ormaxentropy)canreducebiasesinpovertyratesandmeanhouseholdexpenditures.1thoughthebiasesinpovertyrateestimatesandmeansofhouseholdexpendituresdeclined,substantialproportionsremained.However,existingliteraturelacksanassessmentofhowwellreweightingtechniquescanreducethebiasesofpovcrtrprojectionsproducedbySWIFToranyotherPOvCrtyprojectionmethod.UsingphoneorWCbsurveystoestimatepovertynecessitatestheuseofPoVertyprojectionmethods.DrezeandSomanchi(2023)usedactualconsumptionandincomedataandshowedthatalargebiasinthepovertyrateandmeanhouseholdexpenditureremainsevenafterreweightingbutdidnotassessifreweightingcombinedwithpovcm,projectionmethodsiseffectiveinreducingthebias.Infoct,ourstudyfindsthattheperformanceofreweightinginestimatingpovertjrcanbeimprovedwhencombinedwithpovertyimputationmethodssuchasSWIFT.ExperimentsThispaperexamineswhether,andifso,towhatextent,acombinationofreweightingtechniquesandtheSWIFTpovertyprojectionmethodOlOgycaneliminatesamplingbiasesinpovertyestimatesbasedonbiasedsurveysamples.Toseethis,wcfirstusereferencehouseholdsurveysinRwanda,StLucia,andUgandaandconstructsubsamplesbyselectinghouseholdswithatleastamobileorlandlinephone.Thesesamplesaresubjecttosamplingbiasbyconstruction.WithoutreweightingandSWIFTpovertyprojections,thepovertyratesinthesesubsamplcsofphoneownersarclowerthanthoseinthefullsamples.WethenexaminewhetherreweightingandSWIFpovertyprojectionscancorrectfortheabovementionedbias.PhoneandwebSUrVeydatacollectionsfacesamplingandnonresponsebiases,buttheabovementionedexperimentsonlyfocusonsamplingbiasesthatarisefromunevenphoneownership.TounderstandrheabilityoftheSWIFTandreweightingtechniquestoadjustfbrnonresponsebias,thispaperconductsanadditionalexperimentusingthesampleofEthiopiaHigh-FrequencyPhoneSurveyround7(HFPS7),whichisasubsampleofthe2018/19EthiopiaSocioeconomicSUrVeyround4(ESS4).SincethissubsampleofESS4includesonlyphoneowners,itissubjecttosamplingbias.Also,sinceitincludesonlyhouseholdsofESS4thatrespondedtotheHFPS7,itisalsosubjecttononresponsebias.Usingthissubsample,WeconductthesameanalysisasabovetoassesswhetherreweightingtechniquesandSWIFTcancorrectthebiasinpovertyestimatesarisingfrombothsamplingandnonresponsebiases.Thispaperisorganizedasfollows.SectionIIintroducestheSWIFTpovertyprojectionmethodOlOgyandreweightingtechniques,andSectionIIIdisplaysresultsfromfourexperimentalstudies(SaintLucia,Rwanda,Uganda,andEthiopia).SectionIVconcludeswithanassessmentofacombinationofreweighringandtheSWIFT-basedpovertyprojectionmodelineliminatingthebiasinpovertyrestimationbasedonthef()urcasestudies.II. SWIFTandReweightingMethodologyII.1.SllHFTpovertyprojectionmetbodolo>SWIFTisanapplicationofSuney-to-Sureyimputationtechniques(S2S)tomonitorpovertrrapidly.SWIFTtrainsanimputationmodelinatrainingdatasetbyregressinghouseholdcxpcnditurcs/incomcsonpovertyproxies.Householdexpendituresandpovemrratesarethenimputedinanotherdataset,calledoutputdata,“bypluggingpovertyproxiesoftheoutputdataintothemodel.Figure1illustratestheprocess.TherearetwokeyassumptionsinthestandardSvnFTmethodology.First,therelationshipbetweenhouseholdincomeorexpenditureandpovertycorrelatesinrheoutputdatacanbeexpressedinequation(1):lnyh0=Xhoo+UhO(1)wherel0N(0,0)whereIny0referstoanaturallogarithmofthehouseholdincomeorexpenditureofhouseholdbintheoutputdatao.Xfioisa(Ze×1)vectorofpovertycorrelatesofhouseholdhntheoutputdata,o.0isa(k×1)vectorofcoefficientsofpovertycorrelates(x0).UfioreferstoaresidualandisoftenassumedtofollowanormaldistributionOfN(0,0O)JTheoutputdataincludesthePOVertyproxydata冲。%=butdocsnotinclude1Thisnormaldistributionandlinearity-canberelaxed.Forthesakeofexposition,normaldistributionisassumed.householdexpendituresZnyl0)¾=1whicharetobeimputed.Forthesakeofexposition,therelationshipisassumedtobelinear,butthisconditioncanberelaxed(Yoshidaetal.,2022a).Figure11llustrationofbowtheSI11FTworksDatasetJnykrtoh.Relationship如乂%;=M嫄-M九"饶MModelstabilityholdsImputationb:.i>7'i(,uImputeddata11nyL.xk>)fc.SNote:Authors*illustration.ThesecondkeyassumptionisthattherelationshipbetweenhouseholdexpendituresandPOVertyproxiesinthetrainingdataalsofollowsequation(1).Thisassumptioniscalled“modelStabiIity“implyingthatthemodeldocsnotchangefromthetrainingandoutputdata.SWIFTestimatesparametersinequation(15),(0,趣),andtheirdistributions,usingthetrainingdataset,drawthem(0r,偷)randomlyfromtheirestimateddistributions,andsubstitutesthemintoequation(1)toimputehouseholdexpendituresforallhouseholdsintheoutputdata.SWIFTrepeatsthisimputation(typically20-100times),resultingin20to100vectorsofhouseholdexpenditures)intheoutputdata(witheachvectorincludingtheexpenditurefc>rallhouseholds).Povertyrandinequalitymeasuresareestimated