《亚开行-家庭通胀预期调查中的非响应偏差(英)-2023.12_市场营销策划_重点报告20230120.docx》由会员分享,可在线阅读,更多相关《亚开行-家庭通胀预期调查中的非响应偏差(英)-2023.12_市场营销策划_重点报告20230120.docx(47页珍藏版)》请在三一办公上搜索。
1、NONRESPONSEBIASINHOUSEHOLDINFLATIONEXPECTATIONSSURVEYSNO.706ADBECONOMICSDecember2023WORKINGPAPERSERIESADBEconomicsWorkingPaperSeriesNonresponseBiasinHouseholdInflationExpectationsSurveysMeltem Chadwick, Rennae Cherry, and Jaqueson K. GalimbertiNo. 706 I December 2023The ADB Economics Working PaperSe
2、ries presents research in progress to elicit comments and encourage debate on development issues in Asia and the Pacific. The views expressed are those of the authors and do not necessarily reflect the views and policies of ADB or its Board of Governors or the governments they represent.Meltem Chadw
3、ick (meltem.chadwickseacen.org) is a senior economist at the South East Asian Central Banks Research and Training Centre. Rennae Cherry (rennae.cherryrbnz.govt.nz) is an economic analyst at the Reserve Bank of New Zealand. Jaqueson K. Galimberti (jgalimbertiadb.org) is an economist at the Economic R
4、esearch and Development Impact Department, Asian Development Bank.缸_CBMCreativeCommonsAttribution3.0IGOlicense(CCBY3.0IGO)2023AsianDevelopmentBank6ADBAvenue.MandaluyongCity,ISSOMetroManila.PhilippinesTel+63286324444;Fax+63286362444Somerightsreserved.Publishedin2023.ISSN2313-6537(print),2313-6545(ele
5、ctronic)PublicationStockNo.WPS230552-2DOI:WPS230552-2TheviewsexpressedinthispublicationarethoseoftheauthorsanddonotnecessarilyreflecttheviewsandpoliciesOfcheAsianDevelopmentBank(ADB)oritsBoardofGovernorsorthegovernmentstheyrepresent.ADBdoesnotguaranteetheaccuracyofthedataincludedinthispublicationand
6、acceptsnoresponsibilityforanyconsequenceoftheiruse.ThementionofspecificcompaniesorproductsofmanufacturersdoesnotimplythattheyareendorsedorrecommendedbyADBinpreferencetoothersofasimilarnaturethatarenotmentioned.Bymakinganydesignationoforreferencetoaparticularterritoryorgeographicarea,orbyusingtheterm
7、countryinthispublication,ADBdoesnotintendtomakeanyjudgmentsastothelegalorotherstatusofanyterritoryorarea.ThispublicationisavailableundertheCreativeCommonsAttribution3.0IGOlicense(CCBY3.0IGO)Byusingthecontentofthispublication,youagreetobeboundbythetermsofthislicense.Forattribution,translations,adapta
8、tions,andpermissions,pleasereadtheprovisionsandtermsofuseathttps:/www.adb.0rg/terms-use#0penaccess.ThisCChtmaterialsinthispublication.Ifthematerialisattributedcoanothersource,pleasecontactthecopyrightownerorpublisherofthatsourceforpermissiontoreproduceit.ADBcannotbeheldliableforanyclaimsthatariseasa
9、resultofyouruseofthematerial.Pleasecontactpubsmarketingadb.orgifyouhavequestionsorcommentswithrespecttocontent,orifyouwishtoobtaincopyrightpermissionforyourintendedusethatdoesnotfallwithintheseterms,orforpermissiontousetheADBlogo.CorrigendatoADBpublicationsmaybefoundatNoteInthispublication.NZSrefers
10、toNewZealanddollars.ABSTRACTThispaperusesmicrodatafromtheReserveBankofNewZealand,sHouseholdInflationExpectationssurveytoobtainanaccuratereadofhouseholds*trueinflationexpectationsbystudyinghowdifferentdemographicgroupsrespond(ordonotrespond)totheinflationexpectationsquestioninthesurvey.Wefindnonrespo
11、nsesleadtosubstantialunderrepresentationofsomedemographicgroupsinthesurvey:young,female,low-income,andminorityethnicgroupshavelowerresponserates.Howthesurveyisconductedalsoaffectsitemresponserates.Thesurveyresponseratesincreasewhenthesurveyisconductedonlineandwheninflationratesdeviatefromthecentralb
12、ank,stargetrange.Usingasampleselectionmodel,weassesswhetherthesurveyhasitemnonresponsebiasbycomparingthedemographiccharacteristicsofrespondersandnonresponders.Afteraccountingforselection,wefindthatobserveddifferencesininflationexpectationsbygender,ethnicity,andincomedecreasesubstantially,whilediffer
13、encesbyageincrease.Wequantifyanddemonstratehowtoadjustaverageinflationexpectationsforbiascausedbyitemnonresponse.Weshowthatthereisapositivebias,andtheaggregateinflationexpectationseriesshiftsdownaftertheadjustment.Keywords:inflationexpectations,householdsurveys,itemnonresponse,demographicheterogenei
14、tyJELcodes:C83,D84,E31,E71Paperpresentedatthe2023BISAnnualWorkshopoftheAsianResearchNetwork,the2023ConferenceoftheNewZealandAssociationofEconomists,the5thBankofIndonesia-ADBInstitute-AsiaPacificAppliedEconomicsAssociationWorkshop,andattheADB/ERDISeminarSeries.WethankSeunghyeonLeeandBhaveshGargforthe
15、insightfuldiscussionsofourpaperandotherparticipantsofthepresentationsaboveformanycomments.Theviewsexpressedinthispaperarethoseoftheauthorsanddonotnecessarilyrepresenttheviewsoftheircorrespondinginstitutionalaffiliations.Correspondingauthor:JaquesonK.Galimberti(jgalimbertiadb.org)1 IntroductionMeasur
16、esofinflationexpectationsareofparamountimportancetomonetarypolicy.Inflationexpectationsarekeydeterminantsofpricesthroughtheforward-lookingbehaviorofhouseholdsandfirmswhilealsounderpinningwhetherbeliefsareanchoredtothecentralbanksinflationtarget.Surveysarethemostcommoninstrumentusedtomeasureinflation
17、expectations.Whereasthefocusisonaggregatemeasures,suchasaveragesandmedians,itiscrucialtounderstandtheaccuracyofsurveydataintrackingthepopulationsbeliefsaboutinflation.Inthispaper,weusemicrodatafromtheReserveBankofNewZealands(RBNZ)HouseholdInflationExpectationssurveytoinvestigatehownonresponsestoinfl
18、ationexpectationsquestionscanbiasmeasurementsobtainedfromsuchsurveys.Ourmaincontributionisquantifyingnonresponsebiasininflationexpectationsandproposingamethodtoadjustaverageexpectationsfortheeffectsofnonresponsebias.First,weshowthatcertaindemographicgroupstendtonotrespondmorefrequentlythanthetargetp
19、opulationwhenaskedabouttheirinflationexpectations.Thesenonresponsesamounttoabout44%onaveragethroughoutoursample:respondentswhoareyoung,female,havelowincome,orcomefromminorityethnicgroupsendupunderrepresentedduetononresponses.Becausethesenonresponsesarenotrandom,aggregatemeasuresofinflationexpectatio
20、nsderivedfromthesampleofrespondentscanbebiased.Weproposeasampleselectionmodeltoadjustfornonresponsebiasininflationexpectations.Figure1presentstheevolutionofmeanone-year-aheadinflationexpectationsthroughoutoursampleperiod,fromthesecondquarter(Q2)of1998toQ42022.Wefindthatnonresponsesartificiallyraisea
21、verageinflationexpectationsbyabout0.30percentagepoints(pp).Throughout this paper, we focus on so-called item nonresponses to the specific survey question on inflation expectations instead of unit nonresponses to the whole survey. The use of survey weights corrects the incidence of unit nonresponse (
22、see Meyer et al., 2015, for further discussion).Anotherimportantfindingrelatestotheeffectofsurveymodeonnonresponses.StartinginQ32018,thesurveychangedfrombeingconductedbytelephonetoonlinemode.Wefindthatthischangesignificantlyaffectedtheincidenceofnonresponsetotheinflationexpectationsquestion.Theavera
23、geofnonresponsesdecreasedtoabout24%sincethesurveymovedtoonlinemode.AsevidencedinFigure1,thischangealsosignificantlyreducedtheeffectofnonresponsebiasinestimatingaverageinflationexpectations.Accordingtoourestimates,themovetoonlinemodegenerallyreducedthegapsinnonresponsesacrossthedifferentpopulationgro
24、ups.Inotherwords,conductingthesurveyonlinehasmadeitmoreinclusiveforpreviouslyunderrepresenteddemographicgroups.Anotherimportantfindingisthatnonresponsebiasalsodependsontheleveloftheinflationrateatthetimethesurveyisconducted.Wefindthatresponseratestendtoincreasenon-linearlywhenthepreviousquartersinfl
25、ationisawayfromthecentralbank,stargetrange.We used lagged values of the inflation rate because these are the latest available information to respondents at the time the surveys are conducted each quarter.Forexample,aninflationrateincreasefrom2%to7%increasestheaverageresponseprobabilityby12%,whilethi
26、sprobabilitybarelychangesoveraninflationrangebetween0%and4%.ThiseffectisalsoapparentinFigure1,wheretheadjustmentfornonresponsebiasdecreasesinmagnitudesincetheonsetoftherecentincreaseininflationrates.Ourmethodologicalapproachisbasedonsampleselectionmodels.Wefirstidentifypotentialdeterminantsofrespons
27、estotheexpectationsquestionbyestimatingProbitregressionsonseveraldemographicvariablescollectedwiththesurvey.Probitregressionsmodeltheprobabilityofanevent,inourcase,aresponsetotheinflationexpectationsquestion,usingasetofexplanatoryvariables.Theseestimateshelpusdefineaselectionequation,whichdetermines
28、whenarespondentislikelytoanswertheFigure1:TheRBNZSurveyofHouseholdInflationExpectations20002005201020152020UnadjustedaverageAdjustedaverageIIDifference(Ihs)StartofonlinemodeRBNZ=ReserveBankofNewZealand,Ihs=left-handside.Notes:Thelinesdepictquarterlyaveragesofone-year-aheadinflationexpectationsfromth
29、eRBNZhouseholdinflationexpectationssurvey.Theunadjustedaverageisarawweightedaverageacrossrespondents,whiletheadjustedaverageiscalculatedusingourmethodologytoadjustfornonresponsebias.Thegapsinthefourthquarter(Q4)of2008andQ2andQ32010areduetomissingobservations.Thedashedlinedepictswhenthesurveyswitched
30、toonlinemodeinQ32018.BeforeQ32018,thesurveywasconductedbytelephone.Source:ReserveBankofNewZealandHouseholdInflationExpectationsSurvey.inflationexpectationquestiondependingontheircharacteristics.Our regressions include additional macro variables, such as lagged inflation and lagged inflation squared,
31、 an annual linear trend, seasonal dummy variables, and a dummy variable accounting for the change to online mode.Weconsiderseveralspecificationsforthesetofexplanatoryvariablesdependingontheiravailabilityacrossthesampleperiod.Wefindthattheeffectsofvariablesincludedinourbaselinespecificationgender,age
32、,region,ethnicity,income,andemploymentarerobustacrosssampleperiodsandtotheinclusionofmoreinformationsuchasoccupation,exposuretogroceryshopping,whethertherearechildreninthehouseholdandtypesofhomeownership.WethenstudyinflationexpectationsbiasaccountingfornonresponsesusingaHeckmanselectionmodel(Heckman
33、,1974,1979).TheHeckmancorrectionisbasedontheinsightthatsampleselectioncanbeviewedasaformofomittedvariablebiasspecifically,themethoddrawsonProbitestimatesoftheselectionequationtocalculatetheinverseMillsratio,whichisthenusedasanadditionalexplanatoryvariableintheregressionwithmissingobservations.The He
34、ckman selection model estimates the selection equation and uses the predicted probabilities from the selection equation as a correction term in the outcome equation. This correction term, known as the inverse Mills ratio, accounts for the selection bias by adjusting the coefficients in the outcome e
35、quation.Comparedtoestimatesthatdonotaccountforselection,wefindthatmostdifferencesinbiasacrosssubgroupsturninsignificantafteraccountingfornonresponsebias.Observeddifferencesininflationexpectationsbygender,ethnicity,andincometurninsignificantordecreasesubstantiallyinmagnitude.Theonlyexceptionisage,whe
36、reolderindividualstendtoover-predictinflationmorethantheyoung,whichisstrongerafteraccountingforselection.Forrobustnesspurposes,wealsoconsiderdifferentestimationmethodologiesandfindthatourestimatesarenotsensitivetothechoiceofestimationmethod.The Heckman sample selection model can be estimated using e
37、ither a maximum likelihood approach or the original two-step approach (see, e.g., Puhani, 2000, for more details). Here, we extend those methods to account for survey weights and derive weighted estimates.Ourproposedadjustmenttothecalculationofaverageinflationexpectationsgoesalongsimilarlines:averag
38、eindicescanbeeasilyobtainedbyrunningaregressionofsurveyinflationexpectationsonquarterdummyvariables.AfterincludingourbaselineestimatesoftheHeckmancorrectiontermasanadditionalvariableinthisregression,weobtainaverageinflationexpectationsadjustedfornonresponsebiasthesearetheadjustedaverageexpectationsr
39、eportedinFigure1.Thesimplicityofthisapproachmakesitattractiveforoperationalpurposes:toobtainupdatedestimateseveryquarter,allthatisrequiredarenewestimatesoftheinverseM川Sratio,whichcanbeeasilycomputedfromthepre-fittedProbitmodel.Indeed,thefactthattheProbitmodelestimatesarerelativelystableacrosssub-sam
40、plesindicatesthattheadjustmentisunlikelytoundergosevererevisionsovertime.Finally,ourestimatesusesurveyweightstoaccountforunitnonresponsebiasarisingfromdifficultyinobtainingarepresentativepopulationsurveysample.Althoughtheseweightscannotaccountfordeterminantsofnonresponsestotheinflationexpectationsqu
41、estion,wealsofindthattheyarerelevantfortheanalysisofinflationexpectationsbias.1.1 RelatedLiteratureandSurveysThispaperrelatestothebroaderliteratureontheheterogeneityofinflationexpectations.LookingatasamplefromtheUSMichiganSurveyofConsumers,BruinedeBruinetal.(2010)corroboratefindingsthatdemographicva
42、riablesplayasignificantroleindetermininginflationexpectations.PfajfarandSantoro(2010)documentpervasiveheterogeneityinforminginflationexpectationsinthatsamesurvey.MalmendierandNagel(2016)showconsumerinflationexpectationsalsovarywithageduetolearningfromexperience.D,Acuntoetal.(2023)documentthathouseho
43、ldinflationexpectationsareupward-biasedandsystematicallydifferentacrossgender,income,education,andrace.OurfindingsbasedonNewZealanddataaddtothisliteraturebyshowingthatsomedemographicdifferencesareaproductofnonresponsebias.Whenaccountingforselection,wefindthatdifferencesbygender,ethnicity,andincomede
44、creasesubstantially.Ourfindingsabouttheunderrepresentationofsomedemographicgroupsareconsistentwithpreviousstudiesintheliterature.ExploringUnitedKingdom(UK)surveymicrodata,anearlystudybyBlanchflowerandMacCoiIIe(2009)alsofoundsignificantnonresponsebiasfromyoung,female,andlow-incomerespondents.Leung(20
45、09)reportedsimilarfindingswithashortersamplefromtheRBNZhouseholdsurvey.Ourfindingthatonlinesurveymodecanattenuatenonresponsebiasisconsistentwithpreviousstudies.BruinedeBruinetal.(2017),forexample,findthatonlinesurveysachievehigherresponseratestotheinflationexpectationsquestionthanface-to-faceSUrVeys
46、.Noneofthepapersaboveprovidedanadjustmentforthenonresponsebiasininflationexpectationssurveys.Onestandardapproachtodealwithnonresponsesistoreplacethenonresponsesormissingobservationsbyimputation.TheUnitedStates(US)MichiganSurveyofConsumers(MSC),forexample,usesdistribution-basedimputationstoreplaceuDo
47、n,tKnowresponseswithrandomdrawsfromadistributionthatmatchesthepropertiesofobserveddata(Curtin,1996).However,thisimputationmethoddoesnotconsiderthesocio-demographiccompositionofthesampleofrespondents.Itcan,therefore,reinforcetheeffectsofselectionbiasinanalyzingthesurveyofexpectationsdata.Morebroadly,
48、theissueofitemnonresponsehasreceivedincreasedattentioninrecentrelatedstudies.FocusingonaUSlongitudinalsurveyofprofessionalforecasters,Burgi(2023)comparesmethodsforfillinginmissingobservationsduetosurveyattritionnaturally,thisisadifferentproblemthanwhatwefacewithrepeatedcross-sectionalsurveysastheone
49、westudyhere.Analternativeapproachforthatcaseinvolvestheuseofsurveydesignfeatures.McGovernetal.(2018)exploreHIVtestingdatatoshowthatrandomizedincentivesorsurveyinterventionscanprovideidealselectionvariablestocorrectfornonresponsebias.Comerford(2023)proposesusingaverbalquestiontodealwithnonresponsebiasfoundininflationexpectationsderivedfromdensityforecasts.Exante,thesemethodsprovidevitalinsightsintosurveydesign.However,therequir