PolicyResearchWorkingPaper10383IsClimateChangeSlowingtheUrbanEscalatoroutofPoverty?EvidencefromChile,Colombia,andIndonesiaShoheiNakamuraKseniyaAbanokovaHai-AnhDangShinyaTakamatsuChunchenPeiDilouProsperePovertyandEquityGlobalPractice&DevelopmentDataGroupMarch2023PublicDisclosureAuthorizedPublicDisclosureAuthorizedPublicDisclosureAuthorizedPublicDisclosureAuthorizedProducedbytheResearchSupportTeamAbstractThePolicyResearchWorkingPaperSeriesdisseminatesthefindingsofworkinprogresstoencouragetheexchangeofideasaboutdevelopmentissues.Anobjectiveoftheseriesistogetthefindingsoutquickly,evenifthepresentationsarelessthanfullypolished.Thepaperscarrythenamesoftheauthorsandshouldbecitedaccordingly.Thefindings,interpretations,andconclusionsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelopment/WorldBankanditsaffiliatedorganizations,orthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.PolicyResearchWorkingPaper10383Whileurbanizationhasgreatpotentialtofacilitatepovertyreduction,climateshocksrepresentaloomingthreattosuchupwardmobility.Thispaperempiricallyanalyzestheeffectsofclimaticrisksonthefunctionofurbanagglomerationstosupportpoorhouseholdstoescapefrompoverty.Com-bininghouseholdsurveyswithclimaticdatasets,thepanelregressionanalysisforChile,Colombia,andIndonesiafindsthathouseholdsinlargemetropolitanareasaremorelikelytoescapefrompoverty,indicatingbetteraccesstoeconomicopportunitiesinthoseareas.However,theclimateshocksoffsetsuchbenefitsofurbanagglomerations,asextremerainfallsandhighfloodriskssignificantlyreducethechanceofupwardmobility.Thefindingsunderscoretheneedtoenhanceresilienceamongtheurbanpoortoallowthemtofullyutilizethebenefitsofurbanagglomerations.ThispaperisaproductofthePovertyandEquityGlobalPracticeandtheDevelopmentDataGroup,DevelopmentEconomics..ItispartofalargereffortbytheWorldBanktoprovideopenaccesstoitsresearchandmakeacontributiontodevelopmentpolicydiscussionsaroundtheworld.PolicyResearchWorkingPapersarealsopostedontheWebathttp://www.worldbank.org/prwp.Theauthorsmaybecontactedatsnakamura2@worldbank.organdhdang@worldbank.org.IsClimateChangeSlowingtheUrbanEscalatoroutofPoverty?EvidencefromChile,Colombia,andIndonesia1ShoheiNakamura,WorldBankKseniyaAbanokova,WorldBankHai-AnhDang,WorldBankShinyaTakamatsu,WorldBankChunchenPei,BeijingNormalUniversityDilouProspere,OsakaUniversityJELClassification:R23;O18;I32Keywords:Migration;Urbanagglomeration;Poverty;ClimaticChange;Flooding1Correspondingauthor:ShoheiNakamura(snakamura2@worldbank.org).ThispaperwaspreparedasabackgroundpaperfortheWorldBankreportThriving:MakingCitiesGreen,Resilient,andInclusiveinaChallengingClimate.AnupdatedversionofthispaperwillbeforthcomingintheInternationalJournalofEnvironmentalResearchandPublicHealth.WewouldliketothankMarkRoberts,MeghaMukim,SaileshTiwari,LeonardoLucchetti,SomikLall,RinkuMurgai,andCarlosRodriguezCastelanfortheirvaluablecommentsonanearlierdraft.WearealsogratefultoBennyIstanto,ImamSetiawan,andLuisQuinteroforsharingandprocessingdatasetsforus.Thefindings,interpretations,andconclusionsexpressedinthispaperareentirelythoseoftheauthors.TheydonotnecessarilyrepresenttheviewsoftheInternationalBankforReconstructionandDevelopment/WorldBankanditsaffiliatedorganizationsorthoseoftheExecutiveDirectorsoftheWorldBankorthegovernmentstheyrepresent.21.IntroductionUrbanareasaroundtheworldareattractiveplacesforpeoplelookingforopportunitiesforabetterlife.About55percentoftheworld’spopulationlivedinurbanareasin2018andthatnumberislikelytogrowby68percentby2050(UnitedNations2019).Urbanagglomerationsspureconomicgrowththroughproductivitygainswithineconomicsectorsandstructuraltransformation(DurantonandPuga2004;Glaeseretal.1992;GlaeserandGottlieb2009;Michaels,Rauch,andRedding2012).Atthesametime,urbanresidentstendtobevulnerabletoclimaticandenvironmentalshockstriggeredbytheincreaseineconomicactivitiesandareoftenpushedbacktoorremaintrappedinpoverty(Hallegatteetal.2017).Withoutpropermitigationandadaptationmeasuresagainstclimatechange,thebenefitsofurbanizationcouldbenegated(MukimandRoberts2022).Inthiscontext,thispaperattemptstotestthefollowingtwohypotheses.Thefirstisthatpeoplearemorelikelytobecomeorstaynonpoorinlargerordensercities,comparedtosmallerorlessdenselypopulatedtowns.Thesecondhypothesisisthatlargeordensecitiesthataremoreexposedtoclimaticandenvironmentalshocksofferresidentsalowerchancetobecomeorstaynonpoor,comparedtocitiesofsimilarsizebutnotexposedtosuchshocks.Byempiricallytestingthosehypotheses,weinvestigatethefollowingkeyquestion:Doclimaticandenvironmentalshockshamperthekeyfunctionofurbanagglomerationsastheescalatoroutofpovertyinthedevelopingworld?Confirmingthisquestioniscriticallyimportantasitunderscorestheneedforpolicyinterventionstohelpachieveinclusiveandgreengrowththroughurbandevelopment.Wedevelopananalyticalapproachtoaddressingthequestionswithandwithoutpaneldatasetsbycombiningpanel(orsyntheticpanel)householdsurveyswithclimaticdatasets.Thesyntheticpanelmethodisausefulapproachtoanalyzingpovertydynamicswhenpanelhouseholdsurveydatasetsarenotavailable.Wedevelopsyntheticpaneldatasetsoutofrepeatedcross-sectionalhouseholdsurveysinChilebetween2011and2015andColombiabetween2008and2010.Wethenexaminetheassociationbetweenpovertychangesovertimeandcitypopulationsizeaswellastheheterogeneityoftheassociationbyfloodrisks.Wealsoanalyzeanothercountry,Indonesia,toapplyasimilaranalyticalframework.Insteadofdevelopingasyntheticpanel,however,weestimatetwo-wayfixed-effect(FE)regressionsonthefivewavesoftheIndonesiaFamilyandLifeSurveys(IFLS)toanalyzethevariationofprobabilitiesofpoorhouseholdsescapingfrompovertybyurbanagglomerationclassificationsandclimaticshocks/risks.TheIFLSspansfrom1993to2014over298districtsandtracksthesamehouseholdsovertime.Wefocusonfloodastheclimaticfactor,bymeasuringtherainfallanomalyandheavyrainmeasuredbytheStandardizedPrecipitationEvapotranspirationIndex(SPEI).Theresultsofouranalysissupportthehypothesisthatclimaticriskscouldunderminetheupwardmobilityfacilitatedbyurbanagglomerations.ThesyntheticpanelsinChilebetween2011and2015andColombiabetween2008and2010indicateareductioninurbanpovertyratesmeasuredbytheupper-middle-incomeinternationalpovertyline(US$5.5perdayin2011purchasingpowerparity[PPP]).Inthosecountries,7.4percentand4.1percentoftheurbanpoorescapedfrompovertyduringtheperiods,respectively.Theanalysisfindsthattheprobabilitiesofhouseholds’transitionfrompoortononpoorstatuswerepositivelycorrelatedwiththecitypopulationsizeinbothcountries.Moreimportantly,suchupwardmobilitywasobservedonlyinlargercitieswithlowfloodrisk.TheempiricalanalysisofIndonesiafindsthatcomparedtoruralareas,thechanceofgettingoutofvulnerabilityishigherby7.0percentagepointsinmetropolitancoresbasedontheresultswithhouseholdFEs.However,heavyrainfallandhighfloodriskdecreasethechanceofupwardmobilityinthecoresandurbanperipheriesofmetropolitanareas.3Ourpapercontributestotheliteratureonthenexusbetweenurbanizationandpoverty.Severalstudiesshowurban-ruralgapsinproductivity,wages,andamenities(seeLagakos2020forareview).Somestudiesfindearningsandwelfaregainsfromruraltourbanmigrationinthedevelopingworld,suchasruraltourbanmigrationinTanzania(Beegle,deWeerdt,andDercon2011),seasonalmigrationinBangladesh(Bryan,Chowdhury,andMobarak2014),andtheinterplayoflocationsandmigrantcharacteristicsindetermininggainsinChina(Combesetal.2020).Hamoryetal.(2021)analyzedpaneldatasetsinIndonesiaandKenya,findingthatalargepartofthemeasuredreturnsfrommigrationcamefromthesortingofmigrants.However,veryfewhaveanalyzedtheroleofclimatechangeasahindrancetourbanagglomerationastheurbanescalatoroutofpoverty.Therefore,weattempttoshedlightontheroleoftheclimaticandenvironmentalshocksontheurbanescalatoroutofpoverty.Therestofthepaperisorganizedasfollows.Section2explainstheframeworkwithadescriptionofourhypotheses.InSection3,weelaborateonthemethodologywiththedescriptionofthedatasetsandoureconometricapproach.Section4reportstheresultsofoureconometricanalysis.Finally,Section5concludeswiththesummaryoffindingsanddiscussion,policyimplications,andlimitations.2.FrameworkUrbanescalatoroutofpoverty:Hypothesis1Largerordenserurbanareaspotentiallyprovidepeoplewithampleeconomicopportunitiestoescapefrompoverty,throughbetteraccesstohigher-wagejobs,infrastructure,services,andsoon.Asaresult,moreurbanizedareastendtobecharacterizedbyhigherincomesandconsumption,higherproductivity,betteraccesstoservices,andhigherhumancapital.Earlierstudieshavefoundthatnominalwagesarehigherinlargerordensercitiesduetoproductivitygainsfromagglomerationeconomiesindevelopedcountries(GlaeserandMare2001;Melo,Graham,andNoland2009;Puga2010;RosenthalandStrange2004)andinthedevelopingworld(Chauvinetal.2017;Combesetal.2022;Grover,Lall,andTimmis2021;QuinteroandRoberts2022).Ruraltourbanmigrationcanbewelfareenhancing,giventhegapsinincomeandamenitiesinthedevelopingworld(Lagakos2020).Whenbothpovertyandurbanareasaremeasuredinacomparablewayacrosscountries,povertytendstobelowerindenseurbanareas(Combesetal.2022)(Figure1).22Ontheotherhand,somestudieshighlighturbanizationwithoutgrowth(Bloom,Canning,andFink2008;Castells-QuintanaandWenban-Smith2020;FayandOpal1999;Gollin,Jedwab,andVollrath2016).4Figure1.SubnationalextremepovertyratesacrosscountriesSource:Combesetal.2022.Note:Povertyismeasuredwiththeinternationalpovertyline(US$1.9in2011PPP).FollowingtheDegreeofUrbanizationmethodology(Dijkstraetal.2021),urbancenters(clusters)aredefinedbasedonspatiallycontiguoussetsof1km2gridcellsforwhichpopulationdensityofeachcell≥1,500(300)peopleperkm2andaggregatesettlementpopulation≥50,000(5,000).Nevertheless,thecontributionofurbanizationtopovertyreductionisnotself-evident.Largerordensercitiesdonotnecessarilyhelppeopleescapefrompovertyinthepresenceofovercrowding,congestion,crime,airpollution,highcostofliving,insufficientjobsforlow-skilledworkers,segregation,andsoon.Also,cross-sectionalnegativecorrelationsbetweendensityandpovertydonotnecessarilyindicateupwardmobility.Thus,itisnotevidentapriorithatlargerordensercitiesarethebestplaceforpeopletoescapefrompoverty(forexample,theargumentfavoringsecondarytownsinGibson,Jiang,andSusantono(2022)andChristiaensenandTodo(2014)).Moreover,largerordensercitiesaccommodatingalargershareofricherpeopledoesnotnecessarilymeanthatpeoplearemorelikelytobenonpoorinthosecitiesduetothesorting.Therefore,itisanempiricalquestionwhetherurbanagglomerationsfacilitatepovertyreduction.Thefirsthypothesistobetestedispeoplearemorelikelytobecomeorstaynonpoorinlargerordensercities,comparedtosmallerorlessdenselypopulatedtowns.Climaticandenvironmentalstressors:Hypothesis2Evenifurbanagglomerationssupportpovertyreduction,theupwardmobilityofthepoorcouldbehamperedbyclimaticshocksandrisks.Urbanhouseholdsmayfallintopovertyduetohigherexposuretoshocks,assetvulnerability,andlackofsocioeconomicresilience(Hallegatteetal.2017).Inurbanareas,poorhouseholdstendtobeexposedtofloodsduetotheirresidencesfacinghighenvironmentalhazardrisks,highbuildingdensityandovercrowdedness,andinadequateinfrastructure.Themoreurbanalocation,thescarcerandmoreexpensivethelandbecomes,pushingthepoorintoundesirableandriskylocationswithinorattheperipheriesofcities.Asnetconsumers,urbanhouseholdsarealsovulnerabletofoodpriceshockstriggeredbyclimateanomalies.Theliteratureindeedfoundurbanhouseholdsmorevulnerabletoflooding/droughtshocksinseveralcountries.Forexample,Baezetal.(2017)analyzedtheimpactsofaseveretropicalstormthathitGuatemalain2020withthelargestrainfallinthecountryduringthelastfivedecades.Medianpercapitaconsumptionfellby12.6percentinurbanareas,alargerdropthaninruralareas,significantlyincreasingurbanpoverty.Risingfoodpricesduetothedisastershockcontributedtothelossinurbanhouseholds’consumption,whileasocialsafetynetprogramprotectedmainlyruralhouseholds.471222834222913221138024921803504214223244072222328465AngolaBangladeshEgyptEthiopiaGhanaTanzaniaVietnamNationalUrbanUrbancenterUrbanclusterRural5Upwardmobilityofferedbyagglomerationeconomiescouldbeoffsetandhinderedbyclimaticandenvironmentalstressors.Therefore,wehypothesizethatlargeordensecitiesthataremoreexposedtoclimaticandenvironmentalshocksdonotofferresidentsahigherchancetobecomeorstaynonpoor,comparedtocitiesofsmallersize.3.Methodology3.1DataWeselectedChile,Colombia,andIndonesiaasthecasesforthisstudytodemonstratetheapplicationofanalyticalapproacheswithandwithoutpaneldatasets.Analyzingthesecountriesalsomeritsthetestoftheapproachesincountrieswherepovertyismeasuredbyincome(ChileandColombia)andconsumptionexpenditures(Indonesia).ThesettingofIndonesia—itsrapidurbanizationandheterogenousurbanandclimaticcharacteristicsacrosssubnationalregions—isparticularlysuitabletoouranalysis.HighlyurbanizedcountrieslikeChileandColombiahaveusefuldensityvariationstoexploreaswell.Toanswerourresearchquestionandverifyourhypotheses,wecombinedhouseholdsurveyswithclimaticdatasets.ForChileandColombia,weconstructedsyntheticpaneldatasetsoutofrepeatedcross-sectionalhouseholdsurveys.Floodriskisestimatedasakeyclimatefactorforeachtown.ForIndonesia,wereliedonpanelhouseholdsurveys(IFLS),combinedwithtwoclimateindicators:SPEIandthefloodriskindex.SyntheticpaneldataforChileandColombiaFollowingDangetal.(2014)andDangandLanjouw(2013),weappliedthesyntheticpanelmethodtothehouseholdsurveysofChile(EncuestadeCaracterizaciónSociooeconómicaNacional[CASEN])2011and2015andColombia(GranEncuestaIntegradadeHogares[GEIH])2008and2010.3Thismethodessentiallyexploitsthetime-invariantvariablesinthecross-sectionalsurveysandsomecohort-basedassumptionsabouttheerrortermstoconstructthesyntheticpanels.ThemethodologyisdescribedindetailinAnnexB.RecentapplicationsandfurthervalidationsofthesyntheticpanelmethodshavebeenimplementedusinghouseholdsurveydatafromvariouscountriesinSub-SaharanAfrica,EastAsiaandPacific,EuropeandCentralAsia,LatinAmerica,SouthAsia,andtheMiddleEastandNorthAfrica(seeDang,Jolliffe,andCarletto(2019);DangandLanjouw(forthcoming);andGarcés‐Urzainqui,Lanjouw,andRongen(2021)forrecentreviews).Webeginbyidentifyingthepotentialtime-invariantvariablesavailableinthecross-sectionalsurveysattwotimepoints,whichincludehouseholdheads’gender,age,levelofeducation,andresidencearea(thatis,urbanorrural).Thesevariablescanusuallybeassumedtobetimeinvariantiftheunderlyingpopulationremainsunchangedovertime.Onewaytotestthisassumptionistouseat-testfortheequalityofthemeansofthesamevariablesinthetwosurveyrounds.WehaveprovidedtestresultsthatconsiderthecomplexsurveydesigninTableA1forChileandTableA2forColombia.Theassumptionoftheequalityofmeansissatisfiedforhouseholdheads’genderandsecondarylevelofeducationinChileandheads’secondaryandtertiarylevelsofeducationinColombia.Theassumptionissatisfiedforresidenceareasinbothcountries.Althoughthedifferenceinheads’primarylevelofeducationisstatisticallysignificantinColombia,ithassimilarmeansovertime.Thedifferencesbetweensurveysforheads’primaryandtertiarylevelsofeducationinChileandheads’3TheyearsforthehouseholdsurveysforChileandColombiawereselectedbasedontheavailabilityofthesubnationallocationinformationthatcanbematchedwithclimaticlayersonthegeographicinformationsystem(GIS)platform.6genderinColombiaarelessthanfivepercentagepoints.Thus,thesemaynotmakemuchdifferencetothefinalestimatesinpractice.IFLSpanelhouseholdsurveydata:IndonesiaTheIFLSincludesatotalof54,000householdobservationsoverfivewavesfrom2,556subdistrictsin26provinces.Focusingonthesocioeconomicandhealthaspectsofthehouseholds,thesurveywasconductedforthefirsttimein1993covering13ofthetotal26provincesinthecountry.Inasampleof22,000individualsfrom7,224households,thesurveycollecteddataonindividualrespondentsandtheirfamilies(households)inadditiontodataoncommunities,health,andeducationfacilities.In1997/98,thesecondwavewasadministeredtothesamerespondentswitharecontactrateof94.4percent.ThatsurveyhadtheparticularitytoaccountfortheeconomicandpoliticalcrisisinIndonesia.Thethirdwavein2000wasmanagedtorecontact95.3percentofthefirstwavesamplewhilethefourthandthefifthroundsconductedin2007/08and2014/15recontacted93.6and90.5percent,respectively,ofthefirstwavesample(Straussetal.2016,citedinSetiawan,Tiwari,andRizal2018).WemeasuredpovertybasedonpercapitaconsumptionexpenditureswiththenationalpovertylinefollowingSetiawan,Tiwari,andRizal(2018).TheIFLSaccountsfortheexpenditureoftheconsumptionof37fooditemsoveraseven-dayrecallperiodandvariousnonfooditems.4Thenominalconsumptionaggregateisbothtemporallyandspatiallydeflated.5Inadditiontopoverty,weidentifiedvulnerablepeoplewithavulnerabilityline,whichissetat1.5timesthepovertyline.Roberts,Sander,andTiwari(2019)highlighttheimportanceofclassifyingurbanareasbasedontheirfunctionalityinsteadofmerepopulationsizeinIndonesia.FollowingDuranton(2015),wedefinedthefollowingfourlocationcategories:(1)metrocore,whichstandsforJakartaoradistrictwiththehighestpopulationdensityforothermetros;(2)urbanperipheries,whicharepredominantlyurbannon-coredistricts;(3)otherurbanareasthataccountforsingle-districtmetro(predominantlyurbanwithkotas)ornon-metrourban(predominantlyurbannon-metrodistricts);and(4)ruralareas,whichencompasstheruralperiphery(predominantlyruralnon-coredistrict)ornon-metroruralareas(predominantlyruralnon-metrodistricts).Climatedata:FloodriskindexandSPEIToaccountforclimaticandenvironmentalshocks,weusedtwoindicators:floodriskindexandtheSPEI.Thoseclimaticvariablesarepreparedatthesubdistrictlevel.Theprimaryclimaticstressoranalyzedinthisstudyisfloodrisk,givenitspotentialthreattourbanlivelihoods.Tocapturethefloodrisk,weusedtheflooddepthdataprovidedbyFATHOMin2016.Theflooddepthisexpressedinmetersandcomputedat3arc-second(approximately90m)resolutionandhasaglobalcoveragebetween56°Sand60°N.Thecomputationisbasedonpluvialdatawithareturnperiodof100years(1-in-100flooddepth).6The1-in-100flooddepthmeansafloodeventthathasa1percentprobabilityofoccurringinanygivenyearwithin100years.Weclassifiedtheareas4Nonfoodexpendituresincludehouseholdamenities(forexample,refrigerator,TV,andtelephone);housing;assorteditemssuchasclothing,furniture,medical,ceremonies,education(tuition,uniform,transportation,boarding);andothers.Regardingthehousingexpenditure,theactualmonthlyrentpaidwasrecorded.However,ifthehouseholdownsthehouse,theestimatedrentwasimputed.5Temporaldeflationisbasedontheconsumerpriceindexseries;spatialdeflatoriscalculatedbasedontheratiooftheregionalpovertylinestothenationalpovertyline,obtainedfromtheNationalSocioeconomicSurvey(SUSENAS)ofthecorrespondingwave.6Seehttps://agupubs.onlinelibrary.wiley.com/doi/epdf/10.1002/2015WR016954formoredetailsonthecomputationmethod.7withthetop25percentflooddepthineachcountryashighfloodriskareas.Asthisisanindicatoroflong-termfloodrisk,wehaveanindexonlyforonetimepoint.ThefloodriskmapsforthreecasecountriesareshowninFigure2.Takingadvantageofthepaneldataspanningoveralongduration,weadditionallyanalyzedrainfallanomaliesasaclimaticfactorforIndonesia.TheSPEIisamultiscalardroughtindex(Vicente-Serranoetal.2010).Theconstructionrequiresdataontemperature,precipitation,andpotentialevaporation.Accordingly,weprocessedmonthlyprecipitationandpotentialevapotranspirationderivedfromtheterraclimatedatafrom1958to2020.TheSPEIdataarefittedtoagammadistributionandnormalizedtoaflexiblemultipletimescalesuchas1,4,6,12,24,and48months.Forthestudy,weconsidereda12-monthtimescaleforeachyearfrom1993to2015withtwolagperiodsforeachIFLSwave.7NegativeSPEIvaluesrepresentrainfalldeficit—lessthanmedianprecipitation—andhighpotentialevapotranspiration(dry)startswhentheSPEIvalueisequaltoorbelow−1.0.Ontheotherhand,positiveSPEIvaluesindicaterainfallsurplus—greaterthanmedianprecipitation—andlowpotentialevapotranspiration(wet)startswhentheSPEIvalueisequaltoorabove1.0.Figure3illustratestheSPEIvaluesinIndonesiandistrictsasofMarch2014.7Thecomputationinvolvesthefollowingthreesteps(HarariandLaFerrara2018).Wefirstcomputethedifference(D)betweenprecipitationandevapotranspiration(PET)andaccountedfortheclimaticwaterbalancedefinedatthemonthlylevel.ThePenman-MonteithequationisusedtoapproximatethePET(asrecommendedbytheFoodandAgricultureOrganizationoftheUnitedNations(FAO)asthebestmethodfordeterminingreferenceevapotranspiration).Maximumtemperature,minimumtemperature,vaporpressure,precipitationaccumulation,downwardsurfaceshortwaveradiation,andwindspeedareusedasinputdata.Thenextstepistheaggregationoftheclimaticwaterbalanceatdifferenttimescalesandfinally,westandardizethetimeseriesaccordingtoagammadistribution.TheSPEIisthencomputedasthestandardizedvaluesofthegammafunction.8Figure2.Themapsof100-yearfloodrisks(A)Chile(B)Colombia(C)IndonesiaSource:BasedonFATHOMdata.Figure3.12-monthSPEIinIndonesia,March2014Source:Basedonterraclimatedata.DescriptivestatisticsTable1presentsthesummarystatisticsforChile(PanelA)andColombia(PanelB)basedonthesyntheticpaneldata.Households’probabilityoftransitionfrompoortononpooristheoutcomevariable.TheaverageprobabilityofpovertytransitionforChilefrom2011to2015is73.1percent,9andforColombiafrom2008to2010itis16.8percent.About24.2and13.8percentofhouseholdsinChileandColombia,respectively,areinhighfloodriskareas.Table1.Summarystatistics,ChileandColombiaCountMeanSDMinMaxPanelA:ChilePoorin2011(1=yes,0=no)44,6140.0960.2950.0001.000Poorin2015(1=yes,0=no)61,4330.0390.1930.0001.000Probabilityfrompoortononpoorbetween2011and201536,0350.7310.1010.4850.930Logofpopulationsizein201536,03510.8871.3347.43214.434Highfloodrisk(1=yes,0=no)36,0350.2420.4290.0001.000PanelB:ColombiaPoorin2008(1=yes,0=no)188,8010.2760.4470.0001.000Poorin2010(1=yes,0=no)190,3440.2410.4280.0001.000Probabilityfrompoortononpoorbetween2008and2010119,6920.1680.0760.0530.399Logofpopulationsizein2010119,69212.7450.9237.98814.546Highfloodrisk(1=yes,0=no)119,6920.1380.3450.0001.000Sources:BasedonCASEN2011and2015andGEIH2008and2010.Note:Povertymeasureisbasedonpercapitahouseholdincome,withathresholdofUS$5.50perday.TheprobabilityofchangingpovertystatusfrompoortononpoorisestimatedbasedonthesyntheticpanelapproachdescribedinAnnexB).Weclassifytheareaswithtop25percentflooddepthineachcountryashighfloodriskareas(seeSection3.1.3).SD=Standarddeviation.Table2presentsthesummarystatisticsofkeyvariablesforIndonesia.Households’poverty(1=nonpoor;0=poor)andvulnerability(1=neitherpoornorvulnerable;0=poororvulnerable)statusaredummyvariablesusedastheoutcomevariablesforourregressionanalysis.Around88percentofhouseholdobservationsinourfive-wavepaneldataarenonpoor,while69percentareneitherpoornorvulnerable.Theurbanlocationtypology—metrocore,peripheryurban,otherurban,andtheruralarea—arealsodefinedasdummyvariables.Around45percentofhouseholdobservationsarefromperipheryurban,followedbyotherurban(19.2percent),rural(18.6percent),andmetrocoreareas(17.4percent).Households’movementsacrosslocationsbetweeneachofthefiveIFLSwavesaresummarizedinTableA5.About5.9percentofhouseholdsareinSPEI-dryareas,90.7percentinareaswithSPEI-normal,and3.4percentinareasthatexperiencedheavyrains.About25percentofhouseholdsareexposedtohighfloodrisks.Table2.Summarystatistics,IndonesiaCountMeanSDMinMaxNonpoor(1=yes,0=no)47,7960.8770.3280.0001.000Neitherpoornorvulnerable(1=yes,0=no)47,7960.6900.4620.0001.000City:MetroCore(1=yes,0=no)47,7960.1740.3790.0001.000City:Peripheryurban(1=yes,0=no)47,7960.4480.4970.0001.000City:Otherurban(1=yes,0=no)47,7960.1920.3940.0001.000City:Rural(1=yes,0=no)47,7960.1860.3890.0001.000SPEI:Dry(SPEI<−2)(1=yes,0=no)47,7960.0590.2350.0001.000SPEI:Normal(1=yes,0=no)47,7960.9070.2900.0001.000SPEI:Rainy(SPEI>2)(1=yes,0=no)47,7960.0340.1820.0001.000Highfloodrisk(1=yes,0=no)47,7960.2510.4330.0001.000Source:BasedonIFLS1993,1997/98,2000,2007/8,and2014/15.Note:Povertyismeasuredwiththenationalpovertyline;vulnerabilityismeasuredwiththevulnerabilityline,whichissetat1.5timesthepovertyline.Weclassifytheareaswithtop25percentflooddepthineachcountryashighfloodriskareas(seeSection3.1.3).103.2EconometricapproachSincethesyntheticpaneldataonlyallowforstudyingthecorrelationalrelationshipbetweenclimatechangeandpovertymobility,weexaminedthecorrelationalrelationshiptoprovidesupportforthefirsthypothesisforChileandColombia.TotestthesecondhypothesisforColombiaandChilewithsyntheticpaneldatasets,weestimatedthefollowingfirst-differenceregressionmodelforhouseholdiincityjwiththeprobabilityoftransitionfrompoortononpoorstatusbetweenthetwotimepointst0andt1,or𝑦𝑖𝑗(𝑡0−𝑡1):𝑦𝑖𝑗(𝑡0−𝑡1)=𝛼+𝛽1POPSIZE𝑗,𝑡0+𝛽2(POPSIZE𝑗×CLMT𝑗)+𝛽3CLMT𝑗+𝜀𝑖𝑗.(1)WithPOPSIZE𝑗,𝑡0asthevariableindicatingthepopulationsizeofcityjatyeart0andCLMT𝑗asthe1-in-100-yearfloodrisksatcityj,𝜀𝑖𝑗𝑡istheerrorterm.Theparameter𝛽2indicateshowtherelationshipbetweenthecitypopulationsizeandtheupwardmobility(thatis,theprobabilityofpoorpeopleescapingfrompoverty)variesbytheclimaticrisks.TotestourfirsthypothesisthaturbanareassupportupwardmobilityforIndonesiawherethepaneldataareavailable,weestimatedthefollowingtwo-wayFEmodel:where𝑦𝑖𝑗𝑡standsforthepovertystatus(1=nonpoor;0=poor)orvulnerabilitystatus(neitherpoornorvulnerable=1;0=poororvulnerable)ofhousehold(i)incity(j)atyear(t).SincetherearenodataonpopulationsizeforIndonesia,weemployedthevariableCITY𝑗thatindicatesthelocationtypology—metrocore,urbanperiphery,otherurbanareas,andruralareas,withtheruralareasasthereferencecategory—forthiscountry.𝛾𝑖and𝛿𝑡standforthehouseholdFEsandtheyearFEs,respectively.WithhouseholdFEs,wefocusedontheprobabilityofescapingpovertyamongthemovers.Weexpectedtheparameter𝛽4formultidistrictmetropolitanareastobepositivebasedonthefirsthypothesis.Forthesecondhypothesis,weanalyzedtheinteractionofclimaticconditionswiththelocationeffectofurbanareasonpovertybyaddingtoEquation(3)aninteractiontermbetweenthelocationtypologyandtheclimatevariable.whereCLMT𝑗𝑡indicatestheexposuretofloodorfloodrisks.Inthecaseofexposuretoflood,theprecipitationanomaliesweremeasuredforeachIFLSwave(seeSection3.1).Theparameter𝛽6,thecoefficientfortheinteractionterm,capturestheeffectoftheclimateshocksassociatedwiththecityindicators.WeestimatedthepanelregressionsinEquations(2)and(3)aslinearprobabilitymodelswithstandarderrorsclusteredattheenumeratorareas.4.Results4.1ChileandColombia:SyntheticpanelanalysisTheresultsofthesyntheticpanelanalysisforChileandColombiashowthattheprobabilitiesofurbanresidentsescapingpovertyarepositivelyassociatedwiththepopulationsizeoftheircities.From2011to2015inChile,7.4percentofurbanpopulation(ortwo-thirdsoftheurbanpoor)escapedfrompoverty.Asshownincolumn1inTable3,theprobabilityofthetransitionfrompoortononpooris𝑦𝑖𝑗𝑡=𝛼+𝛽4CITY𝑖𝑗+𝛾𝑖+𝛿𝑡+𝜑𝑖𝑗𝑡,(2)𝑦𝑖𝑗𝑡=𝛼+𝛽5CITY𝑖𝑗+𝛽6(CITY𝑖𝑗×CLMT𝑗𝑡)+𝛽7CLMT𝑗𝑡+𝛾𝑖+𝛿𝑡+𝜑𝑖𝑗𝑡,(3)11positivelycorrelatedwiththepopulationsizeofcities.AsimilarcorrelationisobservedforColombiabetween2008and2010(column1inTable4),where4.1percentofpopulationescapedfrompoverty.WethenestimatedregressionmodelsinEquation(1)forChileandColombiatoexaminetheheterogeneityinthelinkbetweentheprobabilityofpovertytransitionandcitypopulationsizebyfloodrisks.AsshowninColumn4inTable3andTable4,theinteractiontermbetweenthelogofcitypopulationsizeandthefloodriskvariablesisnegative(−0.005forChileand−0.003forColombia),indicatingthattheupwardmobilityinlargecitiestendstobelimitediftheyfacehighfloodrisks.Withflatlinesforhigh-riskareasandsteeplinesforlow-riskareas,Figure4clearlyshowssuchheterogeneitybyfloodrisks.Table3.First-differencemodel(depvar:theprobabilityoftransitionfrompoortononpoor),Chile(1)(2)(3)(4)Logpopulation20150.00830.00750.0089(0.0004)(0.0004)(0.0004)Floodriskishigh−0.0174−0.01360.0380(0.0012)(0.0011)(0.0085)Floodriskishigh#logpopulation2015−0.005(0.0008)Constant0.6410.7360.6540.638(0.0041)(0.0006)(0.0040)(0.0050)Observations36,03536,03536,03536,035R-squared0.0120.0060.0160.016Note:ThetablesummarizestheestimationresultsofsyntheticpanelmodelsinEquation(1).Thedependentvariableistheprobabilityofeachhousehold’stransitionfrompoortononpoorbetween2011and2015.Standarderrorsinparenthesesareestimatedwith1,000bootstraps.p<0.1,p<0.05,p<0.01.Table4.First-differencemodel(depvar:theprobabilityoftransitionfrompoortononpoor),Colombia(1)(2)(3)(4)Logofpopulation20100.00290.00210.0025(0.0002)(0.0002)(0.0002)Floodriskishigh−0.0101−0.0090.0223(0.0006)(0.0006)(0.0075)Floodriskishigh#Logpop.2010-0.0025(0.0006)Constant0.1320.1700.1430.138(0.0029)(0.0002)(0.0030)(0.0032)Observations119,692119,692119,692119,692R-squared0.0010.0020.0030.003Note:ThetablesummarizestheestimationresultsofsyntheticpanelmodelsinEquation(1).Thedependentvariableistheprobabilityofeachhousehold’stransitionfrompoortononpoorbetween2008and2010.Standarderrorsinparenthesesareestimatedwith1,000bootstraps.p<0.1,p<0.05,p<0.01.12Figure4.Probabilityofpovertytransitionbycitypopulationsizeandfloodrisk(A)Chile(B)ColombiaSource:Authors’construction.Note:ThepredictedprobabilitiesofbecomingnonpoorarebasedontheresultsofColumns4inTable3(Chile)andTable4(Colombia).Theerrorbarsindicate95%confidenceintervals(CIs).4.2.Indonesia:PaneldataanalysisUrbanescalatoroutofpovertyTable5summarizestheestimationresultsofthelinearprobabilitymodelsinEquation(2)usingthehousehold’snonpoorstatusasthedependentvariable.Columns1and2reportspecificationswithouthouseholdFEs,whereasColumns3and4includehouseholdFEs.YearFEsareincludedinColumns2and4.Themainvariableofinterestistheindicatorformetrocore,itscoefficientindicatingtheprobabilityofhouseholds’becomingnonpoorrelativetothoseinruralareas.InthebaselinespecificationwithnoFEs(Column1),thecoefficientestimateformetrocoreis0.065(90%CI=0.019),meaningthatthemetrocoreoffersa6.6percentagepointhigherprobabilitytogetoutofpovertyincomparisonwiththeruralareas.Otherurbanareasalsohaveapositivecoefficientof0.034(90%CI=0.020).Bycontrast,peri-urbanareasshowanegativecoefficient,indicatingthatthechanceofbeingnonpoorislowerthaninruralareas.AddingyearFEsinColumn2doesnotchangetheresultmuch.WithhouseholdandyearFEs(Column4),thecoefficientestimateformetrocoreisreducedto0.031(90%CI=0.031).Otherurbanareasshowanevensmallercoefficient(0.0007,90%CI=0.036).13Table5.Baselinelinearprobabilitymodels(depvar:nonpoor)(1)(2)(3)(4)City:Core0.0659(0.0118)0.0701(0.0119)0.0145(0.0207)0.0310(0.0193)City:Peripheryurban−0.0268(0.0125)−0.0264(0.0125)−0.0023(0.0165)0.0005(0.0144)City:Otherurban0.0342(0.0128)0.0338(0.0128)0.0054(0.0250)0.0073(0.0225)City:Rural(Reference)HouseholdFENoNoYesYesYearFENoYesNoYesAdjustedR20.01170.0189−0.00000.0058#ofobservations47,79547,79547,79547,795#ofhouseholds18,49018,49018,49018,490Note:ThetablesummarizestheestimationresultsofpanelregressionmodelsinEquation(2)forhouseholdsinthefivewavesofIFLS(1993,1997/8,2000,2007/8,and2014/15).Thedependentvariableisabinaryindicatorabouthousehold’spovertystatus(1=nonpoor;0=poor).Clusterrobuststandarderrorsinparentheses.p<0.1,p<0.05,p<0.01.AnotherseriesofregressionsinTable6replacetheoutcomevariablewithanindicatoraboutvulnerability.InColumns2and4,thecoefficientestimateformetrocoreis−0.141(withouthouseholdFEs)and−0.070(withhouseholdFEs),respectively.Thatis,theprobabilityofbeingneitherpoornorvulnerableincreasesforhouseholdslivinginthemetrocoreareas,regardlessofwithorwithouthouseholdFEs.Inotherwords,peoplelivingin(ormovingto)themetrocoreareasarelesslikelytobecomepoororvulnerable.Table6.Baselinelinearprobabilitymodels(depvar:neitherpoornorvulnerable)(1)(2)(3)(4)City:Core0.136(0.0196)0.141(0.0196)0.0482(0.0286)0.0701(0.0276)City:Peripheryurban−0.0504(0.0194)−0.0503(0.0194)−0.0168(0.0260)−0.0159(0.0242)City:Otherurban0.0610(0.0213)0.0602(0.0212)0.0310(0.0271)0.0325(0.0254)City:Rural(reference)HouseholdFENoNoYesYesYearFENoYesNoYesAdjustedR20.02280.03080.000190.0088#ofobservations47,79547,79547,79547,795#ofhouseholds18,49018,49018,49018,490Note:ThetablesummarizestheestimationresultsofpanelregressionmodelsinEquation(2)forhouseholdsinthefivewavesofIFLS(1993,1997/8,2000,2007/8,and2014/15).Thedependentvariableisabinaryindicatorabouthousehold’svulnerabilitystatus(1=notvulnerable;0=vulnerable).Clusterrobuststandarderrorsinparentheses.p<0.1,p<0.05,p<0.01.ClimaticshockontheurbanescalatorFloodriskasashockindicatorTable7presentstheestimatesforEquation(3),showingtherelationbetweenthefloodrisk(asaclimateshockindicator)andthelocationeffectofurbanareasonpoverty.Column4withtheinteractionbetweenlocationandfloodriskvariables,aswellashouseholdandyearFEs,showsthattheprobabilityofbeingnonpoordecreasesduetohighfloodrisksby7.4percentagepoints(90%CI14=0.066)inmetrocoreand5.33percentagepoints(90%CI=0.040)inperipheryurbanareas,respectively.Thatis,inmetroareas,highfloodrisklowersthechanceofgettingoutofpovertyincomparisonwithlow-riskareas.PredictedprobabilitiesinFigure5showhighfloodriskareasindicatealowerpredictedprobabilityofbeingnonpoorineachlocationindicatorincomparisonwiththelow-riskareasexceptformetrocorewithouthouseholdFEsandruralareaswithhouseholdFEs.Table7.Linearprobabilitymodelswithfloodrisk(depvar:nonpoor)(1)(2)(3)(4)City:Core0.0634(0.0123)0.0601(0.0149)0.0293(0.0194)0.0467(0.0196)City:Peripheryurban−0.0274(0.0124)−0.0299(0.0157)−0.00034(0.0143)0.0176(0.0157)City:Otherurban0.0304(0.0127)0.0349(0.0153)0.0064(0.0227)0.0189(0.0231)City:Rural(reference)Highfloodrisk−0.0257(0.0092)−0.0282(0.0192)−0.0107(0.0111)0.0317(0.0189)City:Core#Highfloodrisk0.0329(0.0253)−0.0743(0.0404)City:Peripheryurban#Highfloodrisk0.00827(0.0235)−0.0533(0.0244)City:Otherurban#Highfloodrisk−0.0232(0.0270)−0.0322(0.0259)HouseholdFENoNoYesYesYearFEYesYesYesYesAdjustedR20.02000.02030.00580.0061#ofobservations47,79547,79547,79547,795#ofhouseholds18,49018,49018,49018,490Note:ThetablesummarizestheestimationresultsofpanelregressionmodelsinEquation(3)forhouseholdsinthefivewavesofIFLS(1993,1997/8,2000,2007/8,and2014/15).Thedependentvariableisabinaryindicatorabouthousehold’spovertystatus(1=nonpoor;0=poor).Clusterrobuststandarderrorsinparentheses.p<0.1,p<0.05,p<0.01.Figure5.PredictedprobabilityofbeingnonpoorbylocationsSource:Authors’construction.Note:ThepredictedprobabilitiesofbecomingnonpoorarebasedontheresultsofColumns2and4inTable7.Errorbarsindicate90%CI.0.750.80.850.90.951LowriskHighriskLowriskHighriskLowriskHighriskLowriskHighriskCorePeripheryurbanOtherurbanRuralPr(nonpoor)(A)WithouthouseholdFEs0.80.820.840.860.880.90.920.94LowriskHighriskLowriskHighriskLowriskHighriskLowriskHighriskCorePeripheryurbanOtherurbanRuralPr(nonpoor)(B)WithhouseholdFEs15SPEIasashockindicatorTurningtoSPEI,wereplacethefloodriskwiththeSPEIasthesecondclimaticshockindicatorinTable8.Wekeepthenonpoorstatusasthedependentvariable.AsexplainedinSection3,wedividetheSPEIintothreecategories:SPEI-rainy(SPEI>2.0),SPEI-normal(−2˃SPEI˃2),andSPEI-dry(SPEI<−2.0).ThecoefficientestimatefortheinteractionbetweenthemetrocoreandSPEI-rainyvariablesis−0.098(90%CI=0.058)inColumn4withhouseholdFEs,suggestingthatmetrocoreareasthatexperiencedheavyrainshave9.8percentagepointslowerchanceofgettingoutofpoverty.ThatmeansSPEIstronglyreducestheurbanescalatorfunctionofmetrocoreareas.Figure6showsthepredictedprobabilitiesconfirmingthepatternofSPEI-rainyformetrocoreareas.Exceptforotherurban,areaswithheavyrainsshowlowerpredictedprobabilitiesofbeingnonpoorwithorwithouthouseholdFEs.WhenwereplacedthenonpoorindicatorwiththestatusofneitherpoornorvulnerableastheoutcomevariableinTable9,wemadeasimilarconclusion.ThecoefficientoftheinteractionbetweenmetrocoreandSPEI-rainy(Column4)is0.131(90%CI=0.036),meaningthatthechanceofbeingneitherpoornorvulnerableforpeoplemovingtometrocoreareasthatexperienceheavyraindecreasesby13.1percentagepointscomparedtothosewhodidnotfaceheavyrains.TheresultislessclearfortheregressionswithouthouseholdFEs(Column2).16Table8.LinearprobabilitymodelswithSPEI(depvar:nonpoor)(1)(2)(3)(4)City:Core0.0676(0.0115)0.0670(0.0112)0.0272(0.0197)0.0273(0.0197)City:Peripheryurban−0.0263(0.0124)−0.0268(0.0118)−0.0025(0.0151)−0.0032(0.0152)City:Otherurban0.0323(0.0125)0.0284(0.0123)0.0047(0.0233)0.0036(0.0233)City:Rural(reference)SPEI:Dry0.0220(0.0104)0.0358(0.0212)0.0256(0.0109)0.0185(0.0200)SPEI:Normal(reference)SPEI:Rainy−0.0406(0.0193)−0.0817(0.0460)−0.0388(0.0155)−0.0445(0.0310)City:Core#SPEI:dry−0.0135(0.0215)0.0090(0.0198)City:Core#SPEI:rainy−0.00019(0.0619)−0.0985(0.0356)City:Peripheryurban#SPEI:dry−0.0323(0.0259)0.0039(0.0252)City:Peripheryurban#SPEI:rainy0.0495(0.0496)0.0078(0.0363)City:Otherurban#SPEI:dry0.0141(0.0226)0.0193(0.0257)City:Otherurban#SPEI:rainy0.128(0.0473)0.0657(0.0317)HouseholdFENoNoYesYesYearFEYesYesYesYesAdjustedR20.01950.01990.00650.0068#ofobservations47,79547,79547,79547,795#ofhouseholds18,49018,49018,49018,490Note:ThetablesummarizestheestimationresultsofpanelregressionmodelsinEquation(3)forhouseholdsinthefivewavesofIFLS(1993,1997/8,2000,2007/8,and2014/15).Thedependentvariableisabinaryindicatorabouthousehold’spovertystatus(1=nonpoor;0=poor).Clusterrobuststandarderrorsinparentheses.p<0.1,p<0.05,p<0.01.Figure6.PredictedprobabilityofbeingnonpoorbylocationsSource:Authors’construction.Note:ThepredictedprobabilitiesofbecomingnonpoorarebasedontheresultsofColumns2and4inTable5.Errorbarsindicate90%CI.0.70.750.80.850.90.951DryNormalRainyDryNormalRainyDryNormalRainyDryNormalRainyCorePeripheryurbanOtherurbanRuralPr(nonpoor)(A)WithouthouseholdFEs0.70.80.91DryNormalRainyDryNormalRainyDryNormalRainyDryNormalRainyCorePeripheryurbanOtherurbanRuralPr(nonpoor)(B)WithhouseholdFEs17Table9.Linearprobabilitymodel,regressionwithSPEI(depvar:neitherpoornorvulnerable)(1)(2)(3)(4)City:Core0.137(0.0193)0.134(0.0193)0.0646(0.0284)0.0637(0.0283)City:Peripheryurban−0.0504(0.0192)−0.0502(0.0190)−0.0207(0.0256)−0.0199(0.0255)City:Otherurban0.0581(0.0210)0.0556(0.0214)0.0285(0.0265)0.0281(0.0264)City:Rural(reference)SPEI:Dry0.0145(0.0181)0.0235(0.0316)0.0273(0.0155)0.00452(0.0271)SPEI:Normal(reference)SPEI:Rainy−0.0610(0.0243)−0.0900(0.0419)−0.0598(0.0186)−0.0229(0.0204)City:Core#SPEI:Dry0.0195(0.0348)0.0433(0.0293)City:Core#SPEI:Rainy0.0266(0.0690)−0.131(0.0362)City:Peripheryurban#SPEI:Dry−0.0369(0.0407)0.0150(0.0354)City:Peripheryurban#SPEI:Rainy0.0359(0.0495)−0.0434(0.0314)City:Otherurban#SPEI:Dry0.0132(0.0384)0.0373(0.0423)City:Otherurban#SPEI:Rainy0.0558(0.0768)0.0541(0.0370)HouseholdFENoNoYesYesYearFEYesYesYesYesAdjustedR20.03120.03130.00970.0099#ofobservations47,79547,79547,79547,795#ofhouseholds18,49018,49018,49018,490Note:ThetablesummarizestheestimationresultsofpanelregressionmodelsinEquation(3)forhouseholdsinthefivewavesofIFLS(1993,1997/8,2000,2007/8,and2014/15).Thedependentvariableisabinaryindicatorabouthousehold’svulnerabilitystatus(1=notvulnerable;0=vulnerable).Clusterrobuststandarderrorsinparentheses.p<0.1,p<0.05,p<0.01.5.DiscussionandconclusionThispaperexaminestheeffectsoftheclimaticandenvironmentalshocksonakeyfunctionofurbanagglomerationstofacilitatepovertyreduction.Ourstudyalsoshowcasesdifferentempiricalapproachesdependingontheavailabilityofpaneldatasets.WeconstructedsyntheticpaneldatasetsforColombiaandChilefromrepeatedcross-sectionalhouseholdsurveysandexaminedtheassociationbetweenpovertychangesovertimeandthecitypopulationsizeaswellastheheterogeneityofsuchassociationbyfloodrisks.Byestimatingtwo-wayFEmodelswithfivewavesofIFLSspanningfrom1993to2015,weanalyzedtheprobabilitiesofhouseholdsescapingpovertyindifferentlocations—metrocore,urbanperiphery,otherurbanareas,andruralareas—andfloodingrisksinIndonesia.Theresultsfromthethreecasecountriesshowsimilarpatterns.ThesyntheticpanelanalysesforColombiaandChileindicateareductioninurbanpovertyratesmeasuredbytheupper-middle-incomeinternationalpovertyline(US$5.5perdayin2011PPP).Theanalysisfindsthattheprobabilitiesofhouseholds’transitionfrompoortononpoorstatuswerepositivelycorrelatedwiththecitypopulationsizeinbothcountries.Moreimportantly,suchupwardmobilitywasobservedonlyinlargercitieswithlowfloodrisk.Theresultsofourtwo-wayFEregressionanalysesforIndonesiasuggestthatdensemetropolitanareashaveprovidedgoodopportunitiesforthosemigrantstoescapepovertyin18Indonesia.However,highfloodriskappearstohavereducedsuchupwardmobilityinlargemetropolitanareas.Thereareseveralpotentialreasons.First,heavyrainfallsandfloodingcosturbanresidentstorepairandreplacetheirdamagedassets,suchasdwellings.Neighborhoodswithhighbuildingdensityandpoorinfrastructurecouldaugmentthedamageandtherebytherecoverycosts.Second,floodingmaylowertheproductivityandoutputsofworkersbydamagingproductiveassets,reducingthetimeallocatedforwork,andconstrainingcommuting.Thefindingssuggesttheimportanceofreducingfloodriskstopromotepovertyreductionthroughmigrationtolargemetroareas.Upgradinghigh-densityinformalsettlementswouldbeaneffectiveapproachforadaptation.IntheIndonesianurbanizationcontext,itwouldalsobeimportanttoinvestintheperipheriesofmetropolitanareas,astheyhavebeenreceivingalargeinfluxofmigration.Itisessentialtoreducecongestionforcesduetotheincreasedmigrationandbetterconnectperipheriestothecoresasthelatterprovidemorepoverty-reducingopportunities.Finally,weclarifysomelimitationsofourstudy.First,althoughweemployedtwo-wayFEregressionmodelsforIndonesia,wecouldnotdistinguishthesortingofmigrantsfromthelocationeffects.Wewouldneedastrongeridentificationstrategy,suchasanaturalexperimentaldesign.Ifthosewithhighcapabilitytoescapefrompovertytendtomovetocities,ourestimationoflocationeffectsmightbeoverestimated.Second,wefocusedonheavyrainfallandfloodingastheclimaticvariable,thoughmanyotherclimaticandenvironmentalstressorsalsopotentiallyunderminethebenefitsofurbanagglomerations.Finally,thisisacasestudyofthethreecountries;thus,wecannotautomaticallygeneralizethefindingstoothercontexts.19ReferencesBaez,J.E.,L.Lucchetti,M.E.Genoni,andM.Salazar.2017.“GonewiththeStorm:RainfallShocksandHouseholdWellbeinginGuatemala.”TheJournalofDevelopmentStudies53(8):1253–1271.Beegle,K.,J.DeWeerdt,andS.Dercon.2011.“MigrationandEconomicMobilityinTanzania:EvidencefromaTrackingSurvey.”ReviewofEconomicsandStatistics93(3):1010–1033.Bloom,D.E.,D.Canning,andG.Fink.2008.“UrbanizationandtheWealthofNations.”Science319(5864):772–775.Bryan,G.,S.Chowdhury,andA.Mobarak.2014.“UnderinvestmentinaProfitableTechnology:TheCaseofSeasonalMigrationinBangladesh.”Econometrica82(5):1671–1748.Castells-Quintana,D.,andH.Wenban-Smith.2020.“PopulationDynamics,UrbanisationwithoutGrowth,andtheRiseofMegacities.”TheJournalofDevelopmentStudies56(9):1663–1682.Chauvin,J.,E.Glaeser,Y.Ma,andK.Tobio.2017.“WhatIsDifferentaboutUrbanizationinRichandPoorCountries?CitiesinBrazil,China,IndiaandtheUnitedStates.”JournalOfUrbanEconomics98:17–49.Christiaensen,L.,andY.Todo.2014.“PovertyReductionduringtheRural-UrbanTransformation:TheRoleoftheMissingMiddle.”WorldDevelopment63:43–58.Combes,P-P.,S.Demurger,S.Li,andJ.Wang.2020.“UnequalMigrationandUrbanisationGainsinChina.”JournalofDevelopmentEconomics142.Combes,P-P.,S.Nakamura,M.Roberts,andB.Stewart.2022.“EstimatingUrbanPovertyConsistentlyAcrossCountries.”WorldBankPovertyandEquityNotes48.Dang,H-A.,P.Lanjouw,J.Luoto,andD.McKensize.2014.“UsingRepeatedCross-SectionstoExploreMovementsintoandoutofPoverty.”JournalofDevelopmentEconomics107:112–128.Dang,Hai-Anh,andPeterLanjouw.Forthcoming.“Regression-BasedImputationforPovertyMeasurementinDataScarceSettings.”InHandbookofResearchonMeasuringPovertyandDeprivation,editedbyJacquesSilber.EdwardElgarPress.Dang,Hai-Anh,andPeterLanjouw.2013.MeasuringPovertyDynamicswithSyntheticPanelsBasedonCross-Sections.WorldBank.Dang,Hai-Anh,DeanJolliffe,andCalogeroCarletto.2019.“DataGaps,DataIncomparability,andDataImputation:AReviewofPovertyMeasurementMethodsforData-ScarceEnvironments.”JournalofEconomicSurveys3:757–797.Dijkstra,L.,Florczyk,A.J.,Kemper,T.,Melchiorri,M.,Pesaresi,M.,Shiavina,M.2021.Applyingthedegreeofurbanisationtotheglobe:Anewharmoniseddefinitionrevealsadifferentpictureofglobalurbanisation.J.UrbanEcon.125.Duranton,G.2015.“AProposaltoDelineateMetropolitanAreasinColombia.”DesarrolloySociedad15:223–64.20Duranton,G.,andD.Puga.2004.“Micro-foundationsofUrbanAgglomerationEconomies.”HandbookofRegionalandUrbanEconomics4(4):2063–2117.Fay,M.,andC.Opal.1999.UrbanizationwithoutGrowth:ANot-So-UncommonPhenomenon.Oxford:WorldBank.Garcés‐Urzainqui,D.,P.Lanjouw,andG.Rongen.2021.“ConstructingSyntheticPanelsforthePurposeofStudyingPovertyDynamics:APrimer.”ReviewofDevelopmentEconomics25(4):1803–1815.Gibson,J.,Y.Jiang,andB.Susantono.2022.“RevisitingtheRoleofSecondaryTowns:EffectsofDifferentTypesofUrbanGrowthonPovertyinIndonesia.”UniversityofWaikatoWorkingPaperinEconomics5/22.Glaeser,E.L.,andJ.D.Gottlieb.2009.“TheWealthofCities:AgglomerationEconomiesandSpatialEquilibriumintheUnitedStates.”JournalofEconomicLiterature47(4):983–1028.Glaeser,E.L.,H.D.Kallal,J.A.Scheinkman,andA.Shleifer.1992.“GrowthinCities.”JournalofPoliticalEconomy100(6):1126–1152.Glaeser,E.,andD.Maré.2001.“CitiesandSkills.”JournalofLaborEconomics19(2).Gollin,D.,R.Jedwab,andD.Vollrath.2016.“Urbanizationwithandwithoutindustrialization.”JournalofEconomicGrowth21(1):35–70.Grover,A.,S.V.Lall,andJ.Timmis.2021.“AgglomerationEconomiesinDevelopingCountries:AMeta-Analysis.”WorldBankPolicyResearchWorkingPaperNo.9730.Hallegatte,S.,A.Vogt-Schilb,M.Bangalore,andJ.Rozenberg.2017.Unbreakable:BuildingtheResilienceofthePoorintheFaceofNaturalDisasters.Washington,DC:WorldBank.Hamory,J.,M.Kleemans,N.Y.Li,andE.Miguel.2021.“ReevaluatingAgriculturalProductivityGapswithLongitudinalMicrodata.”JournalofEuropeanEconomicAssociation19(3):1522–1555.Harari,M.,andE.L.Ferrara.2018.“Conflict,Climate,andCells:ADisaggregatedAnalysis.”TheReviewofEconomicsandStatistics100(4):594–608Lagakos,D.2020.“Urban-RuralGapsintheDevelopingWorld:DoesInternalMigrationOfferOpportunities?”JournalofEconomicPerspectives34(3):174–192.Melo,P.C.,D.J.Graham,andR.B.Noland.2009.“AMeta-AnalysisofEstimatesofUrbanAgglomerationEconomies.”RegionalScienceandUrbanEconomics39(3):332–342.Michaels,G.,F.Rauch,andS.J.Redding.2012.“UrbanizationandStructuralTransformation.”TheQuarterlyJournalofEconomics127(2):535–586.Mukim,M.,andM.Roberts.2022.Thriving:MakingCitiesGreen,Resilient,andInclusiveinaChallengingClimate.Washington,DC:WorldBank.Puga,D.2010.“TheMagnitudeandCausesofAgglomerationEconomies.”JournalofRegionalScience50(1):203–21921Quintero,L.E.,andM.Roberts.2022.“CitiesandProductivity:Evidencefrom16LatinAmericanandCaribbeanCountries.”JohnsHopkinsCareyBusinessSchoolResearchPaperNo.22–13.Roberts,M.,F.G.Sander,andS.Tiwari.2019.TimetoAct:RealizingIndonesia’sUrbanPotential.Washington,DC:WorldBank.Rosenthal,S.S.,andW.C.Strange.2004.“EvidenceontheNatureandSourcesofAgglomerationEconomies.”InHandbookofRegionalandUrbanEconomics(Vol.4),editedbyJ.V.HendersonandJ.F.Thisse,2119–2171.Amsterdam:Elsevier.Setiawan,I.,S.Tiwari,andH.Rizal.2018.EconomicandSocialMobilityinUrbanizingIndonesia.BackgroundpaperforRoberts,Sander,andTiwari(2019).UnitedNations,DepartmentofEconomicandSocialAffairs,PopulationDivision.2019.WorldUrbanizationProspects:The2018Revision(ST/ESA/SER.A/420).NewYork:UnitedNations.Vicente-Serrano,S.M.,S.Beguera,andJ.I.Lopez-Moreno.2010.“AMultiscalarDroughtIndexSensitivetoGlobalWarming:TheStandardizedPrecipitationEvapotranspirationIndex.”JournalofClimate23:1696–1718.22AnnexA:AdditionaltablesTableA1.Summarystatistics,Chile(2011–2015)Difference20112015Thelogarithmofpercapitaincome5.8756.1770.302(0.039)(0.034)(0.016)Head’sage42.94045.6072.667(0.111)(0.121)(0.153)Headisfemale0.3420.340−0.002(0.008)(0.005)(0.009)Headdoesnotcompleteprimaryschool0.1330.118−0.015(0.007)(0.006)(0.004)Head’shighesteducationlevelisprimary0.3000.260−0.039(0.010)(0.008)(0.007)Head’shighesteducationlevelissecondary0.3970.4100.013(0.013)(0.010)(0.008)Head’shighesteducationlevelistertiary0.1700.2110.041(0.018)(0.016)(0.008)Urbanarea0.8790.875−0.004(0.012)(0.012)(0.004)Note:Standarderrorsareinparentheses,andthedifferencesareestimated,consideringthecomplexsurveydesign.p<0.1,p<0.05,p<0.01.Populationweightsareapplied.Wedonottestforthedifferenceinthedistributionsoftheagevariableinthetwosurveyroundssinceitisadeterministicvariable.Householdheads’agesarerestrictedtobetween25and55forthefirstsurveyroundandbetween29and59forthesecondsurveyround.TableA2.Summarystatistics,Colombia(2008–2010)Difference20082010Thelogarithmofpercapitaincome5.3415.4220.081(0.052)(0.044)(0.020)Head’sage41.07342.3171.243(0.104)(0.075)(0.070)Headisfemale0.2610.2850.023(0.008)(0.008)(0.004)Headdoesnotcompleteprimaryschool0.2250.2390.014(0.016)(0.017)(0.008)Head’shighesteducationlevelisprimary0.3850.372−0.013(0.004)(0.005)(0.005)Head’shighesteducationlevelissecondary0.2880.2930.004(0.010)(0.011)(0.005)Head’shighesteducationlevelistertiary0.1010.097−0.005(0.007)(0.007)(0.003)Urbanarea0.8140.8150.002(0.031)(0.028)(0.011)Note:Standarderrorsareinparentheses,andthedifferencesareestimated,consideringthecomplexsurveydesign.p<0.1,p<0.05,p<0.01.Populationweightsareapplied.Wedonottestforthedifferenceinthedistributionsoftheagevariableinthetwosurveyroundssinceitisadeterministicvariable.Householdheads’agesarerestrictedtobetween25and55forthefirstsurveyroundandbetween27and57forthesecondsurveyround.23TableA3.EstimatedOLSmodelofhouseholdincomepercapitain2015,ChileCoef/SEHead’sage0.014(0.00)=1iftheheadisfemale−0.139(0.01)Educationlevel(reference-ifheaddoesnotcompleteprimaryschool)=1ifthehead’shighesteducationlevelisprimary0.129(0.01)=1ifthehead’shighesteducationlevelissecondary0.398(0.02)=1ifthehead’shighesteducationlevelistertiary1.183(0.07)=1iftheareaofresidenceisurban0.016(0.02)Constant5.141(0.04)AdjustedR20.294Numberofobservations45,954Note:Standarderrorsclusteredatprimarysamplingunitsareinparentheses.p<0.1,p<0.05,p<0.01.Thedependentvariableisthelogarithmofhouseholdincomepercapita.Householdheads’agesarerestrictedtobetween29and59.TableA4.EstimatedOLSmodelofhouseholdincomepercapitain2010,ColombiaCoef/SEHead'sage0.019(0.00)=1iftheheadisfemale−0.127(0.02)Educationlevel(reference-ifheaddoesnotcompleteprimaryschool)=1ifthehead’shighesteducationlevelisprimary0.319(0.02)=1ifthehead’shighesteducationlevelissecondary0.767(0.02)=1ifthehead’shighesteducationlevelistertiary1.704(0.05)=1iftheareaofresidenceisurban0.358(0.06)Constant3.915(0.05)AdjustedR20.348Numberofobservations133,483Note:Standarderrorsclusteredatprimarysamplingunitsareinparentheses.p<0.1,p<0.05,p<0.01.Thedependentvariableisthelogarithmofhouseholdincomepercapita.Householdheads’agesarerestrictedtobetween30and60.24TableA5.ResidentialmovementacrossIFLSwaves,Indonesia(percentageofhousehold)Wave2MetrocorePeripheryUOtherURuralWave1Metrocore95.20.73.50.6PeripheryU0.199.40.20.3OtherU0.30.399.30.2Rural0.00.70.399.0Wave3MetrocorePeripheryUOtherURuralWave2Metrocore96.21.62.00.3PeripheryU0.295.71.23.0OtherU0.64.095.00.3Rural0.113.92.983.2Wave4MetrocorePeripheryUOtherURuralWave3Metrocore91.92.24.91.1PeripheryU0.393.12.24.4OtherU1.63.391.04.2Rural0.48.90.889.9Wave5MetrocorePeripheryUOtherURuralWave4Metrocore91.82.54.51.2PeripheryU0.398.60.40.7OtherU2.41.595.50.7Rural0.72.00.796.725AnnexB:SyntheticpanelmethodThisannexprovidesabriefsummaryofthesyntheticpanelmethodbasedonDangetal.(2014)andDangandLanjouw(2013).Let𝑥𝑖𝑗beavectorofhouseholdcharacteristicsobservedinsurveyroundj(j=1or2)thatarealsoobservedintheothersurveyroundforhouseholdi,i=1,…,N.Thesehouseholdcharacteristicscanincludetime-invariantvariablessuchasethnicity,religion,language,placeofbirth,andparentaleducationaswellasothertime-varyinghouseholdcharacteristicsifretrospectivequestionsaboutthefirstroundvaluesofsuchcharacteristicsareaskedinthesecondroundsurvey.Toreducespuriouschangesduetochangesinhouseholdcompositionovertime,weusuallyrestricttheestimationsamplestohouseholdheadsinacertainagerange,forexample,25–55,inthefirstcross-sectionandadjustthisagerangeaccordinglyinthesecondcross-section.Thisrestrictionalsohelpsensurecertainvariablessuchasheads’educationattainmentremainsrelativelystableovertime(assumingmostheadshavecompletedschooling).8Thisagerangeisusuallyusedintraditionalpseudo-panelanalysisbutcanvarydependingontheculturalandeconomicfactorsineachspecificsetting.Populationweightsarethenemployedtoprovideestimatesthatrepresentthewholepopulation.Then,let𝑦𝑖𝑗representhouseholdconsumptionorincomeinsurveyroundj,j=1or2.Thelinearprojectionofhouseholdconsumption(orincome)onhouseholdcharacteristicsforeachsurveyroundisgivenby𝑦𝑖𝑗=𝛽𝑗′𝑥𝑖𝑗+𝜀𝑖𝑗.(B.1)Let𝑧𝑗bethepovertylineinperiodj.Weareinterestedinknowingtheunconditionalmeasuresofpovertymobilitysuchas𝑃(𝑦𝑖1<𝑧1𝑎𝑛𝑑𝑦𝑖2>𝑧2).(B.2)whichrepresentsthepercentageofhouseholdsthatarepoorinthefirstsurveyround(year)butnonpoorinthesecondsurveyroundortheconditionalmeasuressuchas𝑃(𝑦𝑖2>𝑧2𝑦𝑖1<𝑧1),(B.3)whichrepresentsthepercentageofpoorhouseholdsinthefirstroundthatescapepovertyinthesecondround.Iftruepaneldataareavailable,wecanstraightforwardlyestimatethequantitiesinB.2andB.3,butintheabsenceofsuchdata,wecanusesyntheticpanelstostudymobility.Tooperationalizetheframework,wemaketwostandardassumptions.First,weassumethattheunderlyingpopulationbeingsampledinsurveyrounds1and2areidenticalsuchthattheirtime-invariantcharacteristicsremainthesameovertime.Morespecifically,coupledwithEquation(B.1),thisimpliestheconditionaldistributionofexpenditureinagivenperiodisidenticalwhetheritisconditionalonthegivenhouseholdcharacteristicsinperiod1orperiod2(thatis,𝑥𝑖1=𝑥𝑖2implies𝑦𝑖1𝑥𝑖1and𝑦𝑖1𝑥𝑖2haveidenticaldistributions).Second,weassumethat𝜀i1and𝜀i2haveabivariatenormaldistributionwithpositivecorrelationcoefficientρandstandarddeviations𝜎𝜖1andσ𝜖2,respectively.Quantity(B.2)canbeestimatedby8Whilehouseholdheadsmaystillincreasetheireducationachievementintheory,thisrarelyhappensinpractice.26𝑃(𝑦𝑖1<𝑧1𝑎𝑛𝑑𝑦𝑖2>𝑧2)=Φ2(𝑧1−𝛽1′𝑥𝑖2𝜎𝜀1,−𝑧2−𝛽2′𝑥𝑖2𝜎𝜀2,−𝜌),(B.4)where2(.)standsforthebivariatenormalcumulativedistributionfunctionand𝜙2(.)standsforthebivariatenormalprobabilitydensityfunction.InEquation(B.4),theestimatedparametersobtainedfromdatainbothsurveyroundsareappliedtodatafromthesecondsurveyround(x2)(orthebaseyear)forprediction,butwecanusedatafromthefirstsurveyroundasthebaseyearaswell.Itisthenstraightforwardtoestimatequantity(B.3)bydividingquantity(B.2)by(𝑧1−𝛽1′𝑥𝑖2𝜎𝜀1),where(.)standsfortheunivariatenormalcumulativedistributionfunction.InEquation(B.4),theparameters𝛽𝑗and𝜎𝜀𝑗areestimatedfromEquation(B.1),andρcanbeestimatedusinganapproximationofthecorrelationofthecohort-aggregatedhouseholdconsumptionbetweenthetwosurveys(𝜌𝑦𝑐1𝑦𝑐2).Inparticular,givenanapproximationof𝜌𝑦𝑐1𝑦𝑐2,wherecindexesthecohortsconstructedfromthehouseholdsurveydata,thepartialcorrelationcoefficientρcanbeestimatedby𝜌=𝜌𝑦𝑖1𝑦𝑖2√𝑣𝑎𝑟(𝑦𝑖1)𝑣𝑎𝑟(𝑦𝑖2)−𝛽1′𝑣𝑎𝑟(𝑥𝑖)𝛽2𝜎𝜀1𝜎𝜀2.(B.5)Thestandarderrorsofestimatesbasedonthesyntheticpanelscaninfactbeevensmallerthanthoseofthetrue(ordesign-based)rateifthereisagoodmodelfit(orthesamplesizeinthetargetsurveyissignificantlylargerthanthatinthebasesurvey;seeDangandLanjouw(2013)formorediscussion).Equation(B.4)canbeextendedtoincorporatethecaseofextremepoverty.Forexample,wecanestimatethepercentageofextremelypoorhouseholdsinthefirstperiodthatescapeextremepovertybutstillremainmoderatelypoorinthesecondperiod(jointprobability)as𝑃(𝑦𝑖1<𝑚1𝑎𝑛𝑑𝑚2<𝑦𝑖2<𝑧2)=2(𝑚1−𝛽1′𝑥𝑖2𝜎𝜀1,𝑧2−𝛽2′𝑥𝑖2𝜎𝜀2,𝜌)−2(𝑚1−𝛽1′𝑥𝑖2𝜎𝜀1,𝑚2−𝛽2′𝑥𝑖2𝜎𝜀2,𝜌),(B.6)where𝑚1and𝑚2standfortheextremepovertylinesinperiod1andperiod2,respectively.Moredetailedderivationsareprovidedinthecitedstudies.
VIP
