Phillips曲線的斷裂 -來自小組數(shù)據的證據 Breaks in the Phillips Curve - Evidence from Panel Data

NBERWORKINGPAPERSERIES

BREAKSINTHEPHILLIPSCURVE:

EVIDENCEFROMPANELDATA

SimonSmith

AllanTimmermann

JonathanH.Wright

WorkingPaper31153

/papers/w31153

NATIONALBUREAUOFECONOMICRESEARCH

1050MassachusettsAvenue

Cambridge,MA02138

April2023

Wehavenothingtodisclose.TheviewsexpressedhereinarethoseoftheauthorsanddonotnecessarilyreflecttheviewsoftheNationalBureauofEconomicResearch.

NBERworkingpapersarecirculatedfordiscussionandcommentpurposes.Theyhavenotbeenpeer-reviewedorbeensubjecttothereviewbytheNBERBoardofDirectorsthataccompaniesofficialNBERpublications.

?2023bySimonSmith,AllanTimmermann,andJonathanH.Wright.Allrightsreserved.Shortsectionsoftext,nottoexceedtwoparagraphs,maybequotedwithoutexplicitpermissionprovidedthatfullcredit,including?notice,isgiventothesource.

BreaksinthePhillipsCurve:EvidencefromPanelData

SimonSmith,AllanTimmermann,andJonathanH.Wright

NBERWorkingPaperNo.31153

April2023

JELNo.C11,C22,E51,E52

ABSTRACT

Werevisittime-variationinthePhillipscurve,applyingnewBayesianpanelmethodswithbreakpointstoUSandEuropeanUniondisaggregatedata.OurapproachallowsustoaccuratelyestimateboththenumberandtimingofbreaksinthePhillipscurve.Itfurtherallowsustodeterminetheexistenceofclustersofindustries,cities,orcountrieswhosePhillipscurvesdisplaysimilarpatternsofinstabilityandtoexaminelead-lagpatternsinhowindividualinflationserieschange.WefindevidenceofamarkedflatteninginthePhillipscurvesforUSsectoraldataandamongEUcountries,particularlypoorerones.Conversely,evidenceofaflatteningisweakerforMSA-leveldataandforthewagePhillipscurve.USregionaldataandEUdatapointtoakinkinthepricePhillipscurvewhichremainsrelativelysteepwhentheeconomyisrunninghot.

SimonSmithJonathanH.Wright

simon.c.smith@DepartmentofEconomics

JohnsHopkinsUniversity

AllanTimmermann3400N.CharlesStreet

DepartmentofEconomicsBaltimore,MD21218

UniversityofCaliforniaandNBER

LaJolla,CA92093-0508wrightj@

andCEPR

atimmermann@

1

1.Introduction

“Therewasatimewheretherewasatightconnectionbetweenunemploymentandin?ation.Thattimeislonggone.”(JeromePowell,2021.)

1

“...gradualismisawell-establishedprincipleforcentralbanksintimesofuncertainty.Whenfacedwithuncertaintyabouttheresilienceoftheeconomy,itpaystomovecarefully.”(ChristineLagarde,2022.)

2

ThePhillipscurveisakeyelementofthenewKeynesianmacroeconomicmodelandiscriticalinhowcentralbanksthinkofthemacroeconomy.Recentlytherehasbeenmuchdebateaboutapotential?atteningofthePhillipscurve,whichcould,inturn,hinderthecentralbanks’abilitytocontrolin?ation.ThegoalofthispaperistoapplyBayesianpanelmethodswithbreakpointstodisaggregatedatainordertorevisittimevariationintheslopeofthePhillipscurve.

Thereareanumberofmotivationsforlookingatdisaggregatedata,whetherbyindus-try,byregion,orbycountry.First,theremaybesomecross-sectionalheterogeneitywhichmightshedlightonthecausesofchangingPhillipscurves.Second,sincedi?erentregionsandsectorsexperiencedi?erentbusinesscycles,thereisextrainformationindisaggregatedatathatenablesustoidentifyslopecoe?cientsandregimechangesmorepreciselythanusingaggregatedataalone.Forexample,

Baietal.

(

1998

)and

SmithandTimmermann

(

2021

)arguethatpaneldataimposingcommontimingofbreaksincreasestheprecisionofbreakdateestimates,evenwhenthee?ectofsuchbreaksisallowedtovaryacrossindividualunitsorvariables.Third,severalrecentpapers(e.g.

Hooperetal.

(

2020

),

Fitzgeraldetal.

(

2020

)and

McLeayandTenreyro

(

2020

))havepointedoutthatifthecentralbankissuc-cessfullytargetingin?ation,thenthiscreatesanendogeneitybiasintheslopeofthePhillipscurve,biasingthecoe?cienttowardszero.Theuseofdisaggregatedatainconjunctionwiththeinclusionoftime?xede?ectsavoidsthisproblem,becausethecentralbankdoesnotspeci?callytargetin?ationinanyoneparticularregionorsector.

3

1ThisquoteisfromFederalReserveChairJeromePowell’spressConference,March17,2021;

/mediacenter/

?les/FOMCpresconf20210317.pdf.

2Thisquoteistakenfromthespeech“Monetarypolicyinanuncertainworld”byChristineLagarde,PresidentoftheECB,at“TheECBandItsWatchersXXIIconference,17March2022.”

3Theproblemwouldnotbesolvedwithdisaggregatedatawithouttime?xede?ects,becauseinthatcasesomeoftheidenti?cationwouldcomefromthetimeseriesdimensionwherethereisendogeneity.

2

Buildingontheseinsights,inthispaperweapplyrecentBayesianpanelbreakmethodstostudyPhillipscurvein?ationdynamics.OurBayesianpanelbreakapproachestimatesthenumberofbreakstothePhillipscurveandthetimeoftheiroccurrence(location).Further,itendogenouslyidenti?esclustersofin?ationserieswithcommonPhillipscurvesandforwhichtheimpactofbreaksissimilar.ThelatterfeatureallowsustoexamineevidenceofconvergenceinPhillipscurves.Finally,ourapproachallowsustoidentifylead-lage?ectsinthetimingbywhichindividualin?ationseriesgeta?ectedbybreaks.

ExistingworkonestimatingPhillipscurvesinpanelsofdisaggregatedatamostlyimposestherestrictionofacommonslopecoe?cient.AnalternativeistoestimatethePhillipscurveforeachindustryorregionseparately,butthisgivesupalotofinformationandtestsconductedonunivariatePhillipscurvemodelsgenerallylacksu?cientpowertodetectbreaksortoestimatetheirdatesprecisely.Anadvantageofourapproachisthatwecantakeamiddleground,anddopartialpooling,whileallowingforsomecross-sectionalvariationintheslopecoe?cients.Forexample,ourmethodologyallowsustoconsidergroupingsbyindustryorgeographicregion,withdi?erentslopecoe?cientsapplyingtoeachgroup.Wecanimposethegroupingsapriori,orthegroupingstructurecanbeestimatedaspartofthemodelingprocess.IfthedatasupportahomogeneousPhillipscurvethatisidenticalacrossallunits,onlyasinglegroupwillbeidenti?ed.Conversely,verystrongheterogeneityinPhillipscurvesacrossindustriesorregionswillleadtoamodelinwhicheachgroupcomprisesasingleunit.Ourmethodologyendogenouslydetermineswhetheranyofthesespecialcasesoranintermediatescenariowithmultipleunitsineachcluster,issupportedbythedata,thusadaptingtothedegreeofheterogeneityfoundinthedata.

ThefocusofouranalysisisonunderstandinghowthePhillipscurvehaschangedovertimeandidentifyingpossibledriversofsuchchange.Acomplexsetoffactorscouldbeatplay,includingchangesinunionizationandwageindexation,exposuretointernationaltrade,andeveneconomicintegration.TheimpactofbreakstothePhillipscurveis,therefore,likelytodependontheunitatwhichin?ationismeasured.Tohelpidentifythesedrivers,wethereforeapplyourestimationapproachtoavarietyofdatasets.Speci?cally,weconsiderUSpricePhillipscurvesusingdisaggregationattheindustryandMSAlevel,andtowagePhillipscurvesatthestatelevel.WealsoexaminePhillipscurvesatthecountrylevelwithintheEuropeanUnion.

Turningtotheempiricalresults,inUSindustrydatacoveringthesample1959Q1-2022Q3,we?ndtworegimechangesinthePhillipscurve;asteepeningaround1972anda?atteningin2001.Moreover,therecent?atteningofthePhillipscurveismorepronouncedforgoods

3

pricesthanforservicesprices.Thesteepeningaround1972comesafteraperiodwhenin?ationhadbeentrendingupforsomeyearsandwhenindexationofwagecontracts,eitherimplicitorexplicit,becamemorecommon.ThiswouldinturnsteepenthePhillipscurve.Meanwhile,thesubsequent?atteningcorrespondstoatimeofgreaterimportpenetration,especiallyfromChina,withChinajoiningtheWorldTradeOrganizationin2001.

4

WhileeconomicintuitionmightsuggestthatchangestotheslopeofthePhillipscurvewouldoccurgradually,China’saccessiontotheWTOmayhavecausedmoreofasuddenbreakwithquitesharpe?ectsdocumentedinstudiessuchas

BenaandSimintzi

(

2022

).Decliningunionizationandthefactthatin?ationisstableatalowlevelcreatinglessofaneedforpayingattentiontoin?ationinwagesettingareotherpossibleexplanationsforthe?atteningofthePhillipscurve.

5

Theseregimechangesthatwedetectareconsistentwithsomeoftheexistingliterature(e.g.

Hooperetal.

(

2020

)),although

Hazelletal.

(

2022

)argueusingstateleveldatathatthePhillipscurvehasbeenconsistently?at.

Ournon-commonbreakestimatessuggestthatthe?rstbreaktothePhillipscurvebasedontheindustry-levelPCEseriesoccursbetween1972and1973whilethebreaktothePhillipscurvebasedontheCPIseriesoccursbetween1971and1974.Asecondbreakoccursbetween

2001and2002.Inbothcases,thetimingofthebreaksis,thus,quitepreciselyestimated.

6

USregional(MSA)dataarenotavailableasfarbackintime,spanningtheshortersample1980-2022.ThismeansthatwecannotexaminethepresenceofPhillipscurvebreaksinthe70sforthisdata.Still,evenwiththisshortersamplecoverage,wemanagetoidentifyaregimechangearound2000.Further,we?ndagainthatMSAswithabove(below)medianratesofimportpenetrationfromChinahaveexperiencedaconsiderablystronger(weaker)?atteningoftheirpricePhillipscurve.These?ndingsareconsistentwithmoregoodscompetitionfromChinaexplainingapartofthe?atteningofthepricePhillipscurve.

BroadlysimilarpatternsarefoundintheEUforasamplethatbeginsin1986andendsin2021.Forthisdatawe?ndevidenceofasinglebreakwhichweestimateoccursin2004atwhichpointtheslopeofthePhillipscurve?attenssigni?cantly.Usingourclusteringmethodology,we?ndthatthePhillipscurveusedtobeparticularlysteepinpoorer(mostly

4

Aueretal.

(

2017

),

StockandWatson

(

2020

),

GilchristandZakrajˇsek

(2019

)and

Firat

(2020

)allshowhowgreatertradeopennesscan?attenthePhillipscurve.

5Whilewedo?nda?atteningbreakinthewagePhillipscurve(estimatedoverthesample1980Q1-2019Q4),itisofsmallermagnitudethanforthepricePhillipscurve,anditcomesin1989,earlierthanwe?ndwithmostpricedata.

6Thecommonbreakapproachcanbeviewedasasimpli?cationthatplacesthebreaktothePhillipscurveatagivendatewithinthebreakintervalidenti?edbythenoncommonbreakmodel.

4

EastEuropean)countriespriortothe2004break,buthas?attenedbymoreinthosecoun-

tries,consistentwithclearevidenceofPhillipscurveconvergenceacrosscountriesthat,earlyinoursample,usedtodisplayaverydi?erentin?ation-unemploymenttrade-o?.

WealsostudynonlinearityofthePhillipscurve,which,asnotedby

Hooperetal.

(

2020

),ismucheasiertodowithdisaggregatedatasincethenationallabormarkethasnotreallybeentightsincethelate1960s,whereasmanyindividualMSAshavehadtightlabormarketsinthistimeperiod.Sincethecurrentpolicydebateisfocusedonsuchtightvaluesofthelabormarket,regionaldataseemslikelytobehelpfulhere.WeconsiderakinkinthePhillipscurveatathresholdofunemploymentratesof5and4.2percent.

7

Usingthesethresholds,we?ndthatthePhillipscurveissteeperinatightlabormarket.

Hooperetal.

(

2020

),

Babb

andDetmeister

(

2017

)and

Leducetal.

(

2019

)also?ndthatthePhillipscurveissteeperinatightlabormarketbutdonotconsidersubsampleinstability.Ignoringbreakshasthee?ectofleadingtounderestimationoftheadditionalsteepnessintightlabormarketsinthemostrecentperiod.

Next,weexploresomeaggregateimplicationsofourPhillipscurveestimates.Ouresti-matesforboththeUSandtheEUimplyessentiallynomissingdisin?ationduringtheGreatRecessionandnomissingrein?ationduringthesubsequentrecoveryyears.Inaddition,we?ndthatasteeper(nonlinear)Phillipscurveinhotlabormarketscombinedwithahighernaturalrateofunemploymentdrivenbyunusuallystrongwagegrowth(

Crumpetal.

2022

)canexplainalmosthalfofthesurgeinU.S.in?ationbetween2020and2022.

8

Finally,weinvestigatetheimplicationsforoptimalmonetarypolicyofthebreakinthePhillipscurvearoundtheturnofthecenturyusingourMSA-levelestimates.ThebreakinducesadditionalparameteruncertaintyinourBayesianframeworkwhichcausesthecentralbanktorespondmorecautiouslytodeviationsintheunemploymentgapasthepolicymakerisuncertainaboutthelinkbetweeneconomicslackandin?ation,inlinewith

Brainard

(

1967

)’s“conservatismprinciple”thatisencapsulatedintheopeningquotefromECBPresidentChristineLagarde.Thepolicy-makercompensatesforthiscautionbyrespondingmoreaggressivelytodeviationsinin?ationfromtarget.We?ndasimilarpatternusingourEUcountry-levelestimates.Theseresultsarerelevantforrecentglobalmonetarypolicyactionsinresponsetothere-openingoftheeconomyafterthelockdownsinducedby

7Forcomparison,

StockandWatson

(2009

)de?neatightlabormarketasanunemploymentgapbelowminus1.5percentwhile

BabbandDetmeister

(2017

)usethesamethresholdsaswedo.

8Thekinked(nonlinear)Phillipscurvee?ectsareanimportantpartofthisexplanation.Interestingly,ourapproachdoesnotdetectabreakaroundtheCovidlockdown,suggestingthatthelinearPhillipscurveremainedquite?atduringthisperiod.

5

thepandemic.

Ouranalysisisrelatedtoalargebodyofresearchontime-variationinthePhillipscurve.Thisliteraturecanbedividedintotwobroadcategories.

9

The?rstapproachcapturestime-variationbyassumingthattheparametersofthePhillipscurvefollowarandomwalk(

Ball

andMazumder

2011

;

MathesonandStavrev

2013

;

Blanchard

2016

;

InoueandWang

2022

).ThesecondapproachestimatestheparametersofthePhillipscurvesubjecttoanassumptionthatbreakdatesareeitherpre-speci?edordeterminedbasedonthesinglebreakpointtestof

Andrews

(

1993

)orbasedoninformalmodelingtechniques,suchasregressionswithrollingwindows(

Roberts

2006

;

Coibionetal.

2013

;

CoibionandGorodnichenko

2015

;

Leducetal.

2017

;

BallandMazumder

2019

;

Gal′?andGambetti

2019

;

GilchristandZakrajˇsek

2019

;

DelNegroetal.

2020

;

Fitzgeraldetal.

2020

;

Hooperetal.

2020

;

CerratoandGitti

2022

;

Hazelletal.

2022

).

10

However,breaktestsconductedonunivariatetimeserieshavelowpower,makingitdi?culttodetectbreaksinthePhillipscurveonindividualin?ationseries.Exploitingtherichinformationinthecross-sectionofpaneldatasetso?erstheopportunityforincreasedpowerandourstudyisthe?rsttoformallyestimatemultiplebreaksinthePhillipscurveinthecontextofsuchpaneldata.

11

Theremainderofthepaperproceedsasfollows.Section

2

introducesthepaneldatasetsusedinouranalysiswhileSection

3

explainsourBayesianpanelapproach,includingestimation,modelselectionandchoiceofpriors.Section

4

presentsourmainempiricalresultsonbreaksintheindustryandregionalPhillipscurves,andSection

5

discussesaggregateimplicationsofourresults.Section

6

conductsasetofrobustnessexercises,whileSection

7

concludes.AdditionalempiricalresultsaredescribedinanAppendixattheendofthepaper.

2.Data

Thissectionintroducesourdataalongwiththedatasourcesusedinourempiricalanalysis.We?rstdescribeourin?ationexpectationsandaggregateunemploymentgapmeasuresbeforeexplainingthein?ationrate,unemploymentrate,andNAIRUmeasureswhichweusefortheUSMetropolitanStatisticalAreas(MSAs)andindustries,aswellasfortheEU.

9AppendixTable

A1

containsalistofsomeofthemainstudiesonvariationinthePhillipscurve.

10

BarnichonandMesters

(

2021

)useasubsamplesplittotracktime-variationinthePhillipsmultiplier.

11Allowingformultiplebreaksisalsocrucialwhenconsideringlongtimeseriessamplesforwhichstation-arityislesslikelytohold.

6

2.1.In?ationexpectations,unemploymentrates,andNAIRU

Wesourcefour-quarter-aheadConsumerPriceIndex(CPI)in?ationexpectationsfromBlueChipEconomicIndicators.Thesedatagobackto1985.Between1980and1985,weuseProducerPriceIndex(PPI)in?ationexpectationsfromthesamesource.Before1980,weusedatafromLivingstonwhichisonlyupdatedeverysixmonthsandsowesimplyrepeatobservationsinthetwocorrespondingquarters,e?ectivelyassumingthatin?ationexpecta-tionsremainthesameineach6-monthperiod.BecauseU.S.in?ationexpectationsareonlymeasuredfortheaggregatepriceindexasopposedtoattheregionalorsectorallevel,wecanonlyusethesein?ationexpectationsdatainspeci?cationswithouttime?xede?ects.

Weusetheend-of-quartermonthlyaggregateunemploymentgap,measuredasthedif-ferencebetweentheunemploymentratefromtheU.S.BureauofLaborStatistics(BLS)andtheNAIRUestimate(fromtheCongressionalBudgetO?ce).ThesedatabegininJanuary1949andendinSeptember2022.

Wesourcetheannualcountry-levelunemploymentrateandNAIRUestimatesforthe28EUmembercountries(thecurrent27plustheUKwhichwasamemberuntilrecently),andhencetheunemploymentgaps,forthesampleperiod1965-2021fromtheDGECFIN/AMECO—theEuropeanCommission’smacroeconomicdatabase.

12

Fortheregionalanalysis,weobtainannualunemploymentratedatafrom1980to2022for22MSAsfromtheBLS.Wealsousetheendofquartermonthlyunemploymentrateforall51states(includingtheDistrictofColumbia),alsoobtainedfromtheBLS.ThesedatabegininJanuary1980andendinDecember2019.

2.2.Pricedata

2.2.1.MSAlevel

WesourcemonthlytotalCPIsfor22MSAsfromtheBLS.Weconstructannuallevelsastheaverageofallmonthlyobservationsinthecorrespondingyear.

13

Next,weconstructannualin?ationratesaslog(CPIit/CPIit-1)×100inwhichCPIitdenotesthelevelfortheithMSAinyeart.

12WethankMicheleLenzaforhelpingusaccessthesecountry-levelNAIRUestimates.

13DataforallbutafewMSAsarecollectedonlyineitheroddorevenmonths.See

/opub/hom/cpi/pdf/cpi.pdffordetailsofthecompletemethodologyand

/cpi/additional-resources/geographic-sample.htm.forthegeographicde

?nitions.

7

Annualunemploymentratesareconstructedastheaverageofallmonthlyobservations

inthecorrespondingyear.Oursampleforthesedatabeginsin1980andendsin2022,butformanyMSAsthedataonlystartin1990.

2.2.2.Industrylevel

WeusequarterlyPersonalConsumptionExpenditurespriceindexes(PCE)for16industrycomponents,similartothoseanalyzedby

StockandWatson

(

2020

),sourcedfromtheBureauofEconomicAnalysis(BEA).

14

Oursampleis1959:Q1-2022:Q3.Weconstructannualizedquarterlyin?ationratesaslog(PCEi,t/PCEi,t-1)×400.

FromtheBLS,wesourcemonthlyCPIin?ationfor31“l(fā)evel3”industries,ascurrentlyformulated,beginninginJanuary1954andendinginSeptember2022,thoughnotallseriesgoallthewayback.Weconstructourannualizedquarterlyin?ationrateobservationsfromendofquartermonthlyobservationsaslog(CPIi,t/CPIi,t-3)×400.

2.3.ImpliednationalPhillipscurveslopes

Hazelletal.

(

2022

)showthattheregionalPhillipscurveslopecanbedividedbytheexpen-ditureshareonnontradeablestoobtainthenationalPhillipscurveslope.Weusethe31CPIindustryweightstocomputetheexpenditureshareonnontradeables.Wefollow

Hazelletal.

(

2022

)byassigningthefollowingseriestonontradeables:FullServiceMealsandSnacks,LimitedServiceMealsandSnacks,Foodatemployeesitesandschools,Foodfromvend-ingmachinesandmobilevendors,Otherfoodawayfromhome,Electricity,Utility(piped)gasservice,Waterandsewerandtrashcollectionservices,Householdoperations,Medicalcareservices,Transportationsservices,Recreationservices,Educationandcommunicationservices,Otherpersonalservices,andShelter.Theexpenditureshareonnontradeablesistherefore69.1percent.Doingthesameusingthe16PCEcomponentweights,theexpendi-tureshareonnontradeablesis74.3percent.

14Thetwocategories–HousingandHouseholdutilities–havesincebeenreplacedbyone:Housingandutilities.

8

2.4.Wagedata

Following

Hooperetal.

(

2020

),wedirectlycomputeaveragehourlyearnings(AHE)foreachofthe50statesandtheDistrictofColumbiausingthelatest(2019)CEPRuniformextractfromtheCurrentPopulationSurvey(CPS)

15

.Aggregatingfrommonthlydata,weconstructquarterlydatafrom1980:Q1through2019:Q4,fromwhichweconstructquarterlyannualizedwagein?ation.

2.5.EUdata

Wesourceheadline(aswellastotalgoodsandtotalservices)annualin?ationratesforour28countries(the27currentmembersandtheUK)fromtheECBstatisticalwarehouse.Oursamplebeginsin1986andendsin2021.

2.6.Groupstructure

Wewillbeinterestedingroupheterogeneity,witheitherthegroupallocationimposedac-cordingtopre-determinedselectioncriteria,ordeterminedbytheBayesianalgorithmaspartoftheestimationprocess.

The16PCEsectorsaresplitintogoods–Motorvehiclesandparts,Furnishingsanddurablehouseholdequipment,Recreationalgoodsandvehicles,Otherdurablegoods,Foodandbeveragespurchasedforo?-premisesconsumption,Clothingandfootwear,Gasolineandotherenergygoods,andOthernondurablegoods–andservices–Housingandutilities,Healthcare,Transportationservices,Recreationservices,Foodservicesandaccommoda-tions,Financialservicesandinsurance,Otherservices,andNPISH.

Wealsosplitthe28EUcountriesintorichandpoorcountrieswithrichcountriesde?nedascountrieswithrealGDPpercapitade?atedbyPPPin2019abovetheEUaverageandpoorcountriesde?nedastherest.TherichcountriesincludeLuxembourg,Ireland,Denmark,Netherlands,Austria,Germany,Sweden,Belgium,Finland,France,andUK.

16

15Thedataareavailablefrom/cps-uniform-data-extracts/.

16ThepoorcountriesthereforeincludeMalta,Italy,CzechRepublic,Spain,Cyprus,Slovenia,Slovakia,Romania,Portugal,Poland,Bulgaria,Estonia,Lithuania,Latvia,Hungary,Greece,andCroatia.

9

3.Methodology

Ouranalysisexaminesthreedi?erentBayesianpanelspeci?cations.The?rstisourbase-linepooledpanelmodelwithmultiplebreakpoints.Thi