the lost generation? scarring after the great...
TRANSCRIPT
TheLostGeneration?ScarringaftertheGreatRecession
JesseRothstein1
UCBerkeley,NBER,andIZA
May2019
Abstract:
Iinvestigatemedium-andlong-termimpactsoftheGreatRecessiononpost-recessioncollegegraduates. Most so-called “scarring” models emphasize effects of initial conditions thatattenuateoverthefirstdecadeofaworker’scareer.Butearlycareerrecessionsmayalsohavepermanent effects. I decompose the recent cohorts’ experience into transitory time effects,medium-term scarring, and permanent cohort effects. Cohort effects are strongly cyclical.Medium-termscarringexplainsonlyhalfofthiscyclicality.Thelong-runcumulativeeffectoftherecessionongraduates’employmentismorethantwiceaslargeastheimmediateeffect.
[email protected],RachelYoung,andNathanielRubyforexcellentresearchassistance;toSaraHinkley,RobValletta,TillvonWachter,andDannyYaganforhelpfuldiscussionsandcomments;andtoconferenceandseminarparticipantsatBerkeley,Georgetown,theUniversityofIllinois,LSE,Michigan,NBER,UCBerkeley,andUCMerced.
1
I. Introduction
Prolongedunemploymentmayhavelong-runconsequencesforfutureemployabilityand
productivity,perhapsparticularlyso foryoungworkers forwhomweakconditionsmaymean
reducedaccess toon-the-jobexperienceand training.Consistentwith this, collegegraduates
whoenterthelabormarketduringdownturnsare“scarred,”withloweremploymentandwages
throughout theirearlycareers thansimilar individualswhoenteratbetter times (Kahn2010;
Oreopolous,vonWachter,andHeisz2012;SchwandtandvonWachter,2019).
Iexaminethemedium-andlong-runeffectsofthe2007-2009GreatRecessiononnew
entrants’labormarketoutcomes.Followingthescarringliterature,Ifocusoncollegegraduates,
whosehumancapitalisunlikelytohavebecometechnologicallyobsoleteduringthedownturn.
The Great Recession provides a natural laboratory due to its depth and duration. The
unemploymentraterosefrom4.4percentinMay2007to10.0percentinOctober2009,then
didnotreach4.4percentagainuntilMarch2017.Employmentratesworsenedatthesamerate
as unemployment during thedownturn, but havebeenmuch slower to recover; to this day,
prime-ageemploymentremainsbelowitspre-recessionpeak.
Duringthedownturn,newlabormarketentrantshadfewjobopportunities.Jobopenings
wereatleast20percentbelowtheirpre-recessionpeakfornearlyfiveyears,andthehiringrate
remainedsignificantlydepresseduntillate2014.Theshareofindividualsaged20-25whowere
neitherworkingnorinschoolrosebynearlyhalf,from13.1%in2007toahighof19.0%in2013.
Comparingyoungerandolderworkersinthesamelabormarket,theemployment,annual
earnings,andwagesofthosewhoenteredthelabormarketduringandaftertherecessionare
2
lowerthanwouldbeexpectedgiventheexperienceofoldercohortsatthesametime.College
graduates whose careers started in 2010 and 2011 have employment rates two percentage
pointslowerthantheage-andtime-adjustedemploymentratesofpre-recessioncohorts,and
thosewhoworkhaveannualearningsthataretwopercentlower.Employmentdeclinespersist
foreventhemostrecentlabormarketentrants,whileannualearningsandwageshaverecovered
sincetheimmediatepost-recessionentrants.
Employmentratesoftherecessionandpost-recessionentrantshaveremainedloweven
inmore recent data,whenolder cohorts have recovered.Accordingly, in an age-time-cohort
decompositionIfindthatpurelytransitoryeffectsaccountforonlyabout40%oftheemployment
declineforcohortsenteringthelabormarketin2009-2011,andnoneofthedeclineforcohorts
enteringlater.Theremainderisattributedtodeclinesincohorteffects.
I next investigate whether existing scarring specifications capture these patterns. An
importantfeatureofthespecificationsthathavetypicallybeenestimatedisthatscarringeffects,
whilelonglasting,arenotpermanent.Theeffectsoftheinitialemploymentratearefrequently
estimated todissipateby the individual’s10thyear in the labormarket (e.g.,Borgschulteand
Martorell 2018; Kahn 2010;Oreopoulos et al. 2012). Thus, insofar as thesemodels correctly
characterizethescarringeffectoftheGreatRecession,theyimplythatmuchofthedamageto
thepost-recessioncohortshasalreadybeenborne,andthattherecessionwillnotcastalarge
shadowgoingforward.
IreproducescarringestimatesonemploymentinCurrentPopulationSurveydata,using
similarspecificationstothoseusedbypreviousauthors.GreatRecessionscarringiscloselyinline
withextrapolationsofestimatesbasedonpre-2007data.WhetherIusethefullsampleorjust
3
thepre-recessionsubsample,Ifindmeaningfulmedium-termscarringthatfadesoutoverseveral
years.However,thisaccountsforonlyabouthalfofthedepressedemploymentoftherecent
cohorts,relativetoothercohortsinthepost-recessionlabormarket.Theremaindermanifests
aslowercohorteffectsfortherecentcohorts.Thesecohorteffects,whichhavegenerallybeen
treatedasnuisanceparametersinpastscarringwork(e.g.,Oreopoulosetal.,2012),captureany
effectofearlycareerconditionsthatdoesnotfadeoutovertheearlycareer.Ishowthatthese
effectsarecyclicalevenwhenmedium-termscarringiscontrolledseparately,andrepresenta
quantitativelyimportantcomponentoftheeffectsofrecessions.
I closewith a calculation of the contribution of these long-lasting effects to the total
cumulative impact of the Great Recession, extrapolating the cohort coefficients through the
decadesremainingintheaffectedcohorts’potentialcareers.Myestimatesimplythat,through
a combination of short-,medium-, and long-run effects, theGreat Recessionwill reduce the
employment rates of college graduates over the 2007-2060 period by a total of over 15
percentagepoint-years.Nearlyhalfof this loss reflectsmedium-and long-termeffectsof the
recessionthatmanifestafter2014,whentheunemploymentratewasnolongerelevatedbut
workerscarrythescarsoftheearlierperiod.Themedium-termeffectswillgraduallyfadeover
thenextfewyears,whilethelong-runeffectswillpersistaslongastherecessionentrantsremain
inthelabormarket.Cumulatively,thelatteraremuchmoreimportantthantheformer.Effects
thatwillmanifestafter2025accountfornearlyone-thirdofthetotaldamage.
Thisevidencesupportstheviewthatbusinesscycleshaveimportanthysteresiseffects.
DeLongandSummers(2012)arguethatifpost-recessionhangoversarelargeenoughcounter-
cyclical fiscal policies canbe self-financing.Myestimatesof permanent effects onnew labor
4
marketentrantscaptureonlyasmallcomponentofpotentialhysteresis–andthusshouldbe
addedtoYagan’s(forthcoming)complementaryevidenceofeffectsonolderworkers–butare
ontheirownlargeenoughtoimplyameaningfulfiscaloffsetforcountercyclicalpolicies.
II. TheGreatRecessioninthelabormarket
Theunemploymentrateroseby6.5percentagepointsbetweenmid-2007andlate2009,
(AppendixFigureA-1). It thendeclinedroughly linearlyaftermid-2010. Ithasbeenbelow6%
sincethethirdquarterof2014andbelowitspre-recessionlevelsinceearly2017.
Theprime-age(25-54)employmentratefellbyoverfivepercentagepointsbylate2009,
andwasmuchslowertorecover.Onlyhalfofthedeclinehadbeenerasedbytheendof2015;at
thiswritingtheemploymentrateis79.8%,stillhalfapointbelowthe2007peak.
Youngpeoplefaredparticularlypoorly.ThefirstpanelofFigure1showstheemployment
rateseparately for thoseaged25-30andthoseaged31-54.Youngerworkers’employment is
muchmorecyclicallysensitivethanthatofolderworkers;theiremploymentratefellmorethan
fivepercentagepointsduringtheGreatRecession,thenrecoveredrelativelyquickly.Thesecond
paneldistinguishescollegegraduatesandnon-graduateswithinthe25-30group.Employment
ratesarehigherandlesscyclicalforcollegegraduates.However,thegraduateemploymentrate
declinedsubstantiallyduringtherecession,thendidnotreboundatallthroughthefirstseveral
yearsoftherecovery.Bytheendof2014,non-graduateshadmadeupone-thirdoftheground
lostduringtherecession,whileyounggraduates’employmentwasstillbelowitsend-of-recession
level.
5
FigureA-2showsaveragerealhourlywagesforthesamegroupsbrokenoutinFigure1.
Allfourmeanwageseriesfellaftertherecession(thoughnotduringit,duetochangesinthe
compositionofworkers; seeDaly,Hobijn, andWiles2012),witha largerdecline for younger
workersandsimilardeclinesforyounggraduatesandnon-graduates.Olderworkers’wageshave
more than recovered in recent years, but young workers’ wages, both graduates’ and non-
graduates’,haveonlyapproachedtheir2007 localmaximaandremainwellbelowtheir2000
peaks.Incontrasttotheemploymentseries,however,therecoveryofgraduates’wageswasnot
meaningfullyslowerthanthatofnon-graduates.
III. Scarringandhysteresis
Aninfluentialliteratureexaminesthepersistenteffectsoflabormarketconditionsatthe
timeoflabormarketentry(e.g.,Kahn2010;Oreopoulosetal.2012;SchwandtandvonWachter,
2019;BorgschulteandMartorell2018;Oyer2006,2008).Thisresearchgenerallyfindssizable
persistenteffectsonemployment,wages,andearningsthatfadeoutoverthedecadefollowing
graduation.1Oneinterpretationisthatnewgraduatesaredelayedinbeginningtoclimbthejob
ladderduringrecessions;anotheristhatlabormarketweaknessatcareerstarttranslatesintoa
weakbargainingpositionformanyyearsthereafter(BeaudryandDiNardo1991).Manyofthe
specificationsusedinthiswork(e.g.,Oreopoulosetal.2012)includecohorteffectsascontrols,
but treat them as nuisance parameters, focusing on interactions between the entry
1Anotherrelevantliteratureexaminestheeffectsofjobdisplacementonsubsequentearnings(Jacobson,LaLonde,andSullivan1993;vonWachterandBender2006;Schmieder,vonWachter,andHeining2018).Thesestudiesalsofindpersistenteffects.DavisandvonWachter(2012)findthatthedisplacementeffectislargerinrecessions.
6
unemploymentrateandearly-careerpotentialexperience indicators. Incontrast, Iemphasize
thecohorteffectsaskeycomponentsofthepersistenceofrecessionimpacts.
The so-called “scarring” studies are a small component of an immense literature
examininghysteresismoregenerally.ClarkandSummers(1982)concludeparticipationisfairly
persistent:“Workersdrawnintothelabourforcebycyclicalupturnstendtoremainevenafter
theboomhasended.Theconverseistrueforshockswhichreduceemployment,”(p.842).Yagan
(forthcoming) findsthatworkers inareasthatexperienced largerGreatRecessionshockshad
worse outcomes 6-8 years later, not attenuated bymobility to healthier labor markets.My
analysiscomplementsYagan’sbyexaminingyoungworkers–hisestimatesarebasedonworkers
alreadyover30in2007–andrelatingthepersistentimpactsoftheGreatRecessiontothoseof
pre-2007fluctuations.
IV. Data
IuserepeatedcrosssectiondatafromtheCurrentPopulationSurvey(CPS)toexamine
individuals born between 1948 and 1995 and observed at ages 22 to 40 between 1979 and
November2018.
Myprimaryoutcomeistheemploymentrateofyoungworkers,measuredinthemonthly
CPS.Ialsoanalyzelogrealannualearnings,fromtheAnnualSocialandEconomicSupplement
(theMarchCPS),andlogrealhourlywages,fromtheCPSOutgoingRotationGroup(ORG),for
thosewithpositiveweeklyorannualearnings.Dataprocessinganddefinitionsarediscussedin
theAppendix.
7
Foreachoutcomeandsample,Iaggregatetotheeducation-state-year-birthcohortcell.I
usetwoeducationgroups:Bachelor’sdegrees(BAs)andabove,andlessthanaBA.Followingthe
scarringliterature,Ifocusontheformer,butIpresentresultsforthelatteraswell.Imergeonto
eachcellthecontemporaneousstateunemploymentrateandtheunemploymentrateintheyear
ofthecohort’slabormarketentry.2
Thereare three important limitations to theCPSdata formypurposes.First, Idonot
observethedatethatarespondentfirstsearchedforwork.Followingthescarringliterature,I
use the unemployment rate at age 22 (18 for non-graduates). This abstracts from potential
endogeneityintheentrydate.
Second, Iobserveonly the respondent’s current stateof residence,notwherehe/she
lived at labor market entry. Schwandt and von Wachter (2019) explore the bias in similar
specificationsfromfailingtomeasurelocationasofgraduation,andconcludethatitis“unlikely
tobeverylarge.”IexploresensitivitytomobilityintheAppendix.
Third,someofthoseobservedwithoutdegreesat22willearnthemlater.Thiscreates
differentialselectionbyageinthepopulationsoverwhichmeanoutcomesarecomputed.Figure
A-3showsthegraduatesharebybirthcohortandyear.Thedegreeattainmentratefor22-year-
oldsisonlyhalforone-thirdofwhatwillbeseenforthesamecohortatage30.Forcohortsborn
in the 1960s, the greatmajority of degreeswere earned by age 24. Therewas a substantial
increase inpost-24degreeattainmentaround the1968birthcohort (turning22 in1990; see
Bound,Lovenheim,andTurner2012),butstillthegreatmajorityofdegreesareawardedby26.
2Stateunemploymentratesareavailableonlyfrom1976;forcohortsthatenteredbeforethisdate,Iusethenationalrate.
8
Since themid-1990s, final attainmenthas grown substantially,withmostof this growth is in
degrees awarded by age 24.3 There is no indication of a surge in attainment among Great
Recessioncohorts(thoughotheranalyses,notshownhere,indicatethatthereareincreasesin
collegegoingthatevidentlydoesnotleadtoaBAdegree).Together,thesefactsmakeitplausible
thatageeffects,includedinallofmyspecifications,willabsorbmostofthebiasincohortmean
outcomesaveragedacrossvariousages.
I take twoapproaches tominimizing remaining compositionbias. First, I include inall
regressionsaninverseMillsratiocomputedfromthecohort-by-ageattainmentrate(or,fornon-
graduates,fromitscomplement).Becausemyspecificationsincludebothcohortandageeffects,
thecoefficientonthiscontrol is identifiedfromchanges intheattainmentrateoveragethat
differacrosscohorts.Inasimplebivariatenormalselectionmodel,thiswillabsorbanybiasfrom
changing selection (Gronau 1974, Card and Rothstein 2007). Second, I have reestimatedmy
primaryspecificationsexcludingobservationsfromagesunder24.Resultsarelargelyunchanged.
V. Methodsandresults
A. Age-time-cohortdecomposition
LetYsatcrepresenttheemploymentrateofcollegegraduatesinstatesatageaintimet
frombirthcohortc(indexedbytheyearthatthegroupenteredthelabormarket,orc=t-a+22).I
estimateage-time-cohortdecompositions:
3Post-24degreesmaybegrowingforthepost-1988birthcohorts(aged24after2012).Itwillbeseveralyearsuntilwecanfullymeasurethis,andinthemeantimeithaslittleimpactonmyCPSsamples.
9
!"#$% = ' + )$ + *# + +% + ," + -. /#$% + 0"#$%. (1)
Here,)$areasetoffixedeffectsforyears,*#areageeffects,and+% arebirthcohortfixedeffects.
,"representsstateeffects,while. ⋅ istheinverseMillsfunctionappliedtothestate-cohort-
ageattainmentrate,/#$%.Thetimeeffects)$capturebothaggregatedemandandanysupply-
side factors that are common across age groups, potentially including the hysteresis effects
documented by Yagan (forthcoming). Cohort effects +% are permanent, and capture any
differences between one birth cohort and another in the same labormarket, beyond those
reflectedintheageprofile*#.
Thefullsetofage,time,andcohorteffectsisnotidentifiedduetothelineardependency
amongthem(see,e.g.,Schulhofer-Wohl,2018).Inormalizethecohorteffectsforthe1991and
2000entrycohorts(bornin1969and1978,respectively)tobeequal.Thisforcestheestimated
trendtobezeroacrossthesecohorts;allestimatescanbeseenasrelativetothetruetrendover
thisperiod.Myanalysesofcohorteffectsalwaysincludeordifferenceoutlineartimetrendsto
absorbthisnormalization.
I am interested in the degree to which poor employment rates of young graduates
following the 2007-2009 Great Recession can be attributed to transitory demand shocks,
capturedby)$ inequation(1),versuspermanentdifferencesbetweenpost-recessioncohorts
and other cohorts facing the same demand conditions, which would appear in+%. To avoid
confoundingtheageandtimeeffectswiththeexperienceofpost-recessioncohorts,whoare
observed only when young, inmy preferred specifications I estimate (1) using data only on
cohortsenteringthemarketin2000andearlier.Ithenholdthe)$and*#vectorsfixedwhen
10
estimating the cohort effects formore recent birth cohorts. I obtain nearly identical results,
however,whenIsimplyestimate(1)ontheentiresample.
ThesolidlineinFigure2showsthecohorteffectsfromthebasicspecification,appliedto
collegegraduatesandusingonlypre-recessioncohortstoestimatetimeandageeffects.4Cohort
effectsarelargelystable,decliningslightlybutsmoothly,fortheentrycohortsbetween1975and
2004.Theoverallslopeisdefinedbymynormalizationandnotbythedata,sothesmoothdecline
couldequallywellbeanincrease.However,anychangeintheslopeisidentified.Thereisasharp
downwardturnforcohortsenteringafterabout2005.Thecumulativedeclinefromthe2004to
the2015cohorts,relativetotheslopebetween1993and2004,isnearlyfivepercentagepoints,
withmostofthisdeclineoccurringbetween2006and2010.Therearealsosmallerdeclinesin
cohorteffectsinearlierrecessions.
Although themost recent cohorts– those turning22after2014–have faceda labor
marketthatwasmuchstrongerthanthatfacedby2009-2012entrants,theyhavenotrecovered
anyofthe lostground,andcontinuetofarequitepoorlyrelativetopastcohortsobservedat
similarages.Bycontrast,thepuretimeeffects(FigureA-5)hadfullyrecoveredtheir2007levels
by 2014. In other words, the model indicates that the slow recovery of college graduates’
employment(shownasthesolid line inFigureA-5)since2011isnotevenlydistributedbut is
specifictoyoungercohorts.Oldercohortswhowerealreadyestablishedinthelabormarketby
2007hadfullyrecoveredby2013.
4FiguresA-4andA-5showtheageandtimeeffects,respectively,whileFigureA-6showsthatcohorteffectsarenearlyidenticalwhenthefullsampleisusedtoestimate(1).
11
B. Excesssensitivityandscarring
In specification (1), the decline in cohort effects seen in Figure 2 implies that post-
recession cohortswill have lower employment rates in every year of their careers than pre-
recessioncohorts,afteradjustingforanydifferencesintimeeffects.Inotherwords,therecession
has a large, permanent scarring effect, totaling four to five percentage points, on the post-
recessioncohorts’annualemploymentrates.
However,equation(1)doesnotallowfortwoalternativeexplanationsthatwouldpoint
torosierfutures.First,youngpeople’semploymentismorecyclicallysensitivethanthatofolder
people.Becausethepost-recessioncohortsareobservedonlyatyoungages,andprimarilyat
timeswhen the unemployment rate is high, excess sensitivitywouldmanifest as downward-
biasedestimatedcohorteffectsforthesecohorts.
To address this, I augment (1) with unemployment rate-age interactions. I create a
sequenceofageindicatorsgj(a),j=1,…5,whereg1(a)isanindicatorforages22-23,g2(a)for24-
25,g3(a)for26-27,g4(a)for28-29,andg5(a)for30-31.Ithencontrolforinteractionsbetween
theseindicatorsandURst:
!"#$% = ' + )$ + *# + +% + ," + -. /#$%
+ 34 5 ∗ 78"$ ∗ 94:4;< + 78"$ − 78$ > + 0"#$%. (2)
Ialsoincludeamaineffectforthedeviationofthestateunemploymentratefromthenational
rate, 78"$ − 78$ ,tocaptureeffectsoflocalconditionsonthoseaged32andover.5
5The94 coefficientsareinterpretedastheeffectofthestateunemploymentrateonthosefromagegroupj,relativetoitseffectonthoseover32.Becausetheunemploymentratemaineffectisspecifiedasthedeviationofthestatefromthenationalrate,thetimeeffects)$captureanyeffectofthenationalunemploymentrateonthoseaged32andolder.Thus,thespecification
12
Theestimated94s areplotted,with confidence intervals, in Figure3A. They show the
expectedexcess sensitivity:Employmentof theyoungestcollegegraduates fallsbyabout0.5
percentagepoint forevery1.0percentagepoint increase in theunemployment rate,but the
effectdeclinesroughlylinearlywithage.
These estimates are based on the pre-recession cohorts, so generally on lower early-
careerunemploymentrates thanprevailedduring theGreatRecession.Thedots inFigure3A
showestimatesbasedonthefullsample.Theseindicatesomewhatmoresensitivity,suggesting
thattheyoungestgraduatesdidevenworsethanexpectedduringtheGreatRecession.Butthe
extrapolationfrompre-2007patternsgenerallyworkswell.
Theshort-dashedlineinFigure2showstheestimatedcohorteffectsfromspecification
(2).Theadjustmentmakesonlyamodestdifference,withslightlybettercohorteffectsforthe
2001to2010entrycohorts.Itdoesnothingtoimprovetheestimatedcohorteffectsforthemost
recententrants–notsurprising,sincetheunemploymentratehasnotbeenelevatedduringmost
of their time in the market. Overall, this specification does not indicate that the apparent
downturnincohorteffectsinthebasespecificationwasanartifactoffailingtoallowforexcess
sensitivityofyoungerworkerstoeconomicconditions.
A second explanation for the apparent decline in cohort effects is that the recession
cohortsmayhavebeenscarredbytheirinitialexperiencesbut,asimpliedbypastestimates,the
scarringeffectwillfadeawaywithage.Becausethepost-recessioncohortsareobservedonlyat
young ages, the estimated cohort effects from specification (1) will reflect the early-career
allowsthenationalunemploymentrateandthestatedeviationfromthattohavedistinctmaineffects,butforcestheirageinteractionstobethesame.
13
impactsofelevatedunemploymentin2007-2015,andmayoverstatethepermanentdamageto
thepost-recessioncohorts.
Again,Iaddressthisbyincludingtermsinthedecompositionthatalloweffectsofentry
conditions to fade out gradually. The specification replaces the contemporaneous
unemploymentrateURstin(2)withtheunemploymentrateatlabormarketentry(i.e.,atage
22),UR22sc:
!"#$% = ' + )$ + *# + +% + ," + -. /#$%
+ 34 5 ∗ 7822"% ∗ @4:4;< + 7822"% − 7822% > + 0"#$%. (3)
The@4 coefficientsrepresentage-cohort-stateinteractions.6Thus,thismodelhastwowaysthat
economicconditions inthefirstyearofthecareercanhavepersistenteffects:Theycanhave
medium-term consequences through the @4 coefficients, and they can have permanent
consequencesviathecohortcoefficients+%.
Medium-termpersistenceisimportantlydifferentin(2)and(3).In(2),atransitoryshock
inthasauniformeffectvia)$andadditionaleffectsonthoseinthefirsttenyearsoftheircareers
through94,butnoneoftheseeffectspersistint+1.In(3),ashockhasthesameuniformeffect
via)$ butmay have additional effects on the newest entrants; if so, these workers remain
damagedforuptotenyears.Bycontrast,thosewhoenteredint-1facenopersistentdamagein
t+1orthereafter.Inprinciple,aspecificationwithbothtypesofpersistenceisidentified,butin
6Asbefore,Iomitaninteractionwiththeindicatorforthoseaged32andabove,andIincludeamaineffectfortheunemploymentrateatcohortentry,deviatedfromthenationalrateinthatyear.Thisensures thatanypermanenteffectsof thenationalunemployment rateatentry iscapturedbythe+% cohorteffects,whileallowingthedeviationofthestatefromthenationalunemploymentratetohaveitsowneffectthrough>.
14
practicethetwoaredifficulttodistinguishaseachpredictssustainedpooroutcomesforyoung
workers.
Figure3Bshowstheestimated@4 coefficientsfrom(3).Weseestrongeffectsofinitial
conditionsonearlycareeremployment,fadingoutoveraboutsixyears.7Again,theseareslightly
largerbutbasicallysimilarwhen I includethepost-recessioncohorts in thesampleaswhen I
extrapolatefromearliercohorts,indicatingthatnonlinearityintheunemploymentrateeffectis
notamajorsourceofbias.
Thecohorteffectsfromthisspecification,plottedaslongdashedlinesinFigure2,show
a smaller, though still quantitatively important, decline in the period surrounding the Great
Recession. This indicates that the earlier specifications may have overstated the degree of
permanent damage by failing to allow for persistent but non-permanent effects of entry
conditions. In theadjusted series, there is still a sharp change in the series around the2005
cohort,butthesubsequentdeclineismoregradualandlinear.Thedeclinefrom2006-2010is
60%aslargeasinthebaselinespecification.Declinesincohorteffectsduringpriorrecessionsare
alsoattenuated.
AppendixFigureA-13showscohorteffectsseparatelyformenandwomen,usingversions
ofspecifications(1)and(3)thatinteractage,cohort,andscarringvariableswithgender.Figure
A-12showstheassociatedscarringcoefficients.Thepost-recessiondeclineincohorteffectsis
larger for women than for men though apparent for both; semi-transitory scarring is more
7Itispossiblethatthelater-agescarringcoefficientsareattenuatedduetogeographicmobility.My@4 estimatesaresimilarwhen,followingvonWachterandHeisz(forthcoming),IinstrumentforURscwithameasurethatabstractsfrommobility.Seetheappendix.
15
importantformenthanforwomen,butevenwhenitisaccountedforweseedeclinesinboth
maleandfemalecohorteffectsfollowingtherecession.
C. Cyclicalityofestimatedcohorteffects
Toquantify the cyclicalityof the cohorteffects in Figure2, Table1 shows time series
regressionsof the threesetsofestimatedcohorteffectson thenationalunemployment rate
whenthecohortwas22,controllingforaquadratictimetrend(whichabsorbsthenormalization
ofthecohorteffects).Standarderrorsallowforautocorrelationoftheerrortermsatuptothree
lags.
Column1showsresultsfortheinitialdecomposition,(1).Eachpercentagepointincrease
intheunemploymentrateatthetimeoflabormarketentryisassociatedwitha0.22percentage
pointpermanentdeclineinthecohortemploymentrate.Thisisonlyslightlyreducedwhenthe
specificationallowsforexcesssensitivitybutdeclinesto-0.11andbecomesinsignificantwhenit
allowsforfadeoutofearly-careerscarringeffects.Inotherwords,medium-termscarringeffects
accountforabouthalfofthecyclicalityofscarringeffectsfrom(1);theremainderareimprecisely
estimated but still appear modestly cyclical. As I discuss below, the remaining response is
quantitativelyquiteimportant.
Column4presentsresultsfromaspecificationthatcombines(2)and(3),allowingforboth
excess sensitivityandearly-career scarring. In this specification, the94 coefficients, capturing
excesssensitivity,aresimilartothoseseeninFigure3A,butthe@4 coefficientsarenearzero.
Cohorteffectcyclicalityisasinspecification(2).
ThelowerrowsofTable1,PanelA,comparetheestimatedimpactoftheGreatRecession
towhatthecyclicalsensitivitymodelwouldpredict,bycomparingthechangebetweenthe2006
16
and2010entrycohorts,relativetothatbetweenthe2002and2006cohorts,tothepredicted
change.Themodelcapturestheexperienceofthe2010cohortreasonablywell,thoughincolumn
3thepredicteddeclineissomewhatlargerthantheactualone.8This isanindicationthatthe
parameters of the decompositions (1)-(3) are reasonably stable through theGreat Recession
period–estimatesbasedonpre-recessioncohortspredictaccurately theexperienceofmore
recentcohorts.
D. Earningsandwages
Ihavealsoconductedparallelanalysesoflogannualearningsandloghourlywages,in
eachcaseconditionalonemployment.EstimatesareshowninFiguresA-7andA-8.Cohorteffects
onannualearningsdeclinenotablyforthe2007-2012entrycohorts,whileonlythe2009entry
cohortseesadeclineinhourlywages(netoftransitorytimeeffects).Ifindstrongerevidenceof
excesssensitivityandscarringeffectsonhourlywagesthanonannualearnings,thoughforeach
outcome allowing for scarring dramatically improves the estimated cohort effects for post-
recessionentrants.
ThelowerpanelsofTable1showtheassociationbetweeneachsetofcohorteffectsand
the entry unemployment rate. Earnings but not wages covary with unemployment at entry,
thoughtherelationshipisnotlargeandisonlymarginallystatisticallysignificant.Cyclicalityof
earningsisnotablyreducedwhenIallowforeitherearly-careerscarringor(especially)forexcess
8Thedownturnincohorteffectsappearstostartbeforethe2006cohort,perhapsanindicationthatconditionsthroughouttheearly20s,notjustthoseatage22,canhavescarringeffects.Mycalculationoftherecessioneffectasthechangeintrendin2006isthusconservative,particularlyinspecification(3)wherethedeclineincohorteffectsisfairlysmooth.
17
sensitivity.EarningsofthepostGreatRecessioncohorts,whenemployed,arenotablybetterthan
wouldbepredictedbasedonpastcyclicalpatterns.
E. Non-collegeworkers
Appendix Table A-1 shows parallel analyses for non-college workers, relating cohort
effectsonemployment,wages,andearningstotheunemploymentratewhentheywere18years
old. There is little indicationof cyclicalityof cohorteffectsonanyoutcome.There ismodest
evidence of medium-term scarring on non-graduates’ wages and employment, but little for
earnings(AppendixFiguresA-9,A-10,andA-11).Cohortsthatenterthemarketduringrecessions
have lower employment andwages for the first several years of their careers thando those
enteringatlow-unemploymenttimes(consistentwithSchwandtandvonWachter,2019),butin
thelongruntheirwagesareifanythinghigher.
VI. MeasuringtheshadowcastbytheGreatRecession
TheresultsinTable1andFigure2indicatethatrecessions,andtheGreatRecessionin
particular, have hangover effects on college graduates’ employment rates. About half of the
cyclicality isaccounted forbymedium-termscarring,butaportion remainsevenafter this is
accountedfor.Thisportion,because it ispersistent, isquantitativelyextremely important.To
illustrate this, I use the estimated year, medium-term scarring, and cohort effects from
specification(3)toforecastthecumulativeshort-,medium-,andlong-termeffectsoftheGreat
Recession on college graduates’ employment. This is conservative – any of the other
specificationswouldassignlargerrolestodeclinesincohorteffects,andthusindicatelargerlong-
termeffects.
18
Isimulateastylizedversionofthelabormarketforcollegegraduatesfrom2000through
2060.Iassumethatallcohortsarethesamesize,andthatallworkersenterthemarketat22and
retireat65.Toforecastthefuture labormarket, Iassumethatyeareffects)$ reverttotheir
2005-2007averagefrom2015through2060,thattheunemploymentrateissteadyat4.0percent
throughoutthatperiod,andthatcohorteffects+% forpost-2014entrantsreverttotheaverage
for the 2005-2007 entrants. To measure the effect of the Great Recession, I compare to a
counterfactualsimulationthatshutsofftheGreatRecessionbyimposingthisstabilityfrom2008
forwardratherthanjustfrom2015.Theassumedyearandcohortpatternsinthetwosimulations
areshowninAppendixFigureA-12.
ThefirstpanelofFigure4showsaverageemploymentratesofcollegegraduates,age22-
65,underthetwoscenarios.Theycoincidethrough2007,declininginthe2001recessionand
then recovering. In theobserved series, employmentplummets in theGreatRecession, then
recoversabouttwo-thirdsoftheloss.Beginningin2015, itbeginsslowlytrendingdownward.
This reflectsmyassumptionthat futurecohortswill resemble theones thatentered in2005-
2007.Thesecohortshaveloweremploymentratesthanearliercohorts(Figure2),soastheolder
cohortsgraduallyretirethelaborforceaveragetrendsdownward.
The second series shows the counterfactual no-recession scenario.We see the same
gradualdeclinefrom2015onwardasinscenario1,butfromahigherlevel.Thetwoseriesmove
approximately in parallel from 2015 through themid-2050s, when the cohorts that entered
followingtheGreatRecessionwill retire,andthenfinallyconverge. Integratingthedifference
betweenthem,thecumulativeeffectoftherecessionis15.5percentagepoint-years,ofwhich
19
only7.9percentagepointsoccursbetween2007and2014.Theremaining7.6percentagepoints
representsthemedium-andlong-termshadowoftherecession.
TherearethreecomponentstotherecessioneffectindicatedinpanelA,reflectingthe
threetime-varyingtermsinequation(3).PanelBshowsthepureyeareffects)$.Thesearethe
short-termdirecteffectsoftherecession,commontoallparticipantsinthelabormarketatthe
same time but evaporating after it ends. By construction, these are stable after 2014. The
integratedgapbetweenthetwoseriesbetween2008and2014is6.4percentagepoint-years.
PanelCshowsthemedium-termscarringeffects,representedbythecoefficientsA4 in
(3).Thesearespecifiedtolastbeyondtheinitialshockforthosecohortsthatwereexposedtoit
at labormarketentry,buttofadeoutwithintenyears.Totranslatethese intoeffectsonthe
overall labormarketatapoint intime, Iaverage 34 5 ∗ 78"% ∗ @44 overallcohorts inthe
labormarketinthatyear.TheseaveragesaregraphedinPanelC.Thegapbetweenthetwoseries
issmallbecausetheaffectedcohortsareasmallshareofthelaborforceatanytime,andbecause
the@4 coefficientsarenot largetobeginwith.Theseriesnearlyconvergeby2020,asthe@4
coefficients are near zero for those six or more years into their careers. Cumulatively, the
medium-term scarring effects reduce employment by 1.8 percentage point-years, over one-
quarteraslargeintotalastheshort-termeffect.
PanelDshows long-runeffectsoperatingthroughthecohortcoefficients+%,averaged
overallcohortsinthelabormarketineachyear.Theydeclinesteadilyinbothscenariosoverthe
entireperiod,duetothecompositioneffectmentionedabove.Butthereisagapbetweenthe
twoscenariosoverthedurationofthecareersofthe2008-2014entrycohorts.Thegapisnot
largeinanyyear(thoughitislargerthanthemedium-termeffectsinPanelC,evenattheirpeak),
20
butitlastsfordecades.Itintegratesto7.3percentage-point-years,almostone-thirdlargerthan
theshort-termeffectinPanelBandquadruplethecumulativesizeofthemedium-termscarring
effectsinPanelC.
DeLong and Summers (2012) define a hysteresis parameter C, representing the
permanentdeclineinfutureoutputperunitcurrentcyclicalreduction.Usingarealdiscountrate
of2.5%,themedium-andlong-termeffectsfromFigure4,1.8and7.3percentagepoint-years,
respectively, are equivalent in present value terms to a constant, permanent reduction in
employmentof 0.14percentagepoints. Comparing to the short-termeffects, 6.4percentage
point-years, thisyieldsC =0.14/6.4=0.022.9This isat the lowerendofDeLongandSummers’
(2012)“plausiblerange,”butisneverthelessquantitativelyimportant:TheyconcludethatwithC
= 0.025, fiscal stimulus is fully self-financing through its impacton future tax revenues if the
multiplier is at least 1.0 and the real government borrowing rate is less than 3.7%, and
substantiallyself-financingfor lowermultipliersand/orhigherborrowingrates.Moreover,my
estimates capture only a single, limited source of hysteresis: They should be added to any
persistenteffectsoftherecessiononanyworkerswhowerealreadyinthelabormarketwhen
therecessionbegan,suchasthosedocumentedbyYagan(forthcoming),andtoanynon-labor-
sidesourcesofhysteresis.Theimplicationisthatavoidedlong-runconsequencesofdownturns
canmakeasubstantialcontributiontowardthepresent-valuecostofcountercyclicalpolicy.
9 With higher discount rates of 5% or 10%, the permanent equivalents are 0.19 and 0.21percentagepointreductionsperyear,andCrisesto0.029or0.033,respectively.
21
VII. Conclusion
TherecentU.S.economy,havingfinallyemergedfromadecade-longslump,exhibitsboth
historicallylowunemploymentandanunimpressiveemployment-populationratio,particularly
foryoungcollegegraduates.Ifthelatterisaconsequenceofthe2007-2009recession,thishas
importantimplicationsfortheeconomiccostofcyclicalvolatility.
Myanalysisofcollegegraduates’employmentratesindicatesthattheyaresensitiveto
theeconomicconditionsthatprevailedatthetimethatthegraduatesenteredthelabormarket.
Cohortsthatgraduateduringrecessionshaveloweremploymentratesevenaftertherecession
ispast.
Pastresearchhasemphasizedpersistenteffectsthatfadeoutoverthefirsttenyearsafter
graduation.However,theexperienceofthecohortsthatgraduatedcollegeduringandafterthe
GreatRecessionhasnotbeenconsistentwiththat:Theiremploymentrateshaverecoveredmuch
moreslowlythanwouldbepredictedbasedontheseso-calledscarringeffectsalone,giventhe
lowoverallunemploymentsince2015.Myanalysishighlightsthatcohorteffectsareasecond
channel,previouslyoverlooked,bywhichearlycareerweaknesscanhavelong-termeffects.This
channel,becauseitissopersistent,isbyfarthemorequantitativelyimportant.
Ifanything,theGreatRecessioncohortsaredoingbetterthanwouldbepredictedbased
onthelong-runrelationshipbetweenentryconditionsandcohorteffects.However,eventhis
better-than-expectedperformanceimpliesanenormoushangoverfromtherecessionthatwill
persistfordecadestocome.Itisconcentratedontheemploymentmargin,withsomeeffectas
wellonlogannualearnings,andamongcollegegraduates.Althoughitwillbemanyyearsbefore
wecanmeasurethiseffectconclusively,theevidencetodatepointstolong-termeffectsofthe
22
recessiononcollegegraduates’employmentthatarenotablylargerthantheshort-termeffect.
Thistypeofhysteresisisanimportantcomponentoftheimpactofcyclicalvariationineconomic
conditions. The imperative tominimize it constitutes a strong argument for counter-cyclical
policiesaimedatminimizingtheinitialshock.
23
References
Beaudry,Paul,andJohnDiNardo(1991)."Theeffectofimplicitcontractsonthemovementofwagesoverthebusinesscycle:Evidencefrommicrodata."JournalofPoliticalEconomy99.4:665-688.
Borgschulte,Mark,andPacoMartorell(2018)."Payingtoavoidrecession:Usingreenlistmenttoestimatethecostofunemployment."AmericanEconomicJournal:AppliedEconomics,10.3:101-127.
Bound,John,MichaelF.Lovenheim,andSarahTurner(2012).“IncreasingtimetobaccalaureatedegreeintheUnitedStates.”EducationFinanceandPolicy,7(4),375-424.
Card,David,andJesseRothstein."Racialsegregationandtheblack–whitetestscoregap."JournalofPublicEconomics91.11-12(2007):2158-2184.
Clark,KimB.,andLawrenceH.Summers(1982).“Labourforceparticipation:timingandpersistence.”TheReviewofEconomicStudies,49(5):825-844.
Daly,MaryC.,BartHobijn,andTheodoreS.Wiles(2012).“Dissectingaggregaterealwagefluctuations:Individualwagegrowthandthecompositioneffect.”FederalReserveBankofSanFranciscoworkingpaper2011-23,May.
Davis,StevenJ.,andTillM.vonWachter(2012)."RecessionsandtheCostofJobLoss.”BrookingsPapersonEconomicActivity2012.
DeLong,J.Bradford,andLawrenceH.Summers(2012)."Fiscalpolicyinadepressedeconomy"BrookingsPapersonEconomicActivity2012:233-297.
Gronau,Reuben(1974)."Wagecomparisons--Aselectivitybias."JournalofPoliticalEconomy82.6:1119-1143.
Jacobson,LouisS.,RobertJ.LaLonde,andDanielG.Sullivan(1993)."Earningslossesofdisplacedworkers."TheAmericanEconomicReview83.4:685-709.
Kahn,LisaB.(2010)"Thelong-termlabormarketconsequencesofgraduatingfromcollegeinabadeconomy."LabourEconomics17.2:303-316.
Oreopoulos,Philip,TillvonWachter,andAndrewHeisz(2012)."Theshort-andlong-termcareereffectsofgraduatinginarecession."AmericanEconomicJournal:AppliedEconomics4.1:1-29.
Oyer,Paul(2006).“Initiallabormarketconditionsandlong-termoutcomesforeconomists.”JournalofEconomicPerspectives20.3:143-160.
Oyer,Paul(2008).“Themakingofaninvestmentbanker:Stockmarketshocks,careerchoices,andlifetimeincome.”TheJournalofFinance.63.6:2601-2628.
24
Schmieder,JohannesF.,TillvonWachter,andJoergHeining(2018).“Thecostsofjobdisplacementoverthebusinesscycleanditssources:EvidencefromGermany.”Workingpaper,March2018.
Schulhofer-Wohl,Sam(2018).“Theage-time-cohortproblemandtheidentificationofstructuralparametersinlife-cyclemodels.”QuantitativeEconomics9.2:643-658.
Schwandt,Hannes,andTillvonWachter(2019)."UnluckyCohorts:Estimatingthelong-termeffectsofenteringthelabormarketinarecessioninlargecross-sectionaldatasets.”JournalofLaborEconomics,37(1),S161-S198.
vonWachter,Till,&StefanBender(2006).“Intherightplaceatthewrongtime:Theroleoffirmsandluckinyoungworkers'careers.”AmericanEconomicReview,96(5),1679-1705.
Yagan,Danny(forthcoming).“EmploymenthysteresisfromtheGreatRecession.”JournalofPoliticalEconomy,forthcoming.
25
Figure1.Employmentrates,byageandeducation,1979-2018
Notes:Priorweekemploymentrates,computedfrommonthlyCurrentPopulationSurveydata,
areseasonallyadjustedandsmoothedusinga3-monthtrianglesmoother.(epopoverview.gph)
26
Figure2.Cohorteffectsonemploymentratesofcollegegraduatesaged22-40
Notes:Prior-weekemploymentratesareestimatedfromthemonthlyCurrentPopulation
Surveyandaggregatedtotheeducation-cohort-age-year-statecell.Specificationscorrespond
toequations(1),(2),and(3),respectively,andincludeage,year,andstatefixedeffects,plusan
inverseMillsratiointhecohort-by-yearcollegeattainmentrate.Cohortsareindexedbythe
yeartheyturned22;cohorteffectsarenormalizedtozeroforthe1991and2000entrycohorts.
Regressionsareestimatedonasampleofcollegegraduatesage22-40fromthe1970-2000
entrycohortsobservedin1978to2018.Cohorteffectsforpost-2000entrantsareestimated
assumingthatage,time,andothercoefficientsareasestimatedfromtheearliercohorts.
Fig_atcfx_cohortB_ed1_empl
27
Figure3.Excesssensitivityandscarringcoefficientsfromaugmentedage-time-cohort
decompositionsofcollegegraduateemploymentrates
Notes:SeenotestoFigure2.PanelAshowscoefficientsandconfidenceintervalsforθj
coefficients,reflectingcurrentunemploymentrate-ageinteractions,fromequation(2);panelB
showsφjcoefficients,reflectingentryunemploymentrate-ageinteractions,fromequation(3).
Connectedseriesandconfidenceintervalcorrespondtoestimatesbasedon1970-2000entry
cohorts,whiledotscorrespondtoestimatesusingallentrycohortsthrough2017.
Fig_semitransitory_URt_ed1_empl,Fig_semitransitory_UR0_ed1_empl
28
Figure4.Decompositionofthelong-runeffectoftheGreatRecessionontheemploymentrate
ofcollegegraduates
Notes:Seetextfordetails.EstimatesarebasedonthespecificationplottedinFigures2
(“scarring”series),3B,andAppendixFigures6and7.fittedvals_empl_mD1a
29
Table1.Timeseriesregressionsofestimatedcohorteffectsonthenationalunemploymentrate
intheyearthecohortenteredthelabormarket
Notes:Eachentrycorrespondstoaseparateregressionofestimatedcohorteffectsforcollege
graduatesontheunemploymentrateintheyearthecohortwas22,withaquadraticcalendar
yearcontrolandNewey-Weststandarderrorsallowingforautocorrelationsatlagsuptothree
years.Cohorteffectsincolumns1-3areestimatedusingequations(1),(2),and(3),respectively.
Incolumn4,thespecificationcombines(2)and(3),withcontrolsforbothURst-ageandURsc-age
interactions.PanelsBandCshowestimatesforcohorteffectsonloghourlywagesamongthose
employed,fromtheCPSOutgoingRotationGroups,andonlogannualearningsamongthose
withpositiveannualemployment,fromtheMarchCPS.Thebottomrowsofeachpanelshow
thedifferencebetweenthefittedoractualvalueforthe2010and2006entrycohorts,lessthe
differencebetweenthe2006and2002cohorts.SeeexcelspreadsheetDropBox/tables_01142019tab1(fromcohfxregs.txt)
30
Appendices
AppendixA.Dataprocessinganddefinitions
Theprimaryanalysisconcernsthecohort-age-year-stateemploymentrateforcollege
graduates.ThisisestimatedfromthemonthlyCurrentPopulationSurvey(CPS).Iconstructa
syntheticpanelfromdatafromeachmonthfromJanuary1979throughNovember2018,
aggregatingtoeachcalendaryear.
Idividethepopulationintocollegegraduatesandnon-graduates.Before1992,graduatesare
thosewith16ormoreyearsofcompletedschooling;afterward,itisthosewithBachelorsor
graduatedegrees.Individualrespondentsareassignedtheage(andimpliedbirthcohort)that
theyareonthesurveydate,sosomeonebornonJuly1,1980andsurveyedin2010willbe
treatedasage29in2010andinthe1981birthcohort,ifsurveyedinthefirsthalfofthe
calendaryear,andas30andinthe1980cohortifsurveyedinthesecondhalf.
Myfullsampleforage-time-cohortdecompositions(equations1-3)consistsofpeopleborn
1948-1995(whoturned22between1970and2017)andobservedatages22-40.Asdiscussed
inthetext,myprimaryestimatesidentifytheage,time,andauxiliarycontrolcoefficientsonly
frompre-Recessioncohorts.Forthesespecifications,Iexcludepost-1978birthcohorts,who
turned22after2000.Iusetheresultingcoefficientstoformpredictionsforthesubsequent
cohorts,andestimatetheircohorteffectsasthedifferencebetweenobservedoutcomesand
thepredictions.
WhenIexaminenon-high-schoolgraduates,thesampleextendsfromage18-40,andbegins
withthe1952birthcohort.Thefullsampleextendstothe1999birthcohort,whilethepre-
Recessionsampleagainexcludesthosebornafter1978.
Inspecification(2),Icontrolforthecontemporaneousunemploymentrate.Thisisthestate
averageunemploymentrateforthecalendaryear.Specification(3)controlsforthe
unemploymentratewhenthecohortwas22yearsold(18fornon-college-graduates).Iusethe
statewheretherespondentcurrentlylivesforthis.Stateunemploymentratesarenotavailable
priorto1976;forthe1948-1954(or,fornon-graduates,1952-1958)birthcohorts,Iusethe
nationalunemploymentrate.
AllanalysesareweightedbythesumofCPSsampleweightsinthecell.
Ialsopresentresultsforlogannualearningsandloghourlywages.Forannualearnings,Iuse
the1979-2018AnnualSocialandEconomicSupplement(ASEC),alsoknownastheMarchCPS.
EachASECsurveyincludesinformationaboutearningsduringthepriorcalendaryear.Imeasure
educationandageasofthesurveydate,butusethepriorcalendaryear’sunemploymentrate.
Earningsaretopcodedatthe98thpercentile,separatelybyyear;convertedtoreal2015dollars
usingtheCPI;andlogged.Hourlywagesarecomputedfromthe1979-2017MergedOutgoing
RotationGroup(ORG)filesfromthemonthlyCPS,usinganalgorithmadaptedfromCenterfor
31
EconomicandPolicyResearch(2018;seealsoSchmitt2003).Forthosepaidhourly,Iusethe
higherofthereportedhourlywage(withoutovertime,tips,orcommissions)andtheratioof
weeklyearnings(whichinprincipleincludeovertime,tips,andcommissions)tousualweekly
hours.Fornon-hourlyworkers,Iusethelatter.Whennecessary,usualhoursareimputedusing
actualhourslastweekorthemeanbygenderandparttimestatus.Hourlywagesareagain
convertedtoreal2015dollars,settomissingifbelow$1orabove$200,andlogged.
AppendixB.Additionalspecifications
Ipresentarangeofadditionalspecificationsandmodelsintheappendix.Idiscusstheminturn.
Educationalattainment
IrelyoneducationalattainmentasoftheCPSsurveydate.Somepeopleinterviewedatage22
whodonothavecollegedegreesatthattimewillobtainthemlater.Thismeansthatthe
compositionofmycollegegraduatesamplechangesacrossagewithincohorts.FigureA-3
showsmeanattainmentbycohortandyear.Igroupentrycohortsintotwo-yeargroups.For
example,the“1980”seriesincludesthosewhowere22in1980and1981.Ithenshowthe
averageattainmentofeachcohortgroupineachyearuntilthegroupis29/30.Dotsindicate
theobservationswhenthecohortis23/24andwhenitis29/30.
Forthe1980-1989entrycohorts,educationalattainmentwaslargelystable.Beginningwiththe
1990entrycohort,age-30attainmentrosesharply,byabout5percentagepoints,withnearly
allofthisgrowthoccurringamongdegreesawardedafterage24.Attainmentgrewslowlyfrom
the1990tothe2010entrycohorts,withgrowthmostlyamongdegreesawardedbefore24;
between2002and2010,therewasevengrowthinpre-22awards.Post-2010entrantsarenot
observedthroughage30.Thereissomeindicationthatpost-24degreeawardshaverisenfor
thesecohorts,butwewillnotknowforsureuntiltheireducationiscomplete.
Maindecompositions
FiguresA-4andA-5showestimatedageandtimeeffects,respectively,fromthemain
decomposition(1)foremploymentofcollegegraduates.Theseareestimatedonthepre-Great-
Recessioncohorts,turning22from1970through2000.Forcomparison,FigureA-4shows
estimatedageeffectsfromaspecificationthatincludescalendaryearbutnotcohortcontrols.
Ageeffectsarenormalizedtozeroatage32.ThecomparisonseriesinFigureA-5isthe
unadjustedaverageemploymentrateof22-40-year-oldcollegegraduatesineachyear,without
adjustmentforcohortdifferences.Here,yeareffectsarenormalizedtozeroin2007.Ineach
figure,thecohorteffectsinthefulldecompositionarenormalizedtohavezeroslopebetween
the1991and2000entrycohorts.
FigureA-6showstwosetsofestimatesofcohorteffectsfromequations(1)and(3).Oneset,in
dashedlines,usesthepre-recessioncohortstoidentifyage,time,andscarringcoefficients,
thenextrapolatesthesecoefficientstosubsequentcohorts.Theotherusesthefullsampleto
estimatetheregressions.Resultsaregenerallysimilar.
32
FiguresA-7andA-8repeatthedecompositionsinFigures2and3ofthemaintext,firstforlog
annualearningsandthenforloghourlywages.Sampleshereexcludethosewithzeroworkin
theprioryear(A-7)orpriorweek(A-8).
FiguresA-9throughA-11presentdecompositionsfornon-collegegraduates.Thesearesimilar
tocollegegraduates,exceptthatcohortsenterthesampleatage18,theage-18
unemploymentrateistherelevantentrycondition,thegj(a)coefficientsrefertoyearssinceage
18,andtheselectioncontrolisaninverseMillsratiointheshareofthecohortthatdidnotgraduatefromcollege,! 1 − $%&' .Estimatesofthecyclicalityofthecohorteffectestimates,
withrespecttotheage-18unemploymentrate,areinTableA-1.
Genderdifferences
Maleandfemalelaborforceparticipationpatternsarequitedifferent,anditisplausiblethat
recessionscouldhavedistincteffects(bothpermanentandsemi-transitory)onthem.To
explorethis,Iuseemploymentratesattheage-time-cohort-state-genderlevel,andaugment
mybasicspecificationswithgenderinteractions.Equation(3)becomes:
()%&'* = , + .* + /& + 0%* + 1'* + 2) + 3! $%&'
+ 45 6 ∗ 8922)' ∗ ;5*<5=> + 8922)' − 8922' ?* + @)%&'*.
(3’)
Here,.*representsagender-specificintercept.Theage,cohort,andsemi-transitoryscarring
coefficientsarealsoallowedtovaryfreelywithgender,buttime,state,andselection
coefficientsareheldconstantacrossgenders.
Thenon-identificationproblemissomewhatdifferentherethaninthespecificationusedinthe
maintext.Asbefore,Inormalizecohorteffectstozeroforthe1991and2000maleentry
cohorts.However,Ineedonlyonenormalizationofthefemalecohorteffects:Isetthe1991
entrycohorteffecttozero,butallowtheeffectstovaryfreelybeforeandafterward.This
enablesmetopermitlong-runtrendsinfemalerelativelaborforceparticipation,necessaryto
capturetheentryofwomenintothelabormarketovertime.
FigureA-12showsthe;5*coefficientsfromthisspecification.Semi-transitoryscarringismuch
moresevereformenthanforwomen,forwhomitisonlymarginallysignificant.FigureA-13
showsthecohorteffects,bothforthisspecificationandforasimpleronethatexcludesthe;5*and?*effects.Weseetheexpectedlong-runriseinfemalelaborforceparticipationhere,
thoughtheGreatRecessiondropoff,whichislargerforwomenthanformen,erasesessentially
allofthisprogress.Forbothmenandwomen,unadjustedcohorteffects(fromthesimpler
specificationwithoutsemi-transitoryscarring)areabout4percentagepointslowerforpost-
recessionentrantsthanforthe1991entrantswhoarethereferencegroup.(Alternative
normalizationswouldchangethemagnitudeofthedrop,butwouldnotalterthesimilarityof
themaleandfemaledrops.)Allowingforsemi-transitoryscarringreducesthemaledecline
substantiallyfortheimmediatepost-recessioncohorts,explainingabout2/3ofthefalloff
relativetothe1991-2000trend.Thishasmuchlessofaneffectontherecessiondropofffor
women(thoughitdoeshaveameaningfulimpactonthefemalelong-runtrend).Overall,there
33
doappeartobegenderdifferences,butthebasicstorytoldbythepooledspecifications
presentedinthemaintextisnotoverturnedhere.
SimulationoftheeffectoftheGreatRecession
FigureA-14presentstheassumedtimeandcohorteffectsusedinthesimulationofthelong-
termeffectsoftheGreatRecession,Figure4.Intheserieslabeled“actual,”Iuseestimates
fromspecification(3)through2014.Afterthatyear,Iassumeyeareffectsequaltheaverageof
the2005-7estimates,andcohorteffectsequaltheaveragecohorteffectestimatesfor2005-7
entrants.Inthe“norecessioncounterfactual”series,Iusetheseaveragesforeveryyearand
cohortfrom2008onward.
Sensitivitytomobility
Inequation(3),Iusetheunemploymentratewhentheindividualwas22,inthestateofcurrent
residence,asameasureoflabormarketconditionswhenthatindividualenteredthemarket.
Interstatemobilitybetweenage22andthedateoftheCPSinterviewwillmakethisanoisy
measureofinitialconditions.Further,anyendogeneityofthismobilitytolocalconditionscould
biasmyresults.
Toaddressthis,IuseaninstrumentalvariablesstrategyadaptedfromSchwandtandvon
Wachter(2019).1Constructionoftheinstrumenthasthreesteps.
First,usingpooleddatafromthe1980,1990,and2000decennialcensusandfromthe2001-
2016AmericanCommunitySurvey(one-yearsamples)publicusemicrodatafiles,Iestimatethe
numberofcollegegraduatesbornineachstateblivinginstatesatagea(22£a£40),Nbsa.
Thisisalong-runaverage,pooledacrossnearlyfortyyearsofdata,soisnotinfluencedby
economicconditionsrelevantforanysinglecohort.
Second,Iconstructtheaverageage-22unemploymentratetowhichcollegegraduatesbornin
eachstatebandeachbirthcohortcwouldhavebeenexposed.Lett=c+22.Thisaveragerateis
then:
8922C' =DEFFGHIJG,LLG
IJG,LLG.
Third,foreachstates,cohortc,andagea,Iaveragetheaveragebirth-stateexposureacrossall
birthstates,weightingbytheshareofage-agraduatesinstatesbornineachstateb:
8922)'% =DEFFJMIJGNJ
IJGNJ.
1SchwandtandvonWachter(2019)useadouble-weightingestimatortoabstractfrombothendogenousmobility
andendogenousattainmentrates.Itreatattainmentratesasexogenous,usingtheselectioncorrectiondiscussed
inthetexttoaddresschangesincohortcompositionwithage.
34
Last,Iinteractthismeasurewithageindicators45 6 (j=1,…,5)andusetheresultsas
instrumentsforthe45 6 ∗ 8922)' interactionsinequation(3).AppendixTableA-2showstheestimatedscarringcoefficientsfromthebaseOLSspecificationandusingIV,bothusingthefull
sampleandlimitingtothepre-recessionentrants.Coefficientsaresimilar,andifanything
smallerintheIVspecifications.Inotherwords,accountingforendogenousmobilityattributes
moreofthepost-Great-Recessiondeclineinyoungpeople’soutcomestopermanentcohort
effectsthaninmymainresults.
ReferencesCenterforEconomicandPolicyResearch.2018.CEPRORGVersion2.3,Apr10,
2018,http://ceprdata.org/wp-content/cps/programs/org/cepr_org_master.do.
Schmitt,John(2003).“CreatingaconsistenthourlywageseriesfromtheCurrentPopulation
Survey’sOutgoingRotationGroup,1979-2002.”CenterforEconomicPolicyResearch
technicalpaper,August.
Schwandt,Hannes,andTillvonWachter(2019)."UnluckyCohorts:Estimatingthelong-term
effectsofenteringthelabormarketinarecessioninlargecross-sectionaldatasets.”
JournalofLaborEconomics,37(1),S161-S198.
35
AppendixFigureA-1.Unemploymentrateandprime-agenon-employmentrate
Notes:Unemploymentrateisthepublished,seasonallyadjustedrate.Non-employmentrateis
calculatedover25-54yearoldsfromtheCPSmicrodatasamples,thenseasonallyadjustedand
smoothedusingathree-periodtrianglesmoother.epopoverview_alt
36
AppendixFigureA-2.Meanlogrealhourlywages,byageandeducation,1979-2018
Notes:Hourlywagesarecensoredatrealvaluesof$1and$200perhour(in2015dollars).
Monthlyvaluesareseasonallyadjustedandthensmoothedwitha5-monthtrianglesmoother.
realwageoverview
37
AppendixFigureA-3.Shareofcohortwithabachelor’sdegree,bycohortandyear
Notes:Eachseriesrepresentsrepeatedcrosssectionalaveragesforasinglebirthcohort,
followedfrom22through30acrosssuccessiveCPSsurveys.Forexample,theleftmostseries
includesthosewhowere22in1980,observedinthe1980-1988CPSsurveys.Trianglesand
circlesindicateobservationswhenrespondentswere24and30yearsold,respectively.
fig_eduageyear_college_alt
38
AppendixFigureA-4.Ageeffectsonemploymentofcollegegraduates
Notes:“Employmentrate”seriesrepresentsestimatedcohort-adjustedageeffectson
employmentfromarestrictedversionofequation(1)withnocalendaryeareffects,estimated
oncollegegraduatesaged22-40fromthe1970-2000entrycohortsobserved1979-2018.“Age
effects”seriesisdrawnfromequation(1),estimatedonthesamesampleandincluding
calendaryearadjustments.Ageeffectsarenormalizedtozeroatage32(indicatedbyadot);
thefullage-time-cohortdecompositionfurthernormalizesthecohorteffectstobethesamefor
the1991and2000entrycohorts.Fig_atcfx_age_ed1_empl
39
AppendixFigureA-5.Timeeffectsonemploymentofcollegegraduates
Notes:“Employmentrate”seriesistheunadjustedaverageemploymentrateof22-40-year-old
collegegraduatesineachyear,relativeto2007.“Yeareffects”seriesisdrawnfromequation
(1),estimatedoncollegegraduatesaged22-40fromthe1970-2000entrycohortsobserved
1979-2018andincludingcalendaryearandageadjustments.Yeareffectsarenormalizedto
zeroin2007;thefullage-time-cohortdecompositionfurthernormalizesthecohorteffectsto
bethesameforthe1991and2000entrycohorts.Fig_atcfx_time_ed1_empl
40
AppendixFigureA-6.Cohorteffectsonemployment:Extrapolatedfrompre-recessioncohorts
andestimatedusingthefullsample
Notes:Figurepresentsestimatedcohorteffectsfromequations(1)and(3).Solidlinesshow
effectsestimatedusingthefullsample.Dashedlinesshowseriesobtainedbyestimatingthe
age-time-cohortdecompositionusingonlycohortsaged22by2000,thenestimatingpost-2000
cohorteffectsasmeanresidualsbycohortaftersubtractingtimeandageeffectsestimated
fromtheearlierdata.Ineachcase,cohorteffectsarenormalizedtozeroforthe1991and2000
entrycohorts.Fig_atcfx_cohortC_ed1_empl
41
AppendixFigureA-7.Cohorteffects,excesssensitivity,andscarringforcollegegraduates’log
annualearningsifemployed
A.Cohorteffectsonlogannualearningsofcollegegraduates
B.Excesssensitivityandscarringeffectsonlogannualearnings
Notes:AnnualearningsarecomputedfromMarchCPSdata,forthosewithpositiveearnings,
thenloggedandaveragedtotheeducation-cohort-age-year-statecell.Specifications
correspondtoequations(1),(2),and(3),respectively,andincludeage,year,andstatefixed
effects,plusaninverseMillsratiointhecohort-by-yearcollegeattainmentrate.Cohortsare
indexedbytheyeartheyturned22;cohorteffectsarenormalizedtozeroforthe1991and
2000entrycohorts.Decompositionsareestimatedoncollegegraduatesaged22-40fromthe
1970-2000entrycohortsobserved1979-2018,thenextrapolatedtothepost-2000entry
cohortsholdingallbutcohorteffectsfixed.Fig_atcfx_cohortB_ed1_log_pearnval_tc_r_all
Fig_semitransitory_URt_ed1_log_pearnval_tc_r,Fig_semitransitory_UR0_ed1_log_pearnval_tc_r
42
AppendixFigureA-8.Cohorteffects,excesssensitivity,andscarringforcollegegraduates’log
hourlywagesifemployed
A.Cohorteffectsonloghourlywagesofcollegegraduates
B.Excesssensitivityandscarringeffectsonloghourlywages
Notes:HourlywagesarecomputedfromtheMergedOutgoingRotationGroups,forthosewith
positiveprior-yearearnings,thenloggedandaveragedtotheeducation-cohort-age-year-state
cell.Specificationscorrespondtoequations(1),(2),and(3),respectively,andincludeage,year,
andstatefixedeffects,plusaninverseMillsratiointhecohort-by-yearcollegeattainmentrate.
Cohortsareindexedbytheyeartheyturned22;cohorteffectsarenormalizedtozeroforthe
1991and2000entrycohorts.Decompositionsareestimatedoncollegegraduatesaged22-40
fromthe1970-2000entrycohortsobserved1979-2018,thenextrapolatedtothepost-2000
entrycohortsholdingallbutcohorteffectsfixed.fig_atcfx_cohortB_ed1_rw_l_all
Fig_semitransitory_URt_ed1_rw_l,Fig_semitransitory_UR0_ed1_rw_l
43
AppendixFigureA-9.Cohorteffects,excesssensitivity,andscarringfornon-graduates’
employment
A.Cohorteffectsonemploymentratesofnon-collegegraduates
B.Excesssensitivityandscarringeffectsonemploymentofnon-collegegraduates
Notes:Prior-weekemploymentratesareestimatedfromthemonthlyCurrentPopulation
Surveyandaggregatedtotheeducation-cohort-age-year-statecell.Specificationscorrespond
toequations(1),(2),and(3),respectively,andincludeage,year,andstatefixedeffects,plusan
inverseMillsratiointhecohort-by-yearcollegenon-attainmentrate.Cohortsareindexedby
theyeartheyturned18;cohorteffectsarenormalizedtozeroforthe1991and2000entry
cohorts.Regressionsareestimatedonasampleofnon-collegegraduatesage18-40fromthe
1970-1996entrycohortsobservedin1978to2018.Cohorteffectsforpost-1996entrantsare
estimatedassumingthatage,time,andothercoefficientsareasestimatedfromtheearlier
cohorts.fig_semitransitory_URt_ed0_empl,fig_semitransitory_UR0_ed0_emplFig_atcfx_cohortB_ed0_empl_all
44
AppendixFigureA-10.Cohorteffects,excesssensitivity,andscarringfornon-graduates’log
annualearningsifemployed
A.Cohorteffectsonlogannualearningsofnon-collegegraduates
B.Excesssensitivityandscarringeffectsonlogannualearningsofnon-collegegraduates
Notes:AnnualearningsarecomputedfromMarchCPSdata,forthosewithpositiveearnings,
thenloggedandaveragedtotheeducation-cohort-age-year-statecell.Specifications
correspondtoequations(1),(2),and(3),respectively,andincludeage,year,andstatefixed
effects,plusaninverseMillsratiointhecohort-by-yearcollegenon-attainmentrate.Cohorts
areindexedbytheyeartheyturned18;cohorteffectsarenormalizedtozeroforthe1991and
2000entrycohorts.Regressionsareestimatedonasampleofnon-collegegraduatesage18-40
fromthe1970-1996entrycohortsobservedin1978to2018.Cohorteffectsforpost-1996
entrantsareestimatedassumingthatage,time,andothercoefficientsareasestimatedfrom
theearliercohorts._atcfx_cohortB_ed0_log_pearnval_tc_r_all
45
AppendixFigureA-11.Cohorteffects,excesssensitivity,andscarringfornon-graduates’log
hourlywagesifemployed
A.Cohorteffectsonloghourlywagesofnon-collegegraduates
B.Excesssensitivityandscarringeffectsonloghourlywagesofnon-collegegraduates
Notes:HourlywagesarecomputedfromtheMergedOutgoingRotationGroups,forthosewith
positiveprior-yearearnings,thenloggedandaveragedtotheeducation-cohort-age-year-state
cell.Specificationscorrespondtoequations(1),(2),and(3),respectively,andincludeage,year,
andstatefixedeffects,plusaninverseMillsratiointhecohort-by-yearcollegenon-attainment
rate.Cohortsareindexedbytheyeartheyturned18;cohorteffectsarenormalizedtozerofor
the1991and2000entrycohorts.Regressionsareestimatedonasampleofnon-college
graduatesage18-40fromthe1970-1996entrycohortsobservedin1978to2018.Cohort
effectsforpost-1996entrantsareestimatedassumingthatage,time,andothercoefficientsare
asestimatedfromtheearliercohorts.fig_semitransitory_URt_ed0_rw_l,fig_semitransitory_UR0_ed0_rw_lfig_atcfx_cohortB_ed0_rw_l_all
46
AppendixFigureA-12.Scarringeffectsonemploymentratesofcollegegraduates,bygender
Notes:Seetextfordescriptionofspecification.
47
AppendixFigureA-13.Cohorteffectsonemploymentratesofcollegegraduates,bygender
Notes:Seetextfordescriptionofspecification.
48
AppendixFigureA-14.Assumedtimeandcohorteffectsforsimulationofthelong-termeffects
oftheGreatRecession
A. Yeareffects B.Cohorteffects
Notes:“Actual”seriesisestimatedyear(panelA)andcohort(panelB)effectsfromthe
“scarring”specificationplottedinFigure2(equation3),through2014,andmeasuredin
percentagepoints.Theseriesthenreverttotheir2005-2007averages.The“norecession
counterfactual”seriesequalstheactualseriesthrough2007,thenequalsthe2005-2007
averagethereafter.
49
AppendixTableA-1.Timeseriesregressionsofestimatedcohorteffectsfornon-graduateson
thenationalunemploymentrateintheyearthecohortwas18
Notes:Eachentrycorrespondstoaseparateregressionofestimatedcohorteffectsfornon-
graduatesontheunemploymentrateintheyearthecohortwas18,withaquadraticcalendar
yearcontrolandNewey-Weststandarderrorsallowingforautocorrelationsatlagsuptothree
years.Cohorteffectsincolumns1-3areestimatedusingequations(1),(2),and(3),respectively.
Incolumn4,thespecificationcombines(2)and(3),withcontrolsforbothURst-ageandURsc-
ageinteractions.Lowerrowsshowestimatesforcohorteffectsonloghourlywagesamong
thoseemployed,fromtheCPSOutgoingRotationGroups,andonlogannualearningsamong
thosewithpositiveannualemployment,fromtheMarchCPS.Thebottomrowsofeachpanel
showthedifferencebetweenthefittedoractualvalueforthe2010and2006entrycohorts,
lessthedifferencebetweenthe2006and2002cohorts.
50
AppendixTableA-2.Instrumentingfortheage-22unemploymentrate
Notes:Columns1and3correspondtothescarringestimatesfromFigureA-6.Columns2and4
repeatthesamespecifications,instrumentingfortheage-22unemploymentrate-age
interactionswithasimulatedunemploymentrate(againinteractedwithage)thatabstracts
fromcohort-specificmobility.Seetextfordetails.