The Rise in the Disability Rolls and the Decline in Unemployment* David H

THERISE INTHEDISABILITY ROLLS ANDTHE DECLINE INUNEMPLOYMENT*

DAVID H. AUTOR AND MARK G. DUGGAN

Between1984 and 2001, the share of nonelderly adults receiving Social Security DisabilityInsurance income (DI) rose by 60 percent to 5.3 million beneciaries.Rapid program growth despiteimproving aggregate health appears to be explainedby reduced screening stringency, declining demand for less skilled workers, andan unforeseen increase in the earnings replacement rate. We esti- matethat the sum of these forces doubled the labor force exit propensity of displacedhigh school dropouts after 1984, lowering measured U. S.unemploy- mentby one-halfa percentagepoint. Steady state calculations augur a further 40 percentincrease in the rate of DIreceipt.

Thefederal Disability Insurance(DI) programis thelargest incomereplacement program in theUnited States directedto- ward nonelderlyadults, with annual cashtransfers exceeding $54 billion in 2001. Oncebene ts areawarded, recipientsreceive incomereplacement and healthinsurance through Medicare until returnto work, medical recovery, death, or retirement at age65, at whichpoint theyobtain equivalentbene ts from theSocial Security Administration’ s Old Ageand Survivors Insuranceprogram. Over the past twodecades, two important changeshave impacted eligibility and generosityfor this program.The rst,occurring in 1984, liberalized thedisability de- terminationprocess, reversing a dramatic reductionin disability rollsunderway. The second occurred more gradually. Becausethe DIbenets formulais progressiveand indexed tothe mean wage in theeconomy, the widening dispersionof earnings during the 1980s and 1990s substantially raised theratio of disability in-

*Thisresearch was supported in part bytheCenter for Retirement Research atBoston College pursuant to a grant fromthe U. S.SocialSecurity Administra- tionfunded as part ofthe Retirement Research Consortium. The opinions and conclusionsare solely those of theauthors and should not be construedas repre- sentingthe opinions or policyof the Social Security Administrationor anyagency ofthe Federal Government or the Center for Retirement Research at Boston College.We are indebted to Daron Acemoglu, Joshua Angrist, Orley Ashenfelter, JohnBound, Esther Du o, DanielHamermesh, James Heckman, Caroline Hoxby, JuanJimeno, Lawrence Katz, JeffreyKling, Alan Krueger, Steven Levitt, Sendhil Mullainathan,Casey Mulligan, Derek Neal,David Stapleton, Paul Van de Water, numerousseminar participants, three astute referees, and the editors of this Journal fornumerous insightful suggestions. We thank Hung-Lin Chen, Radha Iyengar,and Liz March foroutstanding research assistance, and Alan Krueger, BrooksPierce, Kalman Rupp, Charles Scott, andAlan Shafer for generous assis- tancewith data sources. Authors may be contacted at [email protected], and [email protected].

157 158 QUARTERLY JOURNALOF ECONOMICS cometo prior earnings (the “replacementrate ”)forlow-skilled workers,a trendaugmented by therising real value of medical benets provided.Contemporaneously, the share of nonelderly adults receivingDI incomerose by morethan 60 percentfrom its 1984 troughto 3.7 percentof the adult population ages25 to64 (5.3 millionbene ciaries).Relative to prior DI cohorts,new entrants wereyounger and substantially morelikely to suffer fromlow mortalityimpairments, particularly pain and mentaldisorders. Theobjective of this paper is toassess how these program- maticchanges —reducedscreening stringency and arisingre- placementrate —interactedwith decliningdemand forless- skilledworkers to impact laborforce participation ofthe low skilledduring theperiod of 1978 to1998. Thequestion is intrin- sically difcult.DI isforthe most part auniformlyadministered federalprogram, making for limited treatment-control compari- sons.Additionally, whileaggregate health improved during the past twodecades, reduced DI screeningstringency coincided with aperiodof declining labor market opportunities for less-skilled workers,making it likelythat demand for benets (as aformof incomereplacement) rose while bene ts supply was shifting outward.1 Finally,the administrative de nitionof disability —the inability to “engagein asubstantial gainful activity ”—ensures that holding healthconstant, disability bene ts areeasier to obtain whenjob opportunities are scarcer. Given these confounds, distinguishing theimpacts ofincreasing supply fromincreasing demand forbene ts is amajortheme of our paper. To(attempt to)surmount these issues, we developtwo sets of instrumentalvariables toproxy demand and supply conditions. Toidentify exogenousvariation in thesupply ofdisability benets,we exploit the progressivity of theDI bene ts formula.This formuladoes not account for variation in regionalwage levels, and hencepotential replacement rates are substantially higherin lowwage states. This variation yields differential cross-state benetsupply shifts during programcontractions and expansions (both ofwhichoccurred in oursample frame) in high versuslow replacementrate states. To identify exogenousvariation in the demand forbene ts,we follow Bartik [1991] in constructinga measureof state level labor demand shifts at lowand medium frequency.This measure, a weightedsum of national industry

1.See Juhn, Murphy, andTopel [1991], Juhn [1992], and Katz andAutor [1999]on adversedemand shifts faced by low-skilledworkers inrecent decades. DISABILITYAND UNEMPLOYMENT 159 employmentchanges (excluding ownstate employment) projected ontostate industry composition,performs well in predicting the owof DIapplications by state,indicating that adversedemand shocksraise demand fordisability bene ts.Exploiting this combination ofinstruments, we study theimpact ofthe supply and demand forDI bene ts onthe labor force behavior of low-skilled workersduring theperiod of 1978 –1998. Ourcontribution builds onan in uentialliterature exploring theimpacts ofdisability bene ts onthe level of labor force participation in theUnited States and abroad. 2 Distinct fromthis literature,our paper analyzes theimpact ofthesupply ofdisability benets onthe responses of low-skilled workersto adverse labormarket shocks. The rationale for this focusis that nonem- ploymentis ade factoprecondition for disability application, makingthe opportunity cost of seeking DI bene ts higherfor employedworkers than forjob losers and nonparticipants. Many potentialbene ciarieswill thereforeseek bene tsin theevent of jobloss but nototherwise. We call thesepotential bene ciaries “conditionalapplicants. ” As ourconceptual model demonstrates, thesize of this conditionalgroup is likelyto be relatively elastic toprogram and labormarket conditions. By contrast,direct employmentto disability transitionswill berelatively inelastic. Our frameworktherefore implies that reducedDI screeningstringency,rising replacement rates, and slackeningdemand forlow- skilledworkers should haveraised thepropensity of workers facing an adversedemand shockto exit the labor force to seek disability bene ts. Thisprediction structures our empirical analysis. After es- tablishing that cross-stateshifts in thesupply ofbene ts raised thelabor force participation ofless-skilled workers during the 1979 –1984 disability retrenchmentand loweredthem thereafter,

2.Key papersin thisliterature include Parsons [1980], Bound [1989], Have- man,De Jong, and Wolfe [1991], Bound and Waidman [1992], Gruber andKubik [1997],Aarts, Burkhauser, and De Jong [1996], and Gruber [2000],who study the impactof disability bene tson laborsupply in Canada,the Netherlands, and the UnitedStates. Lewin-VHI, Inc. [1995], Rupp andStapleton [1995], and Stapleton etal. [1998] analyze the importance of the economic climate to disability applicationand receipt. Bound and Burkhauser [1999] provide a comprehensiveover- viewof thelabor market impacts of disabilityprograms. The work closestin spirit tothe approach taken here is theexcellent study byBlack, Daniel,and Sanders [2002],who measure the impact of shocks to coal prices on DIand SSI incomein miningintensive counties. The Black, Daniel,and Sanders study doesnot explore labormarket implications. 160 QUARTERLY JOURNALOF ECONOMICS weturn to an examinationof the interaction of the disability programwith adversedemand shocks.We nd that DIapplica- tionrates for given demand shocksrose secularly after the disability reformsof 1984, reachingtwo to three times their prere- formlevels by thelate 1990s. Parallelingthis change,we estimatea 60 percentincrease in thepropensity of displaced high schooldropouts to exit the labor force. Simple calculations sug- gestthat increaseddisability application propensitycan account forthis behavioralchange. As alternativeexplanations forthe observedlabor force exit of less-skilled workers, we explore the importanceof changing mortality rates, rising immigration and incarcerationrates, uctuationsin UIbene ts,and, logically, falling wages.Although many of these factors appear relevantto cross-statepatterns oflow-skilled laborforce participation, none substantially altersthe nding that theincreasing supply ofDI benets induced substantial laborforce exit of low-skilled workersduring 1984 –1998. Togaugethe importance of changes in theDI programfor the aggregatelabor market, we calculate the demand contraction experiencedby high schooldropouts over the post-disability-re- formyears to form a counterfactuallabor force participation gurenet of disability. Alimitationof ourapproach is that itdoes notallow ustodisaggregate the separate labor market impacts of reducedscreening and risingreplacement rates —the latter a partial functionof declining relative wages for low-skilled workers.Summing over these structural components, our “reduced- form” counterfactualsuggests that thechanges in laborforce behaviorof less-skilledworkers induced by thedisability system loweredthe measured U. S.unemploymentrate of nonelderly adults by half apercentagepoint since1984. Notably,the impact ofdisability ontheaggregate labor market is yetto be fully felt. Steady statecalculations augur a further40 percentincrease in DIrecipiencyrates over the next decade, which is likelyto be concentratedamong less-skilled workers.

I. POST-1984 CHANGES TOTHE FEDERAL DISABILITY SYSTEM: SCREENING, BENEFITS, AND BENEFICIARIES Thefederal government provides cash and medicalbene ts tothe disabled throughthe Social Security Disability Insurance DISABILITYAND UNEMPLOYMENT 161

(DI) and Supplemental SecurityIncome (SSI) programs. 3 The healtheligibility criteriafor the two programs are identical, re- quiring amedicallydeterminable impairment that preventsthe applicant fromengaging in any “substantial gainful activity. ” SSI benetsaremeans-tested and do notrequire prior work history. DI benets,which are not means tested, are set according to a recipient’spriorearnings history. To obtain bene ts,applicants providedetailed medical,income, and assetinformation to a federalSocial Security Administration (SSA) of ce,which makes thedisability determination.Individuals currentlyin thelabor forceare not normally eligible for disability bene ts.4

I.A.Clampdown and Liberalization Duringthe late 1970s, concernover swelling disability rolls spurred theSocial Security Administration totighten medical eligibility criteriaand exercisegreater control over the state boards that interpretSSA ’seligibility standards. Thefraction of applicants awarded bene ts (the “award rate”)fell from45 percentin 1976 to32 percentin 1980. Augmentingthis administra- tiveaction, Congress passed legislationin 1980 mandating that SSA conductmore frequent bene ciaryhealth reassessments knownas ContinuingDisability Reviews.In thesubsequent three years,SSA determinedthat morethan 380,000 DIbene ciaries— 40 percentof thosewhose cases were reviewed —no longer metmedical standards and terminatedtheir bene ts [Rupp and Scott1998]. Congressalso required the Social Security Adminis- trationto further tighten medical criteria, accelerating the declinein award rates.This large-scale curtailment of bene ts, occurringduring thedeepest postwar U.S.recession,was met with intensepublic criticism.Citing violationsof due process, seventeenstates refused to comply with theDI revieweffort during 1983 and 1984. Responding tothe backlash, Congress passed legislationin 1984 that profoundly alteredthe disability determinationsystem,

3.Approximately one-fourth of DIrecipientsalso receive funds from Supple- mentalSecurity Insurance(SSI), which is an entitlement program rather than a laborincome insurance system. Though the two programshave many overlaps, we focuson DI becauseby design, it has far moreinteraction with labor force participants.Bound, Burkhauser, and Nichols [2001] report that approximately 85percent of DI applicantswere employed 36 months prior to application while thecomparable gurefor SSI applicantswas below 30 percent. 4.For example, earnings exceeding $500 per month in 1999 would have automaticallydisquali edaDIapplicant. 162 QUARTERLY JOURNALOF ECONOMICS yielding abroaderde nitionof disability and providing applicants and medicalproviders with greateropportunity to in uence thedecision process. 5 Despiteimproving economic conditions, the numberof DIawards increasedby one-thirdfrom its 1982 trough toa 1986 peak(the highest level reached during the1980s). Contemporaneously,Continuing Disability Reviewscame to a nearhalt. In the veyears from 1985 through1989, SSA termi- nated fewerindividuals forfailing tomeet medical eligibility standards than it had terminatedin the rst vemonthsof 1982.

I.B.Rising ReplacementRates Whilewe have stressed the role of the liberalization of screeningin expanding thesupply ofDI bene ts,potentially as signicant was an unforeseenrise in theearnings replacement rate.This rise was caused by theinteraction between the DI benetsindexationschedule and thegrowth of earningsinequality in theUnited States economy. 6 Determinationof an individual ’s DI benetproceedsin two steps.First, the bene ciary’sAverageIndexed MonthlyEarnings (AIME) is computedas

T 1 YT22 5 z (1) AIMEi O Yit max , 1 , T F Yt G t51 where Yit is equalto an individual ’saveragemonthly earnings (conditional onemployment) in eachyear t, inated tocurrent

5.SSA wasrequired to 1) relax its strict screeningof mental illness by placingless weight on diagnostic and medical factors and relatively more on functionalfactors, such as ability to function in a work orworklike setting; 2) considersource evidence provided by the applicant ’sownhealth care provider priorto the results of SSA consultativeexamination; 3) give additional weight to painand related factors; 4) consider multiple nonsevere impairments as consti- tutinga disabilityduring the initial determination (whereas prior to 1984, applicantswere automatically denied awards during theinitial determination if all impairmentswere judged to be nonsevere); 5) desist from terminating bene ts for anyindividual for whom SSA couldnot demonstrate substantial evidence of medicalimprovement; 6) provide bene tsfor those former recipients whose ter- minationswere under appeal; and 7) suspend Continuing Disability Reviews (CDRs) formental impairments and pain until appropriate guidelines could be developed.In the post-1984 period, two additional administrative factors affected applicationsand terminations. In 1991, due to successful court challengesto SSA’streatmentof source evidence, regulations were adopted placing further weighton the information provided by an SSI or DI applicant ’sownmedical provider.Finally, agency downsizing during the 1980s and increased claims workloadin the 1990s resulted in asubstantialdecrease in thefrequency of CDRs during 1989–1993.See Stapleton et al. [1998, pp. 66 – 69]for a detaileddiscussion. 6.To our knowledge, this increase has been overlooked by theeconomics and DIpolicyliteratures. DISABILITYAND UNEMPLOYMENT 163 dollars by theratio of the average wage in theUnited States economytwo year ’s prior (YT22)tothe average wage in theyear ofearnings. 7 Second,the bene tawarded, the “PrimaryInsurance Amount” (PIA), is computedfrom the AIME using thepiecewise linearformula, (2) 0.9 3 AIME if AIME [ @0,b1# PIA 5 0.9 3 b1 1 0.32 3 ~AIME 2 b1! if AIME [ ~b1,b2# H0.9 3 b1 1 0.32 3 ~b2 2 b1! 1 0.15 3 ~ AIME 2 b2! if AIME . b2, where the “bend points” (b1, b2) arealso rescaled each year by averagewage growth in theeconomy. As visiblein (2), theDI benetsformulais progressive(i.e., concave), ensuring that low- wageworkers replace a greatershare of income. 8 Indexation of the benetformulato the mean wage in theeconomy further ensuresthat bene tlevelskeep pace with aggregateearnings growth. In aneraof stable earningsinequality, this formulahas little impact onthe evolution of replacementrates. Between 1979 and 1995, however,real weekly earnings of full-time,full-year workerswith lessthan ahighschool degree fell by 19.1 percentage points,while the Social Security Administration ’s mean wage seriesincreased by 21.6 percentagepoints in realwage terms. 9 Thisincrease in theproportional difference between mean wages and below-meanwages caused the DI replacementrate for low- wageworkers to rise substantially. Toillustratethe mechanism underlying this rise,it is useful toconsider a workerwhose earnings growth over her career lagged contemporaneousaverage nominal wage growth (as would betrue for many low-skilled workersin this era).Because the

7.In addition, approximately the lowest veyears of earningsare discarded andquarters with earnings in excess of that period ’staxableSocial Security maximumare truncated at the cap. Once theDI bene tisawarded, annual cost-of-livingincreases are tied to the Consumer Price Index. The two-year lag for theearnings indexation factor inequation (1) re ectsthe historic time lag re- quiredfor calculating the numerator of thisseries. 8.For example, a worker withan AIME that did not exceed b1wouldreceive 90percent earnings replacement. Because a 7.65percent Social Security payroll tax isassessedon wagebut not on DIincome,the effective replacement rate would becloser to 100 percent. 9.Real wages are calculated from CPS March AnnualDemographic les as annualearnings of full-time male high school dropouts ages 25 –64for earnings years1979 and 1998. Wages series are de atedby the chain-weighted PCE deator. 164 QUARTERLY JOURNALOF ECONOMICS wageindex applied in (1) and (2) in atesprior earnings to current dollars using thegrowth rate of average nominal wages, the indexed monthlyearnings (AIME) computedin step 1ofthe benetscalculationfor this lowearnings-growth individual would exceedcurrent earnings. Additionally, becausethe bend points in (2) followthe same wage index, these would alsohave risen relativeto individual earnings.A greatershare of the worker ’s AIME would thereforebe replaced on the steeper sections of the curve(i.e., in the90 and 32 percentranges rather than the15 percentrange). These two impacts —arisingAIME and rising bend points—areadditive: indexationof earnings to the mean wageraises the ratio of AIME topriorearnings while indexation ofthe bend points raisesthe share of the AIME replacedon the moregenerous sections of the bene ts formula. Thedistributional impacts ofindexationare seen in TableI. Currentlyemployed male workers ages 55 –61 at thetenth percentileof the(age-speci c)earningsdistribution whoobtained DI benetsin 1979 would havereplaced 52 percentof their current earningswith DIcash transfers.By 1999, this numberhad risen to74 percent.Accounting for the rising value of in-kind Medicare benets,the potential DI replacementrate for a tenthpercentile male age 55–61 rosestill furtherfrom 67 to104 percent.Nor was this riselimited to the older workers. For males at thetenth percentileof earnings,the rise in thereplacement rate exceeded 18 percentagepoints in all agebrackets. At higherpositions in theearnings distribution, however, the growth was far lesspronounced.For workers at theseventy- fth and ninetiethpercen- tilesof earnings, potential replacement rates rose a compara- tivelymodest 4 to8 percentagepoints. 10

I.C.Changing Characteristicsof DI Beneciaries As documentedin TablesII and III, severalmarked demographic shifts in theDI population accompaniedthese bene t

10.Further informationon all data sources and variable construction is providedin theData Appendix. Note that there was also a substantial —albeit far lesspronounced —risein theDI replacementrate for relatively high wage workers dueto a largeincrease in earningssubject to the OASDI tax.Income subject to the OASDI tax wascapped at approximately 1.2 times average earnings during the 1950s,1960s, and early 1970s and then rose to approximately 2.4 times average by1989, where it currently remains.Because bene tsare only computed from taxedrather than total earnings, the increase in thecap causedreplacement rates torise for even relatively high wage workers. DISABILITYAND UNEMPLOYMENT 165

TABLE I POTENTIAL DI INCOME AS A PERCENTAGEOF CURRENT EARNINGS FOR NONELDERLY MALES AT VARIOUS PERCENTILES OFTHE WAGE DISTRIBUTION , 1979 AND 1999

Cash income Adding in-kind replacementrate Medicarebene t Earnings Age percentile 1979 1999 1979 1999

55–61 10 52 74 67 104 25 45 54 48 63 50 37 47 36 47 75 27 32 26 31 90 20 24 19 23 50–54 10 47 57 61 81 25 41 47 44 55 50 34 41 33 42 75 26 32 25 31 90 19 23 18 22 40–49 10 48 53 61 80 25 41 45 44 55 50 34 41 33 42 75 26 33 25 32 90 20 26 19 25 30–39 10 46 54 59 84 25 41 46 44 58 50 36 41 35 44 75 29 36 27 35 90 23 28 21 27

Replacementrates arecalculated using the Social SecurityAdministration Disability Insurancebene t formulafor 1979 and1999 inconjunction with annual earnings data fromthe March CPS les forthe years 1963–1998 andthe 1978 CPS Social SecurityEarnings Records Exact Match lefor the years 1951 –1962. The rst twocolumns represent the ratio of potentialdisability bene ts tocurrent earnings for males in the labor forceand with nonzero earnings in 1978 and1998. Thelatter twocolumns add average Medicare expendi- turesto DI bene ts andaverage percentile-speci cfringebene ts toearnings to estimate a total compen- sation measureof the replacement rate. To calculate Average Indexed Monthly Earnings (AIME) for an M year old Nth percentileworker in year T,weuse equation (1) of thetext and assume that this individualwas an M 2 t year old Nth percentileworker in year T 2 t andthat lowearnings years (those excluded in the AIME calculation)occurred before the age of 25. Percentileranks inthis calculationare year and age speci c. Weuse equation (2) of the text to estimate potential DI bene ts as afunctionof the AIME. The nal two columnsadd average Medicare expenditures for DI bene ciaries inthe relevant year to the DI bene t amount andscale earningsto account for average fringe bene trates forindividuals at eachof the ve earnings percentilesusing data fromPierce [2001]. See the Data Appendixfor further details.

supply shifts. Onewas arapid increasein theshare of younger recipients.Among males ages 25 –39 and 40 –54, DIreceiptrose by 50 percentbetween 1984 and 1999. Theproportional increase formales above age 54 was onlyone-quarter as large.Among women,growth in DIreceiptwas evenmore rapid but also skewedtoward youngerrecipients. On net,the share of DI re- 166 QUARTERLY JOURNALOF ECONOMICS

TABLE II DI RECEIPT AND LABOR FORCE PARTICIPATIONBY GENDER, EDUCATION, AND AGE 1979, 1984, AND 1999

A. Males B. Females

All HSdropoutHS plus All HSdropoutHS plus

Age 79 84 99 79 84 99 79 84 99 79 84 99 79 84 99 79 84 99

A.DIRecipientsper 1000 nonelderlyadults (SSA andSurvey of Income and Program Participation data)

25–39 11 10 15 21 53 5 11 4 4 10 7 21 4 8 40–54 35 28 42 52 105 18 26 15 12 30 35 60 10 21 55–64 113 96 108 148 201 46 59 51 43 72 92 164 29 62

B.Percentof nonelderly,nonparticipants receiving DI bene ts (Surveyof Income and Program Participation data)

25–39 17.2 17.0 23.5 26.8 14.1 13.7 1.3 3.2 1.5 3.9 1.1 3.0 40–54 36.9 32.7 38.5 40.0 35.6 29.6 4.9 10.0 7.4 11.7 3.3 9.5 55–64 30.2 26.6 42.5 43.2 20.3 20.3 8.7 16.8 14.2 24.4 5.1 13.8

C.Percentof nonelderly adults participatingin labor force (CurrentPopulation Survey data)

25–39 95.7 94.7 93.1 91.0 88.1 86.1 96.6 95.8 94.1 63.9 70.0 76.3 49.6 50.3 55.0 66.9 73.2 78.7 40–54 92.7 92.7 90.2 86.5 85.0 76.3 95.4 95.0 91.9 60.3 65.7 77.4 48.8 49.8 54.0 64.9 70.6 80.1 55–64 73.0 68.6 68.1 64.2 60.2 53.2 79.0 73.3 71.2 41.9 41.8 51.6 33.8 33.3 32.4 47.0 46.0 55.7

D.Percentof nonelderly adults unemployed(Current Population Survey data)

25–39 3.7 6.3 3.1 7.0 12.5 6.0 3.0 5.4 2.7 3.9 5.0 3.2 6.2 8.3 6.7 3.4 4.5 2.8 40–54 2.5 4.3 2.4 4.0 7.2 4.5 1.9 3.4 2.2 2.4 3.4 2.1 3.1 4.8 3.8 2.2 2.9 1.9 55–64 1.9 3.4 1.9 2.3 4.7 2.1 1.7 2.6 1.8 1.3 1.8 1.4 1.5 2.4 1.7 1.1 1.4 1.3

DIreceiptrates byage and gender in Panel A arecalculated using recipient counts from Social Security Administration, AnnualStatistical Supplement [variousyears] denominated by population estimates from theCurrent Population Survey Merged Monthly les. Education-speci cDIreceiptin Panel A andthe fractionof labor-force nonparticipants receiving DI inPanel B arecalculated with the Survey of Incomeand ProgramParticipation, 1984 wave1 and1996 wave12. Panels CandD werecalculated from Current PopulationSurvey Merged Monthly les. cipientsbetween the ages of 40 and 54 roseby morethan 50 percentfollowing the 1984 liberalization whilethe share of male beneciariesdeclined by 15 percent[U. S.SocialSecurity Administration, Social SecurityBulletin: Annual Statistical Supplement, variousyears]. TheDI population has always beensubstantially lessedu- cated than average,and as theprogram grew post-1984, it en- compasseda substantially largershare of this population.Using data fromthe Survey of Income and ProgramParticipation (SIPP), weestimatethat theshare of high school dropouts receiv- DISABILITYAND UNEMPLOYMENT 167

TABLE III DISTRIBUTIONOF QUALIFYING IMPAIRMENTS OF DI AWARDEES AT FIVE-YEAR INTERVALS, 1983–1999

Percentof DI awards 4-Yearmortality Qualifyingimpairment rate (%) 19831988 1993 1999

Neoplasms 81.0 16.813.2 12.6 10.6 Circulatorydisorders 19.8 21.917.6 14.0 12.1 Musculo-skeletaldisorders 5.3 13.416.8 14.8 23.7 Mentaldisorders 5.4 16.320.9 26.1 22.5 All others 16.0 31.631.5 32.5 31.1

Source: Social SecurityAdministration, AnnualStatistical Supplement, 1984, 1989, 1994, and2000. Four-yearmortality rate is fromadministrative follow-up of thoseawarded bene ts in1985 [Hennesseyand Dykacz1993].

ing disability bene ts roseby over60 percentafter 1984, and morethan doubled forhigh schooldropout males ages 40 –54.11 In 1999, a 40 –54 yearold high schooldropout was fourto ve times as likelyto receive DI bene ts as amalein thesame age range with at leasta high schooldegree. Therapid growthin DIreceiptamong male high school dropoutsmirrored a substantial declinein theirlabor force participation (panel C). Between1984 and 1999 thelabor force participation rateof high schooldropout males declined by 8.7 percentagepoints amongthose 40 –54, and 7percentagepoints amongthose 55 – 64. Simultaneously,the share of male high schooldropouts in theseage brackets receiving DI roseby 5.3 and 6.3 percentagepoints. Hence, despite thesteep decline in male high schooldropout participation, theshare of male high school dropout nonparticipants receivingDI bene ts in theseage cate- goriesrose to slightly above40 percent. 12 Finally,as youngerbene ciariesentered the DI rolls,the fractionsuffering fromcomparatively low mortality impairments

11.Since SSA doesnot report the educational distribution of DIrecipients, weuse the Survey ofIncome and Program Participation (SIPP) to estimate these numbersfor 1984 and 1999. SIPP data are unfortunately not available for earlier years.Relative to high school dropouts, proportionate growth inDI receiptfor thosewith at leasta highschool degree was substantially smaller among males andslightly larger among females. In all cases, growth amongbetter educated workers wasfrom a lowbase. 12.Unemployment per population rates of high school dropouts also fell steeplyduring these years (panel D). Wediscussthese trends further inSectionV. 168 QUARTERLY JOURNALOF ECONOMICS

FIGURE I DITerminationRates per 1000 Bene ciariesby Reason,1978 –2000 Source: SocialSecurity Bulletin: AnnualStatistical Supplement [variousyears]. Terminationrates are equal to the fraction of DIbene ciariesterminated by cause annually.

grew(Table III). Theshare of DIawardeeswith aprimarydiag- nosisof a mentaldisorder or a diseaseof the musculo-skeletal system(typically backpain) —thetwo disorders with thelowest mortalityamong SSA ’sfourteenmajor diagnostic categories — increasedby 60 percentbetween 1983 and 1999. Thecorrespond- ing award sharesfor neoplasms (cancers) and circulatorysystem diseases(primarily heartdisease), both of which have mortality far in excessof average, declined by 40 percent. Severalconsequences of thesedemographic shifts areseen in FigureI, whichplots theannual rateof DIbene tterminationby causefrom 1978 –2000: death,retirement, and medicaldisquali- cation.Following the 1984 liberalization,DI rollsincreased at an averageannual rateof 4.2 percent( excluding dependentsof DI recipients).As youngercohorts with lowermortality impairments entered,the annual mortalityrate of DI recipientsfell by 35 percent,and theexit rate into the retirement system declined by 40 percent.Accordingly, the expected bene tduration ofnewer DISABILITYAND UNEMPLOYMENT 169 cohortssubstantially exceeds1984 levels[Rupp and Scott 1998].13 Becausemost objective evidence suggests that theprevalence ofdisabling illnessfell during this time[Cutler and Richardson 1997], it is likelythat thegrowth in DIrollsis primarily explained by nonhealthfactors that shifted thesupply and demand fordisability bene ts.These include a moreexpansive de nition ofdisability, changesin theDI award process,rising replacement ratesand, closelyrelated, declining labor market opportunities forless skilled workers. 14

II. WHEN DO SHIFTS IN BENEFITS SUPPLY IMPACT LABOR SUPPLY: DEMAND SHOCKS OR DIRECT QUITS? Beforeanalyzing theimpact ofchanging DI bene ts supply and demand onthe labor market, we offer a briefmodel to motivateour empirical approach. In amarketsetting where all jobseparations were involuntary, workers would apply forDI benetsonlyat theonset of illnessor jobloss. An analysis ofthe impact ofshifts in thesupply ofbene ts onlabor force participa- tionwould thereforefocus on the labor supply decisionsof displaced workers,speci cally whetherthey chose to seek new em- ploymentor apply forDI bene ts instead.Since in realityworkers canendogenously quit jobsto obtain DIbene ts,shifts in bene ts supply mightinstead largelyimpact laborsupply by inducing endogenousquits evenabsent adverse shocks. Below, we write a simpledynamic programmingmodel to explore how elastic these

13.An exceptionto this expansionary trend was Congress ’ 1996discontinua- tion of benetsfor individuals who quali edfor disability on the basis of alcohol anddrug addiction,resulting in the termination of approximately 130,000 beneciaries,visible in FigureI. Itis estimated that approximately two-thirds of those terminatedeventually requali ed for benetsunder a differentimpairment (cf. LewinGroup [1998]). 14.This set of factsdoes not imply that more recent disability bene ciaries are “shirking.” AsBound and Waidmann stress [1992, 2000], disability is a continuousrather than a dichotomousmedical state. A moreexpansive de nition ofdisability will accommodate a greaterrange of illness. Bound and Waidmann [2000]provide evidence that the incidence of self-reported disability among males respondsmarkedly to changes in the generosity of the disability program. Simi- larly,Baker, Stabile, and Deri [2001] ndusing matched medical records and healthself-reports from Canada that individuals who are out of thelabor force are morelikely to report major medical ailments that are not re ectedin objective healthrecords. Notably, despite large increases in disabilityreceipt, Burkhauser, Daly,Houtenville, and Nargis [2001] ndthatprevalence of self-reporteddisabil- ity byincomedecile has changed little over the past two decades. 170 QUARTERLY JOURNALOF ECONOMICS responses—exitafter shocks or direct quits —arelikely to be to shifts in thesupply ofbene ts. Wecomparesteady statesof adiscretetime Markov model in whichindividuals maybe in oneof three states in eachperiod: employed;unemployed and seekingwork; and seekingor receiving DI benets (notparticipating). Employed workersreceive per-periodutility ofemployment v(w,h),wherethe utility ofwork 15 is increasingin wagesand health: vw[, vh[ . 0. Employed workersface a per-periodhazard s ofjobloss and unemployedjob seekers face a per-period hazard q ofreemploy- ment.As an alternativeto employment, individuals mayseek DI benets.Applicants with health h qualify forbene ts after one periodwith probability p 5 p(h),and rejectedapplicants may reapply. Consistentwith DIprogramrules, labor force participants arerejected with probability one. 16 Accepted applicants receiveper-period incomeof d in perpetuity. Weassume a discountfactor of b , 1and setper-period utility ofunemploymentto zero. 17 To(further) simplify theanaly- sis,we restrict attention to the case in whichneither w nor h is timevarying, and weconsider a setof individuals whoface a commonprobability ofDI award conditionalon application: p(hi) 5 p for all i. Usingthese parameters, the asset value of employment re- duces(after somealgebra) tothe following expression:

2 bpVD ~1 2 b~1 2 p!! z vi 1 b psVD (3) V 5 max , , E F1 2 b~1 2 p! ~1 2 b~1 2 s!!~1 2 b~1 2 p!!

~1 2 b~1 2 q!! z vi . ~1 2 b!~1 2 b~1 2 q 2 s!!G Each ofthethree terms inside ofthebrackets in (3) corresponds toone of three decision rules ( “policies”)that aworkermay

15.In modeling the disability application decision as a functionof both healthand the disutility of work, wefollow the approach of Diamondand Sheshin- ski[1995]. See also Hausman and Halpern [1986], Burkhauser, Butler, and Weathers[1999], Kreider [1999], and Benitez-Silva et al. [2000] for theoretical andempirical models of theDI applicationdecision. 16.Although it is possible that some DI applicantscollect unemployment insurancebene tsduring theapplication process, this is technically illegal and likelyrare since UI bene ciariesmust demonstrate that they are active job seekers.UI recipientsmay of courseseek DI bene tswhen UI isexhausted. 17.Setting per-period unemployment utility to zero is an assumptionrather thana normalization.Provided that the owutilityof unemployment is lessthan the owutility of receiving DI bene ts,explicitly parameterizing the level of unemploymentbene tsadds complexity but does not change the core results. DISABILITYAND UNEMPLOYMENT 171 pursue:apply forDI bene ts immediately( I),apply conditional on job loss (C),orremain in thelabor market after job loss ( R). The rstbracketed term in (3) is thevalue of quitting em- ploymentdirectly to seek disability bene ts.Note that if aworker selectsthis policy,she will nd it optimalto quit employment immediately,to reapply forbene ts if rejected,and toremain a beneciaryin perpetuityonce accepted. The asset value of policy

VI is thereforethe present discounted value of permanent benets receipt VD 5 u(d)/(1 2 b)discountedby theexpected time fromquit toaward, where u(d)is theper period utility ofreceiv- ing DI benets. Thesecond term is thevalue of remaining employed until exogenousjob loss and thenseeking disability bene ts thereafter. Werefer to this policyas “conditionalapplication ” and its value

VC incorporatesthe asset value of applying forand ultimately receivingDI bene ts and theexpected owvalue of employment priorto DI application. Thethird policyin bracketsis thevalue of remainingin the labormarket in perpetuity, VR.Becauseworkers pursuing this policyseek reemployment (not DI bene ts) afterjob loss, the reemploymenthazard q appears in this expressionwhile the benets award hazard p does not. Thethree components of (3) arelinked by themax[ z ] opera- torbecause the value of employment for a givenindividual correspondsto thepolicy { I,C,R}that yields thehighest expected utility.For market participants arrayed alongthe distribution of v,thevalue of employment is theupper envelopeof these three policies.Figure II depicts asimulationof the asset value of each policyas afunctionof v.18 Asis visible in the gure,these three slopesgive rise to two thresholds labeled v˜ IC and v˜ CR, in which theoptimal policy shifts from( I) to (C) to (R)as afunctionof v. Solvingfor these thresholds yields (4) bp pu~d! 1 2 b~1 2 q 2 s! v˜ 5 u~d!, v˜ 5 . IC S1 2 b~1 2 p!D CR S q DS 1 2 b~1 2 p! D Equation (4) permitsexploration of the question that moti- vatesthe model: how do changes in programparameters d and p

18.Parameter values used for the simulation are b 5 0.9, p 5 0.5, u(d) 5 1, s 5 0.1, and q 5 0.5. Since p isassumed constant in the subpopulation depicted,variation in v canbe viewed as arising from variation in w. 172 QUARTERLY JOURNALOF ECONOMICS

FIGURE II TheChoice of Disability Applicant Status —Immediately,Conditional, Never—asaFunctionof Earningsand Health

(the “supply” of benets) and labormarket parameters q and s (the “demand” for benets) affect theshare of workers pursuing eachpolicy, in particular quitting workimmediately to apply for benets versusapplying forbene ts conditionalon job loss? Consider rstthe impact ofdemand conditionson laborforce exit.An individual whois indifferent betweenimmediate and conditionaldisability application is by de nitionindifferent toa changein thejob loss hazard that hastensor retardsthe moment ofapplication. Hence, v˜ IC is independentof s.Becauseworkers pursuing (I) and (C)do notreenter the labor market after exit, v˜ IC is alsoindependent of the reemployment hazard q. Conse- quently,changes in labormarket conditions do not impact the sizeof the group that quits employmentimmediately to seek benets (I).Theseparameters do, however, unambiguously impact thesize of theconditional applicant group.Because higher s and lower q reducethe value of job search, more adverse labor DISABILITYAND UNEMPLOYMENT 173 marketconditions increase the share of workers applying for benets in theevent of job loss, policy ( C).19 Next,consider an increasein thesupply ofbene ts. Logi- cally,increases in d and p shift both v˜ IC and v˜ CR rightward, raising thetotal share of workers who eventually seek bene ts, bothimmediately and afterjob loss. The magnitude of theimpact at thetwo thresholds differs, however:

2 2 ]v˜ CR ]v˜ IC ]v˜ CR ]v˜ IC ] v˜ CR ] v˜ IC (5) . . 0 and . . 0 and . . 0. ]d ]d ]p ]p ]d]p ]d]p Increasesin d and p—and theinteraction of the two —shift the conditional/remain-in-labor-forcelocus, v˜ CR,fartherrightward than theimmediate/ conditionallocus, v˜ IC.Providedthat theden- sity of v isweaklyincreasing between the previous v˜ IC locus and the new v˜ CR locus(as would bethe case if thedistribution of v wereuniform), the size of theconditional applicant groupwill rise relativeto the size of the immediate applicant group. Finally,it is straightforward toshow using thecross-partial derivativesof v˜ IC and v˜ CR that an increasein thesupply of benetsinteractspositively with adverselabor market conditions toincrease the relative size of the conditional applicant group. Themore generous are program bene ts orthe less stringent is programscreening, the more that adverselabor market condi- tionsincrease the size of theconditional group (and viceversa). 20 On net, we nd that “conditional ” application is likelyto be

19.It bearsemphasis that this result —thatdirect quitsare not in uenced by labormarket conditions —arisesin part fromour assumption that w is xed; workers keeptheir current wageuntil job loss. While this assumption is clearly toostrong, it isqualitatively consistent with well-known evidence that wages of incumbentworkers aresubstantially more sheltered from labor market conditions thanthose of job seekers (cf. Beaudryand DiNardo [1991], Card andHyslop [1997],and Kahn [1997]). We thereforeexpect the direct impactof s and q on w amongemployed workers tobe secondorder. By contrast,it isquite likely that the expectationof w for job losers fallssubstantially at jobdisplacement due to the destructionof speci ccapitalor other incumbency related rents [Jacobson, La- Londe,and Sullivan 1993]. Broadening the model to permit expectations of w to dependon labormarket conditions and the worker ’semployment/unemployment statewould likely increase thedifferential attractiveness of disabilityapplication forjob losers relative to employed. 2 2 20.More formally, the cross-partial derivatives ] v˜ CR/]d]s and ] v˜ CR/]p]s 2 2 arestrictly positivewhile ] v˜ CR/]d]q and ] v˜ CR/]p]q arestrictly negative.All fourcorresponding cross-partial derivatives for the immediate/ conditionalthresh- old v˜ IC arezero. Note that after suf cienttime elapses, all “conditionalapplicants ” exitthe labor force, at which point the program exerts no further effecton their laborforce participation. It is therefore appropriate to think of the model as applyingto a singlecohort of workers, withnew cohorts entering the market in eachperiod. 174 QUARTERLY JOURNALOF ECONOMICS elasticto three forces: bene tssupply, bene ts demand,and the interactionof thetwo. By contrast,direct quits areonly elastic to the rstof thesethree forces and, ingeneral,less elastic than is conditionalapplication. 21 Ourmodel therefore suggests that less stringentDI screeningand higherreplacement rates coupled with declininglabor market prospects for the low skilled are likelyto have increased the propensity of job losers to exit the laborforce to seek disability bene ts. Weexplore two implications of the model below. In Section III weuse the prediction from (5) that shifts in DIscreening stringencyshould inducelarger application responseswhere replacementrates are higher to test whether DI application rates and laborforce participation amonglow skilled were differentially affected by thepre- and post-1984 shifts in DIscreeningin high versuslow replacement rate states. In SectionIV weexplore the model’smainimplication: the responsiveness of DI applicationand laborforce exit to adverse demand shocksshould have risensecularly post DIliberalization.

III. DISABILITY BENEFITS AND LABOR FORCE PARTICIPATION : INSTRUMENTING FOR BENEFITS SUPPLY Webegin the empirical analysis by exploitingthe disability retrenchmentof 1979 –1984 and subsequentliberalization of 1984 –1998 tostudy theimpact ofshifts in thesupply ofdisability benets onlabor force participation. Acomparisonof the two periodsprovides a usefulcontrast: per-capita DIreceiptamong nonelderlyadults contractedat 0.10 percentagepoints annually during 1979 –1984 and thenexpanded at 0.07 percentagepoints peryear during thesubsequent fourteen years. If thesupply of disability bene tsimpacts laborsupply, it should havehad oppo- siteimpacts onlabor force participation during thesetwo time periods. Suppressing subscripts, wewritethe conditional expectation oflabor force participation as

(6) E@LFPuX,w,d,p,h# 5 a 1 fg~d,p! 1 b1w 1 b2h 1 X9b3,

21.A disadvantageof ourfocus on steadystates is thatwe do not model the labormarket impacts of unanticipated changes in parameter values that may induceimmediate labor force exits that would not be visible in equilibrium. Consequently,our model should not be taken to imply that the impact of DI benetson direct quitsis negligible, only that the response of “conditional ” applicationsis likely to be more elastic than direct exits. DISABILITYAND UNEMPLOYMENT 175 where LFP isadichotomousvariable equal toone if an individual is alaborforce participant, w is theopportunity wage, h is health, and X is avectorof individual characteristics.We denote the DI benets “supply” faced by thepotential labor force participant as g(d,p),assumedto be increasing in boththe replacement rate d and theodds ofobtaining bene tsconditionalon application p. The coefcientof interest in (6) is f,theimpact ofpotential benets onlabor force participation. Thereare a numberof dif cultiesin estimatingthis equation with individual leveldata: wecannot typically observeboth d and w fora givenindividual; objectivemeasures of h arenormally not available fromsurvey data; and, as stressedby Bound [1989], becauseindividuals with poorhealth typically experiencedeclin- ing wagesand hencerising potential DI replacementrates, omis- sion of h from(6) will bias estimatesof f.22 Tosurmount some of thesebiases, we estimate a state-level analogof (6) in rstdifferences:

(7) D~LFP/Pop!jt 5 a 1 fDDIjt 1 b1Dwjt 1 b2Dhjt 1 DXj9tb3 1 ejt, where j subscripts the50 U.S.statesexcluding the District of

Columbia and Dt denotes the rstdifference operator over years t and t.As aproxyfor changes in thesupply ofdisability bene ts, weinitially usethe observed contemporaneous state-level changesin DIrecipientsper 1000 nonelderlypopulation ages

25–64 (DDIjt).Wesubsequently instrument this measure. SinceSSA doesnot report the educational attainment of DI beneciaries,we use the total count of nonelderly recipients by state.OLS estimatesof (7) aregiven in the rstfour columns of TableIV. Here,the dependent variable is thestate-level annualized percentagepoint changein thelabor force participation rate ofthe relevant demographic subgroup calculated fromthe merged monthly lesof theCurrent Population Survey (CPS) for1978 to 1998.23 Column1 ofTable IV indicates that stateswith morerapid

22.An additionaldif culty isthe possibility of unmeasured compositional shiftsin thehigh school dropout population, which declined from 26 to12 percent ofoverall population during 1979 –1998.To (imperfectly) account for composition, wecontrol for high school dropout age structure ineach state in all models. 23.Models also control for the age distribution of the relevant demographic subgroups.We initially assume that, conditional on age and education, average wageand health changes are common across states and hence allow them to be absorbed by a.InSection VI wedirectly examine these (and other) alternative explanatoryvariables. A key assumptionhere is thathealth is not itself procycli- cal.Ruhm [2000] provides evidence that health is countercyclical. 176

TABLE IV CHANGE IN DI ROLLS AND LABOR FORCE PARTICIPATION OF NONELDERLY ADULTS: OLS AND INSTRUMENTAL VARIABLES ESTIMATES DEPENDENT VARIABLE: 100 3 ANNUALIZED CHANGE IN LABOR FORCE PARTICIPATION RATE

A. D Male labor force participation B. D Female labor force participation QUARTERLY

OLS estimates IV estimates OLS estimates IV estimates

High school High school High school High school High school High school High school High school dropouts grad plus dropouts grad plus dropouts grad plus dropouts grad plus

78–84 84–98 78–84 84–98 78–84 84–98 78–84 84–98 78–84 84–98 78–84 84–98 78–84 84–98 78–84 84–98 JOURNAL (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8)

D DI Rolls/ 20.61 20.61 20.06 20.07 21.35 20.51 20.20 0.07 20.22 20.38 0.16 0.00 21.01 20.66 0.28 20.14 1000 Pop (0.15) (0.14) (0.05) (0.04) (0.43) (0.32) (0.13) (0.09) (0.18) (0.15) (0.10) (0.08) (0.47) (0.31) (0.25) (0.15) FOF Intercept 21.24 0.25 20.31 20.09 22.03 0.18 20.47 20.20 20.19 0.58 1.13 0.57 21.04 0.59 1.27 0.65

(0.17) (0.11) (0.06) (0.04) (0.47) (0.25) (0.14) (0.07) (0.22) (0.13) (0.11) (0.07) (0.52) (0.14) (0.27) (0.13) ECONOMICS R 2 0.43 0.59 0.19 0.13 0.08 0.27 0.17 0.02 1st-stage 20.77 0.51 20.86 0.62 20.87 0.53 20.75 0.60 Coefcient (0.24) (0.16) (0.24) (0.14) (0.25) (0.14) (0.25) (0.12) 3 1021

n 5 50 U. S. states. Standard errors are in parentheses. Labor force participation rate is the fraction of nonelderly adults in the relevant gender/education category who are in the labor force (employed or unemployed). All estimates control for changes in the age distribution in the state population for the relevant gender/education group (ages 25–39 and 40–54 with 55–64 omitted). Age controls are demeaned in each time period. Estimates are weighted by mean state share of U. S. population in the two years used to form the dependent variable. D(DI Rolls/Pop) is annualized changes in DI recipients per 1000 state population ages 25–64. In OLS specications, this variable is treated as exogenous. In IV specications, it is instrumented by the potential DI earnings replacement rate for currently employed nonelderly adult males ages 25–61 at the seventy-fth percentile of the state’s earning replacement rate distribution in 1978. The potential replacement rate is calculated from the 1978 Current Population Survey/Social Security Earnings Records Exact Match le. Details on this calculation are provided in the Data Appendix. DISABILITYAND UNEMPLOYMENT 177 declinesin disability rollsper population during 1979 to1984 experiencedsubstantial contemporaneous(relative) increases in thelabor force participation ofhighschool dropout males. During this ve-yearinterval, a one-tenthof a percentagepoint declinein disability receiptis associatedwith arisein malehigh school dropoutlabor force participation ofsix-tenths ofa percentage point.24 Column2 displays theanalogous estimate for the labor forceparticipation ofhigh schooldropout males during thepost- liberalization erafrom 1984 –1998. Althoughthe variation in column1 is drivenby DIprogramcontraction while variation in column2 stemsfrom program expansion, the estimated impact of DI benets onlabor force participation is nearidentical in each period. Subsequentcolumns tabulate therelationship between DI receiptand laborforce participation ofmaleswith at leasta high schooleducation. The point estimatesare always smalland typically insigni cant.Although DI receiptincreased among better educated adults after1984 (Table II), this growthdoes not appear tohave been accompanied by laborforce exit. 25 TheseOLS estimateswill bebiased if shifts in thesupply of benetsdueto program retrenchment and reformare correlated with state-levelshifts in thedemand forbene ts (due,for example,to earnings and healthshocks). To isolate the component of variationin DIrollsthat is dueto bene ts supply, weexploitthe progressivityof the DI bene ts formula.This formula does not accountfor variation in regionalwage levels and thus givesrise to substantial cross-statevariation in potentialreplacement rates. As indicated by equation(5) ofour model, shifts in screening stringencyare predicted tohave a greaterimpact onDI rolls wherethe level of the replacement rate is higher.Hence, the interactionbetween national shifts in theDI screeningregime

24.Since high school dropouts comprise a far largershare of the disability recipientpopulation than their share in the workforce, thiscoef cient is of reasonablemagnitude: male high school dropouts averaged 7.9 percent of the laborforce over 1984 –1998(CPS data)and comprised 25 percent of all disability recipientsin 1999(SIPP data). 25.Estimates found in Table 3 ofthe NBER WorkingPaper version of this paper[Autor andDuggan 2001]indicate that state-level changes in SSI receipt areunrelated to the labor force participation of highschool dropouts (or those with highereducation), in contrastto the results above for DI receipt.This pattern is consistentwith the differing labor force histories of thebene ciariesof these two programs(see footnote 3). 178 QUARTERLY JOURNALOF ECONOMICS and statereplacement rates may serve as an instrumentfor the state-levelshifts in thesupply ofDI bene ts.26 Columns 5– 8ofTableIV implementthis instrumentalvariables approach. UsingSocial Security earnings records from the March 1978 CPS Exact Match le,we calculate that themedian potentialDI earningsreplacement rate of currently employed malesin 1978 rangedfrom a lowof 18 percentin Alaska toa high of38 percentin SouthCarolina, with across-statestandard deviationof 2.6 percentagepoints. The rst-stage coef cients tabulated at thebottom of columns 5 –8 conrmthat per-capita changesin DIrollswere substantially morenegative in high replacement-ratestates during 1978 –1984 and, conversely,were substantially morepositive in 1984 –1998.27 Thecorresponding instrumental variables estimatesof the impact ofthesupply ofDIbene ts onlaborforce participation by educationgroup strongly corroborate the earlier OLS estimates. Inward shifts in thesupply ofDI bene ts raised thelabor force participation oflow-skilled malesand femalesduring 1978 –1984; outward supply shifts reducedthem thereafter. IV estimates again showno impact ofDIbene tsonthelabor force participa- tionof more-skilled workers. Sincethe growth in DIrollswas proportionatelylarger amongfemales than malesduring 1984 to1998, panel BofTable IVrepeatsthese estimates for female labor force participation. As in panel A,we nd that thesupply ofDIbene ts reducesthe labor forceparticipation ofhigh schooldropout females and notthose with highereducation both during theretrenchment and reform periods.These results are especially striking in light ofthe fact that femalelabor force participation roseat all educationlevels over 1984 –1998 (Table II).

26.Provided of course that these interactions are not perversely correlated withstate-level health and income shocks in both retrenchmentand reform periods.This possibility seems particularly unlikely because the supply shifts are ofopposite sign foreach state inthe retrenchment and reform eras. We provide further evidenceon this point in Section VI. 27.The instrumental variable employed is theearnings replacement rate at theseventy- fth percentileof themale potential DI replacementrate distribution calculatedfrom the CPS-SSER. Weusethe earnings of malesexclusively because oftheir high rate of labor force attachment, and use the seventy- fth percentile replacementrate —typicallycorresponding to low earnings workers —because of ourfocus on labor supply of low skilled. Using the median replacement instead doesnot qualitatively impact our results. The interpretation of the rst-stage coefcientin column 5 ofpanelA ofthe table is thatfor each additional percentage pointof replacementrate in 1978,state disability rolls fell by 0.077recipients per 1000nonelderly adults per year during 1978 –1984. DISABILITYAND UNEMPLOYMENT 179

IV. DO DISABILITY BENEFITS AFFECT HOW LOW-SKILLED WORKERS RESPOND TO ADVERSE EMPLOYMENT SHOCKS? The nding that shifts in DIbene ts impact laborsupply is consistentwith two(not mutually exclusive) explanations: a once- and-for-all laborsupply response,corresponding to an increasein directquits in ourmodel; or a secularshift in thesensitivity of laborforce exit to adverse demand shocks,corresponding to an increase in “conditionalexits. ” Ourmodel predicts that thelatter mechanismshould havebecome increasingly signi cant following DIreform:reduced screening stringency, rising replacement rates,and decliningdemand forlow-skilled workersshould each haveraised therelative attractiveness of DIapplication and labor forceexit for workers facing adverseshocks. Totest this implication,we require a measureof plausibly exogenouslabor demand shocks.Following the approach devel- opedby Bartik [1991] and employedby Blanchard and Katz [1992] and Bound and Holzer[2000], weexploit cross-state dif- ferencesin industrial compositionand national-levelchanges in employmentto predict individual stateemployment growth. Spe- cically,we calculatethe predicted logemployment change hˆ jt for each state j betweenyears t and t as 5 (8) hˆ jt Ogjkth–j kt, k where h–j kt is thelog change in two-digit industry k’semployment sharenationally and gjkt is theshare of state employment in industry k in state j in theinitial year.The subscript –j in h–j kt indicates that eachstate ’s industry k employmentis excludedin calculating thenational employmentshare change. 28 Thismethodology predicts what eachstate ’schangein em- ploymentwould beif industry levelemployment changes oc- curreduniformly across states and state-levelindustrial composition was xedin theshort term. States that had arelatively largeshare of workers in decliningindustries will havepredicted employmentdeclines, while states that differentially employed

28.In excluding own-state employment, our projected employment changes differfrom the authors cited earlier. We found that including own-state employ- mentsubstantially increased the predictive power of theemployment projections, whichraised a concernabout a potentialmechanical relationship. Also distinct fromBound and Holzer [2000], we compute h –j kt asthe change in log industry employment shares ratherthan levels. This makes the demand index strictly a relativemeasure. 180 QUARTERLY JOURNALOF ECONOMICS workersin growingindustries will havepredicted increases.Pro- vided that national industry growthrates (excluding ownstate industry employment)are uncorrelated with state-levellabor supply shocks,this approach will identify plausibly exogenous variationin stateemployment.

IV. A.EstimationFramework Weconsidera modi edformof equation(6) in whichwe write theconditional expectation of DI application fora displaced worker as (9) E@ApplyuJob Loss,X,w,d,p,h#

5 a 1 sg~d,p! 1 b1w 1 b2h 1 X9b 3, where g[ again representsthe disability supply function.Our modelimplies that s . 0,greater bene ts supply increasesthe propensityof adverselyshocked workers to exit the labor force. As above,we estimate this equationusing state-leveldata. Here,the movefrom individual toaggregate data is lessstraightforward. Ignoringdemographic characteristics in (9) forthe moment, we writethe model for the owof state DI applications in year t through t as

(10) ~ Apps/Pop!jt 5 a 1 ltD~Emp/Pop!jt 1 ejt, where lt measuresthe direct impact ofemployment losses on DI application ows in year t.Similarly,we modellabor force exit as (11)

D~NILF/Pop!jt 5 g 1 utD~Emp/Pop!jt 1 ltD~Emp/Pop!jt 1 vjt. In this equationdeclines in laborforce participation comefrom twosources: workers who withdraw fromthe labor force but do notseek DI bene ts (measuredby ut);and joblosers who leave thelabor force to seek DI bene ts (lt).As theequation under- scores,our data donot permit a directestimate of s in (9). Instead, estimatesof (11) will yield areducedform that combines bothsources of laborforce exit: (12)

D~NILF/Pop!jt 5 a0 1 ptD~Emp/Pop!jt 1 yt, where pt 5 lt 1 ut. Giventhis constraint,we estimate equations (10) and (12) separatelyin boththe pre- and post-DI liberalization erasto DISABILITYAND UNEMPLOYMENT 181 obtain estimatesof Dlt and Dpt,thechange in theresponsiveness ofDI applications and laborforce exit to demand shocksof given magnitude.With these estimates, we may distinguish three cases.First, if DIapplication sensitivityhas risenbut laborforce ˆ exitpropensity has not( Dlt . 0 and Dpˆ t ’ 0), this would suggest that risingDI bene ts supply has increasedprogram take-up but ˆ notimpacted laborforce exit. Conversely, if Dlt ’ 0 and Dpˆ t . 0, this would indicate that laborforce exit has risenfor given shocks but that this changeis unrelatedto the supply ofDI bene ts. ˆ Finally, if Dlt ’ Dpˆ t . 0,application and laborforce exit propensity haverisen conformably, we will viewthis as evidencethat the risingsupply ofDI bene ts has increasedthe propensity of adverselyshocked workers to withdraw fromthe labor force to apply for DI benets.A limitationof this approach is that it doesnot separatelyidentify theimpacts ofrising replacement rates and decliningscreening stringency on laborforce exit propensity. The effectthat weidentify will bea sumof these structural components.

IV. B.Estimation:Adverse Employment Shocks and DI Application Flows Estimatesof equation(10), theimpact ofemploymentshocks onDI application ows,are given in TableV. Thedependent variable in this table is theannualized statelevel ow of unique DIapplications percapita during 1978 –1984 and 1984 –1998. Theindependent variable is theannualized projectedstate log employmentshock measure hˆ jt forthe corresponding time interval.Models also include time dummies and controlfor changes in theage and genderdistribution ofnonelderly adult stateresidents.As above,we initially makethe assumption that wageand healthvariables areuncorrelated with employmentshocks but examinetheir impact directlyin SectionVI. Column(1) ofTableV providesan estimateof equation (10) forthe pre-DI liberalization periodof 1978 –1984. Thepoint estimate of 20.13 indicates that a1-percentprojected log state employmentcontraction yielded 1.3 additional DIapplications per 1000 nonelderlyadults overthese veyears. The subsequent columnpools data forthe entire time period, adding an interactionbetween the shock measure, hˆ jt,and apost-1984 dummy variable.The coef cienton theinteraction term implies that the impact ofadverse shocks on DI application owsgrew threefold fromthe retrenchment to the reform era. Adding statedummies 182 TABLE V IMPACT OF EMPLOYMENT LOSSES ON DI APPLICATIONS FLOWS 1978 –1998: REDUCED-FORM AND INSTRUMENTAL VARIABLES ESTIMATES DEPENDENT VARIABLE: ANNUALIZED FLOW OF DISABILITY APPLICANTS PER NONELDERLY ADULT

A. OLS reduced-form estimates: impact of predicted D(Emp/ B. IV Estimates: impact of high school dropout D(Emp/Pop), Pop) on DI Adds/Pop instrumented by predicted D(Emp/Pop), on DI Apps/Pop

Long changes Stacked 3-yr diffs Long changes Stacked 3-yr diffs QUARTERLY

(1) (2) (3) (4) (5) (6) (1) (2) (3) (4) (5) (6) 78–84 78–98 78–98 78–84 78–98 78–98 78–84 78–98 78–98 78–84 78–98 78–98

D Emp/Pop 20.13 20.13 20.11 20.12 20.11 0.00 20.25 20.30 20.05 20.14 20.12 0.01 (0.06) (0.07) (0.05) (0.04) (0.04) (0.02) (0.20) (0.22) (0.13) (0.07) (0.07) (0.02) JOURNAL D Emp/Pop 3 20.34 20.26 20.17 20.05 20.68 20.34 20.15 20.05 Post-1984 (0.09) (0.06) (0.05) (0.02) (0.67) (0.23) (0.11) (0.03) State dummies No No Yes No No Yes No No Yes No No Yes R 2 0.29 0.32 0.97 0.31 0.35 0.90

1st-stage coef. 0.51 0.52 1.07 0.85 0.90 1.26 OF (main effect) (0.25) (0.29) (0.69) (0.24) (0.26) (0.33) 1-stage coef. 0.54 1.69 1.01 1.62 ECONOMICS (interaction) (0.28) (1.27) (0.30) (0.42) F-statistic 3.97 7.87 1.19 12.02 17.26 11.41 P-value 0.05 0.00 0.31 0.00 0.00 0.00 n 50 100 100 100 350 350 50 100 100 100 350 350

Standard errors in parentheses account for clustering of state level observations in models that contain repeated observations per state and exclude state dummies (specications 2, 4, 5). Dependent variable is the annualized ow of unique DI applications per nonelderly state population (ages 25–64) over relevant time interval: six years in specication 1; six and fourteen years in specications 2 and 3; and 3-year intervals in specications 4–6. Data points for stacked 3-year changes correspond to years 78–81, 81–84, 84–87, 87–90, 90–93, 93–96, and 95–98. In regressions covering the years 78–98 (specications 2, 3, 5, and 6), observations for 78–84 and 84–98 are stacked. The D(Emp/Pop) variable in Panel A (and its interaction with the post-1984 dummy) is the predicted state level change in the employment to population rate, calculated as the national level change in two-digit industry log employment shares (excluding own-state employment) over the relevant interval projected onto initial state industry shares. See equation (8) of the text for details of this calculation. The D(Emp/Pop) variable used in Panel B is the observed change in the employment to population rate of high school dropouts instrumented by the predicted employment change measure in Panel A (and its interaction with a post-1984 dummy). All models include an intercept, time dummies, and controls for annualized state level changes in population shares by age and gender: male/female 3 ages 25–39/40–54/55–65 (six categories, one omitted). Predicted employment change measure, employment to population ratio of high school dropouts, and state age 3 gender distribution variables are calculated from merged CPS monthly les for years 1979–1998. Disability application counts are provided by the Social Security Administration. DISABILITYAND UNEMPLOYMENT 183 tocontrol for xed statedifferences in DIapplication ows only slightly reducesthe magnitude of the estimated growth, which remainshighly signi cant. Topotentially improveupon the precision of theseestimates, wereestimate these models using stackeddata onapplication owsand projecteddemand shocksby stateat three-yearinter- vals.These models found in thesubsequent three columns provide qualitatively similarresults. The estimated growth in DI application sensitivityis notas sizable in thestacked models, but remainsquite large and signi cant.29 Toillustrate the source of this result,Figure III plots theconditional relationship between estimatedlog demand shocks(multiplied by10 3)and DIapplica- tionsper 1000 population forfour ve-yearsubperiods ofour sample.Consistent with thecontemporaneous rise in replacementrates and thedecline in screeningstringency post 1984, the slopeof the demand shock-DIapplication locusincreases secularly ineachadjacent panel.By 1993 –1998 weestimate that this slopehas increasedsevenfold. Wehave performed numerous robustness tests to verify theserelationships. If thedisability application variable is replaced with the owofDIawards orthe change in DIrecipients perpopulation, we nd qualitatively similarpatterns [Autor and Duggan 2001]. Wehave also utilized countylevel DI receiptdata toexplore the sensitivity of disability rollsover 1984 –1998 to within-state variationin localdemand shocksgenerated by variationin county-levelindustrial structure.Our estimates con rm that within states,counties experiencing relatively negative predicted employmentshocks saw substantially fastergrowth in DI rollsover 1984 –1998. PanelB ofTableV presentsa variant ofthemodels above in whichwe estimate the impact ofchanges in theoverall high schooldropout employment to population ratioon DI application owswhile instrumenting high schooldropout employment with theprojected demand shockmeasure. Because these models only allow demand shocksto impact DIapplications throughtheir covariancewith thehigh school dropout employment to popula-

29.The stacked estimates are not strictly preferablesince they may accen- tuatethe importance of measurement error inthe demand instrument at short intervals,particularly in models with state xedeffects. To implement these estimates,we use changes for 1978 –1981and 1981 –1984for the 1978 –1984 period.For the 1984 –1998estimates, we use changes for 1984 –1987, 1987–1990, 1990–1993, 1993–1996,and (due to sample truncation) 1995 –1998. 184 UREL QUARTERLY ORA JOURNAL FOF ECONOMICS

FIGURE III Impact of Projected Log Employment Shocks on Disability Applications per 1000 Nonelderly Adults at Five-Year Intervals, 1979–1998 DISABILITYAND UNEMPLOYMENT 185 tionratio, it is notablethat IVestimatedimpacts ofdemand shockson DI applications areclosely comparable to the stacked OLSmodelsin panel A.Highschool dropout employment levels and DIapplication owsappear torespond to the same componentof the demand shockmeasure. 30

IV. C.Estimation:Adverse Employment Shocks and Labor ForceExit of High SchoolDropouts Sinceessentially all disability applicants arelabor force nonparticipants, the nding ofrising DI application sensitivityfor givenshocks has twopotential explanations. One is that since 1984, employmentshocks have increasingly spurred laborforce nonparticipants toapply fordisability bene ts,perhaps in antici- pation ofdif culty nding workat labormarket reentry. A second is that theshare of job losers exiting the labor force to seek disability bene tshas risen.In termsof equation(12), thesecases correspondto Dpt ’ 0 versus Dpt . 0. Wedistinguish thesepossibilities by estimating(12) using projecteddemand shocksto instrument for net employment lossesat thestate level. Using the identity that D(Emp/Pop)jt 5 2D(Unemp/Pop)jt 2 D(NILF/Pop)jt,wecan decompose the impact ofemployment shocks into the component accruing to unemploymentand thecomponent accruing to labor force exit. To facilitate comparisonto the DI application regressions,we presentresults for both male and femalehigh school dropouts combinedand disaggregated estimatesby gender.These are found in TableVI. During 1978 –1984 aone-percentage-pointshock to the employmentto population ratioof high schooldropouts induced a 0.46 percentagepoint increasein nonparticipation perpopulation and, by implication,a 0.54 percentagepoint increasein un- employmentper population. In thesucceeding fourteen years, shocksto employment of equal magnitude led tosubstantially largerdeclines in laborforce participation. In both1984 –1990 and 1990 –1998, we nd that essentiallyall declinesin high

30.Because the rst stageof the long change models in columns (1) –(4) is weak,we focus discussion on thestacked models. A signi cant limitationof these modelsis that our DIapplicationmeasure is not disaggregated by education level, andhence it is unlikely that the data satisfy the implied exclusion restriction that demandshocks impact DI applicationsonly through high school dropout employmentchanges. 186 QUARTERLY JOURNALOF ECONOMICS

TABLE VI INSTRUMENTAL VARIABLES ESTIMATES OFTHE IMPACT OF EMPLOYMENT LOSSES ON LABOR FORCE EXIT OF HIGH SCHOOL DROPOUTS, 1979–1998 DEPENDENT VARIABLE: CHANGE IN HIGH SCHOOL DROPOUT NONPARTICIPATION RATE

A.Malesand femalescombined

1978–1984 1984–1990 1990–1998 1984–1998 1978–1998

(1) (2) (3) (4) (5) (6) (7) (8)

D Emp/Pop 20.46 20.73 21.47 20.93 21.10 20.90 20.57 20.58 (0.23) (0.16) (0.52) (0.21) (0.21) (0.18) (0.19) (0.19) D Emp/Pop 3 20.34 20.36 (85–98 dummy) (0.16) (0.19) Statedummies NoYes No NoNo Yes No Yes n 100 100 100 150250 250 350 350

B.Estimatesby gender

1978–1984 1984–1998 1978–1998 1978–1998

(1) (2) (3) (4) (5) (6) (7) (8) MalesFemales Males Females Males Females Males Females

D Emp/Pop 20.45 20.44 20.98 21.23 20.54 20.60 20.54 20.49 (0.21) (0.40) (0.19) (0.34) (0.16) (0.29) (0.22) (0.22) D Emp/Pop 3 20.43 20.60 20.49 20.56 (85–98 dummy) (0.25) (0.33) (0.27) (0.36) Statedummies NoNo No No No No Yes Yes n 100100 250 250 350 350 350 350

Standarderrors in parentheses account for clustering of state levelobservations in models excluding state dummies.All data pointsare stacked three-yearstate levelchanges in the relevant measure for years 78–81, 81–84, 84–87, 87–90, 90–93, 93–96, and 95–98. Modelsalso includeyear dummies and control for changesin the age distribution of state highschool dropout population (ages 25 –39 and 40–54, with 55–64 omitted)for the relevant gender group. Estimates areweighted by state shareof U. S.populationin each year.Instrumental variable used for change in employment to population ratio of highschool dropouts in all modelsis projectedstate levellog employment shock for relevant time interval and, in pooled 1978 –1998 specications, its interactionwith a post-1984dummy. See equation (8) in the text for further details on projectedemployment shock measure.

schooldropout employment led toincreases in nonemployment ratherthan unemployment. Tocontrol for persistent state trends in laborforce exit that would bias theresults if correlatedwith thedemand shockmeasure,we add state xedeffects to models in columns(2), (6), and (8). Theiraddition has littleimpact onthe pattern ofestimates. Tofurther explore robustness, columns (7) and (8) pooldata for 1978 –1998, adding an interactionbetween the employment loss measureand apost-1984 dummyvariable. The pooled estimates conrm a signicant increasein thelabor force exit propensity of DISABILITYAND UNEMPLOYMENT 187 displaced highschool dropouts after 1984, onthe order of 35 percent.Inclusion of state xedeffects raises the standard errors slightly,but doesnot alter the ndings. As shownin panel B,the pattern ofresultsis essentiallyidentical forboth genders. It bearsemphasis that theseestimates do not imply that commencingin 1984, all —or even most—highschool dropouts wholost jobs exited the labor force. Since our demand shock capturesnet rather than grossemployment declines over com- paratively longthree-year intervals, displaced high schooldrop- outswho obtain reemploymentwithin themeasurement window arenot detected. In addition, theshock measure does not impact thelabor market exclusively through job loss. Adverse shocks will inducesome subset of workersto choose to exitemployment and, perhaps moreimportantly, slow the prevailing rateof reentryof nonparticipants intothe labor force. Similarly, some share of thoseexiting may be previously unemployed workers whose re- employmentprospects were harmed by entryof newunemployed. Each ofthese forces will reducethe employment to population ratioand yield correspondingincreases in nonparticipation.What is unambiguousfrom Table VI is that unemploymentrates of low-skilled workershave become signi cantly lesssensitive to adversedemand shockswhile their rate of laborforce nonparticipation has becomesubstantially moreso. Could theincreased generosity of the DI programplausibly accountfor this behavioralchange? To makethis calculation,note ˆ that Dlt in TableV measuresthe impact ofemploymentshocks on DIapplications forthe entire nonelderly population, while Dpˆ t in TableVI is measuredonly for high school dropouts. To makethe magnitudescommensurate, we calculate

D (13) Dlˆ t 5 ~1/d!jDlˆ t, where d is thehigh school dropout share of population in theyears ofestimation, equal to 16 percentagepoints in oursample, and j is theshare of the total owof DI applications submittedby high schooldropouts, which we estimate to be 47 percent. 31 Ourestimate ˆ of Dlt fromTable V rangesfrom 0.05 to0.17, i.e., the fraction of all nonelderlyadults applying forDI bene ts in responseto a onelog

31.The rate of DI receiptamong high school dropouts averaged 4.6 times thatof those with at least a highschool degree during 1984 –1999(Table II). Assumingthat this ratio carried over to applications, high school dropouts would haveaccounted for 47 percentof all DI applicationsreceived in thisperiod. 188 QUARTERLY JOURNALOF ECONOMICS point adversedemand shockrose by 0.5to 1.7 applicants per1000 adults post liberalization.In this equation,1/ d rescalesthe esti- matedincrease in DIapplications perpopulation toan increaseper high schooldropout population,while the parameter j adjusts this estimatedownward toaccount for the fact that onlypart ofthe growthin DI applications is dueto high school dropouts. Substituting into(13), weobtain that theimplied increasein high schooldropout labor force exit propensity for a onelog point ˆ D employmentcontraction, Dlt ,rangesfrom 0.15 to0.50. That is, if theincrease in highschool dropout application propensitywere entirelyaccounted for by laborforce leavers, the estimated increasein application sensitivityimplies a 15 to50 percentage point increasein theirconditional probability oflabor force exit. Thisestimate is in closeaccord with the ndings in TableVI. Specically,columns (7) and (8) ofTable VI indicate that the shareof high schooldropouts exiting the labor force in responseto aoneunit shockpost-liberalization increasedby 34 percentage D ˆ D points (Dpˆ t 5 0.34). Notably,the estimates of Dlt [ [0.15,0.50] from(13) fall symmetricallyat one-standard errorabove and belowthis estimate.Based uponthis (necessarily)rough calcula- ˆ tion,we conclude that Dlt ’ Dpˆ t:theincreasing supply ofDI benets post-1984 canplausibly accountfor both greater DI application owsfor given shocks and greaterpropensity of adverselyshocked low-skilled workersto exitthe labor force. Future analyses using micro-data onthe employment and DIapplica- tionbehavior of low-skilled, displaced, and discouragedworkers should beable toimprove on the precision of this estimate.

V. IMPLICATIONSFOR THE AGGREGATE LABOR MARKET Thereduced sensitivity of high schooldropout unemploy- mentto adverse labor demand shocksimplies that theshare of high schooldropouts unemployed at presentis lowerthan it would havebeen if thesupply ofDI bene ts had notshifted outward after1984. Considerthat thetypical nonelderly(that is, age 25–64) high schooldropout in 1984 was employedin an industry that overthe next fourteen years declined by 8.53 percentagepoints as ashareof aggregate employment. Combining this negativerelative demand contractionwith theestimate of D Dpˆ t 5 0.34, wecalculatethat theshare of nonelderlyadult high schooldropouts who were unemployed would havebeen 2.9 percentagepoints higherin 1998 (8.53 3 0.34 5 2.90) but forthe DISABILITYAND UNEMPLOYMENT 189 liberalization ofDI that occurredin 1984. Sincenonelderly high schooldropouts accounted for 12.1 percentof thenonelderly adult population in 1998, this suggeststhat theaggregate unemploymentrate would havebeen approximately half (0.44) ofa percentagepoint higherin 1998 if it werenot for DI liberalization. 32 Giventhat theoverall unemployment rate of nonelderlyadults in this yearwas 3.4 percent,this appears asizable deviationfrom theconventional metric. Althoughwe have focused on the1984 –1998 period,it would beinaccurate to conclude that thelabor market impact oftheDI programhad culminatedby 1998. As is shownin FigureIV, thoughper-capita DIawards peakedin 1992 and declinedafter 1995 (that is,until the2000 recession),the rate of DIreceiptgrew steadily throughoutthe late 1990s. Thesefacts —risingrecipients despite decliningawards —indicate that theDI recipientpopula- tionwas belowsteady statesize. Due to post-1984 declinesin DI beneciarymortality and retirementrates (Figure I), theex- pectedduration ofDI receiptfor new entrants considerably ex- ceededthat ofprevious cohorts, implying an increasein steady stateprogram population. Tomakea simpleprojection of this long-runsize, we use the fact that in steady state,bene tterminationsmust equal bene t awards. Hence, P*l 5 A, where P*is thesteady statepopulation, l is thetermination rate, and A is thenumber of new awards. Drawing ondata fromSSA ’s Annual StatisticalSupple- ment [variousyears], Figure IV depicts estimatesof P* normal- ized by population forthe years 1976 to2001. Followingthe 1984 programliberalization, the steady stateprogram size rapidly doubled fromits 1982 low.At theonset of the 1991 recession, award ratesincreased by another50 percentand theprojected programsize rose accordingly. By 1995 theprojected steady state programsize had reached5.1 percentof nonelderly adults, where it largelyremained for the balance ofthe decade. 33

32.Because the unemployment rate is denominatedby laborforce and not by population,the estimated impact using the overall nonelderly adult (ages 25 – 64) participationrate of 80.3percent is (2.90 3 0.121)/0.803 5 0.44. 33.By usingaverage termination rates, the calculation does not adjust for thefact thatmore recent DI entrantsare likely to have lower mortality and retirementrates than incumbent recipients. Accounting for these cohort effects wouldlikely raise the steady state projection considerably. The projected steady stateseries in Figure VII isimputedin 1997 using the average of 1996 and 1998 todiscount the one-time termination of bene ciarieswhose impairment was Drug andAlcohol Addiction (see Figure I andfootnote 13). 190 QUARTERLY JOURNALOF ECONOMICS

FIGURE IV DIAwards andCurrent andSteady State Recipients per Nonelderly Adult Population,1976 –2001 Source forbene ciaries andawards: SocialSecurity Bulletin: AnnualStatistical Supplement [variousyears] denominated by estimates of adult population ages 25–64from Current PopulationSurvey. Projectedsteady state bene ciaries per populationis annualawards per adult population ages 25 –64dividedby annual terminationrate.

Giventhat 3.7 percentof nonelder ly adults receivedDI benets in 2001, this calculationaugurs another40 percent growthin DIprogramsize relative to population overapproxi- matelythe next decade. Notably, application ratesper popu- lationrose an additional 22 percentbetween 1999 and 2001, indicating thepotential foranother rise in steady statepro- gramsize. The DI recipientpopulation maynever reach this projectedsize, however. As the gureindicates, steady state programsize in 1976 exceeded its 2001 steady statelevel. It was SSA’srecognitionof this fact that galvanizedtheDI pro- gramretrench mentand subsequentliberalizationthat form thebackdrop toour study. DISABILITYAND UNEMPLOYMENT 191

VI. ALTERNATIVE EXPLANATIONS : DECLINING WAGES AND HEALTH, RISING IMMIGRATION, INCARCERATION , AND UNEMPLOYMENT BENEFITS Ourconclusions above must be tempered to the extent that post-1984 increasesin theDI application and laborforce exit propensitiesof low-skilled workersare explained by factorsother than shifts in thesupply ofDI bene ts.Plausible alternative explanations forthe cross-state patterns oflabor force decline that arethe focus of ourstudy includereal wage declines, differ- entialhealth trends, rising immigration and incarcerationrates, and shifts in stateUnemployment Insurance (UI) bene ts. We exploreeach of these alternatives below. Means and cross-state standard deviationsof these measures are given in the rst columnof TableVII. Webegin with wages.As arguedby Juhn [1992] and Juhn, Murphy, and Topel[1991], asizable elasticityof laborforce participation coupledwith decliningreal wages could explain the substantial declinein laborforce participation oflow skilled malesin the1980s and 1990s. Tobe clear,we viewfalling wages as acomplementrather than asubstitutefor the role of the DI programin reducinglabor force participation. Indeed,an important componentof the rise in replacementrates for low-skilled workersis theinteraction between declining wages and thein- dexationand progressivityof theDI bene ts formula.For declining realwages to be the primary explanation forour ndings, however,it would haveto be thecase that, not implausibly, wages fell by substantially morein statesthat had largerdeclines in laborforce participation. Toexplore this hypothesis,we use CPS MergedOutgoing RotationGroup earnings data for1984 and 1998 toconstruct two measuresof the change in theopportunity wage faced by low- skilledworkers in eachstate: the change in thereal xed- weightedmedian log wage of high school dropout males ages 25–64 and thechange in thereal xed-weighted logwage of primeage males 25 –54 at thetwenty- fth percentile. 34 Because thedeclining wage hypothesis implies that workerswith the lowestpotential earnings are least likely to work, we follow

34.In aneffortto maximize the explanatory power of alternativehypotheses, wechose the more inclusive measure of wagesfor male high school dropouts (i.e., ages 25– 64rather than 25 –54).Empirically, the wage measure for the younger group ofhigh school dropouts works similarlybut has less predictive power in laborforce participation regressions. TABLE VII 192 DETERMINANTS OF GROWTH IN DI ROLLS, 1984 –1998 DEPENDENT VARIABLE: 100 3 ANNUALIZED CHANGE IN SHARE OF NONELDERLY ADULTS RECEIVING DI

Means (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)

DI replacement 0.377 0.67 0.46 0.48 0.71 0.49 0.53 0.63 0.54 0.68 0.65 0.50 0.46 rate (0.028) (0.15) (0.17) (0.22) (0.15) (0.15) (0.16) (0.16) (0.14) (0.15) (0.15) (0.16) (0.17) QUARTERLY E[DEmp/Pop] 20.017 21.01 20.66 21.20 20.65 20.41 3 10 (0.017) (0.24) (0.26) (3.07) (0.25) (0.34) E[DEmp/Pop] 20.006 0.14 3 rep. rate 0.007 (0.78)

D HS dropout 20.186 20.06 20.06 20.04 JOURNAL median wage (0.120) (0.03) (0.03) (0.03) D 25th pcntile 0.084 0.10 wage (0.121) (0.03) Mortality rate 0.510 0.15 0.08 FOF ’84 3 100 (0.063) (0.08) (0.08)

D Mortality rate 20.123 0.10 ECONOMICS 3 100 (0.038) (0.12) D Share foreign 0.049 20.36 20.14 born (0.038) (0.09) (0.16) D Incarceration 0.046 20.32 20.14 rate 3 10 (0.020) (0.18) (0.16) D UI replacement 20.025 0.03 rate (0.118) (0.03) Intercept 0.078 20.17 0.10 20.08 20.09 20.19 20.11 20.21 20.15 20.09 20.16 20.16 20.10 20.11 (0.032) (0.06) (0.02) (0.07) (0.08) (0.06) (0.06) (0.06) (0.07) (0.06) (0.06) (0.06) (0.07) (0.07) R2 0.45 0.44 0.53 0.53 0.50 0.56 0.50 0.46 0.59 0.49 0.48 0.57 0.63 N 50 50 50 50 50 50 50 50 50 50 49 50 50 IAIIYDISABILITY

Standard errors are in parentheses (cross-state standard deviations given for column of means). All models control for changes in age by gender shares of state population ages 25–64 (males/females by ages 25–39/40–54/55–64) and are weighted by states’ average share of population in beginning and end year. DI Replacement rate is the seventy-fth percentile DI potential earnings replacement rate of employed nonelderly males calculated from the 1978 Current Population Survey-Social Security Earnings Records Exact Match le. E[DEmp/Pop] is the annualized state projected log employment change for 1984–1998. See equation (8) in the text. AND DMale high school dropout median wage is the change in the xed weighted median wage of all male high school dropout state residents ages 25–64 excluding the self-employed, those working without pay, and employed workers for whom a wage could not be calculated. Weights correspond to 1984 state male population shares for ages 25–64 in ten year brackets (four categories). UNEMPLOYMENT D twenty-fth percentile wage is the analogous x weighted wage for male prime-age residents ages 25–54. Mortality rate is the number of nonaccidental, nonhomicidal deaths among ages 25–64 divided by the average of state population 25–64 in the previous and current year. D Mortality rate is the change in this number. D Share foreign born is the 1984 to 1998 change in share of population 25–64 that is foreign born calculated from the 1998 Current Population Survey and linearly interpolated for the year 1984 using the 1980 and 1990 Census PUMS 1 percent samples. D Incarceration rate is the change in state prison population (excluding jails) per capita 18–64 from the Bureau of Justice Statistics. D UI Replacement Rate is the 1984–1999 change in state UI replacement rate for a full-time minimum wage worker from the Green Book: Overview of Entitlement Programs [U. S. House of Representatives, various years] and is not available for Michigan. 193 194 QUARTERLY JOURNALOF ECONOMICS

Chandra [2002] in including nonparticipants inthequantile calculations,thus assumingthat thepotential earnings of nonparticipants fall belowthe median (or twenty- fth) percentileof the conditionalstate wage distribution. Median potentialwages of malehigh schooldropouts declined 18.6 logpoints between1984 and 1998. In thesame period there was amodestrise at the twenty-fth percentilefor prime age male workers, re ecting strongearnings growth in thelate 1990s [Katz and Krueger 1999]. Althoughaggregate health improved during thetwo decades ofour sample, differential cross-statehealth trends could be in part responsiblefor our ndings. Totest this possibility, we exploitan objectivehealth measure: the mortality rate (excluding accidental deaths and homicides)among state residents ages 25–64.35 Wecontrol for both levels and changesof this measure sinceeach should impact theresponse of state DI rollsto a reductionin DIscreeningstringency. Mortality ratesdeclined by one-quarterbetween 1984 and 1998. Athird potentialconfound is immigration.Because less- educatedimmigrants typically havegreater labor force attach- mentthan native-bornhigh school dropouts [Trejo 2001] and are noteligible for disability bene ts until veyears after arrival, geographicallyconcentrated immigrant in owsduring the1980s and 1990s couldinduce negative cross-state correlations between high schooldropout labor force participation and DIreceipt.We measurenonelderly foreign-born population using data fromthe 1980 and 1990 Censusof Populations and the1998 CPS.The foreign-bornshare of nonelderly adults roseby closeto 5percent- agepoints during 1984 –1998, with sizable cross-statevariation. Afourthconcern is growingincarceration. Because the in- carceratedpopulation is included in neitherthe numerator nor denominatorof our labor force calculations, it is notintrinsically asourceof bias. However,potential criminals have below average ratesof labor force participation relativeto other high school dropouts[Western, Kling, and Weiman2001]. If stateswith low DIgrowthalso experienced differential increasesin incarceration over 1984 –1998, this would bias ourestimates. We assemble Bureauof Justice Statistics data onstate prison populations

35.We focus on an objective measure due to the evidence in Bound and Waidmann[1992, 2000] that male health self-reports closely track expansionsand contractionsin the supply of disabilitybene ts. DISABILITYAND UNEMPLOYMENT 195

(excluding localjails), denominatedby statepopulation ages 18 –64 [U.S.Departmentof Justice2002]. Consistentwith other authors[Freeman 1991; Katz and Krueger1999; Westernand Pettit2000], we nd that incarcerationrates rose by asizable 0.46 percentagepoint during 1984 –1998. Finally,the UI systemprovides an alternative(albeit short- term)source of wagereplacement for displaced workers.Declines in thevalue of UIbene ts during the1980s and 1990s [Anderson and Meyer1997] couldpotentially inducedisplaced workersto seekDI instead.We measure UI generosityas the1984 –1999 changein thestate UI replacementrates for a full-time minimum wageworker using data fromthe U. S.Houseof Representatives Green Book [variousyears]. Although the mean UI replacement ratedid notchange appreciably between1984 and 1998, some statesmade substantial adjustments. Wepresenttwo sets of estimateswith thesevariables. Table VII showsthe relationship between each measure and theannualized percentagepoint growthin DIrollsper nonelderly state population over1984 –1998. TableVIII presentsanalogous esti- matesfor the annualized changein high schooldropout labor forceparticipation inthesame period. Because results are for the mostpart comparablefor the two dependent variables, we discuss themin combination. The rstthree columns in eachtable includeour two main explanatoryvariables: DI replacementrates (as aproxyfor shifts in DI benets supply) and projectedemployment shocks. Each has a signicant impact onDI growthand high schooldropout laborforce nonparticipation. When entered in combination,the projectedshock measure is notprecisely estimated in thelabor forceparticipation equationwhile the replacement rate measure remainssigni cant at the10 percentlevel. Since our model suggeststhat shocksshould havelarger impacts onprogram uptakeand laborforce participation wherereplacement rates are higher,we interact the replacement rate and shockmeasure in column(4). Withonly 50 data points,the interaction term is very impreciselyestimated. Subsequentcolumns of TablesVII and VIII explorethe relationship betweenthe two outcome measures and theother candidate variables,each entered in combinationwith thereplace- mentrate. Higher initial mortality(though not its change), declinesin medianwages of highschool dropouts, and decreases in theshare of foreign born, each contributed to growth in state 196 TABLE VIII DETERMINANTS OF CHANGE IN HIGH SCHOOL DROPOUT LABOR FORCE PARTICIPATION, 1984–1998 DEPENDENT VARIABLE: 10 3 ANNUALIZED CHANGE IN HIGH SCHOOL DROPOUT LABOR FORCE PARTICIPATION RATE

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)

DI replacement 20.35 20.26 20.21 20.33 20.28 20.32 20.34 20.30 20.32 20.38 20.26 20.34 UREL QUARTERLY rate (0.15) (0.15) (0.20) (0.14) (0.15) (0.16) (0.15) (0.14) (0.15) (0.16) (0.15) (0.18) E[DEmp/Pop] 0.55 0.31 20.36 0.31 20.15 3 10 (0.22) (0.23) (2.68) (0.23) (0.36) E[DEmp/Pop] 0.20 3 rep. rate (0.69) D HS dropout 0.08 0.07 0.04 ORA JOURNAL median wage (0.03) (0.03) (0.03) D 25th pcntile 20.07 male wage (0.03) Mortality rate 20.03 0.00 ’84 100 (0.07) (0.06)

3 OF D Mortality rate 20.04

3 100 (0.10) ECONOMICS D Share foreign 0.31 0.30 born (0.09) (0.17) D Incarceration 0.26 0.00 rate 3 10 (0.19) (0.20) D UI replacement 0.02 rate (0.03) Intercept 0.12 0.00 0.10 0.08 0.13 0.10 0.13 0.11 0.10 0.10 0.13 0.10 0.11 (0.06) (0.01) (0.06) (0.08) (0.06) (0.06) (0.06) (0.06) (0.05) (0.06) (0.06) (0.06) (0.06) R2 0.39 0.39 0.42 0.43 0.47 0.44 0.39 0.39 0.51 0.41 0.40 0.49 0.54 – 50 50 50 50 50 50 50 50 50 50 49 50 50

Standard errors are in parentheses. All models control for changes in age distribution of stage high school dropout population ages (25–39/40–54/55–64 with one category omitted) and are weighted by states’ average share of population in beginning and end year. See notes to Table VIII for variable denitions. DISABILITYAND UNEMPLOYMENT 197

DIrollsand declinesin high schooldropout labor force participa- tionduring thesefourteen years. 36 We do not nd a signicant relationshipbetween incarceration or UI replacementrates and eitherDI receiptor highschool dropout labor force participation. Notably,none of these measures greatly affects theestimated impact ofthe replacement rate on growthin DIrollsand declines in laborforce participation. The nal specicationsin eachtable combinegroups of candidate variables shownto besigni cant in previouscolumns. In all cases,DI bene ts supply appears a robustdeterminant of bothgrowth in DIrollsand declinesin high schooldropout labor force participation.

VII. CONCLUSION Followingthe 1984 liberalization ofthe federal Disability Insuranceprogram, the fraction of nonelderlyadults receivingDI benetsroseby 60 percent.Contemporaneously, the DI recipient population becameyounger, less predominantly male, more likely tosuffer fromlow mortality disorders, and far moreencompass- ing ofthelow-skilled population.We tracethese shifts toreduced screeningstringency, rising cash incomereplacement rates, and theincreasing in-kind valueof Medicare bene ts provided.We explorethe impact ofthe supply ofDIbene ts onboththe level of laborforce participation amongthe less skilled and theiremploy- mentresponses to adverse demand shocks.As ourmodel under- scores,because disability application is moreattractive to the unemployed—whointrinsically facelower opportunity costs of exitingthe labor force to seek bene ts—increasesin thesupply of DI benets arelikely to induce differential exitof low-skilled workersfacing employmentdownturns, thereby reducing mea- suredunemployment. Usingstate level data fromthe Current Population Survey paired toSSA administrative data ondisability applications, awards, and receiptby stateand year,we implement two ap- proachesto identify thelabor market impacts ofshifts in the supply ofDI bene ts.Initially, weexploitthe interaction of state disability replacementrates with national changesin program stringencyto identify plausibly exogenousstate-level shifts in the

36.Surprisingly, growth inthe median wage for prime-age males is posi- tively(i.e., wrong) signedin both the DI recipientand labor force participation regressions. 198 QUARTERLY JOURNALOF ECONOMICS supply ofDI bene ts.This analysis nds that statelevel reduc- tionsin bene ts supply induced large(relative) increases in the laborforce participation ofmaleand femalehigh schooldropouts during theDI retrenchment,followed by largedeclines during the DIsubsequentexpansion. No corresponding relationship is ob- servedfor changes in SSIrolls,or for changes in thelabor force participation ofmore-educatedworkers. Wenext investigate whether the declining stringency and growinggenerosity of theDI systemhave hastened the labor force exitof low-skilled workersfacing adversedemand shocks.To distinguish supply fromdemand in this setting,we use cross- statevariation in industrial compositionto identify potentially exogenousshocks to employment of low-skilled workersboth pre- and post-DI reform.Following the liberalization ofDIin 1984, we nd that DIapplication ratesbecame two to three times as responsiveto adverse labor demand shocksof equal magnitude. Simultaneously,male and femalehigh school dropouts became almosttwice as likelyto exit the labor force in theevent of an adverseshock —and correspondinglyless likely to enter unemployment.While our reduced-form approach doesnot quantify the relativeimportance of reducedDI screeningstringency and rising replacementrates in inducing thesebehavioral shifts, simple calculationssuggest that thenet impact ofoutward shifts inthe supply ofDI bene ts is aplausible explanation forboth phenomena. Accountingfor the role of the DI programin inducing labor forceexit among the low-skilled unemployed,our illustrative calculationssuggest that themeasured U. S.unemploymentrate amongadults ages25 – 64 in 1998 would havebeen about half a percentagepoint (13 percent)higher if it had notbeen for the liberalization ofDIin 1984. Ouranalysis con rms the ndings of Juhn [1992] and Juhn,Murphy, and Topel[1991] that declining realwages contributed to the labor force exit of the low skilled in this timeperiod. But decliningwages do not appear asuf cient explanation forour ndings. Controllingfor cross-state variation in thewage declines faced by low-skilled workers,we nd that nonparticipation and disability receiptgrew substantially more in stateswith largershifts in thesupply ofDI bene ts. In the absenceof a signi cant changein DIprogramparameters, the aggregatesize and accompanyinglabor market impact oftheDI programis likelyto grow substantially in thecurrent decade. DISABILITYAND UNEMPLOYMENT 199

DATA APPENDIX

A.1.Employment, Unemployment, and LaborForce Nonparticipation Annual, state-leveldata onemployment, unemployment, and laborforce nonparticipation by gender,age, and educationcate- goryfor individuals betweenthe ages of 25 and 64 werecalculated using thecomplete Current Population Survey monthly les foryears 1978 –1999. Thenumber of observationsranges from 1.1 to1.3 millionannually. All calculationsuse CPS sampling weights.To attain comparableeducational categories (high school dropout,high school graduate, some college, college-plus graduate) acrossthe rede nitionof Census ’s Bureau’seducationvari- ableintroduced in the1992 CPS,we usethe method proposed by Jaeger[1997]. In particular, priorto 1992, wede nehigh school dropoutsas thosewith fewerthan twelveyears of completed schooling.In 1993 forward,we de nehigh schooldropouts as thosewithout a high schooldiploma orGED certi cate.

A.2.Wage Data fromthe CPS Merged Outgoing Rotation Group Files Wecalculate two measures of thechange in theopportunity wagefaced by low-skilled workersin eachstate between 1984 and 1998: thechange in thereal xed-weighted medianlog wage of high schooldropout males ages 25 –64 and thechange in thereal xed-weighted logwage of primeage males 25 –54 at thetwenty- fth percentile.Our sample includes all malesin theseage and educationcategories regardless of employment status excluding theself-employed (for whomwages are not given in theCPS), thoseworking without pay, and employedworkers for whom a wagecould not be calculated (due tousual hoursreported as “variable”).Forworkers employed in thesurvey reference week, wageswere calculated by dividing usual weeklyearnings by usual weeklyhours. Observations were dropped forwhich calculated wagesfell outsidethe range of $1 and $150 perhour (in year 2000 dollars). In caseswhere state-year-age-education cells were empty,we assigned themedian value of correspondingnonempty cells.We use xedage weights corresponding to the shares of males(male high school dropouts) in statepopulation in 1984 ages 25–34, 35–44, 45–54, and 55–64. Wagesare in ated to 2000 valuesusing thechain weightedPersonal Consumption Expen- diture deator. 200 QUARTERLY JOURNALOF ECONOMICS

A.3.State Level DI and SSIRecipientand Bene ts Data Annual, state-leveldata onDI and SSIrecipients,bene t levels,demographics, and qualifying impairmentswere obtained fromvarious years (1978 –2000) ofthe Social Security Adminis- tration’s Annual StatisticalSupplement [variousyears]. DI data includesonly disabled workersreceiving bene ts whereasSSI data includesonly disabled adult bene ciaries(thus excluding child and aged bene ciaries).

A.4.State Level DI and SSIApplication and Award Data Administrative data onunique disability applications and awards by statefor years 1979 –1998 weregenerously provided to usby Kalman Rupp and David Stapleton (DI), Charles Scott (SSI), and Alan Shafer ofthe Social Security Administration. In ourdata set,the year with whichdisability applications and awards areassociated is theyear of application ratherthan the yearof award.

A.5.Demographic Data onDI and SSIRecipientsfrom the Surveyof Income and ProgramParticipation SinceSSA doesnot report the educational distribution ofDI recipients,we use the Survey of Income and ProgramParticipa- tionto estimate these numbers for 1984 and 1999. SIPPdata are unfortunatelynot available forearlier years. Disability receipt ratesby educationcategory, age, and genderwere estimated using data fromthe Survey of Income and ProgramParticipa- tion1984 wave1 and 1996 wave12. Asurveyrespondent is codedas aDIrecipientif she “did receiveincome from Social Securityfor himself/ herselfin this month, ” and herreason for receiptof Social Security was disability. An individual is classi- ed as an SSIrecipientif she “did receiveany incomefrom Supplemental SecurityIncome for him/ herselfduring therefer- enceperiod. ”

A.6.State Level DI ReplacementRates 1978 and National DI ReplacementRates 1979 and 1999 Tocalculate 1978 state-levelpotential DI replacementrates usedin TablesV, VIII, and IX, weuse the CPS SocialSecurity Earnings RecordsExact Match le,which contains Social Secu- rityearnings records in eachyear 1951 –1977 forcurrently employed malesin theCPS sample.This is completedata forDI DISABILITYAND UNEMPLOYMENT 201 benets calculations.We calculate the potential DI replacement ratefor all employedmales ages 25 –61ineachstate in 1978 and usethe seventy- fth percentilein eachstate as ameasureof the potentialreplacement rate faced by lowwage workers. The 1978 CPS-SSER exactmatch leis usedby Bound and Krueger[1991] and was helpfully provided tous by Alan Krueger. Forthe replacement rate calculations in TableI, weutilize annual earningsdata fromthe 1964 –1999 March CPSAnnual DemographicFiles and the1978 CPS SocialSecurity Earnings RecordsExact Match le(CPS-SSER). TheMarch les provide annual earningsdata forthe years 1963 –1998 and theExact Match leis usedto estimate earnings percentiles from 1951 – 1962. In eachyear between 1951 and 1998, wecalculate the tenth,twenty- fth, ftieth,seventy- fth, and ninetiethpercentile ofearnings for males at eachage 25 to61 inclusiveto constructan age-yearearnings lewith 1776 observations(48 years,37 ages). Becauseonly earnings subject to the Social Security payroll tax areused in calculating SocialSecurity bene ts,all earningsper- centilesin excessof the taxable maximumin aparticular yearare truncatedat themaximum. We assume that an Nth percentile workerwho was M yearsold in year T was an Nth percentile M 2 t yearold workerin year T 2 t and estimateAverage Indexed MonthlyEarnings as follows: (14) M 1 YNmt YT22 AIME 5 z z max , 1 . NMT Smin@35,M 2 24#D O S 12 D Y F t G m5max@25,M234#

In this equation, YNmt is equal totaxable earningsfor an m year old Nthpercentileworker in year t where t is equal to T 2 (M 2 m).Theindexing seriesfactor, max[ YT22/Yt,1], is obtained from theSocial Security Administration and is equalto mean national earningsin year T 2 2divided bymeannational earningsin year t (with aminimumvalue of 1). Weuse the median AIME within eachof four age brackets (30 –39, 40 – 49, 50 –54, 55–61), to cal- culatepotential DI bene ts in year T 1 1by applying thePIA formulain equation(2). TheCash IncomeReplacement Rate givenin TableI equals theratio of DI bene ts in year T 1 1 to after-payroll-tax earningsin year T.Calculation for1999 usethe 1965–1999 March CPS les,while the 1979 calculationsutilize boththe 1964 –1979 March CPS lesand the1978 CPS-SSER le (for earningsfrom 1951 –1962). Becausethe Social Security Ad- 202 QUARTERLY JOURNALOF ECONOMICS ministrationonly considered earnings after 1950 in calculating DI benets in 1979, themaximum number of years of prior earningsin the1979 AIME calculationis 28. Ourcalculation also assumesthat individuals havea maximumof 35 yearsof earnings and that lowestearnings years occur before the age of 25. We donot impute wages for individuals whoare out of thelabor force. Tothe extent that malelabor force nonparticipation has in- creaseddifferentially at thelow end of theearnings distribution, ouralgorithm is likelyto understate the actual increasein replacementrates for low-skilled workers. Thesimulated 1979 replacementrates found in TableI closelyaccord with theadministrative guresfrom the CPS- SSER data. Forexample, the 1979 simulatedreplacement rates formales at thetwenty- fth percentileof the earnings distribu- tionfor age groups 30 –39, 40–49, 50–54, and 55–61 are41, 41, 41, and 45 percent.The corresponding replacement rates from the 1978 CPS-SSER data are43, 41, 43, and 45 percent.We use the simulated replacementrates in 1979 forcomparability with thesimulated valuesfor 1999 (forwhich no CPS-SSER data areavailable). Replacementrates that includethe value of Medicare benets in the nal twocolumns of Table I add averageMedicare expendituresfor DI bene ciariesin therelevant year to the DI benetamountand scaleearnings to account for average fringe benetratesfor individuals at eachof the veearnings percentiles(tenth, twenty- fth, ftieth,seventy- fth, and ninetieth)in 1982 and 1996, thetwo years closest to 1979 and 1999 that were available fromBureau of Labor Statistics data. AverageMedicare expendituresfor DI bene ciariesin 1979 were$2890 (in 1998 dollars) and in$4749 in 1998 (themost recent year for which data wereavailable). Averagebene t-wage ratiosin 1982 were0.0923, 0.1911, 0.2512, 0.2780, 0.2801 forthe tenth, twenty- fth, ftieth, seventy-fth,and ninetiethearnings percentiles, while the corresponding ratiosin 1996 were0.0778, 0.1743, 0.2554, 0.2965, 0.2988. Bene t-wage ratiodata werekindly provided by Brooks Pierceof the Bureau of Labor Statistics and arebased upon calculationsin Pierce[2001].

A.7.Additional Variables Usedin SectionVI: Alternative Explanations a.The rate of nonaccidental, nonhomicidal mortality among stateresidents ages 25 –64 is calculated using mortalitydata fromU. S.Department of Health and HumanServices [1984, 1998]. DISABILITYAND UNEMPLOYMENT 203

b.Nonelderly foreign-born populations bystatefor 1984 and 1998 areestimated using data fromthe 1980 and 1990 Censusof Populationsand the1998 CPS.Prior to 1994, theCPS did not collectinformation on nativity, and hencewe linearlyinterpolate 1984 statelevels using the1980 and 1990 IntegratedCensus PublicUse Micro-samples (IPUMS) 1percent les. c.Incarceration rate data arecalculated using Bureauof JusticeStatistics data onstateprison populations excludinglocal jails [U.S.Departmentof Justice 2002] denominatedby state population ages18 – 64 fromthe CPS. d. UIgenerositycorresponds to the 1984 and 1999 potential stateUI replacementrate for a full-time minimumwage worker. Thesedata areobtained fromthe U. S.Houseof Representatives Green Book [Various years].

MASSACHUSETTS INSTITUTE OF TECHNOLOGY DEPARTMENT OF ECONOMICSAND NBER UNIVERSITY OF CHICAGO DEPARTMENT OF ECONOMICSAND NBER

REFERENCES Aarts,Leo J. M.,Richard V.Burkhauser,and Philip R. DeJong, eds., Curing the Dutch Disease (Brookeld,VT: Avebury,1996). Anderson,Patricia M., andBruce D.Meyer, “UnemploymentInsurance Takeup Ratesand the After-Tax Valueof Bene ts,” Quarterly Journalof Economics, CXII(1997), 913 –937 1997. Autor,David H., andMark G.Duggan, “TheRise in Disability Recipiency and the Declinein Unemployment, ” NBER WorkingPaper No. 8336, June 2001. Baker,Michael, Mark Stabile,and Catherine Deri, “WhatDo Self-Reported, ObjectiveMeasures of Health Measure? ” NBER WorkingPaper No. 8419, August 2001. Bartik, Timothy, Who Benets fromState andLocal Economic Development Pol- icies? (Kalamazoo,MI: W.E.Upjohn Institute for Employment Research, 1991). Beaudry,Paul, and John DiNardo, “TheEffect ofImplicit Contracts on the Movementof Wages over the Business Cycle: Evidence from Micro Data, ” Journalof Political Economy, XCIX(1991), 665 – 688. Benitez-Silva,Hugo, Moshe Buchinsky, Hiu Man Chan, So aCheidvasser,and John Rust, “HowLarge Is the Bias in Self-Reported Disability? ” NBER WorkingPaper No. 7526, February 2000. Black, Dan,Kermit Daniel, and Seth Sanders, “TheImpact of Economic Condi- tionson Participationin DisabilityPrograms: Evidence from the Coal Boom and Bust,” American EconomicReview, XCII(2002), 27 –50. Blanchard,Olivier Jean, and Lawrence F. Katz, “RegionalEvolutions, ” Brookings Papers onEconomicActivity (1992), 1–75. Bound,John, “TheHealth and Earnings of RejectedDisability Insurance Appli- cants,” American EconomicReview, LXXIX(1989), 482 –503. ——, “TheHealth and Earnings of Rejected Disability Insurance Applicants: Reply,” American EconomicReview, LXXXI(1991), 1427 –1434. Bound,John, and Richard V.Burkhauser, “EconomicAnalysis of Transfer Pro- gramsTargeted on People with Disabilities, ” inOrley Ashenfelterand David Card, eds., Handbookof Labor Economics, Volume3 (Amsterdam:North- Holland,1999). Bound,John, Richard V.Burkhauser,and Austin Nichols, “Reportto the Social 204 QUARTERLY JOURNALOF ECONOMICS

Security Administration:Tracking theHousehold Income of SSDI andSSI Applicants. ” mimeograph,April 2001. Bound,John, and Harry J.Holzer, “DemandShifts, Population Adjustments, and LaborMarket Outcomes during the1980s, ” Journalof Labor Economics, XVIII(2000), 20 –54. Bound,John, and Alan B. Krueger, “TheExtent ofMeasurement Error inLongi- tudinalEarnings Data: Do Two Wrongs Make a Right? ” Journalof Labor Economics, IX (1991), 1–24. Bound,John, and Timothy Waidmann, “DisabilityTransfers, Self-Reported Health,and the Labor Force Attachment of Older Men: Evidence from the HistoricalRecord, ” Quarterly Journalof Economics, CVII(1992), 1393 –1419. Bound,John, and Timothy Waidmann, “Accountingfor Recent Declines in Em- ploymentRates among the Working-Aged Disabled, ” NBER WorkingPaper No.7975, October 2000. Burkhauser,Richard V.,Mary C.Daly,Andrew J.Houtenville,and Niger Nargis, “EconomicOutcomes of Working-AgePeople with Disabilities over the Busi- nessCycle: An Examinationof the 1980s and 1990s, ” FederalReserve Bank ofSanFrancisco Working Paper No. 2001-17, March 2001. Burkhauser,Richard V.,J.S.Butler,and Robert R. WeathersII, “How Policy VariablesIn uencethe Timing of Social Security DisabilityInsurance Appli- cations,” mimeograph,Cornell University, August 1999. Card, David,and Dean Hyslop, “Does Ination ‘Grease’ theWheels of theLabor Market?” inChristina D. Romerand David H. Romer,eds., Reducing Ina- tion:Motivation and Strategy (Chicagoand London: University of Chicago Press,1997). Chandra,Amitabh, “Isthe Convergence in theRacial Wage Gap Illusory? ” mimeograph,Dartmouth College, January 2002. Cutler,David, and Elizabeth Richardson, “Measuringthe Health of the U. S. Population. ” BrookingsPapers onEconomic Activity: Microeconomics (1997), 217–271. Diamond,Peter, and Eytan Sheshinski, “EconomicAspects of OptimalDisability Benets,” Journalof PublicEconomics, LVII(1995), 1 –23. Freeman,Richard B., “Crimeand the Employment of Disadvantaged Youths, ” NBER WorkingPaper No. 3875, 1991. Gruber,Jonathan, “DisabilityInsurance Bene tsand Labor Supply, ” Journal of Political Economy, CVIII(2000), 1162 –1183. Gruber,Jonathan, and Jeffrey D. Kubik, “DisabilityInsurance Rejection Rates andthe Labor Supply ofOlderWorkers, ” Journalof PublicEconomics, LXIV (1997), 1–23. Hausman,Jerry, and J. Halpern, “Choiceunder Uncertainty: The Decision to Apply forDisability Insurance, ” Journalof Public Economics, XXX,(1986), 131–161. Haveman,Robert, Philip De Jong, and Barbara Wolfe, “DisabilityTransfers and theWork Decisions of Older Men, ” Quarterly Journalof Economics, CVI (1991), 939–949. Hennessey,John, and Janice Dykacz, “AComparisonof the Recovery Termination Ratesof Disabled-Worker Bene ciariesEntitled in 1972 and 1985, ” Social Security Bulletin, LVI(1993), 58 –69. Jacobson,Louis S., RobertJ. LaLonde,and Daniel G. Sullivan, “EarningsLosses ofDisplacedWorkers, ” American EconomicReview, LXXXIII(1993), 685 –709. Jaeger,David A., “Reconcilingthe Old and New Census Bureau Education Ques- tions:Recommendations for Researchers, ” Journalof Business andEconom- ics Statistics, XV(1997),300 –309. Juhn,Chinhui, “Declineof MaleLabor Market Participation: The Role of Declin- ingMarket Opportunities, ” Quarterly Journalof Economics, CVII(1992), 79–121. Juhn,Chinhui, Kevin Murphy, andRobert Topel, “WhyHas the Natural Rate of UnemploymentIncreased Over Time? ” BrookingsPapers onEconomicActiv- ity (1991), 75–142. Kahn,Shulamit, “Evidenceof Nominal Wage Stickiness from Microdata, ” Ameri- can EconomicReview, LXXXVII(1997), 67 –82. Katz, LawrenceF., and David H. Autor, “Changesin the Wage Structure and DISABILITYAND UNEMPLOYMENT 205

EarningsInequality, ” inOrley Ashenfelterand David Card, eds., Handbook ofLaborEconomics, Volume3A (Amsterdam:Elsevier-North-Holland, 1999). Katz, LawrenceF., and Alan B. Krueger, “TheHigh Pressure U. S.LaborMarket ofthe1990s, ” BrookingsPapers onEconomicActivity (1999), 1–87. Kreider,Brent, “SocialSecurity DisabilityInsurance: Applications, Awards and LifetimeIncome Flows, ” Journalof Labor Economics, XVII(1999), 784 – 827. Lewin-VHI,Inc., “LaborMarket Conditions, Socioeconomic Factors, and the Growth ofApplications and Awards forSSDI andSSI Disability Bene ts,” (HHSContract No.100-0012) (Washington, DC: Of ceof the Assistant Sec- retary forPlanning and Evaluation and Social Security Administration, 1995). LewinGroup, “PolicyEvaluation of the Effect ofLegislation Prohibiting the Paymentof Disability Bene tsto Individuals Whose Disability Is Based on Drug Addictionand Alcoholism: Interim Report, ” LewinGroup, Inc., June 21, 1998. Parsons,Donald O., “TheDecline of MaleLabor Force Participation, ” Journal of Political Economy, LXXXVIII(1980), 117 –134. ——, “TheHealth and Earnings of Rejected Disability Insurance Applicants: Comment,” American EconomicReview, LXXXI(1991), 1419 –1426. Pierce,Brooks, “CompensationInequality, ” Quarterly Journalof Economics, CXVI(2001), 1493 –1525. Ruhm,Christopher, “Are RecessionsGood for Your Health? ” Quarterly Journalof Economics, CXV(2000),617 – 650. Rupp,Kalman, and Charles Scott, “Determinantsof Duration on the Disability Rollsand Program Trends. ” inKalman Rupp andDavid Stapleton, eds., Growth inDisability Bene ts: Explanationsand Policy Implications (Kalama- zoo,MI: W.E.Upjohn Institute for Employment Research, 1998), pp. 139 – 176. Rupp,Kalman, and David Stapleton, “Determinantsof theGrowth inthe Social Security Administration ’sDisabilityPrograms: An Overview, ” SocialSecu- rity Bulletin, LVIII(1995), 43 –70. Stapleton,David, Kevin Coleman, Kimberly Dietrich, and Gina Livermore, “Em- piricalAnalyses of DI andSSI Applicationand Award Growth, ” in Kalman Rupp andDavid Stapleton, eds., Growth inDisability Bene ts: Explanations andPolicy Implications (Kalamazoo,MI: W.E.Upjohn Institute for Employ- mentResearch, 1998), pp. 31 –92. Trejo,Stephen, “Selectivityand Immigrant Employment, ” mimeograph,Univer- sityof Texasat Austin, 2001. U.S.Departmentof Health and Human Services, National Center for Health Statistics, “MultipleCause of Death le, 1984,” Inter-UniversityConsortium forPolitical Science Research, Study No.9811, 1984. U.S.Departmentof Health and Human Services, National Center for Health Statistics, “Deathsfrom 282 Selected Causes by 5-Year Age Groups, Race, andSex, 1998, ” March 11,2002, http:/ /www.cdc.gov/nchs,1998. U.S.Departmentof Justice, Of ceof Justice Programs, Bureau of Justice Pro- grams, “Prisonersunder State or Federal Jurisdiction: National Prisoner DataStatistics Series NPS-1, ” March 8,2002, http:/ /www.ojp.usdoj.gov/bjs/ prisons.htm,2002. U.S.Houseof Representatives Committee on Ways and Means, Green Book: Overview ofEntitlement Programs (Washington,DC: U.S.GovernmentPrint- ing Ofce,various years). U.S.SocialSecurity Administration, SocialSecurity Bulletin: AnnualStatistical Supplement, variousyears. Western,Bruce, JeffreyR. Kling,and David F. Weiman, “TheLabor Market Consequencesof Incarceration, ” PrincetonUniversity Industrial Relations Section,Working Paper No. 450, January 2001. Western,Bruce, andBecky Pettit, “Incarcerationand Racial Inequality in Men ’s Employment, ” Industrial andLabor Relations Review, LIV(2000),3 –16.